text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
COARSENING OF GRADED LOCAL COHOMOLOGY
arXiv:1207.2547v2 [math.AC] 4 Jan 2014
FRED ROHRER
Abstract. Some criteria for graded local cohomology to commute with coarsening functors are proven, and an example is given
where graded local cohomology does not commute with coarsening.
Let G be a commutative group, and let R be a G-graded commutative
ring. The category GrModG (R) of G-graded R-modules is Abelian,
fulfils Grothendieck’s axiom AB5, and has a projective generator, hence
it has enough projectives and injectives. The G-graded Hom bifunctor
G
HomR (•, ) maps two G-graded R-modules M and N onto the Ggraded R-module
M
G
HomR (M, N) =
HomGrModG (R) (M, N(g)).
g∈G
This contra-covariant bifunctor is left exact in both arguments, and
it turns into an exact functor when a projective or injective G-graded
R-module is plugged into its first or second argument, respectively.
Thus, by general nonsense we get the G-graded Ext bifunctors, i.e.,
a contra-covariant bifunctor G ExtiR (•, ) such that
for every i ∈
i
(G ExtR (M, •))i∈Z and (G ExtiR (•, M))i∈Z are the right derived cohomological functors of G HomR (M, •) and G HomR (•, M), respectively, for
every G-graded R-module M. Furthermore, if F is a projective system
in GrModG (R) over a right filtering ordered set J, then composition
yields for every i ∈
a functor limJ G ExtiR (F, •), and it follows by
−→
general nonsense that (limJ G ExtiR (F, •))i∈Z is the right derived coho−→
mological functor of limJ G HomR (F, •). In particular, given a graded
−→
ideal a ⊆ R we can consider the projective system (R/an )n∈N over
(where the morphisms are the canonical ones), and then the above construction yields the G-graded local cohomology functors G Hai , i.e., the
right derived cohomological functor of the G-graded a-torsion functor
G
Γa .
The above evokes the natural question: Do we get the same when
we first apply graded local cohomology and then forget the graduation
Z
Z
N
2010 Mathematics Subject Classification. Primary 13D45; Secondary 13A02.
Key words and phrases. Coarsening, graded local cohomology.
The author was supported by the Swiss National Science Foundation.
1
2
FRED ROHRER
on the resulting module, and when we first forget the graduation on a
given module and then apply ungraded local cohomology? The main
reason for this to be nontrivial is that the functor that forgets the
graduation does not necessarily preserve injectivity of objects – see [7,
A.I.2.6] for a counterexample.
In case G = and R is Noetherian, the above question is extensively
discussed and positively answered in [2, Chapter 12]1. Here we study
this question in case G and R are arbitrary, and we present some criteria for a positive answer. The importance of this stems mainly from
toric geometry, where arbitrary groups of degrees (of finite type) are
ubiquitous (see [3]). Since it seems also useful and enlightening, we
consider a more general situation. Namely, we replace the functor that
forgets the graduation by coarsening functors, of which we remind the
reader now.
Let ψ : G ։ H be an epimorphism of groups. We define an H-graded
L
L
ring R[ψ] = h∈H ( g∈ψ−1 (h) Rg ), and analogously, we get a functor
Z
•[ψ] : GrModG (R) → GrModH (R[ψ] ),
called the ψ-coarsening functor. This functor is faithful, conservative,
exact, has a right adjoint, and thus commutes with inductive limits
and with finite projective limits. Our question is now whether graded
local cohomology commutes with coarsening, i.e., whether the diagram
of categories
GrModG (R)
GH i
a
/
GrModG (R)
•[ψ]
•[ψ]
GrModH (R[ψ] )
H Hi
a[ψ]
/
GrModH (R[ψ] )
quasicommutes.
Our first approach uses the above definition of graded local cohomology and hence comes down to the analogous question for graded Ext
functors. Known facts about the analogous question for graded Hom
functors (see [4], [5]) together with δ-functor techniques will allow us
to quickly get our first two criteria (5) and (6). A second approach
uses the relation between local cohomology and Čech cohomology and
will result in a third criterion (7).
We start by defining some canonical morphisms of functors.
1In
fact, in loc. cit. it is supposed that the ungraded ring underlying R is Noetherian, but in this special situation this is by [7, A.II.3.5] equivalent to R being
Noetherian (as a graded ring, i.e., every increasing sequence of graded ideals is
stationary).
COARSENING OF GRADED LOCAL COHOMOLOGY
3
(1) For G-graded R-modules M and N and g ∈ G, we have a monomorphism of groups
HomGrModG (R) (M, N(g)) HomGrModH (R[ψ] ) (M[ψ] , N[ψ] (ψ(g)))
with u 7→ u[ψ] , inducing a monomorphism of H-graded R[ψ] -modules
hψ (M, N) : G HomR (M, N)[ψ] H HomR[ψ] (M[ψ] , N[ψ] ).
Varying M and N, we get a monomorphism of bifunctors
hψ : G HomR (•, )[ψ] H HomR[ψ] (•[ψ] ,
[ψ] ).
If F is a projective system in GrModG (R) over a right filtering ordered
set J, then hψ induces a monomorphism of functors
hF,ψ := lim hψ (F, •) : lim G HomR (F, •)[ψ] lim H HomR[ψ] (F[ψ] , •[ψ] ).
−→
−→
−→
J
J
J
One may note that the above monomorphisms are not necessarily
isomorphisms. An easy example of this phenomenon is obtained by
considering G infinite, H = 0 and R 6= 0 trivially G-graded (i.e., R[0] =
L
R0 ), for then h0 ( g∈G R(g), R) coincides with the canonical injection
R0⊕G R0G . In fact, it follows from [7, A.I.2.11] and [5, 3.4] that if M
is a G-graded R-module, then h0 (M, •) is an isomorphism if and only
if G is finite or M is small2.
(2) If M is a G-graded R-module, then exactness of •[ψ] yields exact
δ-functors
(G ExtiR (M, •)[ψ] )i∈Z ,
(H ExtiR[ψ] (M[ψ] , •[ψ] ))i∈Z ,
(G ExtiR (•, M)[ψ] )i∈Z ,
(H ExtiR[ψ] (•[ψ] , M[ψ] ))i∈Z ,
and implies effaceability of G ExtiR (M, •)[ψ] and G ExtiR (•, M)[ψ] for i >
0. But H ExtiR[ψ] (•[ψ] , M[ψ] ) is effaceable for i > 0, too, for •[ψ] preserves
projectivity ([7, A.I.2.2]). Thus,
(G ExtiR (M, •)[ψ] )i∈Z ,
(G ExtiR (•, M)[ψ] )i∈Z ,
(H ExtiR[ψ] (•[ψ] , M[ψ] ))i∈Z
are the right derived cohomological functors of G HomR (M, •)[ψ] ,
G
HomR (•, M)[ψ] , and H HomR[ψ] (•[ψ] , M[ψ] ), respectively. So, there are
unique morphisms of δ-functors
(hiψ (M, •))i∈Z : (G ExtiR (M, •)[ψ] )i∈Z → (H ExtiR[ψ] (M[ψ] , •[ψ] ))i∈Z
and
(hiψ (•, M))i∈Z : (G ExtiR (•, M)[ψ] )i∈Z → (H ExtiR[ψ] (•[ψ] , M[ψ] ))i∈Z
2A
G-graded R-module M is called small if HomGrModG (R) (M, •) commutes with
direct sums.
4
FRED ROHRER
with h0ψ (M, •) = hψ (M, •) and h0ψ (•, M) = hψ (•, M), and they define
for every i ∈ a morphism of bifunctors
Z
hiψ : G ExtiR (•, )[ψ] → H ExtiR[ψ] (•[ψ] ,
[ψ] )
with h0ψ = hψ ([6, 2.3]). Furthermore, if hψ (•, M) is an isomorphism
then so is (hiψ (•, M))i∈Z , and thus if hψ is an isomorphism then so is
hiψ for every i ∈ .
If F is a projective system in GrModG (R) over a right filtering ordered
set J, then exactness of •[ψ] yields exact δ-functors
Z
(lim G ExtiR (F, •)[ψ] )i∈Z ,
−→
J
(lim H ExtiR[ψ] (F[ψ] , •[ψ] ))i∈Z ,
−→
J
of which the first is universal, hence a unique morphism of δ-functors
(hiF,ψ )i∈Z : (lim G ExtiR (F, •)[ψ] )i∈Z → (lim H ExtiR[ψ] (F[ψ] , •[ψ] ))i∈Z
−→
−→
J
with
h0F,ψ
J
= hF,ψ .
(3) Let a ⊆ R be a graded ideal. We consider the projective system
R/A = (R/an )n∈N in GrModG (R) over . There is by (2) a canonical
morphism of δ-functors
N
(hiR/A,ψ )i∈Z : (G Hai (•)[ψ] )i∈Z → (H Hai[ψ] (•[ψ] ))i∈Z ,
and hR/A,ψ : G Γa (•)[ψ] H Γa[ψ] (•[ψ] ) is the identity morphism.
Related to graded local cohomology are graded higher ideal transformations: The δ-functor (G Dai (•))i∈Z := (limn∈N G ExtiR (an , •))i∈Z is the
−→
right derived cohomological functor of limn∈N G HomR (an , •). It plays
−→
an important role in the relation between local cohomology and sheaf
([2, 20.4.4]), and
cohomology, on projective schemes in case G =
on toric schemes in case of more general groups of degree. Setting
A = (an )n∈N we get from (2) a morphism of δ-functors
Z
(hiA,ψ )i∈Z : (G Dai (•)[ψ] )i∈Z → (H Dai [ψ] (•[ψ] ))i∈Z .
We show now that local cohomology commutes with coarsening if
and only if higher ideal transformation does so.
(4) Proposition The morphism of δ-functors
(hiA,ψ )i∈Z : (G Dai (•)[ψ] )i∈Z → (H Dai [ψ] (•[ψ] ))i∈Z
is an isomorphism if and only if the morphism of δ-functors
(hiR/A,ψ )i∈Z : (G Hai (•)[ψ] )i∈Z → (H Hai[ψ] (•[ψ] ))i∈Z
is so.
COARSENING OF GRADED LOCAL COHOMOLOGY
5
Proof. Analogously to [2, 2.2.4] we get an exact sequence of functors
0 → G Γa → IdGrModG (R) → G Da0 → G Ha1 → 0
and a morphism of δ-functors (ζai )i∈Z : (G Dai )i∈Z → (G Hai+1 )i∈Z such
that ζai is an isomorphism for i > 0. Now the Five Lemma yields the
claim.
(5) Proposition If Ker(ψ) is finite then
(hiR/A,ψ )i∈Z : (G Hai (•)[ψ] )i∈Z → (H Hai[ψ] (•[ψ] ))i∈Z
is an isomorphism.
Proof. By (2) it suffices to show that
hψ : G HomR (•, )[ψ] H HomR[ψ] (•[ψ] ,
[ψ] )
is an epimorphism. Let M and N be G-graded R-modules, and let
f ∈ H HomR[ψ] (M[ψ] , N[ψ] ). By [7, A.I.2.10], there is a finite subset
P
S ⊆ H with f ((M[ψ] )h ) ⊆
l∈S (N[ψ] )h+l for h ∈ H. Finiteness of
−1
Ker(ψ) implies that T := ψ (S) ⊆ G is finite. If g ∈ G then
P
f (Mg ) ⊆ f ((M[ψ] )ψ(g) ) ⊆ l∈S (N[ψ] )ψ(g)+l =
P
P P
P
l∈T (N[ψ] )ψ(g)+ψ(l) =
l∈T
k∈ψ−1 (ψ(g+l)) Nk =
k∈T Ng+k ,
thus f ∈ G HomR (M, N), again by [7, A.I.2.10].
The hypothesis of (5) is fulfilled if G is of finite type and ψ is the
projection onto G modulo its torsion subgroup. In this sense, local
cohomology does not care about torsion in the group of degrees.
(6) Proposition If an has a projective resolution with components
of finite type for every n ∈ , then
N
(hiR/A,ψ )i∈Z : (G Hai (•)[ψ] )i∈Z → (H Hai[ψ] (•[ψ] ))i∈Z
is an isomorphism.
Proof. It suffices to show that hiψ (M, •) is an isomorphism for every
i ∈ and every G-graded R-module M that has a projective resolution
P with components of finite type. As hiψ (M, •) = H i (hψ (P, •)) for
i ∈ , it suffices to show that hψ (M, N) is an isomorphism for every
G-graded R-module M of finite type and every G-graded R-module N.
In this situation, M[ψ] is of finite type, and hence
Z
Z
h0 (M, N) : G HomR (M, N)[0] 0 HomR[0] (M[0] , N[0] )
and
h0 (M[ψ] , N[ψ] ) :
H
HomR[ψ] (M[ψ] , N[ψ] )[0] 0 HomR[0] (M[0] , N[0] )
6
FRED ROHRER
are isomorphisms ([7, A.I.2.11]). It follows that
(hψ (M, N))[0] : (G HomR (M, N)[ψ] )[0] (H HomR[ψ] (M[ψ] , N[ψ] ))[0]
is an isomorphism, too, and since coarsening functors are conservative,
this implies that hψ (M, N) is an isomorphism as desired.3
The hypothesis of (6) is fulfilled if R is coherent and a is of finite
type, and thus in particular if R is Noetherian.
Our third criterion makes use of the so-called ITI property. The
G-graded ring R is said to have ITI with respect to a if the graded
a-torsion submodule G Γa (I) of an injective G-graded R-module I is
injective. This property, strictly weaker than graded Noetherianness,
lies at the heart of many basic properties of local cohomology and thus
is a very natural hypothesis. For more details and examples about ITI,
we refer the reader to [8].
(7) Proposition If a has a finite set E of homogeneous generators
such that R and R[ψ] have ITI with respect to hai for every a ∈ E, then
(hiR/A,ψ )i∈Z : (G Hai (•)[ψ] )i∈Z → (H Hai[ψ] (•[ψ] ))i∈Z
is an isomorphism.
Proof. We choose a counting a = (ai )ni=1 of E and write (G H i (a, •))i∈Z
for the G-graded Čech cohomology with respect to a. This exact
δ-functor is obtained by taking homology of the G-graded Čech cocomplex, and since that cocomplex is obtained by taking direct sums
of modules of fractions with homogeneous denominators, it commutes
with coarsening. As G Ha0 (•) = G H 0(a, •) there is a unique morphism
of δ-functors
(gai )i∈Z : (G Hai (•))i∈Z → (G H i (a, •))i∈Z
with ga0 = IdG Ha0 . So, it suffices to show that (gai )i∈Z is an isomorphism
if R has ITI with respect to hai i for every i ∈ [1, n]. For this it suffices
to show that G H i (a, •) is effaceable for i > 0. But this can be shown
analogously to the ungraded case ([2, 5.1.19])4.
In [10], Schenzel characterised ideals a such that local cohomology
with respect to a is isomorphic to Čech cohomology with respect to
3Using
[5, 3.1; 3.4] one sees that this proof also applies if every power of a has
a projective resolution with small components. But since a projective G-graded
R-module is small if and only if it is of finite type ([1, II.1.2]), this yields no
improvement.
4In loc. cit. one makes use only of the ITI property, but not of Noetherianness.
COARSENING OF GRADED LOCAL COHOMOLOGY
7
a generating system of a. The above proof together with a graded
analogue of Schenzel’s result yields a further coarsening criterion.
We end this note with an example showing that graded local cohomology does not necessarily commute with coarsening.
(8) Example Let K be a field, let R denote the algebra of the additive
monoid ( ≥0 , +) of positive rational numbers over K, furnished with
its canonical -graduation, and let {eα | α ∈ ≥0 } denote its canonical
basis (as a K-vector space). We consider coarsening with respect to
the zero morphism 0 :
→ 0. The graded ideal m := heα | α > 0iR
of R is idempotent, of countable type, but not of finite type, and thus
it is not small by [9, 2◦ ]. Hence, by [5, 3.4] there exists a -graded
R-module N such that the canonical monomorphism
Q
Q
Q
Q
Q
Q HomR (m, N)[0] 0 HomR (m[0] , N[0] )
[0]
is not an epimorphism. By idempotency of m it equals the canonical
morphism Q Dm0 (N)[0] → 0 Dm0 [0] N[0] , and so the exact sequence
0 → Q Γm → IdGrModQ (R) → Q Dm0 → Q Hm1 → 0
together with the Snake Lemma shows that the canonical morphism
Q H 1 (N)[0] → 0 H 1 (N[0] ) is a monomorphism, but not an epimorm
m[0]
phism.
Acknowledgement. I am grateful to Markus Brodmann for his encouraging support during the writing of this article. I also thank the
referee for his careful reading and the suggested improvements.
References
[1] H. Bass. Algebraic K-theory. W. A. Benjamin, New York, 1968.
[2] M. P. Brodmann, R. Y. Sharp. Local cohomology: an algebraic introduction
with geometric applications. Cambridge Stud. Adv. Math. 60, Cambridge University Press, Cambridge, 1998.
[3] D. A. Cox. The homogeneous coordinate ring of a toric variety. J. Algebraic
Geom. 4 (1995), 17–50.
[4] J. L. Gómez Pardo, C. Nǎstǎsescu. Topological aspects of graded rings. Comm.
Algebra 21 (1993), 4481–4493.
[5] J. L. Gómez Pardo, G. Militaru, C. Nǎstǎsescu. When is HOMR (M, −) equal
to HomR (M, −) in the category R-gr? Comm. Algebra 22 (1994), 3171–3181.
[6] A. Grothendieck. Sur quelques points d’algèbre homologique. Tohoku Math. J.
(2) 9 (1957), 119–221.
[7] C. Nǎstǎsescu, F. van Oystaeyen. Graded ring theory. North-Holland Math.
Library 28, North-Holland Publishing Co., Amsterdam, 1982.
[8] P. H. Quy, F. Rohrer. Bad behaviour of injective modules. Preprint (2011).
8
FRED ROHRER
[9] R. Rentschler. Sur les modules M tels que Hom(M, −) commute avec les
sommes directes. C. R. Acad. Sc. Paris 268 (1969), 930–933.
[10] P. Schenzel. Proregular sequences, local cohomology, and completion. Math.
Scand. 92 (2003), 161–180.
Universität Tübingen, Fachbereich Mathematik, Auf der Morgenstelle 10, 72076 Tübingen, Germany
E-mail address: [email protected]
| 0math.AC
|
Rewriting Logic Semantics of a Plan Execution Language ∗
Gilles Dowek
César Muñoz
École polytechnique and INRIA
LIX, École polytechnique
91128 Palaiseau Cedex, France
[email protected]
NASA
Langley Research Center
MS. 130, Hampton, VA 23681, USA
[email protected]
Camilo Rocha
Department of Computer Science
University of Illinois at Urbana-Champaign
201 Goodwin Ave, Urbana, IL 61801, USA
[email protected]
The Plan Execution Interchange Language (PLEXIL) is a synchronous language developed by NASA
to support autonomous spacecraft operations. In this paper, we propose a rewriting logic semantics
of PLEXIL in Maude, a high-performance logical engine. The rewriting logic semantics is by itself a
formal interpreter of the language and can be used as a semantic benchmark for the implementation of
PLEXIL executives. The implementation in Maude has the additional benefit of making available to
PLEXIL designers and developers all the formal analysis and verification tools provided by Maude.
The formalization of the PLEXIL semantics in rewriting logic poses an interesting challenge due to
the synchronous nature of the language and the prioritized rules defining its semantics. To overcome
this difficulty, we propose a general procedure for simulating synchronous set relations in rewriting
logic that is sound and, for deterministic relations, complete. We also report on two issues at the
design level of the original PLEXIL semantics that were identified with the help of the executable
specification in Maude.
1
Introduction
Synchronous languages were introduced in the 1980s to program reactive systems, i.e., systems whose
behavior is determined by their continuous reaction to the environment where they are deployed. Synchronous languages are often used to program embedded applications and automatic control software.
The family of synchronous languages is characterized by the synchronous hypothesis, which states that
a reactive system is arbitrarily fast and able to react immediately in no time to stimuli from the external
environment. One of the main consequences of the synchronous hypothesis is that components running
in parallel are perfectly synchronized and cannot arbitrarily interleave. The implementation of a synchronous language usually requires the simulation of the synchronous semantics into an asynchronous
computation model. This simulation must ensure the validity of the synchronous hypothesis in the target
asynchronous model.
The Plan Execution Interchange Language (PLEXIL) [9] is a synchronous language developed by
NASA to support autonomous spacecraft operations. Space mission operations require flexible, efficient
and reliable plan execution. The computer system on board the spacecraft that executes plans is called the
executive and it is a safety-critical component of the space mission. The Universal Executive (UE) [20]
∗ Authors
in alphabetical order.
B. Klin, P. Sobociński (Eds.):
6th Workshop on Structural Operational Semantics (SOS’09)
EPTCS 18, 2010, pp. 77–91, doi:10.4204/EPTCS.18.6
c G. Dowek & C. Muñoz & C. Rocha
This work is licensed under the
Creative Commons Attribution License.
78
Rewriting Logic Semantics of a Plan Execution Language
is an open source PLEXIL executive developed by NASA1 . PLEXIL and the UE have been used on midsize applications such as robotic rovers and a prototype of a Mars drill, and to demonstrate automation
for the International Space Station.
Given the critical nature of spacecraft operations, PLEXIL’s operational semantics has been formally
defined [8] and several properties of the language, such as determinism and compositionality, have been
mechanically verified [7] in the Prototype Verification System (PVS) [13]. The formal small-step semantics is defined using a compositional layer of five reduction relations on sets of nodes. These nodes are
the building blocks of a PLEXIL plan and represent the hierarchical decomposition of tasks. The atomic
relation describes the execution of an individual node in terms of state transitions triggered by changes
in the environment. The micro relation describes the synchronous reduction of the atomic relation with
respect to the maximal redexes strategy, i.e., the synchronous application of the atomic relation to the
maximal set of nodes of a plan. The remaining three relations are the quiescence relation, the macro
relation and the execution relation which describe the reduction of the micro relation until normalization, the interaction of a plan with the external environment, and the n-iteration of the macro relation
corresponding to n time-steps, respectively. From an operational point of view, PLEXIL is more complex than general purpose synchronous languages such as Esterel [2] or Lustre [4]. PLEXIL is designed
specifically for flexible and reliable command execution in autonomy applications.
In this paper, we propose a rewriting logic semantics of PLEXIL in Maude [5] that complements
the small-step structural operational semantics written in PVS. In contrast to the PVS higher-order logic
specification, the rewriting logic semantics of PLEXIL is executable and it is by itself an interpreter of
the language. This interpreter is intended to be a semantic benchmark for validating the implementation
of PLEXIL executives such as the Universal Executive and a testbed for designers of the language to
study new features or possible variants of the language. Additionally, by using a graphical interface [15],
PLEXIL developers will be able to exploit the formal analysis tools provided by Maude to verify properties of actual plans.
Rewriting logic is a logic of concurrent change in which a wide range of models of computation and
logics can be faithfully represented. The rewriting semantics of a synchronous language such as PLEXIL
poses interesting practical challenges because Maude implements the maximal concurrency of rewrite
rules by interleaving concurrency. That is, although rewriting logic allows for concurrent synchronous
specifications at the mathematical level, Maude executes the rewrite rules by interleaving concurrency.
To overcome this situation, we develop a serialization procedure that allows for the simulation of a
synchronous relation via set rewriting systems. This procedure is presented in a library of abstract set
relations that we have written in PVS. The procedure is sound and complete for the synchronous closure
of any deterministic relation under the maximal redexes strategy.
We are collaborating with the PLEXIL development team at NASA Ames by using the rewriting
logic semantics of PLEXIL to validate the intended semantics of the language against a wide variety of
plan examples. We report on two issues of PLEXIL’s original semantics that were discovered with the
help of the rewriting logic semantics of PLEXIL presented in this paper: the first was found at the level
of the atomic relation for which undesired interleaving semantics were introduced in some computations,
and the second was found at the level of the micro relation for which spurious infinite loops were present
in some computations. Solutions to both issues were provided by the authors, and have been adopted in
the latest version of the PLEXIL semantics.
Summarizing, the contributions presented in this paper are:
• The rewriting logic specification of the PLEXIL semantics.
1 http://plexil.sourceforge.net.
G. Dowek & C. Muñoz & C. Rocha
79
• A library of abstract set relations suitable for the definition and verification of synchronous relations.
• A serialization procedure for the simulation of synchronous relations by rewriting, and an equational version of it in rewriting logic for deterministic synchronous relations.
• The findings on two issues in the design of the original PLEXIL semantics, and the corresponding
solutions that were adopted in an updated version of the language semantics.
Outline of the paper. Background on rewriting logic, and the connection between this logic and Structural Operational Semantics are summarized in Section 2. In Section 3 we present the library of set relations, including the soundness and completeness proof of the serialization procedure. Section 4 describes
the rewriting logic semantics of PLEXIL. In Section 5 we discuss preliminary results. Related work and
concluding remarks are presented in Section 6.
2
Rewriting Logic and Structural Operational Semantics
Rewriting logic [11] is a general semantic framework that unifies in a natural way a wide range of models
of concurrency. Language specifications can be executed in Maude, a high-performance rewriting logic
implementation, and benefit from a wide set of formal analysis tools available to it, such as Maude’s LTL
Model Checker.
A rewriting logic specification or theory is a tuple R = (Σ, E ∪ A, R) where:
• (Σ, E ∪ A) is a membership equational logic theory with Σ a signature having a set of kinds, a
family of sets of operators, and a family of disjoint sets of sorts; E a set of Σ-sentences, which
are universally quantified Horn clauses with atoms that are equations (t = t 0 ) and memberships
(t : s), with t,t 0 terms and s a sort; A a set of “structural” axioms (typically associativity and/or
commutativity and/or identity) such that there exists a matching algorithm modulo A producing a
finite number of A-matching substitutions; and
• R a set of universally quantified conditional rewrite rules of the form
(∀X) r : t −→ t 0 if
^
i
ui = u0i ∧
^
j
vj : sj ∧
^
wl −→ w0l
l
where X is a set of sorted variables, r is a label, t,t 0 , ui , u0i , v j , wl and w0l are terms with variables
among those in X, and s j are sorts.
Intuitively, R specifies a concurrent system, whose states are elements of the initial algebra TΣ/E∪A
specified by the theory (Σ, E ∪ A) and whose concurrent transitions are specified by the rules R. Concurrent transitions are deduced according to the set of inference rules of rewriting logic, which are described
in detail in [3] (together with a precise account of the more general forms of rewrite theories and their
models). Using these inference rules, a rewrite theory R proves a statement of the form (∀X) t −→ t 0 ,
written as R ` (∀X) t −→ t 0 , meaning that, in R, the state term t can transition to the state term t 0 in a
finite number of steps. A detailed discussion of rewriting logic as a unified model of concurrency and its
inference system can be found in [11].
We have a one-step rewrite [t]E∪A −→R [t 0 ]E∪A in R iff we can find a term u ∈ [t]E∪A such that u can
be rewritten to v using some rule r : a −→ b if C ∈ R in the standard way (see [6]), denoted u −→R v,
and we furthermore have v ∈ [t 0 ]E∪A . For arbitrary E and R, whether [t]E∪A −→R [t 0 ]E∪A holds is in
80
Rewriting Logic Semantics of a Plan Execution Language
general undecidable, even when the equations in E are confluent and terminating modulo A. Therefore,
the most useful rewrite theories satisfy additional executability conditions under which we can reduce
the relation [t]E∪A −→R [t 0 ]E∪A to simpler forms of rewriting just modulo A, where both equality modulo
A and matching modulo A are decidable.
The first condition is that E should be terminating and ground confluent modulo A [6]. This means
that in the rewrite theory RE/A = (Σ, A, E), (i) all rewrite sequences terminate, that is, there are no infinite
sequences of the form [t1 ]A −→RE/A [t2 ]A · · · [tn ]A −→RE/A [tn+1 ]A · · · , and (ii) for each [t]A ∈ TΣ/A there is
a unique A-equivalence class [canE/A (t)]A ∈ TΣ/A called the E-canonical form of [t]A modulo A such that
there exists a terminating sequence of zero, one, or more steps [t]A −→∗RE/A [canE/A (t)]A .
The second condition is that the rules R should be coherent [21] relative to the equations E modulo A.
This precisely means that, if we decompose the rewrite theory R = (Σ, E ∪ A, R) into the simpler theories
RE/A = (Σ, A, E) and RR/A = (Σ, A, R), which have decidable rewrite relations −→RE/A and −→RR/A
because of the assumptions of A, then for each A-equivalence class [t]A such that [t]A −→RR/A [t 0 ]A we can
always find a corresponding rewrite [canE/A (t)]A −→RR/A [t 00 ]A such that [canE/A (t 0 )]A = [canE/A (t 00 )]A .
Intuitively, coherence means that we can always adopt the strategy of first simplifying a term to canonical
form with E modulo A, and then apply a rule with R modulo A to achieve the effect of rewriting with R
modulo E ∪ A.
The conceptual distinction between equations and rules has important consequences when giving the
rewriting logic semantics of a language L as a rewrite theory RL = (ΣL , EL ∪ AL , RL ). Rewriting logic’s
abstraction dial [12] captures precisely this conceptual distinction. One of the key features of Structural
Operational Semantics is that it provides a step-by-step formal description of a language’s evaluation
mechanisms [14]. Setting the level of abstraction in which the interleaving behavior of the evaluations
in L is observable, corresponds to the special case in which the dial is turned down to its minimum
position by having EL ∪ AL = 0.
/ The abstraction dial can also be turned up to its maximal position as the
special case in which RL = 0,
/ thus obtaining an equational semantics of the language. In general, we can
make a specification as abstract as we want by identifying a subset R0L ⊆ RL such that the rewrite theory
(ΣL , (EL ∪ R0L ) ∪ AL , RL \ R0L ) satisfies the executability conditions aforementioned. We refer the reader
to [12, 16, 19] for an in-depth presentation of the relationship between structural operational semantics
and rewriting logic semantics, and the use of equations and rules to capture in rewriting logic the dynamic
behavior of language semantics.
The conceptual distinction between equations and rules also has important practical consequences for
program analysis, because it affords massive state space reduction which can make formal analyses such
as breadth-first search and model checking enormously more efficient. Because of state-space explosion,
such analyses could easily become infeasible if we were to use a specification in which all computation
steps are described with rules.
3
A Rewriting Library for Synchronous Relations
When designing a programming language, it is useful to be able to define its semantic relation, to formally prove properties of this relation and to execute it on particular programs. However, defining such a
semantic relation and formally reasoning about it is generally difficult, time consuming, and error-prone.
This would be a major endeavor if it had to be done from scratch for each language. Moreover, since
programming languages tend to evolve constantly, tools must allow reusing parts of former developments
to support rapid yet correct prototyping.
Fortunately, small-step operational semantic relations are, in general, built from simple relations with
G. Dowek & C. Muñoz & C. Rocha
81
a limited number of operations, such as reflexive-transitive extension, reduction to normal form, parallel
extension, etc. As a minimum, the framework should include a library containing the definitions of
these operations and formal proofs of their properties. This will considerably reduce the amount of work
needed to define the semantic relation of particular programming languages and to formally prove their
properties. Defining the semantic relation of synchronous languages requires defining the synchronous
extension of an atomic execution relation, an operation that has been much less studied formally than
other relation operations such as the reflexive-transitive extension or the parallel extension.
We present in this section a first attempt to design a framework for rapid yet correct prototyping
of semantic relations, in particular of synchronous languages. This framework allows one to define
semantic relations, to execute them on particular programs and to formally prove some of their properties
using general theorems about the operations that permit to build relations from relations. We have been
experimenting with this framework using various versions of the PLEXIL language (see Section 4).
The definitions and properties presented in Section 3.1 have been developed in PVS. The Maude
engine is used for executing the semantic relations on particular programs. The full development of the
framework, including the formal semantics of PLEXIL, is available from http://research.nianet.
org/fm-at-nia/PLEXIL.
3.1
Set Relations and Determinism
Let → be a binary relation on a set T . We say that a ∈ T is a redex if there exists a0 ∈ T such that a → a0 ,
and that it is a normal form otherwise. We denote by →0 , →n , and →∗ , the identity relation, n-fold
composition, and reflexive-transitive closure of →, respectively.
In addition to the above relations, we also define the normalized reduction relation →↓ of →.
Normalized reduction a→↓ a0 if and only if a→∗ a0 and a0 is a normal form.
Henceforth, we assume that the relation → is defined on sets over an abstract type T , i.e., → ⊆
2
P(T ) × P(T ). We define the asynchronous extension of →, denoted →, as the congruence closure of
||
→ and the parallel extension of →, denoted →, as the parallel closure of →.
2
Asynchronous extension a → a0 if and only if there exist sets b and b0 such that b ⊆ a, b 6= ∅, b → b0
and a0 = (a \ b) ∪ b0 .
||
Parallel extension a → a0 if and only if there exist b1 , . . . , bn , nonempty, pairwise disjoint subsets of a,
S
S
and sets b01 , . . . , b0n such that bi → b0i and a0 = (a \ i bi ) ∪ i b0i .
The definition of a synchronous reduction requires the definition of a strategy that selects the redexes
to be synchronously reduced.
Strategy A strategy is a function mapping elements a ∈ P(T ) into b1 , . . . , bn , nonempty, pairwise disjoint subsets of a such that all bi are redexes for →.
s
Synchronous extension Let s be a strategy, a → a0 if and only if there exist b01 , . . . , b0n such that s(a) =
S
S
{b1 , . . . , bn }, bi → b0i and a0 = (a \ i bi ) ∪ i b0i .
A natural way of defining strategies is via priorities. A priority is a function p that maps elements
a ∈ P(T ) into natural numbers.
82
Rewriting Logic Semantics of a Plan Execution Language
Maximal redex Let a ∈ P(T ) and let p be a priority function. A nonempty subset b of a is said to be
a maximal redex of a if it is a redex, and for all nonempty subsets c of a such that c is a redex, c 6= b
and c ∩ b 6= ∅, we have p(b) > p(c). By construction, the set of maximal redexes of a set are pairwise
disjoint. The maximal redexes strategy is the function that, given a priority function, maps elements
a ∈ P(T ) into the set of its maximal redexes.
In addition to the definition of the relation operators presented here, our library includes formal proofs
of properties related to determinism and compositionality for abstract set relations. In this paper, we will
focus on determinism as this property is fundamental to the specification of synchronous relations in
rewriting logic.
Determinism A binary relation → defined on a set T is said to be deterministic if for all a, a0 and a00 in
T , a → a0 and a → a00 implies a0 = a00 .
Determinism is a stronger property than confluence, i.e., a deterministic relation is also confluent,
but a confluent relation is not necessarily deterministic.
s
Proposition 3.1 (Determinism of →n , →↓ and →). If the relation → is deterministic, then so are the
s
relations →n , →↓ , and →.
2
In contrast, even if the relation → is deterministic, the relations →∗ , → and →k are not always
deterministic.
3.2
Executing Semantic Relations
Executing the semantic relation of a programming language is desirable during the design phase of the
language. In particular, it allows the designer of the features to experiment with different semantic
variants of the language before implementing them.
Rewrite systems are a computational way of defining binary relations. Since our formalism is based
on set relations, we consider rewrite systems on an algebra of terms of type T modulo associativity,
commutativity, identity, and idempotence: the basic axioms for the union of sets. We denote the equality
on terms of this algebra by =ACUI . The relation → defined by a rewrite system R is defined as follows.
Relation defined by a rewrite system a → b if and only if there exists a rewrite rule l −→ r in R and a
substitution σ such that a =ACUI σ l and b =ACUI σ r.
We remark that the previous definition uses the substitution closure of the rewrite system, rather than
the more traditional definition based on the congruence closure. For example, if we consider the rewrite
system
A(x) −→ B(x),
we have that A(0) → B(0) and A(1) → B(1). On the other hand, A(0), A(1) is not a redex for →.
The synchronous extension of a relation → challenges the standard asynchronous interpretation of
rewrite systems. Consider again the previous example. The asynchronous extension of → defined in Sec2
2
tion 3.1, which indeed encodes the congruence closure, relates A(0), A(1) → B(0), A(1) and A(0), A(1) →
A(0), B(1). However, it does not relate A(0), A(1) to B(0), B(1), which corresponds to the parallel reduc2
2
tion of both A(0) and A(1). In this particular case, we have that A(0), A(1) → B(0), A(1) → B(0), B(1).
s
2∗
We remark that if a → b, for a strategy s, then a→ b. However, in order to select the redexes to
be reduced, we need additional machinery. In particular, we need to keep a log book of redexes that
G. Dowek & C. Muñoz & C. Rocha
83
need to be reduced and redexes that have been already reduced. We propose the following procedure to
implement in an asynchronous rewrite engine, such as Maude, the synchronous extension of a relation
for a strategy.
Serialization procedure Let → be a relation and s a strategy. Given a term a ∈ P(T ), we compute a
term b as follows.
1. Reduce the pair h s(a) ; ∅i to a normal form h∅ ; a0 i using the following rewrite system:
S
hai , c ; di −→ hc ; a0i , di,
where ai → a0i .
2. The term b is defined as (a \ s(a)) ∪ a0 .
S
Since a strategy is a set of redexes, and this set is finite, the procedure is well-defined, i.e., it always
terminates and returns a term. However, the procedure is not necessarily deterministic.
In our previous example, we want to apply the procedure to A(0), A(1), B(1) using the maximal
redexes strategy max p (assuming that all terms have the same priority). Since max p ({A(0), A(1), B(1)}) =
{A(0), A(1)}, we have to reduce the pair hA(0), A(1) ; ∅i to its normal form h∅ ; B(0), B(1)i. Then,
we compute {A(0), A(1), B(1)} \ {A(0), A(1)} ∪ {B(0), B(1)}, which is equal to B(0), B(1). We check
s
that A(0), A(1), B(1) → B(0), B(1).
Theorem 3.2 (Correctness of serialization procedure). The serialization procedure is sound, i.e., if the
s
procedure returns b from a, then a → b. Furthermore, if → is deterministic, the procedure is complete,
s
i.e., if a → b then the procedure returns b from a.
Proof. Soundness Assume that the procedure returns b = a \ s(a) ∪ a0 from a. We have to prove that
s
a → b. Let s(a) = {a1 , . . . , an }, where ai ⊆ a, for 1 ≤ i ≤ n. From the definition of a strategy, the
elements in s(a) are pairwise disjoint. Then, from the procedure, a0 = a01 , . . . , a0n , where ai → a0i ,
for 1 ≤ i ≤ n. Let c ⊆ a be such that none of the subsets of c is in s(a). Then, a has the form a =
s
s
a1 , . . . , an , c. Hence, b = a01 , . . . , a0n , c. By definition of →, we have that a1 , . . . , an , c → a01 , . . . , a0n , c.
S
s
Completeness In this case, it suffices to note that by Proposition 3.1, if → is deterministic, then → is
S
deterministic. Therefore, the normal form of h s(a) ; ∅i is unique and the procedure returns a
S
s
unique b = a \ s(a) ∪ a0 . This b is the only term that is related to a in the relation →.
4
Rewriting Logic Semantics of PLEXIL
The framework presented in Section 3 is abstract with respect to the elements in the set T and the basic set
relation →. If we consider that T is a set of PLEXIL nodes and → is PLEXIL’s atomic relation, we can
deduce by Proposition 3.1 that, since PLEXIL’s atomic relation is deterministic [7] , PLEXIL’s micro
and quiescence relations are deterministic as well. Therefore, we can use the serialization procedure
presented in Section 3.2 to implement a sound and complete formal interpreter of PLEXIL in Maude.
In this section, we describe in detail the specification of such an interpreter. We only discuss the
atomic and micro relations since they are the most interesting ones for validating the synchronous semantics of PLEXIL. More precisely, we present the rewrite theory RPXL = (ΣPXL , EPXL ∪ APXL , RPXL ),
specifying the rewriting logic semantics for PLEXIL’s atomic and micro relations. We use the determinism property of PLEXIL’s atomic relation to encode it as the computation rules in EPXL because
84
Rewriting Logic Semantics of a Plan Execution Language
it yields a confluent equational specification. Consequently, the serialization procedure for PLEXIL’s
synchronous semantics into rewriting logic can be defined equationally, thus avoiding the interleaving
semantics associated with rewrite rules in Maude. Of course, due to the determinism property of the
language, one can as well turn up the “abstraction dial” to its maximum by making the rewrite rules
RPXL into computational rules. This will result in a faster interpreter, for example. Nevertheless, we are
interested in PLEXIL semantics at the observable level of the micro relation. Therefore, in the rewrite
theory RPXL : (i) the equational theory (ΣPXL , EPXL ∪ APXL ) defines the semantics of the atomic relation
and specifies the serialization procedure for the synchronous semantics of PLEXIL, and (ii) the rewrite
rules RPXL define the semantics of the micro relation.
In this section we assume the reader is familiar with the syntax of Maude [5], which is very close to
standard mathematical notation.
4.1
PLEXIL Syntax
A PLEXIL plan is a tree of nodes representing a hierarchical decomposition of tasks. The interior nodes
in a plan provide the control structure and the leaf nodes represent primitive actions. The purpose of
each node determines its type: List nodes group other nodes and provide scope for local variables,
Assignment nodes assign values to variables (they also have a priority, which serves to solve race
conditions between assignment nodes), Command nodes represent calls to commands, and Empty nodes
do nothing. Each PLEXIL node has gate conditions and check conditions. The former specify when the
node should start executing, when it should finish executing, when it should be repeated, and when it
should be skipped. Check conditions specify flags to detect when node execution fails due to violations
of pre-conditions, post-conditions, or invariants. Declared variables in nodes have lexical scope, that is,
they are accessible to the node and all its descendants, but not siblings or ancestors. The execution status
of a node is given by status such as Inactive, Waiting, Executing, etc. The execution state of a plan
consists of (i) the external state corresponding to a set of environment variables accessed through lookups
on environment variables, and (ii) the internal state which is a set of nodes and (declared) variables.
Figure 1 illustrates with a simple example the standard syntax of PLEXIL. In this particular example,
the plan tasks are represented by the root node SafeDrive, the interior node Loop, and the leaf nodes
OneMeter, TakePic and Counter. OneMeter and TakePic are, for example, nodes of type Command.
The node Counter has two different conditions: Start is a gate condition constraining the execution
of the assignment to start only when the node TakePic is in state Finished, while Pre is a check
condition for the number of pictures to be less than 10. The internal state of the plan at a particular
moment is represented by the set of all nodes of the plan, plus the value of the variable pictures, while
the external state of the plan contains the (external) variable WheelStuck.
The external state of a plan is defined in the functional module EXTERNAL-STATE-SYNTAX. The sort
ExternalState represents sets of elements of sort Pair, each of the form (name, value); we assume that
the sorts Name and Value, specifying names and values, respectively, have been defined previously in
the functional modules NAME and VALUE, respectively.
fmod EXTERNAL-STATE-SYNTAX is
protecting Name . protecting Value .
sort Pair .
op (_,_) : Name Value -> Pair .
sort ExternalState .
subsort Pair < ExternalState .
op mtstate : -> ExternalState .
op _,_ : ExternalState ExternalState -> ExternalState [assoc comm id: mtstate] .
eq ES:ExternalState , ES:ExternalState = ES:ExternalState .
endfm
G. Dowek & C. Muñoz & C. Rocha
85
List SafeDrive {
int pictures = 0;
End:
LookupOnChange(WheelStuck) == true OR pictures == 10;
List Loop {
Repeat-while:
LookupOnChange(WheelStuck) == false;
Command OneMeter {
Command: Drive(1);
}
Command TakePic {
Start: OneMeter.status == FINISHED AND pictures < 10;
Command: TakePicture();
}
Assignment Counter {
Start: TakePic.status == FINISHED;
Pre: pictures < 10;
Assignment: pictures := pictures + 1;
}
}
}
Figure 1: SafeDrive: A PLEXIL Plan Example
The internal state of a plan is represented with the help of Maude’s built-in CONF module supporting
object based programming. The internal state has the structure of a set made up of objects and messages,
called configurations in Maude, where the objects represent the nodes and (declared) variables of a plan.
Therefore, we can view the infrastructure of the internal state as a configuration built up by a binary
set union operator with empty syntax, i.e., juxtaposition, as : Configuration × Configuration −→
Configuration. The operator is declared to satisfy the structural laws of associativity and commutativity and to have identity mtconf. Objects and messages are singleton set configurations and belong to
subsorts Object, Msg < Configuration, so that more complex configurations are generated out of them
by set union. An object, representing a node or a (declared) variable, in a given configuration is represented as a term hO : C | a1 : v1 , . . . , an : vn i, where O is the object’s name or identifier (of sort Oid), C is
its class (of sort Cid), the ai ’s are the names of the object’s attribute identifiers, and the vi ’s are the corresponding values. The set of all the attribute-value pairs of an object state (of sort Attribute) is formed
by repeated application of the binary union operator , which also obeys structural laws of associativity,
commutativity, and identity, i.e., the order of the attribute-value pairs of an object is immaterial. The
internal state of a plan is defined in the functional module INTERNAL-STATE-SYNTAX by extending the
sort Configuration; the sorts Exp and Qualified, which we assume to be defined, are used to specify
expressions and qualified names, respectively.
fmod INTERNAL-STATE-SYNTAX is
extending CONFIGURATION . protecting EXP .
protecting QUALIFIED .
subsort Qualified < Oid .
--- Qualified elements are object identifiers
ops List Command Assignment Empty : -> Cid .
--- Types of nodes
sort Status .
ops Inactive Waiting Executing Finishing Failing Finished IterationEnded Variable : -> ExecState .
sort Outcome .
ops None Success Failure : -> Outcome .
op status: : Status -> Attribute .
--- Status of execution
op outcome: : Outcome -> Attribute .
--- Outcome of execution
ops start: skip: repeat: end: : Exp -> Attribute . --- Gate conditions
86
ops
op
op
ops
...
endfm
Rewriting Logic Semantics of a Plan Execution Language
pre: post: inv: : Exp -> Attribute . --- Check conditions
command: : Exp -> Attribute .
--- Command of a command node
assignment: : Exp -> Attribute .
--- Assignment of an assignment node
initval actval: Exp -> Attribute .
--- Initial and actual values of a variable node
Using the infrastructure in INTERNAL-STATE-SYNTAX, the internal state of SafeDrive in Figure 1, is
represented by the configuration in Figure 2. Observe that the sort Qualified provides qualified names
by means of the operator . : Qualified × Qualified −→ Qualified, which we use to maintain the hierarchical structure of the plans. The dots at the end of each object represent the object’s attributes that
are not explicitly defined by the plan but that are always present in each node such as the status or the
outcome. There is a “compilation procedure” from PLEXIL plans to their corresponding representation
in Maude, that we do not discuss in this paper, which includes all implicit elements of a node as attributes
of the object representation of the node.
<
<
<
<
SafeDrive : List | end: LookupOnChange(WheelStuck) == true OR pictures == 10, ... >
Loop . SafeDrive : List | repeat: LookupOnChange(WheelStuck) == false , ... >
OneMetter . Loop . SafeDrive : Command | command: Drive(1), ... >
TakePic . Loop . SafeDrive : Command | start: OneMeter.Status == Finished and pictures < 10,
command: TakePicture(), ... >
< Counter . Loop . SafeDrive : Assignment | pre: pictures < 10,
assignment: pictures := pictures + 1, ... >
< pictures . SafeDrive : Memory | initval: 0, actval: 0 >
Figure 2: SafeDrive in RPXL
We are now ready to define the sort State representing the execution state of the plans in the functional module STATE-SYNTAX, by importing the syntax of external and internal states:
fmod STATE-SYNTAX is
pr EXTERNAL-STATE-SYNTAX .
pr INTERNAL-STATE-SYNTAX .
sort State .
op _|-_ : ExternalState Configuration -> State .
endfm
We adopt the syntax Γ ` π to represent the execution state of the plans, where Γ and π are the external
and internal states, respectively.
4.2
PLEXIL Semantics
PLEXIL execution is driven by external events. The set of events includes events related to lookup in
conditions, e.g., changes in the value of an external state that affects a gate condition, acknowledgments
that a command has been initialized, reception of a value returned by a command, etc. We focus on the
execution semantics of PLEXIL specified in terms of node states and transitions between node states
that are triggered by condition changes (atomic relation) and its synchronous closure under the maximal
redexes strategy (micro relation). PLEXIL’s atomic relation consists of 42 rules, indexed by the type and
the execution status of nodes into a dozen groups. Each group associates a priority to its set of rules
which defines a linear order on the set of rules.
The atomic relation is defined by (Γ, π) ` P −→a P0 , where P ⊆ π. For instance, the four atomic
rules corresponding to the transitions from Executing for nodes of type Assignment are depicted in
G. Dowek & C. Muñoz & C. Rocha
87
Figure 3. Rule r3 updates the status and the outcome of node A to the values IterationEnded and
Success, respectively, and the variable x to the value v, i.e., the value of the expression e in the state π,
whenever the expressions associated with the gate condition End and the check condition Post of node A
both evaluate to true in π. In rule r1 , AncInv(A) is a predicate, parametric in the name of nodes, stating
that none of the ancestors of A has changed the value associated with its invariant condition to false.
The value ⊥ represents the special value “Unknown”. We use (Γ, π) ` e ; v to denote that expression e
evaluates to value v in state Γ ` π; by abuse of notation, we write (Γ, π) ` e 6; v to denote that expression
e does not evaluate to value v in (Γ, π).
(Γ, π) ` AncInv(A) ; false
A.body = x := e
A.type = Assignment
A.status = Executing
r1
(Γ, π) ` Node A −→a Node A with [status = Finished , outcome = Failure , x = ⊥]
(Γ, π) ` A.Invariant ; false
A.body = x := e
A.type = Assignment
A.status = Executing
r2
(Γ, π) ` Node A −→a Node A with [status = IterationEnded , outcome = Failure , x = ⊥]
(Γ, π) ` A.End ; true
A.body = x := e
(Γ, π) ` A.Post ; true
(Γ, π) ` e ; v
A.type = Assignment
A.status = Executing
r3
(Γ, π) ` Node A −→a Node A with [status = IterationEnded , outcome = Success , x = v]
(Γ, π) ` A.End ; true
(Γ, π) ` A.Post 6; true
A.type = Assignment
A.status = Executing
r4
(Γ, π) ` Node A −→a Node A with [status = IterationEnded , outcome = Failure]
{r4 < r3 < r2 < r1 }
Figure 3: Atomic rules corresponding to the transitions from Executing for nodes of type Assignment
The relation r < s between the labels of two different rules specifies that the rule r is only applied
when the second rule s cannot be applied. That is, the binary relation on rules defines the order of their
application when deriving atomic transitions. So, a rule r can be used to derive an atomic transition if
all its premises are valid and no rule higher than r (in its group) is applicable. In the case of PLEXIL’s
atomic relation, the binary relation < on rules is a linear ordering. This linearity is key to the determinism
of PLEXIL (see [8]).
The micro relation Γ ` π −→m π 0 , the synchronous closure of the atomic relation under the maximal
redexes strategy, is defined as:
(Γ, π) ` P1 −→a P10
...
(Γ, π) ` Pn −→a Pn0
Γ ` π −→m (π \
[
1≤i≤n
Pi ) ∪
[
Pi0
M ICRO
1≤i≤n
where Mπ = {P1 , . . . , Pn } is the set of nodes and variables in π that are affected by the micro relation.
If two different processes in π, say A and B, write to the same variable, only the update of the process with higher priority is considered (assignment nodes have an associated priority always), e.g., A if
A.priority > B.priority, B if B.priority > A.priority, and none otherwise.
88
Rewriting Logic Semantics of a Plan Execution Language
In order to specify the PLEXIL semantics in Maude, we first define the infrastructure for the serialization procedure in the functional module SERIALIZATION-INFRASTRUCTURE.
fmod SERIALIZATION-INFRASTRUCTURE is
inc STATE-SYNTAX .
...
op [_:_|_] : Oid Cid AttributeSet -> Object .
--- New syntactic sugar for objects
op updateStatus
: Qualified Status -> Msg .
--- Update status message
op updateOutcome
: Qualified Outcome -> Msg .
--- Update outcome message
op updateVariable : Qualified Value
-> Msg .
--- Update variable message
...
ops applyUpdates unprime : State -> State .
--- Application of updates and ‘unpriming’
var Γ : ExternalState .
var π : InternalState . var A : Oid . var C : Cid .
var Att : AttributeSet . vars S S’ : Status .
var St : State .
eq applyUpdates( Γ ` [ A : C | status: S , Att ] updateStatus(A , S’) π )
= applyUpdates( Γ ` [ A : C | status: S’ , Att ] π ) .
...
eq applyUpdates(St) = St [owise] .
eq unprime( Γ ` [ A : C | Att ] π ) = unprime( Γ ` < A : C | Att > π ) .
eq unprime(St) = St [owise] .
endfm
Following the idea of the serialization procedure, we distinguish between unprimed and primed
redexes by using syntactic sugar for denoting objects in the Maude specification: unprimed redexes
are identified with the already defined syntax for objects in the form of hO : C | ...i and primed redexes are identified with the new syntax for objects in the form of [O : C | ...]. We use messages,
i.e., elements in the sort Msg, to denote the update actions associated with the reduction rules for
the atomic relation; we accumulate these messages in the internal state of the execution state of the
plans, i.e., we also use the internal state in the spirit of the log book of the serialization procedure.
For example, the configuration updateStatus(A,IterationEnded) updateOutcome(A,Success)
updateVariable(x,v) corresponds to the update actions in the conclusion of rule r3 in Figure 3. The
functions applyUpdates and unprimes apply all the collected updates in the internal state, and “unprimes” the “primed” nodes, respectively. In the specification above, it is shown how the status of a node
is updated and how primed nodes become unprimed.
We give the equational serialization procedure in the general setting in which we consider a linear
ordering on the rules.
Equational serialization procedure (with priorities) Let
{ri : (Γ, π) ` Node A −→a Node A with [updatesi ] i f Ci }1≤i≤n
be the collection of atomic rules (in horizontal notation) defining the transition relation for nodes of
type T in status S, with rn < · · · < ri < · · · < r1 , where updatesi is the set of update actions (the
order in the update actions is irrelevant) in the conclusion of ri and Ci is the set of premises of ri . The
equational serialization procedure is given by the following set of equations, in Maude notation, defining
the function symbol, say, r:
var
var
op
eq
=
eq
Γ : ExternalState .
var A : Oid .
var S : Status .
π : Configuration .
var T : Cid .
var Attr : AttributeSet .
r : State -> State .
r( Γ ` < A : T | status: S , Attr > π )
if C1 == true then r( Γ ` [ A : T | status: S , Attr ] messages(updates1 ) π )
else if C2 == true
... else if Cn == true then r( Γ ` [ A : T | status: S , Attr ] messages(updatesn ) π )
else r( Γ ` [ A : T | status: S , Attr ] π ) fi
fi ...
fi .
r( Γ ` π ) = Γ ` π [owise] .
G. Dowek & C. Muñoz & C. Rocha
89
where messages(updadatesi ) represents the message configuration associated with the update actions
in the conclusion of rule ri .
The equational serialization procedure defines a fresh function symbol, say, r : State −→ State. The
first equation for r tries to apply the atomic rules in the given order, by first evaluating the condition
and then marking the affected node. If the condition evaluates to true, then update messages are generated. The second equation, removes the function symbol r when there aren’t any more possible atomic
reductions with the rules {ri }.
The atomic relation is defined in the functional module ATOMIC-RELATION by instantiating the equational serialization procedure for each one of the twelve groups of atomic rules with a different function
symbol for each one.
Finally, the micro relation is defined by the rule micro in the system module PLEXIL-RLS, which
materializes the rewrite theory RPXL in Maude:
mod PLEXIL-RLS is
pr ATOMIC-RELATION .
pr SERIALIZATION-INFRASTRUCTURE .
rl [micro] : Γ ` π => Γ ` unprime(applyUpdates(a1 (...a12 ( Γ ` π )...))) .
endm
where a1 , . . . , a12 are the function symbols in ATOMIC-RELATION defining the serialization procedure
for each one of the twelve groups of rules.
5
Preliminary Results
We have used RPXL to validate the semantics of PLEXIL against a wide variety of plan examples. We
report on the following two issues of the original PLEXIL semantics that were discovered with the help
of RPXL :
1. Non-atomicity of the atomic relation. A prior version of the atomic rules r3 and r4 for Assignment
nodes in state Executing, presented in Figure 3, introduced an undesired interleaving semantics
for variable assignments, which invalidated the synchronous nature of the language.
2. Spurious non-termination of plans. Due to lack of detail in the original specification of some
predicates, there were cases in which some transitions for List nodes in state IterationEnded
would lead to spurious infinite loops.
Although the formal operational semantics of PLEXIL in [8] has been used to prove several properties of PLEXIL, neither one of the issues was previously found. As as matter of fact, these issues do
not compromise any of the proven properties of the language. Solutions to both issues were provided
by the authors, and have been adopted in the latest version of the formal PLEXIL semantics. We are
currently using RPXL as the formal interpreter of PLEXIL’s Formal Interactive Visual Environment [15]
(PLEXIL5), a prototype graphical environment that enables step-by-step execution of plans for scripted
sequence of external events, for further validation of the language’s intended semantics.
We have also developed a variant of RPXL in which the serialization procedure was implemented
with rewrite rules, instead of equations, and rewrite strategies. In general, RPXL outperforms that variant
by two orders of magnitude on average, and by three orders of magnitude in some extreme cases.
The rewrite theory RPXL has approximately 1000 lines of code, of which 308 lines correspond to the
module ATOMIC-RELATION. The rest corresponds to the syntax and infrastructure specifications.
90
6
Rewriting Logic Semantics of a Plan Execution Language
Related Work and Conclusion
Rewriting logic has been used previously as a testbed for specifying and animating the semantics of
synchronous languages. M. AlTurki and J. Meseguer [1] have studied the rewriting logic semantics of
the language Orc, which includes a synchronous reduction relation. T. Serbanuta et al. [17] define the
execution of P-systems with structured data with continuations. The focus of the former is to use rewriting logic to study the (mainly) non-deterministic behavior of Orc programs, while the focus of the latter
is to study the relationship between P-systems and the existing continuation framework for enriching
each with the strong features of the other. Our approach is based more on exploiting the determinism of
a synchronous relation to tackle the problem associated with the interleaving semantics of concurrency
in rewriting logic. P. Lucanu [10] studies the problem of the interleaving semantics of concurrency in
rewriting logic for synchronous systems from the perspective of P-systems. The determinism property
of the synchronous language Esterel [2] was formally proven by O. Tardieu in [18].
We have presented a rewriting logic semantics of PLEXIL, a synchronous plan execution language
developed by NASA to support autonomous spacecraft operations. The rewriting logic specification, a
formal interpreter and a semantic benchmark for validating the semantics of the language, relies on the
determinism of PLEXIL’s atomic relation and a serialization procedure that enables the specification of
a synchronous relation in an asynchronous computational model. Two issues in the original design of
PLEXIL were found with the help of the rewriting logic specification of the language: (i) there was an
atomic rule with the potential to violate the atomicity of the atomic relation, thus voiding the synchronous
nature of the language, and (ii) a set of rules introducing spurious non-terminating executions of plans.
We proposed solutions to these issues that were integrated into the current semantics of the language.
Although we have focused on PLEXIL, the formal framework that we have developed is presented
in a general setting of abstract set relations. In particular, we think that this framework can be applied
to other deterministic synchronous languages. To the best of our knowledge there was no mechanized
library of abstract set relations suitable for the definition and verification of synchronous relations; neither was there a soundness and completeness proof of a serialization procedure for the simulation of
synchronous relations by rewrite systems.
To summarize, we view this work as (i) a step forward in bringing the use of formal methods closer
to practice, (ii) a contribution to the modular and mechanized study of semantic relations, and (iii) yet
another, but interesting contribution to the rewriting logic semantics project.
We intend to continue our collaborative work with PLEXIL development team with the goal of arriving at a formal environment for the validation of PLEXIL. Such an environment would provide a
rich formal tool to PLEXIL enthusiasts for the experimentation, analysis and verification of PLEXIL
programs, which could then be extended towards a rewriting-based PLEXIL implementation with associated analysis tools. Part of our future work is also to investigate the modularity of the equational
serialization procedure with prioritized rules.
Acknowledgments. This work was supported by the National Aeronautics and Space Administration
at Langley Research Center under the Research Cooperative Agreement No. NCC-1-02043 awarded
to the National Institute of Aerospace, while the second author was resident at this institute. The third
author was partially supported by NSF Grant IIS 07-20482. The authors would like to thank the members
of the NASA’s Automation for Operation (A4O) project and, especially, the PLEXIL development team
led by Michael Dalal at NASA Ames, for their technical support.
G. Dowek & C. Muñoz & C. Rocha
91
References
[1] M. AlTurki & J. Meseguer (2008): Reduction Semantics and Formal Analysis of Orc Programs. Electr. Notes
Theor. Comput. Sci. 200(3), pp. 25–41.
[2] G. Berry (2000): The Foundations of Esterel. In: Proof, Language and Interaction: Essays in Honour of
Robin Milner. MIT Press, Cambridge, MA, USA, pp. 425–454.
[3] R. Bruni & J. Meseguer (2006): Semantic foundations for generalized rewrite theories. Theor. Comput. Sci.
360(1-3), pp. 386–414. Available at http://dx.doi.org/10.1016/j.tcs.2006.04.012.
[4] P. Caspi, D. Pilaud, N. Halbwachs & J. A. Plaice (1987): LUSTRE: a declarative language for real-time
programming. In: POPL ’87: Proceedings of the 14th ACM SIGACT-SIGPLAN symposium on Principles
of programming languages. ACM, New York, NY, USA, pp. 178–188.
[5] M. Clavel, F. Durán, S. Eker, J. Meseguer, P. Lincoln, N. Martı́-Oliet & C. Talcott (2007): All About Maude
- A High-Performance Logical Framework. Springer LNCS Vol. 4350, 1st edition.
[6] N. Dershowitz & J. P. Jouannaud (1990): Rewrite Systems. In: Handbook of Theoretical Computer Science,
Volume B: Formal Models and Sematics (B). The MIT Press, pp. 243–320.
[7] G. Dowek, C. Muñoz & C. Păsăreanu (2007): A Formal Analysis Framework for PLEXIL. In: Proceedings
of 3rd Workshop on Planning and Plan Execution for Real-World Systems. pp. 45–51.
[8] G. Dowek, C. Muñoz & C. Păsăreanu (2008): A Small-Step Semantics OF PLEXIL. Technical Report 200811, National Institute of Aerospace, Hampton, VA.
[9] T. Estlin, A. Jónsson, C. Păsăreanu, R. Simmons, K. Tso & V. Verna (2006): Plan Execution Interchange
Language (PLEXIL). Technical Memorandum TM-2006-213483, NASA.
[10] D. Lucanu (2009): Strategy-Based Rewrite Semantics for Membrane Systems Preserves Maximal Concurrency of Evolution Rule Actions. Electr. Notes Theor. Comput. Sci. 237, pp. 107–125.
[11] J. Meseguer (1992): Conditional Rewriting Logic as a Unified Model of Concurrency. Theoretical Computer
Science 96(1), pp. 73–155.
[12] J. Meseguer & G. Rosu (2007): The rewriting logic semantics project. Theor. Comput. Sci. 373(3), pp.
213–237. Available at http://dx.doi.org/10.1016/j.tcs.2006.12.018.
[13] S. Owre, J. Rushby & N. Shankar (1992): PVS: A Prototype Verification System. In: Deepak Kapur, editor:
11th International Conference on Automated Deduction (CADE), Lecture Notes in Artificial Intelligence
607. Springer-Verlag, Saratoga, NY, pp. 748–752.
[14] G. D. Plotkin (2004): A structural approach to operational semantics. J. Log. Alg. Prog. 60-61, pp. 17–139.
[15] C. Rocha, C. Muñoz & H. Cadavid (2009): A Graphical Environment for the Semantic Validation of a
Plan Execution Language. IEEE International Conference on Space Mission Challenges for Information
Technology 0, pp. 201–207.
[16] T. Serbanuta, G. Rosu & J. Meseguer (2009): A rewriting logic approach to operational semantics. Inf.
Comput. 207(2), pp. 305–340.
[17] T. Serbanuta, G. Stefanescu & G. Rosu (2008): Defining and Executing P Systems with Structured Data in
K. In: David W. Corne, Pierluigi Frisco, Gheorghe Paun, Grzegorz Rozenberg & Arto Salomaa, editors:
Workshop on Membrane Computing, Lecture Notes in Computer Science 5391. Springer, pp. 374–393.
[18] O. Tardieu (2007): A deterministic logical semantics for pure Esterel. ACM Trans. Program. Lang. Syst.
29(2), p. 8.
[19] A. Verdejo & N. Martı́-Oliet (2006): Executable structural operational semantics in Maude. J. Log. Algebr.
Program. 67(1-2), pp. 226–293.
[20] V. Verna, A. Jónsson, C. Păsăreanu & M. Latauro (2006): Universal Executive and PLEXIL: Engine and
Language for Robust Spacecraft Control and Operations. In: Proceedings of the American Institute of
Aeronautics and Astronautics Space Conference.
[21] P. Viry (2002): Equational rules for rewriting logic. Theoretical Computer Science 285, pp. 487–517.
| 6cs.PL
|
JS KETCH: Sketching for Java
Jinseong Jeon†
Xiaokang Qiu‡
†
Jeffrey S. Foster†
Armando Solar-Lezama‡
‡
University of Maryland, College Park
MIT CSAIL
{[email protected], [email protected], [email protected], [email protected]}
arXiv:1507.03577v1 [cs.PL] 13 Jul 2015
Abstract
JSketch
Sketch-based synthesis, epitomized by the S KETCH tool, lets developers synthesize software starting from a partial program, also
called a sketch or template. This paper presents JS KETCH, a tool
that brings sketch-based synthesis to Java. JS KETCH’s input is a
partial Java program that may include holes, which are unknown
constants, expression generators, which range over sets of expressions, and class generators, which are partial classes. JS KETCH
then translates the synthesis problem into a S KETCH problem; this
translation is complex because S KETCH is not object-oriented. Finally, JS KETCH synthesizes an executable Java program by interpreting the output of S KETCH.
.jsk
Parser
AST
.out
Encoder
Sketch
Solver
Figure 1: JS KETCH Overview.
General Terms Design, Languages.
overloading and overriding, anonymous/inner classes—in terms
of the features available in S KETCH. Section 3 briefly explains
several technical challenges addressed in JS KETCH. Section 4 describes our experience with JS KETCH. JS KETCH is available at
http://plum-umd.github.io/java-sketch/.
Keywords Program Synthesis, Programming by Example, Java,
S KETCH, Input-output Examples.
Introduction
Program synthesis [5, 6] is an attractive programming paradigm
that aims to automate the development of complex pieces of code.
Deriving programs completely from scratch given only a declarative specification is very challenging for all but the simplest algorithms, but recent work has shown that the problem can be made
tractable by starting from a partial program—referred to in the literature as a sketch [11], scaffold [13] or template—that constrains the
space of possible programs the synthesizer needs to consider. This
approach to synthesis has proven useful in a variety of domains
including program inversion [14], program deobfuscation [4], development of concurrent data-structures [12] and even automated
tutoring [8].
This paper presents JS KETCH, a tool that makes sketch-based
synthesis directly available to Java programmers. JS KETCH is built
as a frontend on top of the S KETCH synthesis system, a mature synthesis tool based on a simple imperative language that can generate
C code [11]. JS KETCH allows Java programmers to use many of
the S KETCH’s synthesis features, such as the ability to write code
with unknown constants (holes written ??) and unknown expressions described by a generator (written {| e∗ |}). In addition, JSKETCH provides a new synthesis feature—a class-level generator—
that is specifically tailored for object oriented programs. Section 2
walks through JS KETCH’s input and output, along with a running
example.
As illustrated in Figure 1, JS KETCH compiles a Java program with unknowns to a partial program in the S KETCH language and then maps the result of S KETCH synthesis back to
Java. The translation to S KETCH is challenging because S KETCH
is not object oriented, so the translator must model the complex object-oriented features in Java—such as inheritance, method
.java
.sk
Categories and Subject Descriptors I.2.2 [Automatic Programming]: Program Synthesis; F.3.1 [Specifying and Verifying and
Reasoning about Programs]: Assertions, Specification techniques
1.
Decoder
2.
Overview
We begin our presentation with two examples showing JS KETCH’s
key features and usage.
2.1
Basics
The input to JS KETCH is an ordinary Java program that may
also contain unknowns to be synthesized. There are two kinds
of unknowns: holes, written ??, represent unknown integers and
booleans, and generators, written {| e∗ |}, range over a list of expressions. For example, consider the following Java sketch1 , similar
to an example from the S KETCH manual [10]:
2
class SimpleMath {
static int mult2(int x) { return (?? ∗ {| x , 0 |}); }
3
}
1
Here we have provided a template for the implementation of
method mult2: The method returns the product of a hole and either parameter x or 0. Notice that even this very simple sketch has
233 possible instantiations (32 bits of the hole and one bit for the
choice of x or 0).
To specify the solution we would like to synthesize, we provide
a harness containing assertions about the mult2 method:
5
class Test {
harness static void test() { assert(SimpleMath.mult2(3) == 6); }
6
}
4
1 https://github.com/plum-umd/java-sketch/blob/
master/test/benchmarks/t109-mult2.java
Now we can run JS KETCH on the sketch and harness.
$ ./jsk.sh SimpleMath.java Test.java
The result is a valid Java source file in which holes and generators
have been replaced with the appropriate code.
7
8
9
10
11
$ cat result/java/SimpleMath.java
class SimpleMath { ...
static public int mult2 (int x) {
return 2 * x;
}
}
12
13
14
15
16
interface Token{ public int getId (); }
generator class Automaton {
private int state = ??;
static int num state = ??;
harness static void min num state() { minimize(num state); }
public void transition (Token t) {
assert 0 ≤ state && state < num state;
int id = t .getId ();
minrepeat {
if (state == ?? && id == ??) { state = ??; return; }
} }
public void transitions ( Iterator <Token> it) {
while ( it .hasNext()) { transition ( it .next ()); }
}
public boolean accept() { return state ≤ ??; }
17
2.2
Finite Automata
Now consider a harder problem: suppose we want to synthesize
a finite automaton given sample accepting and rejecting inputs.2
There are many possible design choices for finite automata in an
object-oriented language, and we will opt for one of the more
efficient ones: the current automaton state will simply be an integer,
and a series of conditionals will encode the transition function.
Figure 2a shows our automaton sketch. The input to the automaton will be a sequence of Tokens, which have a getId method returning an integer (line 8). An Automaton is a class—ignore the generator keyword for the moment—with fields for the current state
(line 9) and the number of states (line 10). Notice these fields are
initialized to holes, and thus the automaton can start from any arbitrary state and have an arbitrary yet minimal number of states
(restricted by S KETCH’s minimize function on line 11). The class
includes a transition function that asserts that the current state is inbounds (line 13) and updates state according to the current state and
the input Token’s value (retrieved on line 14).
Here we face a challenge, however: we do not know the number
of automaton states or tokens, so we have no bound on the number of transitions. To solve this problem, we use a feature that JSKETCH inherits from S KETCH : the term minrepeat { e } expands to
the minimum length sequence of e’s that satisfy the harness. In this
case, the body of minrepeat (line 16) is a conditional that encodes
an arbitrary transition—if the guard matches the current state and
input token, then the state is updated and the method returns. Thus,
the transition method will be synthesized to include however many
transitions are necessary.
Finally, the Automaton class has methods transitions and accept;
the first performs multiple transitions based on a sequence of input
tokens, and the second one determines whether the automaton is
in an accepting state. Notice that the inequality (line 21) means
that states 0 up to some bound will be accepting; this is fully
general because the exact state numbering does not matter, so the
synthesizer can choose the accepting states to follow this pattern.
Class Generators. In addition to basic S KETCH generators like
we saw in the mult2 example, JS KETCH also supports class generators, which allow the same class to be instantiated differently
in different superclass contexts. In Figure 2a, the generator annotation on line 8 indicates that Automaton is such a class. (Class generators are analogous to the the function generators introduced by
S KETCH [10].)
Figure 2b shows two classes that inherit from Automaton. The
first class, DBConnection, has an inner class Monitor that inherits
from Automaton. The Monitor class defines two tokens, OPEN and
CLOSE, whose ids are 1 and 2, respectively. The outer class has
a Monitor instance m that transitions when the database is opened
(line 34) and when the database is closed (line 35). The goal is to
synthesize m such that it acts as an inline reference monitor to check
that the database is never opened or closed twice in a row, and is
2 Of
course, there are many better ways to construct finite automata—this
example is only for expository purposes.
18
19
20
21
22
}
(a) Automaton sketch.
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
class DBConnection {
class Monitor extends Automaton {
final static Token OPEN =
new Token() { public int getId() { return 1; } };
final static Token CLOSE =
new Token() { public int getId() { return 2; } };
public Monitor() { }
}
Monitor m;
public DBConnection() { m = new Monitor(); }
public boolean isErroneous() { return ! m.accept(); }
public void open() { m.transition (Monitor.OPEN); }
public void close() { m. transition (Monitor.CLOSE); }
}
class CADsR extends Automaton { ...
public boolean accept(String str) {
state = init state backup ;
transitions ( convertToIterator ( str ));
return accept();
} }
(b) Code using Automaton sketch.
Figure 2: Finite automata with JS KETCH.
only closed after it is opened. The harnesses in TestDBConnection in
Figure 3 describe both good and bad behaviors.
The second class in Figure 2b, CADsR, adds a new (overloaded)
accept(String) method that converts the input String to a token iterator (details omitted for brevity), transitions according to that iterator, and then returns whether the string is accepted. The goal is
to synthesize an automaton that recognizes c(a|d)+r. The corresponding harness TestCADsR.examples() in Figure 3 constructs a
CADsR instance and makes various assertions about its behavior.
Notice that this example relies critically on class generators, since
Monitor and CADsR must encode different automata.
Output. Figure 4 shows the output produced by running JSKETCH on the code in Figures 2 and 3. We see that the generator
was instantiated as Automaton1, inherited by DBConnection.Monitor,
and Automaton2, inherited by CADsR. Both automata are equivalent to what we would expect for these languages. Two things were
critical for achieving this result: minimizing the number of states
(line 11) and having sufficient harnesses (Figure 3).
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
We experimented further with CADsR to see how changing
the sketch and harness affects the output. First, we tried running
with a smaller harness, i.e., with fewer examples. In this case,
the synthesized automaton covers all the examples but not the
full language. For example, if we omit the four-letter inputs in
Figure 3 the resulting automaton only accepts three-letter inputs.
Whereas going to four-letter inputs constrains the problem enough
for JS KETCH to find the full solution.
Second, if we omit state minimization (line 11), then the synthesizer chooses large, widely separated indexes for states, and it also
includes redundant states (that could be merged with a textbook
state minimization algorithm).
Third, if we manually bound the number of states to be too small
(e.g., manually set num state to 2), the synthesizer runs for more
than half an hour and then fails, since there is no possible solution.
Of these cases, the last two are relatively easy to deal with since
the failure is obvious, but the first one—knowing that a synthesis
problem is underconstrained—is an open research challenge. However, one good feature of synthesis is that, if we do find cases that
are not handled by the current implementation, we can simply add
those cases and resynthesize rather than having to manually fix the
code (which could be quite difficult and/or introduce its own bugs).
Moreover, minimization—trying to ensure the output program is
small—seems to be a good heuristic to avoid overfitting to the examples.
class TestDBConnection {
harness static void scenario good() {
DBConnection conn = new DBConnection();
assert ! conn.isErrorneous();
conn.open(); assert ! conn.isErroneous();
conn.close(); assert ! conn.isErroneous(); }
// bad: opening more than once
harness static void scenario bad1() {
DBConnection conn = new DBConnection();
conn.open(); conn.open(); assert conn.isErroneous(); }
// bad: closing more than once
harness static void scenario bad2() {
DBConnection conn = new DBConnection();
conn.open();
conn.close(); conn.close(); assert conn.isErroneous();
} }
class TestCADsR {
// Lisp−style identifier : c(a|d)+r
harness static void examples() {
CADsR a = new CADsR();
assert ! a.accept(”c” ); assert ! a.accept(”cr” );
assert a.accept(”car” ); assert a.accept(”cdr” );
assert a.accept(”caar” ); assert a.accept(”cadr” );
assert a.accept(”cdar” ); assert a.accept(”cddr” );
} }
3.
We implemented JS KETCH as a series of Python scripts that invokes S KETCH as a subroutine. JS KETCH comprises roughly 5.7K
lines of code, excluding the parser and regression testing code. JSKETCH parses input files using the Python front-end of ANTLR
v3.1.3 [7] and its standard Java grammar. We extended that grammar to support holes, expression-level generators, minrepeat, and
the harness and generator modifiers.
There are a number of technical challenges in the implementation of JS KETCH; due to space limitations we discuss only the
major ones.
Figure 3: Automata use cases.
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
class Automaton1 {
int state = 0; static int num state = 3;
public void transition (Token t) { ...
assert 0 ≤ state && state < 3;
if (state == 0 && id == 1) { state = 1;
if (state == 1 && id == 1) { state = 2;
if (state == 1 && id == 2) { state = 0;
if (state == 0 && id == 2) { state = 2;
return;
return;
return;
return;
Implementation
}
}
}
}
//
//
//
//
open@
open 2x
( init )@
close 2x
}
public boolean accept() { return state ≤ 1; } ...
}
class DBConnection{ class Monitor extends Automaton1 { ... } ...}
class Automaton2 {
int state = 0; static int num state = 3;
public void transition (Token t) { ...
assert 0 ≤ state && state < 3;
if (state == 0 && id == 99) { state = 1; return; } // c
if (state == 1 && id == 97) { state = 2; return; } // ca
if (state == 1 && id == 100) { state = 2; return; } // cd
if (state == 2 && id == 114) { state = 0; return; } // c(a|d)+r@
}
public boolean accept() { return state ≤ 0; } ...
}
class CADsR extends Automaton2 { ... }
Figure 4: JS KETCH Output (partial).
Class hierarchy. The first issue we face is that S KETCH’s language is not object-oriented. To solve this problem, JS KETCH follows a similar approach to [8] and encodes objects with a new type
V Object, defined as a struct containing all possible fields plus an
integer identifier for the class. More precisely, if C1 , . . . , Cm are
all classes in the program, then we define:
93
struct V Object {
int class id ; fields-from-C1 ... fields-from-Cm
94
}
92
where each Ci gets its own unique id.
JS KETCH also assigns every method a unique id, and it creates
various constant arrays that record type information. For a method
id m, we set belongsTo[m] to be its class id; argNum[m] to be its
number of arguments; and argType[m][i] to be the type of its ith argument. We model the inheritance hierarchy using a twodimensional array subcls such that subcls[i][j] is true if class i is a
subclass of class j.
Encoding names. When we translate the class hierarchy into
JS KETCH, we also flatten out the namespace, and we need to
avoid conflating overridden or overloaded method names, or inner
classes.
Thus, we name inner classes as Inner Outer , where Inner is the
name of the nested class and Outer is the name of the enclosing
class. We also handle anonymous classes by assigning them distinct
numbers, e.g., Cls 1.
To support method overriding and overloading, methods are
named Mtd Cls Params, where Mtd is the name of the method, Cls
is the name of the class in which it is declared, and Params is the
list of parameter types. For example, in the finite automaton example, CADsR inherits method transition from Automaton2 (the second
variant of the class generator), hence the method is named transition Automaton2 Token(V Object self, V Object t) in S KETCH. The first
parameter represents the callee of the method.
Dynamic dispatch. We simulate the dynamic dispatch mechanism of Java in S KETCH. For each method name M (suitably encoded, as above), we introduce a function dyn dispatch M(V Object
self, ...) that dispatches based on the class id field of the callee:
100
void dyn dispatch M(V Object self, ...) {
int cid = self . class id ;
if (cid == R0 id) return M R0 P(self, ...);
if (cid == R1 id) return M R1 P(self, ...);
...
return;
101
}
95
96
97
98
99
Note that if M is static, the self argument is omitted.
Java libraries. To perform synthesis, we need S KETCH equivalents of any Java standard libraries used in the input sketch.
Currently, JS KETCH supports the following collections and APIs:
ArrayDeque, Iterator, LinkedList, List, Map, Queue, Stack, TreeMap,
CharSequence, String, StringBuilder, and StringBuffer. Library classes
are implemented using a combination of translation of the original
source using JS KETCH and manual coding, to handle native methods or cases when efficiency is an issue. Note that several of these
classes include generics (e.g., List), which is naturally handled because the all objects are uniformly represented as V Object.
Limitations and unsupported features. As Java is a very large
language, JS KETCH currently only supports a core subset of Java.
We leave several features of Java to the future versions of JSKETCH , including packages, access control, exceptions, and concurrency.
Additionally, JS KETCH assumes the input sketch is type correct,
meaning the standard parts of the program are type correct, holes
are used either as integers or booleans, and expression generators
are type correct. This assumption is necessary because, although
S KETCH itself includes static type checking, distinctions between
different object types are lost by collapsing them all into V Object.
Using S KETCH. We translate JS KETCH file, which is composed
of the user-given template and examples, as well as supportive
libraries (if necessary) to .sk files as input to S KETCH. For example,
the SimpleMath example in Section 2.1 is translated to
102
103
104
int e h1 = ??;
int mult2 SimpleMath int(int x) { return e h1 ∗ {| x | 0 |}; }
harness void test Test() { assert mult2 SimpleMath int(3) == 6; }
We refer the reader elsewhere [9, 10] for details on how S KETCH
itself works.
After solving the synthesis problem, JS KETCH then unparses
these same Java files, but with unknowns resolved according to the
S KETCH synthesis results. We use partial parsing [1] to make this
output process simpler.
4.
Experience with JS KETCH
We developed JS KETCH as part of the development of another tool,
PASKET [2], which aims to construct framework models, e.g. mock
classes that implement key functionality of a framework but in a
way that is much simpler than the actual framework code and is
more amenable to static analysis. PASKET takes as input a log of
the interaction between the real framework and a test application,
together with a description of the API of the framework and of
the design patterns that the framework uses. PASKET uses these
inputs to automatically generate an input to JS KETCH which is
then responsible for actually synthesizing the models. Through
PASKET, we have used JS KETCH to synthesize models of key
functionality from the Swing and Android frameworks. The largest
JS KETCH inputs generated by PASKET contain 117 classes and
4,372 lines of code, and solve in about two minutes despite having
over 7318 × 16428 possible choices; this is possible thanks to a
new synthesis algorithm called Adaptive Concretization [3] that is
available in S KETCH and was also developed as part of this work.
Acknowledgments
This research was supported in part by NSF CCF-1139021, CCF1139056, CCF-1161775, and the partnership between UMIACS
and the Laboratory for Telecommunication Sciences.
References
[1] A. Demaille, R. Levillain, and B. Sigoure. TWEAST: A Simple and
Effective Technique to Implement Concrete-syntax AST Rewriting
Using Partial Parsing. In SAC ’09, pages 1924–1929, 2009.
[2] J. Jeon, X. Qiu, J. S. Foster, and A. Solar-Lezama. Synthesizing
Framework Models for Symbolic Execution. Unpublished manuscript,
2015.
[3] J. Jeon, X. Qiu, A. Solar-Lezama, and J. S. Foster. Adaptive Concretization for Parallel Program Synthesis. In Computer Aided Verification (CAV ’15), July 2015.
[4] S. Jha, S. Gulwani, S. A. Seshia, and A. Tiwari. Oracle-guided
component-based program synthesis. In ICSE ’10, pages 215–224,
2010.
[5] Z. Manna and R. Waldinger. A Deductive Approach to Program Synthesis. ACM Transactions on Programming Languages and Systems,
2(1):90–121, Jan. 1980.
[6] Z. Manna and R. J. Waldinger. Toward automatic program synthesis.
Communications of the ACM, 14(3):151–165, Mar. 1971.
[7] T. Parr and K. Fisher. LL(*): The Foundation of the ANTLR Parser
Generator. In PLDI ’11, pages 425–436, 2011.
[8] R. Singh, S. Gulwani, and A. Solar-Lezama. Automated Feedback
Generation for Introductory Programming Assignments. In PLDI ’13,
pages 15–26, 2013.
[9] A. Solar-Lezama. Program sketching. International Journal on Software Tools for Technology Transfer, 15(5-6):475–495, 2013.
[10] A. Solar-Lezama. The Sketch Programmers Manual, 2015. Version
1.6.7.
[11] A. Solar-Lezama, L. Tancau, R. Bodik, S. Seshia, and V. Saraswat.
Combinatorial sketching for finite programs. In ASPLOS-XII, pages
404–415, 2006.
[12] A. Solar-Lezama, C. G. Jones, and R. Bodik. Sketching concurrent
data structures. In PLDI ’08, pages 136–148, 2008.
[13] S. Srivastava, S. Gulwani, and J. S. Foster. From program verification
to program synthesis. In POPL ’10, pages 313–326, 2010.
[14] S. Srivastava, S. Gulwani, S. Chaudhuri, and J. S. Foster. Path-Based
Inductive Synthesis for Program Inversion. In PLDI ’11, pages 492–
503, June 2011.
A.
Tool Demonstration Walkthrough
As mentioned in the introduction, JS KETCH is available at http:
//plum-umd.github.io/java-sketch/. The tool is at a
fairly early stage of development, but is robust enough to be used
by the wider research community.
Our demonstration will generally follow the overview in Section 2. Below are more details of what we plan to present.
A.1
Basics
JS KETCH performs program synthesis—generating an output program given an input specification.
Let’s begin with a small example:
$ cat >> SimpleMath.java
class SimpleMath {
static int mult2(int x) {
return ?? * {| x , 0 |};
}
}
This is a sketch (also scaffold or template), which is a partial Java
program. The ?? is a hole—unknown integer—and the other part
of the product is a generator—ranging over the listed expressions.
Notice that this sketch has 233 possible instantiations.
In addition to the template, the other important input to JSKETCH is examples that specify the expected behavior of the template. These are analogous to unit tests. We provide a harness containing assertions about the mult2 method:
states via distinct integers, along with a series of conditionals that
encode state transitions.
Here’s the initial sketch:
1
2
3
4
5
6
7
8
9
10
interface Token { public int getId (); }
class Automaton {
private int state = ??;
static int num state = ??;
harness static void min num state() { minimize(num state); }
public void transition (Token t) {
assert 0 ≤ state && state < num state;
int id = t .getId ();
minrepeat {
if (state == ?? && id == ??) { state = ??; return; }
} }
public void transitions ( Iterator <Token> it) {
while( it .hasNext()) { transition ( it .next ()); }
}
public boolean accept() { return state ≤ ??; }
11
12
13
14
15
16
}
Here are some key things to notice about the source code:
• transitions are taken based on the Token interface
• the initial state is arbitrary (line 3)
• the number of states is arbitrary (line 4)
• states are dense (line 5)
$ cat >> Test.java
class Test {
harness static void test() {
assert SimpleMath.mult2(3) == 6;
}
}
• we use an assertion to check the validity of the current state
(line 7)
• a transition is arbitrary as it depends on an arbitrary current state
and an arbitrary id (line 10)
• minrepeat replicates its body the minimum necessary number of
Now we can run JS KETCH on the sketch and harness:
$ ./jsk.sh SimpleMath.java Test.java
06:07:15 rewriting syntax sugar
06:07:15 specializing class-level generator
06:07:15 rewriting exp hole
06:07:15 semantics checking
06:07:15 building class hierarchy
06:07:15 encoding result/sk_Test/SimpleMath.sk
06:07:15 encoding result/sk_Test/Test.sk
...
06:07:15 sketch running...
06:07:15 sketch done: result/output/Test.txt
06:07:15 replacing holes
06:07:15 replacing generators
06:07:15 semantics checking
06:07:15 decoding result/java/SimpleMath.java
...
06:07:15 synthesis done
The final result synthesized by JS KETCH is a valid Java source
file where unknowns have been replaced with the appropriate code:
$ cat result/java/SimpleMath.java
class SimpleMath { ...
static int mult2 (int x) {
return 2 * x;
}
}
A.2
times.
• the number of transitions is arbitrary (line 9)
• the number of accepting states is arbitrary (line 15); by packing
the states densely, we could use an inequality here
Now we can define an inline reference monitor as follows:
17
18
19
20
21
22
23
}
Monitor m;
public DBConnection() { m = new Monitor(); }
public boolean isErroneous() { return ! m.accept(); }
public void open() { m.transition (Monitor.OPEN); }
public void close() { m. transition (Monitor.CLOSE); }
24
25
26
27
28
29
30
}
The key idea is that each database connection operation is associated with an unique id, and the monitor maintains an automaton
that keeps receiving operation ids. At any point, a client can check
the status of the connection by asking the monitor whether it is in
an accepting state.
As expected, we need to provide a harness:
Database Connection Monitor
Now consider a harder problem: suppose we want to synthesize an
automaton-based inline reference monitor to check basic properties of a database connection, namely that the connection is never
opened or closed twice in a row and is only closed after being
opened. Let’s use a simple, efficient implementation: representing
class DBConnection {
class Monitor extends Automaton {
final static Token OPEN =
new Token() { public int getId() { return 1; } };
final static Token CLOSE =
new Token() { public int getId() { return 2; } };
public Monitor() { }
31
32
33
34
class TestDBConnection {
harness static void scenario good() {
DBConnection conn = new DBConnection();
assert ! conn.isErroneous();
conn.open(); assert ! conn.isErroneous();
conn.close(); assert ! conn.isErroneous();
35
36
}
// bad: opening more than once
harness static void scenario bad1() {
DBConnection conn = new DBConnection();
conn.open(); conn.open(); assert conn.isErroneous();
}
37
38
39
40
41
42
}
These examples illustrate one normal usage—opening a connection
and closing it—and one abnormal usage—opening a connection
twice. Given these harnesses, JS KETCH finds a solution:
43
44
45
46
47
48
49
74
75
}
public boolean accept() { return state ≤ 1; } ... }
51
80
class CADsR extends Automaton { ...
public boolean accept(String str) {
state = init state backup ;
transitions ( convertToIterator ( str ));
return accept();
81
}
76
77
78
79
88
class TestCADsR {
harness static void examples() {
CADsR a = new CADsR();
assert ! a.accept(”c” ); assert ! a.accept(”cr” );
assert a.accept(”car” ); assert a.accept(”cdr” );
assert a.accept(”caar” ); assert a.accept(”cadr” );
assert a.accept(”cdar” ); assert a.accept(”cddr” );
89
}
83
57
class TestDBConnection { ...
// bad: closing more than once
harness static void scenario bad2() {
DBConnection conn = new DBConnection();
conn.open();
conn.close(); conn.close(); assert conn.isErroneous();
58
}
52
53
54
55
56
84
85
86
87
60
}
61
62
63
64
65
66
67
68
69
class Automaton {
int state = 0;
static int num state = 3;
public void transition (Token t) { ...
assert 0 ≤ state && state < 3;
if (state == 0 && id == 1) { state =
if (state == 1 && id == 1) { state =
if (state == 1 && id == 2) { state =
if (state == 0 && id == 2) { state =
90
91
92
93
94
1;
2;
0;
2;
return;
return;
return;
return;
}
}
}
}
//
//
//
//
open@
open 2x
( init )@
close 2x
}
public boolean accept() { return state ≤ 1; } ... }
95
96
97
98
99
100
The resulting automaton is exactly same as what one can write by
hand.
A.3
101
102
103
104
105
106
107
108
109
70
71
generator class Automaton { ... }
class DBConnection {
class Automaton2 {
int state = 106; static int num state = 120;
public void transition (Token t) { ...
if (state == 106 && id == 99) { state = 64; return; }
if (state == 64 && id == 97) { state = 100; return; }
if (state == 100 && id == 114) { state = 50; return; }
if (state == 64 && id == 100) { state = 119; return; }
if (state == 119 && id == 114) { state = 32; return; }
if (state == 64 && id == 114) { state = 72; return; }
//
//
//
//
//
//
c
ca
car@
cd
cdr@
cr
}
public boolean accept() { return state ≤ 50; } ... }
Notice that the synthesizer picked fairly strange numbers for the
states and left a lot of states unused. Moreover, the automaton is
inefficient in that it uses two different paths and final states to accept
“car” and “cdr”.
If we run synthesis again using minimize, we get:
A Regular Language: Lisp-Style Identifier
Now let’s consider synthesizing another automaton, trying to create
a finite automaton given sample accepting and rejecting inputs.
One benefit of Java as an object oriented language is code reuse
via subclassing, so we could just make another class that extends
Automaton, assuming we want this to be part of the same program.
But subclassing won’t quite work here because we need different
automata for each use case.
To solve this problem, we can make Automaton a class generator, so that it can be instantiated differently in different superclass
contexts.
}
If we provide less examples, e.g., if we remove examples about
rejected strings, the synthesizer simply returns an automaton that
does not make any transitions, while the initial state is an accepting
state. This awkward automaton actually conforms to any accepted
strings, and one can easily figure out the necessity of rejected
strings.
To see the advantage of using minimize, let’s run synthesis without line 5. We get:
After adding that abnormal case—closing twice, JS KETCH
finds this solution:
59
}
Note the overloaded accept(String) method. Now, we need to specify
sample strings that are accepted or rejected by the synthesized automaton. Suppose we want to synthesize an automaton that recognizes Lisp-style identifier c(a|d)+r. The following harness constructs a CADsR instance and makes several assertions about its behavior:
82
This sort of looks okay, but it’s odd that there are no transitions
for the close operation. When the monitor is in state 1 and the given
operation is close, it is fine for the monitor to stay at the same
accepting state. But, it is problematic if we close the connection
more than once, which we should have specified:
}
class CADsR extends Automaton { ... }
Now let’s use this to synthesize an example automaton:
class Automaton {
int state = 0;
static int num state = 3;
public void transition (Token t) { ...
assert 0 ≤ state && state < 3;
if (state == 0 && id == 1) { state = 1; return; } // open@
if (state == 1 && id == 1) { state = 2; return; } // open 2x
50
class Monitor extends Automaton { ... }
...
72
73
110
class Automaton2 {
int state = 0; static int num state = 3;
public void transition (Token t) { ...
assert 0 ≤ state && state < 3;
if (state == 0 && id == 99) { state =
if (state == 1 && id == 97) { state =
if (state == 1 && id == 100) { state =
if (state == 2 && id == 114) { state =
1;
2;
2;
0;
return;
return;
return;
return;
}
}
}
}
}
public boolean accept() { return state ≤ 0; } ... }
//
//
//
//
c
ca
cd
c(a|d)+r@
This result is better in the sense that it uses dense states and that
it encompasses only one path and final state to accept the valid
strings.
To double-check whether it is indeed the minimum number of
states, we can test with the bounded number of states:
112
class Automaton {
static int num state = 2; ...
113
}
111
In this case, the synthesizer runs for more than half an hour and
then fails, as there is no possible solution using only two states.
A.4 Internals of JS KETCH
If time permits, we will show a bit of JS KETCH’s translation to
S KETCH. For example, the translation of the mult2 example looks
like:
$ cat result/sk_Test/SimpleMath.sk
...
int e_h1 = ??;
int mult2_SimpleMath_int(int x) {
return e_h1 * {| x | 0 |};
}
$ cat result/sk_Test/Test.sk
...
harness void test_Test() {
assert mult2_SimpleMath_int(3) == 6;
}
JS KETCH extracts the synthesis results by looking at how each
hole was solved by S KETCH:
$ cat result/log/log.txt
...
06:07:15 [DEBUG] java_sk/decode/finder.py:41
hole: SimpleMath.e_h1
06:07:15 [INFO] java_sk/decode/__init__.py:69
replacing holes
06:07:15 [DEBUG] java_sk/decode/replacer.py:72
replaced: SimpleMath.e_h1 = 2
06:07:15 [DEBUG] java_sk/decode/replacer.py:89
replaced: e_h1 @ int SimpleMath.mult2(int) with 2
06:07:15 [DEBUG] java_sk/decode/finder.py:93
generator@mult2: {| x | 0 |}
06:07:15 [INFO] java_sk/decode/__init__.py:79
replacing generators
06:07:15 [DEBUG] java_sk/decode/replacer.py:151
{| x | 0 |} => x
...
Then it traverses the original Java sketch and outputs it, plugging in the solved values for the holes.
| 6cs.PL
|
Inclusion graph of subgroups of a group
arXiv:1604.08259v1 [math.GR] 27 Apr 2016
P. Devi∗, R. Rajkumar†
Department of Mathematics, The Gandhigram Rural Institute – Deemed University,
Gandhigram – 624 302, Tamil Nadu, India.
Abstract
For a finite group G, we define the inclusion graph of subgroups of G, denoted by
I(G), is a graph having all the proper subgroups of G as its vertices and two distinct
vertices H and K in I(G) are adjacent if and only if either H ⊂ K or K ⊂ H. In
this paper, we classify the finite groups whose inclusion graph of subgroups is one of
complete, bipartite, tree, star, path, cycle, disconnected, claw-free. Also we classify
the finite abelian groups whose inclusion graph of subgroups is planar. For any given
finite group, we estimate the clique number, chromatic number, girth of its inclusion
graph of subgroups and for a finite abelian group, we estimate the diameter of its
inclusion graph of subgroups. Among the other results we show that some groups can
be determined by their inclusion graph of subgroups
Keywords: Inclusion graph of subgroups, bipartite graph, clique number, girth,
diameter, planar graph.
2010 Mathematics Subject Classification: 05C25, 05C05, 05C10, 20K27,
20E15.
∗
†
e-mail: [email protected]
e-mail: [email protected]
1
2
1
Introduction
The properties of an algebraic structure can be investigated in several ways. One of the ways
is by associating a suitable graph to that algebraic structure and analyzing the properties of
the associated graph by using graph theoretic methods. The subgroup lattice and subgroup
graph of a group are well known graphs associated with a group (cf. [4], [10], [11], [12]).
The intersection graph of subgroups of a group is another interesting graph associated with
a group (cf. [2], [6], [9]). For a group G, the intersection graph of subgroups of G, denoted
by I (G), is a graph having all the proper subgroups of G as its vertices and two distinct
vertices are adjacent if and only if the corresponding subgroups intersects non-trivially. In
[1], S. Akbari et al assigned a graph to ideals of a ring as follows: For a ring R with unity,
the inclusion ideal graph of R is a graph whose vertices are all non-trivial left ideals of R
and two distinct left ideals I and J are adjacent if and only if I ⊂ J or J ⊂ I.
Motivated by this, in this paper, we define the following: For a finite group G, the inclusion graph of subgroups of G, denoted by I(G), is a graph having all the proper subgroups
of G as its vertices and two distinct vertices H and K in I(G) are adjacent if and only if
H ⊂ K or K ⊂ H.
For a group G, its subgroup lattice is denoted by L(G). The height of L(G) is the length
of the longest chain in L(G) through partial order from greatest to least. We denote the
order of an element a ∈ Zn by ordn (a) . The number of Sylow p-subgroups of a group G is
denoted by np (G).
Now we recall some basic definitions and notations of graph theory. We use the standard
terminology of graphs (e.g., see [8]). Let G be a simple graph with vertex set V (G) and
edge set E(G). G is said to be k-partite if V (G) can be partition into k disjoint subsets Vi ,
i = 1, 2, . . . , k such that every edge joins a vertex of Vi to a vertex of Vj , i 6= j. A k-partite
graph is said to be a complete k-partite if every vertex in each partition is adjacent with
all the vertices in the remaining partitions and is denoted by Km1 ,m2 ,...,mk , where mi = |Vi |,
i = 1, . . . , k. The graph K1,m is called a star and the graph K1,3 is called a claw. A graph in
which any two distinct vertices are adjacent is said to be complete. A graph whose edge set
is empty is said to be totally disconnected. Two graphs G1 , G2 are isomorphic if there exists
3
a bijection from V (G1 ) to V (G2 ) preserving the adjacency. A path connecting two vertices u
and v in G is a finite sequence (u =)v0 , v1 , . . . , vn (= v) of distinct vertices (except, possibly,
u and v) such that ui is adjacent to ui+1 for all i = 0, 1, . . . , n − 1. A path is a cycle if u = v.
The length of a path or a cycle is the number of edges in it. A path or a cycle of length n
is denoted by Pn or Cn , respectively. A graph G is said to be connected if any two vertices
are connected by a path; otherwise G is said to be disconnected. A connected graph with
out a cycle is called a tree. For a connected graph G, its diameter, denoted by diam(G), is
the maximum of length of a shortest path of any two vertices. If G is disconnected, then
we define diam(G) = ∞. The girth of a graph G, denoted by girth(G), is the length of
a shortest cycle in G, if it exist; otherwise, we define girth(G) = ∞. A clique is a set of
vertices in G such that any two are adjacent. The clique number ω(G) of G is the cardinality
of a largest clique in G. The chromatic number χ(G) of G is the smallest number of colors
needed to color the vertices of G such that no two adjacent vertices gets the same color. A
graph is said to be planar if it can be drawn on a plane such that no two edges intersect,
except, possibly at their end vertices. We define a graph G to be X-free if it does not contain
a subgraph isomorphic to a given graph X. G denotes the complement of a graph G. For
two graphs G and H, G ∪ H denotes the disjoint union of G and H.
In this paper, we classify all the finite groups whose inclusion graph of subgroups is one of
complete, bipartite, tree, star, path, cycle, claw-free, disconnected (cf. Theorems 2.4, 2.6, 2.8,
2.12, Corollaries 2.2, 2.4). Also we give the classification of finite abelian groups whose
inclusion graph of subgroups is planar (cf. Theorem 2.10). For any given finite group, we
estimate the clique number, chromatic number, girth of its inclusion graph of subgroups (cf.
Theorem 2.12, Corollary 2.1) and for a finite abelian group, we estimate the diameter of its
inclusion graph of subgroups (cf. Theorem 2.11). Moreover, we show that some groups can
be determined by their inclusion graph of subgroups (cf. Corollary 2.3). In this sequel, we
show some interesting connections between the inclusion graph of subgroups of a group, its
subgroup lattice and its intersection graph of subgroups (cf. Theorems 2.2, 2.5, 2.6, 2.7, 2.9
and Corollary 2.1).
4
2
Main results
Theorem 2.1. Let G1 and G2 be groups. If G1 ∼
= G2 , then I(G1 ) ∼
= I(G2 ).
Proof. If φ is a group isomorphism from G1 into G2 , then the map ψ : V (I(G1 )) → V (I(G2 ))
by ψ(H) = φ(H), for all H ∈ V (I(G1 )) is a graph isomorphism.
Remark 2.1. The converse of the above theorem is not true. For example, let G1 = Z3 × Z3
and G2 = S3 , then it is easy to see that I(G1 ) ∼
= I(G2 ), but G1 G2 .
= K4 ∼
Theorem 2.2. Let G1 and G2 be groups. If L(G1 ) ∼
= L(G2 ), then I(G1 ) ∼
= I(G2 ).
Proof. Let φ be a lattice isomorphism from L(G1 ) into L(G2 ). Define a map ψ : V (I(G1 )) →
V (I(G2 )) by ψ(H) = φ(H), for all H ∈ V (I(G1 )). Since φ is bijective, so is ψ. Suppose
H1 and H2 are adjacent in I(G1 ), then either H1 ⊂ H2 or H2 ⊂ H1 . Since φ is a lattice
isomorphism, so it preserves the meet and order and so either φ(H1 ) ⊂ φ(H2 ) or φ(H2 ) ⊂
φ(H1 ). It follows that ψ(H1 ) and ψ(H2 ) are adjacent in I(G2 ). By following a similar
argument as above, it is easy to see that if ψ(H1 ) and ψ(H2 ) are adjacent in I(G2 ), then H1
and H2 are adjacent in I(G1 ). Thus ψ is a graph isomorphism. Hence the proof.
Theorem 2.3. Let G be any group and N be a subgroup of G. Then I(N ) is a subgraph of
I(G). In addition, N is a normal subgroup, then I(G/N ) is isomorphic (as a graph) to a
subgraph of I(G).
Proof. The first result is obviously true. Any subgroup of G/N is of the form H/N , where
H is a subgroup of G containing N . Here two proper subgroups H/N , K/N are adjacent in
I(G/N ) if and only if either H/N ⊂ K/N or K/N ⊂ H/N . This implies that either H ⊂ K
or K ⊂ H and so H and K are adjacent in I(G). This completes the proof.
Theorem 2.4. Let G be a finite group. Then I(G) is complete if and only if G ∼
= Zpα ,
where p is a prime and α > 1. Moreover, I(Zpα ) ∼
= Kα−1 .
Proof. If G ∼
= Zpα , where p is a prime and α > 1, then L(G) is a chain of length α and
so I(G) ∼
= Kα−1 . If G Zpα , then L(G) is not a chain and so there exists at least two
subgroups H1 and H2 of G such that H1 * H2 and H2 * H1 . It follows that I(G) is not
complete.
5
Theorem 2.5. Let G be a finite group. Then I(G) is (k − 1)-partite, where k is the height
of L(G).
Proof. Let M1 be the set of all maximal subgroups of G and for each i = 2, 3, . . . , k − 1,
let Mi be the set of all maximal subgroups of the subgroups in Mi−1 . Then {Mi }k−1
i=1 is a
partition of the vertex set of I(G). Also no two vertices in a same partition are adjacent
in I(G). Moreover, k − 1 is the minimal number such that a k-partition of the vertex
set of I(G) is having this property, since the height of L(G) is k. It follows that I(G) is
(k − 1)-partite.
The next result is an immediate consequence of Theorem 2.5, the definition of the clique
number and chromatic number of a graph.
Corollary 2.1. Let G be a finite group. Then ω(I(G)) = k − 1 = χ(I(G)), where k is the
height of L(G).
The next two results are immediate consequences of Theorem 2.5 and the definition of
the subgroup lattice of a group.
Theorem 2.6. Let G be a group. Then the following are equivalent.
(i) I(G) is totally disconnected;
(ii) every proper subgroups of G is of prime order;
(iii) height of L(G) is 2.
Theorem 2.7. Let G be a group and e be its identity element. Then the following are
equivalent.
(i) I(G) is bipartite;
(ii) L(G) − {G, e} ∼
= I(G);
(iii) height of L(G) is either 2 or 3.
Corollary 2.2. Let G be a finite group and p, q, r be distinct primes. Then
6
(1) I(G) is totally disconnected if and only if G is one of Zp2 , Zp × Zp , Zpq , Zq o Zp ;
(2) I(G) is bipartite if and only if |G| is one of p2 , pq, p3 , p2 q or pqr.
α
Consider the semi-direct product Zq ot Zpα = ha, b | aq = bp = 1, bab−1 = ai , ordq (i) =
pt i, where p and q are distinct primes with pt | (q − 1), t ≥ 0. Then every semi-direct
product Zq o Zpα is one of these types [4, Lemma 2.12]. Note that, here after we suppress
the subscript when t = 1.
Theorem 2.8. Let G be a finite group and p, q, r be distinct primes. Then
(1) I(G) ∼
= Cn if and only if either n = 3 and G ∼
= Zp4 or n = 6 and G ∼
= Zpqr ;
(2) I(G) is a tree if and only if G is one of Zp3 , Zp2 q , Zp2 × Zp , Q8 , M8 , Zq o Zp2 , Zq o2 Zp2
2
or Zp2 o Zq = ha, b | ap = bq = 1, bab−1 = ai , ordp2 (i) = qi (q |(p − 1));
(3) I(G) is a star if and only if G is either Zp3 or Q8 ;
(4) I(G) ∼
= Pn if and only if either n = 1 and G ∼
= Zp3 or n = 3 and G ∼
= Zp2 q .
Proof. First we claim that I(G) ∼
= Cn , where n is odd if and only if n = 3 and G ∼
= Zp4 .
It is easy to see that G ∼
= Zp4 if and only if I(G) ∼
= C3 . Now suppose that I(G) ∼
= Cn ,
where n is odd, n ≥ 5. Let this cycle be H1 − H2 − · · · − Hn − H1 . Now we have two
possibilities: H1 ⊂ H2 , Hn or H2 , Hn ⊂ H1 .
(i) Suppose that H1 ⊂ H2 , Hn . Since H2 is adjacent to H3 , so either H2 ⊂ H3 or H3 ⊂ H2 .
If H2 ⊂ H3 , then H1 , H2 , H3 forms C3 as a subgraph of I(G), which is a contradiction
and so we must have H3 ⊂ H2 . Next, since H3 is adjacent to H4 , so either H3 ⊂ H4
or H4 ⊂ H3 . If H4 ⊂ H3 , then H2 , H3 , H4 forms C3 as a subgraph in I(G), which is a
contradiction and so we must have H3 ⊂ H4 . If we proceed like this, we get Hi ⊂ Hi−1 ,
Hi+1 , when i is an odd integer (suffixes taken modulo n) and so Hn ⊂ H1 , which is not
possible.
(ii) Suppose that H2 , Hn ⊂ H1 . Since H2 is adjacent to H3 , so either H2 ⊂ H3 or H3 ⊂ H2 .
If H3 ⊂ H2 , then H1 , H2 , H3 forms C3 as a subgraph of I(G), which is a contradiction
and so we must have and so we have H2 ⊂ H3 . Next, since H3 is adjacent to H4 , so
7
either H3 ⊂ H4 or H4 ⊂ H3 . If H3 ⊂ H4 , then H2 , H3 , H4 forms C3 as a subgraph of
I(G) and so we have H4 ⊂ H3 . If we proceed like this, we get Hi ⊂ Hi−1 , Hi+1 , when
i is an even integer (suffixes taken modulo n) and so we have Hn−1 ⊂ Hn . This implies
that Hn−1 , Hn , H1 forms C3 as a subgraph of I(G), which is a contradiction.
Thus I(G) Cn , where n is odd, n ≥ 5 and our claim is now proved.
Next, we start to prove the main theorem. Since every tree, star graph, path, even cycle
are bipartite, so to classify the finite groups whose inclusion graph of subgroups is one of
tree, star graph, path, even cycle, it is enough to consider the groups of order p2 , pq, p3 , p2 q,
pqr, by Corollary 2.2(2).
Case 1: Let |G| = p2 or pq. By Theorem 2.6, I(G) is neither a tree nor a cycle.
Case 2: Let |G| = p3 . Here we use the classification of groups of order p3 .
(i) If G ∼
= Zp3 , then by Theorem 2.4, I(G) ∼
= K2 , which is a path but not a cycle.
(ii) If G ∼
= Zp2 ×Zp , then h(1, 0)i, h(1, 1)i, . . ., h(1, p−1)i, h(p, 0), (0, 1)i, h(p, 0)i, h(p, 1)i, . . .,
h(p, p − 1)i, h(0, 1)i are the only proper subgroups of G. I(G) is shown in Figure 1(a),
which is a tree, but none of a star, path, cycle.
(iii) If G ∼
= Zp × Zp × Zp := ha, b, c | ap = bp = cp = 1, ab = bc, ac = ca, bc = cbi, then I(G)
contains C6 as a subgraph: ha, bi − hbi − hb, ci − hci − ha, ci − hai − ha, bi. So I(G) is
neither a tree nor a cycle.
(iv) If G ∼
= Q8 = ha, b | a4 = b4 = 1, a2 = b2 , ab = ba−1 i, then hai, hbi, habi, ha2 i are the
only proper subgroups of G. It follows that
I(Q8 ) ∼
= K1,3 ,
(2.1)
which is a star; but neither a path nor a cycle.
(v) If G ∼
= M8 = ha, b | a4 = b2 = 1, ab = ba−1 i, then hai, ha2 , bi, ha2 , abi, hbi, ha2 i, habi,
ha2 bi, ha3 bi are the only proper subgroups of G. It follows that I(G) is as shown in
Figure 1(b), which is a tree; but none of a star, path, cycle.
8
(vi) If G ∼
= Mp3 , p > 2, then subgroup lattice of Mp3 and Zp2 × Zp are isomorphic. It follows
from Theorem 2.2 that I(G) is as in Figure 1(a), which is a tree, but none of a star,
path, cycle.
(vii) If G ∼
= (Zp × Zp ) o Zp := ha, b, c | ap = bp = cp = 1, ab = ba, ac = cacbc−1 = abi, then
I(G) contains C6 as a subgraph: ha, bi − hbi − hb, ci − hci − ha, ci − hai − ha, bi. So I(G)
is neither a tree nor cycle.
Case 3: Let |G| = p2 q. Here we use the classification of groups of order p2 q given in [5, pp.
76-80].
Subcase 3a: Let G be abelian.
(i) If G ∼
= Zp2 q , then it is easy to see that
I(Zp2 q ) ∼
= P3 ,
(2.2)
which is a path; but neither a star nor a cycle.
(ii) If G ∼
= Zpq × Zp , then H := Zp × Zp , A := Zq × {e}, Hx := h(1, x)i (x = 0, 1, . . . , p − 1),
Hp = h(0, 1)i, Ax := Hx A, are the only subgroups of G. It follows that I(G) is as
shown in Figure 1(c), which is neither a tree nor a cycle.
Subcase 3b: Let G be non-abelian.
Subcase 3b(I): Let p < q.
(i) If p - (q − 1), by Sylow’s Theorem, there are no non-abelian groups.
(ii) If p | (q −1) but p2 - (q −1), then we have two groups. The first one is G1 ∼
= Zq oZp2 =
2
ha, b | aq = bp = 1, bab−1 = ai , ordq (i) = pi. Here ha, bi, hai bi (i = 1, 2, . . . , q),
habp i, hbp i are the only proper subgroups of G1 . It follows that I(G1 ) is as shown in
Figure 1(d), which is a tree; but none of a star, path, cycle.
The second one is G2 = ha, b, c | aq = bp = cp = 1, bab−1 = ai , ac = ca, bc = cb, ordq (i) =
pi. Then I(G) contains C6 as a proper subgraph: ha, bi − hbi − hb, ci − hci − ha, ci −
hai − ha, bi and so I(G2 ) is neither a tree nor a cycle.
9
(iii) If p2 | (q − 1), then we have both groups G1 and G2 mentioned in (ii) of this subcase,
2
together with the group G3 = Zq o2 Zp2 = ha, b | aq = bp = 1, bab−1 = ai , ordq (i) = p2 i.
In (ii) of this subcase, we already dealt with G1 , G2 . By Sylow’s Theorem, G3 has a
unique subgroup, say H of order q, and has q Sylow-p subgroups of order p2 , say Hi ,
i = 1, 2, . . . , q ; for each i = 1, 2, . . . , q, Hi has a unique subgroup Hi0 of order p and
G3 has a unique subgroup, say K of order pq. Here H is a subgroup of K; for each
i = 1, 2, . . . , q, Hi0 is a subgroup of Hi , K; for each i = 1, 2, . . . , q, Hi is a subgroup
of K; for each i = 1, 2, . . . , q, no two Hi contained in Hj and Hi ’ contained in Hj0 , for
every i 6= j. It follows that I(G3 ) is shown in Figure 1(e), which is neither a tree nor
a cycle.
Subcase 3b(II): Let p > q.
(i) If q - (p2 − 1), then there is no non-abelian group in this case.
2
(ii) If q | (p − 1), then we have two groups. The first group is G4 ∼
= Zp2 o Zq = ha, b | ap =
bq = 1, bab−1 = ai , ordp2 (i) = qi. By Sylow’s Theorem, G4 has a unique subgroup H
of order p2 ; H has a unique subgroup H 0 of order p and p2 Sylow q-subgroups of order
q, say H1 , H2 ,. . ., Hp2 and p subgroups of order pq, say N1 , N2 , . . ., Np ; these are the
only proper subgroups of G4 . It follows that I(G4 ) is as shown in Figure 1(f), which is
a tree but none of a star, path, cycle.
Second, we have a family of groups ha, b, c | ap = bp = cq = 1, cac−1 = ai , cbc−1 =
t
bi , ab = ba, ordp (i) = qi. As q > 2, there are (q + 3)/2 isomorphism types in this family
(one for t = 0 and one for each pair {x, x−1 } in Fpx ). We will refer to all these groups
as G5(t) of order p2 q. Then I(G(5t) ) contains C6 as a subgraph: ha, bi − hbi − hb, ci −
hci − ha, ci − hai − ha, bi. So I(G5(t) ) is neither a tree nor a cycle.
(iii) If q | (p + 1), then we have only one group of order p2 q, given by G6 ∼
= (Zp × Zp ) o Zq =
ha, b, c | ap = bp = cq = 1, ab = ba, cac−1 = ai bj , cbc−1 = ak bl i, where ki jl has order
q in GL2 (p). Then G6 has unique subgroup of order p2 , let it be H; p + 1 subgroups
of order p, let them be Hi , i = 1, 2, . . ., p; p2 subgroups of order q; these are the only
10
ha2 , bi
ha2 i
h(p, 0)i
ha2 , abi
b
b
b
h(p, 0), (0, 1)i
b
H
Hb 0
A0
Hb 1
Ab 1
b
b
b
b
b
b
b
b
b
b
b
Hp−1
b
b
b
b
b
h(1, 0)i h(1, p − 1)i h(p, 1)i h(p, p − 1)i
b
b
ha2 bi
hbi
hai
(a)
b
b
ha3 bi
habi
b
Ap−1
b
b
Hp
Ap
(b)
A
(c)
Hb
b
habp i
hbp i
b
b
b
H
b
b
b
hai
hbi
haq−1 bi
(d)
b
b
H1
N
b
H1′
b
b
H1
Hp
b
Hq′
H2′
b
b
H2
Hq
Np
N
1 H′ b
b
Hp2
b
b
b
H(p−1)p
N2 b
b
Hp+1
(e)
b
H2p
(f)
Figure 1: (a) I(Zp2 × Zp ), (b) I(M8 ), (c) I(Zpq × Zp ), (d) I(Zq o Zp2 ), (e) I(Zq o2 Zp2 ), (f)
I(Zp2 o Zq )
proper subgroups of G6 . It follows that
I(G6 ) ∼
= K1,p+1 ∪ K p2 ,
(2.3)
which is neither a tree nor a cycle.
Note that if (p, q) = (2, 3), then subcases 3b(I) and 3b(II) are not mutually exclusive. Up to
isomorphism, there are three non-abelian groups of order 12: Z3 o Z4 , D12 and A4 . We have
already dealt with Z3 o Z4 , D12 in (ii) of Subcase 3b(I). But for A4 , we cannot use the same
argument as in (iii) of Subcase 3b(II). Now A4 has unique subgroup of order 4, say H ; three
subgroups of order 2, say H1 , H2 , H3 and four subgroups of order 3, say N1 , N2 , N3 , N4 .
Here for each i = 1, 2, 3, Hi is a subgroup of H; for each i = 1, 2, 3, 4, Ni is not contained in
any other subgroups of A4 ; no two remaining subgroups are adjacent in I(A4 ). Therefore,
I(A4 ) ∼
= K1,3 ∪ K 4 ,
which is neither a tree nor a cycle.
Case 4: Let |G| = pqr.
(2.4)
11
If G ∼
= Zpqr , then let H1 , H2 , H3 , H4 , H5 , H6 be the subgroups of G of orders p, q, r, pq,
pr, qr, respectively. Here H1 is a subgroup of H4 , H5 ; H2 is a subgroup of H4 , H6 ; H3 is a
subgroup of H5 , H6 . So it turns out that I(G) ∼
= C6 , which is not a tree.
If G is non-abelian, then G has a Sylow basis containing Sylow p, q, r-subgroups, let them
be H1 , H2 , H3 , respectively and so H1 H2 , H2 H3 , H1 H3 are proper subgroups of G. It follows
that I(G) contains C6 as a proper subgraph: H1 − H1 H2 − H2 − H2 H3 − H3 − H1 H3 − H1 .
So I(G) neither a tree nor a cycle.
Proof follows by combining all the above cases together.
In the next result, we characterize some groups by using their inclusion graph of subgroups.
Corollary 2.3. Let G be a group.
(1) If I(G) ∼
= I(Q8 ), then G ∼
= Q8 ;
(2) If I(G) ∼
= I(M8 ), then G ∼
= M8 .
(3) If I(G) ∼
= I(Z9 o Z2 ), then G ∼
= Z9 o Z2
Proof. By Theorem 2.8(2), I(Q8 ), I(M8 ), I(Z9 o Z2 ) are trees and by (2.1), Figures 1(b)
and 1(f), these trees uniquely determines the corresponding group. Hence the proof.
Theorem 2.9. Let G be a group. Then I(G) is connected if and only if I (G) is connected.
Proof. Since I(G) is a spanning subgraph of I (G), so if I (G) is connected, then so is
I (G). Now let H, K be two adjacent vertices in I (G). Then exactly one of the following
holds: H ⊂ K, K ⊂ H, neither H ⊂ K nor K ⊂ H but |H ∩ K| > 1. If one of the first
two possibilities holds, then H and K are adjacent in I(G). If the third condition holds,
then H − H ∩ K − K is a path in I(G). So it follows that if I (G) is connected, then so is
I(G).
In [9], Rulin Shen et al classified all the finite groups whose intersection graphs of subgroups are disconnected. So as a consequence of Theorem 2.9, we have the following result.
12
Corollary 2.4. Let G be a finite group. Then I(G) is disconnected if and only if G is one
of Zp × Zq , where both p, q are primes, or a Frobenius group whose complement is a prime
order group and the kernel is a minimal normal subgroup.
Theorem 2.10. Let G be a finite abelian group and p, q, r be distinct primes. Then I(G)
is planar if and only if G is one of Zpα (α = 2, 3, 4, 5), Zpα q (α = 1, 2, 3, 4), Zp2 q2 , Zpqr , Zp2 qr ,
Zpqrs , Zp × Zp , Zp2 × Zp or Zpq × Zp .
Proof. Let |G| = pα1 1 pα2 2 . . . pαk k , where pi ’s are distinct primes and αi ≥ 1.
Case 1: Let G be abelian.
Subcase 1a: k = 1. By Theorem 2.4, I(G) is planar if and only if α = 2, 3, 4, 5.
Subcase 1a: k = 2.
(i) If either α1 ≥ 5 or α2 ≥ 5, then L(G) has a chain of length at least four and so I(G)
contains K5 as a subgraph. This implies that I(G) is non-planar.
(ii) If α1 ≥ 3 and α2 ≥ 2, then G has a subgroups H1 , H2 , H3 , N1 , N2 , N3 of orders p1 , p2 ,
p1 p2 , p21 p2 , p21 p22 , p1 p22 , respectively and so K3,3 is a subgraph of I(G) with bipartition
X := {H1 , H2 , H3 } and Y := {N1 , N2 , N3 }. This implies that I(G) is non-planar.
(iii) If α1 = 4 and α2 = 1, then G has subgroups of order p1 , p21 , p31 , p41 , p2 , p1 p2 , p21 p2 , p31 p2 ,
let them be Hi , i = 1, 2, . . . , 8, respectively and these are the only proper subgroups of
G. Here I(G) is planar and a corresponding plane embedding is shown in Figure 2(a).
(iv) If α1 ≤ 3 and α2 = 1, then I(G) is a subgraph of I(Zp41 p2 ) and so I(G) is planar.
(v) If α1 = 2 and α2 = 2, then G has subgroups of order p1 , p21 , p2 , p22 , p1 p2 , p21 p2 , p1 p22 , let
them be Hi , i = 1, 2, . . . , 7 respectively and these are the only proper subgroups of G.
Here I(G) is planar and a corresponding plane embedding is shown in Figure 2(b).
Subcase 1b: k = 3.
(i) If α1 ≥ 3, α2 = 1 = α3 , then let H1 , H2 , H3 , N1 , N2 , N3 be subgroups of G of orders
p1 , p2 , p1 p2 , p1 p2 p3 , p21 p2 , p21 p2 p3 , respectively. Then K3,3 is a subgraph of I(G) with
bipartition X := {H1 , H2 , H3 } and Y := {N1 , N2 , N3 } and so I(G) is non-planar.
13
H8
Hb 7
H1
b
H7
b
b
b
H2
H5
b
H4 b
b
H5
b
b
H4
b
H6
H5
H1
H3
H4
b
H2
b
b
H6
H3
(a)
H3
b
H1
b
b
b
b
H10
H6
b
b
b
b
H9
H7
(b)
b
b
H2
H8
(c)
H9
b
H14
b
H2
b
H8
b
H5
H3
b
b
b
b
b
H6
H13
H12
H1
H11
H10
b
b
H1
H2
H3
b
b
H9
b
b
H6
b
b
H4
b
H8
H7
b
b
b
b
b
H4
(d)
H5
H7
b
H12
b
H10
b
H11
(e)
Figure 2: (a) I(Zp41 p2 ), (b) I(Zp21 p22 ), (c) I(Zp21 p2 p3 ), (d) I(Zp1 p2 p3 p4 ), (e) a subdivision of K3,3
in I(Zp × Zp × Zp )
(ii) If α1 , α2 ≥ 2, α3 ≥ 1, then we can use a similar argument as in (i) above to show I(G)
is non-planar.
(iii) α1 = 2, α2 = 1 = α3 , then G has subgroups of order p1 , p21 , p2 , p3 , p1 p2 , p1 p3 , p21 p2 ,
p2 p3 , p21 p3 , p1 p2 p3 , let them be Hi , i = 1, . . . , 10 respectively and these are the only
proper subgroups of G. Here I(G) is planar and a corresponding plane embedding is
shown in Figure 2(c).
(iv) If α1 = α2 = α3 = 1, then I(G) is a subgraph of I(Zp2 qr ) and so I(G) is planar.
Subcase 1c: k = 4.
(i) If α1 ≥ 2, α2 , α3 , α4 ≥ 1, then let H1 , H2 , H3 , N1 , N2 , N3 be subgroups of G of orders
p1 , p2 , p1 p2 , p1 p2 p3 , p1 p2 p4 , pα1 1 p2 p3 , respectively and so K3,3 is a subgraph of I(G)
with bipartition X := {H1 , H2 , H3 } and Y := {N1 , N2 , N3 }. This implies that I(G)
is non-planar.
(ii) If α1 = α2 = α3 = α4 = 1, then G has subgroups of order p1 , p2 , p3 , p4 , p1 p2 , p1 p3 ,
14
p1 p4 , p2 p3 , p2 p4 , p3 p4 , p1 p2 p3 , p1 p2 p4 , p1 p3 p4 , p2 p3 p4 , let them be Hi , i = 1, 2, . . . , 14
respectively and these are the only proper subgroups of G. Here I(G) is planar and a
corresponding plane embedding is shown in Figure 2(d).
Subcase 1d: k ≥ 5. Let H1 , H2 , H3 , N1 , N2 , N3 be subgroups of G of orders p1 , p2 ,
p1 p2 , p1 p2 p3 , p1 p2 p4 , p1 p2 p5 , respectively and so K3,3 is a subgraph of I(G) with bipartition
X := {H1 , H2 , H3 } and Y := {N1 , N2 , N3 }. This implies that I(G) is non-planar.
Case 2: Let G be non-cyclic.
(i) If G ∼
= Zp × Zp , then by Corollary 2.2(1), I(G) is planar.
(ii) If G ∼
= Zp2 × Zp , then by Figure 1(a), I(G) is planar.
(iii) If G ∼
= Zpq × Zp , then by Figure 1(c), I(G) is planar.
(iv) If G ∼
= Zp2 q × Zp , then Zp2 × Zp is a subgroup of G and by Figure 1(a), G has a unique
subgroup of order p, say H; let H1 , H2 , H3 be subgroups of Zp2 × Zp of order p2 ; let
N be a subgroup of G of order q. Then K3,3 is a subgraph of I(G) with bipartition
X := {H1 N , H2 N , H3 N }, Y := {N , H HN } and so I(G) is non-planar.
2
2
(v) If G ∼
= Zp2 × Zp2 = ha, b | ap = bp , ab = bai, then K3,3 is a subgraph of I(G) with
bipartition X := {ha, bp i hap , bi, hap , bp i}, Y := {hap i, hbp i, hap bp i} and so I(G) is
non-planar.
(vi) If G ∼
= Zp × Zp × Zp , then H1 := h(1, 0, 0)i, H2 := h(0, 1, 0)i, H3 := h(0, 0, 1)i,
H4 := h(0, 1, 1)i, H5 := h(1, 0, 1)i, H6 := h(1, 1, 0)i, H7 := h(1, 0, 0), (0, 1, 1)i, H8 :=
h(0, 1, 0), (1, 0, 1)i, H9 := h(0, 0, 1), (1, 1, 0)i, H10 := h(1, 0, 0), (0, 0, 1)i, H11 := h(1, 0, 0),
(0, 1, 0)i, H12 := h(0, 1, 0), (0, 0, 1)i are subgroups of G. Here I(G) has a subdivision of
K3,3 as a subgraph, which is shown in Figure 2(e) and so I(G) is non-planar.
(vii) If G ∼
= Zpα1 1 × Zpα2 2 × · · · ×Zpαk , where k ≥ 3, pi ’s are primes and at least two of them
k
are equal (since G is non-cyclic, all the primes cannot be distinct here), αi ≥ 1 are
integers, i = 1, . . . , k. Then, G has one of Zp2i pj × Zpi , Zp2i × Zp2i or Zpi × Zpi × Zpi as a
subgroup for some i and j, i 6= j. So by the above arguments, it follows that I(G) is
non-planar.
15
Proof follows by putting together the above cases.
Theorem 2.11. If G is a finite abelian group, then diam(I(G)) ∈ {1, 2, 3, 4, ∞}.
Proof. If I(G) is disconnected, then diam(I(G)) = ∞. Now we assume that I(G) is con-
nected. If G ∼
= Zpα , where p is a prime and α > 1, then by Theorem 2.4, it follows that
diam(I(G)) = 1. Now we assume that G Zpα . Since I(G) is connected, so the order of
not every proper subgroup of G is a prime. Let H, K be two proper subgroups of G. If |H|,
|K| are prime, then HK is a proper subgroup of G and so H − HK − K is a path from
H to K. If |H| is a prime, |K| is not a prime, then H − HK 0 − K 0 − K is a path from H
to K, where K 0 is a proper subgroup of K. If |H|, |K| are not primes and |H ∩ K| = 1,
then H − H 0 − H 0 K 0 − K 0 − K is a path from H to K, where H 0 , K 0 are proper subgroups
of H, K respectively. If |H|, |K| are not primes and |H ∩ K| > 1, then H − H ∩ K − K
is a path from H to K. Thus we have shown that diam(I(G)) ≤ 4. It is easy to see that
diam(I(Zp2 q2 )) = 2, diam(I(Zp2 × Zp )) = 3, diam(I(Zpq × Zpq )) = 4. So this shows that
diam(I(G)) takes all the values in {1, 2, 3, 4, ∞} and the proof is complete.
Our next aim is to prove the following result, which describes the girth of the intersection
graph of subgroups of finite groups and classifies all the finite groups whose inclusion graph
of subgroups is K1,3 -free.
Theorem 2.12. Let G be a finite group and p, q be distinct primes. Then
(1) girth(I(G)) ∈ {3, 6, ∞};
(2) I(G) is K1,3 -free if and only if G is one of Zpα (α = 2, 3, 4), Zpα q (α = 1, 2), Zpqr , Zp × Zp
or Zq o Zp ;
To prove the above theorem, we start with the following result.
Proposition 2.1. Let G be a group of order pα , where p is a prime and α ≥ 2. Then
(1) girth(I(G)) ∈ {3, 6, ∞};
(2) I(G) is K1,3 -free if and only if G is either Zpα (α = 2, 3, 4) or Zp × Zp .
16
Proof. Proof is divided in to two cases.
Case 1: Let α ≥ 4. Then G has a chain of subgroups of length at least four and so I(G)
contains C3 as a subgraph. Hence girth(I(G)) = 3. If G ∼
= Zpα , then by Theorem 2.4, I(G)
is K1,3 -free only when α = 2, 3, 4. If G Zpα , then G has at least two subgroups of order
pα−1 , let them be H1 , H2 . Also since |H1 ∩ H2 | = pα1 −2 , H1 ∩ H2 has a subgroup of order p1 ,
let it be H3 . It follows that I(G) contains K1,3 as a subgraph with bipartition X := {H3 }
and Y := {H1 , H2 , H1 ∩ H2 }.
Case 2: If α ≤ 3, then G is isomorphic to one of Zpα , Zp × Zp , Zp2 × Zp , Zp × Zp × Zp , Q8 ,
M8 , Mp3 (p > 2) or (Zp × Zp ) o Zp .
(i) If G ∼
= Zp × Zp × Zp := ha, b, c | ap = bp = cp = 1, ab = ba, ac = ca, bc = cbi, then
I(G) contains K1,3 as a subgraph with bipartition X := {ha, bi} and Y := {hai, hbi,
habi}. By Case 2 in the proof of Theorem 2.8, we proved that I(G) contains C6 as
a subgraph. Also by Corollary 2.2(2), I(G) is bipartite and so girth(I(G)) = 4 or 6.
Suppose that I(G2 ) contains C4 as a subgraph. Let it be H1 − H2 − H3 − H4 − H1 .
Then either H1 , H3 ⊂ H2 , H4 or H2 , H4 ⊂ H1 , H3 . Without loss of generality, we may
assume that H1 , H3 ⊂ H2 , H4 . Then we must have |H1 | = p = |H3 |, |H2 | = p2 = |H4 |
and this implies that hH1 , H2 i = H2 = H4 , which is not possible. so girth(I(G)) = 6.
(ii) If G ∼
= (Zp × Zp ) o Zp := ha, b, c | ap = bp = cp = 1, ab = ba, ac = ca, cbc−1 = abi, then
I(G) contains K1,3 as a subgraph with bipartition X := {ha, bi} and Y := {hai, hbi,
habi}. In Case 2 in the proof of Theorem 2.8, we already proved that I(G) contains
C6 as a subgraph. Also by Corollary 2.2(2), I(G) is bipartite and so girth(I(G)) = 4
or 6. As in part (i) of this Case, one can easily show that girth(I(G)) = 6.
(iii) By Theorem 2.4, Corollary 2.2(1), Figures 1(a), 1(b) and (2.1), we see that the girth
of the inclusion graph of subgroups of all the remaining groups is infinity and except
Zp2 × Zp , they contains K1,3 as a subgraph.
Proof follows from the above two cases.
Proposition 2.2. Let G be a group of order pα q β , where p, q are distinct primes and
α, β ≥ 1. Then
17
(1) girth(I(G)) ∈ {3, 6, ∞};
(2) I(G) is K1,3 -free if and only if G is either Zpq or Zp2 q .
Proof. Proof is divided in to several cases.
Case 1: If α = 1 = β, then G ∼
= Zpq or Zp o Zq . By Corollary 2.2(1), I(G) is totally
disconnected and so I(G) is K1,3 -free and girth(I(G)) = ∞.
Case 2: If α = 2, β = 1, then we need to consider the following subcases.
Subcase 2a: Let G be abelian. If G ∼
= Zp2 q , then by (2.2), I(G) is K1,3 -free and girth(I(G)) =
∞. If G ∼
= Zpq × Zp , then by Figure 1(c), I(G) has K1,3 as a subgraph and girth(I(G)) = 6.
Subcase 2b: Let G be non-abelian. We proceed with the groups considered in Subcase 3b
in the proof of Theorem 2.8.
(i) Since Zp × Zp is a subgroup of G2 , so Zp × Zp together with its proper subgroups forms
K1,3 as a subgraph of I(G2 ). Moreover, we already proved that I(G2 ) contains C6
as a subgraph. By Corollary 2.2(2), I(G2 ) is bipartite and so girth(I(G2 )) = 4 or 6.
Suppose that I(G2 ) contains C4 as a subgraph. Let it be H1 − H2 − H3 − H4 − H1 .
Then either H1 , H3 ⊂ H2 , H4 or H2 , H4 ⊂ H1 , H3 . Without loss of generality we may
assume that H1 , H3 ⊂ H2 , H4 .
• If |H1 | = p1 = |H3 |, |H2 | = p21 = |H4 |, then hH1 , H2 i = H2 = H4 , which is not
possible.
• If |H1 | = p1 = |H3 |, |H2 | = p21 , |H4 | = p1 p2 , then H3 * H4 if H1 ⊂ H4 ; H1 * H4 ,
if H3 ⊂ H4 , which are not possible.
• If |H1 | = p1 = |H3 |, |H2 | = p1 p2 = |H4 |, then also we get a contradiction as above.
• If |H1 | = p1 , |H3 | = p2 , |H2 | = p1 p2 = |H4 |, then hH1 , H2 i = H2 = H4 , which is
not possible.
Hence girth(I(G2 )) = 6.
(ii) We already proved that I(G5(t) ) contains C6 as a subgraph. Also Zp × Zp is a subgroup
of G5(t) , so Zp × Zp together with its proper subgroups forms K1,3 as a subgraph of
18
I(G5(t) ). By Corollary 2.2(2), I(G5(t) ) is bipartite and so girth(I(G)) = 4 or 6. As in
part (i) of this Subcase, one can easily show that girth(I(G)) = 6.
(iii) By Figures 1(d), 1(e), 1(f), (2.3), (2.4), we see that the girth of the inclusion graph of
subgroups of all the remaining groups is infinity and they contains K1,3 as a subgraph.
Case 3: Let α ≥ 3 and β = 1. Then G has a chain of subgroups of length at least four and
so girth(I(G)) = 3. If G is cyclic, then let H1 , H2 , H3 , H4 be the subgroups of G of orders p,
q, p2 , p2 q respectively. Then I(G) contains K1,3 as a subgraph with bipartition X := {H4 }
and Y := {H1 , H2 , H3 }. If G is non-cyclic abelian, then Zpq × Zp is a proper subgroup of G,
so by Subcase 2a, I(G) contains K1,3 as a subgraph.
Now we assume that G is non-abelian. Let P denote a Sylow p-subgroups of G. We shall
prove that I(G) contains K1,3 as a subgraph. First let α1 = 3. If p > q, then np (G) = 1,
by Sylow’s theorem and our group G ∼
= P o Zq . Suppose I(P ) is contains K1,3 , then
I(G) also contains K1,3 , so it is enough to consider the case when I(G) is K1,3 -free. By
3
Proposition 2.1, we must have P ∼
= Zp3 . Then G ∼
= Zp3 o Zq := ha, b | ap = bq = 1, bab−1 =
ai , ordp3 (i) = qi and I(G) contains K1,3 as a subgraph with bipartition X := {hap , bi} and
2
Y := {hap i, hap i, hbi}.
Now, let us consider the case p < q and (p, q) 6= (2, 3). Here nq (G) = p is not possible.
If nq (G) = p2 , then q|(p + 1)(p − 1) which implies that q|(p + 1) or q|(p − 1). But this
is impossible, since q > p > 2. If nq (G) = p3 , then there are p3 (q − 1) elements of order
q. However, this only leaves p3 q − p3 (q − 1) = p3 elements, and the Sylow p-subgroup
must be normal, a case we already considered. Therefore, the only remaining possibility
is that G ∼
= Zq o P . Suppose I(P ) contains K1,3 , then so is I(G). So it is enough to
consider the case when I(G) is K1,3 -free. By Proposition 2.1, we must have P ∼
= Zp3 . Then
3
G∼
= Zq o Zp3 = ha, b | aq = bp = 1, bab−1 = ai , ordq (i) = pi, p | (q − 1) and I(G) contains
2
K1,3 as a subgraph with bipartition X := {ha, bp i} and Y := {hbp i, hbp i, hai}.
If (p, q) = (2, 3), then [5, p.160] states that the only group of order 24 that is not a
semi-direct product is S4 . Also A4 is a subgroup of S4 . So by (2.4), I(A4 ) contains K1,3 as
a subgraph. Thus the result is true when α = 3.
Next, let α1 > 3. Then P has a chain of subgroups of length at least 4 and so I(G) has
19
K4 as a subgraph. This implies that I(G) has K1,3 as a subgraph.
Case 4: Let α1 , β ≥ 2. Since G is solvable, so it has a normal subgroup with prime index,
say p2 , let it be H. Let H1 , H2 , H3 be subgroups of H of order p, p2 , q respectively, such that
H1 ⊂ H2 . Then H, H1 , H2 forms C3 as a subgraph of I(G) and so girth(I(G)) = 3. Also,
I(G) contains K1,3 as a subgraph with bipartition X := {H} and Y := {H1 , H2 , H3 }.
Proposition 2.3. Let G be a solvable group whose order has at least three distinct prime
factors. Then
(1) girth(I(G)) ∈ {3, 6};
(2) I(G) is K1,3 -free if and only if G ∼
= Zpqr , where p, q, r are distinct primes.
Proof. Let |G| = pα1 1 pα2 2 · · · pαk k , where pi ’s are distinct primes, k ≥ 3, αi ≥ 1.
Case 1: Let k = 3. If α1 = α2 = α3 = 1, then in Case 4 in the proof of Theorem 2.8,
we already proved that I(G) contains C6 as a subgraph. Also by Corollary 2.2(2), I(G) is
bipartite and so girth(I(G)) = 4 or 6. By using a similar argument as in (i) of Subcase 2b
in the proof of Proposition 2.2, one can easily see that girth(I(G)) = 6. If G ∼
= Zpqr , then
by Theorem 2.8(1), I(G) is K1,3 -free. If G is non-abelian, then G has a non-cyclic subgroup
of composite order, let it be H. It follows that H together with its proper subgroups forms
K1,3 as a subgraph of I(G). Now let α1 ≥ 2, α2 , α3 ≥ 1. Since G is solvable, so it has a
subgroup of order pα1 1 p2 , let it be H. Let H1 , H2 , H3 be subgroups of H of order p1 , pα1 1 ,
p2 respectively with H1 ⊂ H2 . It follows that girth(I(G)) = 3 and I(G) contains K1,3 as a
subgraph with bipartition X := {H} and Y := {H1 , H2 , H3 }.
Case 2: k ≥ 4. Since G is solvable, so it has a subgroup of order pα1 1 pα2 2 pα3 3 , let it be H. Let
H1 , H2 , H3 be subgroups of H of order p1 , pα1 1 p2 , p2 respectively with H1 ⊂ H2 . It follows
that girth(I(G)) = 3 and I(G) contains K1,3 as a subgraph with bipartition X := {H} and
Y := {H1 , H2 , H3 }.
Proof follows from the above two cases.
It is well known that any non-solvable group has a simple group as a sub-quotient and
every simple group has a minimal simple group as a sub-quotient. So if we can show that the
inclusion graph of subgroups of a minimal simple group contains a graph X as a subgraph,
20
then by Theorem 2.3, the inclusion graph of subgroups of a non-solvable group also contains
X.
Recall that SLm (n) is the group of m × m matrices having determinant 1, whose entries
are lie in a field with n elements and that Lm (n) = SLm (n)/H, where H = {kI|k m = 1}.
For any prime q > 3, the Suzuki group is denoted by Sz(2q ). For any integer n ≥ 3, the
dihedral group of order 2n is given by D2n = ha, b | an = b2 = 1, ab = ba−1 i. Note that K1,3
is a subgraph of I(D4n ) with bipartition X := {ha2 , bi} and Y := {ha2 i, ha2 bi, hbi}.
Lemma 2.1. I(D4n ) contains K1,3 as a subgraph, for n ≥ 3.
Proposition 2.4. If G is a finite non-solvable group, then I(G) has K1,3 as a subgraph and
girth(I(G)) = 3.
Proof. If we show that I(G) contains K2 + K 3 as a subgraph, then it follows that I(G)
contains K1,3 as a subgraph and girth(I(G)) = 3. As mentioned above, to prove this, it
is enough to show that the inclusion graph of subgroups of minimal simple groups contains
K2 +K 3 as a subgraph. We use the J. G. Thompson’s classification of minimal simple groups
given in [13] and check this for this list of groups. We will denote the image of a matrix A
in Lm (n) by A.
Case 1: G ∼
= L2 (q p ). If p = 2, then the only non-solvable group is L2 (4) and L2 (4) ∼
= A5 .
Also A4 is a subgroup of A5 , and so by (2.4), A4 together with its proper subgroups forms
K2 + K 3 as a subgraph of I(G). If p > 2, then L2 (q p ) contains a subgroup isomorphic to
(Zq )p , namely the subgroup of matrices of the form ( 10 a1 ) with a ∈ Fqp . By Proposition 2.1,
I((Zq )p ), p > 2 contains K1,3 as a subgraph and so (Zq )p together with proper its subgroups
forms K2 + K 3 as a subgraph of I(G).
Case 2:: G ∼
= L3 (3). Note that L3 (3) ∼
= SL3 (3). Let us consider the subgroup consisting
1 a b
of matrices of the form 0 1 c with a, b, c ∈ F3 . This subgroup is isomorphic to the group
0 0 1
(Zp ×Zp )oZp with p = 3. By Proposition 2.1, I((Zp ×Zp )oZp ) contains K1,3 as a subgraph.
So (Zp × Zp ) o Zp together with its proper subgroups forms K2 + K 3 as a subgraph of I(G).
Case 3: G ∼
= L2 (p). Note that here H = {±I}. We have to consider two subcases:
Subcase 3a: p ≡ 1 (mod 4). Then L2 (p) has a subgroup isomorphic to Dp−1 [4, p. 222]. So,
by Lemma 2.1, Dp−1 together with its proper subgroups forms K2 + K 3 as a subgraph of
21
I(G), when p > 5. If p = 5, then L2 (5) ∼
= A5 ∼
= L2 (4), which we already dealt.
Subcase 3b: p ≡ 3 (mod 4). L2 (p) has a subgroup isomorphic to Dp+1 [4, p. 222]. By
Lemma 2.1, Dp+1 together with its proper subgroups forms K2 + K 3 as a subgraph of I(G),
when p > 7. If p = 7, then S4 is a maximal subgroup of L2 (7). Since A4 is a subgroup of
S4 , we already dealt this case.
Case 4: G ∼
= Sz(2q ). Then Sz(2q ) has a subgroup isomorphic to (Z2 )q , q ≥ 3. By Proposition 2.1, I((Z2 )q ), q > 2 contains K1,3 as a subgraph; (Z2 )q together with its proper
subgroups forms K2 + K 3 as a subgraph of I(G).
Proof follows by putting together all the above cases.
Combining the Propositions 2.1 – 2.4, we obtain the proof of Theorem 2.12.
References
[1] S. Akbari, M. Habibi, A. Majidinya, and R. Manaviyat, The inclusion ideal graph of
rings, Comm. Algebra 43 (2015) 2457–2465.
[2] S. Akbari, F. Heydari and M. Maghasedi, The intersection graph of group, J. Algebra.
Appl. 5 (2015) Article No. 1550065, 9 pp.
[3] Atlas of Finite Group Representations, http://brauer.maths.qmul.ac.uk/Atlas/
(version 2)
[4] J. P. Bohanon and Les Reid, Finite groups with planar subgroup lattices, J. Algebraic
Combin. 23 (2006) 207–223.
[5] W. Burnside, Theory of groups of finite order, Dover Publications, Cambridge, 1955.
[6] B. Csákány and G. Pollák, The graph of subgroups of a finite group (Russian),
Czechoslovak Math. J. 19 (1969) 241–247.
[7] F. N. Cole and J. W. Glover. On groups whose orders are products of three prime
factors, Amer. J. Math. 15 (1893) 191–220.
[8] F. Harary, Graph Theory, Addison-Wesley, Philippines, 1969.
22
[9] Rulin Shen and Enshi, Intersection graphs of subgroups of finite groups, Czechoslovak
Math. J. 60 (2010) 945–950.
[10] R. Schmidt, Subgroup lattices of groups, Expositions in Math., vol. 14, de Gruyter,
1994.
[11] R. Schmidt, Planar subgroup lattices, Algebra Universalis 55 (2006) 3-12.
[12] C. L. Starr and G. E. Turner III, Planar groups, J. Algebraic Combin. 19 (2004), 283295.
[13] J. G. Thompson, Nonsolvable finite groups all of whose local subgroups are solvable I,
Bull. Amer. Math. Soc. 74 (1968) 383–437.
| 4math.GR
|
1
On Submodularity and Controllability in
Complex Dynamical Networks
arXiv:1404.7665v3 [math.OC] 30 Sep 2016
Tyler H. Summers, Fabrizio L. Cortesi, and John Lygeros
Abstract—Controllability and observability have long been
recognized as fundamental structural properties of dynamical
systems, but have recently seen renewed interest in the context of
large, complex networks of dynamical systems. A basic problem is
sensor and actuator placement: choose a subset from a finite set of
possible placements to optimize some real-valued controllability
and observability metrics of the network. Surprisingly little is
known about the structure of such combinatorial optimization
problems. In this paper, we show that several important classes of
metrics based on the controllability and observability Gramians
have a strong structural property that allows for either efficient
global optimization or an approximation guarantee by using a
simple greedy heuristic for their maximization. In particular, the
mapping from possible placements to several scalar functions
of the associated Gramian is either a modular or submodular
set function. The results are illustrated on randomly generated
systems and on a problem of power electronic actuator placement
in a model of the European power grid.
I. I NTRODUCTION
Controllability and observability have been recognized as
fundamental structural properties of dynamical systems since
the seminal work of Kalman in 1960 [19], but have recently seen renewed interest in the context of large, complex
networks, such as power grids, the Internet, transportation
networks, and social networks. A prominent example of this recent interest is [26], which, based on Kalman’s rank condition
and the idea of structural controllability [25], presents a graph
theoretic maximum matching method to efficiently identify a
minimal set of so-called driver nodes through which timevarying control inputs can move the system around the entire
state space (i.e., render the system controllable). The method of
[26] is applied across a range of technological and social systems, leading to several interesting and surprising conclusions.
Using a metric of controllability given by the fraction of driver
nodes in the minimal set required for complete controllability,
it is shown that sparse inhomogeneous networks are difficult to
control while dense homogeneous networks are easier. It is also
shown that the minimum number of driver nodes is determined
mainly by the degree distribution of the network. However,
there is an implicit assumption about the diagonal entries of
the dynamics matrix that restrict the application of their result
[10]. Many other studies of controllability in complex networks
have followed, including [37], [31], [45], [47], [43].
The authors are with the Automatic Control Laboratory, ETH Zurich,
Switzerland, email: {tsummers,jlygeros}@control.ee.ethz.ch This work is
partially supported by the ETH Zurich Postdoctoral Fellowship Program.
Preliminary versions of results in this paper appeared in [40], [9]. Here,
we present the work in a unified framework, provide modified and simplified
proofs of the main results, and revise and elaborate on the numerical examples.
One issue with the approach taken by [26] and much of the
follow-up work is that the exclusive focus on structural controllability and the associated quantitative notion of controllability
(namely, the number/fraction of required driver nodes) can be
rather crude and even misleading in some settings. This was
noted, for example, by [29] in response to the surprising result
in [26] that genetic regulatory networks seem to require many
driver nodes, which apparently contradicts other findings in
biological literature on cellular reprogramming. This suggests
that rather than finding a set of driver nodes that would render
a network completely controllable, a more appropriate strategy
might be to choose from a finite set of possible actuator and
sensor placements the subset that optimizes some real-valued
controllability and observability metrics of the network. There
is a variety of more sophisticated metrics for controllability
and observability that have been proposed in the systems and
control literature on sensor and actuator placement or selection
problems in dynamical systems; see, e.g., the survey paper
[44]. One important class of metrics involves the controllability
and observability Gramians, which are symmetric positive
semidefinite matrices whose structure relates energy notions
of controllability and observability. The use of Gramians as
quantitative metrics of controllability in networks is studied
in [37], [46], [4], [42], [34]. Other important studies of
controllability in networks include [39], [36], [33].
While a variety of metrics have been proposed in the
literature [44], the corresponding combinatorial optimization
problems for sensor and actuator placement are less wellunderstood. These can be solved by brute force for small
problems by testing all possible placement combinations.
However, for problems that arise in large networks, testing all
combinations quickly becomes infeasible. Only very recently
in the context of large networks have researchers started to
investigate combinatorial properties of sensor and actuator
placement problems for optimizing system dynamics and control metrics.
Clark et al. have recently considered a related but different problem of leader selection in networks with consensus
dynamics, in which a set of leader states are selected to act
as control inputs to the system [7], [8], [6]. In [7], [8], it is
shown that the minimum mean square error due to link noise is
a supermodular function of the leader set. In [6] it is shown that
a graph controllability index, which is related to the structural
controllability framework of [25], is a submodular function of
the leader set. As discussed below, these properties allow for
suboptimality guarantees using simple greedy algorithms.
In the present paper, we show that one important class of
metrics of controllability and observability, previously thought
to lead to difficult combinatorial optimization problems [44],
2
can be in fact easily optimized, even for very large networks.
In particular, we show that the mapping from subsets of
possible actuator/sensor placements to any linear function of
the associated controllability or observability Gramian has
a strong structural property: it is a modular set function.
Furthermore, we show that the rank of the Gramian, the log
determinant of the Gramian, and the negative trace of the
inverse Gramian are submodular set functions. Although maximization of submodular functions is difficult, submodularity
allows for an approximation guarantee if one uses a simple
greedy heuristic for their maximization [30]. We also describe
how these observations define new dynamic network centrality
measures for networks whose dynamics are described by
linear models, assigning a control energy-related “importance”
value to each node in the network. We illustrate the results
on randomly generated systems and on a problem of power
electronic actuator placement in a model of the European
power grid.
The rest of the paper is organized as follows. Section II
reviews the network model and Gramian-based controllability
metrics. Section III introduces the notions of modular, submodular, and supermodular set functions and shows our main
results that several set functions mapping possible actuator
placements to various functions of the controllability Gramian
are either modular or submodular. Section IV presents illustrative numerical examples. Finally, Section V gives concluding
remarks and and outlook for future research.
II. L INEAR M ODELS OF N ETWORK DYNAMICS
This section defines a linear model for network dynamics
and reviews and interprets metrics for controllability based
on the controllability Gramian. The material in this section
is mostly standard and can be found in many texts on linear
system theory, e.g. [17], [3]; we discuss the material mostly to
set our notation. Since controllability and observability are dual
properties [18], we focus only on controllability and actuator
placement; all of the results have analogous counterparts and
interpretations for observability and sensor placement.
In the literature on controllability in networks, it is common
to start with linear network models. In this spirit, we focus on
linear, time-invariant dynamical network models, in which the
dynamics are given by
ẋ(t) = Ax(t) + Bu(t),
y(t) = Cx(t),
x(0) = x0 ,
(1)
where x(t) ∈ Rn are the states of the network, u(t) ∈ Rm are
the control inputs that can be used to influence the network
dynamics, y(t) ∈ Rp are the outputs, and A, B, and C are
constant matrices of appropriate dimensions. We assume that
C is full row rank. For example, x(t) might represent voltages,
currents, or frequencies in devices in a power grid, chemical
species concentrations in a genetic regulatory network, or
individual opinions or propensities for product adoption in a
social network. The matrix C ∈ Rp×n is typically interpreted
as a set of linear state measurements, but here we will interpret
it as a weight matrix whose rows define important directions
in the state space.
The dynamics matrix A induces a graph G(V, E) of the
network in which the vertices correspond to states, i.e., V =
{1, ..., n} and the edges correspond to non-zero entries of A,
i.e., (i, j) ∈ E whenever aji 6= 0. The non-zero entries of
the input matrix B describe how each actuator affects the
nodes in the network. When optimizing actuator placement,
one effectively designs a network structure by connecting
sets of inputs to sets of network nodes to optimize a metric
controllability for the resulting network.
A. Controllability
Definition 1 (Controllability). A dynamical system is called
controllable over a time interval [0, t] if given any states x0 ,
x1 ∈ Rn , there exists an input u(·) : [0, t] → Rm that drives
the system from x0 at time 0 to x1 at time t.
Kalman’s well-known rank condition states that a linear dynamical system is controllable if and only if
[B, AB, ..., An−1 B] is full rank. Since rank is a generic
property of a matrix, it has the same value for almost all values
of the non-zero entries of A and B (assuming that the nonzero entries are independent). This suggests that controllability
can be cast as a structural property of the graph defined
by A and B, as captured in the graph-theoretic concept
of structural controllability described by Lin in [25], which
underpins the recent results of [26]. However, it is informative
and practically relevant to consider more quantitative metrics
for controllability in complex networks.
B. Energy-related controllability metrics
Actuators in real systems are usually energy limited, so an
important class of metrics of controllability deals with the
amount of input energy required to reach a given state from
the origin. The symmetric positive semidefinite matrix
Z t
T
Wc (t) =
eAτ BB T eA τ dτ ∈ Rn×n
(2)
0
is called the controllability Gramian at time t and provides
an energy-related quantification of controllability. Eigenvectors
of Wc associated with small eigenvalues (large eigenvalues of
Wc−1 ) define directions in the state space that are less controllable (require large input energy to reach), and eigenvectors
of Wc associated with large eigenvalues (small eigenvalues
of Wc−1 ) define directions in the state space that are more
controllable (require small input energy to reach).
For stable systems, the state transition matrix eAt comprises
decaying exponentials, so a finite positive definite limit of the
controllability Gramian always exists and is given by
Z ∞
T
Wc =
eAτ BB T eA τ dτ ∈ Rn×n .
(3)
0
This infinite-horizon controllability Gramian can be computed
by solving a Lyapunov equation
AWc + Wc AT + BB T = 0,
(4)
3
which is a system of linear equations and is therefore easily
solvable. Specialized algorithms have been developed to compute the solution [1], [15], [24] and scale to large networks.
We will focus on the infinite-horizon Gramian due to ease
of computation. However, all of our results also apply to the
finite-horizon Gramian, with the only disadvantage being that
one must evaluate (2) rather than solve (4), which may be more
difficult for large networks. An advantage of the finite-horizon
Gramian is that it can be used for unstable systems.
An alternative definition and interpretation of the Gramian
for unstable systems can be used to quantify controllability
[48], but we do not discuss this interpretation further in the
interest of simplicity. It is worth keeping in mind that though
the results are stated for asymptotically stable systems, they
apply more generally.
The controllability Gramian gives a more sophisticated
energy-related quantitative picture of controllability, but we
still need to form a scalar metric for Wc , which is a positive
semidefinite matrix. We want Wc “large” so that Wc−1 is
“small”, requiring small amount of input energy to move
around the state space. There are a number of possible metrics
for the size of Wc , several of which we now discuss.
1) tr(Wc ): The trace of the Gramian is inversely related
to the average energy and can be interpreted as the average
controllability in all directions in the state space. It is also
closely related to the system H2 norm:
Z
kHk22 = tr C
∞
T
eAt BB T eA t dtC T
0
(5)
T
= tr(CWc C ),
4) λmin (Wc ): The smallest eigenvalue of the Gramian is
a worst-case metric inversely related to the amount of energy
required to move the system in the direction in the state space
that is most difficult to control.
5) rank(Wc ): The rank of the Gramian is the dimension of
the controllable subspace.
Remark 1. Our main results and much of the discussion
generalize straightforwardly to linear time-varying systems.
The only differences are that the Gramian depends on both
initial and final time, rather than just their difference, and
that it must be computed by integrating (2), rather than by
solving a Lyapunov equation.
In the following section, we briefly review the combinatorial
notion of submodularity and consider which of the above
controllability metrics have a submodularity property, which
provides global optimality or approximation guarantees for
associated actuator selection problems.
III.
O PTIMAL S ENSOR AND ACTUATOR P LACEMENT IN
N ETWORKS
A. Set Functions, Modularity, and Submodularity
Sensor and actuator placement problems can be formulated
as set function optimization problems. For a given finite set
V = {1, ..., M }, a set function f : 2V → R assigns a real
number to each subset of V . In our setting, the elements of V
represent potential locations for the placement of actuators in
a dynamical system, and the function f is a metric for how
controllable the system is for a given set of placements.
We consider set function optimization problems of the form
maximize
i.e., the system H2 norm is a weighted trace of the controllability Gramian.
2) tr(Wc−1 ): The trace of the inverse controllability
Gramian is proportional to the energy needed on average to
move the system around on the state space. Note that when the
system is uncontrollable, the inverse Gramian does not exist
and the average energy is infinite because there is at least one
direction in which it is impossible to move the system using
the inputs. In this case, one could consider the trace of the
pseudoinverse, which is the average energy required to move
the system around the controllable subspace.
3) log det Wc : The determinant of the Gramian is related to
the volume enclosed by the ellipse it defines
V (Emin ) =
p
π n/2
n
det Wc ,
Γ(n/2 + 1)
where Γ is the Gamma function. This means that the determinant is a volumetric measure of the set of states that
can be reached with one unit or less of input energy. Since
determinant is numerically problematic in high dimensions,
and because the logarithm is monotone, we will consider
optimizing log det Wc . Note that for uncontrollable systems,
the ellipsoid volume is zero, so log det Wc = −∞. In this case,
one could consider the associated volume in the controllable
subspace.
S⊆V, |S|=k
f (S).
(6)
The problem is to select a k-element subset of V that maximizes f . This is a finite combinatorial optimization problem,
so one can solve it by brute force simply by enumerating all
possible subsets of size k, evaluating f for all of these subsets,
and picking the best subset. However, we are interested in
cases arising from complex networks in which the number of
possible subsets is very large. The number of possible subsets
grows factorially as |V | increases, so the brute force approach
quickly becomes infeasible as |V | becomes large.
We focus instead on structural properties of the set function
f that make it more amenable to optimization. In particular,
submodularity plays a similar role in combinatorial optimization to convexity in continuous optimization and shares other
features of concave functions [27], [21]. It occurs often in
applications [2], [20], [22] (though is underexplored in systems
and control theory); is preserved under various operations,
allowing design flexibility; is supported by an elegant and
practically useful mathematical theory; and there are efficient
methods for minimizing and approximation guarantees for
maximizing submodular functions.
Definition 2 (Submodularity). A set function f : 2V → R
is called submodular if for all subsets A ⊆ B ⊆ V and all
elements s ∈
/ B, it holds that
f (A ∪ {s}) − f (A) ≥ f (B ∪ {s}) − f (B),
(7)
4
or equivalently, if for all subsets A, B ⊆ V , it holds that
f (A) + f (B) ≥ f (A ∪ B) + f (A ∩ B).
(8)
Intuitively, submodularity is a diminishing returns property
where adding an element to a smaller set gives a larger gain
than adding one to a larger set. This is made precise by the
following result from [27], which will be useful for verifying
submodularity of set functions later.
Definition 3. A set function f : 2V → R is called monotone
increasing if for all subsets A, B ⊆ V it holds that
A ⊆ B ⇒ f (A) ≤ f (B),
(9)
and is called monotone decreasing if for all subsets A, B ⊆ V
it holds that
A ⊆ B ⇒ f (A) ≥ f (B).
(10)
Theorem 1 ([27]). A set function f : 2V → R is submodular
if and only if the derived set functions fa : 2V −{a} → R
fa (X) = f (X ∪ {a}) − f (X)
are monotone decreasing for all a ∈ V .
A set function is called supermodular if the reversed inequalities in (7) and (8) hold, and is called modular if it is both
submodular and supermodular, i.e., for all subsets A, B ⊆ V ,
we have f (A ∩ B) + f (A ∪ B) = f (A) + f (B). A modular
function has the following simple, equivalent characterization
[27]:
Theorem 2 (Modularity [27]). A set function f : 2V → R
is modular if and only if for any subset S ⊆ V it can be
expressed as
X
f (S) = w(∅) +
w(s)
(11)
s∈S
for some weight function w : V → R.
Modular set functions are analogous to linear functions and
have the property that each element of a subset gives an
independent contribution to the function value. Consequently,
one can see that if f is modular, the optimization problem (6)
is easily solved by simply evaluating the set function for each
individual element, sorting the result, and then choosing the
top k individual elements from the sorted list to obtain the best
subset of size k.
Maximization of monotone increasing submodular functions
is NP-hard, but the so-called greedy heuristic can be used to
obtain a solution that is provably close to the optimal solution
[30]. The greedy algorithm for (6) starts with an empty set,
S0 ← ∅, computes the gain ∆( a | Si ) = f (Si ∪ {a}) − f (Si )
for all elements a ∈ V \Si and adds the element with the
highest gain:
Si+1 ← Si ∪ {arg max ∆( a | Si ) | a ∈ V \Si }.
a
The algorithm terminates after k iterations.
Performance of the greedy algorithm is guaranteed by a well
known bound [30]:
Theorem 3 ([30]). Let f ∗ be the optimal value of the set
function optimization problem (6), and let f (Sgreedy ) be
the value associated with the subset Sgreedy obtained from
applying the greedy algorithm on (6). If f is submodular and
monotone increasing, then
k
f ∗ − f (Sgreedy )
1
k−1
≤ ≈ 0.37.
≤
(12)
f ∗ − f (∅)
k
e
This is the best any polynomial time algorithm can achieve
[11], assuming P 6= N P . Note that this is a worst-case bound;
the greedy algorithm often performs much better than the
bound in practice.
We now demonstrate the modularity or submodularity of
several classes of controllability metrics involving functions
of the controllability Gramian.
Recall that the space of symmetric n × n matrices S n is
partially ordered by the semidefinite partial order: W1 W2
if W1 − W2 0. The space of symmetric positive definite
n
matrices is denoted S++
and the space of symmetric positive
n
semidefinite matrices is denoted S+
.
B. Trace of the Gramian
Suppose A ∈ Rn×n is a stable system dynamics matrix
and V = {b1 , ..., bM } is a set of possible columns that
can be used to form or modify the system input matrix B.
The problem is to choose a subset of V to maximize a
metric of controllability. We now consider a linear function
of the controllability Gramian, which can be expressed as a
weighted trace of the controllability Gramian. For a given
S ⊆ V , we form BS = [B0 bs ] given a (possibly empty)
existing matrix B0 and using the associated columns defined
by s ∈ S,
and we denote the associated controllability Gramian
R∞
T
WS = 0 eAτ BS BST eA τ dτ , which is the unique positive
semidefinite solution the Lyapunov equation
AWS + WS AT + BS BST = 0.
(13)
To simplify notation, we write Ws for W{s} . We have the
following result.
Theorem 4. The set function mapping subsets S ⊆ V to
a linear function of the associated controllability Gramian,
f (S) = tr(CWS C T ) for any weighting matrix C ∈ Rp×n , is
modular.
Proof: We will prove the result directly using Theorem 2.
For any S ⊆ V it is easy to see that the controllability Gramian
associated with BS is simply a sum of the controllability
Gramians associated with the individual columns of BS :
Z ∞
T
WS =
eAτ BS BST eA τ dτ
Z0 ∞
X
T
=
eAτ
bs bTs eA τ dτ
0
=
=
s∈S
XZ
∞
s∈S
0
X
Ws .
s∈S
T
eAτ bs bTs eA
(14)
τ
dτ
5
Now since trace is a linear matrix function, we have for any
weight matrix C ∈ Rp×n
f (S) = tr(CWS C T )
!
X
= tr
CWs C T
(15)
s∈S
=
X
tr(CWs C T ).
s∈S
Thus, for any s ∈ V , we can define the weight function
w(s) = tr(CWs C T ). Defining w(∅) = 0, Theorem 2 implies
that f (S) = tr(CWS C T ) is a modular set function.
Theorem 4 shows that each possible actuator placement
gives an independent contribution to the trace of the controllability Gramian. Because of this, the actuator placement
problem using this metric is easily solved: one needs only
to compute the metric individually for each possible actuator
placement, sort the results, and choose the best k. Based on the
interpretations in the previous section, this means that placing
actuators in a complex network to maximize the average
controllability available to move the system around the state
space, or to maximize the energy in the system response to a
unit impulse, is easily done. Since the result holds for the
weighted trace, this gives considerable design freedom for
actuator placement; important directions in the state space
can be weighted and actuator placement done based on the
weighted metric.
C. Trace of the inverse Gramian
We now consider the trace of the inverse of the controllability Gramian. We assume in this subsection that for any S ⊆ V
the associated Gramian WS is invertible. This is the case, for
example, if the network already has a set of actuators that
provide controllability and we would like to add additional
actuators to improve controllability. We discuss how to deal
with non-invertibility of the Gramian in Section III.E.
Theorem 5. Let V = {b1 , ..., bM } be a set of possible input
matrix columns and WS the controllability Gramian associated
with S ⊆ V . The set function f : 2V → R defined as
f (S) = − tr(WS−1 )
is submodular and monotone increasing.
Proof: We will use Theorem 1 to prove the result. Fix
an arbitrary a ∈ V and consider the derived set function fa :
2V −{a} → R defined by
−1
−1
fa (S) = − tr((WS∪{a} ) ) + tr((WS ) )
= − tr((WS + Wa )−1 ) + tr((WS )−1 ).
Take any S1 ⊆ S2 ⊆ V − {a}. By the additivity property of
the Gramian noted in Theorem 4 in (14), it is clear that S1 ⊆
S2 ⇒ WS1 WS2 . Define W (γ) = WS1 + γ(WS2 − WS1 )
for γ ∈ [0, 1]. Clearly, W (0) = WS1 and W (1) = WS2 . Now
define
fˆa (W (γ)) = − tr((W (γ) + Wa )−1 ) + tr((W (γ))−1 ).
Note that fˆa (W (0)) = fa (S1 ) and fˆa (W (1)) = fa (S2 ). We
have
d ˆ
d
fa (W (γ)) =
− tr((W (γ) + Wa )−1 ) + tr(W (t)−1 )
dγ
dγ
= tr (W (γ) + Wa )−1 (WS2 − WS1 )(W (γ) + Wa )−1
− tr W (γ)−1 (WS2 − WS1 )W (γ)−1
−2
−2
= tr (W (γ) + Wa ) − W (γ)
(WS2 − WS1 ) ≤ 0.
To obtain the second equality we used the matrix derivative
d
d
X(γ)−1 = X(γ)−1 dγ
(X(γ))X(γ)−1 [35]. To
formula dγ
obtain the third equality we used the cyclic property of trace.
Since (W (γ) + Wa )−2 − W (γ)−2 0 and WS2 − WS1 0,
the last inequality holds because the trace of the product of a
positive and negative semidefinite matrix is non-positive. Since
Z 1
d ˆ
ˆ
ˆ
fa (W (1)) = fa (W (0)) +
fa (W (γ))dγ,
dγ
0
it follows that fˆa (W (1)) = fa (S2 ) ≤ fˆa (W (0)) = fa (S1 ).
Thus, fa is monotone decreasing, and f is submodular by
Theorem 1.
Finally, it can be seen from additivity of the Gramian (14)
that f is monotone increasing, which just means that adding
an actuator to the system cannot decrease its controllability.
D. Log determinant of the Gramian
We now consider the log determinant of the controllability
Gramian. We assume again that for any S ⊆ V the associated
Gramian is invertible. We have the following result.
Theorem 6. Let V = {b1 , ..., bM } be a set of possible input
matrix columns and WS the controllability Gramian associated
with S ⊆ V . The set function f : 2V → R, defined as
f (S) = log det WS
is submodular and monotone increasing.
Proof: The proof uses the same idea as before, namely,
showing that the derived set functions fa : 2V −{a} → R
fa (S) = log det WS∪{a} − log det WS
= log det(WS + Wa ) − log det WS
are monotone decreasing for any a ∈ V . For arbitrary a ∈ V ,
S1 ⊆ S2 ⊆ V − {a}, define again W (γ) = WS1 + γ(WS2 −
WS1 ) for γ ∈ [0, 1] and
fˆa (W (γ)) = log det(W (γ) + Wa ) − log det W (γ).
We have
d ˆ
fa (W (γ))
dγ
d
=
[log det(W (γ) + Wa ) − log det W (γ)]
dγ
= tr[(W (γ) + Wa )−1 (WS2 − WS1 )]
− tr[W (γ)−1 (WS2 − WS1 )]
= tr[((W (γ) + Wa )−1 − W (γ)−1 )(WS2 − WS1 )]
≤ 0.
6
d
We used the matrix derivative formula dγ
log det X(γ) =
−1 d
tr[X(γ) dγ X(γ)] [35] to obtain the second equality. The
remainder of the proof follows the previous proof.
with S ⊆ V . The set function f : 2V → R, defined as
f (S) = rank(WS )
is submodular and monotone increasing.
V
Corollary 1.√The related set function g : 2 → R defined by
g(S) = log n det WS is submodular and monotone increasing.
Proof: We have
1
g(S) = log det WS .
n
Thus, from Theorem 6 g is a non-negatively scaled version of
a submodular and monotone increasing function and therefore
also submodular and monotone increasing.
Not all directions in the state space may be of equal
importance, one might want to use a weight matrix as an
additional design parameter for an actuator selection problem.
In a simple case, the weight matrix could be a diagonal
matrix, assigning a relative weight to every state. We have
the following corollary; the proof follows exactly the same
arguments as in the previous theorems.
Corollary 2. Let V = {b1 , ..., bM } be a set of possible input
matrix columns and WS the controllability Gramian associated
with S ⊆ V . The set functions g1 , g2 : 2V → R defined as
g1 (S) = log det(CWS C T ),
g2 (S) = − tr[(CWS C T )−1 ],
where C ∈ Rp×n with p ≤ n and rank(C) = p, are
submodular and monotone increasing.
Remark 2. Interestingly, other combinatorial network design
problems unrelated to controllability have a strikingly similar
mathematical structure. Specifically, in [38] it is shown that
problems in which one chooses sets of nodes or edges to
optimize certain rigidity properties of a network, which relate
to formation control and network localization objectives, are
also submodular set function optimization problems for which
greedy algorithms yield solutions with suboptimality guarantees. In that setting, one can define a “rigidity Gramian”
to quantify desirable rigidity properties, and the results and
proofs techniques are nearly identical to what we present in
Theorems 4-6. Furthermore, problems in which one adds sets
of edges to a network to optimize the coherence of the resulting
network, which relates to robustness of consensus process to
additive noise, also have a similar structure [41].
E. Rank of the Gramian
The controllability metrics − tr(WS−1 ) and log det WS fail
to distinguish amongst subsets of actuators that do not yield a
fully controllable system. In particular, these functions are undefined, or are interpreted to return −∞, when the Gramian is
not full rank. One way to handle cases where the controllability
Gramian in not invertible is to consider its rank. The following
result shows that this is also a submodular set function.
Theorem 7. Let V = {b1 , ..., bM } be a set of possible input
matrix columns and WS the controllability Gramian associated
Proof: For two linear transformations V1 , V2 ∈ Rn×n , we
have
rank(V1 + V2 )
= rank(V1 ) + rank(V2 ) − dim(range(V1 ) ∩ range(V2 )).
We can form gain functions fa : 2V −{a} → R
fa (S) = rank(WS∪{a} ) − rank(WS )
(16)
= rank(Wa ) − dim(range(WS ) ∩ range(Wa )).
It is now easy to see that fa is monotone decreasing: the
first term in the second line is constant and the second term
decreases because dim(range(WS )) only increases with S.
That f is monotone increasing is clear from additivity of the
Gramian (14).
Note that Olshevsky has analyzed a greedy algorithm for
maximizing the rank of the controllability matrix, though not
in a submodularity framework [33].
Another way to handle uncontrollable systems is to work
with related continuous metrics defined for uncontrollable
systems, such as the trace of the pseudoinverse tr(WS† ), which
corresponds to the average energy required to move the system
around the controllable subspace, or the log product of nonWS
zero eigenvalues log Πrank
λi (WS ), which relates to the
i=1
“volume” of the subspace reachable with one unit of input
energy.
F. Smallest eigenvalue of the Gramian
We have seen so far that the trace of the Gramian is a
modular (and thus both sub- and supermodular) set function
of actuator subsets and that the trace of the inverse Gramian,
the log determinant of the Gramian, and the rank of the
Gramian are submodular set functions. The first two functions
are also concave matrix functions. Given the connections
between submodular functions and concave functions, one
might be tempted to conjecture that any concave function of
the Gramian corresponds to a submodular function of actuator
subsets. However, we now show by counterexample that this
is false. Consider the set function f (S) = λmin (WS ), which
corresponds to the concave matrix function that returns the
smallest eigenvalue of the Gramian.
We show an example where this function violates the
diminishing gains property (7) of a set function f (S)
∆( s | A ) ≥ ∆( s | B ),
where ∆( s | A ) =
defined by
"
−8
A= 0
7
A ⊆ B ⊆ V, s ∈
/ B,
f (A ∪ {s}) − f (A). Consider the system
0
−2
0
#
−2
−8 ,
−3
BV = [bV ] = I3 .
7
We see that the diminishing returns property holds in some
cases, e.g.,
∆( b3 | {b1 } ) = 0.037 ≥ 0.033 = ∆( b3 | {b1 , b2 } ),
but is violated in others
∆( b3 | {b2 } ) = 0.001 ≤ 0.033 = ∆( b3 | {b1 , b2 } ).
G. Dynamic network centrality measures
Network centrality measures are real-valued functions that
assign a relative “importance” to each node within a graph.
Examples include degree, betweenness, closeness, and eigenvector centrality [32]. The meaning of importance and the
relevance of various metrics depends highly on the modeling
context. For example, PageRank, a variant of eigenvector
centrality, turns out to be a much better indicator of importance
than vertex degree in the context of networks of web pages,
one of the core factors leading to Google’s domination of web
search.
In the context of complex dynamical networks, the controllability metrics described above can be used to define a control
energy-based centrality measures, describing the importance of
a node in terms of its ability to move the system around the
state space with a low-energy time-varying control input. In
particular, given a system dynamics matrix A ∈ Rn×n , imagine that it is possible to place an actuator at each individual
node in the network; thus, define V = {e1 , ..., en }, where ei
is the standard unit basis vector in Rn , i.e., ei has a 1 in
the ith entry and zeros elsewhere. We define several Control
Energy Centrality measures for a complex dynamical network
as follows.
Definition 4 (Control Energy Centralities). Given a complex
network with n nodes and an associated stable linear dynamics
matrix A ∈ Rn×n , we define the following Control Energy
Centrality measures for each node i
• Average Controllability Centrality
CAC (i) = tr(Wi ),
•
i∈V
Average Control Energy Centrality
CACE (i) = −tr(Wi† ),
•
(17)
i∈V
(18)
Volumetric Control Energy Centrality
CV CE (i) = log
rank
YWi
λj (Wi ),
i ∈ V,
(19)
j=1
where Wi is the infinite-horizon controllability Gramian that
satisfies AWi + Wi AT + ei eTi = 0.
These measures provide more relevant quantities of centrality than purely graph based measures in the context of
dynamical systems and control, and can give quite a different
view of what nodes are important. The greedy algorithm for
choosing nodes in which to inject control signals can be
interpreted as choosing the most central node at each iteration,
given the current set of controlled nodes. An interesting topic
for future work would be to explore the distribution of the
Control Energy Centrality measure in random networks and
networks from various application domains.
Pasqualetti et al. have also defined a different network
centrality measure based on the controllability Gramian [34].
In the context of networks with consensus dynamics, Chapman
and Mesbahi have also defined a related network centrality
measure that quantifies effectiveness of each agent in tracking
the mean of a noisy signal [5]. It is possible to define many
other network centrality measures related to network dynamics
and control, e.g., based on the leader selection metrics of Clark
et al. [7], [6].
H. Computational scaling for large networks
In this subsection we discuss computational techniques
that can be used to scale the greedy algorithm described in
Section III-A to large structured networks. First, specialized
algorithms can be used to exploit sparsity and compute low
rank solutions to Lyapunov equations. In particular, when
computing the Gramian associated with an individual actuator,
the Cholesky factor-alternating direction implicit algorithm
of Li and White [24] allows one to exploit both the rankone structure of the constant term in the Lyapunov equation
(viz., bbT ) and any sparsity structure in the network dynamics
matrix A. Moreover, it is often the case in large networks
that the Gramian associated with an individual actuator is low
rank or approximately low rank itself. One can obtain further
computational benefits by computing low rank approximations
of these Gramians also using methods from [24].
Second, several techniques can be used to improve the
greedy algorithm. Each iteration in the standard version is
trivially parallelizable. The Gramians associated with each
possible actuator can be pre-computed using the specialized
methods mentioned above independently and in parallel. Because the Gramian is additive in the actuators, effectively one
can solve the Lyapunov equation for any set of actuators
by solving it in parallel for each individual actuator and
summing the results. Then at each iteration, the marginal
gain of each actuator can also be computed in parallel by
adding its Gramian to the Gramian of the current set of added
actuators and evaluating the metric (trace, logdet, etc.). When
the individual Gramians are low rank, the marginal gains
can be computed more efficiently by using low rank update
formulae, e.g., the Sherman-Morrison formula for trace of
the inverse Gramian or the matrix determinant lemma for log
determinant of the Gramian.
Alternatively, there is also an accelerated version of the
greedy algorithm, originally due to Minoux [28], in which
one can exploit the submodularity of the set functions to
significantly reduce the number of times that marginal gains
for the actuators need to be computed. This can lead to orders
of magnitude speedups in practice; see, e.g., [21].
IV. N UMERICAL E XAMPLES
In this section, we illustrate the results on randomly generated systems and on a problem of power electronic actuator
placement in a model of the European power grid. The problem
data is a system dynamics matrix A ∈ Rn×n , a set of possible
8
4
input matrix columns V = {b1 , ..., bM }, and an integer number
k of actuators to choose from this set to form an input matrix
that maximizes a controllability metric.
f (Sgreedy )
≈ 99%
f (Sopt )
frequency
3
2
histogram
greedy result
1
0
69.4
69.6
69.8
70.0
70.2
70.4
70.6
log det metric
Fig. 1.
Histogram displaying the shifted log determinant metric for all
possible selections of 7 actuators from a set of 25. The result achieved by
greedy optimization is displayed by the red line, which is better than 99.93%
of all other selections.
1
10
0
10
−1
10
λi(Wc)
A. Greedy performance on random systems
To evaluate performance of the greedy algorithm and to
compare the various controllability metrics, we first consider
randomly generated data. We use Matlab’s rss routine to generate a stable dynamics matrix with random stable eigenvalues.
We use V = {e1 , ..., en }, where ei is the ith unit vector in Rn ,
i.e., we assume one can choose states in which a control input
can be injected.
Figure 1 shows the result of applying the greedy algorithm to
maximize the log determinant metric with n = 25 and k = 7.
This problem is small enough to evaluate every possible 7element actuator subset, and this result is also shown in a
histogram, shifted so that minS f (S) = 0. The support of
f (S) for |S| = k, is relatively narrow and close to the optimal
value. Hence, the greedy bound is not informative in this case,
as 63% of the optimum is lower than the values achieved by
all of the size k subsets. Nevertheless, our algorithm finds a
good set Sgreedy scoring
x 10
tr(Wc)
−tr(W−1
)
c
−2
10
log det(Wc)
λmin(Wc)
−3
of the optimum value f (Sopt ), where Sopt is an optimal subset,
and is better than 99.93% of all other n choose k possible
choices. We repeated the greedy algorithm for 500 randomly
generated stable dynamic matrices and found that in all cases
the greedy algorithm returned a selection better than 99.5%
of all possible choices. In other words, for this example the
greedy algorithm provides a near-optimal selection and also
one that performs much better than the worst case bound.
We also compare for this example the four continuous
metrics tr(WS ), tr(WS−1 ), log det WS , and λmin (WS ). Figure
2 shows the eigenvalues of the Gramians resulting from
applying the greedy algorithm with each metric. The results
are averages over 10,000 random samples of stable dynamics
matrices. For the trace metric, Theorem 4 guarantees that the
greedy algorithm finds the globally optimal subset. We can
see that this metric tends to focus on making the largest few
eigenvalues large at the expense of the smaller eigenvalues.
In contrast, the trace of the inverse Gramian and the log
determinant strike a compromise, with both resulting in similar
eigenvalue distributions. The largest eigenvalues are not as
large as the ones achieved by optimizing the trace metric,
but they do better on average on the smaller eigenvalues.
Although the globally optimal subset is not guaranteed to be
found, the submodularity of these metrics guarantees that a
near optimal subset is produced by the greedy algorithm, as
proved in Theorems 5 and 6. On the other hand, the final
metric focuses exclusively on the smallest eigenvalue, but for
this example actually does slightly worse on average than the
trace of the inverse Gramian on the smallest eigenvalue. This
may result from the fact that the smallest eigenvalue metric
is not submodular, and so thus is not always guaranteed to
produce a near optimal selection. However, even using the
10
−4
10
0
5
10
15
20
25
i
Fig. 2. Eigenvalues of the Gramian averaged over 10,000 random samples
of stable dynamics matrices for several continuous metrics resulting from
applying the greedy algorithm to select 7 actuators from a set of possible
25.
greedy algorithm on this non-submodular metric does not do
too much worse than trace of the inverse Gramian and the log
determinant on other eigenvalues, and does better than trace
of the Gramian on most of the smaller eigenvalues.
B. Power electronic actuator placement in the European
power grid
Recently developed power electronic devices, such as high
voltage direct current (HVDC) links or flexible alternating
current transmission devices (FACTS), can be used to improve
transient stability properties in power grids by modulating
active and reactive power injections to damp frequency oscillations and prevent rotor angle instability [12]. In this section,
we illustrate our results by using them to place such power
electronic actuators in a model of the European power grid.
We emphasize that this section is intended to show that there
are practical and relevant problems that could be studied using
the theory in the preceding sections; however, many important
political and economic issues are neglected, and placements
are evaluated purely from a controllability perspective.
We consider a simplified model of the European grid derived
from [16] with 74 buses, each of which is connected to a
9
1 Ideally, one would of course want to evaluate actuator placement on the
nonlinear model, but even evaluating controllability metrics can be extremely
difficult computationally, even for small-scale nonlinear systems. This section
is intended to illustrate the theory from the previous section, so we focus on a
linearized model, though actuator placement problems for nonlinear networks
are an important topic for future work.
Fig. 3. Network of the 74-bus European grid model. The red dots show the
buses, and the black lines between buses show normal AC transmission lines.
The best 10 HVDC line placements according to the controllability Gramian
trace metric are shown by the bold blue lines.
45
controllability Gramian trace metric
generator and a constant impedance load. We consider the
placement of HVDC links, which are modeled as ideal current
sources that can instantaneously inject AC currents into each
of their terminal buses; for modeling details see [12], [13],
[14]. The system dynamics we consider here are based on the
swing equations, a widely-used nonlinear model for the time
evolution of rotor angles and frequencies of each generator
in the network [23]. Each HVDC link has three degrees of
freedom that allow influence of the frequency dynamics at
the corresponding buses. The nonlinear model is linearized1
about a desired operating condition for each possible HVDC
link placement, and the placements are evaluated based on
the linearized model using the infinite-horizon controllability
Gramian. In principle one can easily work with the finitehorizon Gramian; again we chose to use the infinite-horizon
due to the simplicity of its computation, and the results are
qualitatively similar when the finite-horizon Gramian is used.
Each generator has two associated states: rotor angle and
frequency, which gives a 148-dimensional state space model,
i.e., A ∈ R148×148 , which always turns out to be stable. Since
an HVDC link could be placed in principle between any two
distinct nodes in the network, there are 2701 possible locations.
Consider the problem of finding the best subset of size 10. This
gives approximately 5.6 × 1027 possible combinations, far too
many for a brute force search.
Figure 3 shows the network and the 10 best placements
according to the controllability Gramian trace metric with all
state space directions weighted equally, i.e., C = I148 . The
best three are relatively long lines connecting the northeastsouthwest and northwest-southeast quadrants of the network,
respectively. A modal analysis of the dynamics matrix reveals
that these choices correlate well with directions associated
with lightly damped modes in the rotor angle dynamics. The
next group of placements is concentrated in the southeast,
indicating that there is room to improve control authority by
increasing connectivity in this sparsely connected region. This
also indicates a potential weakness in the trace metric, which
may cluster actuators to get high controllability in a few of the
more controllable directions at the expense of controllability in
other directions. Figure 4 shows the sorted values of the metric,
with the top few placements giving a substantial benefit over
other placements.
Figure 5 shows the placement obtained by using the greedy
algorithm with the log determinant metric, using the rank
metric until the system becomes controllable. Compared to
the trace metric, we see that the lines are in general longer,
connecting buses that are further apart, and more evenly
distributed in the network, and no node is part of more than one
HVDC line. These placements can also be seen to align with
directions corresponding to lightly damped modes in the rotor
angle dynamics, though with a different distribution across
the modes than with the trace. Although both metrics tend
40
35
30
25
20
15
0
500
1000
1500
2000
2500
HVDC placement location index
Fig. 4. Sorted values of the controllability Gramian trace metric. The vertical
axis shows the amount each particular actuator placement adds to the trace of
the Gramian. The optimal value is the sum of the first 10 amounts; the top
few placement give substantial benefit over other placements.
to produce placements with a similar qualitative function, the
two sets of obtained placements are quite different.
V. C ONCLUSIONS AND O UTLOOK
We have considered optimal actuator placement problems in
complex dynamical networks. These problems are in general
difficult combinatorial optimization problems; however, we
have shown that an important class of metrics related to the
controllability Gramians yield modular and submodular set
functions. This allows globally optimal or near optimal placements to be obtained with a simple greedy algorithm. By duality, all of the results hold for corresponding sensor placement
10
model discussed in Section IV.
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
Fig. 5. Network of the 74-bus European grid model.. The best 10 HVDC line
placements achieved by the greedy algorithm maximizing the log determinant
of the controllability Gramian are shown by the bold blue lines.
[7]
[8]
problems using metrics of the observability Gramian. To our
knowledge, this is the first such investigation of submodularity
in the context of controllability and observability in dynamical
systems. We also defined several dynamic Control Energy
Centrality measures, which assigns an importance value to
each node in a dynamical network based on its ability to
move the system around the state space with a low-energy
time-varying control input. The results were illustrated via
placement of power electronic actuators in a model of the
European power grid.
There are many open problems involving the structure of
combinatorial optimization problems in the optimal placement
of sensors and actuators in complex networks. For example,
there are many other quantitative metrics of controllability and
observability, such as those associated with optimal control
and filtering design problems, that may be more appropriate
in certain settings. Nothing is known about modularity or submodularity for any other metrics. Further future work involves
exploring other case studies in power networks, biological
networks, social networks, and discretized models of infinitedimensional systems. For more complicated system models,
such as constrained, nonlinear, hybrid, etc., corresponding
controllability questions are much more complicated, and the
available tools do not scale well computationally, but one
could explore how efficient methods could be used to obtain
approximate metrics in these types of systems. Finally, an
important and interesting topic for future work is to investigate
how various graphical properties of the network structure affect
the actuator placement results, which may lead to insights
about energy-related controllability for complex dynamical
networks.
ACKNOWLEDGEMENTS
The authors would like to thank Dr. Alex Fuchs for providing details and helpful discussion about the power system
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
R.H. Bartels and G.W. Stewart. Solution of the matrix equation ax+
xb= c. Communications of the ACM, 15(9):820–826, 1972.
Y.Y. Boykov and M.-P. Jolly. Interactive graph cuts for optimal boundary & region segmentation of objects in nd images. In Proceedings
of the Eighth IEEE International Conference on Computer Vision,
volume 1, pages 105–112. IEEE, 2001.
F.M. Callier and C.A. Desoer. Linear system theory. 1991.
A. Chapman and M. Mesbahi. System theoretic aspects of influenced
consensus: Single input case. IEEE Transactions on Automatic Control,
57(6):1505–1511, 2012.
A. Chapman and M. Mesbahi. Semi-autonomous consensus: network
measures and adaptive trees. IEEE Transactions on Automatic Control,
58(1):19–31, 2013.
A. Clark, L. Bushnell, and R. Poovendran. On leader selection for
performance and controllability in multi-agent systems. In IEEE
Conference on Decision and Control, pages 86–93. IEEE, 2012.
A. Clark, L. Bushnell, and R. Poovendran. A supermodular optimization
framework for leader selection under link noise in linear multi-agent
systems. IEEE Transactions on Automatic Control, 59(2):283–296,
2014.
A. Clark and R. Poovendran. A submodular optimization framework
for leader selection in linear multi-agent systems. In IEEE Conference
on Decision and Control, pages 3614–3621. IEEE, 2011.
F.L. Cortesi, T.H. Summers, and J. Lygeros. Submodularity of energy
related controllability metrics. In IEEE Conference on Decision and
Control, Los Angeles, CA, USA, 2014.
N.J. Cowan, E.J. Chastain, D.A. Vilhena, J.S. Freudenberg, and C.T.
Bergstrom. Nodal dynamics, not degree distributions, determine
the structural controllability of complex networks. PLOS ONE,
7(6):e38398, 2012.
U. Feige. A threshold of ln n for approximating set cover. Journal of
the ACM, 45(4):634–652, 1998.
A. Fuchs and M. Morari. Grid stabilization through VSC-HVDC using
wide area measurements. In IEEE PowerTech Conference, Trondheim,
Norway. IEEE, 2011.
A. Fuchs and M. Morari. Actuator performance evaluation using LMIs
for optimal HVDC placement. In European Control Conference, Zurich,
Switzerland. IEEE, 2013.
A. Fuchs and M. Morari. Placement of HVDC links for power
grid stabilization during transients. In IEEE PowerTech Conference,
Grenoble, France. IEEE, 2013.
S.J. Hammarling. Numerical solution of the stable, non-negative definite
Lyapunov equation. IMA Journal of Numerical Analysis, 2(3):303–323,
1982.
T. Hasse.
Anforderung an eine durch Erneuerbare Energien
geprägeigen Energieversorgung - Untersuchung des Regelverhaltens
von Kraftwerken und Verbundnetzen. PhD thesis, University of Rostock,
2006.
T. Kailath. Linear systems, volume 1. Prentice-Hall Englewood Cliffs,
NJ, 1980.
R.E. Kalman. On the general theory of control systems. IRE
Transactions on Automatic Control, 4(3):110–110, 1959.
R.E. Kalman. Contributions to the theory of optimal control. Boletin
de la Sociedad Matematica Mexicana, 5(1):102–119, 1960.
D. Kempe, J. Kleinberg, and É. Tardos. Maximizing the spread of
influence through a social network. In Proceedings of the Ninth ACM
SIGKDD International Conference on Knowledge Discovery and Data
Mining, pages 137–146. ACM, 2003.
A. Krause and D. Golovin. Submodular function maximization.
Tractability: Practical Approaches to Hard Problems, 3, 2012.
11
[22]
A. Krause, A. Singh, and C. Guestrin. Near-optimal sensor placements
in gaussian processes: Theory, efficient algorithms and empirical studies. The Journal of Machine Learning Research, 9:235–284, 2008.
[23]
P. Kundur. Power system stability and control. McGraw-Hill, Inc.,
1993.
[24]
J.-R. Li and J. White. Low rank solution of lyapunov equations. SIAM
Journal on Matrix Analysis and Applications, 24(1):260–280, 2002.
[25]
C.-T. Lin. Structural controllability. IEEE Transactions on Automatic
Control, 19(3):201–208, 1974.
[26]
Y.-Y. Liu, J.-J. Slotine, and A.-L. Barabási. Controllability of complex
networks. Nature, 473(7346):167–173, 2011.
[27]
L. Lovász. Submodular functions and convexity. Mathematical
programming: the state of the art, pages 235–257, 1983.
[28]
M. Minoux. Accelerated greedy algorithms for maximizing submodular
set functions. In Optimization Techniques, pages 234–243. Springer,
1978.
[29]
F.J. Müller and A. Schuppert. Few inputs can reprogram biological
networks. Nature, 478(7369):E4–E4, 2011.
[30]
G.L. Nemhauser, L.A. Wolsey, and M.L. Fisher. An analysis of approximations for maximizing submodular set functions—I. Mathematical
Programming, 14(1):265–294, 1978.
[31]
T. Nepusz and T. Vicsek. Controlling edge dynamics in complex
networks. Nature Physics, 8(7):568–573, 2012.
[32]
M. Newman. Networks: an introduction. Oxford University Press, 2010.
[33]
A. Olshevsky. Minimal controllability problems. IEEE Transactions on
Control of Network Systems, 1(3):249–258, 2014.
[34]
F. Pasqualetti, S. Zampieri, and F. Bullo. Controllability metrics,
limitations and algorithms for complex networks. IEEE Transactions
on Control of Network Systems, 1(1):40–52, 2014.
[35]
K.B. Petersen and M.S. Pedersen. The matrix cookbook. Technical
University of Denmark, 2008.
[36]
A. Rahmani, M. Ji, M. Mesbahi, and M. Egerstedt. Controllability of
multi-agent systems from a graph-theoretic perspective. SIAM Journal
on Control and Optimization, 48(1):162–186, 2009.
[37]
I. Rajapakse, M. Groudine, and M. Mesbahi. Dynamics and control
of state-dependent networks for probing genomic organization. Proceedings of the National Academy of Sciences, 108(42):17257–17262,
2011.
[38]
I. Shames and T.H. Summers. Rigid network design via submodular
set function optimization. submitted to IEEE Transactions on Network
Science and Engineering, 2014.
[39]
F. Sorrentino, M. di Bernardo, F. Garofalo, and G. Chen. Controllability of complex networks via pinning. Physical Review-Series E,
75(4):046103, 2007.
[40]
T.H. Summers and J. Lygeros. Optimal sensor and actuator placement
in complex dynamical networks. In IFAC World Congress, Cape Town,
South Africa, pages 3784–3789, 2014.
[41]
T.H. Summers, I. Shames, J. Lygeros, and F. Dörfler. Topology design
for optimal network coherence. In European Control Conference, 2015.
[42]
J. Sun and A.E. Motter. Controllability transition and nonlocality in
network control. Physical Review Letters, 110(20):208701, 2013.
[43]
Y. Tang, H. Gao, W. Zou, and J. Kurths. Identifying controlling nodes in
neuronal networks in different scales. PLOS ONE, 7(7):e41375, 2012.
[44]
M. Van De Wal and B. De Jager. A review of methods for input/output
selection. Automatica, 37(4):487–510, 2001.
[45]
W.-X. Wang, X. Ni, Y.-C. Lai, and C. Grebogi. Optimizing controllability of complex networks by minimum structural perturbations. Physical
Review E, 85(2):026115, 2012.
[46]
G. Yan, J. Ren, Y.-C. Lai, C.-H. Lai, and B. Li. Controlling complex
networks: How much energy is needed? Physical Review Letters,
108(21):218703, 2012.
[47]
T. Yang, G. Huijun, J. Kurths, and F. Jian-an. Evolutionary pinning
control and its application in UAV coordination. Industrial Informatics,
IEEE Transactions on, 8(4):828–838, 2012.
[48] K. Zhou, G. Salomon, and E. Wu. Balanced realization and model
reduction for unstable systems. International Journal of Robust and
Nonlinear Control, 9(3):183–198, 1999.
| 3cs.SY
|
arXiv:1606.05807v1 [math.GR] 18 Jun 2016
IRREDUCIBLE CHARACTERS OF EVEN DEGREE
AND NORMAL SYLOW 2-SUBGROUPS
NGUYEN NGOC HUNG AND PHAM HUU TIEP
Abstract. The classical Itô-Michler theorem on character degrees of finite groups
asserts that if the degree of every complex irreducible character of a finite group
G is coprime to a given prime p, then G has a normal Sylow p-subgroup. We
propose a new direction to generalize this theorem by introducing an invariant
concerning character degrees. We show that if the average degree of linear and
even-degree irreducible characters of G is less than 4/3 then G has a normal Sylow 2subgroup, as well as corresponding analogues for real-valued characters and strongly
real characters. These results improve on several earlier results concerning the ItôMichler theorem.
1. Introduction
The celebrated Itô-Michler theorem [15, 20] is one of the deep and fundamental
results on the relation between character degrees and local structure of finite groups.
It asserts that if a prime p does not divide the degree of every complex irreducible
character of a finite group G, then G has a normal abelian Sylow p-subgroup.
Over the past decades, there have been several variations and refinements of this
result by considering Brauer characters [17, 18, 21], nonvanishing elements [5, 25],
fields of character values [4, 25, 13], and Frobenius-Schur indicator [19, 28]. One
primary direction is to weaken the condition that all irreducible characters of G have
degree coprime to p, and assume instead only that a subset of characters with a
specified field of values has this property, see [4, 19] for real-valued characters and
[25] for p-rational characters.
In this paper, we introduce an invariant concerning character degrees and propose
to generalize the Itô-Michler theorem in a completely new direction.
Date: June 21, 2016.
2010 Mathematics Subject Classification. Primary 20C15, 20D10, 20D05.
Key words and phrases. finite groups, character degrees, normal subgroups, Sylow subgroups,
real characters, strongly real characters.
N. N. Hung was partially supported by the NSA Young Investigator Grant #H98230-14-1-0293
and a Faculty Scholarship Award from Buchtel College of Arts and Sciences, The University of
Akron.
P. H. Tiep was partially supported by the NSF grant DMS-1201374 and the Simons Foundation
Fellowship 305247.
1
2
NGUYEN NGOC HUNG AND PHAM HUU TIEP
Given a finite group G, let Irr(G) denote the set of all complex irreducible characters of G, then write
Irrp (G) := {χ ∈ Irr(G) | χ(1) = 1 or p | χ(1)}
and
acdp (G) :=
P
χ∈Irrp (G)
χ(1)
|Irrp (G)|
so that acdp (G) is the average degree of linear characters and irreducible characters
of G with degree divisible by p. Then the Itô-Michler theorem can be reformulated
in the following way:
If acdp (G) = 1 then G a has normal abelian Sylow p-subgroup.
Our first result significantly improves this for the prime p = 2.
Theorem 1.1. Let G be a finite group. If acd2 (G) < 4/3 then G has a normal Sylow
2-subgroup.
Theorem 1.1 basically says that even when a group has some irreducible characters
of even degree, it still has a normal Sylow 2-subgroup as long as the number of
linear characters of the group is large enough. This of course implies the Itô-Michler
theorem for p = 2 where it is required that the group has no irreducible characters
of even degree at all.
One of the key steps in the proof of Theorem 1.1 is to establish the solvability of
the groups in consideration. In fact, we can do more.
Theorem 1.2. Let G be a finite group. If acd2 (G) < 5/2 then G is solvable.
We remark that both bounds in Theorems 1.1 and 1.2 are optimal, as shown by the
groups S3 and A5 . Furthermore, acd2 (G) of non-abelian 2-groups G can get as close
to 1 as we wish – just consider the extraspecial 2-groups, and therefore one can not
get the commutativity of the Sylow 2-subgroup in Theorem 1.1 as in the Itô-Michler
theorem.
In fact, we can also improve on some main results of [4] and [19], by restricting our
attention to only real-valued characters or even strongly real characters. A character
χ ∈ Irr(G) is called real-valued if χ(g) ∈ R for all g ∈ G, and strongly real if it has
Frobenius-Schur indicator 1, equivalently, if it is afforded by a real representation.
Let
P
χ∈Irrp,R (G) χ(1)
Irrp,R (G) := {χ ∈ Irrp (G) | χ is real-valued}, acdp,R (G) :=
,
|Irrp,R (G)|
and
Irrp,+ (G) := {χ ∈ Irrp (G) | χ is strongly real}, acdp,+ (G) :=
P
χ∈Irrp,+ (G)
χ(1)
|Irrp,+ (G)|
CHARACTERS OF EVEN DEGREE AND NORMAL SYLOW 2-SUBGROUPS
3
Theorem 1.3. Let G be a finite group. We have:
(i) If acd2,+ (G) ≤ 2 then G is solvable.
(ii) If acd2,+ (G) < 4/3 then G has a normal Sylow 2-subgroup.
Theorem 1.3(ii) immediately implies [4, Theorem A] and [19, Theorem B]. Furthermore, since any real-valued character of degree 1 is automatically strongly real,
it has the following consequence.
Corollary 1.4. Let G be a finite group. We have:
(i) If acd2,R (G) ≤ 2 then G is solvable.
(ii) If acd2,R (G) < 4/3 then G has a normal Sylow 2-subgroup.
Again, the example of S3 shows that the bounds in Theorem 1.3(ii) and Corollary 1.4(ii) are optimal.
To prove Theorems 1.2 and 1.3(i), we have to use the classification of finite simple
groups to show that every nonabelian finite simple group S possesses an irreducible
character of even and large enough degree which is extendible to its stabilizer in
Aut(S), cf. Theorem 2.1. Together with Proposition 2.3, this result allows us to
bound the number of (strongly real) irreducible characters of small degrees of a finite
group with a nonabelian minimal normal subgroup, and then to control the invariant acd2 (G) of such a group, see Section 2 and Proposition 3.1. We hope that the
techniques developed here will be useful in the future study of other problems involving the average degree of a certain set of characters and other invariants concerning
character degrees like the largest character degree [3, 7, 8, 10] or the character degree
ratio [2, 16].
One obvious question that one may ask is: is there an analogue of Theorem 1.1
for odd primes? Although our ideas in the proof for the prime 2 do not carry out
smoothly to odd primes, we believe that the following is true.
Conjecture 1.5. Let p be a prime and G a finite group. If acdp (G) < 2p/(p + 1)
then G has a normal Sylow p-subgroup.
The bound in Conjecture 1.5 perhaps is not optimal for all primes. If Cp (the cyclic
group of order p) can act nontrivially on an abelian group of order p + 1, then the
bound clearly can not be lower (for instance p = 3, 7, or any Mersenne prime). But
when p = 5 for example, C5 can only act trivially on an abelian group of order 6,
and we think that the best possible bound is not 10/6.
Theorems 1.2, 1.1, and 1.3 are respectively proved in Sections 3, 4, and 5.
2. Extendibility of some characters of even degree
Throughout the paper, for a finite group G and a positive integer k, we write
nk (G) to denote the number of irreducible complex characters of G of degree k, and
nk,+ (G) to denote the number of strongly real, irreducible complex characters of G
4
NGUYEN NGOC HUNG AND PHAM HUU TIEP
of degree k. Furthermore, if N is a normal subgroup of G, then nk (G|N) denotes the
number of irreducible characters of G of degree k whose kernels do not contain N,
and similarly for nk,+ (G|N). If θ is a character of a normal subgroup of G, we write
IG (θ) to denote the stabilizer or the inertia subgroup of θ in G. Other notation is
standard (and follows [11]) or will be defined when needed.
We need the following result, whose proof relies on the classification of finite simple
groups.
Theorem 2.1. Every nonabelian finite simple group S has an irreducible character
θ of even degree such that θ(1) ≥ 4 and θ is extendible to a strongly real character of
IAut(S) (θ). Furthermore, if S ≇ A5 then θ can be chosen so that θ(1) ≥ 8.
Proof. (i) The cases where S ∼
= PSL2 (q) with q ≤ 19,
= An with 5 ≤ n ≤ 8, or S ∼
∼
or S = PSU3 (3), PSp4 (3), Sp6 (2), 2 F4 (2)′ , or S is one of the 26 sporadic finite
simple groups, can be checked directly using [1]. (We note that in all these cases but
S = A6 , we can always find θ so that it has a rational-valued extension to IAut(S) (θ).)
In what follows, we may therefore assume that S is not isomorphic to any of the listed
groups. In particular, it follows from the main results of [26, 27] that the degree of
any nontrivial complex irreducible character of S is at least 8.
(ii) Certainly, one can find many different choices for the desired character θ. In
what follows, having in mind some other applications, we will try to construct θ in
such a way that its extension to IAut(S) (θ) is rational-valued if possible.
Assume now that S ∼
= An with n ≥ 9. Consider the irreducible characters α, β ∈
Irr(Sn ) labeled by the partitions (n − 2, 2) and (n − 2, 12), of degree n(n − 3)/2 and
(n − 1)(n − 2)/2 respectively. Since the given partitions are not self-conjugate, α and
β both restrict irreducibly to S. Furthermore, β(1) = α(1) + 1 ≥ 28, and so exactly
one of α, β has even degree. As Aut(S) ∼
= Sn , we are done in this case by choosing
θ ∈ {αS , βS } of even degree.
Next we consider the case S is a finite simple group of Lie type in characteristic 2.
As shown in [6], the Steinberg character of S, of degree |S|2, extends to a character
of a rational representation of Aut(S), whence we are done again.
(iii) From now on we may assume that S is a finite simple group of Lie type, defined
over a finite field Fq of odd characteristic p.
Consider the cases where S = PSUn (q) with n ≥ 4, PSp2n (q) with n ≥ 2, Ω2n+1 (q)
−
with n ≥ 3, PΩ+
2n (q) with n ≥ 5, or PΩ2n (q) with n ≥ 4. As shown in pp. 1–4
of the proof of [4, Theorem 2.1], Aut(S) has a rank 3 permutation character ρ =
1 + α + β, where the characters α and β both restrict nontrivially and irreducibly
to S. Furthermore, exactly one of α and β is of even degree (and both are afforded
by rational representations). Hence we are done by choosing θ ∈ {αS , βS } of even
degree.
The same argument, but applied to a rank 5 permutation character of Aut(S), see
p. 5 of the proof of [4, Theorem 2.1], handles the cases S = E6 (q) or 2 E6 (q).
CHARACTERS OF EVEN DEGREE AND NORMAL SYLOW 2-SUBGROUPS
5
Suppose that S = PSLn (q) with n ≥ 4. As shown in the proof of [24, Proposition 5.5], Aut(S) has a permutation representation, whose character contains a
rational-valued irreducible character γ of (even) degree
n
(q − 1)(q n−1 − 1)/(q − 1)2 , q ≡ 1(mod 4),
(q n − 1)(q n−1 − 1)/(q 2 − 1), q ≡ 3(mod 4),
with multiplicity one that restricts irreducibly to S. It follows that γ is also afforded
by a rational representation, and we can choose θ = γS .
(iv) In the remaining cases, our choice of θ yields a not necessarily rational, but
still admits a strongly real extension to J := IAut(S) (θ). First, if S = PSL3 (q),
then the (unique) character θ ∈ Irr(S) of degree q 2 + q extends to Aut(S) by [28,
Lemma 6.2]. Next, if S = G2 (q), 2 G2 (q), 3 D4 (q), F4 (q), or E8 (q), then the proof of
[19, Proposition 4.4] yields a strongly real character θ ∈ Irr(S) of even degree, with
J = S.
If S = PΩ+
8 (3) then by p. 1 of the proof of [19, Proposition 4.9] we can choose θ of
degree 300.
Suppose now that S = PSL2 (q), PSU3 (q), or (a simple group of type) D4 (q) with
q ≥ 5, or E7 (q). Then we can view S as the derived subgroup of a finite Lie-type
group G of adjoint type: G = PGL2 (q), PGU3 (q), D4 (q)ad , or E7 (q)ad , respectively.
As shown in the proof of [19, Proposition 4.5] (for types A1 and E7 ), Case IIb of
the proof of [19, Proposition 4.7] (for PSU3 ), and the proof of [19, Proposition 4.9]
(for type D4 ), G contains a strongly real character of even degree that restricts to
an irreducible character θ ∈ Irr(S) with J = G. So we are done in these cases as
well.
When a finite group G has a nonabelian minimal normal subgroup N, by using
Theorem 2.1, we can produce an irreducible character of N of even degree that is
extendible to (a strongly real character of) its stabilizer in G.
Theorem 2.2. Let G be a finite group with a nonabelian minimal normal subgroup
N ≇ A5 . Then there exists ϕ ∈ Irr(N) of even degree such that ϕ(1) ≥ 8 and ϕ is
extendible to a strongly real character of IG (ϕ).
Proof. Since N is a nonabelian minimal normal subgroup of G, we have N ∼
=S×
S × · · · × S, a direct product of k copies of a nonabelian simple group S. Replacing
G by G/CG (N) if necessary, we may assume that CG (N) = 1. Then we have
N ✂ G ≤ Aut(N) = Aut(S) ≀ Sk .
Let θ be an irreducible character of S found in Theorem 2.1 and let OS be the orbit of
θ in the action of Aut(S) on Irr(S). Consider the character ϕ := θ × · · · × θ ∈ Irr(N).
Then the orbit of ϕ under the action of Aut(N) is
ON := {θ1 × θ2 × · · · × θk ∈ Irr(N) | θi ∈ OS }.
6
NGUYEN NGOC HUNG AND PHAM HUU TIEP
Clearly ϕ is invariant under IAut(S) (θ) ≀ Sk . On the other hand,
|Aut(N) : IAut(S) (θ) ≀ Sk | = |Aut(S) : IAut(S) (θ)|k = |OS |k = |ON |.
Therefore we deduce that IAut(N ) (ϕ) = IAut(S) (θ) ≀ Sk .
Assume that θ extends to a strongly real character α of J := IAut(S) (θ), say afforded
by an RJ-module V . Then IAut(N ) (ϕ) acts naturally on the R-space V ⊗k , on which
N acts with character ϕ. It follows that ϕ is extendible to the strongly real character
of IG (ϕ) = G ∩ IAut(N ) (ϕ) afforded by V ⊗k .
Since θ(1) ≥ 4 in general and θ(1) ≥ 8 when S ≇ A5 , we observe that ϕ(1) =
θ(1)k ≥ 8 as long as (S, k) 6= (A5 , 1), and we are done.
Theorem 2.2 can be used to bound the number of irreducible characters of degrees
1 and 2 of finite groups with a nonabelian minimal normal subgroup, as shown in the
next proposition. This should be compared with Proposition 3.2 of [9].
Proposition 2.3. Let G be a finite group with a nonabelian minimal normal subgroup
N. Assume that there is some ϕ ∈ Irr(N) such that ϕ is extendible to IG (ϕ). Let
d := ϕ(1)|G : IG (ϕ)|. Then the following hold.
(i) n1 (G) ≤ nd (G)|G : IG (ϕ)|.
(ii) If ϕ extends to a strongly real character of IG (ϕ), then
n1,+ (G) ≤ nd,+ (G)|G : IG (ϕ)|.
(iii) n2 (G) ≤ n2d (G)|G : IG (ϕ)| + 21 nd (G)|G : IG (ϕ)|.
Moreover, if ϕ is invariant in G then n2 (G) ≤ n2d (G).
Proof. For simplicity we write I := IG (ϕ).
(i) First, since n1 (G) = |G : G′ | and n1 (I) = |I : I ′ |, we have n1 (G) ≤ |G : I|n1 (I).
Therefore, we wish to show that n1 (I) ≤ nd (G) where d := ϕ(1)|G : I|.
As N = N ′ ⊆ I ′ , the normal subgroup N is contained in the kernel of every linear
character of I so that
n1 (I) = n1 (I/N).
Recall that ϕ ∈ Irr(N) is extendible to I and so we let χ ∈ Irr(I) be an extension
of ϕ. Using Gallagher’s theorem and Clifford’s theorem (see [11, Corollary 6.17 and
Theorem 6.11]), we see that each linear character λ of I/N produces the irreducible
character λχ of T of degree ϕ(1), and this character in turns produces the irreducible
character (λχ)G of G of degree (λχ)G (1) = ϕ(1)|G : I| = d. As the maps λ 7→ λχ 7→
(λχ)G are both injective, it follows that
n1 (I/N) ≤ nd (G)
and we therefore have n1 (I) ≤ nd (G), as desired.
(ii) For any group X, let X ∗ denote the subgroup generated by all x2 , x ∈ X.
Then note that n1,+ (G) = |G : G∗ |, n1,+ (I) = |I : I ∗ |, and I ∗ ≤ G∗ , whence
CHARACTERS OF EVEN DEGREE AND NORMAL SYLOW 2-SUBGROUPS
7
n1,+ (G) ≤ |G : I|n1,+ (I). Furthermore, for any χ ∈ Irr(G) and any strongly real
linear character λ of G, λ2 = 1G and so χ and χλ have the same Frobenius-Schur
indicator. In particular, χλ is strongly real if and only if χ is. Furthermore, if ρ is
a strongly real character of a subgroup T ≤ G, then so is the induced character ρG .
Now we can argue as in (i) to complete the proof.
(iii) We first claim that
1
n2 (G) ≤ n2 (I)|G : I| + n1 (I)|G : I|.
2
Let χ ∈ Irr(G) with χ(1) = 2. Take φ to be an irreducible constituent of χ ↓I .
Frobenius reciprocity then implies that χ in turn is an irreducible constituent of φG .
If φ(1) = 2 then as φG (1) = 2|G : I|, there are at most |G : I| irreducible constituents
of degree 2 of φG . We deduce that there are at most n2 (I)|G : I| irreducible characters
of degree 2 of G that arise as constituents of φG with φ(1) = 2. On the other hand, if
φ(1) = 1 then, as φG (1) = |G : I|, there are at most |G : I|/2 irreducible constituents
of degree 2 of φG . As above, we deduce that there are at most n1 (I)|G : I|/2
irreducible characters of degree 2 of G that arise as constituents of φG with φ(1) = 1.
So the claim is proved.
Since we have already proved in (i) that n1 (I) = n1 (I/N) ≤ nd (G), to prove
Proposition 2.3(iii) it suffices to show that n2 (I) ≤ n2d (G).
We claim that
n2 (I) = n2 (I/N)
or in other words, N is contained in the kernel of every irreducible character of degree
2 of I. Let φ ∈ Irr(I) with φ(1) = 2. Since N has no irreducible character of degree 2
and has only one linear character, which is the trivial one, it follows that φN = 2 · 1N .
We then have N ⊆ Ker(φ), as claimed.
Recall that χ ∈ Irr(I) is an extension of ψ. Using Gallagher’s theorem and Clifford’s theorem again, we obtain that each irreducible character µ ∈ Irr(I/N) of degree
2 produces the character (µχ)G ∈ Irr(G) of degree (µχ)G (1) = 2ψ(1)|G : I| = 2d. It
follows that
n2 (I/N) ≤ n2d (G),
and thus n2 (I) ≤ n2d (G), as desired.
If ϕ is invariant in G then G = I, yielding immediately that n2 (G) ≤ n2d (G).
3. Solvability - Theorem 1.2
In this section, we use the results in Section 2 to prove Theorem 1.2. The next
proposition handles an important case of this theorem.
Proposition 3.1. Let G be a finite group with a nonabelian minimal normal subgroup. Then acd2 (G) ≥ 5/2.
8
NGUYEN NGOC HUNG AND PHAM HUU TIEP
Proof. Let N be a nonabelian minimal normal subgroup of G. First we assume
that N ≇ A5 . Then Theorem 2.2 guarantees that there is ϕ ∈ Irr(N) of even
degree such that ϕ(1) ≥ 8 and ϕ is extendible to the inertia subgroup IG (ϕ). Using
Proposition 2.3, we then have
n1 (G) ≤ nd (G)|G : IG (ϕ)|
and
1
n2 (G) ≤ n2d (G)|G : IG (ϕ)| + nd (G)|G : IG (ϕ)|,
2
where d := ϕ(1)|G : IG (ϕ)|. It follows that
7
3n1 (G) + n2 (G) ≤ nd (G)|G : IG (ϕ)| + n2d (G)|G : IG (ϕ)|.
2
Since ϕ(1) ≥ 8, we have d ≥ 8|G : IG (ϕ)| ≥ 8 and hence we can check that (7/2)|G :
IG (ϕ)| ≤ (7/16)d < 2d − 5 and |G : IG (ϕ)| ≤ d/8 < 4d − 5. It follows that
X
3n1 (G) + n2 (G) <
(2k − 5)nk (G),
2|k≥4
and thus
X
(2k − 5)nk (G) > 0.
2|k or k=1
Therefore
2
X
2|k or k=1
knk (G) > 5
X
nk (G),
2|k or k=1
and we have acd2 (G) > 5/2, as desired.
So it remains to consider N ∼
= A5 . Then N has an irreducible character ϕ of degree
4 that is extendible to Aut(N) (see [1, p. 2]), and hence extendible to G as well. It
follows from Proposition 2.3 that n1 (G) ≤ n4 (G) and n2 (G) ≤ n8 (G). Thus
X
3n1 (G) + n2 (G) ≤ 3n4 (G) + n8 (G) ≤
(2k − 5)nk (G).
2|k≥4
Again this yields acd2 (G) ≥ 5/2 and the proof is completed.
We are now ready to prove Theorem 1.2, which we restate below.
Theorem 3.2. Let G be a finite group. If acd2 (G) < 5/2 then G is solvable.
Proof. Assume, to the contrary, that the theorem is false, and let G be a minimal
counterexample. In particular acd2 (G) < 5/2 and G is nonsolvable. Let H E G be
minimal such that H is nonsolvable. Then clearly H is perfect and contained in the
last term of the derived series of G. Choose a minimal normal subgroup N of G such
that N ⊆ H, and when [H, O∞ (H)] 6= 1 we choose N ⊆ [H, O∞ (H)], where O∞ (H)
denotes the largest normal solvable subgroup of H.
CHARACTERS OF EVEN DEGREE AND NORMAL SYLOW 2-SUBGROUPS
9
In view of Proposition 3.1, we can assume that N is abelian. It follows that the
quotient G/N is nonsolvable since G is nonsolvable. By the minimality of G, we must
have acd2 (G/N) ≥ 5/2. So
acd2 (G) < 5/2 ≤ acd2 (G/N).
Since nk (G/N) ≤ nk (G) for every positive integer k and n1 (G/N) = n1 (G) as N ⊆ G′ ,
it follows that n2 (G/N) < n2 (G) and thus there exists χ ∈ Irr(G) such that χ(1) = 2
and N * Ker(χ).
It then has been shown in the proof of [14, Theorem 2.2] that G = LC is a central
product with the amalgamated subgroup N = Z(L) of order 2, where
L = SL2 (5) and C/ Ker(χ) := Z(G/ Ker(χ)).
Since G = LC is a central product with the central amalgamated subgroup N,
there is a bijection (α, β) 7→ τ from Irr(L|N) × Irr(C|N) to Irr(G|N) such that
τ (1) = α(1)β(1). If (α, β) 7→ χ under the above bijection, we must have β(1) = 1
since L ∼
= SL(2, 5) and there are only three possibilities for α(1), namely 2, 4, and 6.
So β ∈ Irr(C|N) is an extension of the unique non-principal linear character of N.
Using Gallagher’s theorem, we then have a degree-preserving bijection from Irr(C/N)
to Irr(C|N). In particular, n1 (C|N) = n1 (C/N). Since G/L ∼
= C/N and L ⊆ G′ , we
have
n1 (C|N) = n1 (C/N) = n1 (G).
Employing the arguments in the proof of [22, Theorem B], we can evaluate and
estimate n2 (G), n4 (G), and n6 (G) in terms of n1 (G) as follows:
(i) n2 (G) = 2n1 (G) + n2 (C/N),
(ii) n4 (G) ≥ 2n1 (G), and
(iii) n6 (G) ≥ n1 (G) + 2n2 (C/N).
Now putting all things together, we have
X
(2k − 5)nk (G) ≥ 3n4 (G) + 7n6 (G)
2|k≥4
≥ 6n1 (G) + 7(n1 (G) + 2n2 (C/N)
= 13n1 (G) + 14n2 (C/N)
> 3n1 (G) + n2 (G).
It then follows that acd2 (G) > 5/2 and this is a contradiction.
4. Normal Sylow 2-subgroups - Theorem 1.1
The next lemma is crucial in the proof of Theorem 1.1.
Lemma 4.1. Let G = N ⋊ M where N is an abelian group. Assume that no nonprincipal irreducible character of N is invariant under M. If acd2 (G) < 4/3 then
10
NGUYEN NGOC HUNG AND PHAM HUU TIEP
there is no orbit of even size in the action of M on the set of irreducible characters
of N.
Proof. Let {1N = α0 , α1 , . . . , αf } be a set of representatives of the action of M on
Irr(N). For each 1 ≤ i ≤ f , let Ii be the inertia subgroup of αi in G. Since no
nonprincipal irreducible character of N is invariant under M, we observe that every
Ii is a proper subgroup of G.
Assume, to the contrary, that there is some orbit of even size in the action of M
on Irr(N). Then there exists some index 1 ≤ i ≤ f such that |G : Ii | is even. For
0 ≤ i ≤ f , set
X
X
ni,1 := n1 (Ii /N), ni,even :=
nk (Ii /N), si,even :=
λ(1).
2|k
λ∈Irr(Ii /N ), 2|λ(1)
Since G splits over N, it is clear that every Ii also splits over N. It follows that
αi extends to a linear character, say βi , of Ii as αi is linear. Gallagher’s theorem
then implies that the mapping λ 7→ λβi is a bijection from Irr(Ii /N) to the set of
irreducible characters of Ii lying above αi . Using Clifford’s theorem, we then obtain
a bijection λ 7→ (λβi )G from Irr(Ii /N) to the set of irreducible characters of G lying
above αi . We note that (λβi )G (1) = |G : Ii |λ(1) and hence (λβi )G (1) is even if and
only if either |G : Ii | is even or λ(1) is even.
We have
f
X
X
X
|G : Ii |si,even
χ(1) = n1 (G/N) +
|G : Ii |ni,1 +
χ∈Irr(G),χ(1)=1 or even
i=0
|G:Ii | even
X
≥ n1 (G/N) +
|G : Ii |ni,1 + 2
f
X
|G : Ii |ni,even .
i=0
|G:Ii | even
On the other hand,
X
χ∈Irr(G),χ(1)=1 or even
χ(1) = acd2 (G) n1 (G/N) +
X
|G:Ii | even
ni,1 +
f
X
i=0
Therefore, we deduce that
X
(|G : Ii | − acd2 (G))ni,1 +
|G:Ii | even
f
X
ni,even .
(2|G : Ii | − acd2 (G))ni,even
i=0
≤ (acd2 (G) − 1)n1 (G/N).
Since acd2 (G) < 4/3 and |G : Ii | ≥ 1 for every 0 ≤ i ≤ f , it follows that
X
(|G : Ii | − acd2 (G))ni,1 ≤ (acd2 (G) − 1)n1 (G/N)
|G:Ii | even
CHARACTERS OF EVEN DEGREE AND NORMAL SYLOW 2-SUBGROUPS
11
and hence
(|G : Ij | − acd2 (G))nj,1 ≤ (acd2 (G) − 1)n1 (G/N)
for some 1 ≤ j ≤ f such that |G : Ij | is even.
Observe that n1 (G/N) = |(G/N) : (G/N)′ | and nj,1 = n1 (Ij /N) = |(Ij /N) :
(Ij /N)′ |. Therefore n1 (G/N) ≤ |G : Ij |nj,1. It follows from the above inequality that
(|G : Ij | − acd2 (G))nj,1 ≤ (acd2 (G) − 1)|G : Ij |nj,1,
and thus
acd2 (G)
.
2 − acd2 (G)
This is impossible since |G : Ij | ≥ 2 and acd2 (G) < 4/3, and the proof is complete.
|G : Ij | ≤
We now prove the main Theorem 1.1, which is restated below.
Theorem 4.2. Let G be a finite group. If acd2 (G) < 4/3 then G has a normal Sylow
2-subgroup.
Proof. We will argue by induction on the order of G. We have acd2 (G) < 4/3 and
therefore by Theorem 1.2, G is solvable. If G is abelian then there is nothing to prove,
so we assume that G is nonabelian. We then choose a minimal normal subgroup N
of G such that N ⊆ G′ . As G is solvable, we have that N is abelian.
Since N ⊆ G′ , if an irreducible character of G has kernel not containing N, its
degree must be at least 2. We therefore deduce that acd2 (G/N) ≤ acd2 (G) < 4/3 and
it follows from the induction hypothesis that G/N has a normal Sylow 2-subgroup,
say Q/N.
If N is a 2-group then Q is a normal Sylow 2-subgroup of G, and we are done. So we
assume from now on that N is an elementary abelian group of odd order. The SchurZassenhaus theorem then implies that Q = N ⋊ P where P is a Sylow 2-subgroup
of Q (and G as well). By Frattini’s argument, we have G = QNG (P ) = NNG (P ).
If N ⊆ NG (P ) then it follows that G = NG (P ) and we are done. So we can assume
that N * NG (P ). This implies that N ∩ NG (P ) < N. As N is abelian, N ∩ NG (P )
is a normal subgroup of G, and we deduce that N ∩ NG (P ) = 1 by the minimality
of N. We conclude that G = N ⋊ NG (P ).
If N ⊆ Z(G), then we would have Q = N × P , so that P ✂ G, and we are done
again. So we assume that N is noncentral in G. Thus, by the minimality of N,
we have [N, G] = N. It follows that no nonprincipal irreducible character of N is
invariant under NG (P ).
We are now in the situation of Lemma 4.1, and therefore we conclude that there is
no orbit of even size in the action of NG (P ) on the set of irreducible characters of N.
In particular, there is no orbit of even size in the action of P on the set of irreducible
characters of N. This means that P acts trivially on N since P is a 2-group. Now
we have Q = N × P ✂ G and, as N has odd order, we deduce that P ✂ G and this
completes the proof of the theorem.
12
NGUYEN NGOC HUNG AND PHAM HUU TIEP
5. Strongly real characters - Theorem 1.3
In this section we prove Theorem 1.3. We first prove Theorem 1.3(i).
Theorem 5.1. Let G be a finite group. If acd2,+ (G) ≤ 2 then G is solvable.
Proof. Since there is nothing to prove if G is abelian, we assume that G′ is nontrivial. Let N ⊆ G′ be a minimal normal subgroup of G. If a strongly real character
χ ∈ Irr(G) has kernel not containing N, its degree must be at least 2, which is above
the average acd2,+ (G). We therefore deduce that
acd2,+ (G/N) ≤ acd2,+ (G) ≤ 2.
Working by induction, we now can assume that N is nonabelian.
By Theorem 2.2, there exists ϕ ∈ Irr(N) of even degree such that ϕ(1) ≥ 4 and
ϕ extends to a strongly real character of IG (ϕ). (Note that we assumed N ≇ A5 in
Theorem 2.2, but when N ∼
= A5 one can choose ϕ to be the irreducible character of
degree 4, and this character is extendible to a strongly real character of S5 .) We then
apply Proposition 2.3 to have
n1,+ (G) ≤ nd,+ (G)|G : IG (ϕ)|,
where d := ϕ(1)|G : IG (ϕ)|. Since ϕ(1) ≥ 4, it follows that
|G : IG (ϕ)| < 4|G : IG (ϕ)| − 2 ≤ d − 2,
and so
n1,+ (G) < (d − 2)nd,+ (G) ≤
X
(k − 2)nk,+ (G).
2|k
Therefore we have
acd2,+ (G) =
n1,+ (G) +
and this completes the proof.
P
n1,+ (G) +
P2|k
knk,+ (G)
2|k
nk,+ (G)
> 2,
To prove Theorem 1.3(ii), we begin with two known observations on strongly real
characters.
Lemma 5.2. Let N be a normal subgroup of a finite group G such that G/N has odd
order. Then every strongly real character ϕ ∈ Irr(N) lies under a unique strongly
real irreducible character of G.
Proof. This is [19, Lemma 2.1(ii)].
Lemma 5.3. Let N ✁ G be such that G/N is a 2-group. Assume that ϕ ∈ Irr(N) has
2-defect 0, and that ϕ is G-conjugate to ϕ. Then there exists a strongly real character
χ ∈ Irr(G) such that [χN , ϕ]N = 1.
Proof. This is [28, Lemma 2.5].
CHARACTERS OF EVEN DEGREE AND NORMAL SYLOW 2-SUBGROUPS
13
We also need the following observation.
Lemma 5.4. Let P be a 2-group acting on an abelian group N of odd order such
that N is the unique minimal normal subgroup of N ⋊ P . Then |N| − 1 ≥ |P : Φ(P )|,
where Φ(P ) is the Frattini subgroup of P .
Proof. Suppose that N is a product of n copies of the cyclic groups Cq , where q is a
prime. Since N is the only minimal normal subgroup of N ⋊ P , we have CP (N) =
1, and the action of P on N induces a faithful irreducible representation X of P
over the field Fq . We extend this representation to the representation XFq over the
algebraically closure Fq . Since |P | is even and q is odd, by Maschke’s theorem, XFq
is completely reducible, and moreover, is faithful since X is faithful. Using the FongSwan theorem [23, Theorem 10.1] on lifts of irreducible Brauer characters in solvable
groups, we conclude that P has a complex faithful character, say χ, of degree n.
Now we apply [12, Theorem A] to deduce that the number of generators in a
minimal generating set for P , say d(P ), is at most (3/2)(n − s) + s, where s is the
number of linear constituents of χ. In particular, d(P ) ≤ [3n/2], and it follows that
|P : Φ(P )| ≤ 2[3n/2] .
As it is easy to check that 2[3n/2] ≤ 3n − 1 for every positive integer n, we then have
|P : Φ(P )| ≤ 3n − 1 ≤ q n − 1 = |N| − 1,
and the lemma follows.
Lemmas 5.3 and 5.4 allow us to control acd2,+ (G) in the following special situation.
Proposition 5.5. Let G = N ⋊P be a split extension of an elementary abelian group
of odd order N by a nontrivial 2-group P . Assume that N is the unique minimal
normal subgroup of G. Then acd2,+ (G) ≥ 4/3.
Proof. It is well known that groups of odd order have no non-principal real irreducible
character. Therefore 1N is the only real irreducible character of N. It follows that
every (strongly) real linear character of G has N inside its kernel. In particular,
n1,+ (G) = n1,+ (G/N) = n1,+ (P ) = |P : Φ(P )|.
Together with Lemma 5.4, this implies that
n1,+ (G) ≤ |N| − 1.
Hence, we can find s disjoint P -orbits Ω1 , . . . , Ωs on Irr(N) r {1N } of size k1 , . . . , ks ,
such that
k1 + . . . + ks ≥ n1,+ (G).
From the hypotheses, we know that P acts irreducibly and faithfully on N, and
therefore on Irr(N) as well since the two actions are permutationally isomorphic.
Therefore, each nontrivial character in Irr(N) is inverted by a central involution of P .
14
NGUYEN NGOC HUNG AND PHAM HUU TIEP
In other words, for any nontrivial ϕ ∈ Irr(N), ϕ and ϕ are P -conjugate. Lemma 5.3
then guarantees that there is a strongly real character χ{ϕ,ϕ} ∈ Irr(G) lying above
both ϕ and ϕ with [χN , ϕ]N = 1. Since IG (ϕ) splits over N, ϕ is extendible to
IG (ϕ), and hence χ{ϕ,ϕ} (1) = |G : IG (ϕ)|, which is an even number since ϕ is not
G-invariant.
Applying the above argument to λi ∈ Ωi , we get
Psa strongly real character χi ∈
Irr(G) of even degree ≥ ki ≥ 2; in particular, s ≤ ( i=1 ki )/2 and so
s
s
X
X
ki ≥ n1,+ (G).
ki − 4s ≥
3
i=1
i=1
It follows that the average degree of these s characters χ1 , . . . , χs and the (strongly)
real linear characters of G is at least
P
n1,+ (G) + si=1 ki
4
≥ .
n1,+ (G) + s
3
Since any other strongly real irreducible character of G has degree at least 2, we
conclude that acd2,+ (G) ≥ 4/3.
The proof of Theorem 1.3 is completed by
Theorem 5.6. Let G be a finite group. If acd2,+ (G) < 4/3 then G has a normal
Sylow 2-subgroup.
Proof. We assume that the statement is false, and let G be a minimal counterexample.
In particular, G is non-abelian. Let N be a minimal normal subgroup of G such
that N ⊆ G′ . As before, it follows that the degree of every strongly real, irreducible
character of G whose kernel does not contain N is at least 2, and hence acd2,+ (G/N) ≤
acd2,+ (G) < 4/3. By the minimality of G, we then deduce that G/N has a normal
Sylow 2-subgroup, and thus NP ✂ G, where P is a Sylow 2-subgroup of G.
Since acd2,+ (G) < 4/3 < 2, Theorem 5.1 guarantees that G is solvable, and so N
is elementary abelian. If N has even order, then NP = P and we would be done. So
we may assume that N is an elementary abelian group of odd order, and moreover,
G has no non-trivial normal 2-subgroup.
We observe that each strongly real linear character of G restricts to a strongly
real linear character of NP , and moreover, by Lemma 5.2, each strongly real linear
character of NP lies under a unique strongly real linear character of G. It follows that
n1,+ (NP ) = n1,+ (G). This and Lemma 5.2 imply that acd2,+ (NP ) ≤ acd2,+ (G) <
4/3. Thus, if NP < G, then by the minimality of G, we have P ✁ NP so that
NP = N × P , and hence P ✁ G, a contradiction. So we may assume that G = N ⋊ P .
Since G has no non-trivial normal 2-subgroup, we see that N is the unique minimal
normal subgroup of G.
Now we have all the hypotheses of Proposition 5.5, and therefore we conclude that
acd2,+ (G) ≥ 4/3. This contradiction completes the proof.
CHARACTERS OF EVEN DEGREE AND NORMAL SYLOW 2-SUBGROUPS
15
Acknowledgement
We are grateful to the referee for helpful comments and suggestions that have
significantly improved the exposition of the paper.
References
[1] J. H. Conway, R. T. Curtis, S. P. Norton, R. A. Parker, and R. A. Wilson, Atlas of Finite Groups,
Clarendon Press, Oxford, 1985.
[2] J. P. Cossey and H. N. Nguyen, Controlling composition factors of a finite group by its character
degree ratio, J. Algebra 403 (2014), 185–200.
[3] J. P. Cossey, Z. Halasi, A. Maróti, and H. N. Nguyen, On a conjecture of Gluck, Math. Z. 279
(2015), 1067–1080.
[4] S. Dolfi, G. Navarro, and Pham Huu Tiep, Primes dividing the degrees of the real characters,
Math. Z. 259 (2008), 755–774.
[5] S. Dolfi, E. Pacifici, L. Sanus, and P. Spiga, On the orders of zeros of irreducible characters, J.
Algebra 321 (2009), 345–352.
[6] W. Feit, Extending Steinberg characters, Linear algebraic groups and their representations,
Contemp. Math. 153 (1993), 1–9.
[7] D. Gluck, The largest irreducible character degree of a finite group, Canad. J. Math. 37 (1985),
442–451.
[8] Z. Halasi, C. Hannusch, and H. N. Nguyen, The largest character degrees of the symmetric and
alternating groups, Proc. Amer. Math. Soc. 144 (2016), 1947–1960.
[9] N. N. Hung, Characters of p′ -degree and Thompson’s character degree theorem, Rev. Mat.
Iberoam., to appear. http://arxiv.org/abs/1506.06450
[10] N. N. Hung, M. L. Lewis, and A. A. Schaeffer Fry, Finite groups with an irreducible character
of large degree, Manuscripta Math. 149 (2016), 523–546.
[11] I. M. Isaacs, ‘Character Theory of Finite Groups’, AMS Chelsea Publishing, Providence, Rhode
Island, 2006.
[12] I. M. Isaacs, The number of generators of a linear p-group, Canad. J. Math. 24 (1972), 851–858.
[13] I. M. Isaacs and G. Navarro, Groups whose real irreducible characters have degree coprime to
p, J. Algebra 356 (2012), 195–206.
[14] I. M. Isaacs, M. Loukaki, and A. Moretó, The average degree of an irreducible character of a
finite group, Israel J. Math. 197 (2013), 55–67.
[15] N. Itô, Some studies on group characters, Nagoya Math. J. 2 (1951), 17–28.
[16] M.L. Lewis and H. N. Nguyen, The character degree ratio and composition factors of a finite
group, Monatsh. Math. 180 (2016), 289–294.
[17] O. Manz, On the modular version of Ito’s theorem on character degrees for groups of odd order,
Nagoya Math. J. 105 (1987), 121–128.
[18] O. Manz and T. R. Wolf, Brauer characters of q ′ -degree in p-solvable groups, J. Algebra 115
(1988), 75–91.
[19] S. Marinelli and Pham Huu Tiep, Zeros of real irreducible characters of finite groups, Algebra
Number Theory 7 (2013), 567–593.
[20] G. O. Michler, Brauer’s conjectures and the classification of finite simple groups, Lecture Notes
in Math. 1178, Springer, Berlin, 1986, 129–142.
[21] G. O. Michler, A finite simple group of Lie type has p-blocks with different defects, p 6= 2, J.
Algebra 104 (1986), 220–230
16
NGUYEN NGOC HUNG AND PHAM HUU TIEP
[22] A. Moretó and H. N. Nguyen, On the average character degree of finite groups, Bull. Lond.
Math. Soc. 46 (2014), 454–462.
[23] G. Navarro, ‘Characters and Blocks of Finite Groups’, London Mathematical Society Lecture
Note Series 250, Cambridge University Press, Cambridge, 1998.
[24] G. Navarro and Pham Huu Tiep, Degrees of rational characters of finite groups, Adv. Math.
224 (2010), 1121–1142.
[25] G. Navarro and Pham Huu Tiep, Degrees and p-rational characters, Bull. Lond. Math. Soc. 44
(2012), 1246–1250.
[26] R. Rasala, On the minimal degrees of characters of Sn , J. Algebra 45 (1977), 132–181.
[27] G. Seitz and A. E. Zalesskii, On the minimal degrees of projective representations of the finite
Chevalley groups, II, J. Algebra 158 (1993), 233–243.
[28] Pham Huu Tiep, Real ordinary characters and real Brauer characters, Trans. Amer. Math. Soc.
367 (2015), 1273–1312.
Department of Mathematics, The University of Akron, Akron, Ohio 44325, USA
E-mail address: [email protected]
Department of Mathematics, University of Arizona, Tucson, Arizona 85721, USA
E-mail address: [email protected]
| 4math.GR
|
arXiv:1607.05341v1 [math.AC] 18 Jul 2016
A NOTE ON ITOH (e)-VALUATION RINGS OF
AN IDEAL
Youngsu Kim, Louis J. Ratliff, Jr. and David E. Rush
February 26, 2018
Abstract
Let I be a regular proper ideal in a Noetherian ring R, let e ≥ 2 be an integer, let
1
1
1
Te = R[u, tI, u e ]′ ∩ R[u e , t e ] (where t is an indeterminate and u = 1t ), and let re =
1
u e Te . Then the Itoh (e)-valuation rings of I are the rings (Te /z)(p/z) , where p varies
over the (height one) associated prime ideals of re and z is the (unique) minimal prime
ideal in Te that is contained in p. We show, among other things:
(1) re is a radical ideal if and only if e is a common multiple of the Rees integers of I.
(2) For each integer k ≥ 2, there is a one-to-one correspondence between the Itoh (k)valuation rings (V ∗ , N ∗ ) of I and the Rees valuation rings (W, Q) of uR[u, tI]; namely,
1
if F (u) is the quotient field of W , then V ∗ is the integral closure of W in F (u k ).
(3) For each integer k ≥ 2, if (V ∗ , N ∗ ) and (W, Q) are corresponding valuation rings,
as in (2), then V ∗ is a finite integral extension domain of W , and W and V ∗ satisfy the
Fundamental Equality with no splitting. Also, if uW = Qe , and if the greatest common
divisor of e and k is d, and c is the integer such that cd = k, then QV ∗ = N ∗ c and
[(V ∗ /N ∗ ) : (W/Q)] = d. Further, if uW = Qe and k = qe is a multiple of e, then there
1
exists a unit θe ∈ V ∗ such that V ∗ = W [θe , u k ] is a finite free integral extension domain
1
of W , QV ∗ = N ∗ q , N ∗ = u k V ∗ , and [V ∗ : W ] = k.
(4) If the Rees integers of I are all equal to e, then V ∗ = W [θe ] is a simple free integral
1
extension domain of W , QV ∗ = N ∗ = u e V ∗ , and [V ∗ : W ] = e = [(V ∗ /N ∗ ) : (W/Q)].
1
INTRODUCTION
All rings in this paper are commutative and have an identity element 1 6= 0, and our
terminology is mainly as in Nagata [7]. Thus a basis of an ideal is a generating set of the
ideal, the term altitude refers to what is often called dimension or Krull dimension, and
for a pair of local rings (R, M ) and (S, N ), S dominates R in case R ⊆ S and N ∩ R =
M , and we then write R ≤ S or (S, N) ≥ (R, M).
In 1988, Shiroh Itoh proved the following interesting and useful theorem in [5, p. 392,
lines 3-11] (the terminology is defined in Section 2):
1
Theorem 1.1. Let I be a regular proper ideal in a Noetherian ring R, let R = R[u, tI]
be the Rees ring of R with respect to I, let (W1 , Q1 ), . . . , (Wn , Qn ) be the Rees valuation
rings of uR, for j = 1, . . . , n, let uWj = Qj ej (so e1 , . . . , en are the Rees integers of I),
1
and let e ≥ 2 be an arbitrary common multiple of e1 , . . . , en . Also, let S = R[u e ], let T =
1
1
1
S′ ∩ R[u e , t e ], and let r = u e T. Then:
1
(1.1.1) r is a radical ideal, so the Rees integers of r and of (u e S)a are all equal to one.
(1.1.2) There is a one-to-one correspondence between the Rees valuation rings (V ∗ , N ∗ ) of
r and the Rees valuation rings (W, Q) of uR; namely, if F (u) is the quotient field of W ,
1
then V ∗ is the integral closure of W in F (u e ).
(1.1.3) Let (V ∗ , N ∗ ) and (W, Q) be corresponding Rees valuation rings of r and uR, ree
spectively, as in (1.1.2 ), so W = Wj for some j ∈ {1, . . . , n}. Then QV ∗ = N ∗ ej , so the
ramification index of V ∗ relative to W is equal to
e
ej .
Actually, the only part of this theorem that S. Itoh specifically stated in [5] was that
r is a radical ideal when e is the least common multiple of e1 , . . . , en . His proof of this
essentially shows that (1.1.1) - (1.1.3) hold, but his goals in [5] were to prove several nice
applications of the radicality of the ideal r, not to find additional properties of the Rees
valuation rings of this ideal.
However, it turns out that the Rees valuation rings of ideals like r have some additional
nice properties, and the goal of this present paper is to derive some of these properties. To
facilitate discussing these valuation rings we make the following definition.
Definition 1.2. For an arbitrary integer e ≥ 2, the Itoh (e)-valuation rings of I are
1
1
1
1
the Rees valuation rings of re = u e Te , where Te = R[u, tI, u e ]′ ∩ R[u e , t e ].
2
DEFINITIONS AND KNOWN RESULTS
In this section we recall the needed definitions and mention the needed known results
concerning them.
Definition 2.1. Let I be an ideal in a ring R. Then:
(2.1.1) R′ denotes the integral closure of R in its total quotient ring.
2
(2.1.2) Ia denotes the integral closure of I in R, so Ia is the ideal {x ∈ R | x is a root of an
equation of the form X n + i1 X n−1 + · · · + in = 0}, where ij ∈ I j for j = 1, . . . , n. The ideal
I is integrally closed in case I = Ia .
(2.1.3) The Rees ring of R with resect to I is the graded subring R(R, I) = R[u, tI] of
R[u, t], where t is an indeterminate and u = 1t .
(2.1.4) Assume that R is Noetherian and that I is a regular proper ideal in R (that is, I
contains a regular element of R and I 6= R), and let b1 , . . . , bg be regular elements in R that
generate I. For i = 1, . . . , g, let Ci = R[I/bi ], let pi,1 , . . . , pi,ni be the (height one) associated
prime ideals of bi Ci ′ (see (2.1.1)), let zi,j be the (unique; see Remark 2.2.1 below) minimal
prime ideal in Ci ′ that is contained in pi,j (possibly zi,h = zi,j for some h 6= j or zi,j =
(0)), and let Vi,j = (Ci ′ /zi,j )(pi,j /zi,j ) , so Vi,j is a DVR. Then the set RV(I) of all Vi,j (i =
1, . . . , g and j = 1, . . . , ni ) is the set of Rees valuation rings of I. (The Rees valuation
rings of I are well defined by I; they do not depend on the basis b1 , . . . , bg of I.)
(2.1.5) If (V, N ) is a Rees valuation ring of I (see (2.1.4)), then the Rees integer
of I with respect to V is the positive integer e such that IV = N e . If RV(I) =
{(V1,1 , N1,1 ), . . . , (Vg,ng , Ng,ng )}, and if IVi,j = Ni,j ei,j for i = 1, . . . , g and j = 1, . . . , ni ,
then e1,1 , . . . , eg,ng are called the Rees integers of I.
(2.1.6) If (W, Q) ≤ (V ∗ , N ∗ ) are DVRs such that V ∗ is a localization of an integral extension
domain of W , then the ramification index of V ∗ relative to W is the positive integer k
such that QV ∗ = N ∗k .
Remark 2.2. Let I be a regular proper ideal in a Noetherian ring R. Then:
(2.2.1) Concerning (2.1.4), it is shown in [9, Definition p. 213 and Propositions 2.7 and
2.13] that if b is a regular nonunit in the integral closure A′ of a Noetherian ring A, then
bA′ has a finite primary decomposition, each associated prime ideal p of bA′ has height one,
p contains exactly one associated prime ideal z of (0), and (A′ /z)(p/z) is a DVR.
(2.2.2) It is shown in [11, Proposition 10.2.3] that if V1 , . . . , Vn are the Rees valuation rings
of I, then (I k )a = ∩{I k Vj ∩ R | j = 1, . . . , n} for all k ∈ N>0 . (N>0 denotes the set of
positive integers.)
(2.2.3) It readily follows from Definition 2.1.4 that the set RV(I) of Rees valuation rings
of I is the disjoint union of the sets RV((I + z)/z), where z runs over the minimal prime
3
ideals w in R such that I + w 6= R. Also, if V is a Rees valuation ring of I and of (I + z)/z,
then the Rees integer of I with respect to V is the Rees integer of (I + z)/z with respect to
V.
(2.2.4) It follows from Definition 2.1.4 (and is proved in [11, Example 10.3.2]) that if I =
bR is a regular proper principal ideal in R, then the Rees valuation rings of I are the rings
(R′ /z)(p/z) , where p varies over the (height one) associated prime ideals of bR′ and z is the
(unique) minimal prime ideal in R′ that is contained in p. (It is readily checked that if A =
R′ ∩ R[ 1b ], then there is a one-to-one correspondence between the associated prime ideals p
of bR′ and the associated prime ideals q of bA; namely, q = p ∩ A, and then (R′ /z)(p/z) =
(A/z ′ )(q/z ′ ) , where z ′ = z ∩ A.)
(2.2.5) It follows from Definitions 2.1.4 and 2.1.5 that if A is a ring such that R ⊆ A ⊆ R′ ,
then I and IA have the same Rees valuation rings and the same Rees integers. And it also
follows that, for all positive integers k, IA and I k A have the same Rees valuation rings.
(2.2.6) It is well known (and is readily proved, much as in the proof of [1, (2.5)]), that if R
is a Noetherian integral domain and A is an integral extension domain of R, then the Rees
valuation rings of IA are the extensions of the Rees valuation rings of I to the quotient
field of A.
The following theorem is a special case of [13, Theorems 19 and 20, pp. 55 and 60-61];
the terminology “Fundamental Inequality” is due to O. Endler in [2, pp. 127-128].
Theorem 2.3. (Fundamental Inequality): Let (V, N ) be a DVR with quotient field F , let
E be a finite algebraic extension field of F , let [E : F ] = e, let (V1 , N1 ), . . . , (Vn , Nn ) be all
of the valuation rings of E that are extensions of V to E (so the integral closure V ′ of V
in E has exactly n maximal ideals M1 , . . . , Mn and Vj = V ′ Mj , j = 1, . . . , n), and for j =
1, . . . , n, let N Vj = Nj ej and [(Vj /Nj ) : (V /N )] = fj . Then
(FI)
n
X
ej fj ≤ e,
j=1
and the equality holds if the integral closure V ′ of V in E is a finite V -module.
Terminology 2.4. If (V, N ) and (V1 , N1 ), . . . , (Vn , Nn ) are as in Theorem 2.3, if n = 1,
and if the equality in (FI) holds, then it will be said that (V, N ) and (V1 , N1 ) satisfy the
Fundamental Equality with no splitting.
4
Notation 2.5. If R is an integral domain, then R(0) denotes the quotient field of R.
Therefore, if S is a finite algebraic extension domain of R, then [S(0) : R(0) ] denotes the
dimension of the quotient field S(0) of S over the quotient field R(0) of R. If S is a finite free
integral extension domain of R, and if it is clear that the rank of S is equal to [S(0) : R(0) ],
then we often write [S : R] in place of [S(0) : R(0) ].
The next three propositions are known, but we do not know specific references for them,
so we sketch their proofs.
Proposition 2.6. Let (W, Q) be a DVR, let Q = πW , let f ≥ 2 be an integer, let D =
1
1
W [π f ], and let P ′ = π f D. Then D is a DVR that is a simple free integral extension domain
of W , P ′ is the maximal ideal in D, QD = P ′ f , [D(0) : W(0) ] = f , and D/P ′ ∼
= W/Q.
Therefore W and D satisfy the Fundamental Equality with no splitting (see Terminology
2.4).
1
Proof. P ′ = (Q, π f )D is a maximal ideal in D, since D/P ′ ∼
= W/Q is a field, and P ′ =
1
π f D is a principal ideal, since Q is generated by π. Also, D is integral over W and Q is
the unique maximal ideal in W , so every maximal ideal in D contains QD, so it follows
that P ′ is the only maximal ideal in D, hence D is a DVR. It therefore follows that QD =
P ′ f . Therefore, since [D(0) : W(0) ] ≤ f , it follows from the Fundamental Inequality FI that
[D(0) : W(0) ] = f , hence D is a simple free integral extension domain of W . Thus, by (2.4),
the last statement is clear.
Proposition 2.7. Let M be a maximal ideal in a Noetherian ring R, and let m(X) be a
monic polynomial in R[X]. If the image m(X) of m(X) in (R/M )[X] is irreducible, then
m(X) is irreducible in R[X], M R[x] is a maximal ideal in R[x] = R[X]/(m(X)R[X]), and
R[x] is a simple free integral extension ring of R of rank equal to deg(m(X)).
Proof. By considering the maps R[X] → (R/M )[X] → (R/M )[χ], where χ is a root of the
irreducible (by hypothesis) polynomial m(X) in (R/M )[X], it follows that: m(X) is irreducible in R[X]; (M, m(X))R[X] is a maximal ideal; M R[x] = ((M, m(X))R[X])/(m(X)R[X])
is a maximal ideal in R[x]; and, R[x] is a simple free integral extension ring of R of rank
equal to deg(m(X)).
5
Proposition 2.8. Let I be a regular proper ideal in a Noetherian ring R, let b1 , . . . , bg
be regular elements in I that are a basis of I, and let R = R[u, tI] be the Rees ring of R
with respect to I. Then there exists a one-to-one correspondence between the Rees valuation rings (V, N ) of I and the Rees valuation rings (W, Q) of uR[u, tI]. Namely, if V =
(R[I/bi ]′ /z)(p/z) (where i ∈ {1, . . . , g}, p is an associated prime ideal of bi R[I/bi ]′ , and z is
the minimal prime ideal in R[I/bi ]′ that is contained in p), then W = V [tbi ]N V [tbi ] , where
tbi = t(bi + z), and Q = N W . (Note that tbi = tbi + z ′ (where z ′ = zR[u, t] ∩ R[u, tI])
corresponds to t · bi in the isomorphism between R[u, tI]/z ′ and (R/z)[u, t((I + z)/z)]; see
[10].)
Proof. Since t is transcendental over R and u =
1
t,
there is a one-to-one correspondence
between the minimal prime ideals z in R and the minimal prime ideals z ′ in R; namely, z ′
= zR[u, t] ∩ R, and then z = z ′ ∩ R. Thus it follows from Remark 2.2.3 that it suffices to
prove this proposition for the case when R is a Noetherian integral domain.
Therefore assume that R is a Noetherian domain, fix b ∈ {b1 , . . . , bg }, let C = R[I/b], and
let C′ = R[ tb1 ]′ , so R′ [ tb1 ] = C′ = C ′ [tb, tb1 ]. Also, uC′ = bC′ , so, since tb is transcendental
over C ′ , there exists a one-to-one correspondence between the (height one) associated prime
ideals p of bC ′ and the (height one) associated prime ideals p′ of uC′ ; namely, p′ = pC′
(and p = p′ ∩ C ′ ). Therefore there exists a one-to-one correspondence between the DVRs
V = C ′ p and the DVRs W = C′ p′ , and for corresponding V and W , if V = R[I/b]′ p and N
= pV , then W = V [tb]N V [tb] and Q = N W = pW .
It follows that, for each Rees valuation ring (V, N ) of I of the form V = R[I/b]′ p , the
ring W = V [tb]pV [tb] is a Rees valuation ring of uR, and W ∩ R(0) = V .
Finally, let W be a Rees valuation ring of uR, say W = R′ p′ , where p′ is a (height one)
associated prime ideal of uR′ (see Remark 2.2.4). To complete the proof of the one-to-one
correspondence, it suffices to show that there exists b ∈ {b1 , . . . , bg } and a Rees valuation
ring (V, N ) of I such that V = R[I/b]′ p and W = V [tb]N V [tb] .
For this, tb ∈
/ p′ for some b ∈ {b1 , . . . , bg }, by [8, Lemma 3.2] (the assumption in [8]
that R is analytically unramified is not used in the proof of Lemma 3.2). Therefore W =
R[ tb1 ]′ p′′ , where p′′ = p′ W ∩ R[ tb1 ]′ (so p′′ is a (height one) associated prime ideal of uR[ tb1 ]′ =
bR[ tb1 ]′ , and R[ tb1 ]′ = C ′ [tb, tb1 ], where C = R[I/b]. Let p = p′′ ∩ C ′ . Then p is a (height one)
6
associated prime ideal of bC ′ and p′′ = pC ′ [tb, tb1 ], so it follows from the second preceding
paragraph that W = V [tb]N V [tb] , where V = C ′ p and N = pV .
3
PROPERTIES OF ITOH (e)-VALUATION RINGS
In this section we show that Itoh (e)-valuation rings have several nice properties. For this,
we need the following proposition, which is essentially a corollary of Proposition 2.7.
Proposition 3.1. Let I be a regular proper ideal in a Noetherian ring R, let b1 , . . . , bg be
regular elements in I that generate I, and let R = R[u, tI] be the Rees ring of R with respect
to I. Let (V, N ) and (W, Q) be corresponding (as in Proposition 2.8) Rees valuation rings
of I and uR, respectively, say V = (R[I/b]′ /z)(p/z) , where b ∈ {b1 , . . . , bg }, p is a (height
one) associated prime ideal of bR[I/b]′ , and z is the minimal prime ideal in R[I/b]′ that
is contained in p, so W = V [tb]N V [tb] , where b = b + z. Let e ≥ 2 be an integer, let v ∈
V − N , let me (X) = X e −
v
,
tb
and let θe be a root of me (X) in a fixed algebraic closure of
the quotient field of W . Then U = W [θe ] is a DVR that is a simple free integral extension
domain of W , QU is its maximal ideal, and [U : W ] = e = [U/(QU ) : W/Q]. Therefore W
and U satisfy the Fundamental Equality with no splitting (see Terminology 2.4).
Proof. Since W = V [tb]N V [tb] (with tb transcendental over V ) and Q = N W is an extended
ideal from V , tb + Q ∈ W/Q is transcendental over V /N , so it follows that the image
me (X) ∈ (W/Q)[X] of me (X) is irreducible, hence [(W/Q)[y] : (W/Q)] = e, where y =
X + me (X)(W/Q)[X]. Therefore it follows from Proposition 2.7 that: me (X) is irreducible
in the UFD W [X]; M = QU is a (principal) maximal ideal, where U = W [θe ]; and, [U : W ]
= e. Further, U is integral over W and Q is the unique maximal ideal in W , so every
maximal ideal in U contains the maximal ideal QU , hence U is a DVR that is a simple free
integral extension domain of W and M is its maximal ideal. Therefore, by (2.4), the last
statement is clear.
Our first theorem, Theorem 3.2, is an expanded version of Itoh’s Theorem (see Theorem
1.1; note that (1.1.3) is proved in (3.2.5)).
Theorem 3.2. Let I be a regular proper ideal in a Noetherian ring R, let b1 , . . . , bg be
regular elements in I that are a basis of I, let (V1 , N1 ), . . . , (Vn , Nn ) be the Rees valuation
7
rings of I, let e1 , . . . , en be the Rees integers of I, let e be an arbitrary common multiple of
e1 , . . . , en , and let fj =
e
ej
(j = 1, . . . , n). Also, let R = R[u, tI] be the Rees ring of R with
1
e
1
1
1
respect to I, let S = R[u ], let T = S′ ∩ R[u e , t e ], and let r = u e T. Then:
(3.2.1) r is a radical ideal.
(3.2.2) There is a one-to-one correspondence between the Itoh (e)-valuation rings (V ∗ , N ∗ )
of I and the Rees valuation rings (W, Q) of uR; namely, given W , if F (u) is the quotient
1
field of W , then V ∗ is the integral closure of W in F (u e ).
(3.2.3) Let (W, Q) and (V ∗ , N ∗ ) be corresponding (as in (3.2.2 )) valuation rings, and let
(V, N = πV ) be the Rees valuation ring of I that corresponds (as in Proposition 2.8) to W ,
so V ≤ W ≤ V ∗ and N W = Q. Assume that V = Vj , so IV = N ej (= Nj ej ).
Then there exists a unit θ ∈ V ∗ such that (U, P ) = (W [θ], N ∗ ∩ W [θ]) is an Itoh (ej )valuation ring of I that is a simple free integral extension domain of W and P = QU is
the maximal ideal in U , so the ramification index of U relative to W is equal to one (see
Definition 2.1.6 ).
1
Also, QU = πU = yU , where y = u ej , and [U : W ] = ej = [(U/P ) : (W/Q)]. Therefore
W and U satisfy the Fundamental Equality with no splitting.
(3.2.4) Let (V, N ) ≤ (U, P ) ≤ (V ∗ , N ∗ ) (with V = Vj ) be as in (3.2.3 ). Then there exists
a nonunit x ∈ V ∗ such that V ∗ = U [x] is a simple free integral extension domain of U , N ∗
1
1
1
= xV ∗ (where x = y fj with y = u ej (as in (3.2.3 )), so x = u e ), and P V ∗ = N ∗fj , so the
ramification index of V ∗ relative to U is equal to fj . Also, [V ∗ : U ] = fj , and V ∗ /N ∗ ∼
=
U/P , so U and V ∗ satisfy the Fundamental Equality with no splitting.
1
(3.2.5) V ∗ = W [θ, x], N ∗ = xV ∗ = u e V ∗ , where θ is as in (3.2.3 ) and x is as in (3.2.4 ),
and V ∗ is a finite free integral extension domain of W . Also, the ramification index of V ∗
relative to W is equal to fj , [(V ∗ /N ∗ ) : (W/Q)] = ej , and [V ∗ : W ] = e, so W and V ∗
satisfy the Fundamental Equality with no splitting.
(3.2.6) Assume that e1 = · · · = en = e, and let (W, Q) and U = W [θ] be as in (3.2.3 ).
1
Then V ∗ = U is a simple free integral extension domain of W , P = QU = u e U is the
maximal ideal in U , and [U : W ] = ej = [(U/P ) : (W/Q)]. Therefore W and V ∗ (= U )
1
satisfy the Fundamental Equality with no splitting, and u e T is a radical ideal, where T =
1
1
1
R[u, tI, u e ]′ ∩ R[u e , t e ].
8
Proof. We first prove (3.2.3) - (3.2.5). For this, fix an Itoh (e)-valuation ring (V ∗ , N ∗ ) of I.
Then, by Definitions 1.2 and 2.1.4, there exists a (height one) associated prime ideal p∗ of
1
1
1
u e T and a minimal prime ideal z ∗ ⊂ p∗ such that V ∗ = (T/z ∗ )(p∗ /z ∗ ) . Since T ⊂ R[u e , t e ]
1
1
1
1
and T and R[u e , t e ] have the same total quotient ring, it follows that z = z ∗ R[u e , t e ] ∩ R
1
1
is a minimal prime ideal in R and z ∗ = zR[u e , t e ] ∩ T. Also, z ′ = zR[u, t] ∩ R is a minimal
prime ideal in R and z ∗ ∩ R = z ′ .
1
Let F be the quotient field of R/z, so F (u) (resp., F (u e )) is the quotient field of R/z ′
1
(resp., V ∗ and T/z ∗ ). Also, by Definition 1.2, V ∗ is a Rees valuation ring of u e T, so V ∗ is
1
also a Rees valuation ring of (u e T + z ∗ )/z ∗ , by Remark 2.2.3, so V ∗ is also a Rees valuation
ring of (uT + z ∗ )/z ∗ , by Remark 2.2.5.
Therefore W = V ∗ ∩ F (u) is a Rees valuation ring of the ideal (uR + z)/z, by Remark
1
2.2.6 (and so also of uR, by Remark 2.2.3), hence V ∗ is an extension of W to F (u e ).
Let V be the Rees valuation ring of I that corresponds (as in Proposition 2.8) to W ,
and assume that V = Vj = (C ′ /z)(p/z) , where C ′ is the integral closure of C = R[I/b] in its
total quotient ring, b ∈ {b1 , . . . , bg }, p is a (height one) associated prime ideal of bC ′ , and
z is the (unique) minimal prime ideal in C ′ that is contained in p. Let N and Q be the
maximal ideals in V and W , respectively. Then Q = N W by Proposition 2.8, so
(3.2.7) tb + Q ∈ W/Q is transcendental over V /N, where b = b + z.
Let N = πV , so, by hypothesis, bV = IV = N ej = π ej V . Therefore uW = bW = IW
= Qej = π ej W , so
(3.2.8) there exist units v ∈ V and w ∈ W such that b = vπ ej and u = wπ ej .
Since b = u(tb) (see the last part of Proposition 2.8), it follows from (3.2.8) that
(3.2.9)
w=
v
.
tb
1
Let θ = w ej (in a fixed algebraic closure (F (u))∗ of F (u)), and let U = W [θ]. Then it follows
from Proposition 3.1 that: U is a DVR that is a simple free integral extension domain of
W ; QU is its maximal ideal; [U : W ] = ej = [(U/(QU )) : (W/Q)]; and, W and U satisfy
the Fundamental Equality with no splitting.
9
Continuing with the proof of (3.2.3), we next show that U ≤ V ∗ , and we first show that
1
1
1
1
θ ∈ V ∗ . For this, wπ ej = u, by (3.2.8), θ = w ej , and u e ∈ V ∗ , so θπ = (wπ ej ) ej = u ej
1
1
= (u e )fj ∈ V ∗ , hence θπ = u ej ∈ V ∗ . Further, π ∈ Q ⊂ V ∗ , so θ is in the quotient field
1
1
F (u e ) of V ∗ . Moreover, θ = w ej is integral over W and is a unit, W ≤ V ∗ , and V ∗ is
1
integrally closed in F (u e ), so θ ∈ V ∗ and is a unit in V ∗ . Therefore U = W [θ] ≤ V ∗ , and
1
u ej U = θπU = πU = QU = N ∗ ∩ U is the maximal ideal in U . Thus, to complete the proof
of (3.2.3), it remains to show that U is an Itoh (ej )-valuation ring of I.
1
1
1
For this, U = W [θ] ⊇ W [u ej ]′ ⊇ R[u ej ]′ ⊇ Tej + R′ , where Tej = R[u, tI, u ej ]′ ∩
1
1
R[u ej , t ej ], and P ∩ R′ = (P ∩ W ) ∩ R′ = Q ∩ R′ is a height one associated prime ideal of
1
1
uR′ , so P ∩ Tej = q, say, is a height one associated prime ideal of u ej Tej (since u ej ∈ P ).
Therefore U ≥ (Tej )q , (Tej )q is an Itoh (ej )-valuation ring of I, and U and (Tej )q are
DVRs with the same quotient field, so U = (Tej )q is an Itoh (ej )-valuation ring of I. Thus
(3.2.3) holds.
1
1
1
For (3.2.4), let y = u ej , let x = y fj (so x = u e , since ej · fj = e), and let D = U [x].
By (3.2.3), P = yU is the maximal ideal in U , so it follows from Proposition 2.6 that: D
is a DVR that is a simple free integral extension domain of U ; (QU )D = P ′ fj , where P ′ =
1
π fj D is the maximal ideal in D (so D/P ′ ∼
= U/P ); [D : U ] = fj ; and, U and D satisfy the
Fundamental Equality with no splitting.
1
1
Further, since u e = x ∈ D, D and V ∗ have the same quotient field F (u e ). Thus, V ∗
1
and D are DVRs in F (u e ), and D ≤ V ∗ , so it follows that D = V ∗ and P ′ = N ∗ . Therefore
(3.2.4) holds.
1
It follows from (3.2.3) and (3.2.4) that V ∗ = W [θ, x] (= W [θ, u e ]) is a finite free integral
1
extension domain of W , that N ∗ = xV ∗ = u e V ∗ , and that [V ∗ : W ] = ej · fj = e. Also,
since the ramification index of U over W is equal to one, by (3.2.3), and the ramification
index of V ∗ = D over U is equal to fj , by (3.2.4), it follows that the ramification index
of V ∗ over W is equal to fj . Further, [(U/P ) : (W/Q)] = ej , by (3.2.3), and V ∗ /N ∗ ∼
=
U/P , by (3.2.4), so [(V ∗ /N ∗ ) : (W/Q)] = ej . Thus it follows that W and V ∗ satisfy the
Fundamental Equality with no splitting, hence (3.2.5) holds.
Moreover, since V ∗ is a finite free integral extension domain of W , and since V ∗ is
integrally closed, it follows that V ∗ is the integral closure of W (= Wj ) in the quotient field
10
1
F (u e ) of V ∗ . Therefore it has been shown that if V ∗ is an Itoh (e)-valuation ring of I,
1
then: V ∗ = (T/z ∗ )(p∗ /z ∗ ) (for some (height one) associated prime ideal p∗ of u e T, where
z ∗ is the minimal prime ideal in T that is contained in p∗ ); z = z ∗ ∩ R and z ′ = z ∗ ∩ R are
minimal prime ideals in R and R, respectively; and, if F (resp., F (u)) is the quotient field
1
of R/z (resp., R/z ′ ), then V ∗ is the integral closure of W (= V ∗ ∩ F (u)) in F (u e ), and W
is a Rees valuation ring of uR (and of (uR + z ′ )/z ′ ).
1
It follows that each Itoh (e)-valuation ring V ∗ (with quotient field F (u e )) corresponds
to the Rees valuation ring W (= V ∗ ∩ F (u)) of uR. On the other hand, if W is a Rees
valuation ring of uR, then: W = (R′ /z ′ )(p′ /z ′ ) for some (height one) associated prime ideal
p′ of uR, where z ′ is the minimal prime ideal in R that is contained in p′ ; z = z ′ ∩ R and
1
1
z ∗ = zR[u e , t e ] ∩ T are minimal prime ideals in R and T, respectively; and, if F (resp.,
1
F (u e )) is the quotient field of R/z (resp., T/z ∗ ) and W ′′ is the integral closure of W in
1
F (u e ), then for each maximal ideal M in W ′′ , W ′′ M is a Rees valuation ring of (uT+z ∗ )/z ∗
1
and of (u e T + z ∗ )/z ∗ , by Remarks 2.2.6 and 2.2.5, so each V ′′ M is a Rees valuation ring
1
of uT and of u e T, by Remark 2.2.3, hence each such W ′′ M is an Itoh (e)-valuation ring of
I, by Definition 1.2. It therefore follows from the first part of this paragraph that W ′′ has
exactly one maximal ideal, so the one-to-one correspondence of (3.2.2) holds.
1
For (3.2.1), u e V ∗ = N ∗ , by (3.2.4). Therefore, since V ∗ is an arbitrary Itoh (e)-valuation
1
ring of I, it follows from Definitions 1.2 and 2.1.5 that the Rees integers of u e T are all
1
1
1
equal to one. Also, since T = R[u e ]′ ∩ R[u e , t e ], by [9, Remarks (ii) p. 215] it follows
1
1
1
1
that u e T = (u e T)a . Therefore it follows from Remark 2.2.2 that r = u e T = (u e T)a =
1
∩{u e Vj ∗ ∩ T | j = 1, . . . , n} = ∩{Nj ∗ ∩ T | j = 1, . . . , n} is a radical ideal. Thus (3.2.1)
holds, so (3.2.1) - (3.2.5) hold.
Finally, for (3.2.6), let W be a Rees valuation ring of uR, let V ∗ be the corresponding
(as in (3.2.2)) Itoh (e)-valuation ring of I, and let U be as in (3.2.3), so W ≤ U ≤ V ∗ .
Then there exists j ∈ {1, . . . , n} such that [U : W ] = ej (by (3.2.3)) = e (by the hypothesis
in (3.2.6)) = [V ∗ : W ] (by (3.2.5)), so V ∗ = U is a simple free integral extension domain of
1
1
W . Also, P = N ∗ = u e V ∗ (by (3.2.5)) = u e U . Since this holds for all Itoh (e)-valuation
1
rings (U, P ) of I, u e T is a radical ideal, hence (3.2.6) holds.
Concerning Theorem 3.2.6, it is shown in Corollary 4.12 below that there always exists
11
a finite integral extension ring A of R such that the Rees integers of IA are all equal to e,
where e is an arbitrary common multiple of the Rees integers e1 , . . . , en of I, so Theorem
3.2.6 directly applies to IA in place of I.
Remark 3.3. Let (V, N ) be a Rees valuation ring of I and assume that IV = N k . Let
(W, Q) be the Rees valuation ring of uR[u, tI] that corresponds (as in Proposition 2.8) to
V , and let h be an arbitrary positive integer. Then it follows from Theorem 3.2.3 and 3.2.4
1
1
1
1
that: U = W [u k ]′ and D = W [u hk ]′ = U [(u k ) h ]′ are DVRs that are finite free integral
1
1
extension domains of W ; the maximal ideal of U (resp., D) is P = u k U (resp., M = u hk D);
the ramification index of D relative to U (resp., U relative to W ) is equal to h (resp., 1);
[D : U ] = h and [U : W ] = k; and, D/M ∼
= U/P and [(U/P ) : (W/Q)] = k. It therefore
follows that W and U (resp., W and D, U and D) satisfy the Fundamental Equality with
no splitting. Also, it follows from Theorem 3.2.3 (resp., 3.2.4) that U (resp., D) is an Itoh
(k)-valuation (resp., (hk)-valuation) ring of I.
Remark 3.4. The following result is called the Theorem of Independence of Valuations,
and it is proved in [7, (11.11)]: Let (V1 , N1 ), . . . , (Vn , Nn ) be valuation rings with the same
quotient field F , and assume there are no containment relations among the Vj . Let V =
V1 ∩ · · · ∩ Vn and let Pj = Nj ∩ V (j = 1, . . . , n). Then P1 , . . . , Pn are the maximal ideals
in V and Vj = VPj for j = 1, . . . , n.
Corollary 3.5. Let I = (b1 , . . . , bg )R be a nonzero proper ideal in a Noetherian integral
domain R, let F be the quotient field of R, let R = R[u, tI], let (W1 , Q1 ), . . . , (Wn , Qn ) be
the Rees valuation rings of uR, let e be a positive common multiple of the Rees integers
1
e1 , . . . , en of uR, let W = W1 ∩ · · · ∩ Wn , and let D = W[x], where x = u e in a fixed
algebraic closure F (u)∗ of F (u). Then:
(1) W is a semi-local Dedekind domain with exactly n maximal ideals Pj = Qj ∩ W, j =
1, . . . , n.
(2) D is a simple free integral extension domain of W of rank e, and D has exactly n
maximal ideals Mj = (Pj , x)D.
(3) There exist distinct elements θ1 , . . . , θn in the quotient field F (x) of D such that:
(3.1) D′ is the intersection of the Itoh (e)-valuation rings (V1 ∗ , N1 ∗ ), . . . , (Vn ∗ , Nn ∗ )
of I, where Vj ∗ = Wj [θj , x] = D′ Mj , where Mj = Nj ∗ ∩ D′ .
12
(3.2) D′ = D[θ1 , . . . , θn ] is a semi-local Dedekind domain that is a finite integral
extension domain of D, and D′ has exactly n maximal ideals Mj = Nj ∗ ∩ D′
(Pj , x, θ1 , . . . , θj−1 , θj+1 , . . . , θn )D′ .
(3.3) The Jacobson radical of D′ is xD′ .
(4) Assume that e1 = · · · = en = e and that there exists b ∈ I such that bWj = IWj for
j = 1, . . . , n. Then, with D′ , as in (3.1 ), and Qj = πj Wj for j = 1, . . . , n,
x
π1 ···πn
is a
x
x
unit in D′ , D′ = D[ π1 ···π
] = W[ π1 ···π
], and xD′ is the Jacobson radical of D′ .
n
n
Proof. (1) and part of (3.1) follow immediately from the Independence of Valuations Theorem (see Remark 3.4). The first part of (2) is clear. Also, each Mj is a maximal ideal in
D, since D/Mj ∼
= Wj /Qj . Further, u (resp., x) is in the Jacobson radical of W (resp., D),
xD ∩ W = uW (since D is a free W-algebra), and D/(xD) ∼
= W/(uW), so it follows that
W, W/(uW), D, and D/(xD) each have exactly n maximal ideals, so (2) holds.
For (3), for j = 1, . . . , n, let (Vj , Nj ) be the Rees valuation ring of I that corresponds (as
in Proposition 2.8) to (Wj , Qj ) (so there exists bσ(j) (where σ : {1, . . . , n} → {1, . . . , g})
I ′
I ′
] )pj , for some (height one) prime ideal pj in R[ bσ(j)
] ),
in {b1 , . . . , bg } such that Vj = (R[ bσ(j)
so Wj = Vj [tbσ(j) ]Nj Vj [tbσ(j) ] , Wj ∩ F = Vj , and Qj ∩ F = Nj .
Let V = V1 ∩ · · · ∩ Vn , and for j = 1 . . . , n, let qj = Nj ∩ V. Then V = W ∩ F and W =
V[u]S , where S = V[u] − (q1 V[u] ∪ · · · ∪ qnV[u]), and qj = (Qj ∩ F ) ∩ V = (Qj ∩ W) ∩ V =
Pj ∩ V, for j = 1, . . . , n. Therefore it follows from the Independence of Valuations Theorem
(see Remark 3.4) that V is a semi-local Dedekind domain with exactly n maximal ideals
qj (j = 1, . . . , n). Thus V is a Principal Ideal Domain, by [12, Theorem 16, p. 278 ], so
for j = 1, . . . , n, there exists πj ∈ qj such that qj = πj V, so: πj Vj = (Nj ∩ V)Vqj = Nj ;
πj W = qj W = (Pj ∩ V)W = Pj ; and, πj Wj = qj WQj ∩W = Pj Wj = Qj . Then, since the
Rees integers of uR are e1 , . . . , en (by hypothesis), uW = P1 e1 ∩ · · · ∩ Pn en = P1 e1 · · · Pn en
= π1 e1 · · · πn en W, so there exists a unit w ∈ W such that u = wπ1 e1 · · · πn en , hence
(*1)
there exists a unit wj ∈ Wj such that u = wj πj ej (in Wj = WPj ), for j = 1, . . . , n,
where
(*2)
wj =
Y e
\
u
i
=
w
π
∈
(
Pi ) − Pj , for j = 1, . . . , n.
e
i
πj j
i6=j
i6=j
13
Also, since ej is the Rees integer of I with respect to Vj and IVj = bσ(j) Vj , there exists
a unit vj ∈ Vj such that
(*3)
bσ(j) = vj πj ej in Vj , for j = 1, . . . , n.
Therefore, since bσ(j) = u(tbσ(j) ) (in R ⊆ Wj ), it follows from (*3) and (*1) that vj πj ej =
bσ(j) = u(tbσ(j) ) = wj πj ej (tbσ(j) ), hence
(*4)
wj =
vj
and wj + Pj is transcendental over V/(Pj ∩ V) for j = 1, . . . , n.
tbσ(j)
Therefore
(*5)
mej (X) = X ej − (wj + Pj ) is irreducible in (W/Pj )[X] f or j = 1, . . . , n.
1
For j = 1, . . . , n, let θj = wj ej in the (fixed) algebraic closure F (u)∗ of F (u). Then it
is shown in Theorem 3.2.3 - 3.2.5 that, for the Itoh (e)-valuation ring Vj ∗ of I,
(*6)
Vj ∗ = Wj [θj , x] and θj is a unit in Vj ∗ , for j = 1, . . . , n.
Also, V1 ∗ ∩ · · · ∩ Vn ∗ is the integral closure W∗ of W in F (x), by Theorem 3.2.2, and x =
1
u e ∈ W∗ (since u ∈ W), so it follows that W∗ = W[x]′ = D′ .
1
Further, for j = 1, . . . , n, θj = wj ej ∈ W∗ , by (*2) and (*6), and by (*2)
θ1 , . . . , θj−1 , θj+1 , . . . , θn ∈ Nj ∗ ∩ W∗ = Nj ∗ ∩ D′ = Mj , for j = 1, . . . , n.
Hence by (*2) and (*6) we see that
(*7) for i = 1, . . . , n, θi ∈ ∩{Nj ∗ ∩ D′ | j = 1, . . . , i − 1, i + 1, . . . , n} − (Ni ∗ ∩ D′ ).
Let E = D[θ1 , . . . , θn ]. Then, for j = 1, . . . , n, x ∈ D ⊆ E ⊆ W∗ (by (*2) and (*6))
⊆ Vj ∗ = Wj [θj , x]. Therefore Wj = WNj ∗ ∩W ⊆ DNj ∗ ∩D ⊆ ENj ∗ ∩E ⊆ W∗ Nj ∗ ∩W∗ = Vj ∗
= (Wj [x])[θj ] ⊆ ENj ∗ ∩E , so Vj ∗ = ENj ∗ ∩E , for j = 1, . . . , n. Also, D and W∗ have exactly
n maximal ideals, so it follows from integral dependence that E has exactly n maximal
ideals. Further, for each integral domain A, A = ∩{AM | M is a maximal ideal in A}, by
[7, (33.9)]. Therefore it follows that E = D[θ1 , . . . , θn ] = V1 ∗ ∩ · · · ∩ Vn ∗ = D′ . Thus (3.1)
and (3.2) hold.
14
(3.3) follows immediately from Theorem 3.2.4.
Finally, for (4), let (Vj , Nj ) be the Rees valuation ring of I that corresponds (as in
Proposition 2.8) to (Wj , Qj ). Then it follows from the hypothesis on b that, for j = 1, . . . , n,
Wj = Vj [tb]Nj Vj [tb] and Qj = Nj Wj .
Let V = V1 ∩ · · · ∩ Vn , so by Remark 3.4 V is a semi-local Dedekind domain with exactly
n maximal ideals Nj ∩ V (= Pj ∩ V). Thus V is a Principal Ideal Domain, by [12, Theorem
16, p. 278], so for j = 1, . . . , n there exists πj ∈ Nj ∩ V such that πj V = Nj ∩ V.
Let α = π1 · · · πn , let N = N1 ∩ · · · ∩ Nn , and let Q = Q1 ∩ · · · ∩ Qn . Then, since the
Rees integers of I are all equal to e, it follows that
(3.5.4.1) IV = bV = Ne = αe V and IW = uW = Qe = αe W.
It follows from (3.5.4.1) that
(3.5.4.2) there exist units v ∈ V and w ∈ W such that b = vαe ∈ V and u = wαe ∈ W.
Since b = u(tb) in R ⊆ W, and since Qj = Nj Vj [tb]Nj Vj [tb] for j = 1, . . . , n, it follows
from (3.5.4.2) that
(3.5.4.3) w =
v
and w + Pj is transcendental over V/(Pj ∩ V) for j = 1, . . . , n.
tb
Therefore
(3.5.4.4) me (X) = X e − (w + Pj ) is irreducible in (W/Pj )[X] for j = 1, . . . , n.
1
Let θ = w e in the (fixed) algebraic closure F (u)∗ of F (u). Then it follows from Proposition 2.7, together with (3.5.4.4), that: me (X) is irreducible in the UFD W[X]; for j =
1, . . . , n, Mj = Pj E is a (principal) maximal ideal, where E = W[θ]; and, [E : W] = e.
Also, E is integral over W, so it follows that M1 , . . . , Mn are the only nonzero prime
ideals in E, hence E is a semi-local Dedekind domain that is a simple free integral extension
domain of W and QE = (Q1 ∩ · · · ∩ Qn )E = (M1 ∩ · · · ∩ Mn )E = π1 · · · πn E = αE is its
Jacobson radical J. Therefore D′ = E = W[θ] = D[θ], and since w =
1
e
it follows that θ = w =
x
α
u
αe
is a unit in W,
is a unit in D′ and that J = αE = xE.
The next remark lists several well known facts concerning finite field extensions and
ramification.
15
Remark 3.6. Let (U1 , P1 ) ≤ (U2 , P2 ) ≤ (U3 , P3 ) be DVRs such that U3 is a finite integral
extension of U1 . Then:
(1) [(U3 )(0) : (U1 )(0) ] = [(U3 )(0) : (U2 )(0) ] · [(U2 )(0) : (U1 )(0) ].
(2) [(U3 /P3 ) : (U1 /P1 )] = [(U3 /P3 ) : (U2 /P2 )] · [(U2 /P2 ) : (U1 /P1 )].
Also, if we let r3,1 (resp., r3,2 , r2,1 ) denote the ramification index of U3 relative to U1 (resp.,
U3 relative to U2 , U2 relative to U1 ), then:
(3) r3,1 = r3,2 · r2,1 .
Further, if U1 and U3 satisfy the Fundamental Equality with no splitting, then
(4) r3,1 · [(U3 /P3 ) : (U1 /P1 )] = [(U3 )(0) : (U1 )(0) ].
It then follows from (1) - (4) and the Fundamental Inequality, (2.3), that:
(5) r3,2 · [(U3 /P3 ) : (U2 /P2 )] = [(U3 )(0) : (U2 )(0) ].
(6) r2,1 · [(U2 /P2 ) : (U1 /P1 )] = [(U2 )(0) : (U1 )(0) ].
(7) Hence, if U1 and U3 satisfy the Fundamental Equality with no splitting, then both (U1
and U2 ) and (U2 and U3 ) satisfy the Fundamental Equality with no splitting.
The next proposition shows that Theorem 3.2.2 holds for all integers greater than one.
Proposition 3.7. With the notation of Theorem 3.2, let k ≥ 2 be an arbitrary integer, let
1
1
1
Sk = R[u k ], and let Tk = Sk ′ ∩ R[u k , t k ]. Then:
(1) There exists a one-to-one correspondence between the Itoh (k )-valuation rings (U, P ) of
I and the Rees valuation rings (W, Q) of uR; namely, given W , if F (u) is the quotient
1
field of W , then U is the integral closure of W in F (u k ).
(2) Let W and U be corresponding (as in (1 )). Then U is a finite integral extension domain
of W , and W and U satisfy the Fundamental Equality with no splitting,
Proof. Let m be the least common multiple of e1 , . . . , en , and let e = k · m. Then it is clear
1
1
1
that R ⊆ Tk ⊆ Te = R[u e ]′ ∩ R[u e , t e ] and that Te is integral over R. Also, by Theorem
3.2.2 there exists a one-to-one correspondence between the Rees valuation rings (W, Q) of
uR and the Itoh (e)-valuation rings (V ∗ , N ∗ ) of I, and V ∗ is the integral closure of W in
the quotient field of V ∗ . It follows from this, and integral dependence, that (1) holds.
For (2), it follows from the proof of (1) that if W and V ∗ are as in (1), then there exists
an Itoh (k)-valuation ring U of I such that W ≤ U ≤ V ∗ . Also, V ∗ is a finite free integral
16
extension domain of W , by Theorem 3.2.5, hence U is a finite integral extension domain of
W . Further, W and V ∗ satisfy the Fundamental Equality with no splitting, by Theorem
3.2.5, so it follows from Remark 3.6.7 that (2) holds.
Terminology 3.8. If (W, Q) is a Rees valuation ring of uR[u, tI] and (U, P ) is the corresponding (as in Proposition 3.7.1) Itoh (k)-valuation ring of I (so U ≥ W and U is the
1
integral closure of W in the quotient field F (u k ) of U ), then we say that (U, P) is the
Itoh (k)-valuation ring of I that corresponds to (W, Q). Also, if (V, N ) is the Rees
valuation ring of I that corresponds (as in Proposition 2.8) to (W, Q), then we say that
(U, P) is the Itoh (k)-valuation ring of I that corresponds to (V, N).
Proposition 3.9. Let I be a regular proper ideal in a Noetherian ring R, let (V, N ) be a
Rees valuation ring of I, let (W, Q) be the Rees valuation ring of uR[u, tI] that corresponds
(as in Proposition 2.8) to (V, N ), and let (Uk , Pk ) be the Itoh (k )-valuation ring of I that
corresponds to (V, N ) (see (3.8)), where k ≥ 2 is an arbitrary integer. Assume that IV =
N e . Then the following hold:
(1) If e is a multiple of k, then Pk = N Uk and [(Uk )(0) : W(0) ] = k = [(Uk /Pk ) : (W/Q)].
(2) If e and k are relatively prime, then N Uk = Pk k , [(Uk )(0) : W(0) ] = k, and Uk /Pk ∼
=
W/Q.
(3) If the greatest common divisor of e and k is d, and if c ∈ N>0 is such that cd = k, then
N Uk = Pk c , [(Uk )(0) : W(0) ] = k, and [(Uk /Pk ) : (W/Q)] = d.
Proof. For (1), let (Ue , Pe ) be the Itoh (e)-valuation ring of I that corresponds to (V, N ).
Then it follows from Remark 3.3 that Pe = N Ue and [(Ue )(0) : W(0) ] = e = [(Ue /Pe ) :
(W/Q)]. Also, it is clear that W ≤ Uk ≤ Ue and that [(Uk )(0) : W(0) ] = k. Item (1) readily
follows from this, together with Remark 3.6.1 - 3.6.3.
For (2), let (Uke , Pke ) be the Itoh (ke)-valuation ring of I that corresponds to (V, N ).
Then it follows from Remark 3.3 that Pke k = N Uke , [(Uke )(0) : W(0) ] = ke, and [(Uke /Pke ) :
(W/Q)] = e. Also, it is clear that W ≤ Uk ≤ Uke and that [(Uk )(0) : W(0) ] = k. Since e and
k are relatively prime, (2) readily follows from this, together with Remark 3.6.2, 3.6.3, and
3.6.7.
17
For (3), let (Uc , Pc ) (resp., (Ud , Pd )) be the Itoh (c)-valuation (resp., (d)-valuation) ring
of I that corresponds to (V, N ). Since cd = k, it follows that W ≤ Uc ≤ Uk and W ≤ Ud
≤ Uk . Since e is a multiple of d, it follows from (1) that
(a)
QUd = Pd and [(Ud )(0) : W(0) ] = d = [(Ud /Pd ) : (W/Q)].
Since c and e are relatively prime, it follows from (2) that
QUc = Pc c and Uc /Pc ∼
= W/Q.
(b)
It follows from (a), (b), and Remark 3.6.2, 3.6.3, and 3.6.7 that (3) holds.
Proposition 3.10 gives several equivalences to Property 3.2.1 of Theorem 3.2 and also
shows that the hypothesis of Theorem 3.2 is necessary for 3.2.1.
Proposition 3.10. Let I be a regular proper ideal in a Noetherian ring R, let R = R[u, tI]
1
be the Rees ring of R with respect to I, let k ≥ 2 be an arbitrary integer, let Sk = R[u k ],
1
1
and let Tk = Sk ′ ∩ R[u k , t k ]. Then the following statements are equivalent:
1
(1) u k Tk is a radical ideal.
1
(2) The Rees integers of u k Tk are all equal to one.
1
(3) The Rees integers of (u k Sk )a are all equal to one.
(4) k is a common multiple of the Rees integers of I.
1
Proof. A primary decomposition of the regular proper principal radical ideal u k Tk , together
with Remark 2.2.4, shows that (1) ⇒ (2), and the next to last paragraph of the proof of
Theorem 3.2 shows that (2) ⇒ (1).
Also, (2) ⇔ (3), by Remark 2.2.5, since Sk ⊆ Tk ⊆ Sk ′ .
Assume that (2) holds and let (V1 ∗ , N1 ∗ ), . . . , (Vn ∗ , Nn ∗ ) be the Itoh (k)-valuation rings
1
of I, so u k Vj ∗ = Nj ∗ for j = 1, . . . , n. For j = 1, . . . , n, let uWj = Qj ej , where RV(uR[u, tI])
= {(Wj , Qj ) | j = 1, . . . , n}. Suppose that k is not a multiple of ej for some j, let d be the
greatest common divisor of k and ej , and let c ≥ 1 and h > 1 be integers such that cd =
k and hd = ej . Then it follows from Proposition 3.9(3) (with (Wj , Qj ) (resp., (Vj ∗ , Nj ∗ ))
in place of (W, Q) (resp., (Uk , Pk )) that Qj Vj ∗ = (Nj ∗ )c . Since uWj = Qj ej , it follows
1
that u k Vj ∗ = (Nj ∗ )
cej
k
= (Nj ∗ )h . However, h > 1, and this contradicts (2). Therefore the
supposition that k is not a multiple of ej leads to a contradiction, hence (2) ⇒ (4).
18
Finally, (4) ⇒ (1) by Theorem 3.2.1.
Corollary 3.11. Let I be a regular proper ideal in a Noetherian ring R, let R = R[u, tI]
1
be the Rees ring of R with respect to I, and for each integer k ≥ 2 let Sk = R[u k ] and let
1
1
Tk = Sk ′ ∩ R[u k , t k ]. Then the following statements are equivalent:
(1) The Rees integers of I are all equal to one.
1
(2) For all integers k ≥ 2, the ideal u k Tk is a radical ideal.
Proof. This follows immediately from Proposition 3.10(1) ⇔ (4).
4
A RELATED THEOREM
In this section, we first prove an expanded version of Corollary 3.5 and Theorem 3.2.6, and
we then prove a closely related and more general theorem.
The next theorem is an expanded version of Corollary 3.5 and Theorem 3.2.6. One of the
main reasons for including this theorem is to show that it displays, except for separability,
a realization (see Definition 4.3 below) of a powerful classical method. This is explained
somewhat more fully in Remark 4.4 below.
Theorem 4.1. Let I be a nonzero proper ideal in a Noetherian integral domain R, let F
be the quotient field of R, let R = R[u, tI], let (W1 , Q1 ), . . . , (Wn , Qn ) be the Rees valuation
rings of uR, let e be a positive common multiple of the Rees integers e1 , . . . , en of uR, let
W = W1 ∩ · · · ∩ Wn , let Q = Q1 ∩ · · · ∩ Qn , and for j = 1, . . . , n, let Pj = Qj ∩ W, so W is
a semi-local Dedekind domain, Q is its Jacobson radical, and the ideals P1 , . . . , Pn are the
maximal ideals in W. Then there exists an integral domain E with an ideal J such that:
(4.1.1) E = W[x]′ is a semi-local Dedekind domain that is a finite integral extension domain
1
of W, where x = u e in a fixed algebraic closure F (u)∗ of F (u).
(4.1.2) [E(0) : W(0) ] = e.
(4.1.3) J = xE is the Jacobson radical of E, E has exactly n maximal ideals M1 , . . . , Mn ,
and Mj ∩ W = Pj for j = 1, . . . , n.
(4.1.4) [(E/Mj ) : (W/Pj )] = ej for j = 1, . . . , n.
(4.1.5) uE = J e .
(4.1.6) The Rees integers of IE = uE are all equal to e.
19
(4.1.7) If e1 = · · · = en = e and if there exists b ∈ I such that bV = IV for each Rees
valuation ring (V, N ) of I, then there exists a unit θ ∈ E such that E = W[θ], so E is a
simple free integral extension domain of W of rank e.
1
1
1
(4.1.8) E = TS ′ , where T = R[u, tI, u e ]′ ∩ R[u e , t e ] and S ′ = T − ∪{q | q is a (height
1
one) associated prime ideal of u e T}.
Proof. It is shown in Corollary 3.5(1) that: W is a semi-local Dedekind domain; Q is its
Jacobson radical, and, the ideals P1 , . . . , Pn are the maximal ideals in W.
(4.1.1) follows from Corollary 3.5(1) and Corollary 3.5(3.2).
(Note: E is frequently denoted by D′ in Corollary 3.5 and its proof.)
(4.1.2) follows from Corollary 3.5(2).
(4.1.3) is proved in Corollary 3.5(3.3) and Corollary 3.5(3.2).
For (4.1.4), it is shown in Corollary 3.5(3) and 3.5(3.2) that there exist θ1 , . . . , θn ∈ E
such that E = W[x, θ1 , . . . , θn ] and Mj = (Pj , x, θ1 , . . . , θj−1 , θj+1 , . . . , θn )E.
Therefore E/Mj ∼
= W[θj ]/Pj ∼
= (Wj /Qj )[θj + Pj ], and Vj ∗ /Nj ∗ = Wj [x, θj ]/Nj ∗ ∼
=
(Wj /Qj )[θj + Pj ], so E/Mj ∼
= Vj ∗ /Nj ∗ for j = 1, . . . , n.
Also, W/Pj ∼
= Wj /Qj for j = 1, . . . , n. Therefore (4.1.4) follows from Theorem 3.2.5.
Since IE = uE, (4.1.5) follows from Corollary 3.5(3.3), and then (4.1.6) follows from
(4.1.5) and Definition 2.1.5. (4.1.7) is proved in Corollary 3.5(4).
Finally, for (4.1.8), E = W[x]′ is the intersection of the Itoh (e)-valuation rings of I,
by Corollary 3.5(3.1). Also, TS ′ is the intersection of the Itoh (e)-valuation rings of I, by
Definition 1.2, so E = TS ′ , so (4.1.8) holds.
We next consider a powerful classical theorem of Krull, and to state the theorem, we
use the following terminology of Gilmer in [3].
Definition 4.2. Let (V1 , N1 ), . . . , (Vn , Nn ) be distinct DVRs of a field F and for j =
1, . . . , n, let Kj = Vj /Nj denote the residue field of Vj . Let m be a positive integer. By
an m-consistent system for {V1 , . . . , Vn }, we mean a collection of sets S = {S1 , . . . , Sn }
satisfying the following conditions:
(1) For each j, Sj = {(Kj,i , fj,i , ej,i ) | i = 1, . . . , sj }, where Kj,i is a simple algebraic field
extension of Kj with fj,i = [Kj,i : Kj ], and sj , ej,i ∈ N>0 .
20
(2) For each j, the sum
Ps j
i=1 ej,i fj,i
= m.
Definition 4.3. The m-consistent system S as in Definition 4.2 is said to be realizable if
there exists a separable algebraic extension field L of F such that:
(1) [L : F ] = m.
(2) For j = 1, . . . , n, Vj has exactly sj extensions Vj,1 , . . . , Vj,sj to L.
(3) For j = 1, . . . , n, the residue field of Vj,i is Kj -isomorphic to Kj,i , and the ramification
index of Vj,i relative to Vj is equal to ej,i (so Nj Vj,i = Nj,i ej,i ).
If S and L are as above, we say the field L realizes S or that L is a realization of S.
Remark 4.4. (4.4.1) With Definition 4.3 in mind, and with the notation of Theorem 4.1,
the semi-local Dedekind domain E = W[x]′ (or its quotient field F (x)) in Theorem 4.1 is,
except for separability, a realization of the e-consistent system S = {S1 , . . . , Sn } for the
Rees valuation rings (W1 , Q1 ), . . . , (Wn , Qn ) of uR[u, tI]. Here, e is an arbitrary positive
common multiple of the Rees integers e1 , . . . , en of uR[u, tI] (and of I), and for j = 1, . . . , n,
Sj = {(Wj /Qj )[θj ], ej , eej )}, where θj is a root of X ej − wj (with wj playing the role of w
in (3.5.4.3) in the proof of Corollary 3.5), and θj = θj + Qj Wj [θj ].
(4.4.2) More generally, let I be a nonzero proper ideal in a Noetherian integral domain R,
let R = R[u, tI] be the Rees ring of R with respect to I, let (W1 , Q1 ), . . . , (Wn , Qn ) be the
Rees valuation rings of uR, let W = W1 ∩· · ·∩Wn , and let k ≥ 2 be an arbitrary integer. For
j = 1, . . . , n, let ej be the Rees integer of uR with respect to Wj , and let dj be the greatest
common divisor of k and ej . Also, let F be the quotient field of R, let F (u)∗ be an algebraic
1
closure of F (u), and let E be the integral closure of W[xk ], where xk = u k ∈ F (u)∗ . Then
E is a realization of the k-consistent system S′ = {S1 ′ , . . . , Sn ′ } for the valuation rings
(W1 , Q1 ), . . . , (Wn , Qn ). Here, for j = 1, . . . , n, Sj ′ = {(Wj /Qj )[θj,k ], dj , dkj )}, where θj,k is
a root of X ej,k − wj,k (with wj,k playing the roll of w in (3.5.4.3) in the proof of Corollary
3.5), and θj,k = θj,k + Qj Wj [θj,k ].
Proof. Item (4.4.1) follows immediately from Theorem 4.1 (it follows from Proposition 3.9(1)
that Qj Wj [θj ] is a prime ideal), and Item (4.4.2) follows from Proposition 3.9(3).
Theorem 4.5. (Krull [6]): Let (V1 , N1 ), . . . , (Vn , Nn ) be distinct DVRs of a field F with
Kj = Vj /Nj for j = 1, . . . , n, let m be a positive integer, and let S = {S1 , . . . , Sn } be an
21
m-consistent system for {V1 , . . . , Vn } with Sj = {(Kj,i , fj,i , ej,i ) | i = 1, . . . , sj } for j =
1, . . . , n. Then S is realizable if one of the following conditions is satisfied:
(1) sj = 1 for at least one j.
(2) F has at least one DVR V distinct from V1 , . . . , Vn .
(3) For each monic polynomial X t + a1 X t−1 + · · · + at with ai ∈ ∩nj=1 Vj = D, and for each
h ∈ N>0 there exists an irreducible separable polynomial X t + b1 X t−1 + · · · + bt ∈ D[X]
with bl − al ∈ Nj h for each l = 1, . . . , t and j = 1, . . . , n.
Observation.
(a) Condition (1) of Theorem 4.5 is a property of the m-consistent system S = {S1 , . . . , Sn }.
(b) Condition (2) of Theorem 4.5 is a property of the family of DVRs of the field F .
(c) Condition (3) of Theorem 4.5 is a property of the family (V1 , N1 ), . . . , (Vn , Nn ).
Remark 4.6. Let R be a Noetherian integral domain with quotient field F , and assume
that altitude(R) ≥ 2. Then there exist infinitely many height one prime ideals in R, so
R′ has infinitely many height one prime ideals, by the Lying-Over Theorem ([7, (10.8)]).
Therefore, since R′ is a Krull domain, by ([7, (33.10)]), there exist infinitely many distinct
DVRs with quotient field F , hence Theorem 4.5(2) is always satisfied for such fields F .
We can now state and prove the first new result in this section. It is closely related to
Theorem 4.1, and it is also considerably more general.
Theorem 4.7. Let R be a Noetherian integral domain with quotient field F , and assume
that altitude(R) ≥ 2. Let I be a nonzero proper ideal in R, let (V1 , N1 ), . . . , (Vn , Nn ) (n
≥ 2) be the Rees valuation rings of I, let D = V1 ∩ · · · ∩ Vn (so D is a Dedekind domain
with exactly n maximal ideals Mj = Nj ∩ D (j = 1, . . . , n)), and let M1 e1 · · · Mn en (=
M1 e1 ∩ · · · ∩ Mn en ) be an irredundant primary decomposition of ID (so e1 , . . . , en are the
Rees integers of I). Let m be the least common multiple of e1 , . . . , en , let dj =
m
ej
(j =
1, . . . , n), and let e = km be a positive multiple of m. Then there exists an integral domain
E with an ideal J such that:
(4.7.1) E = D[θ] is a semi-local Dedekind domain that is a simple free separable integral
extension domain of D.
(4.7.2) [E : D] = e.
22
(4.7.3) J is the Jacobson radical of E, and E has exactly k(e1 + · · · + en ) maximal ideals.
(4.7.4) [(E/Qj,i ) : (D/Mj )] = 1 for j = 1, . . . , n and for the kej associated prime ideals
Qj,i of Mj E.
(4.7.5) IE = J m .
(4.7.6) The Rees integers of IE are all equal to m.
Proof. Let S = {S1 , . . . , Sn } with Sj = {(Kj,i , 1, dj ) | i = 1, . . . , kej } (j = 1, . . . , n). Observe
that dj ·kej = km = e for j = 1, . . . , n, so S is an e-consistent system for {DM1 , . . . , DMn }, so
S is a realizable e-consistent system for {DM1 , . . . , DMn }, by Theorem 4.5(2) and Remark
4.6. Therefore the integral closure E of D in a realization L of S for {DM1 , . . . , DMn }
is a simple free separable integral extension domain of D such that [E : D] = e and E
is a Dedekind domain (by [12, Theorem 19, p. 281]). Also, since Sj = {(Kj,i , 1, dj ) |
i = 1, . . . , kej } (j = 1, . . . , n): each Vj has exactly kej extensions (Vj,i , Nj,i ) to L; E =
V1,1 ∩ · · · ∩ Vn,ken and Qj,i = Nj,i ∩ E; E/Qj,i ∼
= Vj,i /Nj,i ∼
= Vj /Nj ∼
= D/Mj (j = 1, . . . , n
and i = 1, . . . , kej ); and, Nj Vj,i = Nj,i dj (j = 1, . . . , n and i = 1, . . . , kej ).
Therefore (4.7.1) - (4.7.4) hold, and
IE =
n
Y
(Mj
j=1
where J =
Qn
j=1
Qkej
i=1 Nj,i
ej
kej
kej
n Y
n Y
Y
Y
dj ej
Nj,i )m = J m ,
( (Nj,i )) = (
E) =
j=1 i=1
j=1 i=1
is the Jacobson radical J of E. Thus IE = J e , so (4.7.5) holds,
and since E is a semi-local domain with Jacobson radical J, (4.7.6) follows immediately
from (4.7.5), Definition 2.1.5, and Remark 2.2.4.
Remark 4.8. In the proof of Proposition 4.7, there are many cases when the simpler econsistent system T = {T1 , . . . , Tn } with Tj = {(Kj,1 , kej , dj )} (j = 1, . . . , n) can be used in
place of Sj = {(Kj,i , 1, dj ) | i = 1, . . . , kej } (j = 1, . . . , n), and then the resulting realization
E of T for {DM1 , . . . , DMn } has the same number of maximal ideals as D. However, T
cannot always be used, since, for example, if D/Mj is algebraically closed and kej ≥ 2, then
there are no extension fields Kj,1 of Kj = D/Mj such that [Kj,1 : Kj ] = kej .
Remark 4.9. In the proof of Proposition 4.7, if Sj = {(Kj,i , 1, dj ) | i = 1, . . . , kej } is
replaced with Uj = {(Kj,i , 1, kdj ) | i = 1, . . . , ej } (j = 1, . . . , n), then the same conclusions
hold, but replace:
23
(a) “each Vj has exactly kej extensions (Vj,i , Nj,i ) to L” with “each Vj has exactly ej
extensions (Vj,i , Nj,i ) to L”;
(b) “the Rees integers of IE are all equal to m,” with “the Rees integers of IE are all equal
to e”;
(c) “IE = J m ” with “IE = J e ”;
(d) “Nj Vj,i = Nj,i dj (j = 1, . . . , n and i = 1, . . . , kej )” with “Nj Vj,i = Nj,i kdj (j = 1, . . . , n
and i = 1, . . . , ej )”;
Q
Q
Qke
Q
Qke
(e) “IE = nj=1 (Mj ej E) = nj=1 ( i=1j (Nj,i dj ))ej = ( nj=1 i=1j Nj,i )m ” with “IE =
Qe j
Q
Qn Qe j
Qn
kdj ej
ej
Nj,i )e ”.
)) = ( nj=1 i=1
j=1 ( i=1 (Nj,i
j=1 (Mj E) =
Our first corollary of Theorem 4.7 is a more complete and detailed version of [4, Theorem
2.8(1)].
Corollary 4.10. Let I be a nonzero proper ideal in a Noetherian domain R, let (V1 , N1 ),
. . . , (Vn , Nn ) (n ≥ 2) be the Rees valuation rings of I, let ej be the Rees integer of I with
respect to Vj (j = 1, . . . , n), and let e = km be a positive multiple of the least common
multiple m of e1 , . . . , en . Then there exists an integral domain Be such that:
(1) Be is a semi-local Dedekind domain that is a simple free separable integral extension
domain of R.
(2) [Be : R] = e.
(3) The Rees integers of IBe are all equal to m.
Proof. Let D be the intersection of the n Rees valuation rings (Vj , Nj ) of I, and let Mj =
Nj ∩ D (j = 1, . . . , n), so D is a semi-local Dedekind domain with exactly n maximal ideals
Mj and ID = M1 e1 ∩ · · · ∩ Mn en = M1 e1 · · · Mn en . Therefore by Proposition 4.7 there exists
a simple free separable extension field Le of the quotient field F of R such that: [Le : F ]
= e; the integral closure Ee of D in Le is a finite (by free separability) integral extension
domain of D and is a semi-local Dedekind domain with exactly kej maximal ideals Qj,i
lying over each Mj (j = 1, . . . , n); and, (ID)Ee = Je m , where Je is the Jacobson radical of
Ee . By separability Le = F [θ], so there exists r ∈ R such that r · θ is integral over R, so let
Be = R[r · θ]. Then Be is a simple free separable integral extension domain of R, [Be : R]
= e, and Ee is the intersection of the Rees valuation rings of IBe by Remark 2.2.6. Since
24
(IBe )Ee = (ID)Ee = Je m and since Je is the Jacobson radical of Ee , it follows that the
Rees integers of IBe are all equal to m.
Remark 4.11. (4.11.1) With the notation of Corollary 4.10, there exists an integral
domain Ce such that:
(1) Ce is a semi-local Dedekind domain that is a simple free separable integral extension
domain of R.
(2) [Ce : R] = e.
(3) The Rees integers of ICe are all equal to e.
(4.11.2) Let I be a regular proper ideal in a Noetherian ring R, let (V1 , N1 ), . . . , (Vn , Nn )
(n ≥ 2) be the Rees valuation rings of I, let ej be the Rees integer of I with respect to Vj
(j = 1, . . . , n), and let e = km be a positive multiple of the least common multiple m of
e1 , . . . , en . Assume that Rad(0R ) is prime, say Rad(0R ) = z. Then there exists a ring Be
such that:
(1) Be is a simple free integral extension ring of R.
(2) [Be : R] = e.
(3) The Rees integers of IBe are all equal to m.
(4) zBe = Rad(0Be ), so zBe is the only minimal prime ideal in Be .
(4.11.3) With the notation of (4.11.2), there exists a ring Ce such that:
(1) Ce is a simple free integral extension ring of R.
(2) [Ce : R] = e.
(3) The Rees integers of ICe are all equal to e.
(4) zCe = Rad(0Ce ), so zCe is the only minimal prime ideal in Ce .
Proof. The proof of (4.11.1) is the same as the proof of Corollary 4.10, but use Remark 4.9
in place of Proposition 4.7.
For (4.11.2), since Rad(0R ) = z is the only minimal prime ideal in R, the Rees valuation
rings of I are the Rees valuation rings of I = (I + z)/z (see Remark 2.2.3). Also, by
Corollary 4.10 there exists an integral domain Be such that:
(1’) Be is a semi-local Dedekind domain that is a simple free separable integral extension
domain of R = R/z.
(2’) [Be : R] = e.
25
(3’) The Rees integers of IBe are all equal to m.
Let fe (X) be the pre-image in R[X] of the irreducible monic polynomial fe (X) (of degree
e) in R[X] such that Be = R[X]/(fe (X)R[X]). Then Be = R[X]/((fe (X), z)R[X]) =
R[x]/(zR[x]), where x = X + (fe (X)R[X]).
Let Be = R[x]. Then, since Be = R[x]/(zR[x]) is an integral domain, it follows that
zBe is a prime ideal. Also, by hypothesis, there exists r ∈ N>0 such that z r = (0) in R, so
it follows that zBe is the only minimal prime ideal in Be , so (4) holds. Therefore the Rees
valuation rings and Rees integers of IBe are the Rees valuation rings and Rees integers of
IBe , by Remark 2.2.3, so (1) - (3) follow immediately from (1’) - (3’).
The proof of (4.11.3) is the same as the proof of (4.11.2), but use Remark 4.11.1 in place
of Corollary 4.10.
The next corollary of Theorem 4.7 extends Corollary 4.10 to Noetherian rings.
Corollary 4.12. Let I be a regular proper ideal in a Noetherian ring R, let e1 , . . . , en be
the Rees integers of I, let m be the least common multiple of e1 , . . . , en , and let e = km be
a positive multiple of m. Then there exist rings R∗ and Be such that:
(1) R ⊆ R∗ ⊆ Be and Be is a finite integral extension ring of R.
(2) IR∗ and IBe are regular proper ideals.
(3) There is a one-to-one correspondence between the minimal prime ideals w∗ in Be such
that IBe + w∗ 6= Be , the minimal prime ideals z ∗ in R∗ such that IR∗ + z ∗ 6= R∗ , and the
minimal prime ideals z in R such that I + z 6= R; namely, w∗ = z ∗ Be and z = z ∗ ∩ R.
(4) The Rees integers of IBe are all equal to e.
Proof. Let (0) = ∩{qh | h = 1, . . . , k} be an irredundant primary decomposition of the zero
ideal in R and let Rad(qh ) = zh (h = 1, . . . , k). Assume that the zh are re-ordered so that
z1 , . . . , zd1 are the minimal prime ideals z in R such that I + z 6= R and zd1+1 , . . . , zd2 are
the remaining minimal prime ideals in R (so d1 ≤ d2 ≤ k).
Rewrite ∩{qh | h = 1, . . . , k} as ∩{Zh | h = 1, . . . , d2 }, where Zh is the intersection
of all qi such that zi ⊇ zh , for h = 1, . . . , d2 . For i = 1, . . . , d2 , let Ri = R/Zi , let zi =
zi /Zi (so the unique minimal prime ideal in Ri is zi , therefore, zi = Rad(0Ri )), and let Ii
= (I + Zi )/Zi (so Ii is a regular proper ideal in Ri for i = 1, . . . , d1 and Ii = Ri for i =
d1 + 1, . . . , d2 ).
26
By Remark 2.2.3 the set of Rees valuation rings of I is the disjoint union of the sets
of Rees valuation rings of the ideals Ii (i = 1, . . . , d1 ), so the Rees integers of the Ii are
among the Rees integers of I. Hence e is a multiple of the least common multiple of the
Rees integers of the ideals Ii (i = 1, . . . , d1 ).
Therefore, by Remark 4.11.3, for i = 1, . . . , d1 , there exists a simple free integral extension ring Bi,e of Ri such that [Bi,e : Ri ] = e, the Rees integers of Ii Bi,e are all equal to e,
and zi Bi,e is the only minimal prime ideal in Bi,e .
L L
L L
Let R∗ = R1 · · · Rd2 , let I ∗ = I1 · · · Id2 so I ∗ = IR∗ , and for i = 1, . . . , d2 , let
L L
L
L
L
L
zi ∗ = R1
···
Ri−1
zi
Ri+1
· · · Rd2 . Then R∗ is a finite integral extension
ring of R, I ∗ is a regular proper ideal in R∗ , and for i = 1, . . . , d1 , the zi ∗ are the minimal
prime ideals z ∗ in R∗ such that I ∗ + z ∗ 6= R∗ . Also, zi ∗ ∩ R = zi and R∗ /zi ∗ = Ri /zi =
R/zi (i = 1, . . . , d2 ), so the Rees valuation rings of I ∗ are the Rees valuation rings of I, so
the ideals I and I ∗ have the same Rees integers.
L L
Let Be = B1,e · · · Bd2 ,e , where, for notational convenience, we let Bh,e = Rh for h =
L
L L
L L
L
d1 +1, . . . , d2 . Also, for i = 1, . . . , d2 , let wi ∗ = B1,e · · · Bi−1,e zi Bi,e Bi+1,e · · · Bd2 ,e ,
so wi ∗ ∩ R∗ = zi ∗ and wi ∗ = zi ∗ Be . Then Be is a finite R∗ -module (and is also a finite
L
L
L
L L
integral extension ring of R), IBe = I ∗ Be = IB1,e
···
IBd1 ,e
Bd1 +1,e
···
Bd2 ,e
is a regular proper ideal in Be , the wi ∗ are d2 minimal prime ideals in Be , and Be /wi ∗ =
Bi,e /(zi Bi,e ) for i = 1, . . . , d2 .
Since the ideals zi ∗ are the minimal prime ideals z ∗ in R∗ such that I ∗ + z ∗ 6= R∗ (for i
= 1, . . . , d1 ), and since wi ∗ is a minimal prime ideal in Be such that wi ∗ = zi ∗ Be , it follows
that the wi ∗ (i = 1, . . . , d1 ) are the minimal prime ideals w∗ in Be such that IBe + w∗
= I ∗ Be + w∗ 6= Be . So the set of Rees valuation rings of I ∗ Be is the disjoint union of
the sets of Rees valuation rings of the ideals (I ∗ Be + wi ∗ )/wi ∗ = (I ∗ Be + zi ∗ Be )/wi ∗ =
L
L L
L
L L
(B1,e · · · Bi−1,e (Ii + zi )Bi,e Bi+1,e · · · Bd2 ,e )/wi ∗ = (Ii Bi,e + zi Bi,e )/(zi Bi,e ).
Therefore, since zi Bi,e is the unique minimal prime ideal in Bi,e , it follows that, for i =
1, . . . , d1 , Ii Bi,e and (Ii Bi,e + zi Bi,e )/(zi Bi,e ) have the same Rees valuation rings and the
same Rees integers.
Finally, for i =1, . . . , d1 , the Rees integers of Ii Bi,e are all equal to e (by the last sentence
in the second preceding paragraph), so it follows that the Rees integers of IBe are all equal
27
to e.
To state an additional corollary, we need the following definition.
Definition 4.13. Let I and J be ideals in a ring R. Then:
(4.13.1) I and J are projectively equivalent in case there exist i, j ∈ N>0 such that
(I i )a = (J j )a (see (2.1.2)).
(4.13.2) I is projectively full in case, for each ideal J in R that is projectively equivalent
to I (see (4.13.1)), Ja = (I k )a for some k ∈ N>0 .
Remark 4.14. With Definition 4.13 in mind, it should be noted that Theorem 4.7.6 shows
that the Jacobson radical J of E is projectively equivalent to IE, and since E is a semi-local
Dedekind domain, it follows that J is a projectively full radical ideal whose Rees integers
are all equal to one.
The next corollary of Theorem 4.7 was proved in [4, Theorem 2.8(2)] in the case when
R is a Noetherian domain of altitude one.
Corollary 4.15. If R is a Noetherian ring of altitude one, then for each regular proper
ideal I in R there exists a finite integral extension ring A of R with an ideal J such that J
and IA are projectively equivalent, J is a projectively full radical ideal, and the Rees integers
of I are all equal to one.
Proof. It is shown in [4, Theorem 2.8(2)] that this result holds for nonzero proper ideals in
Noetherian domains of altitude one. So a proof similar to the proof of Corollary 4.12 shows
that it continues to hold for regular proper ideals in Noetherian rings of altitude one.
References
[1] C. Ciupercǎ, W. J. Heinzer, L. J. Ratliff, Jr., and D. E. Rush, Projectively full ideals
in Noetherian rings (II), J. Algebra 305 (2006), 974-992.
[2] O. Endler, Valuation Theory, Springer-Verlag, New York, 1972.
28
[3] R. Gilmer, Prüfer domains and rings of integer-valued polynomials, J. Algebra, 129
(1990), 502-517.
[4] W. J. Heinzer, L. J. Ratliff, Jr., and D. E. Rush, Projective equivalence of ideals in
Noetherian integral domains, J. Algebra 320 (2008), 2349-2362.
[5] S. Itoh, Integral closures of ideals generated by regular sequences, J. Algebra 117 (1988),
390-401.
[6] W. Krull, Über einen Existensatz der Bewertungstheorie, Abh. Math. Sem. Univ. Hamburg, 23 (1959), 29-55.
[7] M. Nagata, Local Rings, Interscience, John Wiley, New York, 1962.
[8] L. J. Ratliff, Jr., A Characterization of analytically unramified semi-local rings and
applications, Pacific J. Math. 27 (1968), 127 - 143.
[9] L. J. Ratliff, Jr., On prime divisors of the integral closure of a principal ideal, J. Reine
Angew. Math. 255 (1972), 210 - 220.
[10] D. Rees, A note on form rings and ideals, Mathematika 4 (1957), 51-60.
[11] I. Swanson and C. Huneke, Integral Closure of Ideals, Rings and Modules, Cambridge
Univ. Press, Cambridge, 2006.
[12] O. Zariski and P. Samuel, Commutative Algebra, Vol. 1, D. Van Nostrand, New York,
1958.
[13] O. Zariski and P. Samuel, Commutative Algebra, Vol. 2, D. Van Nostrand, New York,
1960.
Department of Mathematics, University of California, Riverside, California 92521-0135
E-mail address: [email protected]
Department of Mathematics, University of California, Riverside, California 92521-0135
E-mail address: [email protected]
29
Department of Mathematics, University of California, Riverside, California 92521-0135
E-mail address: [email protected]
30
| 0math.AC
|
Advanced Signal Processing Techniqes to Study
Normal and Epileptic EEG
Debadatta Dash1
Dept. of Electrical and Electronics Engineering
Veer Surendra Sai University of Technology (VSSUT)
Burla, India
[email protected]
Abstract—In this paper human normal and epileptic
Electroencephalogram signals are analyzed with popular and
efficient signal processing techniques like Fourier and Wavelet
transform. The delta, theta, alpha, beta and gamma sub bands of
EEG are obtained and studied for detection of seizure and
epilepsy. The extracted feature is then applied to ANN for
classification of the EEG signals.
epileptogenic focus being the hippocampal formation. Each
set contains 100 single channel EEG segments of 23.6s
duration sampled at 173.61 Hz. [6] Hence each data segments
contain 4097 samples collected at the intervals of 1/173.61th of
1s. Exemplary (Z038, S001) EEGs are plotted below.
Keywords—EEG, Epilepsy, FourierTransform, DWT, db-4
I.
INTRODUCTION
In 1875, the English physician, Richard Canton,
discovered electrical currents in the brain and in 1929 the
German psychiatrist, Hans Berger, recorded these currents,
and named them the Electro-encephalon-gram (EEG) (Berger,
1929).The Electroencephalogram is a record of time series of
evoked potentials caused by systematic neural activities in a
brain. [1]
EPILEPTIC seizures are the result of the transient and
unexpected electrical disturbance of the brain. Approximately
one in every 100 persons will experience a seizure at some
time in their life. [2] .EEG monitoring has an important
milestone provide valuable information of those candidates
who suffer from epilepsy.[3] Now a days however Magnetic
Resonance(MR) Brain Tomography(BT) are used for
diagnosis of structural disorders of brain to observe some
special illness like epilepsy. The measurement of human EEG
signals is done by recording the voltage difference on the
scalp by placing electrodes on or inside the brain as specified
by internationally recognized 10-20 electrode placement
system. [4] There are five major brain waves distinguished by
their different frequency ranges. They are: delta (δ) (3Hz and
below), theta (θ) (3.5Hz-7.5Hz), alpha (α) (7.5-13Hz), beta (β)
(14Hz and greater) and Gamma (γ) (31 – 100 Hz) waveforms.
[5]
Types of epileptic seizures and epileptic syndromes are
classified by the Commission of Classification and
Terminology of the International League against Epilepsy.
II.
Fig. 1(a). Normal EEG signal, 1(b). Epileptic EEG signal
From Fig. 1. it is seen that the epilectic signal with its higher
amplitude and distinct transients is complete different from
normal EEG.
III.
FOURIER ANALYSIS
To compare the frequency distribution of both EEG signals
we plotted log-log plot of the Periodogram or power vs.
frequency plot. Power is explained as the square of the absolute
value of Fourier Transform. In each plot so called 1/(f)α
fluctuation can be seen as the function of frequency.
DATA
We used two sets of EEG data; A/Z (Healthy Volunteer,
Eyes open) and E/S (Epileptic Subjects during seizure) which
is publicly available given by Andrzejak et al [6]. The type of
epilepsy was diagnosed as temporal lobe epilepsy with the
Fig 2. Comparison of log-scale Periodograms of both EEG signals
The peaks that indicate alpha and beta rythms are clearly
visible for normal EEG periodogram but not seen in epileptic
signal,as the epileptic EEG shows abrupt increase in frequency
at different ranges.
IV.
WAVELET ANALYSIS
The basic idea underlying wavelet analysis consists of
expressing a signal as a linear combination of a particular set
of functions (wavelet transform, WT), obtained by shifting
and dilating one single function called a mother wavelet. Since
the sampling frequency of the EEG is 173.61 Hz, according to
the Nyquist sampling theorem, the maximum useful frequency
is half of the sampling frequency or 86.81 Hz. Again to
correlate the wavelet decomposition with the frequency
changes of the physiological sub bands the EEG band is
convolved with a low pass FIR filter and band is limited in 060Hz. The basic methodology is as shown below.
EEG LPF WT
Feature Epilepsy
Detection
Extraction
The band limited EEG is decomposed through db-4
wavelet up to 4th level and the components retained are a4 (04Hz), d4 (4-8Hz), d3 (8-15Hz), d2 (15-30Hz), d1 (30-60Hz).
Reconstructions of these five components using the inverse
wavelet transform approximately correspond to the five
physiological EEG sub bands delta, theta, alpha, beta, and
gamma (Fig. 3).
The comparison of the features of both normal and epileptic
EEG signal of two exemplary data (S001 vs. Z038) is given in
the table below.
Table: 1
EEG
Signal
Normal
Epileptic
Max
Mean
Median
Mode
Std
-190
-1765
185
1027
6.816
47.1
7
187
-1
399
478.5
42.6
V.
CONCLUSION
Detecting epileptic seizures from long EEG recordings is
a very time consuming and costly task if done visually by a
trained professional. The paper proposes a technique for
detection and feature extraction of Epilepsy using Fourier,
Wavelet Transform with MATLAB. The local low frequency
components in each normal and epileptic EEG have been
clearly detected by Wavelet Transform but not by Fourier
Transform. We conclude that Wavelet Transform is wellsuited method for local-time frequency analysis of EEG
signals. Once the delta wave is identified and measured, the
features like Mean, Max, Min and median was calculated and
then ANN is used for the classification with 97% accuracy.
VI.
[1]
[2]
[3]
[4]
[5]
Fig.3. Level 4 decomposition of the band-limited EEG into five EEG sub
bands using fourth-order Daubechies wavelet (s=a4 +d4 +d3 +d2 +d1).
[6]
Detail parts (d1-d4) shows high frequency components
while a4 shows low frequency components. Original signal S
is reconstructed by adding all the components as s=a4 +d4 +d3
+d2 +d1 .The local low frequency waves concerned with
epileptic seizures in the EEG are clearly detected by DWT.
The same analysis is done for 100 sets of data of both healthy
and epileptic database and the feature is extracted from the
delta waves of all the decomposed EEG signals which helped
in classification of EEG signals into Epileptic or normal
category. The feature extracted are in the category of mean,
median, mode, min, max and standard deviation (Std) which
showed 97% accuracy for classification of the EEG signals.
Min
[7]
[8]
REFERENCES
Suhanya.S 1, Manimegalai.P2 -- (March-2012), DETECTION OF
EPILEPSY DISORDER IN EEG SIGNAL, International Journal of
Emerging trends in Engineering and Development Issue 2, Vol.2
L.D. Iasemidis, D.S. Shiau, W.Chaovalitwongse, J.C. Sackellares,
P.M. Pardalos, and J.C. Principe, "Adaptive epileptic seizure
prediction system," IEEE Transactions on Biomedical Engineering,
vol. 50(5), pp. 616–627, 2003.
Alexandros T. Tzallas, 2007-- The Use of Time-Frequency Distributions
for Epileptic SeizureDetection in EEG Recordings, IEEE
Maan M. Shaker, 2005 EEG Waves Classifier using Wavelet Transform
and Fourier Transform, International Journal of Biological
and Life Sciences 1:2
Alexandros T. Tzallas, Markos G. Tsipouras, Dimitrios G. Tsalikakis,
Evaggelos C. Karvounis, Loukas Astrakas, Spiros Konitsiotis and
Margaret Tzaphlidou-- Automated Epileptic Seizure DetectionMethods:
A Review Study.
R.G. Andrzejak, G.Widman, K. Lehnertz, C. Rieke, P. David, C.E.
Elger, "The epileptic process as nonlinear deterministic dynamics in a
stochastic environment: an evaluation on mesial temporal lobe
epilepsy," Epilepsy Res. Vol. 44, pp. 129-140, 2001.
Shaik.Jakeer Husain 1 and K.Srinivasa.Rao 2 (2012)-- Epileptic Seizures
Classification from EEG Signals using Neural NetworksInternational
Conference on Information and Network Technology, IPCSIT vol. 37
Singapore.
Chiri Yamaguchi :Fourier and Wavelet analysis of normal and epileptic
Electroencephalogram,Memoirs of Fukui University of Technology,Vol
33,part1,2003
| 5cs.CE
|
Facets, Tiers and Gems: Ontology Patterns for
Hypernormalisation
Phillip Lord 1∗and Robert Stevens2
1
arXiv:1711.07273v1 [cs.AI] 20 Nov 2017
School of Computing Science,
Newcastle University,
Newcastle-upon-Tyne
2
School of Computer Science,
University of Manchester,
Manchester
ABSTRACT
There are many methodologies and techniques for easing the
task of ontology building. Here we describe the intersection of two
of these: ontology normalisation and fully programmatic ontology
development. The first of these describes a standardized organisation
for an ontology, with singly inherited self-standing entities, and a
number of small taxonomies of refining entities. The former are
described and defined in terms of the latter and used to manage
the polyhierarchy of the self-standing entities. Fully programmatic
development is a technique where an ontology is developed using a
domain-specific language within a programming language, meaning
that as well defining ontological entities, it is possible to add arbitrary
patterns or new syntax within the same environment. We describe
how new patterns can be used to enable a new style of ontology
development that we call hypernormalisation.
1
INTRODUCTION
Building ontologies is a difficult and time-consuming business for a
number of reasons: from an abstract point-of-view knowledge about
the domain can be difficult to gather, to understand and to represent
ontologically; more, immediately, ontologies, especially those with
a complex representation, can be taxing to describe and define
consistently, to update, expand or change when that representation
needs to change.
There have been numerous attempts to simplify and clarify
this process including: the development of methodologies such as
OntoClean that defines a set of meta-properties that can inform
ontological modelling (Guarino and Welty, 2002); upper ontologies
such as DOLCE or BFO (Grenon et al., 2004) that provide a
pre-made upper classification.
Another approach that can leverage both of these techniques is
ontology normalisation (Rector, 2002). Originally intended as a
mechanism for “untangling” existing hierarchies or classifications
being reused as the basis for an ontology, it also has significant use
as a pattern for building ontologies de novo.
Broadly, a normalised ontology is defined using a skeleton that is
a strict tree (i.e. not a acyclic graph) of concepts differentiated using
an inheritance (i.e. not a partonomy) relationship. These are further
split into: a set of self-standing entities in which children are disjoint
from each other, but do not cover the parent, and partitioning or
∗ To
whom
correspondence
[email protected]
should
be
addressed:
refining concepts that form closed, covering and disjoint hierarchies.
Building an ontology in this way, allows the ontology developer to
exploit the reasoner to build a polyhierarchy by using classes that
define the self-standing entity in terms of the refining partitions.
Polyhierarchies are difficult to build manually, as human ontology
developers, no matter how good their domain knowledge, find it
hard to ensure all possible parents of an entity are taken into account.
The normalisation approach uses defined classes and reasoning to
remove this chore. Creating the tree of self-standing entities still,
however, remains as a task for the developer. The normalisation
approach can significantly increase the robustness and reduce the
work of manual maintenance (Wroe et al., 2003). In this latter form,
ontological normalisation has been widely, if implicitly, used.
While the term “ontology normalisation” has been borrowed,
somewhat metaphorically, from database engineering, the process
of building ontologies using a set of standard design patterns
has a rather more direct relationship to the software engineering
equivalent. By reusing a standard set of patterns, it is possible to
build an ontology both rapidly, and consistently. This has manifested
itself in a number of different ways, with a number of different tools,
such as TermGenie (Dietze et al., 2014), or Populus (Jupp et al.,
2011) which can generate ontologies according to a pattern.
We have previously described a fully programmatic methodology
for ontology development (Lord, 2013), using the Tawny-OWL
environment. This is built around the programming language
Clojure and enables the ontology to take advantage of all the features
of a programming language and its environment, including unit
testing (Warrender and Lord, 2015), build, evaluation and, of course,
pattern-driven development by simple use of functions (Warrender
and Lord, 2013). With respect to patterns, this environment has
several advantages. First, and unlike tools such as Populous and
OPPL (Egana Aranguren et al., 2009), patterns are developed in
the same environment and syntax as simple ontology concepts; it
is, therefore, as easy to define a pattern as it is to define a class.
Second, being based on Clojure, a language which is homoiconic
and has very little syntax of its own, it is possible to build arbitrary
syntactic constructions to represent patterns in a way that is both
convenient and attractive to the developer.
In this paper, we describe an extension of the normalisation
technique that we call hypernormalisation. This technique is
typified by the (near or complete) absence of asserted hierarchy
among the self-standing entities. We describe how this allows
construction of an exemplar ontology of amino-acids (Stevens and
Lord, 2012). We then move on to describe recent developments
1
Lord and Stevens
Top Level
Self-Standing Entities
Body Substance
Protein
Steroid
Person
Organic Ion
Doctor Role
Refining Type
Role
Care Role
Nurse Role
Value Types
Patient Role
Adult
Age Type
Sex
Child
Male
Female
Fig. 1. A normalised ontology slightly modified from Rector (2002). The graph does not necessarily reflect subsumption, see text for details.
in the Tawny-OWL environment, including the definition of two
new design patterns, the tier and the facet, and one syntactic
abstraction, the gem, can be used to enable hypernormalised
ontology development. Finally, we discuss the application of this
approach to other ontologies.
2
HYPERNORMALISATION AND AMINO ACIDS
Normalisation is a methodology that aims to disentangle an
ontological structure, in the process managing its maintainability,
utility and expressivity of the ontology generated. To achieve
this, the ontology is split into two main hierarchies: self-standing
entities and refining types, see Figure 2 for an example. The
self-standing hierarchy contains entities with a central hierarchy
or skeleton. In this part of the ontology, we would expect that
hierarchy contains levels that are not-exhaustive – that is the
children do not cover the parents, and parents are not closed to new
children. This is contrasted by the refining hierarchy that consists of
classes that are exhaustive; in many cases, children will be nonoverlapping and, therefore, disjoint. This is not to say that the
refining types hierarchy are necessarily complete: in Figure 2, for
example, the representation of Sex is too simple for many medical
uses, but might be sufficient for a customer relations system. In
general, the self-standing entities will be defined in terms of the
refining types, while polyhierarchical relationships between the
self-standing entities will be determined through use of a reasoner.
This form of ontology development is quite different from an
upper ontology and agnostic to the choice of upper ontology or
none. While Rector (2002) suggests only that self-standing entities
and refining types should be “made clear by some mechanism”; in
OWL, it could be an upper ontological term, or an annotation.
We next introduce the amino-acid, used here as an exemplar,
which defines the biological amino-acids in terms of the
physiochemical properties most relevant to their biological role.
It is a structurally interesting ontology because it is normalised,
with a clear and clean separation between the self-standing entities
and the five refining concepts. It is rather more than this, though;
2
the self-standing entities are split into only three sets: the aminoacids themselves (e.g. Alanine); a (very large) set of defined
classes describing the refined types of amino-acid (e.g. Small
Neutral Amino Acid); and, finally, the single class Amino Acid.
Or, stated alternatively, it contains no skeleton hierarchy at
all, and all relationships between the self-standing classes are
arrived at through reasoning. This is particularly relevant for
the amino acid ontology as it contains over 500 defined classes,
with subsumption relationships to the amino acids and between
themselves. Maintaining this form of ontology by hand would be
impractical.
We call this style of ontology development hypernormalised.
We believe that it is a natural extension of normalisation. Rector
notes, for example, that the choice of aspect to form the skeleton
is “to some degree arbitrary”, but that they should be rigid (after
OntoClean (Guarino and Welty, 2002)) and pragmatically stable
(i.e. unlikely to change during the evolution of the ontology). Both
of these are, however, true for all the refining concepts in the aminoacids. In short, not only is the choice of skeleton arbitrary it is
actually unnecessary and brings no further utility to the ontology
than that which can be achieved by use of reasoning.
We note that the distinction between normalisation and
hypernormalisation is not absolute, but one of degree; we are simply
describing the tendency toward an ontology with an flat asserted
hierarchy.
Having introduced the notions of a hypernormalised ontology, we
next consider a set of new patterns in Tawny-OWL that enable this
style of ontology development.
3
PATTERNISING AND TAWNY-OWL
The Tawny-OWL environment (Lord, 2013) and its ability to
support patterns (Warrender and Lord, 2013) has been described
elsewhere in detail; here, we provide a quick overview, so that the
rest of the paper is clear. Tawny-OWL is implemented as a DSL
(domain-specific language) in Clojure, which is a Lisp-like language
implemented in Java, and running on the Java virtual machine.
Facets, Tiers and Gems
Top Level
Self-Standing Entities
Defined Class
Small AA
S. Neutral AA
Amino Acid
S. N. Aliphatic AA
Alanine
Arginine
500 more
Refining Type
18 more
Size
Charge
Tiny
Small
Large
Hydro.
Polarity
SideChain
Fig. 2. A hypernormalised ontology representing the amino-acids using the same terminology as Figure 2. Some labels have been abbreviated.
Tawny-OWL itself wraps the OWL API (Horridge and Bechhofer,
2011); this is the same library that underpins Protg, and from it,
Tawny-OWL gains much of its functionality. Simple sections of the
ontology can be generated using a syntax based on a “lispified”
version of Manchester OWL Notation; for example, the following
code:
(defclass A
:super B)
This declares a new class A that has the pre-existing class B
as a superclass 1 which in Manchester OWL notation would be
expressed as:
Class: o:A
SubClassOf:
o:B
This code is entirely valid Clojure and can be evaluated in any
Clojure environment, such as CIDER/Emacs or Cursive/IntelliJ. It
is also possible to define new patterns: for example the following
pattern definition:
(defn some-only [property & clazzes]
(list (some property clazzes)
(only property (or clazzes))))
defines the some-only pattern which generates a set of
existential restrictions and one universal with the union of the
existential fillers as its filler, which implements the ontological
closure pattern. This is a function definition in Clojure terms: defn
introduces the function, property & clazzes is the argument
list, some, only and or are functions provided by Tawny-OWL
and list returns, prosaically, a list2 . Critically, it is possible to
define this pattern in the same environment, or side-by-side in the
same file as a simple class definition; with Tawny-OWL it is as easy
to define a class, as to define and use a new pattern. Ontologies
such as the Karyotype ontology make extensive use of this facility
moving freely between ontology and pattern definitions, as well as
literal data structures, utility functions and unit tests (Warrender and
Lord, 2013).
Tawny-OWL is now a mature and used software product; the first
alpha release of Tawny-OWL was in Nov 2012, first full release,
Nov 2013, followed by four point releases to 2016. This paper
describes mostly the upcoming v2.0 release, although some of the
features described were available in earlier versions.
4
THE VALUE PARTITION
A common pattern for building a normalised ontology is called the
value partition. This pattern (Rector, 2005) addresses the problem
of the ontological modelling of a continuous range. For example, in
modelling the amino-acids, we can consider the concept of Size;
this could be described directly using the molecular weight of the
amino-acid. However, for the purpose of the amino-acids, it is both
easy and general practice to split size into three categories: tiny,
small and large. In Tawny-OWL, this can be achieved straightforwardly using the defpartition function3 .
(defpartition Size
[Tiny Small Large]
:domain AminoAcid
:super PhysioChemicalProperty)
Axiomatically, this expands into: a class Size; three subclasses,
Tiny, Small and Large; and, a property hasSize. The
3
1
See Lord (2014) an explanation of why :super is used rather than
:subclass.
2 The function shown here is a slightly simplified version of one provided
in Tawny-OWL.
For those with knowledge of Lisp, this is actually a macro; the main
implementation is in the value-partition function. Tawny-OWL
provides support for implementing syntactic macros whose function is
simply to allow the use of bare symbols. For those without knowledge of
Lisp, the distinction is not important!
3
Lord and Stevens
property is functional, has range of Size and domain of
AminoAcid. Expanded, this would be expressed as follows4 :
Class: o:Large
SubClassOf:
o:Size
Class: o:Size
EquivalentTo:
o:Large or o:Small or o:Tiny
SubClassOf:
o:PhysioChemicalProperty
Class: o:Small
SubClassOf:
o:Size
Class: o:Tiny
SubClassOf:
o:Size
DisjointClasses:
o:Large,o:Small,o:Tiny
The subclasses are disjoint and cover the parent. Following the
terminology from Rector (2002), the value partition is useful for
defining partitioning or refining concepts.
5
THE TIER
The value partition is a pattern aimed at a specific purpose –
segmenting a continuous range. In practice, though, we have found
that the axiomatization of this pattern is more generally useful. For
example, considering the amino-acid ontology, it is natural to model
the chemistry of the side-chain as such:
(defpartition SideChainStructure
[Aromatic Aliphatic]
:domain AminoAcid
:super PhysicoChemicalProperty)
While this is intuitive, ontologically, SideChainStructure
is actually of a very different form from Size, as it does not reflect
a spectrum. Either the side-chain contains a benzene ring, making
it aromatic, or it does not. This form of partition was also noted
in Rector (2002) which includes the classes Male and Female
which is not a spectrum, at least in this simplified representation.
We introduce here, therefore, the more general notion of the tier:
a small set of concepts in a one-deep hierarchy. The tier function
supports a range of options:
(deftier Charge
[Positive Neutral Negative]
:domain AminoAcid
:super PhysioChemicalProperty
:suffix true)
The use of :suffix true causes a simple change to the
naming of the entities: Positive will become PositiveCharge
which would be expanded as follows:
Class: o:PositiveCharge
SubClassOf:
o:Charge
Other names are modified equivalently. By default, this will
manifest both when referring to the class in the Tawny-OWL
environment, in the IRI of the concept when serialized as OWL, and
in the value of an annotation on the concepts5 . In addition to naming,
it is also possible to optionalise: whether or not the subclasses are
disjoint, covering, whether the property is functional or whether it
is created at all.
The tier is a more general pattern than the value-partition; in fact,
in the current version of Tawny-OWL, the latter is defined in terms
of the former.
6
Both the value partition and tier introduce a new object property
named after the tier, and with a range limited to the classes defined
within the tier. The converse is also true; where we use one of the
tier classes, such as PositiveCharge it is most likely that we
wish to use it with the hasCharge property defined as part. Taken
together, we describe the combination of classes and a property as a
facet. Facets are a well known technique, first proposed in a library
classification (the Colon Classification (Ranganathan, 1933), named
after the use of “:” as a separator). They are now common-place as
seen with facetted browsers used by many websites for navigation
of complex product catalogues.
Tawny-OWL provides explicit support for facets, allowing the
association of a property and a set of classes, as demonstrated by
the following code:
(as-facet
hasCharge
Positive Neutral Negative)
The practical implication of this is that we can now use the
facet function to return an existential restriction providing just a
class. We can express this programmatically; for example, we might
use the assert function provided by Clojure’s unit test framework.
(assert
(= (some hasCharge Positive)
(facet Positive)))
By itself, this ability is only slightly more succinct. However,
when used with multiple facetted classes, the advantages become
considerably clearer, as can be shown by the following assertion.
(assert
(= (list (some hasCharge
5
4
4
Tawny-OWL also adds annotations which have been elided
THE FACET
The duplication between the annotation and the IRI fragment is there
because IRI schemes such as numeric style OBO IDs; annotations have been
elided for brevity
Facets, Tiers and Gems
Neutral)
(some hasHydrophobicity
Hydrophobic)
(some hasPolarity
NonPolar)
(some hasSideChainStructure
Aliphatic)
(some hasSize
Tiny))
(facet Neutral Hydrophobic NonPolar
Aliphatic Tiny)))
In addition to succinctness, this pattern also reduces the risk of
errors; a class such as Tiny will always be used with its correct
property. Without the use of facets, the ontology developer must
achieve this by hand. It would also be possible to detect the error
using reasoning, although this will only succeed if appropriate range
and disjoint restrictions are in the ontology. The defpartition
and deftier functions, of course, both add these range and
disjoint restrictions and declare their classes as facets of their
properties.
7
THE GEM
Finally, we define the gem that provides a syntactic abstraction for
a class composed entirely or mainly from facets. Following the
terminology from Rector (2002), this abstraction would be useful
mostly for self-standing concepts. For example, we could define the
amino acid alanine using the following defgem statement.
(defgem Alanine
:comment "An amino acid with a single
methyl group as a side-chain."
:facet Neutral Hydrophobic NonPolar
Aliphatic Tiny)
The other amino-acids can be likewise defined as a series of
gems. In fact, the amino acids are so regular, all having the same
five facets, that we use a further syntactic abstract specific to
the amino-acid ontology – a form of pattern that we describe as
localized (Warrender, 2015). The gem represents generalised syntax
useful for developing any ontology.
8
ON ANNOTATION
We have previously discussed the relationship between a design
methodology such as normalisation and the use of an upper
ontology. The Tawny-OWL patterns described here are all
orthogonal and agnostic to the choice of an upper ontology or to
none. They do not place their entities in any particular part of the
class hierarchy nor define classes outside of those required for the
domain ontology, although they could be easily extended to do so
should the ontology developer require.
However, we agree with Rector (2002) that the use of patterns
should “made clear” and be explicit within the ontology. For
this reason, all of the patterns described here also make use of
annotations, using annotation properties defined using its own
internal annotation ontology. For example, all entities generated
as a result of a pattern such as deftier are explicitly annotated
as such. This means that the use of these patterns is (informally)
explicit in the OWL serialization. Tawny-OWL actually uses
these annotations internally, for example, to enable the facet
functionality by providing a relationship between the classes and
the appropriate object property. This is a strictly an implementation
detail and could have been achieved without annotations; however,
we believe that it shows the value of having this knowledge explicit
in OWL.
9
DISCUSSION
In this paper, we describe how we have used Tawny-OWL to provide
higher-level patterns which can be applied to ontology development.
The patterns provide both functionality and syntactic abstraction
over the underlying OWL implementation. In the process, they
enable the easy and accurate construction of ontologies.
More specifically, we demonstrate two new patterns: the tier and
the facet. The tier is an extension of the existing value partition
pattern and can be used for the generation of many small hierarchies
that can be used as refining properties. The facet borrows from the
library sciences notion of a facetted classification, and is used to
associate a set of classes with a specific set of values. This form of
classification is very common in the web; the majority of web stores,
for example, offer facetted browsing, often with the facets changing
for different subsections of the catalogue.
Taken together, these two patterns enable a new form of ontology
development, hypernormalisation, which is an extreme form of
normalisation. In this form of normalisation, we do away with
the creation of a tree of self-standing entities and instead rely
on the reasoner to build all the hierarchy. As well as making
the ontologist’s task easier, it makes the characteristic that would
have been used to create the tree of self-standing entities explicit
in the form of a refining characteristic. Here, we have described
the application of this methodology to the exemplar amino-acid
ontology. Of course, it is dangerous to extrapolate to generality from
an exemplar, but we have also started to apply hypernormalisation
to ontologies of other, more real, domains including clouds (in
the meterological sense), cell lines and a reworking of the Gene
Ontology. The tier has been made generic; it does not require, for
example, that all refining types are closed (i.e. all possibilities are
known in advance) nor disjoint.
Clearly, not all forms of ontology will naturally be represented
in a hypernormalised form. For example, the Karyotype
ontology (Warrender and Lord, 2013) is far from this form; here, we
define the self-standing concepts and then use reasoning over a set of
defined classes which effectively operate as facets (Warrender and
Lord, 2015). However, the popularity of the facetted browsers shows
that is possible to use this form of classification in many areas. We
believe that the introduction of the concept of hypernormalisation
and the implementation of it in Tawny-OWL could have significant
implications for the future development of ontologies.
REFERENCES
Dietze, H., Berardini, T. Z., Foulger, R. E., Hill, D. P., Lomax, J.,
OsumiSutherland, D., Roncaglia, P., and Mungall, C. J. (2014).
Termgenie - a web application for pattern-based ontology class
generation. Journal of Biomedical Semantics, 5(1), 48.
Egana Aranguren, M., Stevens, R., and Antezana, E. (2009).
Transforming the axiomisation of ontologies: The ontology preprocessor language. Nature Precedings.
5
Lord and Stevens
Grenon, P., Smith, B., and Goldberg, L. (2004). Biodynamic
ontology: applying BFO in the biomedical domain. Stud Health
Technol Inform, 102, 20–38.
Guarino, N. and Welty, C. (2002). Evaluating ontological decisions
with ontoclean. Commun. ACM, 45(2), 61–65.
Horridge, M. and Bechhofer, S. (2011). The OWL API: A Java API
for OWL Ontologies. Semantic Web Journal, 2.
Jupp, S., Horridge, M., Iannone, L., Klein, J., Owen, S., Schanstra,
J., Wolstencroft, K., and Stevens, R. (2011). Populous: a tool for
building owl ontologies from templates. BMC Bioinformatics,
13(Suppl 1), S5.
Lord, P. (2013). The Semantic Web takes Wing: Programming
Ontologies with Tawny-OWL. OWLED 2013.
Lord, P. (2014). Manchester syntax is a bit backward. http:
//www.russet.org.uk/blog/2985.
Ranganathan, S. (1933). Colon Classification.
Rector, A. (2005). Representing specified values in owl: “value
partitions” and “value sets”. W3C Working Group Note.
Rector, A. L. (2002). Normalisation of ontology implementations:
Towards modularity, re-use, and maintainability. Proceedings
Workshop on Ontologies for Multiagent Systems (OMAS) in
6
conjunction with European Knowledge Acquisition Workshops.
Siguenza, Spain.
Stevens, R. and Lord, P. (2012). Semantic publishing of knowledge
about amino acids.
http://ceur-ws.org/Vol-903/
paper-06.pdf.
Warrender, J. (2015). The Consistent Representation of Scientific
Knowledge: Investigations into the Ontology of Karyotypes and
Mitochondria. Ph.D. thesis, School of Computing Science,
Newcastle University.
Warrender, J. and Lord, P. (2013). A pattern-driven approach to
biomedical ontology engineering. SWAT4LS 2013.
Warrender, J. D. and Lord, P. (2013). The Karyotype Ontology:
a computational representation for human cytogenetic patterns.
Bio-Ontologies 2013.
Warrender, J. D. and Lord, P. (2015). How, What and Why to test
an ontology.
Wroe, C., Stevens, R., Goble, C., and Ashburner, M. (2003).
A methodology to migrate the gene ontology to a description
logic environment using daml+oil.
Pacific Symposium on
Biocomputing.
| 2cs.AI
|
1
Who Killed Albert Einstein?
From Open Data to Murder Mystery Games
arXiv:1802.05219v1 [cs.AI] 14 Feb 2018
Gabriella A. B. Barros1 , Michael Cerny Green1 , Antonios Liapis2 and Julian Togelius1
1: Tandon School of Engineering, New York University, New York, USA
2: Institute of Digital Games, University of Malta, Msida, Malta
[email protected], [email protected], [email protected], [email protected]
Abstract—This paper presents a framework for generating
adventure games from open data. Focusing on the murder
mystery type of adventure games, the generator is able to
transform open data from Wikipedia articles, OpenStreetMap
and images from Wikimedia Commons into WikiMysteries.
Every WikiMystery game revolves around the murder of a person
with a Wikipedia article, and populates the game with suspects
who must be arrested by the player if guilty of the murder or
absolved if innocent. Starting from only one person as the victim,
an extensive generative pipeline finds suspects, their alibis, and
paths connecting them from open data, transforms open data
into cities, buildings, non-player characters, locks and keys and
dialog options. The paper describes in detail each generative
step, provides a specific playthrough of one WikiMystery where
Albert Einstein is murdered, and evaluates the outcomes of games
generated for the 100 most influential people of the 20th century.
Index Terms—Data Games, Open Data, Murder Mystery,
Adventure Games, Data Adventures, Game Generation
I. I NTRODUCTION
Games that cast the player in the role of a detective, where
the gameplay and main challenge revolve around solving a
crime or mystery, have been popular for many decades. Some
games, such as Where in the World is Carmen Sandiego?
(Brøderbund Software 1985), task the player with finding a
fugitive criminal. Other games, such as Indiana Jones and
the Fate of Atlantis (LucasArts 1992) or the series of Tomb
Raider (Eidos 1996) and Uncharted (Naughty Dog 2007), see
the player embark on an adventure to solve ancient mysteries
in the face of opposition from shadowy goons. It is common
for these games to feature frequent in-game travel to exotic
locales around the world to interact with colorful people and
to gather clues, solve puzzles and overcome resistance.
These games often make heavy use of real-world locations,
stories, items and characters to build their narrative. Authoring
such games is complex, time-consuming and requires considerable skill. However, the fact that these games depend on
so much real-world information, and that such information is
freely available in structured or semi-structured form from resources such as Wikipedia and OpenStreetMap, suggests that it
would be possible to somehow automatically incorporate realworld information in these games. Furthermore, murder mysteries and similar adventure games are often highly structured,
suggesting the possibility of generating the game itself. But
could we do this in practical manner, and if so, how? While
research projects in AI-based game generation such as Gameo-Matic [1] and A Rogue Dream [2] create simple arcade-like
and rogue-like games respectively, a murder mystery requires a
much larger volume and variety of content such as locations,
people, dialog and clues for solving the mystery. Moreover,
consistency of the content is necessary, both internally within
the game narrative and externally as fidelity with the real
world. The contribution of a generator of murder mysteries
from open data is three-fold: (a) it explores how disparate data
can be connected together to create and represent plot-lines,
(b) it identifies a design “formula” and structure for murder
mysteries and their constituent elements (such as dialog) which
can be used to generate mystery games from any Wikipedia
entry, (c) it tests the limits of autonomous game generation
and issues that arise from absurd or incomplete source data
(or from their algorithmic combination).
This paper presents WikiMystery, a framework for generating complete, playable point-and-click adventure games with
minimal human input: in this case, the name of a person who
has a Wikipedia page. The WikiMystery generative system
featured in this paper builds upon earlier work [3], extending it
significantly with a description of the full generative pipeline,
a more sophisticated dialog system and a broader evaluation
of nearly 100 games for solving the murder of the most
influential people of the 20th century. The WikiMystery game
generator is built on the previous project Data Adventures [4]
and reuses much of that technology to discover paths between
victim and suspects of the murder. The current framework
however offers a much more engaging, coherent and complete
experience with a clear goal to arrest the culprit of a murder.
This is facilitated by extensive story branching towards several
suspects, enhanced ludic elements as game objects that unlock
certain locations, and enriched dialog elements that allow NonPlayer Characters (NPCs) to share facts both about the mystery
and about themselves (based on open data).
The paper starts with a brief survey in Section II on game,
plot and dialog generation. The paper provides an overview of
the generative pipeline of WikiMystery in Section III and the
specifics of culprit and evidence selection in Section IV-A,
path generation in Section IV-B, location, NPC and item
generation in Section V and finally NPC dialog generation
in Section VI. To assess the generated games, a sample
playthrough is described in Section VII, while Section VIII
analyzes the games created for murders of the 100 most
influential people of the 20th century. The paper leads to a
2
discussion in Section IX and concludes with Section X.
II. BACKGROUND
The WikiMystery system is a framework for transforming
open data into adventure games. This section discusses the
domains of data games, game plot, and dialog generation.
A. Data Games
In a world of ever-more ubiquitous technology, the amount
of data we consume daily is rapidly increasing. One data
category that is growing exponentially is open data, i.e. information that can be freely used, re-used, and redistributed
by anyone. Creating games out of data can be seen as a form of
visualization, where instead of using charts or figures to make
the information easier to grasp, one creates playable media.
Data games use real world information (such as open data)
to automatically generate game content [5]. Players interact
with this content during gameplay, and more often than not
must learn how to understand the data in order to play the
game well. Typically, to use data as game content one must
select what parts of the data are useful for content generation,
and structurally transform it into applicable game content.
Some example data games are discussed below.
Open Trumps [6] is a data game where the cards’ content
is based entirely on published governmental data. Its generator creates a balanced Top Trumps deck using evolutionary
algorithms. While it is not required to learn the data, it helps
when playing Open Trumps. In MuseumVILLE1 , content is
selected from Europeana2 , and the user, playing as a museum
curator, must theme their museum based of their interests.
BarChartBall [7] is a physics game that uses UK census
data to transform the playable level. It is necessary to infer
how the data would affect the playfield, which is modified
based upon how high or low the selected attribute was. A
Rogue Dream [2] uses the auto-complete results of Google
queries (using templates) to choose names for player abilities,
enemies and healing items; these names are then given a visual
found via Google image search. Finally, geographical data
from OpenStreetMap have been used to generate maps and
players’ initial positions for FreeCiv 3 [8].
B. Story, Quest and Dialog Generation in Games
Research on story generation tends to focus on the textual
form. In BRUTUS [9], the generator creates dark stories
about characters with backgrounds and narratives. In MINSTREL [10], the generator uses author-level problem solving
to write short stories based on King Arthur and his knights.
Bardiche [11] acts as a collaborative tool to create good
stories based on user input, improvising based on user input.
Creating games from stories, or stories from games, is part
of the broader subject of recontextualizing data from one
medium into another. Examples include the transformation of
1 MuseumVILLE:
(https://github.com/bogusjourney/museumville/)
is a portal for accessing digitised cultural heritage material,
such as paintings and books, from more than 2,000 institutions across Europe.
3 FreCiv is an open souce version of Civilization. (http://www.freeciv.org/)
levels in Sonancia [12] and text in Audio Metaphor [13] into
soundscapes, or news articles into games in ANGELINA [14].
In adventure games, stories and quests are crucial for
progression and this needs to be taken into account when
generating them. The game Charbitat [15] procedurally generates the environment as the player explores it, but it lacks
a sense of progression. In order to anchor the player, a
quest generator was introduced to Charbitat that uses key-lock
mechanisms to advance in the game world [16]. Similarly,
a two-tiered procedural generator was built for Mystery of
Solaris [17], constructing maps and missions via two separate
grammars. First, a mission graph is constructed containing
all quest information. The mission graph is then taken as
input to the map generator, which builds a map around the
mission. Symon [18], a point-and-click adventure game, uses
procedural content generation to create meaningful puzzles.
The generation in Symon was expanded into the Puzzle Dice
System [19] for generating puzzles in adventure games.
Since murder mystery games largely rely on interaction
with non-player characters, the quality of NPC dialog is an
important factor to gameplay experience. Dialog generation
is hardly a new topic of research, going back all the way to
ELIZA [20]. Another example is the Text2dialog system [21]
which transforms monological text into dialog, and has agents
act out the dialog. It uses textual coherence relationships to
map text to question-answer pairs, and is able to create fairly
believable dialog. However, we are less interested in creating
realistic sequences of chat (and responding accommodatingly
to any player request), but rather in driving NPC interactions
towards a specific direction, i.e. providing clues to the player.
Interactive storytelling has studied dialog generation and delivery extensively, e.g. combining implicit forms of character
expression, overall narrative goals, and emotional relationships
between characters to generate realistic dialog [22].
III. OVERVIEW OF THE G AME AND THE G ENERATOR
WikiMystery is a data-based procedurally generated pointand-click adventure game. It uses data from Wikipedia, OpenStreetMap4 and Wikimedia Commons to automatically create
different game content, from plot progressing to images. It was
strongly inspired by classic adventure games such as Where in
the World is Carmen Sandiego? (Brøderbund Software 1985).
Fernández-Vara and Osterweil [23] describe some of the main
characteristics of an adventure game: gameplay driven by
the story, puzzle-solving core mechanics, interaction with the
game world through object manipulation, and an in-world
player-controlled character motivated to explore and interact
with her surroundings.
In the game, the player assumes the role of a detective trying
to solve a murder case. The victim is the central point of the
story, and suspects are based on people related —somehow—
to her. We use Wikipedia to identify possible suspects, out of
which five are selected. The game plot is tree-structured: the
victim is the root, each suspect is a leaf, and the path between
2 Europeana
4 OpenStreetMap is a open source project that attempts at mapping the world
(https://www.openstreetmap.org/).
3
Selecting
Victim
DBpedia
WikiMedia
Commons
Open
Street
Maps
Allen
Shenstone
Rosa
Beddington
Evolving
Set of Suspects
Generating
Paths to Suspects
Generating
Items
Generating Cities
& Buildings
Generating
NPCs
Adding Blocks
& Solutions
Adding
Dialogue
received Royal
Society award
David
MacKay
received Royal
Society award
died in Princeton,
New Jersey
Albert
Einstein
field of physics
born in the
German Empire
Jakob
Meisenheimer
William
Rankine
Playable
Adventure Game
Fig. 1: Flowchart of WikiMystery and its open data sources.
them is a representation of the hyperlinks between the victim’s
and each suspect’s Wikipedia articles.
Initially, the only location available is the victim’s house,
where the player can talk to people related to the victim. The
player also becomes aware of who the five suspects for the
murder are. As they interact with people inside the house, new
locations, objects and NPCs become available. As the player
explores and interacts with the world, they collect information
about suspects’ characteristics (e.g. year of death, occupation
etc.). Every suspect except the culprit has a value for one
characteristic (e.g. “died in 1980” for suspect 1 or “birth place
in Cleveland” for suspect 2) which acts as an evidence of
innocence (identified in the dialog with a note “... couldn’t
have done it.”). The player receives no such information about
the culprit; the culprit also does not share the same value
in a characteristic which is evidence of innocence for other
suspects (i.e. the culprit did not die in 1980 and was not born
in Cleveland). The game ends when the player issues an arrest
warrant, identifying the culprit and specifying the values acting
as evidence of innocence of the other suspects. If the player
correctly finds the culprit and provides the correct evidence
for the other suspects, the game is won. If the player does not
specify the right culprit, or if the evidence for the innocence
of any other suspect is incorrect, then the game is lost.
The game generation involves several steps, as shown in
Figure 1, the first of which is selecting the victim. From
the victim, the system uses DBpedia5 to find a set of five
suspects via artificial evolution presented in Section IV-A and
to generate paths between each suspect and the victim via
a constructive algorithm presented in Section IV-B. Once all
paths are generated, the system creates locations, items and
NPCs through constructive processes covered in Section V.
Finally, it generates puzzles for accessing locations and dialog
options for learning about clues or general information from
NPCs; we discuss the latter in Section VI.
IV. C RAWLING DB PEDIA
The plot of WikiMystery is created from a series of hyperlinks in Wikipedia, generated using several consecutive queries
to DBpedia. Once a victim has been introduced, the system
tries to find suspects and pinpoint a culprit among them. Then,
it searches for paths between the victim and the suspects.
5 DBpedia is a project which extracts information from Wikipedia in a
structured manner (http://dbpedia.org/).
Fig. 2: Selecting suspects from DBpedia to use in the game
and finding related between victim and suspects. Initially, the
system has a single node: the victim (black node). Suspects
related to the victim are selected with a genetic algorithm
(white nodes), and paths between the victim and suspects are
created from DBpedia (see Fig. 3). All suspects share direct
connections with the victim, shown on the arrows.
A. Finding Suspects and Culprit
Selection of a set of suspects involves identifying who is
related to the victim, and out of those, which subset is the
most interesting. Given a Wikipedia article about a person, the
system queries DBpedia to find anyone who has something in
common with the victim. It can be as common as living in
the same place, or as specific as being in the same band. This
list is our pool of suspects. For each one, we query DBpedia
to find everything known about them. At this point, we have
a list of suspects, each containing a list of characteristics.
Each characteristic can have multiple values. For example,
a suspect could be “Albert Einstein”, who would have the
characteristic “Field” with values “Physics” and “Philosophy”.
Figure 2 shows a simplified selection of suspects. From a
victim (black node), the system finds suspects (white nodes),
who are related somehow to the victim (arrows). The system
must also find a set of characteristics that can single out the
culprit among suspects. A characteristic and a value of that
characteristic, together, form an evidence of innocence, which
is used to identify the culprit and issue an arrest warrant.
The list of possible subjects, characteristics and values
can be very large, at times. Characteristics may have multiple values available, of which we will only use one per
characteristic, and suspects can have multiple characteristics.
For example, as of the publication date of this paper, there
are 15,300,451 distinct people related to Albert Einstein in
some way, constituting his possible suspect pool. Each one
has, on average, at least five characteristics, which may or
may not have multiple values. Selecting five suspects, four
characteristics and their values is therefore challenging. To
select which subset of this list is interesting, we turn to an
µ + λ evolutionary algorithm. Our goal is to have a finite set
of suspects n (typically, 5) and a finite set of characteristics
n − 1, such that we can pair each characteristic to a person.
The leftover person is the culprit, and the characteristics are
evidence of innocence for the n − 1 suspects. This allows the
player to eliminate innocent suspects by finding the clue paired
to them. The remaining suspect who does not have the same
value with any evidence of innocence must be the killer.
4
Our fitness function evaluates solvability and diversity.
Solvability favors complete solutions, where the player can
identify the culprit by excluding the n − 1 characteristics she
knows the killer does not have. Depth-First Search is applied:
for every chromosome, it marks one of the suspects in the
chromosome as the killer. The search states are characteristics
in the chromosome, and they are paired with one suspect once
they are visited. Valid states have three properties: (a) the killer
has at least one value for the characteristic; (b) one or more
suspects have value(s) for the characteristic; (c) at least one
suspect has one value different from the killer. The algorithm
tries to pair each suspect to one characteristic if the suspect
has at least one value that is different from that of the killer
(for that specific characteristic). If it cannot match a pair, it
backtracks and tries a different suspect. No characteristic can
be paired with more than one suspect, and vice-versa. The
optimal solution is a leaf where all characteristics are paired
successfully to suspects; the fitness is the depth of the leaf.
Diversity evaluates how different the characteristics and
values are. The game only outputs one value per characteristic/suspect pair, so it is necessary to optimize which
value to use. For example, in a game with 5 suspects and 4
characteristics, a solution with 20 values, one per pair, is better
than one where one suspect has no value for any characteristic
but the one matched to him. Additionally, a solution where all
suspects have the same job and live in the same city is less
diverse than a solution where they all have different jobs and
live in different cities. Even though they still use the same
characteristics (“job” and “residence”), the second one has
more diversity of values. The actual fitness value is given by:
Vi
P
X
X
Qi × −
fD =
pij × log2 (pij )
(1)
i=0
the root of the tree is the crime scene, and each branch leads
to a possible suspect. This tree represents most plot points the
player is able to unravel in the game in order to move the plot
forward, such as locations, NPCs and clues. The only clues
not present in this tree are those of evidence of innocence.
For each suspect, the system queries DBpedia multiple
times, searching all possible paths between the victim and said
suspect. It rates each path based on how diverse it is, i.e. the
type of articles and links in the path. For example, a path that
only has articles about locations is less diverse than one with
an even number of articles about people and locations. This
process can be computationally expensive, since it is necessary
to create one query per node in the path, and per direction of
the edge. In practice, using paths longer than 5 nodes has
proven to be too time consuming [25]. To bypass this, we
divided the search into two steps. The first finds a path of
length no longer than 5 nodes, as described above, which we
call the major path. Then, for each consecutive pair of nodes in
the major path, the system searches for a minor path between
those nodes. The minor path replaces the edge between the
two nodes in the major path. Figure 3 shows an example path
between the victim (Albert Einstein) and a suspect (William
J. M. Rankine), identifying major and minor paths.
The system measures path quality based on its length and
uniqueness. Longer paths are preferred because they will
extend the game: each node will be transformed into a city,
NPC or item in the story. Uniqueness is calculated as the
entropy of each node/edge in the path, compared to all nodes
and edges found in all paths in that particular search. Thus
a path where the type of edge is not found in other possible
paths is better than typical edges (e.g. for scientists it is typical
to have edges of type “influenced by” or “influenced”).
j=0
where P is the number of characteristics; Vi is the number of
values for characteristic i; Qi is the number of people with
characteristic i. We multiply with Qi to reward more suspects
sharing characteristic i. pij is calculated as the number of
people that have value j in characteristic i divided by Qi .
The system uses cascading elitism [24] over a population of
100 individuals for 500 generations, with a mutation chance
of 20%. Cascading elitism uses both fitness functions: it
sorts the population using the solvability fitness, removes the
worse 50% individuals, and then sorts the remaining using
the diversity fitness. The highest 25% of the population is
duplicated and mutated until the new population is filled. It
is far more important that games are solvable for playability’s
sake (rather than diverse); solvability is applied first during
cascading elitism so that it introduces a stronger genetic bias.
B. Finding Paths to Suspects
Once the system has a victim, suspects and clues as evidence
of innocence, it weaves them into a plot by searching DBpedia
for a path of hyperlinks between the victim and each suspect.
A path consists of nodes (Wikipedia articles) and edges (links
between them). The set of all paths can be seen as a tree if we
merge the initial node in the paths (i.e. the victim). Therefore
V. E NRICHING THE DATA
The system transforms the set of paths (tree) obtained from
Wikipedia into gameplay objects that the player can interact
with. Each node in the tree becomes a location, an item or
a NPC. To do so, it creates all necessary game objects, then
generates dialogs and links between them, verifies if all objects
appear in the correct order, and add puzzles.
Nodes in the tree can be roughly categorized into places
(e.g. “London” or “Canada”), people (e.g. “Albert Einstein”)
and everything else (e.g. “Mathematicians of the 20th Century”). The system begins by transforming the nodes into the
simplest objects possible: locations, NPCs and items. For each
node based on an article about a place, it generates a city
(if the place contains a geographic coordinate) or a building.
In the game’s logic, the world contains cities, and buildings
are places inside cities. Buildings can also contain items and
characters. If the system generates a building, it tries to place
it into its respective city. If it cannot find any city related to
that building, it will randomly pick a place from Wikipedia
and generate a city for it, placing the building in it.
After buildings and cities are created, the system takes
all nodes based on real people and generates one NPC for
each. The NPC gets the original person’s name and a small
description. Any node that is not a person or a location is
5
influenced
Albert
Einstein
d
nce
lue
inf
doctoral advisor
Nathan
Rose
lac
thp
bir
e
Brooklyn
y
ntr
cou
United
States
is
io
nat
ity
nal
field
of
John C.
Slater
l
ora
oct of
is d visor
ad
William
Shockley
ced
uen
nfl
i
s
i
by
is field of
Clinton
Davisson
Physics
d
fiel
i
is f
o
eld
f
William
Rankine
Fig. 3: The major and minor paths between Albert Einstein and William J. M. Rankine. Major paths have dotted arrows, minor
paths have black arrows. Locations are represented as hexagons, NPCs as circles and items (books or photographs) as squares.
transformed into an item: either a book, a list, a letter or
a photograph. Depending on the type of item, different text
templates are generated to explain it.
It is not possible to transform the tree’s root into an NPC,
because he/she is supposed to be murdered. WikiMystery
attempts to solve this by adding people related to the victim
instead. For each suspect, it searches for a person directly
connected to the victim, and transforms him/her into an NPC.
If it cannot find enough people, it generates “random” NPCs,
whose sole purpose is to give a clue about the following node.
Once all objects needed for the plot have been generated,
it is necessary to create a logical sequence of steps from the
victim to each suspect. The system traverses each branch in the
tree, and adds clues and conditions from one node to the next.
If the current node is a location, an NPC or item is generated
and placed in it. If it is a person, dialog is created directing the
player to the next node. We discuss dialog generation in more
detail in Section VI. Otherwise, the clue is added to the item’s
text description. Additionally, at random times the game may
generate a “fake” NPC, the sole purpose of which is to provide
a red herring. It is given a random name, no description,
and dialog that is less than helpful. A condition manager
guarantees that non-root game objects are only available after
they have been triggered by another object.
Finally, the system adds “puzzles”. One of the most wellknown puzzles in adventure games is the “lock-and-key”, i.e.
a location that is inaccessible unless the player uses some
specific item to unlock it. WikiMystery generates this kind of
puzzles, creating items that are able to unlock buildings, such
as flashlights for dark places and crowbars for chained gates.
Puzzle objects are placed via a variation of the Breadth-First
Search algorithm. First, nodes in the tree are separated by their
depth, so depth of 0 will have only the location of the root
NPCs, depth of 1 would contain all locations available after
talking to the root NPCs, and so on. We simulate a playthrough
to perform said separation. It also maintains an array with
all possible keys (e.g. “keys”,“crowbars”, etc) and an initially
empty stack of locks. At every depth, it randomly chooses
whether to put a key in a building of that depth. If it does so,
it adds the respective lock to the stack of locks. Additionally, it
may randomly pop a lock from the stack and add it to another
building. For example, at depth 0 it may chose to put the
“flashlight” key in the root building. It will automatically add
the lock “darkness” to the stack. Because the victim’s house is
the only building at depth 0, and it has already been chosen,
the algorithm goes to depth 1. It randomly chooses to put no
key and no lock, so it skips straight to depth 2, where it finds a
Church and a House. It randomly decides to not put any keys,
but decides to pop the lock from the stack (i.e. the “darkness”
lock) and adds it to the Church. Adding the key before the
lock guarantees that the puzzle will be solvable.
VI. D IALOG G ENERATION
The game’s dialog has two goals: to advance the game by
giving hints and evidence needed to win, and to provide a
sense of depth and immersion, which can be hard to capture
in a data game. Each NPC has their own dialog tree, i.e. the
lines of dialog that both the player and the NPC use when
interacting, along with the dialog options for the player. The
root of this tree is a simple “Hello”, and the choices that follow
are called dialog branches. There are two types of branches:
the main branch containing information necessary to complete
the game, and the side branch (of which there are several
subtypes), which contains information that is not necessary to
complete the game but increases immersion.
A. Main Branch
The main dialog branch contains hints that will allow the
player to advance in the game. For every NPC, the generator
parses data and stores information from this person’s DBpedia
page: anything from places, persons, items, and concepts
associated with the person, as well as personal information
like birthplace and birthday, is stored. The dialog generator
then takes sentence templates and replaces placeholder text.
For example, Rosa Beddington’s person object might have
“Jamaica” stored within it as an associated place. When a
player talks to another NPC who is associated with Rosa
Beddington, they might have a dialog node telling the player
where they think Rosa Beddington is.
I saw PERSON in PLACE. You should probably look there.
will now become:
I saw Rosa Beddington in Jamaica. You should probably
look there.
After packaging a sentence of dialog into a dialog node, the
generator adds this node as a child of the dialog root. Dialog
choices are hidden by default unless its parent has been visited.
Thus, the player must select the root “Hello” option before any
branches are revealed to them.
B. Side Branches
Beyond the main dialog branch, the generator also selects
randomly from a set of “side branches”, which have no effect
on the overall story. These branches provide extra information
about the NPC that the player is speaking to. There are future
plans to use these for an educational purpose, where players
6
Fig. 4: An example of a dialog side branch. The player is
speaking to Hermann Einstein in the Albert Einstein game.
can learn more about characters’ backgrounds by talking to
them. Currently, there are 3 possible side branches: birth-dates
and birth-places, current residence, and lifetime achievement.
When data is originally parsed from DBpedia, birth, current
residency and overview information are stored. When creating
a side branch, this data replaces the placeholder words in
templates (as in the main branch); see Fig. 4 for an example
side branch. After the main branch is created, the generator
randomly selects up to two topics for which to generate side
branches, or none. If generated, the system shows an option
on the dialog screen that leads the user to these branches.
(a) Initial World Map
(b) Initial city map of Switzerland
(c) House of Albert Einstein
(d) Clue given during dialog
(e) Travel between cities
(f) Icon of the Photograph item
(g) Zoomed-in Photograph item
(h) Dialog with a Suspect
(i) Finding an Evidence of Innocence
(j) Issuing a warrant
VII. E XAMPLE P LAYTHROUGH
As an indicative playthrough, we describe the first few
minutes of WikiMystery gameplay; this game uses as input
the text “Albert Einstein”, identified by the TIME magazine
[26] as the “Person of the [20th] Century”. Once the game
launches, the user can load any of the 100 most influential
people of the 20th century, which were pre-generated for the
purposes of the analysis of Section VIII.
The game starts at the world map (see Fig. 5a), where
only one point can be visited: Switzerland, chosen as the
birthplace of Albert Einstein. Clicking on that point of interest,
the user moves to a map of a location in Switzerland collected
from OpenStreetMap6 , where a single location titled “House
of Albert Einstein” (see Fig. 5b) can be visited. The player
also has access to a backpack in this screen (bottom right of
Fig. 5b), which is currently empty but can store items that can
be used to access locked locations. When the player clicks on
the house of Albert Einstein, they move to the building screen
which shows a background of a single-story house, coupled
with informative text about Switzerland in the bottom area (see
Fig. 5c) and six different game icons that can be interacted
with to the right. The first five icons are NPCs, while the last
icon displays a crowbar which can be stored in the inventory
by clicking on the hand button on the crowbar icon.
As noted above, in the House of Albert Einstein there are
five NPCs which the player has the option of observing (eye
button under each icon in Fig. 5c) or talking to (dialog button
under each icon in Fig. 5c). These NPCs are Leo Szilard,
David Joseph Bohm, Jean Gebser, Riazuddin, and James; the
last NPC was randomly generated and given a random name,
while the remaining NPCs are physicists except Gebser who
is a philosopher. Clicking on the eye button gives information
about each of these NPCs (for the random NPC James the text
6 Apparently what is shown in Fig. 5b is an open area near the Melchtal
Valley, as the DBpedia entry places the coordinates of the country of
Switzerland at its center.
Fig. 5: In-game screenshots of the mystery around the murder
of Albert Einstein.
says “There is no information available for this character”).
Note that the images chosen for all NPCs are not images of
people in the case of Szilard and Bohm (instead the images
are related to the atomic bomb), and not images of the correct
people in the case of Gebser and Riazuddin.
Clicking on the dialog button of any NPC icon moves the
user to the dialog overlay (see Fig. 5d), where the NPC’s name
is followed by their response text, followed in turn by a set of
dialog options. The general dialog sequence for any NPC in
the house of Albert Einstein revolves around first asking for
help, then asking their name, and then about any information
they might have and where the player should look for it. The
responses of each NPC depend on which path they are on, and
towards which suspect they will guide the player. Indicatively,
if the player initiates dialog with Leo Szilard then the NPC will
respond to the player’s question “Please state your name.” with
“Maybe... You can call me Leo Szilard. I was one influence.”;
in this case, Leo Szilard was an influence to Albert Einstein,
7
which (subtly) explains why this character is in this game.
When the player asks “Is there something you think I should
know?”, Leo Szilard responds “He talked to Hermann Einstein
a lot.”; the player can then ask “Where is Hermann Einstein?”
to which Leo Szilard responds “Hermann Einstein went to the
United States not long ago.” (see Fig. 5d). This immediately
adds “the United States” as a location in the world map, puts
the house of Hermann Einstein on its city map, and an NPC
named Hermann Einstein within it.
After the player has talked to all five NPCs in the House
of Albert Einstein, there are five more locations in the world
map that they can visit by clicking on them, at which point a
small plane will be shown traveling from the player’s current
location to the selected one (see Fig. 5e). These locations on
the world map are ‘the United States’7 (containing the house
of Hermann Einstein), ‘Princeton, New Jersey’ (containing
a Tunnel building), ‘Israel’ (containing the house of Nathan
Rosen), ‘Wrttemberg’ (containing a Stadium building) and
‘Swiss Federal Institute of Technology in Zurich’ (containing a
building of the same name, and placed in Zurich on the map).
Similarly to the house of Albert Einstein, there is one or more
NPCs or clues in each building listed above. For instance, in
the Tunnel building of Princeton, New Jersey there is a random
NPC named Vlad and a photograph icon (left-most in Fig. 5f).
Clicking on the photograph shows an image of Jewish people
(see Fig. 5g), and its description says:
It is a photograph about Jews.
In a corner, it says: The Jews, also known as the Jewish
people, are an ethnoreligious group originating from the
Israelites, or Hebrews, of the Ancient Near East. Jewish
ethnicity, nationhood and religion are strongly interrelated,
as Judaism is the traditional faith of the Jewish nation,
while its observance varies from strict observance to
complete nonobservance.
There are names written behind:
Canada
Israel
This information is based on the abstract of the Wikipedia
article8 regarding the ‘category: Jews’ (as stored in DBpedia)
which is used to link different NPCs in the mystery together.
These NPCs are Nathan Rosen in Israel and a random NPC
(also named Vlad) in the University building in a world map
location named ‘Canada’; Vlad reveals that Allen Goodrich
Shenstone is located in his house in Princeton, New Jersey.
After an extensive investigation taking the player to many
different cities around the globe, and slowly revealing more
and more buildings, NPCs and clues in previously visited
cities, the player finds the five suspects. In this mystery the
suspects are Sir David MacKay (whose dialog and pleas for innocence are shown in Fig. 5h), Allen Shenstone, William John
Macquorn Rankine, Jakob Meisenheimer, Rosa Beddington.
Each of these five names are also provided — after questioning
— by one of the five NPCs in the player’s starting location:
the house of Albert Einstein. Of those suspect NPCs, Rosa
Beddington (linked to Einstein as a fellow scientist and having
been awarded by the Royal Society) is the culprit. Other
7 The marker based on DBpedia information is again placed at the center
of the U.S.A.
8 https://en.wikipedia.org/wiki/Jews, accessed 13 March 2017.
suspects such as Sir David MacKay can be absolved by finding
an evidence of innocence, in this case provided by chemist
Jacob Meisenheimer (see Fig. 5i). Once the player is confident
they have collected enough evidence, they can click on the
cellphone (bottom left corner of Fig. 5d) to choose the guilty
person as in Fig. 5j. The player chooses the guilty person and
once they do so, the remaining suspects are placed in another
window (bottom half of Fig. 5j); the player must then specify
one characteristic and the correct value for each person which
make them incapable of having committed the murder. The
characteristics and values that absolve all suspects except Rosa
Beddington are included in Table II. If the player selects the
culprit and chooses values for the remaining suspects, they can
click on the “arrest” button (bottom right of Fig. 5j) at which
point the game ends with a message of success or failure.
VIII. E VALUATION
While the playthrough of Section VII provides a glimpse of
what it means to play a generated murder mystery, this section
evaluates the content generated from a broader set of murdered
Wikipedia persons. The goal is two-fold: estimate the number
of interactions afforded in each game (e.g. dialogs with NPCs,
visits to cities, item pickups), and assess the sensitivity of
the system to different inputs (i.e. Wikipedia persons). For
the former, several metrics regarding instances of specific
elements (cities, NPCs, dialog lines) per generated game are
listed. For the latter, we describe which Wikipedia persons
were murdered in games with the highest and lowest values
in these metrics. While this paper does not perform an user
playtest of such generated games to assess e.g. how intuitive
the connections between NPCs are, the provided evaluation
is vital in understanding how complex the generated games
are and which of the generated gameplay elements contribute
most to this complexity. This evaluation is thus a first step
prior to a playtest, to assess for instance the minimum number
of player clicks (via the tree size metric combined with the
dialoge nodes metric) for a game to be completed. Such
metrics can then be compared with actual metrics derived
during playtests, but can also inform changes to the generative
algorithms before such playtests can take place.
To assess a broad range of games, based on persons with
a strong presence in Wikipedia, we used the list of the TIME
magazine’s 100 most influential people of the 20th century
[26] as input. Each person in the list became the victim
in a procedurally generated game, some after preprocessing,
excluding two: “American G.I.” and “Unknown Rebel”. The
system was not able to generate games with them, as the first
represents a whole category (we could not choose a single
person that represented this category), and the latter represents
an unknown person who does not contain the tag “Person”
in his DBpedia page. Additionally, the system cannot process
groups of people, so inputs such as “The Kennedy Family” had
to be transformed into a single individual. Entries about groups
were transformed into one of the most known people of the
group. For example, “The Beatles” became “John Lennon” and
“The Kennedy Political Family” became “John F. Kennedy”.
The system generated a total of 98 games, one per input.
Table I shows the quantitative results.
8
TABLE I: Average metrics of all generated adventure games
for the 98 most influential people.
Location metrics
Cities
Buildings
Average buildings per city
Item & Puzzle metrics
All Items
Books
Photographs (torn or not)
Torn Photographs
Key items
Locked buildings
NPC metrics
All NPCs
NPCs based on real people
Average ratio of real NPCs over all NPCs
Average NPCs per building
Dialog metrics
All dialog nodes
Average dialog nodes per NPC
All side-branches
Achievement side-branches
Residence side-branches
Birth side-branches
Complexity
Average length of paths
Tree size
18.07
46.37
2.88
23.91
9.45
7.24
2.55
2.71
2.67
46.53
24.03
52%
1.02
208.33
4.45
35.47
9.05
8.56
17.86
12.07
60.34
A. Game Content
Based on Table I, the average tree size of the generated
games is 60.3 nodes. The game with the smallest tree size
had “Robert H. Goddard” as input and 20 nodes. “Marlon
Brando”, “Martin Luther King, Jr.”, “Richard Rodgers” and
“Willis Carrier” tied for the most nodes in the tree, with 65. On
average, the length of paths between victim and each suspect
was 12 nodes. Six games had the lowest path length with 5
nodes, while 39 had the highest path length with 13 nodes.
Each game had on average around 18 cities and 46 buildings, with approximately 2.9 buildings per city. The most
common cities amongst all games was “The United States”,
appearing in 73 out of the 98 games, followed by “New York
City” (43) and “District of Columbia” (37). North American
locations dominated the top 10 most common cities, with 8
locations. The remaining two were “London” and “Germany”.
Note that while the game only represents locations as cities
and buildings, the in-game city category may include countries
(e.g. The United States), states and actual cities.
On average, 24 items were generated per game, mostly
books (9.45). Games with most and fewest books were created
from, respectively, “Le Corbusier” (20 books) and “Theodore
Roosevelt” (1). Key items and locked buildings tend to appear
together, with an average of 2.71 key items and 2.67 buildings.
Every game had at least one key and one locked building,
while at most there were three keys and three locked buildings
in a single game. The number of keys was always equal or
higher to that of locked buildings, ensuring solvability.
An average of 46 NPCs were created per game, and on
average 24 were based on real people (ratio of 52%). While
this ratio is not optimal, we believe it can be improved in
future versions by being more lenient in the NPC generation:
now, we only look at people with one-degree distance from
the article that originated the node. If there is no person, we
could expand the search to 2 or 3 degrees distances, which we
believe can improve this ratio. We believe that increasing the
percentage of NPCs based on real people over “random” NPCs
would provide more interesting, full-fledged characters and
interactions. The ratio of NPCs based on real people ranged
from 28% to 76% of all NPCs. The game with most real NPCs
was generated from “Lech Wałesa” with 41 real NPCs, and
the one with the least from “Walter Reuther” with 9.
Based on Table I, there are 208.33 dialog nodes on average
in a WikiMystery game, distributed across all NPCs in the
game. Results show an average of 4.45 dialog nodes per
person. Every person has a main branch in their dialog tree,
so the number of main branches is equal to the number of
NPCs. In addition to those, there are on average 35.5 sidebranches in a game. Of those, an average of 9 side-branches
refer to the NPC’s personal achievements, 8.6 side-branches
concern the NPC’s current residence, and 17.9 side-branches
are associated with the person’s birth. There are nearly twice
as many side-branches on birth than the other two types, since
the generator creates two branches (birth date and birth place)
when selecting a side-branch on birth.
B. Suspects, Direct Connections and Evidence
Each generated game must have a set of suspects, evidence
of innocence and direct connections between suspect and
victim (i.e. the reason for selecting those suspects). Table II
shows the set of suspects, evidence of innocence and direct
connection between victim and suspect of the three most
influential people on TIME’s list: Albert Einstein, Franklin
D. Roosevelt and Mahatma Gandhi. Values in italics are used
as evidence for innocent suspects, allowing the player to
differentiate between them and the culprit. Notice that when
the culprit has no value (e.g. in the game generated from
Franklin Roosevelt, Daniel Poulter has no value for the “Party”
characteristic), any value would fit to differentiate between any
suspect, but the game will only check this characteristic for the
specific paired suspect (in the game generated from Franklin
Roosevelt, Gwendolyn Garcia is paired up with the “Party”
characteristic). Additionally, if two innocent suspects share the
same value for a characteristic, it is used as evidence for one
of them, as long as it is different from the culprit. That merely
means that it is evidence that only one suspect is innocent, but
does not absolve the other suspect. An example is shown in
Mahatma Gandhi’s game (see Table II) where Tex Avery and
Jhunnilal Verma both died in 1980, but this is only evidence
of innocence for Avery (note that Eddie Lyons, the culprit,
died in 1926). In some cases, the actual value appeared as
a consequence of Wikipedia’s own organization. In Mahatma
Gandhi’s game, the primary reason for selecting Tex Avery and
Eddie Lyons as suspects is them belonging (like Gandhi) to the
category “Articles containing video clips”, indicating that they
both appear in the Wikipedia list of articles containing video
clips. The secondary reason was that this set of these five
suspect allowed for a solvable and somewhat diverse game,
according to the GA.
Direct connections are relations between the victim and each
suspect. Since they depend on hyperlinks from the victim’s
9
TABLE II: Solution from games generated with the top 3 most influential people: Albert Einstein, Franklin D. Roosevelt and
Mahatma Gandhi. Each innocent suspect is paired to one characteristic (in blue italics) that will differentiate him/her from the
killer (whose name is shown with an asterisk, and is last in each list). Empty values appear in the game as “Unknown”. The
direct connection column shows the primary criteria for choosing the suspect in relationship to the victim (e.g. “politicians”
means that both the victim and that suspect were politicians. thus the suspect was directly connected to the victim).
Albert Einstein
Suspects
Jakob Meisenheimer
Death Place
Field
Subject
1934 deaths
Nazi Germany
Sir David MacKay
Information theory
Living people
Thermodynamicists
Fellows of the Royal
Society
20th-century women
scientists
Glasgow
Physics
Allen G. Shenstone
United States
Physics
Rosa Beddington*
Great Tew
Developmental
biology
Suspects
Term end
Kevin Cahill
1994-12-31
Gwendolyn Garcia
William J. M. Rankine
AlmaMater
Ludwig Maximilian
University of Munich
California Institute of
Technology
University of Edinburgh
Brasenose College,
Oxford
Direct Connection
Born in the German
Empire
Received the Royal
Society award
Physicists
Died in Princeton, New
Jersey
Received the Royal
Society award
AlmaMater
Direct Connection
State University of New
York
Part of Democratic Party
Princeton University
Franklin D. Roosevelt
Party
Democratic Party
(US) from the 103rd
District
Office
Member of the New
York Assembly at
New Paltz
2013-06-30
One Cebu
Governor of Cebu
Johnny Ellis
1993-01-11
Democratic Party
(US)
Jane Griffiths
2005-04-11
Labour Party (UK)
Daniel Poulter*
2015-05-12
Majority Leader of
the Alaska Senate
Member of
parliament
Member of
parliament
Suspects
Death year
University of the
Philippines Diliman
Claremont McKenna
College
Part of Democratic Party,
politicians
Durham University
Politicians
University of Bristol
Politicians
Direct Connection
Appear in the Wikipedia
category of articles
containing video clips
Politicians
Mahatma Gandhi
Birth place
Subject
Occupation
Articles containing
video clips
Animator, cartoonist,
voice actor, director
Tex Avery
1980
Taylor, Texas
Stanley Rosen
2014
Cleveland
Jhunnilal Verma
1980
Eddie Lyons*
1926
Writer
20th century philosophers
Damoh, India
Jewish American
writers
20th-century
philosophers
People from Damoh
Lawyer
Beardstown, Illinois,
USA
Articles containing
video clips
Actor, director,
screenwriter, producer
Indian lawyers
Appear in the Wikipedia
category of articles
containing video clips
Landau
Volker Zotz
Wikipedia page, they can only be as varied as the article
itself. Therefore, games usually have an emergent underlying
theme. In the game generated for Albert Einstein, both he,
Allen Shenston and William J. M. Rankine were in the field
of Physics, and he and Shenston both also died in Princeton,
New Jersey; Einstein, David Mackay and Rosa Beddington
received the Royal Society award, and both Einstein and Jakob
Meisenheimer were born in the German Empire. In Franklin
D. Roosevelt’s game, Roosevelt and all suspects except Kevin
Cahill are politicians9 . Roosevelt, Cahill and Johnny Ellis were
all part of the U.S.A. Democratic Party. Finally, in Mahatma
Gandhi’s game, both he and Jhunnilal Verma were Indian
lawyers; Gandhi, Volker Zotz and Stanley Rosen were 20th
century philosophers; and he, Eddie Lyons and Tex Avery
all appear in the Wikipedia category of articles containing
video clips. The last connection is arguably poorer than others,
demonstrating how some of the source data can be difficult to
tailor to our needs; this is discussed extensively in Section IX.
9 Cahill
actually is a politician, he is not tagged as one in DBpedia.
20th century philosophers
IX. D ISCUSSION
The sample playthrough of Section VII and the numerical
evaluations of Section VIII provide a high-level overview of
the types of games generated by the current WikiMystery
prototype. Contrary to our early attempts at adventure generation, which created one path between two people [4], the
murder mystery is far less linear and includes more dialog and
gameplay options. The fact that paths can be traversed nonsequentially (inevitably so, as it is difficult to keep track of
which NPC or object forms a path towards which suspect)
increases the exploration and branching factor in terms of
decision-making on the part of the player. This in turn leads
to more interesting gameplay as it gives a greater sense
of player agency. The gameplay has been improved with
more branching, better visual presentation of results, more
interesting dialog options, and a concrete winning condition.
While it was a priority for the authors to improve on the
gameplay quality of the broader Data Adventures project, the
biggest appeal remains its link to real-world data accessed
10
(a) Darling house of early colonial (b) The image of Confucius chosen
Australia highlighted... and placed in for Hermann Einstein, and the ‘ask
Israel.
about Jews’ dialog are... unfortunate.
Fig. 6: Absurd and potentially offensive combinations of data
can occur with WikiMystery.
via open data repositories. Based on the metrics of Table I,
that aspect of WikiMystery has been strengthened as well,
with each game containing a multitude of cities placed in
their real-world locations and with a city map showing their
street view (based on OpenStreetMap). The ratio of random
NPCs to “real” NPCs, which are based on Wikipedia articles,
is also kept in balance, while the introduction of photograph
in-game objects increases the modes through which open
data can be experienced (i.e. through images rather than text
information found in book in-game objects). Most importantly,
the improved NPC dialog allows not only for a more engaging
and intuitive way to solve the mystery but also allows for
yet another way to present open data, as a player can choose
what questions to ask of the NPC (e.g. regarding their life
achievements) rather than being presented the data as a large
chunk of text when observing the NPC, for example.
Although there have been substantial improvements in the
presentation of content from earlier iterations of Data Adventures [4], [25], the very nature of generating games from
open data hinges on the uncontrollable nature of such data.
This allows for nigh-infinite expressivity, as any person with
a Wikipedia presence can potentially star in a generated game
(where he is murdered), but this lack of control can lead to
unexpected, unintended or even unwanted outcomes. On the
one hand, the ongoing efforts of the Data Adventures line of
research focus on controlling this vast repository of data and
transforming it into intuitive and playable objects; for instance,
attempting to find unique connections between people rather
than trivial ones such as “they are both human”. It is, however,
impossible to ever fully control or constrain the experience,
as doing so would obfuscate its origins from a living, vast
knowledge base rooted deeply in the real world. It is that very
absurdity that makes the outcomes appealing in their own way;
as the user of another data-based game titled A Rogue Dream
states, it feels “like playing a videogame against The Internet”
[2], and at least in the case of WikiMystery that is intentional.
This absurdity, however, causes some hilarious, and sometimes appalling, outcomes. It has been noted in the playthrough
of Section VII that most NPCs’ images were not correct,
which is either due to the lack of appropriate images for those
people in WikiMedia Commons, or flaws in the image parsers
currently at hand. In such cases, a random search for an image
of a man (for male NPCs) and a woman (for female NPCs)
is used instead. For buildings, moreover, the image search
is based on the name of the building without context of its
geographical location. This can lead to results such as that
of Fig. 6a, where not only is the building’s background an
old photograph with an actual highlighted building with a red
circle, but on closer inspection the chosen building (result of
a search for “house of Nathan Rose”) is the Darling House,
which holds historical significance for early colonial Australia,
but in the game is used as a domicile for Nathan Rosen
in Israel. Our choice of using only freely available sources,
such as Wikimedia Commons, complicates the retrieval of
specific images. A source such as Google Images could
improve results, but contradicts the scope of freely available
solutions. Therefore, future work should improve the search
for appropriate images, possibly by increasing the breadth of
searches in more repositories or by performing some computer
vision verification that e.g. the image is one of a person.
Additionally, more problematic are instances where an
unforeseen combination of content and their transformation
can lead to insensitive or offensive results. As an example,
Fig. 6b shows the dialog with Hermann Einstein as part of
the playthrough of Section VII where the player is seeking the
culprit of Albert Einstein’s murder. The image, unsurprisingly,
is not that of Hermann Einstein; instead, the random search
for an image of a man serendipitously ended up being a
drawing of Confucius. On the other hand, the dialog has
chosen to highlight that the connection between this person
and the next along the path is the ‘category: Jews’; this
category was also cued by the photograph of Fig. 5g as
discussed extensively in Section VII. In this case, the player
interacts with Hermann Einstein with the dialog line “hAsk
about Jewsi”. It is certainly true that the actual story of Albert
Einstein was deeply affected by him being Jewish and the
events of World War II, so the category and the path found
is accurate (perhaps desirable), however the random choice of
this dialog line10 and the random assignment of an image of
Confucius for avatar are a very unfortunate, insensitive and
likely offensive combination. It is difficult to envision how
such instances could be avoided, as it was largely an issue
with simple transformations of data and their combination
going awry. While it is not the case here, one should also
not underestimate that the nature of open online data “is
often tainted by popular belief, misconception, stereotype and
prejudice, as opposed to purely factual information” [2], and
thus such unfortunate instances may actually occur due to
prejudice in the source data before they are even transformed.
There are still important directions for future work, in order
to improve both the usability of the game and its narrative
consistency. For example, some interface additions, such as a
travel diary, could help the player to keep track of clues and
connections between NPCs, objects, locations and characteristics. Moreover, the current dialog format uses fixed templates
and sequences; perhaps a grammar-based approach such as
Tracery [27] could result in more diverse, life-like dialogs.
Furthermore, an important missing component in the narrative
of the murder is the culprit’s motive (and possible motives for
other suspects). It is unlikely that a motive such as jealousy
10 Consider how inoffensive a similar line saying “hAsk about Physicsi”
would be.
11
would be based on real facts and data, although it could be
generated based on relationships of people (e.g. siblings or
spouses) as in [28]. Exploring the relations between the NPCs,
their personalities and goals seems promising. Using data to do
so is not trivial, as even a sentiment analysis over a Wikipedia
article about a person would only express the writer’s feelings,
not the actual subject’s. On the other hand, we have yet to fully
understand how players interact and view the data presented
in WikiMystery. It is one of our priorities to release a playable
version of the game online, setting up a logging system so that
we can perform user studies. We also intend to investigate
the possibility of using WikiMystery to gain insight on the
correctness of data in DBpedia and Wikipedia.
X. C ONCLUSION
This paper presented the latest installment of the WikiMystery game, and detailed its complex generation pipeline, from
the name of a person with a Wikipedia article to a full
interactive murder mystery game. Open data is used in a
multitude of ways in order to find NPC suspects for an in-game
murder of a specified person, to find paths linking these NPCs,
to place them in locations around the globe and to provide a
way for the player to absolve innocents and deduce the culprit.
Moreover, open data is used to create the ‘levels’ (i.e. cities
and buildings) in which NPCs are found, to create in-game
objects with photographs and books that act as clues, and to
enhance dialog options of NPCs beyond the merely functional
needs of completing the game. While there are many directions
of future work in order to increase gameplay intuitiveness, to
provide a better link between visuals and other content, and to
reduce the absurdity of combinations, the current WikiMystery
generator is the first to create fully playable adventure games
with minimal human authorship and curation.
XI. ACKNOWLEDGMENTS
NPCs discussed in the generated adventures are instantiated
from real people, but the similarities end there. The NPCs’
actions in the game (as victims or culprits) in no way reflect the
real-world people they are based on. The generator’s output in
no way accuses or misrepresents these real-world individuals.
WikiMystery creates fictional counterparts of public figures
who have a presence in Wikipedia: any similarity between the
(fictional) NPCs in the game and real-world people is therefore
due to the data available in these open, online, freely accessible
and editable repositories.
We thank Ahmed Khalifa and Scott Lee for helpful insight.
Gabriella Barros acknowledges financial support from CAPES
and Science Without Borders program, BEX 1372713-3.
R EFERENCES
[1] M. Treanor, B. Blackford, M. Mateas, and I. Bogost, “Game-o-matic:
Generating videogames that represent ideas,” in Proceedings of the FDG
workshop on Procedural Content Generation, 2012.
[2] M. Cook and S. Colton, “A rogue dream: Automatically generating
meaningful content for games,” in Proceedings of the AIIDE workshop
on Experimental AI & Games, 2014.
[3] G. A. B. Barros, A. Liapis, and J. Togelius, “Murder mystery generation
from open data,” in Proceedings of the Seventh International Conference
on Computational Creativity, 2016.
[4] ——, “Data adventures,” in Proceedings of the FDG workshop on
Procedural Content Generation in Games, 2015.
[5] M. G. Friberger, J. Togelius, A. B. Cardona, M. Ermacora, A. Mousten,
M. M. Jensen, V. Tanase, and U. Brøndsted, “Data games,” in 4th
Workshop on Procedural Content Generation, 2013.
[6] A. B. Cardona, A. W. Hansen, J. Togelius, and M. Gustafsson Friberger,
“Open trumps, a data game,” in Proceedings of the Foundations of
Digital Games, 2014.
[7] J. Togelius and M. Gustafsson Friberger, “Bar chart ball, a data game,”
in Proceedings of the Foundations of Digital Games, 2013.
[8] G. A. Barros and J. Togelius, “Balanced civilization map generation
based on open data,” in IEEE Congress on Evolutionary Computation
(CEC). IEEE, 2015, pp. 1482–1489.
[9] S. Bringsjord and D. A. Ferrucci, “Artificial intelligence and literary
creativity: Inside the mind of brutus, a storytelling machine,” Computational Linguistics, vol. 26, no. 4, pp. 642–647, 2000.
[10] S. R. Turner, “Minstrel: A computer model of creativity and storytelling,” Ph.D. dissertation, Los Angeles, CA, USA, 1993.
[11] G. Vink, “Bardiche: An interactive online narrative generator,” B.S.
thesis, 2015.
[12] P. Lopes, A. Liapis, and G. N. Yannakakis, “Sonancia: Sonification
of procedurally generated game levels,” in Proceedings of the 1st
Computational Creativity and Games Workshop, 2015.
[13] M. Thorogood, P. Pasquier, and A. Eigenfeldt, “Audio metaphor: Audio
information retrieval for soundscape composition,” Proceedings of the
Sound and Music Computing Cong.(SMC), 2012.
[14] M. Cook, S. Colton, and J. Gow, “Automating game design in three
dimensions,” in Proceedings of the AISB Symposium on AI and Games,
2014, pp. 20–24.
[15] J. Alderman, C. Ashmore, K. Compton, M. Shapiro, and M. Nitsche,
“The many worlds of charbitat,” in Game Set and Match II, 2006.
[16] C. Ashmore and M. Nitsche, “The quest in a generated world,” in
Proceedings of the DiGRA Conference, 2007, pp. 503–509.
[17] B. Lavender and T. Thompson, “Adventures in hyrule: Generating
missions & maps for action adventure games,” 2015.
[18] C. Fernández-Vara and A. Thomson, “Procedural generation of narrative
puzzles in adventure games: The puzzle-dice system,” in Proceedings of
the The third workshop on Procedural Content Generation in Games.
ACM, 2012, p. 12.
[19] C. Fernández-Vara, “Creating dreamlike game worlds through procedural content generation,” in Seventh Intelligent Narrative Technologies
Workshop, 2014.
[20] J. Weizenbaum, “ELIZA - A computer program for the study of natural
language communication between man and machine,” vol. 10, no. 8,
1966.
[21] H. Hernault, P. Piwek, H. Prendinger, and M. Ishizuka, “Generating
dialogues for virtual agents using nested textual coherence relations,” in
International Workshop on Intelligent Virtual Agents. Springer, 2008.
[22] M. Cavazza and F. Charles, “Dialogue generation in character-based
interactive storytelling.” in Proceedings of the AI in Interactive Digital
Entertainment conference, 2005, pp. 21–26.
[23] C. Fernandez-Vara and S. Osterweil, “The key to adventure game design:
Insight and sense-making,” in Proceedings of Meaningful Play, 2010.
[24] J. Togelius, R. De Nardi, and S. M. Lucas, “Towards automatic personalised content creation for racing games,” in IEEE Symposium on
Computational Intelligence and Games. IEEE, 2007, pp. 252–259.
[25] G. A. B. Barros, A. Liapis, and J. Togelius, “Playing with data:
Procedural generation of adventures from open data,” in Proceedings
of the International Joint Conference of DiGRA and FDG, 2016.
[26] F. Golden, “Albert Einstein,” vol. 154, no. 27, 1999.
[27] K. Compton, B. Kybartas, and M. Mateas, “Tracery: an author-focused
generative text tool,” in International Conference on Interactive Digital
Storytelling. Springer, 2015, pp. 154–161.
[28] A. Stockdale, “Cluegen: An exploration of procedural storytelling in
the format of murder mystery games,” in Proceedings of the AIIDE
workshop on Experimental AI in Games, vol. 2, 2003.
| 2cs.AI
|
arXiv:1705.07347v3 [stat.ML] 22 Nov 2017
Ensemble Sampling
Xiuyuan Lu
Stanford University
[email protected]
Benjamin Van Roy
Stanford University
[email protected]
Abstract
Thompson sampling has emerged as an effective heuristic for a broad range of
online decision problems. In its basic form, the algorithm requires computing
and sampling from a posterior distribution over models, which is tractable only
for simple special cases. This paper develops ensemble sampling, which aims to
approximate Thompson sampling while maintaining tractability even in the face
of complex models such as neural networks. Ensemble sampling dramatically
expands on the range of applications for which Thompson sampling is viable. We
establish a theoretical basis that supports the approach and present computational
results that offer further insight.
1 Introduction
Thompson sampling [8] has emerged as an effective heuristic for trading off between exploration
and exploitation in a broad range of online decision problems. To select an action, the algorithm
samples a model of the system from the prevailing posterior distribution and then determines which
action maximizes expected immediate reward according to the sampled model. In its basic form,
the algorithm requires computing and sampling from a posterior distribution over models, which is
tractable only for simple special cases.
With complex models such as neural networks, exact computation of posterior distributions becomes
intractable. One can resort to to the Laplace approximation, as discussed, for example, in [2, 5], but
this approach is suitable only when posterior distributions are unimodal, and computations become
an obstacle with complex models like neural networks because compute time requirements grow
quadratically with the number of parameters. An alternative is to leverage Markov chain Monte
Carlo methods, but those are computationally onerous, especially when the model is complex.
A practical approximation to Thompson sampling that can address complex models and problems
requiring frequent decisions should facilitate fast incremental updating. That is, the time required per
time period to learn from new data and generate a new sample model should be small and should not
grow with time. Such a fast incremental method that builds on the Laplace approximation concept is
presented in [5]. In this paper, we study a fast incremental method that applies more broadly, without
relying on unimodality. As a sanity check we offer theoretical assurances that apply to the special
case of linear bandits. We also present computational results involving simple bandit problems as
well as complex neural network models that demonstrate efficacy of the approach.
Our approach is inspired by [6], which applies a similar concept to the more complex context of deep
reinforcement learning, but without any theoretical analysis. The essential idea is to maintain and
incrementally update an ensemble of statistically plausible models, and to sample uniformly from
this set in each time period as an approximation to sampling from the posterior distribution. Each
model is initially sampled from the prior, and then updated in a manner that incorporates data and
random perturbations that diversify the models. The intention is for the ensemble to approximate
the posterior distribution and the variance among models to diminish as the posterior concentrates.
We refine this methodology and bound the incremental regret relative to exact Thompson sampling
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
for a broad class of online decision problems. Our bound indicates that it suffices to maintain a
number of models that grows only logarithmically with the horizon of the decision problem, ensuring
computational tractability of the approach.
2 Problem formulation
We consider a broad class of online decision problems to which Thompson sampling could, in principle, be applied, though that would typically be hindered by intractable computational requirements.
We will define random variables with respect to a probability space (Ω, F, P) endowed with a filtration (Ft : t = 0, . . . , T ). As a convention, random variables we index by t will be Ft -measurable,
and we use Pt and Et to denote probabilities and expectations conditioned on Ft . The decisionmaker chooses actions A0 , . . . , AT −1 ∈ A and observes outcomes Y1 , . . . , YT ∈ Y. There is a
random variable θ, which represents a model index. Conditioned on (θ, At−1 ), Yt is independent
of Ft−1 . Further, P(Yt = y|θ, At−1 ) does not depend on t. This can be thought of as a Bayesian
formulation, where randomness in θ reflects prior uncertainty about which model corresponds to the
true nature of the system.
We assume that A is finite and that each action At is chosen by a randomized policy π =
(π0 , . . . , πT −1 ). Each πt is Ft -measurable, and each realization is a probability mass function over
actions A; At is sampled independently from πt .
The agent associates a reward R(y) with each outcome y ∈ Y, where the reward function R
is fixed and known. Let Rt = R(Yt ) denote the reward realized at time t. Let Rθ (a) =
E [R(Yt )|θ, At−1 = a]. Uncertainty about θ induces uncertainty about the true optimal action, which
we denote by A∗ ∈ arg max Rθ (a). Let R∗ = Rθ (A∗ ). The T -period conditional regret when the
a∈A
actions (A0 , .., AT −1 ) are chosen according to π is defined by
#
" T
X
Regret(T, π, θ) = E
(R∗ − Rt ) θ ,
(1)
t=1
where the expectation is taken over the randomness in actions At and outcomes Yt , conditioned on
θ.
We illustrate with a couple of examples that fit our formulation.
Example 1. (linear bandit) Let θ be drawn from ℜN and distributed according to a N (µ0 , Σ0 )
prior. There is a set of K actions A ⊆ ℜN . At each time t = 0, 1, . . . , T − 1, an action At ∈ A is
2
selected, after which a reward Rt+1 = Yt+1 = θ⊤ At +Wt+1 is observed, where Wt+1 ∼ N (0, σw
).
Example 2. (neural network) Let gθ : ℜN 7→ ℜK denote a mapping induced by a neural network
with weights θ. Suppose there are K actions A ⊆ ℜN , which serve as inputs to the neural network,
and the goal is to select inputs that yield desirable outputs. At each time t = 0, 1, . . . , T − 1,
an action At ∈ A is selected, after which Yt+1 = gθ (At ) + Wt+1 is observed, where Wt+1 ∼
2
N (0, σw
I). A reward Rt+1 = R(Yt+1 ) is associated with each observation. Let θ be distributed
according to a N (µ0 , Σ0 ) prior. The idea here is that data pairs (At , Yt+1 ) can be used to fit a
neural network model, while actions are selected to trade off between generating data pairs that
reduce uncertainty in neural network weights and those that offer desirable immediate outcomes.
3 Algorithms
Thompson sampling offers a heuristic policy for selecting actions. In each time period, the algorithm
samples an action from the posterior distribution pt (a) = Pt (A∗ = a) of the optimal action. In
other words, Thompson sampling uses a policy πt = pt . It is easy to see that this is equivalent to
sampling a model index θ̂t from the posterior distribution of models and then selecting an action
At = arg max Rθ̂t (a) that optimizes the sampled model.
a∈A
Thompson sampling is computationally tractable for some problem classes, like the linear bandit
problem, where the posterior distribution is Gaussian with parameters (µt , Σt ) that can be updated
incrementally and efficiently via Kalman filtering as outcomes are observed. However, when deal2
Algorithm 1 EnsembleSampling
1: Sample: θ̃0,1 , . . . , θ̃0,M ∼ p0
2: for t = 0, . . . , T − 1 do
3:
Sample: m ∼ unif({1, . . . , M })
4:
Act: At = arg max Rθ̃t,m (a)
a∈A
5:
Observe: Yt+1
6:
Update: θ̃t+1,1 , . . . , θ̃t+1,M
7: end for
ing with complex models, like neural networks, computing the posterior distribution becomes intractable. Ensemble sampling serves as an approximation to Thompson sampling for such contexts.
The posterior can be interpreted as a distribution of “statistically plausible” models, by which we
mean models that are sufficiently consistent with prior beliefs and the history of observations. With
this interpretation in mind, Thompson sampling can be thought of as randomly drawing from the
range of statistically plausible models. Ensemble sampling aims to maintain, incrementally update,
and sample from a finite set of such models. In the spirit of particle filtering, this set of models
approximates the posterior distribution. The workings of ensemble sampling are in some ways
more intricate than conventional uses of particle filtering, however, because interactions between
the ensemble of models and selected actions can skew the distribution.
While elements of ensemble sampling require customization, a general template is presented as Algorithm 1. The algorithm begins by sampling M models from the prior distribution. Then, over
each time period, a model is sampled uniformly from the ensemble, an action is selected to maximize expected reward under the sampled model, the resulting outcome is observed, and each of
the M models is updated. To produce an explicit algorithm, we must specify a model class, prior
distribution, and algorithms for sampling from the prior and updating models.
For a concrete illustration, let us consider the linear bandit (Example 1). Though ensemble sampling
is unwarranted in this case, since Thompson sampling is efficient, the linear bandit serves as a useful
context for understanding the approach. Standard algorithms can be used to sample models from
the N (µ0 , Σ0 ) prior. One possible procedure for updating models maintains a covariance matrix,
updating it according to
⊤
2 −1
,
Σt+1 = Σ−1
t + At At /σw
and generates model parameters incrementally according to
2
θ̃t+1,m = Σt+1 Σ−1
t θ̃t,m + At (Rt+1 + W̃t+1,m )/σw ,
2
for m = 1, . . . , M , where (W̃t,m : t = 1, . . . , T, m = 1, . . . , M ) are independent N (0, σw
)
random samples drawn by the updating algorithm. It is easy to show that the resulting parameter
vectors satisfy
!
t−1
1 X
⊤ 2
⊤ −1
θ̃t,m = arg min
(Rτ +1 + W̃τ +1,m − Aτ ν) + (ν − θ̃0,m ) Σ0 (ν − θ̃0,m ) ,
2
σw
ν
τ =0
which admits an intuitive interpretation: each θ̃t,m is a model fit to a randomly perturbed prior
and randomly perturbed observations. As we establish in the appendix, for any deterministic sequence A0 , . . . , At−1 , conditioned on Ft , the models θ̃t,1 , . . . , θ̃t,M are independent and identically
distributed according to the posterior distribution of θ. In this sense, the ensemble approximates
the posterior. It is not a new observation that, for deterministic action sequences, such a scheme
generates exact samples of the posterior distribution (see, e.g., [7]). However, for stochastic action
sequences selected by Algorithm 1, it is not immediately clear how well the ensemble approximates
the posterior distribution. We will provide a bound in the next section which establishes that, as
the number of models M increases, the regret of ensemble sampling quickly approaches that of
Thompson sampling.
3
The ensemble sampling algorithm we have described for the linear bandit problem motivates an
analogous approach for the neural network model of Example 2. This approach would again begin
with M models, with connection weights θ̃0,1 , . . . , θ̃0,M sampled from a N (µ0 , Σ0 ) prior. It could
be natural here to let µ0 = 0 and Σ0 = σ02 I for some variance σ02 chosen so that the range of probable
models spans plausible outcomes. To incrementally update parameters, at each time t, each model
m applies some number of stochastic gradient descent iterations to reduce a loss function of the
form
t−1
1 X
Lt (ν) = 2
(Yτ +1 + W̃τ +1,m − gν (Aτ ))2 + (ν − θ̃0,m )⊤ Σ−1
0 (ν − θ̃0,m ).
σw τ =0
We present computational results in Section 5.2 that demonstrate viability of this approach.
4 Analysis of ensemble sampling for the linear bandit
Past analyses of Thompson sampling have relied on independence between models sampled over
time periods. Ensemble sampling introduces dependencies that may adversely impact performance.
It is not immediately clear whether the degree of degradation should be tolerable and how that
depends on the number of models in the ensemble. In this section, we establish a bound for the
linear bandit context. Our result serves as a sanity check for ensemble sampling and offers insight
that should extend to broader model classes, though we leave formal analysis beyond the linear
bandit for future work.
Consider the linear bandit problem described in Example 1. Let π TS and π ES denote the Thompson
and ensemble sampling policies for this problem, with the latter based on an ensemble of M models,
generated and updated according to the procedure described in Section 3. Let R∗ = mina∈A θ⊤ a
denote the worst mean reward and let ∆(θ) = R∗ − R∗ denote the gap between maximal and
minimal mean rewards. The following result bounds the difference in regret as a function of the gap,
ensemble size, and number of actions.
Theorem 3. For all ǫ > 0, if
M≥
then
4|A|T
4|A|
log
,
2
ǫ
ǫ3
Regret(T, π ES , θ) ≤ Regret(T, π TS , θ) + ǫ∆(θ)T.
This inequality bounds the regret realized by ensemble sampling by a sum of the regret realized by
Thompson sampling and an error term ǫ∆(θ)T . Since we are talking about cumulative regret, the
error term bounds the per-period degradation relative to Thompson sampling by ǫ∆(θ). The value
of ǫ can be made arbitrarily small by increasing M . Hence, with a sufficiently large ensemble, the
per-period loss will be small. This supports the viability of ensemble sampling.
An important implication of this result is that it suffices for the ensemble size to grow logarithmically
in the horizon T . Since Thompson sampling requires independence between models sampled over
time, in a sense, it relies on T models – one per time period. So to be useful, ensemble sampling
should operate effectively with a much smaller number, and the logarithmic dependence is suitable.
The bound also grows with |A| log |A|, which is manageable when there are a modest number of
actions. We conjecture that a similar bound holds that depends instead on a multiple of N log N ,
where N is the linear dimension, which would offer a stronger guarantee when the number of actions
becomes large or infinite, though we leave proof of this alternative bound for future work.
The bound of Theorem 3 is on a notion of regret conditioned on the realization of θ. A Bayesian
regret bound that removes dependence on this realization can be obtained by taking an expectation,
integrating over θ:
E Regret(T, π ES , θ) ≤ E Regret(T, π TS , θ) + ǫE [∆(θ)] T.
We provide a complete proof of Theorem 3 in the appendix. Due to space constraints, we only offer
a sketch here.
Sketch of Proof. Let A denote an Ft -adapted action process (A0 , . . . , AT −1 ). Our procedure for
generating and updating models with ensemble sampling is designed so that, for any deterministic
4
A, conditioned on the history of rewards (R1 , . . . , Rt ), models θ̃t,1 , . . . , θ̃t,M that comprise the
ensemble are independent and identically distributed according to the posterior distribution of θ.
This can be verified via some algebra, as is done in the appendix.
Recall that pt (a) denotes the posterior probability Pt (A∗ = a) = P (A∗ = a|A0 , R1 , . . . , At−1 , Rt ).
To explicitly indicate dependence on the action process, we will use a superscript:
pt (a) = pA
t (a).
P
M
1
A
A
A
⊤
Let p̂t denote an approximation to pt , given by p̂t (a) = M m=1 I a = arg maxa′ θ̃t,m a′ .
Note that given an action process A, at time t Thompson sampling would sample the next action
A
from pA
t , while ensemble sampling would sample the next action from p̂t . If A is deterministic
then, since θ̃t,1 , . . . , θ̃t,M , conditioned on the history of rewards, are i.i.d. and distributed as θ, p̂A
t
represents an empirical distribution of samples drawn from pA
t . It follows from this and Sanov’s
Theorem that, for any deterministic A,
A
|A| −Mǫ
P dKL (p̂A
e
.
t kpt ) ≥ ǫ|θ ≤ (M + 1)
A naive application of the union bound over all deterministic action sequences would establish that,
for any A (deterministic or stochastic),
A
a
a
kp
)
≥
ǫ
θ
≤ |A|t (M + 1)|A| e−Mǫ
)
≥
ǫ|θ
≤
P
max
d
(p̂
P dKL (p̂A
kp
KL t
t
t
t
t
a∈A
A
However, our proof takes advantage of the fact that, for any deterministic A, pA
t and p̂t do not
depend on the ordering of past actions and observations. To make it precise, we encode the sequence
of actions in terms of action counts c0 , . . . , cT −1 . In particular, let ct,a = |{τ ≤ t : Aτ = a}| be
the number of times that action a has been selected by time t. We apply a coupling argument
that introduces dependencies between the noise terms Wt and action counts, without changing the
distributions of any observable variables. We let (Zn,a : n ∈ N, a ∈ A) be i.i.d. N (0, 1) random
variables, and let Wt+1 = Zct,At ,At . Similarly, we let (Z̃n,a,m : n ∈ N, a ∈ A, m = 1, . . . , M )
be i.i.d N (0, 1) random variables, and let W̃t+1,m = Z̃ct,At ,At ,m . To make explicit the dependence
on A, we will use a superscript and write cA
t to denote the action counts at time t when the action
process is given by A. It is not hard to verify, as is done in the appendix, that if a, a ∈ AT are two
deterministic action sequences such that cat−1 = cat−1 , then pat = pat and p̂at = p̂at . This allows us
to apply the union bound over action counts, instead of action sequences, and we get that for any A
(deterministic or stochastic),
!
A
A
a
a
P dKL (p̂t kpt ) ≥ ǫ|θ ≤ P a max t dKL (p̂t kpt ) ≥ ǫ θ ≤ (t + 1)|A| (M + 1)|A| e−Mǫ .
ct−1 :a∈A
Now, we specialize the action process A to the action sequence At = AES
t selected by ensemble
A
sampling, and we will omit the superscripts in pA
and
p̂
.
We
can
decompose
the per-period regret
t
t
of ensemble sampling as
E R∗ − θ⊤ At |θ = E (R∗ − θ⊤ At )I (dKL (p̂t kpt ) ≥ ǫ) |θ
+ E (R∗ − θ⊤ At )I (dKL (p̂t kpt ) < ǫ) |θ .
(2)
The first term can be bounded by
E (R∗ − θ⊤ At )I (dKL (p̂t kpt ) ≥ ǫ) |θ
≤
≤
∆(θ)P (dKL (p̂t kpt ) ≥ ǫ|θ)
∆(θ)(t + 1)|A| (M + 1)|A| e−Mǫ .
To bound the second term, we will use another coupling argument that couples the actions that would
be selected by ensemble sampling with those that would be selected by Thompson sampling. Let
ATS
denote the action√that Thompson sampling would select at time t. On {dKL (p̂t kpt ) ≤ ǫ}, we
t
have kp̂t − pt kTV ≤ 2ǫ by Pinsker’s inequality. Conditioning on p̂t and pt , if dKL (p̂t kpt ) ≤ ǫ,
TS
ES
we can construct random variables ÃES
t and Ãt such that they have the same distributions as At
ES
TS
and ATS
t , respectively. Using maximal coupling, we can make Ãt = Ãt with probability at least
5
p
1 − 21 kp̂t − pt kTV ≥ 1 − ǫ/2. Then, the second term of the sum in (2) can be decomposed into
E (R∗ − θ⊤ At )I (dKL (p̂t kpt ) ≤ ǫ) |θ
h h
i i
ES
TS
= E E (R∗ − θ⊤ ÃES
p̂t , pt , θ θ
t )I dKL (p̂t kpt ) ≤ ǫ, Ãt = Ãt
i i
h h
ES
TS
p̂
,
p
,
θ
θ ,
+E E (R∗ − θ⊤ ÃES
)I
d
(p̂
kp
)
≤
ǫ,
Ã
=
6
Ã
t
t
KL
t
t
t
t
t
which, after some algebraic manipulations, leads to
p
ǫ/2 ∆(θ).
E (R∗ − θ⊤ At )I (dKL (p̂t kpt ) < ǫ) |θ ≤ E R∗ − θ⊤ ATS
t |θ +
The result then follows from some straightforward algebra.
5 Computational results
In this section, we present computational results that demonstrate viability of ensemble sampling.
We will start with a simple case of independent Gaussian bandits in Section 5.1 and move on to
more complex models of neural networks in Section 5.2. Section 5.1 serves as a sanity check for
the empirical performance of ensemble sampling, as Thompson sampling can be efficiently applied
in this case and we are able to compare the performances of these two algorithms. In addition, we
provide simulation results that demonstrate how the ensemble size grows with the number of actions.
Section 5.2 goes beyond our theoretical analysis in Section 4 and gives computational evidence of the
efficacy of ensemble sampling when applied to more complex models such as neural networks. We
show that ensemble sampling, even with a few models, achieves efficient learning and outperforms
ǫ-greedy and dropout on the example neural networks.
5.1 Gaussian bandits with independent arms
We consider a Gaussian bandit with K actions, where action k has mean reward θk . Each θk is drawn
i.i.d. from N (0, 1). During each time step t = 0, . . . , T − 1, we select an action k ∈ {1, . . . , K}
and observe reward Rt+1 = θk + Wt+1 , where Wt+1 ∼ N (0, 1). Note that this is a special case of
Example 1. Since the posterior distribution of θ can be explicitly computed in this case, we use it as
a sanity check for the performance of ensemble sampling.
Figure 1a shows the per-period regret of Thompson sampling and ensemble sampling applied to a
Gaussian bandit with 50 independent arms. We see that as the number of models increases, ensemble
sampling better approximates Thompson sampling. The results were averaged over 2,000 realizations. Figure 1b shows the minimum number of models required so that the expected per-period
regret of ensemble sampling is no more than ǫ plus the expected per-period regret of Thompson
sampling at some large time horizon T across different numbers of actions. All results are averaged
over 10,000 realizations. We chose T = 2000 and ǫ = 0.03. The plot shows that the number of
models needed seems to grow sublinearly with the number of actions, which is stronger than the
bound proved in Section 4.
5.2 Neural networks
In this section, we follow Example 2 and show computational results of ensemble sampling applied
to neural networks. Figure 2 shows ǫ-greedy and ensemble sampling applied to a bandit problem
where the mapping from actions to expected rewards is represented by a neuron. More specifically,
we have a set of K actions A ⊆ ℜN . The mean reward of selecting an action a ∈ A is given by
gθ (a) = max(0, θ⊤ a), where weights θ ∈ ℜN are drawn from N (0, λI). During each time period,
we select an action At ∈ A and observe reward Rt+1 = gθ (At ) + Zt+1 , where Zt+1 ∼ N (0, σz2 ).
We set the input dimension N = 100, number of actions K = 100, prior variance λ = 10, and noise
variance σz2 = 100. Each dimension of each action was sampled uniformly from [−1, 1], except for
the last dimension, which was set to 1.
In Figure 3, we consider a bandit problem where the mapping from actions to expected rewards is
represented by a two-layer neural network with weights θ ≡ (W1 , W2 ), where W1 ∈ ℜD×N and
W2 ∈ ℜD . Each entry of the weight matrices is drawn independently from N (0, λ). There is a set of
K actions A ⊆ ℜN . The mean reward of choosing an action a ∈ A is gθ (a) = W2⊤ max(0, W1 a).
6
Ensemble sampling on an independent Gaussian bandit with 50 arms
80
Number of models
2.0
Per-period regret
Ensemble sampling on independent Gaussian bandits
100
Thompson sampling
5 models
10 models
20 models
30 models
1.5
1.0
0.5
0.0
60
40
20
0
100
200
300
400
t
500
600
0
700
(a)
0
25
50
75
100
125
Number of actions
150
175
200
(b)
Figure 1: (a) Ensemble sampling compared with Thompson sampling on a Gaussian bandit with 50
independent arms. (b) Minimum number of models required so that the expected per-period regret
of ensemble sampling is no more than ǫ = 0.03 plus the expected per-period regret of Thompson
sampling at T = 2000 for Gaussian bandits across different numbers of arms.
During each time period, we select an action At ∈ A and observe reward Rt+1 = gθ (At ) + Zt+1 ,
where Zt+1 ∼ N (0, σz2 ). We used N = 100 for the input dimension, D = 50 for the dimension of
the hidden layer, number of actions K = 100, prior variance λ = 1, and noise variance σz2 = 100.
Each dimension of each action was sampled uniformly from [−1, 1], except for the last dimension,
which was set to 1.
Ensemble sampling with M models starts by sampling θ̃m from the prior distribution independently
for each model m. At each time step, we pick a model m uniformly at random and apply the greedy
action with respect to that model. We update the ensemble incrementally. During each time period,
we apply a few steps of stochastic gradient descent for each model m with respect to the loss function
Lt (θ) =
t−1
1 X
1
(Rτ +1 + Z̃τ +1,m − gθ (Aτ ))2 + kθ − θ̃m k22 ,
2
σz τ =0
λ
where perturbations (Z̃t,m : t = 1, . . . , T, m = 1, . . . , M ) are drawn i.i.d. from N (0, σz2 ).
Besides ensemble sampling, there are other heuristics for sampling from an approximate posterior
distribution over neural networks, which may be used to develop approximate Thompson sampling.
Gal and Ghahramani proposed an approach based on dropout [4] to approximately sample from a
posterior over neural networks. In Figure 3, we include results from using dropout to approximate
Thompson sampling on the two-layer neural network bandit.
To facilitate gradient flow, we used leaky ReLUs of the form max(0.01x, x) internally in all agents,
while the target neural nets still use regular ReLUs as described above. We took 3 stochastic gradient
steps with a minibatch size of 64 for each model update. We used a learning rate of 1e-1 for ǫgreedy and ensemble sampling, and a learning rate of 1e-2, 1e-2, 2e-2, and 5e-2 for dropout with
dropping probabilities 0.25, 0.5, 0.75, and 0.9 respectively. All results were averaged over around
1,000 realizations.
Figure 2 plots the per-period regret of ǫ-greedy and ensemble sampling on the single neuron bandit. We see that ensemble sampling, even with 10 models, performs better than ǫ-greedy with the
best tuned parameters. Increasing the size of the ensemble further improves the performance. An
ensemble of size 50 achieves orders of magnitude lower regret than ǫ-greedy.
Figure 3a and 3b show different versions of ǫ-greedy applied to the two-layer neural network model.
We see that ǫ-greedy with an annealing schedule tends to perform better than a fixed ǫ. Figure
3c plots the per-period regret of the dropout approach with different dropping probabilities, which
seems to perform worse than ǫ-greedy. Figure 3d plots the per-period regret of ensemble sampling
on the neural net bandit. Again, we see that ensemble sampling, with a moderate number of models,
outperforms the other approaches by a significant amount.
7
(a) Epsilon-greedy
40
(b) Ensemble sampling
agent name
epsilon=0.05
instant regret
30
epsilon=0.1
epsilon=0.2
epsilon=50/(50+t)
20
epsilon=150/(150+t)
epsilon=300/(300+t)
ensemble=5
10
ensemble=10
ensemble=30
ensemble=50
0
0
500
1000
1500
2000
t
0
500
1000
1500
2000
Figure 2: (a) ǫ-greedy and (b) ensemble sampling applied to a single neuron bandit.
(a) Fixed epsilon
(b) Annealing epsilon
60
agent name
40
epsilon=0.05
epsilon=0.1
epsilon=0.2
epsilon=0.3
epsilon=10/(10+t)
epsilon=30/(30+t)
epsilon=50/(50+t)
epsilon=70/(70+t)
dropout=0.25
dropout=0.5
dropout=0.75
dropout=0.9
ensemble=5
ensemble=10
ensemble=30
ensemble=50
instant regret
20
0
(c) Dropout
(d) Ensemble sampling
60
40
20
0
0
100
200
300
400
500
t
0
100
200
300
400
500
Figure 3: (a) Fixed ǫ-greedy, (b) annealing ǫ-greedy, (c) dropout, and (d) ensemble sampling applied
to a two-layer neural network bandit.
6 Conclusion
Ensemble sampling offers a potentially efficient means to approximate Thompson sampling when
using complex models such as neural networks. We have provided an analysis that offers theoretical
assurances for the case of linear bandit models and computational results that demonstrate efficacy
with complex neural network models.
We are motivated largely by the need for effective exploration methods that can efficiently be applied in conjunction with complex models such as neural networks. Ensemble sampling offers one
approach to representing uncertainty in neural network models, and there are others that might also
be brought to bear in developing approximate versions of Thompson sampling [1, 4]. The analysis
of various other forms of approximate Thompson sampling remains open.
Ensemble sampling loosely relates to ensemble learning methods [3], though an important difference in motivation lies in the fact that the latter learns multiple models for the purpose of generating
a more accurate model through their combination, while the former learns multiple models to reflect uncertainty in the posterior distribution over models. That said, combining the two related
approaches may be fruitful. In particular, there may be practical benefit to learning many forms
8
of models (neural networks, tree-based models, etc.) and viewing the ensemble as representing
uncertainty from which one can sample.
Acknowledgments
This work was generously supported by a research grant from Boeing and a Marketing Research
Award from Adobe.
References
[1] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty
in neural networks. In Proceedings of the 32Nd International Conference on International
Conference on Machine Learning - Volume 37, ICML’15, pages 1613–1622. JMLR.org, 2015.
[2] Olivier Chapelle and Lihong Li. An empirical evaluation of Thompson sampling. In J. ShaweTaylor, R. S. Zemel, P. L. Bartlett, F. Pereira, and K. Q. Weinberger, editors, Advances in Neural
Information Processing Systems 24, pages 2249–2257. Curran Associates, Inc., 2011.
[3] Thomas G Dietterich. Ensemble learning. The handbook of brain theory and neural networks,
2:110–125, 2002.
[4] Yarin Gal and Zoubin Ghahramani. Dropout as a Bayesian approximation: Representing model
uncertainty in deep learning. In Maria Florina Balcan and Kilian Q. Weinberger, editors, Proceedings of The 33rd International Conference on Machine Learning, volume 48 of Proceedings
of Machine Learning Research, pages 1050–1059, New York, New York, USA, 20–22 Jun 2016.
PMLR.
[5] Carlos Gómez-Uribe. Online algorithms for parameter mean and variance estimation in dynamic
regression. arXiv preprint arXiv:1605.05697v1, 2016.
[6] Ian Osband, Charles Blundell, Alexander Pritzel, and Benjamin Van Roy. Deep exploration
via bootstrapped DQN. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems 29, pages 4026–4034. Curran
Associates, Inc., 2016.
[7] George Papandreou and Alan L Yuille. Gaussian sampling by local perturbations. In J. D.
Lafferty, C. K. I. Williams, J. Shawe-Taylor, R. S. Zemel, and A. Culotta, editors, Advances in
Neural Information Processing Systems 23, pages 1858–1866. Curran Associates, Inc., 2010.
[8] W.R. Thompson. On the likelihood that one unknown probability exceeds another in view of
the evidence of two samples. Biometrika, 25(3/4):285–294, 1933.
A
Proof of Theorem 3
2
= 1. Recall from Section 3 the procedure for generWithout loss of generality, assume that σw
ating and updating models with ensemble sampling. First, θ̃0,1 , . . . , θ̃0,M are sampled i.i.d. from
N (µ0 , Σ0 ). Then, these vectors are adapted according to
θ̃t,m = arg min
ν
t−1
X
(Rτ +1 + W̃τ +1,m −
τ =0
2
A⊤
τ ν)
⊤
+ (ν − θ̃0,m )
Σ−1
0 (ν
!
− θ̃0,m ) .
Note that we have not yet specified how actions are selected. In the formulation we have put forth,
each At could be any Ft -measureable random variable. We will denote by A the Ft -adapted process
(A0 , . . . , AT −1 ). We say A is deterministic if there exist a0 , . . . , aT −1 ∈ A such that At = at for
t = 0, . . . , T − 1 with probability one.
Lemma 4. If A is deterministic, then, conditioned on (R1 , . . . , Rt ), θ̃t,1 , · · · , θ̃t,M are i.i.d.
N (µt , Σt ) random variables, where µt = E[θ|Ft ] and Σt = E[(θ − µt )(θ − µt )⊤ |Ft ].
9
Proof. Say A = (a0 , . . . , aT −1 ), where a0 , . . . , aT −1 ∈ A. Let X be an t × N matrix with the j th
row equal to aj−1 . Let y = (R1 , . . . , Rt )⊤ . Then,
!
t−1
X
⊤ 2
⊤ −1
(Rτ +1 − aτ ν) + (ν − µ0 ) Σ0 (ν − µ0 )
µt = arg min
ν
τ =0
= X X + Σ−1
0
⊤
and
−1
X ⊤ y + Σ−1
0 µ0 ,
Σt = X ⊤ X + Σ−1
0
For any m = 1, . . . , M , we have
θ̃t,m = arg min
ν
t−1
X
(Rτ +1 + W̃τ +1,m −
τ =0
= X ⊤ X + Σ−1
0
−1
2
a⊤
τ ν)
.
⊤
+ (ν − θ̃0,m )
Σ−1
0 (ν
!
− θ̃0,m )
−1 ⊤
X (y + W̃m ) + Σ−1
0 θ̃0,m ,
where W̃m = (W̃1,m , . . . , W̃t,m )⊤ . We first observe that, conditioned on y, θ̃t,m follows a normal
distribution, since it is affine in θ̃0,m and W̃m . Next, we check its mean and covariance. Since W̃m
and θ̃0,m are independently sampled, we have
h
i
h
h
i
i
−1 ⊤
y + E W̃m y + Σ−1
X
E θ̃t,m y = X ⊤ X + Σ0−1
= µt ,
0 E θ̃0,m y
and
h
i
i
−1 ⊤ h
⊤
X E W̃m W̃m
Cov θ̃t,m y = X ⊤ X + Σ0−1
y X
h
i
−1
−1
⊤
= Σt .
X ⊤ X + Σ−1
+ Σ−1
E
(
θ̃
−
µ
)(
θ̃
−
µ
)
y
Σ
0,m
0
0,m
0
0
0
0
Therefore, if A is deterministic, then for each m = 1, . . . , M , θ̃t,m is a N (µt , Σt ) random variable
conditioned on (R1 , . . . , Rt ). Further, since W̃m and θ̃0,m , m = 1, . . . , M are all independent,
θ̃t,1 , . . . , θ̃t,M are independent.
Recall that pt (a) denotes the posterior probability Pt (A∗ = a) = P (A∗ = a|A0 , R1 , . . . , At−1 , Rt ).
To explicitly indicate dependence on the action process, we will use a superscript:
pt (a) = pA
t (a).
P
M
1
⊤
A
A
A
Let p̂t denote an approximation to pt , given by p̂t (a) = M m=1 I a = arg maxa′ θ̃t,m a′ .
Note that given an action process A, at time t Thompson sampling would sample the next action
A
from pA
t , while ensemble sampling would sample the next action from p̂t .
The following lemma shows that for any deterministic action sequence, conditioned on θ, the action distribution that ensemble sampling would sample from is close to the action distribution that
Thompson sampling would sample from with high probability.
Lemma 5. For any deterministic action sequence a ∈ AT ,
P (dKL (p̂at kpat ) ≥ ǫ | θ) ≤ (M + 1)|A| e−Mǫ .
Proof. If a ∈ AT is deterministic, then conditioned on (R1 , . . . , Rt ), p̂at and pat are independent of
θ. Thus, we have
P (dKL (p̂at kpat ) ≥ ǫ | θ) = E [E [I (dKL (p̂at kpat ) ≥ ǫ) |R1 , . . . , Rt ] | θ] .
By Lemma 4, θ̃t,1 , . . . , θ̃t,M , conditioned on (R1 , . . . , Rt ), are i.i.d. and distributed as the posterior
of θ. Thus, p̂at represents an empirical distribution of samples drawn from pat . Sanov’s Theorem
then implies that
P (dKL (p̂at kpat ) ≥ ǫ|R1 , . . . , Rt ) ≤ (M + 1)|A| e−Mǫ .
The result follows.
Next, we will establish results for any Ft -adapted action process A, deterministic or stochastic. To
do so, it is useful to introduce the notion of action counts. One way of encoding the sequence of
10
actions is in terms of counts c0 , . . . , cT −1 . In particular, let ct,a = |{τ ≤ t : Aτ = a}| be the
number of times that action a has been selected by time t. Each ct takes values in a set
(
)
X
|A|
Ct = c ∈ N :
ca = t + 1 .
a∈A
Since ct has |A| components, and each component takes a value in {0, . . . , t + 1}, we have
|Ct | ≤ (t + 2)|A| .
Sometimes, we use a superscript and write cA
t to explicitly denote the dependence on action process
A.
We now introduce dependencies between the noise terms Wt and action counts, without changing
the distributions of any observable variables. This will turn out to be useful when we take the
union bound later. We let (Zn,a : n ∈ N, a ∈ A) be i.i.d. N (0, 1) random variables, and let
Wt+1 = Zct,At ,At . Similarly, we let (Z̃n,a,m : n ∈ N, a ∈ A, m = 1, . . . , M ) be i.i.d N (0, 1)
random variables, and let W̃t+1,m = Z̃ct,At ,At ,m .
The following lemma establishes that, for any deterministic action sequence a ∈ AT , pat and p̂at
depend on a only through its action counts, cat−1 ; in other words, pat and p̂at do not depend on the
ordering of past actions and observations.
Lemma 6. For any t = 0, . . . , T −1, if a, a ∈ AT are deterministic sequences such that cat−1 = cat−1 ,
then pat = pat and p̂at = p̂at .
Proof. Recall that Rτ +1 = θ⊤ Aτ + Wτ +1 , where Wτ +1 = Zcτ,Aτ ,Aτ . This means that we observe
the same reward the first time we take an action, regardless of where that action appears in the action
sequence. Similarly, for all action sequences, we observe the same reward the second time we take
that action, and so on. Therefore, if cat−1 = cat−1 , we have µat = µat and Σat = Σat , which implies
that pat = pat . By the same reasoning, since W̃τ +1,m = Z̃cτ,Aτ ,Aτ ,m for all τ and m, both action
sequences would yield the same model parameters, and it follows that p̂at = p̂at .
Lemma 7. For any Ft -adapted process A,
A
|A|
P dKL (p̂A
(M + 1)|A| e−Mǫ .
t kpt ) > ǫ | θ ≤ (t + 1)
Proof. We have
P
A
dKL (p̂A
t kpt )
> ǫ|θ
≤
(a)
=
(b)
≤
(c)
≤
P
max
a∈AT
dKL (p̂at kpat )
max
P
T
ca
t−1 :a∈A
X
T
ca
t−1 :a∈A
>ǫ θ
dKL (p̂at kpat )
>ǫ θ
!
P (dKL (p̂at kpat ) > ǫ | θ)
(t + 1)|A| (M + 1)|A| e−Mǫ ,
where (a) follows from Lemma 6, (b) follows from the union bound, and (c) follows from Lemma
5 and the fact that the total number of counts |Ct−1 | ≤ (t + 1)|A| .
Now, we specialize the action process A to the action sequence At = AES
t selected by ensemble
A
sampling, and we will omit the superscripts in pA
t and p̂t .
11
The expected cumulative regret of ensemble sampling conditioned on θ can be decomposed as
Regret(T, π ES , θ) =
T
−1
X
t=0
=
E R∗ − θ⊤ AES
t |θ
T
−1
X
t=0
E (R∗ − θ⊤ AES
t )I (dKL (p̂t kpt ) > ǫ) |θ
.
+ E (R∗ − θ⊤ AES
t )I (dKL (p̂t kpt ) ≤ ǫ) |θ
We will bound the per-period regret for the case where the divergence dKL (p̂t kpt ) is large and the
case where the divergence is small, respectively.
Lemma 8. For any t = 0, . . . , T − 1,
|A|
E (R∗ − θ⊤ AES
(M + 1)|A| e−Mǫ ∆(θ).
t )I (dKL (p̂t kpt ) > ǫ) |θ ≤ (t + 1)
Proof. This follows directly from the definition of ∆(θ) and Lemma 7.
Assumption 9. For simplicity, assume 0 < ǫ < 1 and 0 < δ < 1 are such that
|A|T
ǫδ
≥ 9.
Lemma 10. If the size of the ensemble satisfies
M≥
then
T
−1
X
t=0
|A|T
2|A|
log
,
ǫ
ǫδ
E (R∗ − θ⊤ AES
t )I (dKL (p̂t kpt ) > ǫ) |θ ≤ δ∆(θ)T.
Proof. We show that if M satisfies the condition above, then
T |A| (M + 1)|A| e−Mǫ ≤ δ,
or, equivalently,
1
|A|
log(M + 1) ≥
M−
ǫ
ǫ
We have
1
|A| log T + log
.
δ
|A|
1
1
M−
|A| log T + log
log(M + 1) −
ǫ
ǫ
δ
2|A|
|A|
|A|T
|A|
|A|T
T
4|A|
≥
−
log
−
log
log
log
ǫ
ǫδ
ǫ
ǫ
ǫδ
ǫ
δ
|A|
|A|T
|A| |A|T
4|A|
|A|
−
log
·
log
log
=
ǫ
ǫ
ǫδ
ǫ
ǫ
ǫδ
≥ 0,
since Assumption 9 implies that
|A|T
ǫδ
≥ 4 log |A|T
ǫδ . The result then follows from Lemma 8.
Lemma 11. Let π TS denote the Thompson sampling policy. We have
T
−1
X
t=0
p
TS
E (R∗ − θ⊤ AES
, θ) + ǫ/2 ∆(θ)T.
t )I (dKL (p̂t kpt ) ≤ ǫ) |θ ≤ Regret(T, π
Proof. We apply a coupling argument that couples the actions that would be selected by ensemble
sampling with those that would be selected by Thompson sampling. Let ATS
t denote the action that
Thompson sampling would select at time t. On {dKL (p̂t kpt ) ≤ ǫ}, Pinsker’s inequality implies that
√
kp̂t − pt kTV ≤ 2ǫ.
12
TS
Conditioning on p̂t and pt , if dKL (p̂t kpt ) ≤ ǫ, we construct random variables ÃES
such
t and Ãt
ES
TS
that they have the same distribution as At and At , respectively. Using
maximal
coupling,
we
can
p
1
TS
make ÃES
=
Ã
with
probability
at
least
1
−
kp̂
−
p
k
≥
1
−
ǫ/2.
Then,
t
t
TV
t
t
2
∗
⊤ ES
E (R − θ At )I (dKL (p̂t kpt ) ≤ ǫ) |θ
= E E (R∗ − θ⊤ AES
t )I (dKL (p̂t kpt ) ≤ ǫ) p̂t , pt , θ θ
h h
i i
= E E (R∗ − θ⊤ ÃES
t )I (dKL (p̂t kpt ) ≤ ǫ) p̂t , pt , θ θ
i i
h h
ES
TS
p̂t , pt , θ θ
= E E (R∗ − θ⊤ ÃES
t )I dKL (p̂t kpt ) ≤ ǫ, Ãt = Ãt
i i
h h
ES
TS
p̂t , pt , θ θ
+E E (R∗ − θ⊤ ÃES
t )I dKL (p̂t kpt ) ≤ ǫ, Ãt 6= Ãt
On the first part of the sum, we have
i
h h
ES
TS
p̂t , pt , θ
E E (R∗ − θ⊤ ÃES
t )I dKL (p̂t kpt ) ≤ ǫ, Ãt = Ãt
i
h h
ES
TS
p̂
,
p
,
θ
=
Ã
= E E (R∗ − θ⊤ ÃTS
)I
d
(p̂
kp
)
≤
ǫ,
Ã
t
t
KL
t
t
t
t
t
h h
i i
∗
⊤ TS
≤ E E (R − θ Ãt ) p̂t , pt , θ θ
= E E (R∗ − θ⊤ ATS
t ) p̂t , pt , θ θ
= E R∗ − θ⊤ ATS
t |θ ,
θ
θ
i
i
where the inequality follows from the nonnegativity of R∗ − θ⊤ ÃTS
t . On the second part of the sum,
we have
i i
h h
ES
TS
p̂
,
p
,
θ
E E (R∗ − θ⊤ ÃES
)I
d
(p̂
kp
)
≤
ǫ,
Ã
=
6
Ã
θ
t t
KL t
t
t
t
t
h h
i i
TS
≤ E E (R∗ − R∗ )I dKL (p̂t kpt ) ≤ ǫ, ÃES
p̂t , pt , θ θ
t 6= Ãt
i i
h h
TS
p̂
,
p
,
θ
θ
= ∆(θ)E E I dKL (p̂t kpt ) ≤ ǫ, ÃES
=
6
Ã
t
t
t
t
p
ǫ/2 ∆(θ),
≤
ES
where the last inequality follows from the way we couple ÃES
t and Ãt . Thus, the result follows.
Combining Lemma 10 and Lemma 11 delivers a proof for our main result. In particular, we have
Regret(T, π ES , θ)
=
T
−1
X
t=0
+
2
E (θ⊤ A∗ − θ⊤ AES
t )I dKL (p̂t kpt ) > ǫ /2 |θ
T
−1
X
t=0
≤
=
2
E (θ⊤ A∗ − θ⊤ AES
t )I dKL (p̂t kpt ) ≤ ǫ /2 |θ
1
1
ǫ∆(θ)T + Regret(T, π TS , θ) + ǫ∆(θ)T
2
2
Regret(T, π TS , θ) + ǫ∆(θ)T,
where the inequality follows from Lemma 10, with δ = ǫ/2, and Lemma 11.
13
| 2cs.AI
|
Published as a conference paper at ICLR 2017
D EEP B IAFFINE ATTENTION FOR N EURAL
D EPENDENCY PARSING
Christopher D. Manning
Stanford University
[email protected]
arXiv:1611.01734v3 [cs.CL] 10 Mar 2017
Timothy Dozat
Stanford University
[email protected]
A BSTRACT
This paper builds off recent work from Kiperwasser & Goldberg (2016) using neural attention in a simple graph-based dependency parser. We use a larger but more
thoroughly regularized parser than other recent BiLSTM-based approaches, with
biaffine classifiers to predict arcs and labels. Our parser gets state of the art or
near state of the art performance on standard treebanks for six different languages,
achieving 95.7% UAS and 94.1% LAS on the most popular English PTB dataset.
This makes it the highest-performing graph-based parser on this benchmark—
outperforming Kiperwasser & Goldberg (2016) by 1.8% and 2.2%—and comparable to the highest performing transition-based parser (Kuncoro et al., 2016),
which achieves 95.8% UAS and 94.6% LAS. We also show which hyperparameter
choices had a significant effect on parsing accuracy, allowing us to achieve large
gains over other graph-based approaches.
1
I NTRODUCTION
Dependency parsers—which annotate sentences in a way designed to be easy for humans and computers alike to understand—have been found to be extremely useful for a sizable number of NLP
tasks, especially those involving natural language understanding in some way (Bowman et al., 2016;
Angeli et al., 2015; Levy & Goldberg, 2014; Toutanova et al., 2016; Parikh et al., 2015). However,
frequent incorrect parses can severely inhibit final performance, so improving the quality of dependency parsers is needed for the improvement and success of these downstream tasks.
The current state-of-the-art transition-based neural dependency parser (Kuncoro et al., 2016) substantially outperforms many much simpler neural graph-based parsers. We modify the neural graphbased approach first proposed by Kiperwasser & Goldberg (2016) in a few ways to achieve competitive performance: we build a network that’s larger but uses more regularization; we replace the
traditional MLP-based attention mechanism and affine label classifier with biaffine ones; and rather
than using the top recurrent states of the LSTM in the biaffine transformations, we first put them
through MLP operations that reduce their dimensionality. Furthermore, we compare models trained
with different architectures and hyperparameters to motivate our approach empirically. The resulting parser maintains most of the simplicity of neural graph-based approaches while approaching the
performance of the SOTA transition-based one.
2
BACKGROUND
AND
R ELATED WORK
Transition-based parsers—such as shift-reduce parsers—parse sentences from left to right, maintaining a “buffer” of words that have not yet been parsed and a “stack” of words whose head has not
been seen or whose dependents have not all been fully parsed. At each step, transition-based parsers
can access and manipulate the stack and buffer and assign arcs from one word to another. One can
then train any multi-class machine learning classifier on features extracted from the stack, buffer,
and previous arc actions in order to predict the next action.
Chen & Manning (2014) make the first successful attempt at incorporating deep learning into a
transition-based dependency parser. At each step, the (feedforward) network assigns a probability
to each action the parser can take based on word, tag, and label embeddings from certain words
1
Published as a conference paper at ICLR 2017
root
nsubj
dobj
root/ROOT Casey/NNP hugged/VBD Kim/NNP
Figure 1: A dependency tree parse for Casey hugged Kim, including part-of-speech tags and a special
root token. Directed edges (or arcs) with labels (or relations) connect the verb to the root and the
arguments to the verb head.
on the stack and buffer. A number of other researchers have attempted to address some limitations of Chen & Manning’s Chen & Manning parser by augmenting it with additional complexity:
Weiss et al. (2015) and Andor et al. (2016) augment it with a beam search and a conditional random
field loss objective to allow the parser to “undo” previous actions once it finds evidence that they may
have been incorrect; and Dyer et al. (2015) and (Kuncoro et al., 2016) instead use LSTMs to represent the stack and buffer, getting state-of-the-art performance by building in a way of composing
parsed phrases together.
Transition-based parsing processes a sentence sequentially to build up a parse tree one arc at a
time. Consequently, these parsers don’t use machine learning for directly predicting edges; they
use it for predicting the operations of the transition algorithm. Graph-based parsers, by contrast,
use machine learning to assign a weight or probability to each possible edge and then construct a
maximum spaning tree (MST) from these weighted edges. Kiperwasser & Goldberg (2016) present
a neural graph-based parser (in addition to a transition-based one) that uses the same kind of attention
mechanism as Bahdanau et al. (2014) for machine translation. In Kiperwasser & Goldberg’s 2016
model, the (bidirectional) LSTM’s recurrent output vector for each word is concatenated with each
possible head’s recurrent vector, and the result is used as input to an MLP that scores each resulting
arc. The predicted tree structure at training time is the one where each word depends on its highestscoring head. Labels are generated analogously, with each word’s recurrent output vector and its
gold or predicted head word’s recurrent vector being used in a multi-class MLP.
Similarly, Hashimoto et al. (2016) include a graph-based dependency parser in their multi-task neural model. In addition to training the model with multiple distinct objectives, they replace the traditional MLP-based attention mechanism that Kiperwasser & Goldberg (2016) use with a bilinear
one (but still using an MLP label classifier). This makes it analogous to Luong et al.’s 2015 proposed attention mechanism for neural machine translation. Cheng et al. (2016) likewise propose a
graph-based neural dependency parser, but in a way that attempts to circumvent the limitation of
other neural graph-based parsers being unable to condition the scores of each possible arc on previous parsing decisions. In addition to having one bidirectional recurrent network that computes a
recurrent hidden vector for each word, they have additional, unidirectional recurrent networks (leftto-right and right-to-left) that keep track of the probabilities of each previous arc, and use these
together to predict the scores for the next arc.
3
3.1
P ROPOSED D EPENDENCY PARSER
D EEP BIAFFINE
ATTENTION
We make a few modifications to the graph-based architectures of Kiperwasser & Goldberg (2016),
Hashimoto et al. (2016), and Cheng et al. (2016), shown in Figure 2: we use biaffine attention instead of bilinear or traditional MLP-based attention; we use a biaffine dependency label classifier;
and we apply dimension-reducing MLPs to each recurrent output vector ri before applying the biaffine transformation.1 The choice of biaffine rather than bilinear or MLP mechanisms makes the
classifiers in our model analogous to traditional affine classifiers, which use an affine transformation
over a single LSTM output state ri (or other vector input) to predict the vector of scores si for all
classes (1). We can think of the proposed biaffine attention mechanism as being a traditional affine
1
In this paper we follow the convention of using lowercase italic letters for scalars and indices, lowercase
bold letters for vectors, uppercase italic letters for matrices, uppercase bold letters for higher order tensors. We
also maintain this notation when indexing; so row i of matrix R would be represented as ri .
2
Published as a conference paper at ICLR 2017
H (arc-dep) ⊕ 1
U (arc)
H (arc-head)
S (arc)
⊤
1
1
·
·
=
1
1
(arc-dep)
MLP: hi
(arc-head)
, hi
...
BiLSTM: ri
Embeddings: xi
root
ROOT
Kim
NNP
Figure 2: BiLSTM with deep biaffine attention to score each possible head for each dependent,
applied to the sentence “Casey hugged Kim”. We reverse the order of the biaffine transformation
here for clarity.
classifier, but using a (d × d) linear transformation of the stacked LSTM output RU (1) in place of
the weight matrix W and a (d × 1) transformation Ru(2) for the bias term b (2).
si = W ri + b
Fixed-class affine classifier
(1)
(arc)
(2)
(1)
Variable-class biaffine classifier
(2)
ri + Ru
= RU
si
In addition to being arguably simpler than the MLP-based approach (involving one bilinear layer
rather than two linear layers and a nonlinearity), this has the conceptual advantage of directly mod(2)
eling both the prior probability of a word j receiving any dependents in the term r⊤
and the
j u
⊤ (1)
likelihood of j receiving a specific dependent i in the term rj U ri . Analogously, we also use a
biaffine classifier to predict dependency labels given the gold or predicted head yi (3).
(label)
(1)
Fixed-class biaffine classifier
(3)
ri + (ryi ⊕ ri )⊤ U (2) + b
= r⊤
si
yi U
This likewise directly models each of the prior probability of each class, the likelihood of a class
given just word i (how probable a word is to take a particular label), the likelihood of a class given
just the head word yi (how probable a word is to take dependents with a particular label), and the
likelihood of a class given both word i and its head (how probable a word is to take a particular label
given that word’s head).
Applying smaller MLPs to the recurrent output states before the biaffine classifier has the advantage
of stripping away information not relevant to the current decision. That is, every top recurrent state
ri will need to carry enough information to identify word i’s head, find all its dependents, exclude all
its non-dependents, assign itself the correct label, and assign all its dependents their correct labels, as
well as transfer any relevant information to the recurrent states of words before and after it. Thus ri
necessarily contains significantly more information than is needed to compute any individual score,
and training on this superfluous information needlessly reduces parsing speed and increases the risk
of overfitting. Reducing dimensionality and applying a nonlinearity (4 - 6) addresses both of these
problems. We call this a deep bilinear attention mechanism, as opposed to shallow bilinear attention,
which uses the recurrent states directly.
(arc-dep)
hi
(arc-head)
hj
(arc)
si
= MLP(arc-dep) (ri )
(4)
= MLP(arc-head) (rj )
(5)
=H
(arc-dep)
U (1) hi
(arc-head) (2)
(arc-head)
+H
(6)
u
We apply MLPs to the recurrent states before using them in the label classifier as well. As with other
graph-based models, the predicted tree at training time is the one where each word is a dependent of
its highest scoring head (although at test time we ensure that the parse is a well-formed tree via the
MST algorithm).
3
Published as a conference paper at ICLR 2017
3.2
H YPERPARAMETER CONFIGURATION
Param
Embedding size
LSTM size
Arc MLP size
Label MLP size
LSTM depth
α
Annealing
Value
100
400
500
100
3
2e−3
t
.75 5000
Param
Embedding dropout
LSTM dropout
Arc MLP dropout
Label MLP dropout
MLP depth
β1 ,β2
tmax
Value
33%
33%
33%
33%
1
.9
50,000
Table 1: Model hyperparameters
Aside from architectural differences between ours and the other graph-based parsers, we make a
number of hyperparameter choices that allow us to outperform theirs, laid out in Table 1. We use
100-dimensional uncased word vectors2 and POS tag vectors; three BiLSTM layers (400 dimensions
in each direction); and 500- and 100-dimensional ReLU MLP layers. We also apply dropout at every
stage of the model: we drop words and tags (independently); we drop nodes in the LSTM layers
(input and recurrent connections), applying the same dropout mask at every recurrent timestep (cf.
the Bayesian dropout of Gal & Ghahramani (2015)); and we drop nodes in the MLP layers and
classifiers, likewise applying the same dropout mask at every timestep. We optimize the network
with annealed Adam (Kingma & Ba, 2014) for about 50,000 steps, rounded up to the nearest epoch.
4
E XPERIMENTS & R ESULTS
4.1
DATASETS
We show test results for the proposed model on the English Penn Treebank, converted into Stanford
Dependencies using both version 3.3.0 and version 3.5.0 of the Stanford Dependency converter
(PTB-SD 3.3.0 and PTB-SD 3.5.0); the Chinese Penn Treebank; and the CoNLL 09 shared task
dataset,3 following standard practices for each dataset. We omit punctuation from evaluation only
for the PTB-SD and CTB. For the English PTB-SD datasets, we use POS tags generated from the
Stanford POS tagger (Toutanova et al., 2003); for the Chinese PTB dataset we use gold tags; and for
the CoNLL 09 dataset we use the provided predicted tags. Our hyperparameter search was done with
the PTB-SD 3.5.0 validation dataset in order to minimize overfitting to the more popular PTB-SD
3.3.0 benchmark, and in our hyperparameter analysis in the following section we report performance
on the PTB-SD 3.5.0 test set, shown in Tables 2 and 3.
4.2
4.2.1
H YPERPARAMETER CHOICES
ATTENTION MECHANISM
We examined the effect of different classifier architectures on accuracy and performance. What we
see is that the deep bilinear model outperforms the others with respect to both speed and accuracy.
The model with shallow bilinear arc and label classifiers gets the same unlabeled performance as the
deep model with the same settings, but because the label classifier is much larger ((801 × c × 801) as
opposed to (101 × c × 101)), it runs much slower and overfits. One way to decrease this overfitting
is by increasing the MLP dropout, but that of course doesn’t change parsing speed; another way is
to decrease the recurrent size to 300, but this hinders unlabeled accuracy without increasing parsing
speed up to the same levels as our deeper model. We also implemented the MLP-based approach
to attention and classification used in Kiperwasser & Goldberg (2016).4 We found this version to
2
We compute a “trained” embedding matrix composed of words that occur at least twice in the training
dataset and add these embeddings to their corresponding pretrained embeddings. Any words that don’t occur
in either embedding matrix are replaced with a separate OOV token.
3
We exclude the Japanese dataset from our evaluation because we do not have access to it.
4
In the version of TensorFlow we used, the model’s memory requirements during training exceeded the
available memory on a single GPU when default settings were used, so we reduced the MLP hidden size to 200
4
Published as a conference paper at ICLR 2017
Classifier
UAS
LAS
95.75
94.22
95.74
94.00*
95.73
94.05*
95.63* 93.86*
95.53* 93.91*
Sents/sec
410.91
298.99
300.04
373.24
367.44
Recurrent Cell
UAS
LAS
95.75
94.22
93.18* 91.08*
95.67
94.06*
Sents/sec
410.91
435.32
463.25
Model
Deep
Shallow
Shallow, 50% drop
Shallow, 300d
MLP
Model
LSTM
GRU
Cif-LSTM
Model
3 layers, 400d
3 layers, 300d
3 layers, 200d
2 layers, 400d
4 layers, 400d
Size
UAS
95.75
95.82
95.55*
95.62*
95.83
LAS
94.22
94.24
93.89*
93.98*
94.22
Sents/sec
410.91
460.01
469.45
497.99
362.09
Table 2: Test accuracy and speed on PTB-SD 3.5.0. Statistically significant differences are marked
with an asterisk.
Input Dropout
Model
UAS
Default
95.75
No word dropout 95.74
No tag dropout
95.28*
No tags
95.77
LAS
94.22
94.08*
93.60*
93.91*
Model
β2 = .9
β2 = .999
Adam
UAS
95.75
95.53*
LAS
94.22
93.91*
Table 3: Test Accuracy on PTB-SD 3.5.0. Statistically significant differences are marked with an
asterisk.
likewise be somewhat slower and significantly underperform the deep biaffine approach in both
labeled and unlabeled accuracy.
4.2.2
N ETWORK
SIZE
We also examine more closely how network size influences speed and accuracy.
In
Kiperwasser & Goldberg’s 2016 model, the network uses 2 layers of 125-dimensional bidirectional
LSTMs; in Hashimoto et al.’s 2016 model, it has one layer of 100-dimensional bidirectional LSTMs
dedicated to parsing (two lower layers are also trained on other objectives); and Cheng et al.’s 2016
model has one layer of 368-dimensional GRU cells. We find that using three or four layers gets
significantly better performance than two layers, and increasing the LSTM sizes from 200 to 300 or
400 dimensions likewise signficantly improves performance.5
4.2.3
R ECURRENT
CELL
GRU cells have been promoted as a faster and simpler alternative to LSTM cells, and are used in
the approach of Cheng et al. (2016); however, in our model they drastically underperformed LSTM
cells. We also implemented the coupled input-forget gate LSTM cells (Cif-LSTM) suggested by
Greff et al. (2015),6 finding that while the resulting model still slightly underperforms the more
popular LSTM cells, the difference between the two is much smaller. Additionally, because the
gate and candidate cell activations can be computed simultaneously with one matrix multiplication,
the Cif-LSTM model is faster than the GRU version even though they have the same number of
parameters. We hypothesize that the output gate in the Cif-LSTM model allows it to maintain a
sparse recurrent output state, which helps it adapt to the high levels of dropout needed to prevent
overfitting in a way that GRU cells are unable to do.
5
The model with 400-dimensional recurrent states significantly outperforms the 300-dimensional one on
the validation set, but not on the test set
6
In addition to using a coupled input-forget gate, we remove the first tanh nonlinearity, which is no longer
needed when using a coupled gate
5
Published as a conference paper at ICLR 2017
Type
Model
English PTB-SD 3.3.0
UAS
LAS
Chinese PTB 5.1
UAS
LAS
Transition
Ballesteros et al. (2016)
Andor et al. (2016)
Kuncoro et al. (2016)
93.56
94.61
95.8
91.42
92.79
94.6
87.65
–
–
86.21
–
–
Graph
Kiperwasser & Goldberg (2016)
Cheng et al. (2016)
Hashimoto et al. (2016)
Deep Biaffine
93.9
94.10
94.67
95.74
91.9
91.49
92.90
94.08
87.6
88.1
–
89.30
86.1
85.7
–
88.23
Table 4: Results on the English PTB and Chinese PTB parsing datasets
Model
Catalan
UAS
LAS
Chinese
UAS
LAS
Czech
UAS
LAS
Andor et al.
Deep Biaffine
92.67
94.69
84.72
88.90
88.94
92.08
Model
English
UAS
LAS
German
UAS
LAS
Spanish
UAS
LAS
Andor et al.
Deep Biaffine
93.22
95.21
90.91
93.46
92.62
94.34
89.83
92.02
91.23
93.20
80.85
85.38
89.15
91.44
84.56
87.38
89.95
91.65
Table 5: Results on the CoNLL ’09 shared task datasets
4.2.4
E MBEDDING D ROPOUT
Because we increase the parser’s power, we also have to increase its regularization. In addition to
using relatively extreme dropout in the recurrent and MLP layers mentioned in Table 1, we also
regularize the input layer. We drop 33% of words and 33% of tags during training: when one is
dropped the other is scaled by a factor of two to compensate, and when both are dropped together,
the model simply gets an input of zeros. Models trained with only word or tag dropout but not
both wind up signficantly overfitting, hindering label accuracy and—in the latter case—attachment
accuracy. Interestingly, not using any tags at all actually results in better performance than using
tags without dropout.
4.2.5
O PTIMIZER
We choose to optimize with Adam (Kingma & Ba, 2014), which (among other things) keeps a moving average of the L2 norm of the gradient for each parameter throughout training and divides the
gradient for each parameter by this moving average, ensuring that the magnitude of the gradients will
on average be close to one. However, we find that the value for β2 recommended by Kingma & Ba—
which controls the decay rate for this moving average—is too high for this task (and we suspect more
generally). When this value is very large, the magnitude of the current update is heavily influenced
by the larger magnitude of gradients very far in the past, with the effect that the optimizer can’t adapt
quickly to recent changes in the model. Thus we find that setting β2 to .9 instead of .999 makes a
large positive impact on final performance.
4.3
R ESULTS
Our model gets nearly the same UAS performance on PTB-SD 3.3.0 as the current SOTA model
from Kuncoro et al. (2016) in spite of its substantially simpler architecture, and gets SOTA UAS
performance on CTB 5.17 as well as SOTA performance on all CoNLL 09 languages. It is worth
noting that the CoNLL 09 datasets contain many non-projective dependencies, which are difficult
or impossible for transition-based—but not graph-based—parsers to predict. This may account for
some of the large, consistent difference between our model and Andor et al.’s 2016 transition-based
model applied to these datasets.
7
We’d like to thank Zhiyang Teng for finding a bug in the original code that affected the CTB 5.1 dataset
6
Published as a conference paper at ICLR 2017
Where our model appears to lag behind the SOTA model is in LAS, indicating one of a few possibilities. Firstly, it may be the result of inefficiencies or errors in the GloVe embeddings or POS tagger,
in which case using alternative pretrained embeddings or a more accurate tagger might improve
label classification. Secondly, the SOTA model is specifically designed to capture phrasal compositionality; so another possibility is that ours doesn’t capture this compositionality as effectively, and
that this results in a worse label score. Similarly, it may be the result of a more general limitation of
graph-based parsers, which have access to less explicit syntactic information than transition-based
parsers when making decisions. Addressing these latter two limitations would require a more innovative architecture than the relatively simple one used in current neural graph-based parsers.
5
C ONCLUSION
In this paper we proposed using a modified version of bilinear attention in a neural dependency
parser that increases parsing speed without hurting performance. We showed that our larger but more
regularized network outperforms other neural graph-based parsers and gets comparable performance
to the current SOTA transition-based parser. We also provided empirical motivation for the proposed
architecture and configuration over similar ones in the existing literature. Future work will involve
exploring ways of bridging the gap between labeled and unlabeled accuracy and augment the parser
with a smarter way of handling out-of-vocabulary tokens for morphologically richer languages.
R EFERENCES
Daniel Andor, Chris Alberti, David Weiss, Aliaksei Severyn, Alessandro Presta, Kuzman Ganchev, Slav Petrov, and Michael Collins.
Globally normalized transitionbased neural networks.
In Association for Computational Linguistics, 2016.
URL
https://arxiv.org/abs/1603.06042.
Gabor Angeli, Melvin Johnson Premkumar, and Christopher D Manning. Leveraging linguistic
structure for open domain information extraction. In Proceedings of the 53rd Annual Meeting of
the Association for Computational Linguistics (ACL 2015), 2015.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly
learning to align and translate. International Conference on Learning Representations, 2014.
Miguel Ballesteros, Yoav Goldberg, Chris Dyer, and Noah A Smith. Training with exploration
improves a greedy stack-LSTM parser. Proceedings of the conference on empirical methods in
natural language processing, 2016.
Samuel R Bowman, Jon Gauthier, Abhinav Rastogi, Raghav Gupta, Christopher D Manning, and
Christopher Potts. A fast unified model for parsing and sentence understanding. ACL 2016, 2016.
Danqi Chen and Christopher D Manning. A fast and accurate dependency parser using neural
networks. In Proceedings of the conference on empirical methods in natural language processing,
pp. 740–750, 2014.
Hao Cheng, Hao Fang, Xiaodong He, Jianfeng Gao, and Li Deng. Bi-directional attention with
agreement for dependency parsing. arXiv preprint arXiv:1608.02076, 2016.
Chris Dyer, Miguel Ballesteros, Wang Ling, Austin Matthews, and Noah A Smith. Transitionbased dependency parsing with stack long short-term memory. Proceedings of the conference on
empirical methods in natural language processing, 2015.
Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model
uncertainty in deep learning. International Conference on Machine Learning, 2015.
Klaus Greff, Rupesh Kumar Srivastava, Jan Koutnı́k, Bas R Steunebrink, and Jürgen Schmidhuber.
LSTM: A search space odyssey. IEEE Transactions on Neural Networks and Learning Systems,
2015.
7
Published as a conference paper at ICLR 2017
Kazuma Hashimoto, Caiming Xiong, Yoshimasa Tsuruoka, and Richard Socher. A joint many-task
model: Growing a neural network for multiple nlp tasks. arXiv preprint arXiv:1611.01587, 2016.
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International
Conference on Learning Representations, 2014.
Eliyahu Kiperwasser and Yoav Goldberg. Simple and accurate dependency parsing using bidirectional LSTM feature representations. Transactions of the Association for Computational Linguistics, 4:313–327, 2016.
Adhiguna Kuncoro, Miguel Ballesteros, Lingpeng Kong, Chris Dyer, Graham Neubig, and Noah A.
Smith. What do recurrent neural network grammars learn about syntax? CoRR, abs/1611.05774,
2016. URL http://arxiv.org/abs/1611.05774.
Omer Levy and Yoav Goldberg. Dependency-based word embeddings. In ACL 2014, pp. 302–308,
2014.
Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attentionbased neural machine translation. Empirical Methods in Natural Language Processing, 2015.
Ankur P Parikh, Hoifung Poon, and Kristina Toutanova. Grounded semantic parsing for complex
knowledge extraction. In Proceedings of North American Chapter of the Association for Computational Linguistics, pp. 756–766, 2015.
Kristina Toutanova, Dan Klein, Christopher D Manning, and Yoram Singer. Feature-rich part-ofspeech tagging with a cyclic dependency network. In Proceedings of the 2003 Conference of the
North American Chapter of the Association for Computational Linguistics on Human Language
Technology-Volume 1, pp. 173–180. Association for Computational Linguistics, 2003.
Kristina Toutanova, Xi Victoria Lin, and Wen-tau Yih. Compositional learning of embeddings for
relation paths in knowledge bases and text. In ACL, 2016.
David Weiss, Chris Alberti, Michael Collins, and Slav Petrov. Structured training for neural network
transition-based parsing. Annual Meeting of the Association for Computational Linguistics, 2015.
8
| 9cs.NE
|
1
Minimax Design of Nonlinear Phase FIR Filters
with Optimality Certificates
arXiv:1608.03161v1 [cs.SY] 10 Aug 2016
Sefa Demirtas
Abstract—The Parks-McClellan algorithm provides an efficient
method for designing a linear phase FIR filter with a pre-specified
weight function on the approximation error. For the given filter
order and the specified weight function, the filter designed with
this algorithm will have the unique optimal frequency response
that approximates a desired filter response as certified by the
alternation theorem. In this paper, a nonlinear phase FIR filter
design algorithm is provided that allows the specification of a
piecewise constant weight function on the approximation error
in an analogous manner to linear phase FIR filters. For the given
filter order and weight function, the resulting filter will provably
have the unique optimal magnitude response that approximates
a desired filter response, where the certification of optimality
is given and is also based on the alternations that the weighted
error function exhibits. Furthermore, the method is applicable to
designing filters with both real- and complex-valued coefficients,
which in turn determines the number of required alternations.
Index Terms—Nonlinear phase FIR design, minimax optimality, alternation theorem
I. I NTRODUCTION
Despite their desirable properties, FIR filters have certain
disadvantages when compared to their IIR counterparts. For
example, the minimum order required for an FIR filter to
approximate a desired filter response within pre-specified
bounds is usually much higher than the order of an IIR filter
for the same task, which translates into more multiplications
and additions per input sample in a hardware implementation
and requires larger power and memory. Furthermore, even
though linear phase FIR filters introduce no dispersion but
only a uniform latency, the amount of this latency in samples
is half the filter order and this may become unacceptably
high for latency-sensitive applications. Therefore it is crucial
to optimize an FIR filter for a given order while keeping
this order as low as possible. The optimality can be stated
with respect to a particular norm, and several design methods
exist to meet different optimality criteria. For example, the
windowing methods can be used to minimize the sum of
absolute squares (l2 -norm) of the approximation error between
the filter response and the desired response for a given order.
On the other hand, the Parks-McClellan method [1] aims to
minimize the maximum absolute error (l∞ -norm), also known
as the minimax or Chebyshev error. Since minimizing the l2 norm is equivalent to minimizing the energy in the error, it
does not prevent large but very narrow deviations between the
filter frequency response and the desired response that result
in a small l2 -norm. Therefore, even though it is much easier
to design for l2 -norm optimality, the minimax-optimality have
The author is with Analog Devices Lyric Labs. Email (work):
[email protected], (personal): [email protected].
been adopted widely in the signal processing community due
to its superior worst-case performance, and it will be the focus
in this paper.
The Parks-McClellan design method yields the unique
globally-optimal linear phase FIR filter for a given order
N by exploiting a theorem that provides the necessary and
sufficient conditions for global minimax-optimality, which is
known as the alternation theorem [2]. Even though it is
desirable, asserting linear phase restricts the coefficients of
the filter to have even- or odd-symmetry if they are realvalued, or conjugate-symmetry if they are complex-valued.
This restriction exhibits itself as halving the number of degrees
of freedom available to approximate the desired response
because each choice of a filter coefficient already determines
another coefficient to generate symmetric pairs. For a variety of applications where linear phase is not crucial, this
becomes an unnecessary constraint. Filters with magnitude
responses that provide a much better approximation to the
desired response can be obtained by removing the linear
phase constraint. However, this renders the characterization
of optimality by the alternation theorem inapplicable in its
currently known form, which in turn prevents direct utilization
of the Parks-McClellan method.
In this paper, we first provide a characterization method for
the global minimax-optimality of FIR filters if no restrictions
exist on its phase. In other words, we state the necessary and
sufficient conditions for the magnitude response of an FIR
filter to be the unique best approximation to a desired response.
Of course, due to the restrictions that are imposed in a linear
phase design, the optimal filter with unrestricted phase will
always be at least as good an approximation to the desired
response as the optimal linear phase solution. Since this is a
characterization stated in terms of the magnitude response, all
the FIR filters sharing the same order and magnitude response
will be optimal. Therefore, although the magnitude response
is the unique optimal response, there are a finite number of
distinct optimal FIR filters related to each other through a
cascade with an all-pass filter. Secondly, in this paper, the
arguments of the characterization will be shown to naturally
lead to a minimax-optimal design method involving the computation of an autocorrelation sequence as an intermediate
step, at the end of which the designer will be able to choose
from a variety of options for the phase including a minimum
phase and a maximum phase design without compromising
the global optimality of the magnitude response. Since it
was originally introduced in Hermann and Schuessler’s work
[3], designing a nonlinear phase FIR filter by first designing
an autocorrelation sequence and then finding a filter that
admits this as its autocorrelation is a widely known technique
2
[4]–[10]. Spectral factorization is an obvious first choice to
obtain the filter coefficients from the designed autocorrelation
sequence. This requires finding the roots of the polynomial
the coefficients of which are the same as the autocorrelation
sequence. Since this is a highly impractical approach for designing high order filters, for minimum phase designs, several
algorithms have been proposed as an alternative to polynomial
root finding [4]–[8]. However, in most of these earlier methods
for nonlinear phase FIR filter design, it is the autocorrelation
sequence of the filter impulse response that is designed to have
minimax optimality in approximating the desired response,
which does not necessarily imply the optimality of the filter
itself. Furthermore, since the magnitude response of the filter
and that of its autocorrelation sequence are related through
squaring, the weight function applied during the design of the
autocorrelation will not match the desired weight function on
the approximation error attained by the final design. In this
paper, we first characterize the optimality of the nonlinear
phase FIR filter instead of its autocorrelation sequence and
then provide a method to compute the correct weight to be
applied during the computation of the autocorrelation sequence
so that the resulting filter exhibits the desired ratio of passband
to stopband deviations. Spectral factorization using polynomial
root finding or any of the alternative methods in the current
literature [4]–[8] can then be used to obtain the final design
from the autocorrelation sequence leading to the FIR design
with the desired error weights, where the global optimality is
certified by our characterization of optimality.
II. L INEAR P HASE
VS
N ONLINEAR P HASE FIR F ILTERS
Linear phase FIR filters with real-valued coefficients can
be expressed as a real-valued amplitude function A(ω) multiplied by a linear phase term e−jαω . Since this linear phase
corresponds to a time delay by α samples, a usual approach
to designing linear phase FIR filters is to first design the zerophase filter response A(ω), which is necessarily symmetric
in time domain for it to be zero phase. Afterwards, the
filter is time delayed until it is causal, which corresponds to
multiplying with a phase of the form e−jαω .
Parks and McClellan [1] exploited the fact that the frequency response A(ω) of a symmetric zero-phase filter a[n]
of even order N can be expressed in terms of real sinusoids.
More specifically, for example if a[n] = a[−n], then
N/2
N/2
A(ω) =
X
−N/2
a[n]e
−jωn
= a[0] +
X
2a[n]cos(nω).
(1)
n=1
In other words, A(ω) can be expressed as a linear combination
of the basis functions {cos(nω), n = 0, 1, . . . , N/2}. In
order to approximate an ideal filter response D(ω), we search
for the optimal set of linear combination coefficients for
these basis functions such that the maximum absolute error
is minimized. The alternation theorem asserts that we need
N/2 + 2 alternations for the optimal filter, and the Remez
Exchange algorithm can be used to efficiently find this set of
coefficients.
Since nonlinear phase filters cannot be time-shifted by
any amount to exhibit a real-valued (zero-phase) response,
and since the alternation theorem and the Remez Exchange
Algorithm apply only to real-valued functions, they cannot
be used to characterize or design minimax optimal nonlinear
phase FIR filters directly.
III. C HARACTERIZATION T HEOREM FOR N ONLINEAR
P HASE F ILTERS
In this section, we describe a very straightforward method
to characterize the global minimax optimality of a given
nonlinear phase FIR filter h[n], n = 0, 1, . . . , N , where the
optimality is implied for the magnitude response of this filter,
|H(ejω )|, as compared to the desired filter response, D(ω),
which is unity in the passband and zero in the stopband. In
other words, by examining the magnitude response |H(ejω )|,
we will be able to tell whether this is a filter the magnitude
response of which is the best approximation to D(ω) in
that no other magnitude response achievable with the same
order can attain a smaller infinity norm on the weighted
approximation error W (ω)(|H(ejω )|−D(ω)). We will be able
to characterize optimality for both real-valued and complexvalued filter coefficients, and we do not require any symmetry
in the coefficients, therefore it applies in the most general case.
In the next section, the arguments of this characterization will
enable us to find an efficient algorithm to design nonlinear
phase FIR filters in cases where only the magnitude response is
specified and the phase is not restricted. However we will still
be able to choose among different available phase characteristics including, for example, a minimum phase design without
compromising global optimality with respect to magnitude.
Assume that an FIR filter with coefficients h[n], n =
0, 1, 2, . . . , N and a frequency response H(ejω ) is provided
with the passband ΩP and the stopband ΩS , both of which
are closed subsets of (−π, π], to approximate the desired
magnitude response
1, ω ∈ ΩP
D(ω) =
.
(2)
0, ω ∈ ΩS
Further assume that a desired weight function Wdes (ω) is
provided that expresses the relative emphasis on the error in
the stopband as compared to the passband. More specifically,
1,
ω ∈ ΩP
Wdes (ω) =
,
(3)
Kdes , ω ∈ ΩS
where Kdes is a scalar given as a part of the filter specifications. The weighted error function, and the bounds on
passband and stopband errors are defined as
EW (ω) = Wdes (ω) H(ejω ) − D(ω) ,
(4)
δP =
max
ω∈(ΩP ∪ΩS )
and
δS =
|EW (ω)|
δP
,
Kdes
(5)
(6)
respectively.
Theorem 1. |H(ejω )| is the unique minimax-optimal magnitude response that can be attained by any FIR filter of order
N in order to approximate the ideal filter magnitude response
3
D(ω) with a desired weight function Wdes (ω) if and only if
the adjusted weighted error function
′
′
EW
(ω) = Wdes
(ω) H(ejω ) − D′ (ω)
(7)
exhibits at least N +2 alternations if the filter coefficients h[n]
are restricted to be real-valued, or at least 2N +2 alternations
′
if they are not restricted to be real-valued. Here, Wdes
(ω) and
′
D (ω) are defined as
1,
ω ∈ ΩP
′
(ω) =
Wdes
,
(8)
2Kdes , ω ∈ ΩS
where the subscripts re and im refer to the real and imaginary
parts, respectively. The frequency response of a conjugatesymmetric autocorreation sequence p[n] can be represented as
jω
P (e ) =
N
X
p[n]e−jωn
n=−N
= p[0] +
= p[0] +
N
X
n=1
N
X
p[n]e−jωn + p∗ [n]ejωn
2pre [n]cos(nω) +
n=1
and
′
D (ω) =
1,
δS
2 ,
ω ∈ ΩP
.
ω ∈ ΩS
(9)
Note that in the context of linear phase FIR filters where the
alternation theorem is used for characterization, the optimality
is characterized by counting the alternations in the weighted
error function which is computed using the desired response
D(ω) and the weight function Wdes (ω). The characterization
given here for nonlinear phase FIR filters is based on the
′
adjusted weighted error function EW
computed using D′ (ω)
′
and Wdes (ω) as in equation (7).
The formal proof of the characterization theorem (Theorem
1) is given in [11] and will be excluded here for brevity.
However, intuitively, the optimality of h[n] can be related to
the optimality of its autocorrelation sequence p[n] as follows.
Due to the specific choice of values in the stopband for
′
D′ (ω) and Wdes
(ω), the number of alternations and the points
at which alternations occur are the same for |H(ejω )| and
P (ejω ) = |H(ejω )|2 . More specifically, if |H(ejω )| attains
its extreme value at a specific frequency and hence form
′
an alternation in EW
(ω), then the Fourier transform of the
autocorrelation function P (ejω ) = |H(ejω )|2 will also attain
its extremal value and form an alternation in a related weighted
error function at the same frequency. Therefore, the number
of required alternations in the magnitude response of the filter
can be related to that of the autocorrelation sequence, which
is zero-phase and in turn can be characterized for optimality
using the traditional form of the alternation theorem.
The number of required alternations for filters with
complex-valued coefficients are larger than that of filters with
real-valued coefficients. This also is consistent with intuition
because it reflects the additional degrees of freedom in choosing the filter coefficients by relaxing the constraint to be realvalued. More formally, the autocorrelation function of such
filters also have complex-valued coefficients in general, and
they exhibit conjugate-symmetry instead of the even-symmetry
in the real case. This means that the autocorrelation sequence
p[n] of a filter h[n] with complex-valued coefficients satisfy
p[n] = p∗ [−n], n = 0, 1, 2, . . . , N,
(10)
or equivalently
pre [n] = pre [−n], n = 0, 1, 2, . . . , N
(11)
pim [n] = −pim [−n], n = 0, 1, 2, . . . , N,
(12)
and
N
X
(13)
2pim [n]sin(nω)
n=1
This implies that the frequency response P (ejω ) can be
represented as a linear combination of basis functions given
as
{cos(nω), n = 0, 1, 2, . . . , N } ∪ {sin(nω), n = 1, 2, . . . , N }
(14)
This means the flexibility in choosing the coefficients of h[n]
as complex-valued results in an addition of N basis functions
to the set of available functions to represent P (ejω ), potentially leading to smaller approximation errors as intuitively
expected. Furthermore, the basis set in (14) also satisfies the
Haar condition and therefore leads to a unique optimal solution
[2]. These additional N basis functions manifest themselves
as an increase by N in the number of required alternations to
satisfy the traditional form of the alternation theorem for the
design of the autocorrelation. Therefore, Theroem 1 applies to
complex-valued filters with 2N + 2 alternations as opposed to
N + 2 alternations.
Example: Before proceeding to the design procedure, we
close this section with an example of an FIR filter that is
globally minimax optimal to illustrate the computation of
the adjusted desired response D′ (ω), the adjusted weight
′
function Wdes
(ω) and the alternation counting process in the
characterization of nonlinear phase FIR filters for optimality.
Our example design with real-valued coefficients in Figure 1a
has smaller ripple sizes than the one designed with MATLAB’s
firpm function which is based on the Parks-McClellan design
[1]. The filter order is N = 26 with the passband and the
stopband specified as ΩP = [0, 0.36π] and ΩS = [0.42π, π].
The weight Kdes is chosen as 3 in this example, meaning the
degrees of freedom will be chosen to suppress the stopband
error more than the passband error by this factor. We compute
EW (ω) as in equation (4), which is provided in Figure 1b.
From this computation, δP and δS are computed as 0.12 and
′
0.04, respectively. Therefore, D′ (ω) and Wdes
(ω) become
1,
ω ∈ [0, 0.36π]
D′ (ω) =
(15)
0.02, ω ∈ [0.42π, π]
and
′
Wdes
(ω) =
1, ω ∈ [0, 0.36π]
.
6, ω ∈ [0.42π, π]
(16)
′
This leads to an adjusted weight function EW
(ω) illustrated
in Figure 1c, which was obtained as in (7). This error indeed
exhibits N + 2 = 28 points, which is a necessary and
sufficient condition for the unique and global optimality of
the magnitude response |H(ejω )| asserted by Theorem 1. The
4
IV. O PTIMAL N ONLINEAR P HASE FIR F ILTER D ESIGN
A LGORITHM
1.2
Ours
Parks-McClellan
magnitude response
1
In this section, we describe the design algorithm for filters
restricted to having real-valued coefficients, therefore we will
require N +2 alternations. The same arguments apply to filters
with complex-valued coefficients simply by requiring 2N + 2
alternations and including sines in the basis functions for the
computation of optimal squared response.
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
0.8
1
A. Design Constraints
ω (× π)
(a)
weighted error
0.2
0.1
0
-0.1
-0.2
0
0.2
0.4
0.6
0.8
1
0.6
0.8
1
0.6
0.8
1
ω (× π)
adjusted weighted error
(b)
0.2
0.1
0
-0.1
-0.2
0
0.2
0.4
ω (× π)
adjusted weighted error
(c)
0.2
0.1
0
-0.1
-0.2
0
0.2
0.4
ω (× π)
(d)
Fig. 1. (a) The magnitude responses of two 26-th order FIR filters with
real-valued coefficients, one designed with the algorithm proposed in Section
IV and the other using firpm function of MATLAB (b) Weighted function
′ (ω) for our design
EW (ω) for our design (c) Adjusted weighted error EW
(d) Adjusted weighted error for the firpm design.
firpm design exhibits only 23 alternations in the adjusted
weighted error which is computed similarly and illustrated in
Figure 1d; and is clearly suboptimal.
The formal steps provided for testing optimality can be
bypassed by a more practical observation. Setting the maximum error in the passband to δP , and the maximum error
in the stopband to δS , one can directly verify whether the
filter satisfies the desired weight by checking if δδPS equals
Kdes . If this is the case, then the alternations can be also
counted directly on the magnitude response as the points
where the function reaches its extreme points in an alternating
′
fashion including the band edges since Wdes
(ω) and D′ (ω)
are tailored to turn these points into alternation points of the
′
adjusted weighted error EW
(ω). This approach can be verified
in Figure 1a where the extremal points are indeed alternations
and lead to 28 alternations including those occur at the band
edges.
Designing a zero-phase sequence g[n] that approximates
an ideal filter response, lifting its frequency response G(ejω )
until it is nonnegative and treating the lifted sequence p[n]
as the autocorrelation of an FIR filter h[n] has been used
as a nonlinear phase FIR filter design method at least since
1970 [3]. However, since the design specifications such as
relative weight on the stopband versus passband deviation in
the autocorrelation domain do not remain the same for the
filter due to the squaring relationship between P (ejω ) and
|H(ejω )|, the resulting filter h[n] does not necessarily reflect
the desired weight. Furthermore, no optimality arguments are
available for the final design h[n] because the optimality of
the autocorrelation sequence for one set of metrics does not
make the corresponding filter optimal for the same metrics.
We now provide a design method that correctly accounts for
this relationship and computes the weight to be applied in
the design of the autocorrelation sequence such that the final
design exhibits the desired ratio between passband and stopband deviations. Furthermore, the characterization theorem of
Section III will certify the optimality of the filter itself as
opposed to the optimality of the autocorrelation sequence.
Since the alternation frequencies and the number of alternations are the same for |H(ejω )| and P (ejω ) due to the specific
′
choice of D′ (ω) and Wdes
(ω) in the characterization theorem,
designing the autocorrelation sequence p[n] instead of the filter
h[n] itself with the correct number of alternations satisfies the
conditions of the characterization theorem for the optimality
of h[n] itself. Therefore we can design an autocorrelation
function that satisfies the required number of alternations and
recover the filter coefficients that will accept this function as
its autocorrelation function using either spectral factorization,
or if a minimum phase filter is particularly desired, any of the
methods in [4]–[8].
An autocorrelation sequence with at least N +2 alternations
can be designed by first by computing the coefficients of an optimal even-symmetric sequence g[n] of length 2N +1 such that
its Fourier transform G(ejω ) approximates the ideal filter, and
then by scaling and shifting to obtain P (ejω ) = aG(ejω ) + b
such that the following constraints are satisfied:
p
P (ejω ) swings symmetrically around
(i) |H(ejω )| =
unity in the passband, i.e. its extremal values become
1 + δP and 1 − δP for some positive δP ,
(ii) the minimum value of P (ejω ) is zero, p
(iii) the maximum value δS of |H(ejω )| = P (ejω ) in the
stopband satisfies the desired weight constraint, i.e., δδPS =
Kdes .
5
The first condition guarantees that the effects of squaring are
properly taken into account in the autocorrelation domain; the
second constraint ensures p[n] = ag[n] + bδ[n] is a proper
autocorrelation sequence; and the third constraint ensures that
the desired weight is not compromised through squaring.
The relationship between G(ejω ), P (ejω ) and |H(ejω )| is
illustrated with an example in Figure 2. Referring to the
passband and stopband deviations of |H(ejω )| as δP and δS
and those of G(ejω ) as ∆P and ∆S respectively, these three
constraints can be represented mathematically in terms of the
scale and shift coefficients a, b and the weight K that will
be applied in the design of g[n]. More specifically, we choose
the scaling coefficient a and the shifting coefficient b such that
the midpoints of the passband and stopband ranges of G(ejω )
and P (ejω ) match as in
a · 1 + b = 1 + δP2
and
a·0+b=
(17)
(18)
δS2
2
(19)
a = 1 + δP2 −
1.2
1
0.8
0.6
0.4
0.2
0
0
δS2
,
2
which yields
G(e j! )
0.2
0.4
0.6
0.8
1
0.6
0.8
1
0.6
0.8
1
(a)
P(ej! )
1.2
1
and
δS2
.
(20)
2
The relative weight between passband and stopband does
not change after scaling G(ejω ), therefore the weights are
identical in both:
0.8
b=
2
2
(1 + δP ) − (1 − δP )
4δP
∆P
=K
= 2 =
δS2
δS
∆S
(21)
0.6
0.4
0.2
0
0
0.2
0.4
Since δP /δS = Kdes , we can write this as
K=
(b)
4Kdes
δS
|H(e j! )|
(22)
1.2
or
4Kdes
.
(23)
K
In order to match the upper bound of the filter response in
the stopband to that of the autocorrelation after the scale and
shift, we have
δS2 = a · ∆S + b.
(24)
δS =
Inserting the values of a and b from equations
16K 2
and inserting δS2 = K 2des from equation (23),
4
2
2
16Kdes
8Kdes
16Kdes
∆S +
= 1+
−
K2
K2
K2
2
8Kdes
4 − 8K 2 .
K 2 + 16Kdes
des
2
8Kdes
K2
(25)
(26)
and, since ∆P = K∆S , we obtain
∆P =
K2
2
8Kdes
K
4 − 8K 2 .
+ 16Kdes
des
0.8
0.6
0.4
0.2
(19) and (20),
we obtain
Solving this for ∆S yields
∆S =
1
(27)
Finally, once the appropriate weight K that satisfies this
equation is found, the scaling and shifting coefficients can be
0
0
0.2
0.4
(c)
Fig. 2.
An example for (a) the Fourier Transform of the symmetric
sequence g[n] that approximates an ideal filter response, (b) that of the
autocorrelation p[n] obtained from g[n] through scaling and shifting, i.e.
p[n] = ag[n] + bδ[n], (c) the magnitude response of a nonlinear phase
filter h[n] the autocorrelation of which is p[n]. In this example, N = 20,
ΩP = [0, 0.30π], ΩS = [0.35π, π] and Kdes = 0.5.
6
computed directly from the parameters of this filter. Using
equation (23) and (20), we obtain
2
8Kdes
,
K2
and from equation (28) and (24), we have
b=
(28)
2
2
8Kdes
8Kdes
=
.
(29)
2
K ∆S
K∆P
Equation (27) is an implicit and nonlinear equation that
expresses the correct weight K that needs to be applied
in the design of G(ejω ) in terms of the actual passband
deviation ∆P obtained using the Remez Exchange Algorithm.
It can be solved efficiently for example using an iterative
procedure where we cut the search space on K each time
in a binary search fashion or using Newton-Raphson method.
Once the appropriate K is found, we design G(ejω ) with that
weight, scale by a and shift by b to obtain the autocorrelation
response P (ejω ), and then recover the filter coefficients h[n]
using spectral factorization where we can choose the filter
to be minimum phase, or maximum phase or anything in
between. If the desired filter is minimum phase, then more
efficient methods in [4]–[8] can be used instead of spectral
factorization.
a=
B. The Overall Algorithm
Assume that we want to compute the minimax-optimal
magnitude response |H(ejω )|, which is of order N (therefore
has N + 1 coefficients). Given ΩS , ΩP and Kdes , start with
an initial guess for K such that K ≥ 4Kdes (Kdes + 1)
for a physically meaningful design. (See Appendix A for
the derivation of this lower bound for weight as well as the
justification for why the following iterations converge.)
1. Compute the coefficients of the minimax-optimal evensymmetric (therefore zero-phase) filter g[n] of order 2N
to approximate the target function
1, ω ∈ ΩP
D(ω) =
(30)
0, ω ∈ ΩS
with the weight function
1, ω ∈ ΩP
W (ω) =
.
K, ω ∈ ΩS
(31)
This can be done directly using the Remez Exchange
algorithm, or modifying the Parks-McClellan algorithm,
etc.
2. Compute G(ejω ), the frequency response of g[n], and
compute the maximum value of passband error ∆P for
this g[n]. (This is also equivalent to the maximum value
of the absolute weighted error |W (ω)(G(ejω ) − D(ω))|).
3. If the resulting ∆P from step 2 satisfies the equality in
equation (27), go to step 4. Otherwise
(i) If ∆P is smaller than the expression in equation (27),
then increase the value of K and go to step 1.
(i) If ∆P is greater than the expression in equation (27),
then decrease the value of K and go to step 1.
The amount by which K is increased or decreased,
or the bounds of the search space, can be decided in
several ways, including methods such as binary search
(or bisection method), Newton-Raphson method, or any
other appropriate numerical method. Any choice of K
must satisfy K ≥ 4Kdes (Kdes + 1) for a physically
meaningful design.
4. Compute scale and shift coefficients a and b using equations (28) and (29).
5. Compute the function p[n] from g[n] as p[n] = a · g[n] +
b · δ[n], where δ[n] is the unit impulse function (not to be
confused with passband or stopband ripples of |H(ejω )|,
namely δP or δS .). This p[n] is the autocorrelation that
we were looking for.
6. Using any method including but not limited to spectral
factorization, obtain the coefficients of h[n] from the
autocorrelation sequence p[n]. There will be more than
one filter for which the autocorrelation sequence is p[n],
all of which are related to one another through a cascade
with an allpass filter. Choosing the zeros in the unit circle
as well as one from each pair of zeros located on the
unit circle leads to a minimum phase design. Choosing
all the zeros outside the unit circle and one pair from
each pair on the unit circle leads to a maximum phase
design. If a minimum phase design is desired but the filter
order is too high to perform spectral factorization using
polynomial root finding, methods in [4]–[8] can be used
to find the minimum phase solution. Other solutions can
then be found by finding the roots of this lower order
polynomial and reflecting the zeros from inside the unit
circle to outside of the unit circle as desired.
Example: Figure 3 illustrates the design of a high pass
filter where N = 500, ΩP = [0.40π, π], ΩS = [0, 0.39π] and
Kdes = 2. In this particular design, for the desired relative
ratio of Kdes = 2 between the passband and the stopband
deviations, the weight K to be applied during the the design
of the zero phase sequence g[n] was computed as 9801.96
corresponding to ∆P = 3.2646E −3 in the iterative procedure
in Section IV-B. This pair of (K, ∆P ) can be verified to satisfy
equation (27) which was obtained through only few iterations
using the bisection method.
Figure 3a illustrates the minimum phase impulse response
obtained using our design algorithm, which is the globally
optimal solution that approximates the desired response if no
restrictions on phase exist as certified by Theorem 1, and that
is designed as a linear phase filter using the firpm function
in MATLAB based on the Parks-McClellan algorithm [1].
Figure 3b, 3c and 3d illustrate the comparisons of magnitude
responses in the entire frequency range, in the passband and
in the stopband respectively. Removing the restrictions on
the phase has clearly led to a sharper magnitude response
characteristic. Furthermore, since this is a minimum phase
design, the group delay in the entire passband is much less than
that of the linear phase design at the expense of a non-constant
delay profile. Since this is a very high order filter, a polynomial
root finding based spectral factorization was impractical and
we exploited the algorithm in [4] instead to compute h[n] from
the computed p[n], where we used a matlab implementation
of this algorithm provided in [12].
7
0.8
P ROOF
our design
firpm design
0.6
Impulse Response
0.4
0.2
0
-0.2
-0.4
-0.6
0
100
200
300
400
500
Samples
(a)
0
our design
firpm design
Magnitude Response (dB)
-10
-20
-30
FOR
A PPENDIX A
C ONVERGENCE OF I TERATIONS
First we show that the resulting passband ripple ∆P,res of
a zero-phase G(ejω ) obtained by applying a weight K is an
increasing function of K, i.e. if we design a filter for the same
passband and stopband specifications but with a larger K, then
∆P,res will be also higher. Here the subscript res stands for
“resulting”. We can prove that ∆P,res (K) is an increasing
function of K by contradiction.
Assume that the maximum weighted error in a filter
G1 (ejω ) designed with a weight function
1,
ω ∈ ΩP
(32)
W1 (ω) =
K 1 , ω ∈ ΩS
is given by
-40
∆P 1 =
-50
max
ω∈(ΩP ∪ΩS )
|EW 1 (ω)|
(33)
-60
where
-70
0
0.2
0.4
0.6
0.8
EW 1 (ω) = W1 (ω) G1 (ejω ) − D(ω) .
1
ω (× π)
(b)
0.1
our design
firpm design
0.08
Magnitude Response (dB)
0.06
0.04
0.02
0
-0.02
-0.04
-0.06
-0.08
-0.1
0.4
0.5
0.6
0.7
ω (× π)
0.8
0.9
1
(c)
(34)
Similarly define W2 (ω), EW 2 (ω) and ∆P 2 for another filter
G2 (ejω ) and weight scalar K2 , which is greater than K1 . Since
G1 and G2 are real-valued frequency responses, they can be
designed using the Remez Exchange Algorithm and they will
be the respective unique optimal-minimax designs since they
satisfy the alternation theorem with at least N +2 alternations.
In order to obtain the contradiction, assume ∆P 2 ≤ ∆P 1 .
We can compute another weighted error function E12 (ω) using
the weight W1 (ω) along with G2 (ejω ) as
E12 (ω) = W1 (ω) G2 (ejω ) − D(ω) .
(35)
The maximum value of this weighted error ∆P 12 becomes
our design
firpm design
-50
Magnitude Response (dB)
∆P 12 =
-55
-60
-65
-70
0
0.05
0.1
0.15
0.2
ω (× π)
0.25
0.3
0.35
(d)
250
Samples
200
150
our design
firpm design
100
50
0
0.4
0.5
0.6
0.7
ω (× π)
0.8
0.9
1
max
ω∈(ΩP ∪ΩS )
|E12 (ω)| = max {∆P 2 , K1 ∆S2 } . (36)
Since both ∆P 2 and K1 ∆S2 = K1 ∆KP22 are at most ∆P 2 , the
filter G2 (ejω ) is at least as good as G1 (ejω ) in approximating
the desired function with the weight function W1 (ω), which
contradicts the unique optimality of the latter which had been
established by the alternation theorem. Therefore, for K2 >
K1 , the passband error ∆P 2 cannot be equal or smaller than
∆P 1 . This implies ∆P (K) is a strictly increasing function of
K.
Now consider the range of weights K that results in physically meaningful designs. For a given weight Kdes , we can
always design a trivial filter h[n] = 1+K1 des δ[n], which leads to
Kdes
stopband error δS = 1+K1 des and passband error δP = 1+K
,
des
δP
which satisfy the desired weight constraint δS = Kdes . The
sum of these errors is exactly unity. Therefore, in an optimal
design with N + 1 coefficients, the sum of these errors can
never be larger than unity, and we have
(e)
Fig. 3. Comparison of our optimal design to a linear phase design for a
high-pass filter where N = 500, ΩP = [0.40π, π], ΩS = [0, 0.39π] and
Kdes = 2. (a) the impulse responses, (b) the magnitude responses, (c) the
magnitude response zooming into the passband, (d) the magnitude response
zooming into the stopband, (e) the group delays in samples in the passband.
δP + δS = (Kdes + 1)δS ≤ 1.
(37)
Inserting δS from equation (23), we obtain
(Kdes + 1)
4Kdes
≤ 1,
K
(38)
8
1
∆ P (K)
∆ P,res (K)
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
40
60
80
100
120
140
160
180
200
K
Fig. 4. An example of the resulting passband deviation ∆P,res as a function
of weight K and the function ∆P given in equation (27).
which yields the physically meaningful lower bound for the
weight K as
K ≥ 4Kdes (Kdes + 1).
(39)
We need to find the weight K that yields a resulting value of
weighted error ∆P,res that is equal to the expression for ∆P
in equation (27). We have already established ∆P,res (K) is an
increasing function of K and is always lower than 1−∆S (K).
On the other hand, ∆P (K) stated in equation (27) equals
1 − δS at the lower bound we just found in equation (39), is
monotonically decreasing for weights greater than this lower
bound and asymptotically approaching zero. This means these
two functions do intersect once and only once in this regime
of weights, and that is the weight we are looking for. This
is also another proof that an optimum exists and it is unique.
Figure 4 illustrates an example of ∆P (K) and ∆P,res curve
for wdes = 2, ΩS = [0, 0.36π] and ΩP = [0.42π, π] where
the physically meaningful values of K start as 24. During
the iterations, if ∆P,res (K) > ∆P (K), the current value of
K is to the right of the intersection and therefore needs to
be decreased. Otherwise, it is increased, and the iterations
continue until the two values meet. The search is particularly
efficient if the search space is halved each time, corresponding
to the binary search (or bisection) scheme.
R EFERENCES
[1] T. Parks and J. McClellan, “Chebyshev approximation for nonrecursive
digital filters with linear phase,” Circuit Theory, IEEE Transactions on,
vol. 19, no. 2, pp. 189 – 194, Mar 1972.
[2] E. W. Cheney, Introduction to approximation theory. McGraw-Hill,
1966.
[3] O. Herrmann and W. Schuessler, “Design of nonrecursive digital filters
with minimum phase,” Electronics Letters, vol. 6, no. 11, pp. 329–330,
May 1970.
[4] R. Boite and H. Leich, “A new procedure for the design of high order
minimum phase fir digital or ccd filters.” Signal Processing, vol. 3, pp.
101 – 108, 1981.
[5] X. Chen and T. W. Parks, “Design of optimal minimum phase fir filters
by direct factorization,” Signal Process., vol. 10, no. 4, pp. 369–383,
Jun. 1986.
[6] N. Damera-Venkata and B. L. Evans, “Optimal design of real and complex minimum phase digital fir filters,” in Acoustics, Speech, and Signal
Processing, 1999. Proceedings., 1999 IEEE International Conference
on, vol. 3, Mar 1999, pp. 1145–1148 vol.3.
[7] Y. Kamp and C. Wellekens, “Optimal design of minimum-phase fir
filters,” IEEE Transactions on Acoustics, Speech, and Signal Processing,
vol. 31, no. 4, pp. 922–926, Aug 1983.
[8] G. Mian and A. Nainer, “A fast procedure to design equiripple minimumphase fir filters,” IEEE Transactions on Circuits and Systems, vol. 29,
no. 5, pp. 327–331, May 1982.
[9] H. Samueli, “On the design of optimal equiripple fir digital filters
for data transmission applications,” IEEE Transactions on Circuits and
Systems, vol. 35, no. 12, pp. 1542–1546, Dec 1988.
[10] S.-P. Wu, S. Boyd, and L. Vandenberghe, “Fir filter design via semidefinite programming and spectral factorization,” in Decision and Control,
1996., Proceedings of the 35th IEEE Conference on, vol. 1, Dec 1996,
pp. 271–276 vol.1.
[11] S. Demirtas, “Characterization and design of nonlinear phase fir filters
for global minimax-optimality,” in preparation.
[12] J. O. Smith, Introduction to Digital Filters with Audio Applications.
http://ccrma.stanford.edu/˜jos/filters/, accessed (May 2016), online book.
| 3cs.SY
|
Statement A: Approved for Release. Distribution is unlimited.
Presented In Partial Fulfillment of the Requirements of a Master of Science Degree
Penn State Department of Mechanical Engineering
February 29, 2016
APPLYING ARTIFICIAL NEURAL NETWORKS TO PREDICT NOMINAL VEHICLE
PERFORMANCE
Adam J. Last
Applied Research Laboratory
P.O. Box 30
State College, PA 16804
e-mail: [email protected]
ABSTRACT
This paper investigates the use of artificial neural
networks (ANNs) to replace traditional algorithms and manual
review for identifying anomalies in vehicle run data. The
specific data used for this study is from undersea vehicle
qualification tests. Such data is highly non-linear, therefore
traditional algorithms are not adequate and manual review is
time consuming. By using ANNs to predict nominal vehicle
performance based solely on information available pre-run,
vehicle deviation from expected performance can be
automatically identified in the post-run data. Such capability is
only now becoming available due to the rapid increase in
understanding of ANN framework and available computing
power in the past decade. The ANN trained for the purpose of
this investigation is relatively simple, to keep the computing
requirements within the parameters of a modern desktop PC.
This ANN showed potential in predicting vehicle performance,
particularly during transient events within the run data.
However, there were also several performance cases, such as
steady state operation and cases which did not have sufficient
training data, where the ANN showed deficiencies. It is
expected that as computational power becomes more readily
available, ANN understanding matures, and more training data
is acquired from real world tests, the performance predictions
of the ANN will surpass traditional algorithms and manual
human review.
Timothy F. Miller
Applied Research Laboratory
P.O. Box 30
State College, PA 16804
e-mail: [email protected]
service lives to ensure proper functionality. Results from these
qualification runs can include hundreds of channels of data,
including speeds, pressures, temperatures, flow rates,
vibrations, and many others. Further, each qualification run
usually has a unique mission profile, which is the progression
of commanded speeds throughout the run. This makes direct
comparison between qualification runs difficult.
Therefore, identifying anomalous behavior from the
post-run data requires many highly trained and experienced
engineers manually reviewing each data channel. Even then,
due to human error, subtle anomalies can be overlooked or,
conversely, normal behavior can be flagged as anomalous. To
counter this, creating an algorithm or method to predict results
based solely on the mission profile is highly desirable. The
current best method is to average the measured values for a
given data channel at a given speed from all previous runs, and
assume that is the “normal” operating condition for that speed.
However, the response of each data channel is highly nonlinear, as not only the current speed command but also the
cumulative effect of all prior operation determines current
performance. As a result, traditional algorithms become
unwieldy and unreliable for this application.
PROBLEM
Undersea vehicles, like any high performance systems,
require many qualification runs prior to and throughout their
INTRODUCTION
Overcoming the breakdown of traditional algorithms
on large non-linear data sets has been a focus for many
industries in recent years. Data center optimization [1],
handwriting recognition [2], language translation [3], and
image classification [4] are examples of problems that
traditional algorithms cannot effectively solve. Decades of
Statement A: Approved for Release.
Distribution is unlimited.
1
exponential growth in available computing power, along with
recent developments in mathematical methods, has enabled a
new approach to handling large data sets- artificial neural
networks (ANNs). ANNs have been deployed in the last few
years to enable everything from real time language translation
[5] to instant handwriting recognition [6] to financial market
analysis [7].
This paper outlines the approach to using the existing
undersea vehicle qualification run data to train an ANN to
predict future hypothetical run results, with the only input being
the mission profile. Then, when the hypothetical qualification
run is actually performed in the real world, the results can be
fed back to the computer. The ANN can then identify where
differences exist between its predicted result and the actual
results to automatically flag anomalous behavior.
TECHNICAL BACKGROUND
The first work on artificial neural networks began in
the mid-20th century with the concept of perceptrons. The
perceptron is essentially a logic gate, where multiple inputs are
taken in, individually weighted, and summed. If the sum of the
weighted inputs are greater than a threshold value, then the
output of the perceptron is “1”, otherwise the output is “0”.
This is shown graphically in Figure 1 (graphic adapted from
Nielson [8]).
0 𝑖𝑓 ∑ 𝑤𝑗 𝑥𝑗 ≤ 𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑
𝑜𝑢𝑡𝑝𝑢𝑡 = {
1 𝑖𝑓 ∑ 𝑤𝑗 𝑥𝑗 > 𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑
Figure 1. Perceptron logic
The perceptron was later refined to allow for nonbinary output (i.e. output other than just 1 and 0). This is
accomplished by replacing the step function logic with a
sigmoid function. For this reason, the new model for the
perceptron was named the “sigmoid neuron”. The “neuron”
portion of the name derives from the sigmoid neuron’s
similarity in operation to a biological neuron, which on a very
basic level activates an output given a sufficient input [8].
Sigmoid neurons become useful when linked together
to perform complex logic, and were renamed “hidden neurons”
in the scope of neural networks. Figure 2 (graphic adapted
from Nielson [8]) shows an artificial neural network containing
3 inputs, 4 hidden neurons, and one output.
Statement A: Approved for Release.
Distribution is unlimited.
Figure 2. Simple ANN
To orient Figure 2 to the current problem at hand, the
input layer would be known values (such as the commanded
speed). The output layer would be whatever data channel you
are trying to simulate, the most intuitive being actual vehicle
speed. There will obviously need to be far more than 4 hidden
neurons for any meaningful results.
For a neural network to be useful, the weighting
factors for each input to each hidden layer sigmoid neuron must
be defined. To do so, a training process must be undertaken.
The training process requires a large set of already known
input/output pairings. In this case, this known set is the
available previous vehicle qualification runs, which have
known inputs (such as commanded speed) and measured output
(such as actual speed). Initially in the training process, a
random set of weights is defined. The output is then calculated
and compared to the actual output. This is the error in the
neural network for the current “evolution”. The weights are
then adjusted (shown as Δw in Figure 2) and the ANN resolved.
This process is repeated until the error is below an acceptable
value or a preset number of evolutions has been achieved [9].
There are several methods for determining the
amounts by which to adjust each individual weight. One of the
easiest to implement and most computationally efficient is the
optimization method of stochastic gradient decent. Gradient
descent is a well-known mathematical model for finding
minima and maxima. Stochastic methods are employed to
introduce an element of randomness into the optimization
routine. This allows for the optimization to escape local
minima and maxima during the solution and continue
improving on the solution through many evolutions [10].
While stochastic gradient decent is efficient, it is still
very computationally intensive when solving the weights for
many (hundreds or more) hidden layer sigmoid neurons. That
is why the recent sustained exponential growth in available
computing power has been critical to making ANNs practical.
Once the ANN weights are solved, it takes very little
computational resources to run new inputs through the ANN to
find the output.
There have been several advances in mathematical
methods in recent years that have made ANNs much more
powerful and extensible to more applications. However, for the
purpose of this paper, using the methods already outlined is
sufficient.
2
ANN SETUP
A simple ANN was set up that was capable of being
run on a single desktop computer with 4 GB RAM. This ANN
was built using the Netlab MATLAB toolbox developed by
Nabney [11]. The training data was taken from 20 qualification
runs of an undersea vehicle. This gave 64,779 unique “time
steps”. Each “time step” represents a discrete amount of
runtime. Time and speeds have been non-dimensionalized.
For each “time step”, input values must be provided to
the ANN. For the engine of this vehicle, like many other
vehicles, the only true control input is commanded speed. All
other engine parameters are adjusted internally for the goal of
matching that commanded speed as closely as possible.
Though not strictly a “control” parameter, the other engine
input is the amount of energy reserves remaining (and, by
extension, the amount of reserves that has already been used).
For various systems, this could take the form of percentage of
combustible fuel remaining in a tank, remaining battery power,
etc. For many systems, such as the one under study, this energy
reserves input also functions as a “state of health” parameter
(i.e. for electric systems, motors may not perform optimally
once the batteries drop below a threshold voltage).
From these general “inputs”, the specific input data
that was deemed pertinent to be provided to the ANN is as
follows…
1.
2.
Speed Command Information…
a. Current Commanded Speed
b. Time at Current Commanded Speed
c. Previous Commanded Speed
% of Energy Reserves Depleted (termed “Utilization”
for the purposes of this paper)
Also provided for training the ANN was the actual measured
vehicle speed for each data point. The “success” of the ANN
training would be determined by comparing the actual vehicle
measured speed to the ANNs predicted vehicle speed.
The ANN was initialized to have 750 hidden sigmoid
neurons. The training was set to terminate after 750,000
evolutions. This is about 48 hours of computation time on a PC
with an Intel Core i5-2400 CPU running at 3.1 GHz.
To assist with visualization of the training process of
this specific ANN, the notation from Figure 2 is used in Figure
3 below, depicting the ANN setup. The routine of Figure 3
would be run once for each time step each evolution (64,779
times every evolution for this data set).
Figure 3. ANN setup and training visualization
Statement A: Approved for Release.
Distribution is unlimited.
3
1
Nondimensionalized Speed
0.9
0.8
0.7
0.6
0.5
0.4
0
0.1
0.2
0.3
0.4
0.5
0.6
Nondimensionalized Time
0.7
0.8
0.9
1
Figure 4. Full ANN results
Statement A: Approved for Release.
Distribution is unlimited.
Consistently low prediction of high speed command
performance
0.9
0.8
Nondimensionalized Speed
ANN RESULTS
Figure 4 above shows the commanded speed (dark
blue line) and the ANN’s prediction for actual speed (red line).
Also given is the true actual speed as measured in the
qualification tests (light blue line).Quick inspection of Figure 4
shows that the ANN has largely succeeded in matching the
general speed response. However, there are several spans of
time where the ANN fails to predict any more accurately than
simply guessing that the actual speed is equal to the
commanded speed. The most obvious example is during long
periods of constant speed, known as steady state. Conversely,
there are several spans where the ANN prediction is much more
accurate than any traditional algorithm could be, such as during
speed changes (transients)
It is critical to attempt to understand the reasons why
the ANN is very accurate or not accurate in various situations.
If these reasons can be well understood, then adjustments can
be made to the ANN training routine to address deficiencies
and promote strengths.
Examples would be providing
additional input information, working to obtain test data for
situations where the ANN shows deficiencies, changing the
number of hidden neurons, or adjusting the evolutions to
simulation completion. To that end, the remainder of this
section of the paper examines three specific performance cases
where the ANN shows interesting prediction behavior. The
upcoming Figures 5, 6, and 7 all show magnified subsets of
data from Figure 4, to allow for in-depth analysis.
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.42
0.44
0.46
0.48
0.5
Nondimensionalized Time
0.52
0.54
0.56
Figure 5. ANN results, low predictions of high speed operation
Figure 5 shows a magnified portion of the ANN data
of Figure 4, from 0.42 to 0.56 of the nondimensional time.
Added to the plot, that was not present in Figure 4, is the nondimensionalized utilization (% of energetics used), shown in
magenta. Note that there are three separate “runs” shown, as the
utilization resetting to zero is indicative of a new run. In this
timeframe there are three instances of the vehicle being
commanded to its highest speed. In all three instances, the
ANN predicts the actual speed will only achieve about 80% of
the full speed value. However, this is only true for one of the
three occurrences; in the other two the vehicle does achieve
4
high speed operation. It is probable that the ANN is chronically
under predicting the actual high speed operation as a “safe”
prediction. This is because the speed sometimes undershoots,
but almost never overshoots, a high speed command. The
likelihood that the vehicle will undershoot the high speed
command is increased at high vehicle utilization (near the end
of an individual run- because effective control is more
difficult), as it was in the one instance of undershooting in
Figure 5. The ANN shows slight sensitivity to this as the high
speed command in Figure 5 furthest to the right occurs at the
lowest utilization; the ANN displays the highest predicted
speed. However, the prediction is still too low, and the ANN
clearly does not fully incorporate the link between utilization
and the potential for undershooting. More run data which
contains more high speed command operation at various
utilizations would likely remedy this issue.
Data overfitting during “staircase” speed increase
Accurate prediction of transient performance
Figure 7 shows a brief speed command of 0.8 nondimensionalized speed. The ANN predicted speed appears to
match the actual speed of the vehicle closely over this time.
However, the predicted speed is slowly increasing the entire
time, as if “anticipating” the next speed command. The reason
is that this speed change is part of a “staircase” pattern of speed
changes, where the speed goes from the minimum to the
maximum over a series of small changes in speed in rapid
succession. These “staircase” patterns are very common in the
training run data. Therefore, the ANN learned that when the
speed changes from a moderate speed to just a slightly higher
speed, in all likelihood it was in the middle of a “staircase”
speed profile and should anticipate another small speed step up.
However, it is obviously not always true that these small speed
steps must be part of a “staircase” sequence. Therefore, the
“anticipation” of the ANN for another immediate speed change
is unwarranted and indicative of overfitting of the data [12].
This would be remedied by additional run data which showed
small speed steps which are not part of a “staircase” pattern.
Nondimensionalized Speed
0.65
0.6
0.55
0.5
0.45
0.4
0.45
0.452
0.454
0.456
0.458
Nondimensionalized Time
0.46
0.462
0.464
Figure 6. ANN results, transient performance
Figure 6 shows one of the speed command changes in
the data. During the transient, the ANN predicted speed
showed a sharp drop then gradual step down to the final steady
state speed, consistent with actual vehicle operation. This is a
feature that a traditional algorithm would have trouble
matching, particularly once data other than speed
(temperatures, pressures, etc.), which may take a much longer
time than speed to reach steady state, are considered.
Statement A: Approved for Release.
Distribution is unlimited.
Nondimensionalized Speed
0.86
0.84
0.82
0.8
0.78
0.76
0.74
0.72
0.144
0.146
0.148
0.15
Nondimensionalized Time
0.152
0.154
Figure 7. ANN results, overfit solution
CONCLUSIONS
The results show that even simple ANNs show
promise of outperforming traditional prediction methods for
times in the data which are highly non-linear (such as speed
transients). However, the simple ANN does not have adequate
performance for all times, most notably during steady state
operation. It is highly probably that the ANN would meet or
exceed the performance of traditional algorithms if a more
complex ANN were solved on a more powerful computer.
A more fundamental limitation of the ANN
performance may be the potential for overfitting of data. This
is a result of not having a sufficiently large set of training data.
Specifically, if the same sequence of speed steps is present in
most data sets, the ANN will be highly accurate in predicting
results for that exact sequence, but not when the sequence is
altered slightly. Therefore, it is possible that even with an
advanced ANN solved on a powerful computer, the size of the
training data set could prove a limitation to ANN effectiveness.
The mitigation for this problem is the inclusion of additional
5
data sets in the training data, especially those with a diverse
mission profiles.
FUTURE WORK
The size of the training data set continues to increase,
as qualification runs of the vehicle are ongoing. Additionally,
computational power continues to become more and more
readily available. Therefore, it is prudent to develop an
advanced ANN framework now and run it with the existing
data set with the understanding that the results may not yet
outperform traditional algorithms or manual data analysis.
However, with that framework in place, retraining the ANN
with new data becomes a trivial exercise. As more data and
computational power becomes available it can be expected that
at some point the ANN will be provide the best predictive
capability. At that point additional data channels (temperatures,
pressures, etc.) can all be predicted using the same ANN
framework.
Once ANNs have been proven to reliably identify
anomalies from post run data, it will be possible to extend their
use into identifying anomalies during a run itself. For instance,
if a vehicle behaves off-prediction by some margin, the ANN
can alert the vehicle control which can then evaluate whether to
issue an early termination of the run. Even further in the future,
ANNs should be capable of assuming control for adjusting
vehicle run parameters (flow rates, pressures, etc.) in real time
to optimize vehicle performance.
CLOSING REMARKS
The conclusions of this paper (requiring more
computational power and better mathematical techniques, but
more importantly additional training data) are consistent with
the needs of nearly all ANN projects currently in development.
On the computational side, chipmakers are rushing to
provide hardware designed specifically for ANN training and
deployment [13].
This hardware leverages architecture
traditionally used for intensive graphics processing. Brand new
architecture is under development which will be even further
optimized for machine learning and neural networks [14].
On the training data side, companies have been using
devices owned and used by individuals to collect data to
improve their algorithms. The most relevant of these to this
paper might be Tesla Motors, which receives data from every
one of its cars on the road to improve its “autopilot”
(autonomous driving) feature [15]. Interestingly, this type of
training data collection is a driving force in the recent public
debate about the privacy of personal electronic devices. For
instance, Google’s efforts to track cell phone activity stems
from their desire to expand the training sets for the algorithms
that control search result prioritization, traffic congestion
prediction, language translation, and more [16].
NOMENCLATURE
x = ANN training data input
w = ANN sigmoid neuron weight
Statement A: Approved for Release.
Distribution is unlimited.
ACKNOWLEDGMENTS
The data for this paper was provided by ARL PSU from
experiments performed under NAVSEA contract. We thank the
entirety of ARL’s Energy Science and Power Systems division
for the support and resources. The Netlab MATLAB toolbox
which was utilized to perform the ANN training was developed
at Aston University by Ian Nabney and Christopher Bishop.
REFERENCES
[1] Gao, Jim. Machine Learning Applications for Data Center Optimization.
Rep.
Mountain
View:
Google,
2013.
Web.
<https://docs.google.com/a/google.com/viewer?url=www.google.com/about/dat
acenters/efficiency/internal/assets/machine-learning-applicationsfor-datacenteroptimization-finalv2.pdf>.
[2] Le Cun, Y. "Handwritten Character Recognition Using Neural Network
Architectures." 4th USPS Advanced Technology Conference (1990): n. pag.
AT&T
Bell
Laboratories.
Web.
20
Oct.
2015.
<http://yann.lecun.com/exdb/publis/pdf/matan-90.pdf>.
[3] Auli, Michael. "Joint Language and Translation Modeling with Recurrent
Neural Networks." Microsoft Research (n.d.): n. pag. Microsoft. Web. 20 Oct.
2015.
<http://research-srv.microsoft.com/enus/um/people/gzweig/Pubs/EMNLP2013RNNMT.pdf>.
[4] Szegedy, Christian. "Building a Deeper Understanding of Images."
Research
Blog.
Google,
n.d.
Web.
20
Oct.
2015.
<http://googleresearch.blogspot.com/2014/09/building-deeper-understandingof-images.html#uds-search-results>.
[5] Bergen, Mark. "Google's Neural Nets Can Now Translate Text Instantly,
Practically Everywhere." Recode. Recode, 29 July 2015. Web. 20 Oct. 2015.
<http://recode.net/2015/07/29/googles-neural-nets-can-now-translate-textinstantly-practically-everywhere/>.
[6] Martendale, Jon. "Now Our Computers Can Forge Our Letters and Checks."
Digital Trends. Digital Trends, 22 July 2015. Web. 20 Oct. 2015.
<http://www.digitaltrends.com/computing/researcher-computer-handwriting/>.
[7] "Neural Networks - Applications." Neural Networks - Applications. N.p.,
n.d.
Web.
20
Oct.
2015.
<http://cs.stanford.edu/people/eroberts/courses/soco/projects/neuralnetworks/Applications/stocks.html>.
[8] Nielson, Michael A. Neural Networks and Deep Learning. N.p.:
Determination, 2015. Determination Press. Web. 20 Oct. 2015.
<http://neuralnetworksanddeeplearning.com/chap1.html>.
[9] Bishop, Christopher M. Neural Networks for Pattern Recognition. Oxford:
Clarendon, 1995. Print.
[10] Bottou, Leon. "Stochastic Gradient Descent Tricks." Microsoft Esearch
(n.d.):
n.
pag.
Web.
20
Oct.
2015.
<http://research.microsoft.com/pubs/192769/tricks-2012.pdf>.
[11] Nabney, Ian T. NETLAB: Algorithms for Pattern Recognition. London:
Springer, 2003. Print.
[12] "Documentation." Improve Neural Network Generalization and Avoid
Overfitting.
Mathworks,
n.d.
Web.
20
Oct.
2015.
<http://www.mathworks.com/help/nnet/ug/improve-neural-networkgeneralization-and-avoid-overfitting.html>.
[13] Ferrucci, David. "AI MAGAZINE." Building Watson: An Overview of the
DeepQA Project. AAAI, Aug. 2007. Web. 03 Feb. 2016.
<http://www.aaai.org/ojs/index.php/aimagazine/article/view/2303/2165>.
[14] Colaner, Seth. "Movidius, Google Team Up To Leverage Deep Learning
On Mobile Devices." Tom's Hardware. Tom's Hardware, 27 Jan. 2016. Web. 3
Feb.
2016.
<http://www.tomshardware.com/news/movidius-google-deeplearning-mobile-devices,31078.html>.
[15] Fehrenbacher, Katie. "How Tesla Is Ushering in the Age of the Learning
Car." How Tesla Is Ushering in the Age of the Learning Car Comments.
Fortune,
16
Oct.
2015.
Web.
03
Feb.
2016.
<http://fortune.com/2015/10/16/how-tesla-autopilot-learns/>.
[16] Lafrance, Adrienne. "The Convenience-Surveillance Tradeoff." The
Atlantic. Atlantic Media Company, 14 Jan. 2016. Web. 03 Feb. 2016.
<http://www.theatlantic.com/technology/archive/2016/01/the-conveniencesurveillance-tradeoff/423891/>.
6
ANNEX A
INPUT FILE TO NETLAB TOOLBOX (MATLAB CODE)
% Running this script requires installation of Netlab 3.3 MATLAB toolbox
% Netlab toolbox available at
http://www.aston.ac.uk/eas/research/groups/ncrg/resources/netlab/downloads/
% Script derived from standard Netlab input file, demmlp1.m, by Ian T Nabney
clear all;
clc
% Input arrays must be initialized to variable 'x' as follows...
%
First Column = Current Commanded Speed
%
Second Column = Previous Commanded Speed (0 if run just started)
%
Third Column = Time Since Last Speed Command Change
%
Fourth Column = Current Utilization
% Set up network parameters.
nin = 4; % Number of inputs. (see input array comment above)
nhidden = 750; % Number of hidden sigmoid neurons.
nout = 1; % Number of outputs.
alpha = .1; % Coefficient of weight-decay prior.
% Create and initialize network weight vector.
net = mlp(nin, nhidden, nout, 'linear', alpha);
% Set up vector of options for the optimiser.
options = zeros(1,18);
options(1) = 1; % This provides display of error values.
options(14) = 75000; % Number of training cycles.
% Run Netlab to train neural network
[net, options] = netopt(net, options, x, t, 'scg');
% Use neural network to generate predicted speeds
y = mlpfwd(net, x);
Statement A: Approved for Release.
Distribution is unlimited.
7
ANNEX B
VALIDATION SET
Several months after the ANN described in this paper had been trained, new field test data became available.
This new data was used as a “validation set”, in order to test the efficacy of the ANN without re-training. The
results are shown in Figure B-1 below.
1
0.9
0.8
Nondimensionalized Speed
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
Nondimensionalized Time
0.7
0.8
0.9
1
Figure B-1. Validation Set Results
A few observations from the ANN results are given below.
Non-dimensionalized time = 0. The ANN is unable to capture the speed change at “startup”. This
is likely because there are relatively few “startups” in the training data (there is only 1 startup per field test,
with 20 field tests used in the training data).
Non-dimensionalized time = 0.25. Through this speed step, the ANN predicts a near-constant
speed that is the average of the speed transient. This is likely due to an insufficiently advance algorithm, as
the ANN “settled” into minimizing error by picking a mid-point speed, rather than accurately modeling the
transient.
Non-dimensionalized time = 0.3. The ANN anticipated a speed “overshoot” during this transient,
which did not occur. Analysis of the training data shows that, on several occasions, an “overshoot” did
happen at this speed. Therefore, this ANN error is the results of an insufficient number of training sets that
do not have an overshoot during a similar speed step.
Non-dimensionalized time = 0.55. The ANN correctly predicts a speed “undershoot” during this
transient.
Non-dimensionalized time = 0.6 through end of run. The ANN predicts that the speed will be
much less than the actual speed through the end of the run. This is likely due to previous runs in the
training set that have failed to maintain speed late in run. Therefore, more training data is necessary for the
ANN to be able to distinguish the characteristics of runs that will and will not be able to maintain speed.
From the results, we can calculate the accumulated error using two speed prediction methods. The first method
is simply presuming the actual speed is equal to the setpoint (this will be referred to as the “Non-ANN Method”). It
should be noted that this method could be improved by simple algorithms, such as assuming a time constant for
speed changes, but that is not the focus of this analysis. The second method is the prediction of the ANN. The error
Statement A: Approved for Release.
Distribution is unlimited.
8
of each of these methods can be calculated as the absolute value of prediction minus the actual speed. A graph of
accumulated error is shown in Figure B-2 below.
Accumulated Error (|Prediction Speed - Actual Speed|)
300
250
200
150
100
50
0
0
0.1
0.2
0.3
0.4
0.5
0.6
Nondimensionalized Time
0.7
0.8
0.9
1
Figure B-2. Accumulated Error from Non-ANN and ANN Speed Prediction Methods on Validation Set
Figure B-2 shows that the Non-ANN and ANN methods for speed prediction have nearly the same error
accumulation rate until late in the data set (second to last speed change). This is an encouraging sign that with
additional training-set data and more advanced ANN methods, neural networks will be practical tools for predicting
vehicle performance in the near future. In summary, the results of this validation set reinforce the assertions made in
the conclusion of this paper that ANN performance can be improved with more advanced methods and more
computing power, but more training data will be needed as well.
Statement A: Approved for Release.
Distribution is unlimited.
9
| 9cs.NE
|
Feature Engineering for Predictive Modeling using Reinforcement Learning
Udayan Khurana, Horst Samulowitz, Deepak Turaga
Abstract
1.00
100
0.75
80
0.50
sin(0.32x)
Feature engineering is a crucial step in the process of predictive modeling. It involves the transformation of given feature space, typically using mathematical functions, with the
objective of reducing the modeling error for a given target.
However, there is no well-defined basis for performing effective feature engineering. It involves domain knowledge, intuition, and most of all, a lengthy process of trial and error.
The human attention involved in overseeing this process significantly influences the cost of model generation. We present
a new framework to automate feature engineering. It is based
on performance driven exploration of a transformation graph,
which systematically and compactly enumerates the space of
given options. A highly efficient exploration strategy is derived through reinforcement learning on past examples.
60
x
arXiv:1709.07150v1 [cs.AI] 21 Sep 2017
{ukhurana,samulowitz,turaga}@us.ibm.com
IBM TJ Watson Research Center
40
0.25
0.00
0.25
0.50
20
0
0.75
1.00
0
20
40
t
60
(a) Original data
80
100
0
20
40
t
60
80
100
(b) Engineered data.
Figure 1: Illustration of different representation choices.
Introduction
Predictive analytics are widely used in support for decision
making across a variety of domains including fraud detection, marketing, drug discovery, advertising, risk management, amongst several others. Predictive models are constructed using supervised learning algorithms where classification or regression models are trained on historical data
to predict future outcomes. The underlying representation
of the data is crucial for the learning algorithm to work effectively. In most cases, appropriate transformation of data
is an essential prerequisite step before model construction.
For instance, Figure 1 depicts two different representations for points belonging to a classification problem dataset.
On the left, one can see that instances corresponding to the
two classes are present in alternating small clusters. For most
machine learning (ML) algorithms, it is hard to draw a reasonable classifier on this representation that separates the
two classes On the other hand if the feature x is replaced by
its sine, as seen in the image on the right, it makes the two
classes reasonably separable by most classifiers. The task or
process of altering the feature representation of a predictive
modeling problem to better fit a training algorithm is called
feature engineering (FE). The sine function is an instance of
a transformation used to perform FE. Consider the schema
of a dataset for forecasting hourly bike rental demand1 in
Figure 2(a). Deriving several features (Figure 2(b)) dramatically reduces the modeling error. For instance, extracting
1
https://www.kaggle.com/c/bike-sharing-demand
the hour of the day from the given timestamp feature helps
to capture certain trends such as peak versus non-peak demand. Note that some valuable features are derived through
a composition of multiple simpler functions. FE is perhaps
the central task in improving predictive modeling performance, as documented in a detailed account of the top performers at various Kaggle competitions (Wind 2014).
In practice, FE is orchestrated by a data scientist, using
hunch, intuition and domain knowledge based on continuously observing and reacting to the model performance
through trial and error. As a result, FE is often timeconsuming, and is prone to bias and error. Due to this inherent dependence on human decision making, FE is colloquially referred to as “an art/science” 2 3 , making it difficult
to automate. The existing approaches to automate FE are
either computationally expensive evaluation-centric and/or
lack the capability to discover complex features.
We present a novel approach to automate FE based on
reinforcement learning (RL). It involves training an agent
on FE examples to learn an effective strategy of exploring
available FE choices under a given budget. The learning
and application of the exploration strategy is performed on
2
http://www.datasciencecentral.com/profiles/blogs/
feature-engineering-tips-for-data-scientists
3
https://codesachin.wordpress.com/2016/06/25/
non-mathematical-feature-engineering-techniques-for-data-science/
(a) Original features and target count.
(b) Additionally engineered features using our technique.
Figure 2: In Kaggle’s biking rental count prediction dataset,
FE through our technique reduced Relative Absolute Error
from 0.61 to 0.20 while retaining interpretability of features.
a transformation graph, a directed acyclic graph representing relationships between different transformed versions of
the data. To the best of our knowledge, this is the first work
that learns a performance-guided strategy for effective feature transformation from historical instances. Also, this is
the only work in FE space that provides an adaptive budget
constrained solution. Finally, the output features are compositions of well-defined mathematical functions which make
them human readable and usable as insights into a predictive
analytics problem, such as the one illustrated in Figure 2(b).
Related Work
Given a supervised learning dataset, FICUS (Markovitch
and Rosenstein 2002) performs a beam search over the space
of possible features, constructing new features by applying “constructor functions” (e.g. inserting an original feature into a composition of transformations). FICUS’s search
for better features is guided by heuristic measures based
on information gain in a decision tree, and other surrogate measures of performance. In constrast, our approach
optimizes for the prediction performance criterion directly,
rather than surrogate criteria, and that we require no constructor functions. Note that FICUS is more general than
a number of less recent approaches (Ragavan et al. 1993;
Bagallo and Haussler 1990; Yang, Rendell, and Blix 1991;
Matheus and Rendell 1989; Hu and Kibler 1996).
Fan et al. (Fan et al. 2010) propose FCTree uses a decision tree to partition the data using both original and constructed features as splitting points (nodes in the tree). As in
FICUS (Markovitch and Rosenstein 2002), FCTree uses surrogate tree-based information-theoretic criteria to guide the
search, as opposed to the true prediction performance. FCTree is capable of generating only simple features, and is not
capable of composing transformations, i.e. it is search in a
smaller space than our approach. They also propose a weight
update mechanism that helps identify good transformations
for a dataset, such that they are used more frequently.
The Deep Feature Synthesis component of Data Science
Machine (DSM) (Kanter and Veeramachaneni 2015) relies
on applying all transformations on all features at once (but
no combinations of transformations), and then performing
feature selection and model hyper-parameter optimization
over the combined augmented dataset. A similar approach
is adopted by One Button Machine (Lam et al. 2017). We
will call this category as the expansion-reduction approach.
This approach suffers performance performance and scalability bottleneck due to performing feature selection on a
large number of features that are explicitly generated by simultaneous application of all transforms. In spite of the expansion of the explicit expansion of the feature space, it does
not consider the composition of transformations.
FEADIS (Dor and Reich 2012) relies on a combination
of random feature generation and feature selection. It adds
constructed features greedily, and as such requires many
expensive performance evaluations. A related work, ExploreKit (Katz et al. 2016) expands the feature space explicitly. It employs learning to rank the newly constructed
features and evaluating the most promising ones. While this
approach is more scalable than the expand-select type, it still
is limited due to the explicit expansion of the feature space,
and hence time-consuming. For instance, their reported results were obtained after running FE for days on moderately
sized datasets. Due to the complex nature of this method, it
does not consider compositions of transformations. We refer
to this approach of FE as evolution-centric.
Cognito (Khurana et al. 2016) introduces the notion of
a tree-like exploration of transform space; they present a
few simple handcrafted heuristics traversal strategies such as
breadth-first and depth-first search that do not capture several factors such as adapting to budget constraints. This paper generalizes the concepts introduced there. LFE (Nargesian et al. 2017) proposes a learning based method to predict
the most likely useful transformation for each feature. It considers features independent of each other; it is demonstrated
to work only for classification so far, and does not allow for
composition of transformations.
Other plausible approaches to FE are hyper-parameter optimization (Bergstra et al. 2011) where each transformation
choice could be a parameter, black-box optimization strategies (Li et al. 2016), or bayesian optimization such as the
ones for model- and feature-selection (Feurer et al. 2015).
To the best of our knowledge, these approaches have been
employed for solving FE. (Smith and Bull 2003) employ a
genetic algorithm to determine a suitable transformation for
a given data set, but is limited to single transformations.
Certain ML methods perform some level of FE indirectly.A recent survey on the topic appears can be found
here (Storcheus, Rostamizadeh, and Kumar 2015). Dimensionality reduction methods such as Principal Component Analysis (PCA) and its non-linear variants (Kernel PCA) (Fodor 2002) aim at mapping the input dataset
into a lower-dimensional space with fewer features.Such
methods are also known as embedding methods (Storcheus,
Rostamizadeh, and Kumar 2015). Kernel methods (ShaweTaylor and Cristianini 2004) such as Support Vector Machines (SVM) are a class of learning algorithms that use kernel functions to implicitly map the input feature space into a
higher-dimensional space.
Multi-layer neural networks allow for useful features to
be learned automatically, such that they minimize the training loss function. Deep learning methods have made remarkable successes on various data such as video, image
and speech, where manual FE is very tedious. (Bengio,
Courville, and Vincent 2013). However, deep learning methods require massive amounts of data to avoid overfitting and
are not suitable for problems instances of small or medium
sizes, which are quite common. Additionally, deep learning
has mostly been successful with video, image, speech and
natural language data, whereas the general numerical types
of data encompasses a wide variety of domains and need
FE. Our technique is domain, and model independent and
works generally irrespective of the scale of data. Also, the
features learned by a deep network may not always be easily explained, limiting application domains such as healthcare (Che et al. 2015). On the contrary, features generated
by our algorithm are compositions of well-understood mathematical functions that can be analyzed by a domain expert.
Overview
The automation of FE is challenging computationally, as
well as in terms of decision-making. First, the number of
possible features that can be constructed is unbounded since
transformations can be composed, i.e., applied repeatedly
to features generated by previous transformations. In order to confirm whether a new feature provides value, it requires training and validation of a new model upon including the feature. It is an expensive step and infeasible to perform with respect to each newly constructed feature. The
evolution-centric approaches described in the Related Work
section operate in such manner and take days to complete
even on moderately-sized datasets. Unfortunately, there is
no reusability of results from one evaluation trial to another.
On the other hand, the expansion-reduction approach performs fewer or only one training-validation attempts by first
explicitly applying all transformations, followed by feature
selection on the large pool of features. It presents a scalability and speed bottleneck itself; in practice, it restricts
the number of new features than can be considered. In both
cases, there is a lack of performance oriented search. With
these insights, our proposed framework performs a systematic enumeration of the space of FE choices for any given
dataset through a transformation graph. Its nodes represent
different versions of a given datasets obtained by the application of transformation functions (edges). A transformation
when applied to a dataset, applies the function on all possible features (or sets of features in case non-unary functions),
and produces multiple additional features, followed by optional feature selection, and training-evaluation. Therefore,
it batches the creation of new features by each transformation function. This lies somewhat in the middle of the
evolution-centric and the expansion-reduction approaches.
It not only provides a computation advantage, but also a
logical unit of measuring performance of various transforms, which is used in composing different functions in a
performance-oriented manner. This translates the FE problem to finding the node (dataset) on the transformation graph
with the highest cross-validation performance, while only
exploring the graph as little as possible. It also allows for
composition of transformation functions.
Secondly, the decision making in manual FE exploration
involves initiation and complex associations, that are based
on a variety of factors. Some examples are: prioritizing
transformations based on the performance with the given
dataset or even based on past experience; whether to explore
different transformations or exploit the combinations of the
ones that have shown promise thus far on this dataset, and so
on. It is hard to articulate the notions or set of rules that are
the basis of such decisions; hence, we recognize the factors
involved and learn a strategy as a function of those factors in
order to perform the exploration automatically. We use reinforcement learning on FE examples on a variety of datasets,
to find an optimal strategy. This is based on the transformation graph. The resultant strategy is a policy that maps each
instance of the transformation graph to the action of applying a transformation on a particular node in the graph.
Notation and Problem Description
Consider a predictive modeling task consisting of (1) a set
of features, F = {f1 , f2 . . . fm }; (2) a target vector, y. A
pair of the two is specified as a dataset, D = hF, yi. The nature of y, whether categorical or continuous, describes if it
is a classification or regression problem, respectively. For an
applicable choice of learning algorithm L (such as Random
Forest or Linear Regression) and a measure of performance,
m (such as AUROC or -RMSE). We use Am
L (F, y) (or simply, A(F ) or A(D)) to signify the cross-validation performance of a the model constructed on given data with using
the algorithm L through the performance measure m.
Additionally, consider a set of k transformation functions
at our disposal, T = {t1 , t2 . . . tk }. The application of a
transformation on a set of features, t(F ), suggests the application of the corresponding function on all valid input feature subsets in F , applicable to t. For instance, a square
transformation applied to a set of features, F with eight numerical and two categorical features will produce eight new
output features, fo = square(fi ), ∀fi ∈ F, fi ∈ Rn . This
extends to k-ary functions, which work on k input features.
A derived feature is recognized with a hat, fˆ. The entire
(open) set of derived features from F is denoted by F̂ .
A ‘+’ operator on two feature sets (associated with the
same target y) is a union of the two feature sets, Fo =
F1 + F2 = F1 ∪ F2 , preserving row order. Generally, transformations add features; on the other hand, a feature selection operator, which is to a transformation in the algebraic
notation, removes features. Note that all operations specified on a feature set , T (F ), can exchangeably be written for
a corresponding dataset, D = hF, yi, as, T (D), where it is
implied that the operation is applied on the corresponding
feature set. Also, for a binary, such as sum, Do = D1 + D2 ,
it is implied that the target is common across the operands
and the result.
The goal of feature engineering is stated as follows. Given
a set of features, F , and target, y, find a set of features, F ∗
where each feature is in F ∗ = F1 ∪ F2 , where F1 ⊆ F
(original) and F2 ⊂ F̂ (derived), to maximize the modeling
accuracy for a given algorithm, L and measure, m.
F ∗ = arg max Am
L (F1 ∪ F2 , y)
F1 ,F2
(1)
even a height bounded tree with a modest number of transforms, computing the cross-validation performance across
the entire graph is computationally expensive.
D0
log
square
sum
D1
sum
D2
square
FS1
square
log
D6
D4
D5
Graph Exploration under a Budget Constraint
D3
D7
log
sum
D8
D9
D10
D4,9 =D4+D9
D5,3 =D5+D3
Figure 3: Example of a Transformation Graph (a DAG). The
start node D0 corresponds to the given dataset; that and the
hierarchical nodes are circular. The sum nodes are rectangular. In this example, we can see three transformations, log,
sum, and square, as well as a feature selection operator F S1 .
Transformation Graph
Transformation Graph, G, for a given dataset, D0 , and a finite set of transformations, T , is a directed acyclic graph
in which each node corresponds to a either D0 or a dataset
derived from it through transformation path. Every node’s
dataset contains the same target and number of rows. The
nodes are divided into three categories: (a) The start or the
root node, D0 corresponding to the given dataset; (b) Hierarchical nodes, Di , where i > 0, which have one and only
one incoming node from a parent node Dj , i > j ≥ 0, and
the connecting edge from Dj to Di corresponds to a transform T ∈ T (including feature selection), i.e., Dj = T (Di );
+
(c) Sum nodes, Di,j
= Di + Dj , a result of a dataset sum
such that i 6= j. Similarly, edges correspond to either transforms or ‘+’ operations, with children as type (b) or type
(c) nodes, respectively. The direction of an edge represents
the application of transform from source to a target dataset
(node). Height (h) of the transformation graph refers to the
maximum distance between the root and any other node.
A transformation graph is illustrated in Figure 3. Each
node of a transformation graph is a candidate solution for the
FE problem in Equation 1. Also, a complete transformation
graph must contain a node that is the solution to the problem, through a certain combination of transforms including
feature selection. The operator θ(G) signifies all nodes of
graph G, and θh (G) signifies all hierarchical nodes. Also,
λ(Di , Dj ) signifies the transformation T , such that its application on Di created Dj as its child; alternatively if Dj is a
sum node and Di is one of its parents, then λ(Di , Dj ) = +.
A complete transformation graph is unbounded for a nonempty transformation set. A constrained (bounded height,
h) complete transformation graph for t transformations will
have th+1 − 2 hierarchical nodes (and an equal number
h+1
h+1
−2)
of corresponding edges), and (t −1)×(t
sum nodes
2
(and 2 times corresponding edges). It can be seen that for
It should be emphasized that the exhaustive exploration of
a transformation graph is not an option, given its massive
potential size. For instance, with 20 transformations and a
height = 5, the complete graph contains about 3.2 million
nodes; an exhaustive search would imply as many model
training and testing iterations. On the other hand, there is
no known property that allows us to deterministically verify
the optimal solution in a subset of the trials. Hence, the focus of this work is to find a performance driven exploration
policy, which maximizes the chances of improvement in accuracy within in a limited time budget. The exploration of
the transformation graph begins with a single node, D0 , and
grows one node at a time from the then current state of the
graph. In the beginning it is reasonable to perform exploration of the environment, i.e., stumble upon the transforms
that signal an improvement. Over time (elapsed budget) it
is desirable to reduce the amount of exploration and focus
more on exploitation.
Algorithm 1: Transformation Graph Exploration
Input: Dataset D0 , Budget Bmax ;
Initialize G0 with root D0 ;
while i < Bmax do
N ← θ(Gi );
i
bratio = Bmax
n∗ , t∗ ←
arg max
R(Gi , n, t, bratio )
n,t@n0 ∀t=λ(n,n0 )
Gi+i ← Apply t∗ to n∗ in Gi
i ← i + 1;
Output: argmax A(θ(Gi ))
D
Algorithm 1 outlines our general methodology for exploration. At each step, an estimated reward from each posi
sible move, R(Gi , n, t, Bmax
) is used to rank the options
of actions available at each given state of the transformation graph Gi , ∀i ∈ [0, Bmax ), where Bmax is the overall allocated budget in number of steps4 . Note that the algorithm allows for different exploration strategies, which
is left up to the definition of the function R(. . . ), which
defines the relative importance of different steps at each
step. The parameters of the function suggest that it depends on the various aspects of the graph at that point, Gi ,
the remaining budget, and specifically, attributes of the action (node+transformation) being characterized. Below, we
briefly discuss such factors that influence the exploration
choice at each step. These factors are compared across all
choices of node-transformation pairs < n, t > at Gi :
4
Budget can be considered in terms of any quantity that is monotonically increasing in i, such as time elapsed. For simplicity, we work with “number of steps”.
1. Node n’s Accuracy: Higher accuracy of a node incentives
further exploration from that node, compared to others.
2. Transformation, t’s average immediate reward till Gi .
3. Number of times transform t has already been used in the
path from root node to n.
4. Accuracy gain for node n (from its parent) and gain for
n’s parent, i.e., testing if n’s gains are recent.
5. Node Depth: A higher value is used to penalize the relative complexity of the transformation sequence.
6. The fraction of budget exhausted till Gi .
7. Ratio of feature counts in n to the original dataset: This
indicates the bloated factor of the dataset.
8. Is the transformation a feature selector?
9. Whether the dataset contain (a) numerical features; or (b)
datetime features; or (c) string features?
Simple graph traversal strategies can be handcrafted. A
strategy essentially translates to the design of the reward estimation function, R(. . . ). In line with Cognito (Khurana et
al. 2016), a breadth-first or a depth-first strategy, or perhaps
a mix of them can be described. While such simplistic strategies work suitably in specific circumstances, it seems hard
to handcraft a unified strategy that works well under various
circumstances. We instead turn to machine learning to learn
the complex strategy from several historical runs.
Traversal Policy Learning
So far, we have discussed a hierarchical organization of FE
choices through a transformation graph and a general algorithm to explore the graph in a budget allowance. At the
heart of the algorithm is a function to estimate reward of an
action at a state. The design of the reward estimation function determines the strategy of exploration. Strategies could
be handcrafted; however, in this section we try to learn an
optimal strategy from examples of FE on several datasets
through transformation graph exploration. Because of the
behavioral nature of this problem - which can be perceived
as continuous decision making (which transforms to apply
to which node) while interacting with an environment (the
data, model, etc.) in discrete steps and observing reward (accuracy improvement), with the notion of a final optimization
target (final improvement in accuracy), it is modeled as a RL
problem. We are interested in learning an action-utility function to satisfy the expected reward function, R(. . . ) in Algorithm 1. In the absence of an explicit model of the environment, we employ Q-learning with function approximation
due to the large number of states (recall, millions of nodes
in a graph with small depth) for which it is infeasible to learn
state-action transitions explicitly.
Consider the graph exploration process as a Markov Decision Process (MDP) where the state at step i is a combination of two components: (a) transformation graph after i
node additions, Gi (G0 consists of the root node corresponding to the given dataset. Gi contains i nodes); (b) the remaini
ing budge at step i, i.e., bratio = Bmax
. Let the entire set of
states be S. On the other hand, an action at step i is a pair of
existing tree node and transformation, i.e., < n, t > where
n ∈ θ(Gt ), t ∈ T and @n ∈ Gi ∀λ(n, n0 ) = t; it signifies
the application of the one transform (which hasn’t already
been applied) to one of the exiting nodes in the graph. Let
the entire set of actions be C. A policy, Π : S → C, determines which action is taken given a state. Note that the
objective of RL here is to learn the optimal policy (exploration strategy) by learning the action-value function, which
we elaborate later in the section.
Such formulation uniquely identifies each state. Considering the “remaining budget” as factor in the state of the
MDP helps us address the runtime exploration versus exploitation trade-off for a given dataset. Note that this runtime
explore/exploit trade-off is not identical to the commonly referred trade-off in RL training in context of selecting actions
to balance reward and not getting stuck in a local optimum.
At step i, the occurrence of an action results in a new
node, ni , and hence a new dataset on which a model is
trained and tested, and its accuracy A(ni ) is obtained. To
each step, we attribute an immediate scalar reward:
ri =
max
n0 ∈θ(Gi+1 )
A(n0 ) − max A(n)
n∈θ(Gi )
with r0 = 0, by definition. The cumulative reward over time
from state si onwards is defined as:
R(si ) =
BX
max
γ i .ri+j
j=0
where γ ∈ [0, 1) is a discount factor, which prioritizes early
rewards over the later ones. The goal of RL here is to find the
optimal policy Π∗ that maximizes the cumulative reward.
We use Q-learning (Watkins and Dayan 1992) with function approximation to learn the action-value Q-function. For
each state, s ∈ S and action, c ∈ C, Q-function with respect
to policy Π is defined as:
Q(s, c) = r(s, c) + γRΠ (δ(s, c))
where δ : S × C → S is a hypothetical transition function,
and RΠ (s) is the cumulative reward following state s. The
optimal policy is:
Π∗ (s) = arg max[Q(s, c)]
(2)
c
However, given the size of S, it is infeasible to learn
Q-function directly. Instead, a linear approximation the Qfunction is used as follows:
Q(s, c) = wc .f (s)
(3)
c
where w is a weight vector for action c and f (s) =
f (g, n, t, b) is a vector of the state characteristics described
in the previous section and the remaining budget ratio.
Therefore, we approximate the Q-functions with linear combinations of characteristics of a state of the MDP. Note that,
in each heuristic rule strategy, we used a subset of these state
characteristics, in a self-conceived manner. However, in the
ML based approach here, we select the entire set of characteristics and let the RL find the appropriate weights of those
characteristics (for different actions). Hence, this approach
generalizes the other handcrafted approaches.
The update rule for wc is as follows:
wcj ← wcj + α.(rj + γ. max
Q(g 0 , c0 ) − Q(g, c)).f (g, b)
0 0
n ,t
(4)
where g 0 is the state of the graph at step j + 1, and α is
the learning rate parameter. The proof follows from (Irodova
and Sloan 2005).
A variation of the linear approximation where the coefficient vector w is independent of the action c, is as follows:
Q(s, c) = w.f (s)
Handcrafted
(5)
This method reduces the space of coefficients to be learnt
by a factor of c, and makes it faster to learn the weights. It is
important to note that the Q-function is still not independent
of the action c, as one of the factors in f (s) or f (g, n, t, b) is
actually the average immediate reward for the transform for
the present dataset. Hence, Equation 5 based approximation
still distinguishes between various actions (t) based on their
performance in the transformation graph exploration so far;
however, it does not learn a bias for different transformations
in general and based on the feature types (factor #9). We
refer to this type of strategy as RL2 . In our experiments RL2
efficiency is somewhat inferior to the strategy to the strategy
learned with Equation 3, which we refer to as RL1 .
Experiments
Training: We used 48 datasets (not overlapping with test datasets) to select training examples using different values for maximum budget,
Bmax ∈ {25, 50, 75, 100, 150, 200, 300, 500} with each
dataset, in a random order. We used the discount factor,
γ = 0.99, and learning rate parameter, α = 0.05. The
weight vectors, wc or w, each of size 12, were initialized
with 1’s. The training example steps are drawn randomly
with the probability = 0.15 and the current policy with
probability 1 − . We have used the following transformation functions in general (except when specified a smaller
number): Log, Square, Square Root, Product, ZScore,
Min-Max-Normalization, TimeBinning, Aggregation (using
Min,Max,Mean,Count,Std), Temporal window aggregate,
Spatial Aggregation, Spatio Temporal Aggregation, k-term
frequency, Sum, Difference, Division, Sigmoid, BinningU,
BinningD, NominalExpansion, Sin, Cos, TanH.
Comparison: We tested the impact of our FE on 24 publicly available datasets (different from the datasets used in
training) from a variety of domains, and of various sizes.
We report the accuracy of (a) base dataset; (b) Our FE routine with RL1 , Bmax = 100; (c) Expansion-reduction implementation where all transformations are first applied separately and add to original columns, followed by a feature
selection routine; (d) Random: randomly applying a transform function to a random feature(s) and adding the result
to the original dataset and measuring the CV performance;
this is repeated 100 times and finally, we consider all the
new features whose cases showed an improvement in performance, along with the original features to train a model
(e) Tree-Heur: our implementation of Cognito’s (Khurana
et al. 2016) global search heuristic for 100 nodes. We used
Random Forest with default Weka parameters as our learning algorithm for all the comparisons as it gave us the
strongest baseline (no FE) average. A 5-fold cross validation using random stratified sampling was used. The results
for a representative are captured in Table 1. It can be seen
Learned
Figure 4: Comparing efficiencies of exploration policies.
that our FE outperforms others in most of the cases but one
(where expand-reduce is better) and tied for two with Cognito global search. Our technique reduces the error (relative
abs. error, or 1- mean unweighted FScore) by 23.8% (by
median) for the 24 datasets presented in Table 1. For reference to time taken, it took the Bikeshare DC dataset 4 minutes, 40 seconds to run for 100 nodes for our FE, on a single
thread on a 2.8GHz processor. Times taken by the Random
and Cognito were similar to our FE for all datasets, while
expand-reduce took 0.1 to 0.9 times the time of our FE, for
different datasets.
Traversal Policy Comparison: In Figure 4, we see that on
an average for 10 datasets, the RL-based strategies are 4-8
times more efficient than any handcrafted strategy (breadthfirst, depth-first and global as described in (Khurana et al.
2016)), in finding the optimal dataset in a given graph with 6
transformations and bounded height, hmax = 4. Also, Figure 5 tells us that while RL1 (Eqn. 3) takes more data to
train, it is more efficient than RL2 (Eqn. 5), demonstrating
that learning a general bias for transformations and one conditioned on data types makes the exploration more efficient.
Figure 5: Policy effectiveness with training dataset sources
Dataset
Higgs Boson
Amazon Employee
PimaIndian
SpectF
SVMGuide3
German Credit
Bikeshare DC
Housing Boston
Airfoil
AP-omentum-ovary
Lymphography
Ionosphere
Openml 618
Openml 589
Openml 616
Openml 607
Openml 620
Openml 637
Openml 586
Credit Default
Messidor features
Wine Quality Red
Wine Quality White
SpamBase
Source
UCIrvine
Kaggle
UCIrvine
UCIrvine
LibSVM
UCIrvine
Kaggle
UCIrvine
UCIrvine
OpenML
UCIrvine
UCIrvine
OpenML
OpenML
OpenML
OpenML
OpenML
OpenML
OpenML
UCIrvine
UCIrvine
UCIrvine
UCIrvine
UCIrvine
C/R
C
C
C
C
C
C
R
R
R
C
C
C
R
R
R
R
R
R
R
C
C
C
C
C
Rows
50000
32769
768
267
1243
1001
10886
506
1503
275
148
351
1000
1000
500
1000
1000
500
1000
30000
1150
999
4900
4601
Features
28
9
8
44
21
24
11
13
5
10936
18
34
50
25
50
50
25
50
25
25
19
12
12
57
Base
0.718
0.712
0.721
0.686
0.740
0.661
0.393
0.626
0.752
0.615
0.832
0.927
0.428
0.542
0.343
0.380
0.524
0.313
0.547
0.797
0.691
0.380
0.677
0.955
Our RL1
0.729
0.806
0.756
0.788
0.776
0.724
0.798
0.680
0.801
0.820
0.895
0.941
0.587
0.689
0.559
0.647
0.683
0.585
0.704
0.831
0.752
0.387
0.722
0.961
Exp-Red
0.682
0.744
0.712
0.790
0.711
0.680
0.693
0.621
0.759
0.725
0.727
0.939
0.411
0.650
0.450
0.590
0.533
0.581
0.598
0.802
0.703
0.344
0.654
0.951
Random
0.699
0.740
0.709
0.748
0.753
0.655
0.381
0.637
0.752
0.710
0.680
0.934
0.428
0.571
0.343
0.411
0.524
0.313
0.549
0.766
0.655
0.380
0.678
0.937
Tree-Heur
0.702
0.714
0.732
0.780
0.766
0.662
0.790
0.652
0.794
0.758
0.849
0.941
0.532
0.644
0.450
0.629
0.583
0.582
0.647
0.799
0.762
0.386
0.704
0.959
Table 1: Comparing performance for base dataset (no FE), Our FE, Expansion-Reduction style FE, Random FE, and Tree
heuristic FE, using 24 datasets. Performance here is FScore for classification (C) and (1−rel. absolute error) for regression (R).
exploration cost due to a higher depth. Also, using feature
selection (compared to none) as a transform improves the
final gain in performance by about 51%, measured on the 48
datasets (aforementioned 24 + another 24). Finally, the use
of different models (learning algorithms) lead to different
optimal features being engineered for the same dataset, even
for similar improvements in performance.
Conclusion and Future Work
Figure 6: Performance vs. hmax
Internal System Comparisons: We additionally performed experimentation to test and tune the internals of
our system. Figure 6 shows the maximum accuracy node
(for 5 representative datasets) found when the height was
constrained to a different numbers, using Bmax = 100
nodes; hmax = 1 signifies base dataset. Majority of
datasets find the maxima with hmax = 4 with most find
it with hmax = 5. For hmax = 6, a tiny fraction shows
deterioration, which can be interpreted as unsuccessful
In this paper, we presented a novel technique to efficiently
perform feature engineering for supervised learning problems. The cornerstone of our framework are, a transformation graph that enumerates the space of feature options and
a RL-based, performance-driven exploration of the available
choices to find valuable features. The models produced using our proposed technique considerably reduce the error
rate (24% by median) across a variety of datasets, for a relatively small computational budget. This methodology can
potentially save a data analyst hours to weeks worth of time.
One direction to further improve the efficiency of the system
is through a complex non-linear modeling of state variables.
Additionally, extending the described framework to other aspects of predictive modeling, such as missing value imputation or model selection, is of potential interest as well. Since
optimal features depend on model type (learning algorithm),
a joint optimization of the two is particularly interesting.
References
[Bagallo and Haussler 1990] Bagallo, G., and Haussler, D.
1990. Boolean feature discovery in empirical learning. Machine learning 5(1):71–99.
[Bengio, Courville, and Vincent 2013] Bengio,
Y.;
Courville, A.; and Vincent, P. 2013. Representation
learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence
35(8):1798–1828.
[Bergstra et al. 2011] Bergstra, J. S.; Bardenet, R.; Bengio,
Y.; and Kégl, B. 2011. Algorithms for hyper-parameter optimization. In Shawe-Taylor, J.; Zemel, R. S.; Bartlett, P. L.;
Pereira, F.; and Weinberger, K. Q., eds., Advances in Neural
Information Processing Systems 24. Curran Associates, Inc.
2546–2554.
[Che et al. 2015] Che, Z.; Purushotham, S.; Khemani, R.;
and Liu, Y. 2015. Distilling knowledge from deep networks
with applications to healthcare domain. arXiv preprint
arXiv:1512.03542.
[Dor and Reich 2012] Dor, O., and Reich, Y. 2012. Strengthening learning algorithms by feature discovery. Information
Sciences 189:176–190.
[Fan et al. 2010] Fan, W.; Zhong, E.; Peng, J.; Verscheure,
O.; Zhang, K.; Ren, J.; Yan, R.; and Yang, Q. 2010. Generalized and heuristic-free feature construction for improved
accuracy. 629–640.
[Feurer et al. 2015] Feurer, M.; Klein, A.; Eggensperger, K.;
Springenberg, J. T.; Blum, M.; and Hutter, F. 2015. Efficient
and robust automated machine learning. NIPS.
[Fodor 2002] Fodor, I. K. 2002. A survey of dimension reduction techniques.
[Hu and Kibler 1996] Hu, Y.-J., and Kibler, D. 1996. Generation of attributes for learning algorithms. AAAI.
[Irodova and Sloan 2005] Irodova, M., and Sloan, R. H.
2005. Reinforcement learning and function approximation.
In FLAIRS Conference, 455–460.
[Kanter and Veeramachaneni 2015] Kanter, J. M., and
Veeramachaneni, K. 2015. Deep feature synthesis: Towards
automating data science endeavors. IEEE Data Science and
Advanced Analytics 1–10.
[Katz et al. 2016] Katz, G.; Chul, E.; Shin, R.; and Song, D.
2016. Explorekit: Automatic feature generation and selection. In IEEE ICDM, 979–984.
[Khurana et al. 2016] Khurana, U.; Turaga, D.; Samulowitz,
H.; and Parthasarathy, S. 2016. Cognito: Automated feature
engineering for supervised learning. In IEEE ICDM (Demo).
[Lam et al. 2017] Lam, H. T.; Thiebaut, J.-M.; Sinn, M.;
Chen, B.; Mai, T.; and Alkan, O. 2017. One button machine
for automating feature engineering in relational databases.
arXiv preprint arXiv:1706.00327.
[Li et al. 2016] Li, L.; Jamieson, K. G.; DeSalvo, G.; Rostamizadeh, A.; and Talwalkar, A. 2016. Efficient hyperparameter optimization and infinitely many armed bandits.
CoRR abs/1603.06560.
[Markovitch and Rosenstein 2002] Markovitch, S., and
Rosenstein, D. 2002. Feature generation using general
constructor functions. Machine Learning.
[Matheus and Rendell 1989] Matheus, C. J., and Rendell,
L. A. 1989. Constructive induction on decision trees. IJCAI.
[Nargesian et al. 2017] Nargesian, F.; Samulowitz, H.; Khurana, U.; Khalil, E. B.; and Turaga, D. 2017. Learning feature engineering for classification. In IJCAI, 2529–2535.
[Ragavan et al. 1993] Ragavan, H.; Rendell, L.; Shaw, M.;
and Tessmer, A. 1993. Complex concept acquisition through
directed search and feature caching. IJCAI.
[Shawe-Taylor and Cristianini 2004] Shawe-Taylor, J., and
Cristianini, N. 2004. Kernel methods for pattern analysis.
Cambridge university press.
[Smith and Bull 2003] Smith, M. G., and Bull, L. 2003. Feature Construction and Selection Using Genetic Programming and a Genetic Algorithm. Berlin, Heidelberg: Springer
Berlin Heidelberg. 229–237.
[Storcheus, Rostamizadeh, and Kumar 2015] Storcheus, D.;
Rostamizadeh, A.; and Kumar, S. 2015. A survey of modern
questions and challenges in feature extraction. Proceedings
of The 1st International Workshop on Feature Extraction,
NIPS.
[Watkins and Dayan 1992] Watkins, C. J., and Dayan, P.
1992. Q-learning. Machine learning 8(3-4):279–292.
[Wind 2014] Wind, D. K. 2014. Concepts in predictive machine learning. Master’s thesis, Technical University of Denmark.
[Yang, Rendell, and Blix 1991] Yang, D.-S.; Rendell, L.; and
Blix, G. 1991. Fringe-like feature construction: A comparative study and a unifying scheme. ICML.
| 2cs.AI
|
Finite Blocklength Rates over a Fading Channel
with CSIT and CSIR
arXiv:1705.07863v3 [cs.IT] 13 Nov 2017
Deekshith P K and Vinod Sharma
ECE Dept., Indian Institute of Science, Bangalore, India.
Email: {deeks, vinod}@ece.iisc.ernet.in
Abstract—In this work, we obtain lower and upper bounds
on the maximal transmission rate at a given codeword length
n, average probability of error and power constraint P̄ ,
over a finite valued, block fading additive white Gaussian noise
(AWGN) channel with channel state information (CSI) at the
transmitter and the receiver. These bounds characterize deviation
of the finite blocklength coding rates from the channel capacity
which is in turn achieved by the water filling power allocation
across time. The bounds obtained also characterize the rate
enhancement possible due to the CSI at the transmitter in the
finite blocklength regime. The results are further elucidated via
numerical examples.
I. I NTRODUCTION
Next generation cellular networks ought to handle mission
critical data with delay requirements far more stringent than
that in present day cellular networks [1]. Refined engineering
insights to build such delay critical systems can be obtained
using the analytical methods pioneered in [2], [3], [4]. In this
work, we characterize data rate enhancement in a wireless
system with delay constraints by means of power adaptation,
when the transmitter has certain side information about the
channel. We restrict our attention to the delay incurred at the
physical layer in sending the codeword to the receiver.
In a cellular system, if the instantaneous channel gain can
be fed back to the transmitter, the transmitter can use this
knowledge to perform power control so as to increase the
overall data rate. In particular, under the assumption of perfect
CSI at the transmitter (CSIT) and the receiver (CSIR), the
optimal power allocation with no delay constraints over a flat
fading AWGN channel has the well known interpretation of
water filling in time ([5]).
With delay constraints imposed at the physical layer, the
traditional approach to study rate enhancement due to the
knowledge of CSIT is to first characterize either the delay
limited, outage or average capacity (see [6] for details) and
then obtain a power allocation strategy that maximizes the
required one of these quantities. In this regard, [7] obtains the
optimal power allocation that maximises the outage capacity
under the assumption of non-causal CSIT. In [8], the authors
obtain the optimal power allocation scheme maximizing the
average capacity with causal CSIT. Nonetheless, the above
mentioned schemes do not provide a realistic metric to evaluate the performance of actual delay sensitive systems. This
is because, the various notions of capacity used therein are
inherently asymptotic.
In this work, we provide lower and upper bounds on the
maximal channel coding rate over a block fading AWGN
channel with CSIT and CSIR in finite blocklength regime
under two kinds of constraints on the transmitted codewords.
Consequently, we characterize the rate enhancement possible
due to power adaptation at the transmitter. Our assumption of
perfect CSIT is idealistic. Nevertheless, rates obtained under
this assumption provide upper bounds for rates achievable
without CSIT or with imperfect CSIT and is commonly made
in literature. Also, knowledge of the power control strategies
suitable for delay constrained systems sheds insights into how
system energy is to be used in such systems. Efficient usage
of system energy can be beneficial for energy constrained
transmitters ought to become prominent in future wireless
networks [9].
An overview of the preceeding works is presented next. A
scalar coherent fading channel with stationary fading (generalization of block fading) without CSIT is considered in [10]
and the dispersion term is characterized. In [11], the authors
show that the (second order) optimal power allocation scheme
over a quasi static fading channel with CSIT and CSIR is
truncated channel inversion. The quasi static fading model is
a special case of the block fading model that we consider.
However, our bounds involve asymptotic terms, asymptotic in
the number of blocks and is derived under the assumption of a
finite valued fading process. A MIMO Rayleigh block fading
channel with no CSIT and CSIR is considered in [12] and
achievability and converse bounds are derived for the short
packet communication regime. In [13], the authors consider
a discrete memoryless channel with non-causal CSIT and
CSIR, and obtains the second order coding rates under general
assumptions on the state process. In contrast, we consider
a cost constrained setting with real valued channel inputs.
This renders a direct translation of the techniques used therein
infeasible. A high-SNR normal approximation of the maximal
coding rate over a block fading Rayleigh channel without CSIT
and CSIR is obtained in [14].
Our contribution is a finer characterization of the delay
limited performance of a wireless link with channel state
feedback, under two kinds of power constraints the wireless
transmitters are normally subjected to. The bounds obtained
(on the maximal coding rate) characterize the rate enhancement due to power adaptation for a given codeword length
and error probability. In deriving these bounds, the CSIT
assumption makes the analysis involved and non trivial. In
particular, in obtaining the upper bounds, the dependence of
the channel input on the fading states makes the corresponding
optimization problems difficult to solve. To circumvent this,
we derive alternate bounds utilizing the properties of asymptotically optimal power allocation scheme, viz, the water filling
scheme.
The paper is organized as follows. In Section II, we introduce the system model and notation. We provide lower bounds
on the maximal channel coding rate in Section III. Next, in
Section IV, we provide upper bounds for the maximal coding
rate. In Section V, we compare the bounds numerically and
exemplify the utility of the bounds derived. We conclude in
Section VI. Proofs are delegated to the appendices.
II. M ODEL AND N OTATION
We consider a point to point discrete time, memoryless
block fading channel subject to AWGN noise with density
2
N (0, σN
), where N (a, b) denotes Gaussian density with mean
a and variance b. The noise is independent and identically
distributed (i.i.d.) across channel uses. Let Tc denote the
channel coherence time which is the duration over which
the gain of the underlying physical channel remains constant.
Let D (in appropriate time units) denote the delay constraint
imposed on the communication. Then, B = d TDc e denotes
the number of blocks over which the communication spans
(where dxe denotes the smallest integer ≥ x). Let nc denote
the number of times the channel is used within a block. Then,
the number of channel uses for the whole of communication,
or equivalently, the codeword length n = Bnc .
The channel gain or the fading coefficient in block b is
denoted as Hb , which is a random variable taking values
in a finite set H = {η1 , . . . , η|H| } such that min{ηi :
1 ≤ i ≤ |H|} > 0. Here |H| (the cardinality of the set) denotes
the number of fading states. Let qi denote the probability of
Hb taking value ηi and qi > 0. The channel gains are i.i.d.
across blocks and is independent of the additive noise process.
The instantaneous channel gains are assumed to be known to
the transmitter as well as the receiver and the transmitter gets
to know them only causally (we refer to it as the full CSIT
and CSIR assumption).
Let X(b−1)nc +k denote the channel input corresponding to
the k th channel use in the bth block, where k ∈ {1, . . . , nc },
b ∈ {1, . . . , B}. For convenience, from here on, [bk] ≡
2
)) and
(b − 1)nc + k. Let Z[bk] (distributed as N (0, σN
Y[bk] denote the corresponding noise variable and the channel
output, respectively. Then, Y[bk] = Hb X[bk] + Z[bk] . If the
delay D tends to infinity (and hence, the number of blocks B
tends to infinity), it is well known that the channel capacity
is given by ([5])
"
#
1
H 2 PWF (H)
C(P̄ ) , EH
log 1 +
,
(1)
2
2
σN
+
σ2
where PWF (H) , λ − HN2 , (.)+ = max(0, .), EH [.]
denotes the expectation with respect to the distribution of the
random
variable
h
i H and λ is obtained by solving the equation
EH PWF (H) = P̄ . Here, PWF (.) is the water filling solution
with average power constraint P̄ .
We make note of certain notations
that we use throughout.
Let C(x) , 12 log 1 + σx2 , L(x) , σ2 x+x , V (x) ,
N
N
2
1 − 1 − L(x) . A set of positive integers {1, 2, . . . , N }
is denoted as [1: N ]. Wherever required,
Pnwe denote a sum
of n numbers ak , k ∈ [1 : n] ,
k=1 ak , as Sn (a).
We choose to represent the channel input and output vectors
conveniently as a collection of B vectors, each of length nc .
Thus, the channel input x ≡ (x1 , . . . , xn ) = (x1 , . . . , xB ),
where xb = (x[b1] , . . . , x[bnc ] ), b ∈ [1 : B]. Similarly, we
have the noise vector z = (z 1 , . . . , z B ) and the channel
output vector y = (y 1 , . . . , y B ). Also, the vector of channel
fading gains h = (h1 , . . . , hB ) is a collection of B scalars.
Corresponding random variables are denoted as X, Z, Y and
H. Let ηmin , min{ηi : i ∈ [1 : |H|]}, qmin , min{qi :
i ∈ [1 : |H|]}. Similarly, the corresponding maximum values
are denoted as ηmax and qmax , respectively. The Cartesian
product of two sets S1 , S2 is denoted as S1 × S2 and n
fold Cartesian product of sets S1 , . . . , Sn is denoted as S n .
The set of integers is denoted as Z and the set of positive
integers as Z+ . The real line is denoted as R, positive real
line as R+ , the n dimensional Euclidean space by Rn and
Rn+ , R+ ×. . .×R+ (n times). Given vectors x, y ∈ Rn , ||x||
denotes the Euclidean norm of x and hx, yi denotes the inner
product between x and y. The variance of a random variable X
is denoted as V[X]. The function Φ(.) denotes the cumulative
distribution function (cdf) of a standard Gaussian random
variable, φ(.) denotes the corresponding density function and
Φ−1 (.) denotes its inverse cdf. The notation fn = o(gn ) is
equivalent to lim fn /gn = 0. Also, fn = O(gn ) is equivalent
n→∞
to |fn | ≤ K|gn |, for some constant K, for all n sufficiently
large. The notation lim ≡ lim sup. To indicate relations that
hold almost surely, we use the abbreviation a.s.. We use the
D
notation = to mean equivalence in distribution. We denote the
indicator function of an event as 1A . The exponential function
is denoted as exp(.). For any set A, Ac denotes its complement.
All logarithms are taken to the natural base.
1
2
III. M AXIMAL C ODING R ATE : L OWER BOUNDS
In this section, we will obtain lower bounds on the maximal
coding rate for a given codeword length and average probability of error, under two different kinds of power constraints on
the transmitted codewords. We have the following definitions.
Let M ≡ [1 : M ] denote the message set. Let S be a
uniformly distributed random variable, corresponding to the
message transmitted, taking values in M. Given S ∈ M and
the fading coefficients (H1 , . . . , Hb ), at k th instance of channel
use in block b, the output of the encoder ζ[bk] : M × Hb → R
is denoted as X[bk] . The decoder ψ : Rn × HB → M,
on obtaining the (Y, H) pair, outputs an estimate of the
message Ŝ. The encoding and decoding is done
so that the
average probability of error P ψ(Y, H) 6= S ≤ , where
is prefixed. Throughout this work, we adhere to the average
probability of error formalism.
In this work, we consider two types of power constraints:
ST :
nc
B X
X
b=1 k=1
a.s
2
ζ[bk]
(m, H b ) ≤ Bnc P̄ , ∀m ∈ M.
(2)
"
LT : EH
nc
B X
X
#
2
(m, H b )
ζ[bk]
≤ Bnc P̄ , ∀m ∈ M. (3)
b=1 k=1
The constraint in equation (2) is referred to as the short
term power constraint (ST) and that in (3), as the long term
power constraint (LT). Studying a communication system
under the ST constraint is motivated by the peak power
limitations of the circuitry involved. Whereas, imposing the
LT constraint captures the requirement of power utilization
efficiency in a communication device (for instance, battery
powered mobile radio transmitters). In addition, in a wireless
communication setting, in studying systems which allocate
resources (e.g., rate, power) dynamically, the LT constraint
is a natural metric to consider. Though in reality these constraints are simultaneously present, in this work, we study
them in isolation. For each of the above constraint, the goal
is to characterize the maximum size (or cardinality) of the
codebook with block length n and average probability of error
, denoted as M ∗ (n, , P̄ ) ≡ M ∗ . The maximal coding rate
R∗ (n, , P̄ ) ≡ R∗ = n−1 log M ∗ .
Our first result gives a lower bound on log M ∗ under ST
constraint.
Theorem 1. For a block fading channel with input subject to
a short term power constraint P̄ and average probability of
error , the maximal codebook size M ∗ satisfies
r
q
√
n
∗
log M ≥ nC(P̄ ) −
+ nVBF (P̄ )Φ−1 () + o( n).
2
(4)
Here, with G , H 2 PWF (H),
1
VBF (P̄ ) , EG V (G) + nc V C(G) + V L(G) .
2
Proof. See Appendix A.
The following result provides a lower bound on log M ∗
subject to LT constraint.
Theorem 2. For a block fading channel with input subject
to a long term power constraint P̄ and average probability of
error , the maximal codebook size M ∗ at a given blocklength
n satisfies
q
√
log M ∗ ≥ nC(P̄ ) + nVBF (P̄ )Φ−1 () + o( n), (5)
where, the notation is as in Theorem 1.
Proof. See Appendix B.
IV. M AXIMAL C ODING R ATE : U PPER BOUNDS
In this section, we provide upper bounds on the maximal
coding rate under the ST and LT constraints. In deriving the
upper bounds, we assume that CSIT is known non-causally.
First, we obtain an upper bound for the ST case. Next, we
proceed to provide an upper bound for the LT case.
Theorem 3. For a block fading channel with input subject to
a short term power constraint P̄ , average probability of error
0 < < 21 , the maximal codebook size M ∗ satisfies
q
0 (P̄ )Φ−1 () + O(log n),
log M ∗ ≤ nC(P̄ ) + nVBF
(6)
0
where, with G has in Theorem 1 and iλ as in (1), VBF
(P̄ ) ,
L(G)
P̄
EG [V (G)] + V nc C(G) + 2λ − 2 .
Proof. See Appendix C.
Theorem 4. For a block fading channel with input subject to
a long term power constraint P̄ and average probability of
error 0 < < 12 , the maximal codebook size M ∗ satisfies
r
q
n
∗
0 (P̄ )Φ−1 () + O(log n),
+
nVBF
log M ≤ nC(P̄ ) +
4λ2
(7)
0
where, λ is as in (1) and VBF
(P̄ ) is as in Theorem 3.
Proof. See Appendix D.
V. N UMERICAL E XAMPLES
In this section, we compare numerically, the various bounds
obtained thus far. To that end, we need a refined √
characterization of the lower bounds. In particular, the o( n) terms
need
√ to be explicitly identified. Terms contributing to the
o( n) expression in the lower bound for the LT constraint
are same as that in the no CSIT case with ST constraint
([10]). Hence, the terms can be identified by revisiting the
analysis in [10] and computing the constants therein, explicitly√ for the i.i.d. √
block fading case. Doing so, we obtain
o( n) = log2 n − n1−β + O(nβ ), for some small β > 0.
For our computation, we fix β = 0.01.
Next, consider the lower bound for the ST constraint case.
Note that we are interested in error probability < 1/2 for
which, Φ−1 () < 0. Under this assumption, the right hand
side of (9) can be readily simplified using Taylor’s theorem
t
and the fact that VBF
(.) therein, is a monotonically increasing
function. By virtue of choice of δB (as mentioned in Appendix
0
−1
A), Φ−1 ( − δB
) = Φ
O(1/n1.5 ). Thus, we obtain
p () + p
√
−1
∗
log M ≥ nC(
√ P̄ ) − n/2 + nVBF (P̄ )Φ () + o( n),
where the o( n) term is same as in the LT case. The
coefficient of log n/n in the upper bound for the ST and LT
case is |H|/2. We compute the terms excluding the O(1/n1−β )
terms in the lower bound and O(1/n) term in the upper bound.
Thus, the approximate bound that we compute is akin to the
normal approximation for AWGN channels ([3]). With this,
we proceed to compute the bounds.
We consider the following discrete version of the Rayleigh
distribution: fix η0 = 0.1, |H| = 10, η|H|−1 = 4.1 and
∆ = (η|H|−1 − η0 )/(|H| − 1). Define ηi = ηi−1 + i∆, for
i ∈ [1 : |H| − 2] and qi = P (ηi ≤ HR ≤ ηi+1 ), for i ∈
[0 : |H| − 1] and q|H|−1 = P (HR ≥ η|H|−1 ) where HR ∼
2
Rayleigh distribution with parameter unity. Fix σN
= 1 and
= 0.01 and nc = 1 (which corresponds to the fast fading
case). By fixing P̄ = 5dB, we plot the convergence of various
bounds to the channel capacity C(P̄ ) = 0.6502 nats/channel
use (equation (1)), as B increases, in Figure 1. The acronyms
LB and UB refer to lower bound and upper bound. We have
also plotted the rates with no CSIT and ST constraint.
In Figure 2, we fix nc = 1, B = 4000 and compare the
various bounds for different values of P̄ . By comparing the
lower bound under the ST constraint with the rate for no CSIT
use C(M, n, , 1) in conjunction with a power controller. Next,
we explain the power control scheme.
σ 2 +
Choose λB such that EH λB − HN2
= P̄ − δB , for
some δB (to be chosen appropriately later with P̄ > δB > 0).
(B)
σ 2 +
Let PWF (H) = λB − HN2 . At the beginning of block b,
the transmitter checks if the constraint
CST :
b
X
||x0 ||2
l
l=1
Fig. 1. Rate versus blocklength.
nc
λB −
2 +
σN
≤ B P̄
Hl2
is met. If CST is not met, the ST constraint is violated. The
transmission is halted and an error is declared (a short error
message is sent to the receiver which it can receive without
error). Else, for the k th transmission in blockq
b, the channel
(B)
input X[bk] is chosen such that X[bk] = x0[bk] PWF (Hb ). If
the transmitter sends codeword symbols successfully in all B
blocks without violating CST for all b ∈ [1 : B],
B nc
B
1 X ||x0b ||2 (B)
1 XX
2
x[bk] =
PWF (Hb ) ≤ P̄
Bnc
B
nc
b=1 k=1
b=1
and the ST constraint is met. If transmission error does not
occur, for all
q b ∈ [1 : B], k ∈ [1 : nc ], the channel output
Fig. 2. Rate versus input power.
case under ST constraint, the rate enhancement possible due
to the knowledge of CSIT (via power control) is observed.
VI. C ONCLUSION
In this paper, we have obtained upper and lower bounds
for the maximal coding rate over a block fading channel with
short term and long term average power constraints on the
transmitted codeword. The bounds obtained shed light on the
rate enhancement possible due to the availability of CSIT.
The bounds also characterize the performance of water filling
power allocation in the finite block length regime.
ACKNOWLEDGEMENT
The authors would like to thank Gautam Konchady Shenoy,
ECE Dept., IISc, Bangalore and Wei Yang, Princeton University, for valuable discussions and useful comments.
A PPENDIX A
Lower bound: ST constraint
Let F (n) = {x0 ∈ Rn : ||x0 ||2 = n}. Let C(M, n, , 1) be
a codebook with M codewords, codeword length n, average
probability of error and codewords belonging to F (n) , for
the same block fading channel under consideration with CSIR
but no CSIT. These codewords satisfy the ST constraint. We
(B)
PWF (Hb )x0[bk] + Z[bk] .
q
D
(B)
Let G0 = G0b , Hb PWF (Hb ). Any rate achievable over
a block fading channel with i.i.d. fading process {G0b } with
CSIR and no CSIT satisfying an ST constraint of unity, is
also achievable, over a channel subject to block fading process
{Hb } with CSIT and CSIR satisfying an ST constraint of P̄ .
Henceforth, we consider the channel with block fading process
{G0b } (having CSIR but no CSIT).
0
denote an upper bound on the probability of the
Let δB
0
event that CST is violated. A specific choice of δB
will be
0
0
made later. Let τ = τ + δB , for some τ > 0. It is immediate
to see that the following bound (Theorem 25, [3]) holds: for
0
, 0 < τ < 0 ,
0 , − δ B
κτ F (n) , QY,G
,
M≥
(8)
supx0 ∈F (n) β1−0 +τ x0 , QY,G
Y[bk] = Hb
where βα (x0 , Q) ≡ βα (PY,G|X=x0 (y, g|x0 ), Q), QY,G is an
auxiliary output distribution, κτ (F (n) , Q) are as defined in [3]
and PY,G|X=x0 (y, g|x0 ) corresponds to the channel transition
probability. Now, from the analysis in [10], it follows that
q t −1 0
√
Φ ( − τ ) + o( n), (9)
log M ∗ ≥ nC P̄ − δB + nVBF
t
t
where
VBF
≡
VBF
(P̄ − δB ) = EG0 V (G02 ) + nc V C(G02 ) +
1
02
t
2 V L(G ) . Note that the difference between VBF and
t
VBF (P̄ ) (defined in the statement of Theorem 1) is that, VBF
0
is evaluated according to the distribution of G rather than G
defined in Theorem 1. As per definition, G0 is evaluated such
(B)
that EH [PWF (H)] = P̄ − δB . The first order term in the lower
bound obtained does not match with the channel capacity and
hence we proceed to rectify this shortcoming. For convenience,
00
00
let δB
= δB /P̄ . Here, δB
∈ (0, 1). Note that, for G as defined
in Theorem 1,
i
h
h
i
2 (a)
00
= EG C G(1 − δB
) .
C(P̄ − δB ) = EG0 C G0
Here (a) follows because of the following reason:
h
i
h
i
00
= P̄ − δB .
EH PWF (H) = P̄ ⇒ EH PWF (H) 1 − δB
For 00 > 0 and function η, Taylor’s theorem applied to the
functional F (.) about a certain function f is given by
F (f + 00 η) = F (f ) + 00
dF (f + yη)
dy
probability of violating the constraint CST in the following
way:
( b
)!
B
2 +
[
X ||x0 ||2
σN
b
λB − 2
P
> B P̄
nc
Hl
b=1
l=1
!
B
X
||x0b ||2 (B)
=P
PWF (Hb ) > B P̄
nc
b=1
!
B n
o
X
||x0b ||2 (B)
≤P
PWF (Hb ) − (P̄ − δB ) > BδB
nc
b=1
!
B
Bδ 2 (b) κ
X
(a)
B
≤ ,
=P
Sb > BδB ≤ exp −
2λ2
B
b=1
,
y=x
for some x ∈ (0, 00 ). Applying Taylor’s theorem to the
functional C(P̄ − δB ) yields
||x0 ||2
(B)
where, Sb , nbc PWF (Hb ) − (P̄ − δB ) and (a) follows from
Hoeffding’s inequality ([15]) as {Sb } is a set of i.i.d. random
B
P
Sb ∈ [−λB , λB ] ⊆
variables with zero mean such that
b=1
[−λ,
p λ]. For some constant κ, (b) follows by choosing δB =
h
i
λ 2/B (1−α) , for an arbitrary, small α > 0. For the specified
00
00
00
C(P̄ − δB )=C(P̄ ) − δB EG L(G(1 − α1 δB )) 2(1 − α1 δB ) ,
choice of δB , by Taylor’s theorem,
h
i
h
i
h
i
−1
00
for α1 ∈ p
(0, 1). Next, we apply Taylor’s theorem to the EĜ L(Ĝ)(2δ̂B )
= EG L(G)/2 − δB
EG L01 G(1 − u) ,
t (P̄ − δ ), and obtain
functional VBF
B
00
) and L01 G(1 − u) denotes the first
for some u ∈ (0, α1 δB
q
q
derivative of L1 G(1
t (P̄ − δ )= V (P̄ ) − δ 00 F (P̄ , α , δ )
√ − x) with respect to x at∗ x = u.
VBF
B
BF
2 B
B 1
Gathering all the o( n) terms,
we obtainpthat log M is upper
p
(1+α) /2 +
bounded
by
nC(
P̄
)
−
n
nVBF (P̄ )Φ−1 () +
√
for some α2 ∈ (0, 1). Here,
o( n). Since α is chosen arbitrarily small, the result follows.
q
F1 (P̄ , α2 , δB ) = F2 (P̄ , α2 , δB ) 2 VBF (P̄ − α2 δB ),
A PPENDIX B
t
where, VBF (P̄ − α2 δB ) is evaluated in the same way as VBF
2 +
0 σN
and VBF , but such that EH λ − H 2
= P̄ − α2 δB . Also,
00
2
2
with δ̄B , (1 − α2δB ), Ḡ , G δ̄B , G̃ , (σN
+ Ḡ) and
4
2
3
L1 , EG σN G G̃ , F2 (P̄ , α2 , δB ) ≡ F2 is given by
Lower bound: LT constraint
As in Appendix A, we fix the constraint set F (n) . Let
C(M, n, , 1) be a codebook with M codewords, codeword
length n, average probability of error and codewords belonging to F (n) , for a block fading channel (same as the one in
2
2
consideration)
with CSIR but no CSIT. By virtue of belonging
L1 + nc Co C(Ḡ), L(Ḡ) δ̄B + Co 1 − L(Ḡ), σN L Ḡ) G̃ .
to F (n) , the codewords inherently satisfy the ST constraint.
Here Co(., .) denotes the covariance function. Finally, we We use this codebook in conjunction with a power controller
apply Taylor’s theorem to the function Φ−1 (x), x ∈ [0 , ] in the following way: at the beginning of block b, the channel
0
0
so that Φ−1 (0 ) = Φ−1 () − δB
g1 (, δB
), for some α3 ∈ gain Hb for the next nc channel uses will be available to the
√
√ 2
transmitter. Fix the water filling power allocation PWF (.) in
0
0
) = 2πexp Φ−1 ( − α3 δB
) 2 . For
(0, 1) and g1 (, δB
equation (1). For the k th transmission in block
p b, the channel
0
0
0
notational convenience, let g2 ≡ g2 (, δB
) , δB
g1 (, δB
), input X
0
is
chosen
such
that
X
=
x
[bk]
[bk]
[bk] PWF (Hb ). Note
00
00
F3 ≡ F3 (P̄ , α2 , δB ) , δB
F1 (P̄ , α2 , δB ), δ̂B , (1 − α1 δB
), that
−1
Ĝ , δ̂B G2 and L1 (Ĝ) , L(Ĝ)(2δ̂B ) . Substituting these
nc
B X
hX
i
2
values in (9), we obtain
E
P (H )x0
= E P (H) ||x0 ||2 = nP̄ .
H
h
i q
L1 (Ĝ) + nVBF (P̄ )Φ−1 ()
00
log M ∗ ≥ nC(P̄ ) − nδB
EĜ
q
√
√
√
+ nF3 Φ−1 () − nVBF (P̄ )g2 − ng2 F2 + o( n).
Next, we fix the choice of δB so as to get an exact
0
00
expression for δB
and δB
. To that end, we upper bound the
WF
b
[bk]
H
WF
b=1 k=1
For each joint fading state H ∈ RB , the above scheme
generates a codebook satisfying LT constraint for the channel
with CSIT and CSIR on the fly, in a causal manner. The
channel output
p
Y[bk] = Hb PWF (Hb )x0[bk] + Z[bk] .
Now, nconsider a block
o fading channel with i.i.d. fading
p
process Hb PWF (Hb ) . Assume no CSIT and full CSIR.
Any rate achievable over such a channel with ST constraint
is achievable over the original channel (with full CSIT and
CSIR) subject to the LT constraint as well. It follows from
the analysis in [10] that
q
√
log M ∗ ≥ nC(P̄ ) + nVBF (P̄ )Φ−1 () + o( n).
Optimizing over the choice of input distributions, for the
auxiliary channel under consideration, we have
+
log M ∗ ≤ log γ − inf log P Iγ − .
(11)
FX
With the prescribed choice of the auxiliary channel, for H =
B
D P
h, X = x, under PX,H,Y , i(x, h, Y) =
Wb0 , where,
b=1
Wb0 = nc C(gb2 ) +
2
2 2
hxb , Z b i − gb2 ||Z b ||2
σN
hb ||xb ||2 + 2hb σN
,
2
2
2σN σN + gb2
A PPENDIX C
p
D
Upper bound: ST constraint
and gb = hb PWF (hb ). Next, it can be seen that Wb0 =
Let M ∗ ≡ M ∗ (n, , P̄ ) denote the maximal codebook Wb00 (Phb (x, h)) ≡ Wb00 , where, for p ≥ 0,
size for the channel with average probability of error and
√ 2
2 2
nc σN
hb p + 2hb pσN
h1b , Z b i − gb2 ||Z b ||2
D
2
satisfying the ST constraint. In particular, assume that the ST W 00 (p) =
n
C(g
)
+
,
c
b
b
2
2
2σN σN + gb2
constraint is satisfied with equality. Note that we can make this
assumption without loss of optimality (see, for instance, [16],
Lemma 65). Next, assume that CSIT is known non-causally. where, Phb (x, h) is as defined in (10) and 001b is the nc length
Bound derived under this assumption will give a valid upper vector (1, 1, . . . , 1) for all b. Note that {Wb } are independent
the conditional
bound for the causal case under consideration. Corresponding random variables. Let Fh ≡ FX|H=h denote
distribution
of
X
given
H
=
h.
Then,
P
Iγ
to each codeword x (satisfying ST constraint) and h (obtained
Z
i
h
X
non causally), we can identify a power
allotment vector
P(h)
P i(x, h, Y) ≤ log γ x, h dFh (x)
=
P(x, h) , P1 (x, h), . . . , P|H| (x, h) , where,
Rn
h∈HB
" B
#
0
,
n
(h)
=
0,
Z
i
X
X
nP
(h)
i
Pi (x, h) =
(10)
=
P(h)
P
Wb0 ≤ log γ x, h dFh (x)
x2kj (i,h) , ni (h) > 0.
ni1(h)
n
R
B
b=1
h∈H
j=1
" B
#
Z
X
X
(a)
B
P
0
≥
P(h)
inf P
Wb ≤ log γ x, h dFh (x),
Here, ni (h) = nc
1{hb =ηi } and kj (i, h) = {k ∈ [1 : n] :
n x∈Sh
b=1
hk = ηi }. Also, P(x, h) is such that
|H|
P
R
h∈HB
b=1
(12)
ti (h)Pi (x, h) = P̄ ,
i=1
where ti (h) ≡ ti = ni (h)/n. Since we assume perfect CSIT
and CSIR, let X̄ = (X, H) and Ȳ = (Y, H) denote the
equivalent channel input and output, respectively. Then, we
have ([3], Theorem 26) that
0
β1− (PX,H,Y , QX,H,Y ) ≤ 1 − ,
where βα (P, Q) is the minimum false alarm probability in
deciding between P and Q, subject to a minimum detection
probability α > 0, Q denotes an auxiliary channel and 0 is
the average probability of error for the auxiliary channel. We
choose the auxiliary channel where the channel output Y has
the distribution
nc
B Y
Y
2
QX,H,Y =
N 0, σN
+ Hb2 PWF (Hb ) .
where, in (a), Sh denotes the support of Fh . Next, note that,
" B
#
X
0
inf P
Wb ≤ log γ x, h
x∈Sh
(a)
=
b=1
"
inf
P(h)∈Πh
P
B
X
#
Wb00
Phb (h) ≤ log γ h
b=1
"
#
B
1 X
0
= inf P √
Wb ≤ γB h, P(h) h
νB
P(h)∈Πh
b=1
h
(c)
i
0
≥ inf
Φ γB
h, P(h) − B h, P(h)
(b)
P(h)∈Πh
(d)
≥
inf
P(h)∈Πh
0
Φ γB
h, P(h) −
b=1 k=1
B1 h, P(h)
P(h)∈Πh
0
= min Φ γB
h, P(h) − max B1 h, P(h)
P(h)∈Πh
P(h)∈Πh
(f )
0
= Φ γB
h, P1∗ (h) − B1 h, P2∗ (h)
(13)
(e)
Here, EH [PWF (H)] = P̄ . Since the message M ∈ M is
independent of H, the output of the auxiliary channel is
independent of M and hence
1
β1− (PX,H,Y , QX,H,Y ) ≤ 1 − 1 − ∗ .
M
Next, β1− can
bounded (as in [3], n
equation (102))
be lower
as, β1− ≥ γ1 P Iγ − , where γ > 0, Iγ , i(X, H, Y) ≤
o
P
log γ and i(X, H, Y) , log QX,H,Y
.
X,H,Y
sup
D
where, (a) follows from the fact that Wb0 = Wb00 and the
definition of Wb00 and Πh , where,
|H|
n
o
X
Πh , P(h) ∈ R|H| :
ti (h)Pi (h) = P̄ .
i=1
0
In (b), Wb = (Wb00 Phbp
(h) − µb ), γB
h, P(h) ≡
0
γB
(h) = (log γ − µ(h))/ ν(h), the mean of SB (W 0 )
(as noted, SB (.) is our alternate notation for summation),
µ(h) = SB (µB ), where,
µb ≡ µhb (h) = nc C(gb2 ) +
2
2
2
(hb nc Phb (h) − gb nc ) 2 σN + gb2 . Similarly, the variance of
SB (W 0 ), ν(h) = SB (ν), where,
2
2 2
2
νb ≡ νhb (h) = 2σN
hb nc Phb (h) + nc gb4 2 σN
+ gb2 .
(14)
Next, (c) follows from Berry Esseen Theorem for independent,
but not identically distributed random variables ([17], Chapter
16) with B(h) ≡ B h, P(h) , where,
B(h) = Cτ (h)
B
X
3/2
, τ (h) ,
ν(h)
EZ b |Wb |3 , (15)
b=1
for some constant
C > 0 (not depending on h). The term
B1 h, P(h) in (d) is defined in the proof of Lemma 1
provided in Appendix E. Also,
from the definition therein,
B1 h, P(h) ≥ B h, P(h) . Together with the fact that,
taking infimum of the first term and supremum of the second
term (on the right hand side of (c)) separately can only lower
the value, (d) follows. In (e), the infimum and supremum
are replaced with minimum and maximum respectively, by
noting the following. The functions B1 (from the definition in
Lemma 1 in Appendix E )) and Φ, are continuous in P(h).
Also, Πh ⊂ R|H| is a compact set. Hence the minimum and
the maximum respectively, are attained. Also, note that the
minimum in (e) is not 0 due to the equality constraint in
the definition of Πh . Finally, in (f ), P1∗ (.) and P2∗ (.) denote
the power allocations that attain the minimum and maximum,
respectively.
From (12) and (13), we have that
h
X
i
0
P Iγ ≥
P(h) Φ γB
h, P1∗ (h) − B1 h, P2∗ (h)
h∈HB
(a)
h
i C2
0
≥ EH Φ γB
H, P1∗ (H)
−√ ,
n
(16)
where, (a) follows from Lemma 1 stated and proved in
Appendix E. Here, C2 is a universal positive constant.
Next, we proceed
to obtain a tractable lower bound for
0
γB
H, P1∗ (H) . To that end, we make use of the following
two facts. First, since PWF (hb ) ≥ (λ −
2
σN
h2b
,
2
2
2
2
σN
+ gb2 ≥ σN
+ h2b λ − σN
hb = λh2b .
(17)
Also, using the fact that PWF (hb ) = 1{h2b ≥λ} PWF (hb ), it can
be seen that
L(gb2 ) =
PWF (hb )
.
λ
Using the above two relationships, it can be seen that
nc C(gb2 )
∗
h2b nc P1,h
(h) − gb2 nc
b
+
≤ µ0b (hb ),
2 + g2
2 σN
b
(18)
where, µ0b (hb ) ≡ µ0b = nc C(gb2 ) + nc P̄ − nc PWF (hb ) /2λ.
B
P
Define
µ0b (hb ) , µ0 (h). Let R , log γ/n, µ00 (h) =
b=1
µ0 (h)/n, ν 0 h, P1∗ (h) = ν h, P1∗ (h) /n. Then,
!
√
R − µ00 (h)
0
∗
Φ γB h, P1 (h) ≥ Φ
nq
, Ψ(h).
ν 0 h, P1∗ (h)
(19)
Here, the inequality follows from the fact that Φ(.) is a
monotonically increasing function.
Now, using Lemma 2 provided in Appendix F, we obtain,
!
√ R − C(P̄ )
log n
EH Ψ(H) ≥ Φ
− C3
np 0
,
n
VBF (P̄ )
0
where, VBF
(P̄ ) is definedqas in the statement of Theorem
0 (P̄ )
VBF
log n
2C2
−1
√
.
3. Choose R = C(P̄ ) +
Φ
+
C
+
3
n
n
n
With the prescribed choice of R, combining (11) and (16) and
applying Taylor’s theorem
to the function Φ−1 (.), we obtain,
p
∗
0
log M ≤ nC(P̄ ) + nVBF (P̄ )Φ−1 () + O(log n).
A PPENDIX D
Upper bound: LT constraint
Let M ∗ ≡ M ∗ (n, , P̄ ) denote the maximal codebook size
for the channel with average probability of error satisfying
the LT constraint. In particular, as in the LT case, assume that
the LT constraint is satisfied with equality. As mentioned, this
assumption can be made without loss of optimality (see, for
instance, [16], Lemma 65). Next, assume that CSIT is known
non-causally. Bound derived under this assumption will give
a valid upper bound for the causal case under consideration.
While ST constraints insists that ||x||2 = nP̄ for every
fading realization h and channel input vector x, the LT
constraint only requires ||x||2 = nP(h), for some function
P : HB → R+ such that EH [P(H)] = P̄ . Fix δn > 0, κ > 0
and c > P̄ + δ. For deriving the upper bound, we will first
restrict to the class of policies Πκ,c , where,
n
o
Πκ,c , P : HB → 0, κ + c : EH [P(H)] = P̄ .
Note that, for any P : HB → R+ such that EH [P(H)] = P̄ ,
P̂(h) = min{κ + c, P(h)} belongs to Πκ,c . As κ → ∞,
Πκ,c converges to the original class of policies on R+ . Any
P ∈ Πκ,c is a function of the realizations of an i.i.d. random
vector H, such that 0 ≤ P(H) ≤ κ+c (where, the inequalities
hold in a.s. sense). Hence, in particular, P satisfies the bounded
difference condition ([15], Chapter 3). Let En , {h : P(h) >
P̄ + δn }. From McDiarmid’s inequality ([15]),
2nδn2
P(En ) ≤ δn0 , δn0 ≡ δn0 (κ, c) , exp −
.
(κ + c)2
(20)
Note that, under the event Enc , the channel input vector x
satisfies the ST constraint with power P̄ + δn . Hence, in
deriving the upper bound for the LT case, we can lower bound
the probability term P Iγ in (11) by P Iγ ∩ Enc . It follows where, Ψk (H, ν1 ) is the k th derivative of Ψ(H, x) at x = ν1 .
from the analysis for the ST case that
Let S∞ (H) denote the summation on the right hand side
!
of the above equation, Note that |S∞ (H)| ≤ 2. Hence,
√ R − C(P̄ + δn ) c
log n E Ψ(H, ν 0 ) is upper bounded by
c
c
p
En −C3
P Iγ ∩En ≥ P(En )P Z ≤ n
, H
0 (P̄ + δ )
n
VBF
n
4|H|
where, Z ∼ N (0, 1). Using the fact that, for events A, B
EH Ψ(H, ν1 ) + 2 + EH S∞ (H) TB .
B
P(A ∩ B) ≥ P(A) − P(B c ), we obtain
!
Now, we proceed to obtain a tractable upper bound on (ν 0 −
√ R − C(P̄ + δn )
log n
− δn0 − C3
np 0
P Iγ ≥ Φ
.
ν1 )/ν1 . Note that EH [ν 0 ] 6= ν1 and hence we cannot invoke the
n
VBF (P̄ + δn )
same lines of arguments as in [13]. To that end, observe that,
∗
1
0
Taylor’s theorem, Vi∗ ≤ Vi + Vi0 (P1,i
− PWF,i ), where,
Now, we will choose n so that +δn < 2 . Noting the fact that from
∗
∗
P
≡
P
(i,
h)
denotes
the
power
allocated
to
ηi according to
1
1,i
VBF (P̄ ) is monotonically increasing in P̄ , adopting the same
2 ∗
∗
∗
),
V
,
V (ηi2 PWF (ηi ))
P
,
V
(η
P
,
P
≡
P
(η
),
V
i
WF,i
WF i
lines of arguments as in the ST case, we obtain,
1
i 1,i
i
2
0
2 2
2
3
0
r
log n and Vi , σN ηi (σN + ηi PWF (ηi )) . Thus, ν − ν1 does not
0 (P̄ )
log M ∗
VBF
|H|
P
.
≤ C(P̄ +δn )+
Φ−1 +δn0 (κ, c) +O
∗
exceed
(ti − qi )Vi + ti [P1,i
− PWF,i ]Vi0 . Since we would
n
n
n
−(1−α)
Applying Taylor’s theorem to C(P̄ + δn ), for δn = n 2 ,
for some small q
α > 0, as in Appendix A, we obtain C(P̄ +
δn ) ≤ C(P̄ ) + 4λ2 n1(1−α) . By virtue of choice of δn , δn0 =
∗
exp − 2nα /(κ + c)2 . Thus, lognM is upper bounded by
r
log n
0 (P̄ )
VBF
1
+
Φ−1 +δn0 (κ, c) +O
.
C(P̄ )+
(1−α)
n
n
2λn 2
Note that, δn0 (κ, c) is monotonically increasing in κ. Since the
bound holds good for all κ, we can take the infimum over
κ to obtain the tightest bound. Next, from Taylor’s theorem,
Φ−1 ( + δn0 ) ≤ Φ−1 () + δn0 1/φ(Φ−1 () . Finally,
perform √
(1−α) ) +
ing
Taylor’s
expansion
of
the
function
(1
2λ
n
q
0 (P̄ ) [nφ2 (Φ−1 ())] exp(−2nα /c2 ) around α = 0
VBF
yields the required result.
A PPENDIX E
Lemma 1. Let Ψ(h) ≡ Ψ h, ν 0 h, P1∗ (h) be as defined in
0
(P̄ ) as in the statement of Theorem 3. Then,
(19) and VBF
!
√ R − C(P̄ )
log n
,
EH Ψ(H) ≥ Φ
np
− C3
0
n
VBF (P̄ )
where, C3 is some positive constant. .
Proof. The proof follows along similar lines of arguments as
in [13] (Lemma 17 and Lemma 18). However,
in contrast to the
case therein, the function ν 0 H, P1∗ (H) is not a sum of i.i.d.
random variables. This is due to the dependence of the power
allocation policy P1∗ non causally on H. We will show how
to circumvent this problem and adapt the proof ton our case.
To that end, define the high probability set TB , h ∈ RB :
q
o
max ti (h) − qi ≤ logBB . From Hoeffding’s inequality
i∈[1:|H|]
([15]), it follows that P(TB c ) ≤ (2|H|)/B 2 .
For notational convenience, let ν 0 h, P1∗ (h) ≡ ν 0 . Then,
Taylor’s expansion of Ψ h, ν 0 , as a function of ν 0 h, P1∗ (h)
around ν1 yields
∞ 0
k
X
ν − ν1 Ψk (H, ν1 )
Ψ H, ν 0 = Ψ H, ν1 +
,
ν1
2k k!
k=1
i=1
like to upper bound (ν 0 − ν1 ) under the event TB , noting the
fact that Vi ≤ 1, we obtain that under TB ,
r
0
ν − ν1 ≤ |H|
|H|
log B X
+
ti [P1∗ (i, h) − PWF (ηi )]Vi0 .
B
i=1
p
∗
−PWF,i < δB }.
Now, for δB = (log B)/B, let E , {h : P1,i
Then, note that the summation in the above inequality can be
upper bounded in the following way:
|H|
X
∗
∗
ti [P1,i
− PWF,i ]Vi0 1E + ti [P1,i
− PWF,i ]Vi0 1E c
i=1
(a)
≤ δB +
|H|
X
(b)
∗
ti [P1,i
− PWF,i ]1E c ≤ δB + P̄ −
i=1
|H|
X
ti PWF,i ,
i=1
(21)
where, in obtaining (a), we have used the definition of E and
that Vi0 ≤ 1. In getting (b), we made use of the facts that,
∗
under E c , P1,i
− PWF,i > 0, Vi0 ≤ 1 and P1∗ satisfies the
P|H|
ST constraint with equality. Finally, note that, i=1 ti PWF,i
is bounded below by
|H|
X
(a)
ti PWF,i 1TB ≥
i=1
|H|
X
(qi − δB )PWF,i 1TB
(22)
i=1
(b)
≥ P̄ − λδB |H| −
λ|H|
,
B2
(23)
where, (a) follows from the fact, under the event TB , ti ≥
qi − δB . In (b), λ is as in (1) and follows from the fact that
2
PWF,i ≤ λ and P(TB c ) ≤ (2|H|)/B
p . Using (21) and (22),
0
we see that ν − ν1 ≤ 2(1 + λ|H|) (log B)/B.
As for obtaining a similar upper bound for ν1 − ν 0 , let Vi∗ 0
∗
be the function obtained by replacing PWF,i with P1,i
in the
0
function Vi defined earlier. Again, using Taylor’s theorem, it
∗
follows that Vi ≤ Vi∗ + Vi∗ 0 (PWF,i − P1,i
). Note that, Vi∗ 0 ≤
q
|H|
P
2
∗
ηmax
. Thus, ν1 − ν 0 ≤ |H| logBB +
qi [PWF,i − P1,i
]Vi∗ 0 .
i=1
|H|
P
Now,
i=1
∗
2
qi [PWF,i − P1,i
]Vi∗ 0 1E ≤ δB ηmax
, where we have
2
used the fact that Vi∗ 0 ≤ ηmax
and the definition of E. Now,
|H|
X
(a)
∗
2
qi [PWF,i − P1,i
]Vi∗ 0 1E c ≤ ηmax
i=1
|H|
X
∗
qi [PWF,i − P1,i
]
i=1
|H|
(b)
2
P̄ −
= ηmax
X
∗
qi P1,i
P̄ −
X
∗
qi Pi,1
1 TB
i=1
(c)
2
≤ ηmax
c
Proof. Since we are interested in the asymptotic as a function
of B, we will prove the case for nc = 1 (so as to avoid
unnecessary notation cluttering). The result for nc > 1 can be
proved in an identical way. First, we will consider the term
B
P
ν(h) =
νhb (h), where νhb (h) as defined in (14). Note that
b=1
i=1
|H|
2
=ηmax
on h. Let Bb1 (p) , EH B1 (H, p) and P2∗ be as defined in
1
(13). Then, for Bb1 ≡ Bb1 (P2∗ ), Bb1 = O √Bn
.
−
|H|
X
∗
qi Pi,1
1TB c
,
i=1
|H|
X
∗
∗
P̄ −
(ti − δB )Pi,1
1TB − (ti + δB )Pi,1
1 TB c
|H|
B
2 2 ∗
2 2 ∗
X
σN
ηi P2,hb (h)
σN
hb P2,hb (h) X
=
B
(h)
ν(h) ≥
i
2
∗ (η ) 2
2
2
2
σN + gb
σN + ηi2 P2,h
i
i=1
b=1
b
|H|
(a) X
≥
i=1
Bi (h) ∗
P2,hb (h)Cmin = B P̄ Cmin ,
B
i=1
(d)
|H|
X
(e)
δB P̄
≤
≤
q
min − δB
i=1
r
(f )
δB P̄
P̄
log B
≤
=
,
qmin (1 − α)
qmin (1 − α)
B
2
ηmax
∗
δB Pi,1
1TB
2
where, in (a) we have used the fact that Vi∗ 0 ≤ ηmax
and the
definition of E, (b) follows from the fact that EH [PWF (H)] =
P̄ , in (c) we have made use of the definition of TB , in (d),
we have used the fact that P1∗ satisfies the ST constraint with
∗
equality, (e) follows from the fact that, under TB , P1,i
≤
P̄ /(qi − δB ) ≤ P̄ /(qmin − δB ) and (fp
) holds for some fixed
α ∈ (0, 1) and B large enough so that (log B)/B < αqmin .
Let c1 = max 2(1 + λ|H|), P̄p
/qmin (1 − α) . Thus, we
have shown that |ν 0 − ν1 | ≤ c1 (log B)/B.
Also, for
2
G
2
as in Theorem 1,
note
that
ν
≥
E
G
/2(1
+
G)
=
1
G
2
EH PWF
(H)/2λ2 ≥ P̄ 2 /2λ2 , where, the first inequality follows from the definition of ν1 , the equality follows by invoking
(17) and the last bound follows from Jensen’s inequality.
Thus, we have obtained a lower bound p
on 1/ν1 . Hence, we
have obtained that |(ν 0 − ν1 )/ν1 | ≤ c2 (log B)/B, where
c2 = 2c1λ2 /P̄ 2 . It then follows
from
the analysis in [13]
that EH Ψ(H, ν 0 ) − EH Ψ(H, ν1 ) ≤ C1 logn n , for some
constant C1 > 0.
Next, note that µ00 (H) in the function Ψ(H) is a sum of
i.i.d. random variables taking values in H. Hence, along with
the above bound, directly invoking the result in [13] (Lemma
18), we obtain that
!
√ R − C(P̄ )
log n
EH Ψ(H) ≥ Φ
np 0
.
− C3
n
VBF (P̄ )
where, in (a), Cmin =
2 2
σN
ηi
min
1≤i≤|H|
2 +η 2 P
σN
i WF (ηi )
that Cmin > 0, as ηi > 0, for all i. Thus
1
2 . Also, note
≤
ν H
1
B P̄ Cmin
a.s..
Now, we consider the term τ (h) as defined in (15). Note
that |Wb | =
q
∗ (h)σ 2 Z − g 2 (Z 2 − 1)
2hb P2,h
N b
b
b
b
∗
≤
−
µ
Wb00 P2,h
(h)
b
b
2
2
2
2σN σN + gb
q
∗ (h)
hb P2,h
b
gb2
2
≤ Zb
+
(Z
−
1)
b
2 + g2
2 σ2 + g2
σN
2σN
N
b
b
∗
(h)
h2b P2,h
Zb
(Zb2 − 1)
b
+
+
2
2 + g2
2 + g2
2σN
σN
σN
b
b
∗
2
(a)
P2,hb (h)
Zb
(Zb − 1)
≤ Zb
+ 2 +
,
2
λ
σN
σN
≤ Zb
where, in (a) we have made use
first term.
of (17)
to get the
3
∗
Thus, it can be seen that EZ |Wb |3 ≤ K3 P2,h
(h) +
b
∗
2
∗
K2 P2,h
(h) + K1 P2,h
(h) + K0 , for some constants
b
b
Ki > 0, not depending on h. Next, we will show that
B
3
P
∗
(H) is O(B). To that end, since P2∗ (h) ∈ Πh ,
EH
P2,h
b
b=1
lim
B→∞
|H|
X
∗
ti (h)P2,i
(h) =
i=1
|H|
X
i=1
∗
qi lim P2,i
(h) = P̄
B→∞
This implies that
∗
lim P2,i
(h) < P̄ /qi , for all i.
B→∞
Hence,
A PPENDIX F
|H|
B
X
∗
3 X
3
Lemma 2. Given a realization of the fading vector h and lim 1
∗
qi lim P2,i
P2,hb (h) =
(h) = C1 . (24)
B→∞
a vector p , (p1 , . . . , pB ) (where pb corresponds to the B→∞ B b=1
i=1
power allocated to the state ηib which occurs in block b), let
−1
B
P
Here, C1 is a constant less than qmin
|H|P̄ 2 , where qmin is as
B1 (h, p) , (B P̄ Cmin )−1
K3 p3b +K2 p2b +K1 pb +K0 , for
|H|
S
b=1
defined
in
Section
II.
Next,
denote
S
,
h : |ti (h)−qi | <
δ
some constants C , K > 0, i = 0, . . . , 3, not depending
min
i
i=1
δ , for 0 < δ < qmin . Also, for any power allocation vector
|H|
P
P(H), let π H, P(H) ,
ti (H)P(H). Note that
i=1
π H, P2∗ (H) = π H, P2∗ (H)1{ti (H)>0} = P̄ .
Hence, P2∗ (H)1{ti (H)>0} ≤ P̄ , for all i ∈ {1, . . . , |H|}.
Since, ni (H) ≥ 1 under the event {ti (H) > 0}, we obtain
P2∗ (H)1{ti (H)>0} ≤ nP̄ .
"
(25)
#
3
Using (25), EH π H, P2∗ (H)1{ti (H)>0}
1Sδ is upper
"
#
bounded by EH π H, B 3 P̄ 3 1Sδ . But, π H, B 3 P̄ 3 =
B 3 P̄ 3 P Sδ . By using the union bound and Hoeffding’s
inequality ([15]), we have
P Sδ ≤
|H|
X
P |ti (H) − qi | > δ ≤ 2|H|e−2δB .
i=1
Thus, with C2 , 2|H|P̄ 3 ,
"
#
3
∗
EH π H, P2 (H)1{ti (H)>0}
1Sδ ≤ C2 B 3 e−2δB . (26)
Next,
P ∗ (H) ∈ Πh , note that,
from the fact that
3 2
P̄ 3
∗
π H, P2 (H)1{ti (H)>0}
1Sδc ≤ qmin
−δ . Then, using (24)
and bounded convergence theorem, it follows that
"
#
3
∗
EH π H, P2 (H)1{ti (H)>0}
1Sδc ≤ C1 .
(27)
From (26) and (27), it follows that EH
B
3
P
∗
(H) is
P2,h
b
b=1
O(B). Hence, EH
B
i
P
∗
P2,h
(H) , i = 1, 2 are O(B) as well.
b
b=1
Combining this with the bound on ν(h), we get the required
result.
R EFERENCES
[1] M. Agiwal, A. Roy, and N. Saxena, “Next Generation 5G Wireless
Networks: A Comprehensive Survey,” IEEE Communications Surveys
& Tutorials, vol. 18, no. 3, pp. 1617–1655, 2016.
[2] V. Strassen, “Asymptotische Abschätzungen in Shannon’s Informationstheorie,” in Trans. Third Prague Conf. Inf. Theory, 1962, pp. 689–723.
[3] Y. Polyanskiy, H. V. Poor, and S. Verdú, “Channel Coding Rate in the
Finite Blocklength Regime,” IEEE Transactions on Information Theory,
vol. 56, no. 5, pp. 2307–2359, 2010.
[4] M. Hayashi, “Information Spectrum Approach to Second-Order Coding
Rate in Channel Coding,” IEEE Transactions on Information Theory,
vol. 55, no. 11, pp. 4947–4966, 2009.
[5] A. J. Goldsmith and P. P. Varaiya, “Capacity of Fading Channels with
Channel Side Information,” IEEE Transactions on Information Theory,
vol. 43, no. 6, pp. 1986–1992, 1997.
[6] A. El Gamal and Y.-H. Kim, Network Information Theory. Cambridge
university press, 2011.
[7] G. Caire, G. Taricco, and E. Biglieri, “Optimum Power Control over
Fading Channels,” IEEE Transactions on Information Theory, vol. 45,
no. 5, pp. 1468–1489, 1999.
[8] R. Negi and J. M. Cioffi, “Delay-Constrained Capacity with Causal
Feedback,” IEEE Transactions on Information Theory, vol. 48, no. 9,
pp. 2478–2494, 2002.
[9] R. Mahapatra, Y. Nijsure, G. Kaddoum, N. U. Hassan, and C. Yuen,
“Energy Efficiency Tradeoff Mechanism Towards Wireless Green Communication: A Survey.” IEEE Communications Surveys and Tutorials,
vol. 18, no. 1, pp. 686–705, 2016.
[10] Y. Polyanskiy and S. Verdú, “Scalar Coherent Fading Channel: Dispersion Analysis,” in Information Theory Proceedings (ISIT), 2011 IEEE
International Symposium on. IEEE, 2011, pp. 2959–2963.
[11] W. Yang, G. Caire, G. Durisi, and Y. Polyanskiy, “Optimum Power Control at Finite Blocklength,” IEEE Transactions on Information Theory,
vol. 61, no. 9, pp. 4598–4615, 2015.
[12] G. Durisi, T. Koch, J. Östman, Y. Polyanskiy, and W. Yang, “ShortPacket Communications over Multiple-Antenna Rayleigh-Fading Channels,” IEEE Transactions on Communications, vol. 64, no. 2, pp. 618–
629, 2016.
[13] M. Tomamichel and V. Y. Tan, “Second-order Coding Rates for Channels
with State,” IEEE Transactions on Information Theory, vol. 60, no. 8,
pp. 4427–4448, 2014.
[14] A. Lancho, T. Koch, and G. Durisi, “A High-SNR Normal Approximation for Single-Antenna Rayleigh Block-Fading Channels,” in Information Theory (ISIT), 2017 IEEE International Symposium on. IEEE,
2017, pp. 1773–1777.
[15] S. Boucheron, G. Lugosi, and P. Massart, Concentration Inequalities: A
Nonasymptotic Theory of Independence. Oxford university press, 2013.
[16] Y. Polyanskiy, “Channel Coding: Non-Asymptotic Fundamental Limits,”
Ph.D. dissertation, Princeton Univ., Princeton, NJ, USA, 2010.
[17] W. Feller, Introduction to Probability Theory and Its Applications. John
Wyley, 1971, vol. 2.
| 7cs.IT
|
A Genetically Modified Hoare Logic
G. Bernot1 , J.-P. Comet1 , Z. Khalis1 , A. Richard1 , O. Roux2
1
University Nice-Sophia Antipolis
I3S laboratory, UMR CNRS 7271, Les Algorithmes, bât. Euclide B, BP.121,
06903 Sophia Antipolis Cedex, France
2
arXiv:1506.05887v1 [cs.CE] 19 Jun 2015
IRCCyN UMR CNRS 6597, BP 92101,
1 rue de la Noë, 44321 Nantes Cedex 3, France
Abstract: An important problem when modeling gene networks lies in the identification of parameters, even if we consider a purely discrete framework as the one of
René Thomas. Here we are interested in the exhaustive search of all parameter values
that are consistent with observed behaviors of the gene network. We present in this
article a new approach based on Hoare Logic and on a weakest precondition calculus
to generate constraints on possible parameter values. Observed behaviors play the
role of “programs” for the classical Hoare logic, and computed weakest preconditions represent the sets of all compatible parameterizations expressed as constraints
on parameters. Finally we give a proof of correctness of our Hoare logic for gene
networks as well as a proof of completeness based on the computation of the weakest
precondition.
1
Introduction
Gene regulation is a complex process where the expression level of a gene at each time depends
on a large amount of interactions with related genes. Hence regulations between genes can be
seen as a gene network. Different methods for studying the behavior of such gene networks in
a systematic way have been proposed. Among them, ordinary differential equations played an
important role which however mostly lead to numerical simulations. Moreover, the nonlinear
nature of gene regulations makes analytic solutions hard to obtain. Besides, the abstraction
procedure of René Thomas [Tho91], approximating sigmoid functions by step functions, makes
it possible to describe the qualitative dynamics of gene networks as paths in a finite state space.
Nevertheless this qualitative description of the dynamics is governed by a set of parameters
which remain difficult to be deduced from classical experimental knowledge. Therefore, even
when modeling with the discrete approach of René Thomas, the main difficulty lies in the
identification of these parameters. In this context, we are interested in the exhaustive search of
parameter values which are consistent with specifications given by the observed behavior of gene
regulatory networks. Because of the exponential number of parameterizations to consider, two
main kinds of approaches have emerged. On the one hand, information about cooperation or
concurrence between two regulators of a same target can be taken into account in order to reduce
the number of parameterizations to consider, see for example [KCRB09] and also [CTF+ 09]
in which this notion of cooperation is treated via a grouping of states. On the other hand,
using constraints can be helpful to represent the set of consistent parameterizations see for
example [FCT+ 04, CTF+ 09, MGCLG07].
In this paper, we present a new approach based on Hoare Logic [Hoa69] and on weakest
precondition calculus [Dij75] to generate constraints on parameters. A feature of this approach
lies in the fact that specifications are partially described by a set of paths, seen as “programs.”
Since this method avoids building the complete state graph, it results in a powerful tool to find
out the constraints representing the set of consistent parameterizations with a tangible gain for
1
CPU time. Indeed, the weakest precondition computation which builds the constraints, goes
through the “program” but is independent of the size of the gene network.
Other works were undertaken with such objectives. The application of temporal logic to biological regulatory networks was presented in [BCRG04]. Constraint programming was used for
biological systems in [BPGT01] and these ideas were continued specifically for genetic regulatory
networks in [Cor08, CTF+ 09].
The paper is organized as follows. The basic concepts of Hoare logic and Dijkstra weakest precondition are quickly reminded in Section 2. The formal definitions for discrete gene
regulatory networks are given in Section 3. Section 4 gives the way to describe properties of
states, then presents the path language, and finally introduces the notion of Hoare triplet. The
semantics of these extended Hoare triples is given in Section 5. With the previous material,
in Section 6 an extended Hoare logic for gene networks is defined for Thomas’ discrete models. In Section ??, the example of the “incoherent feedforward loop of type 1” (made popular
by Uri Alon in [SOMMA02, MSOI+ 02]) highlights the whole process of our approach to find
out the suitable parameter values. Section 8 contains a proof of correctness of our Hoare logic
for gene networks as well as a proof of completeness based on the computation of the weakest
precondition. We conclude in Section 9.
2
Reminders on standard Hoare logic
Hoare logic is a formal system for reasoning about the correctness of imperative programs.
In [Hoa69], C. A. R. Hoare introduced the notation “{P } pgm {Q}” to mean “If the assertion
P (precondition) is satisfied before performing the program pgm and if the program terminates,
then the assertion Q (postcondition) will be satisfied afterwards.” This constitutes de facto a
specification of the program under the form of a triple, called the Hoare triple.
In [Dij75], E. W. Dijkstra has defined an algorithm taking the postcondition Q and the
program pgm as input and computing the weakest precondition P0 that ensures Q if pgm terminates. In other words, the Hoare triple {P0 } pgm {Q} is satisfied and, for any precondition
P , {P } pgm {Q} is satisfied if and only if P ⇒ P0 .
Hoare logic and weakest preconditions are now widely known and teached all over the world.
The basic idea is to stamp the sequential phases of a program with assertions that are infered
according to the instruction they surround. There are several equivalent versions of Hoare logic
and our prefered one is the following because it offers a simple proof strategy to compute the
weakest precondition via a proof tree. Here, p, p1 and p2 stand for programs, P , P1 , P2 , I
and Q stand for assertions, v stands for a declared variable of the imperative program, and
Q[v ← expr] means that expr is substituted to each free occurrence of v in Q.
Assignment rule:
{Q[v←expr]} v:=expr {Q}
Sequential composition rule:
{P2 } p2 {Q}
{P1 } p1 {P2 }
{P1 } p1 ;p2 {Q}
Alternative rule:
{P1 } p1 {Q}
{P2 } p2 {Q}
{(e∧P1 )∨(¬e∧P2 )} if e then p1 else p2 {Q}
Iteration rule:
{e∧I} p {I}
{I} while e with I do p {¬e∧I}
Empty program rule:
P ⇒Q
{P } ε {Q}
(where ε stands for the empty program)
There are standard additional rules: first order logic (to establish “P ⇒ Q” introduced by
the Empty program rule), and in practice some reasonnings about the data structures of the
2
program (e.g. integers) in order to simplify the expressions as much as possible “on the fly” in
the proof trees.
The Iteration rule requires some comments. The assertion I is called the loop invariant and
it is well known that finding the weakest loop invariant is undecidable [Hat74, BG01]. It has
been included within the programming language for this reason; we ask the programmer to give
a loop invariant explicitely after the with keyword, although it may appear redundant as it is
also the precondition of the Hoare triple. By doing so, within a program, each while instruction
carries its own (sub)specification and it can consequently be proved apart from the rest of the
program.
When using these Hoare logic rules, the following proof strategy builds a proof tree that
performs the proof by computing the weakest precondition [Dij75]:
1. For each while statement within a Hoare triple
{P } p1 ; while e with I do p2 ; p3 {Q}
perform 3 independent sub-proofs:
• {¬e ∧ I} p3 {Q}
• {I} while e with I do p2 {¬e ∧ I} (i.e. {e ∧ I} p2 {I} according to the Iteration
rule)
• {P } p1 {I}
This first step of strategy leads to proofs on subprograms that do not contain any while
instruction.
2. Apply the Sequential composition rule only when the program p2 of this rule is reduced
to an instruction, which leads to perform the proof starting from postcondition Q at the
end and treat the instructions backward.
3. Never apply the Empty program rule, except when the leftmost instruction has been
treated, that is when all instructions have been treated.
Since the Assignment rule, which is central, makes it possible to precisely define one precondition from its postcondition, and since the other rules relate (but do not evaluate) the
conditions, the proof tree is done from the end to the beginning of the program and computes
a unique assertion. By doing so, the precondition obtained just before applying the last Empty
program rule is actually the weakest precondition (assuming that the programmer has given
the weakest loop invariants). In the remainder of the article, we call this strategy the backward
strategy. In Section ??, we always follow the backward strategy.
The most striking feature of Hoare logic and weakest precondition is that proofs according
to the backward strategy consist in very simple sequences of syntactic formula substitutions and
end with first order logic proofs. Nevertheless, it is worth noticing that it is only a question of
partial correctness since Hoare Logic does not give any proof of the termination of the analyzed
program (while instructions may induce infinite loops).
3
Discrete gene regulatory networks with multiplexes
This section presents our modeling framework, based on the general discrete method of René
Thomas [Td90, TK01] and introduced in [KCRB09, Kha10].
3
The starting point consists in a labeled directed graph in which vertices are either variables
or multiplexes. Variables abstract genes or their products, and multiplexes contain propositional
formulas that encode situations in which a group of variables (inputs of multiplexes) influence
the evolution of some variables (outputs of multiplexes). Hence multiplexes represent biological
phenomena, such as the formation of complexes to activate some genes. In the next definition,
this labeled directed graph is formally defined, and it is associated with a family K of integers.
As we will see later, these integers correspond to parameters that drive the dynamics of the
network.
Definition 3.1 A gene regulatory network with multiplexes ( grn for short) is a tuple N =
(V, M, EV , EM , K) satisfying the following conditions:
• V and M are disjoint sets, whose elements are called variables and multiplexes respectively.
• Ñ = (V ∪ M, EV ∪ EM ) is a labeled directed graph such that:
– edges of EV start from a variable and end to a multiplex, and edges of EM start from
a multiplex and end to either a variable or a multiplex.
– every directed cycle of Ñ contains at least one variable.
– every variable v of V is labeled by a positive integer bv called the bound of v.
– every multiplex m of M is labeled by a formula ϕm belonging to the language Lm
inductively defined by:
− If v → m belongs to EV and s ∈ IN belongs to the interval [1, bv ], then v ≥ s is
an atom of Lm .
− If m0 → m belongs to EM then m0 is an atom of Lm .
− If ϕ and ψ belong to Lm then ¬ϕ, (ϕ ∧ ψ) and (ϕ ∨ ψ) also belong to Lm .
• K = {Kv,ω } is a family of integers indexed by v ∈ V and ω ⊂ N −1 (v), where N −1 (v) is
the set of predecessors of v in Ñ (that is, the set of multiplexes m such that m → v is an
edge of EM ). Each Kv,ω must satisfy 0 ≤ Kv,ω ≤ bv .
Notation 3.2 The flaten version of a formula ϕm , denoted ϕm , is obtained by applying the following algorithm: while the formula contains a multiplex atom m0 , substitute m0 by its associated
formula ϕm0 . The formula ϕm exists since Ñ has no directed cycle with only multiplexes. As a
result, all the atoms of ϕm are of the form v ≥ s.
A state is an assignment of integer values to the variables of V . Such an assignment allows
a natural evaluation of any formula ϕm : by replacing variables by their values, ϕm becomes a
propositional formula whose atoms are integer inequalities.
Definition 3.3 (states, satisfaction relation and resources). Let N be a grn and V be its set
of variables. A state of N is a function η : V → IN such that η(v) ≤ bv for all v ∈ V . Let
L be the set of propositional formula whose atoms are of the form v ≥ s with v ∈ V and s be
a positive integer (so that ϕm is a formula of L for every multiplex m of N ). The satisfaction
relation |=N between a state η of N and a formula ϕ of L is inductively defined by:
• if ϕ is reduced to an atom of the form v ≥ s, then η |=N ϕ if and only if η(v) ≥ s.
• if ϕ ≡ ψ1 ∧ ψ2 then η |=N ϕ if and only if η |=N ψ1 and η |=N ψ2 ; and we proceed similarly
for the other connectives.
4
Given a variable v ∈ V , a multiplex m ∈ N −1 (v) is a resource of v at state η if η |=N ϕm . The
set of resources of v at state η is defined by ρ(η, v) = {m ∈ N −1 (v) | η |=N ϕm }.
From a dynamical point of view, at a given state η, each variable v is supposed to evolve
in the direction of a specific level (between 0 and bv ) that only depends on the set ρ(η, v).
This focal level is given by the logical parameter Kv,ρ(η,v) . Hence, at state η, v can increase if
η(v) < Kv,ρ(η,v) , it can decrease if η(v) > Kv,ρ(η,v) , and it is stable if η(v) = Kv,ρ(η,v) .
Suppose for instance that v has two input multiplexes mab and mcd with formula (a ≥ 1∧b ≥
1) and (c ≥ 1 ∧ d ≥ 1) respectively. Then mab and mcd may be seen as complexes (dimers)
regulating the level of v. Suppose, in addition, that Kv,∅ = 0, Kv,{mab } = Kv,{mcd } = 1 and
Kv,{mab ,mcd } = 2. Then, complexes mab and mcd specify activator complexes, with an individual
effect which is less than a cumulated effect (the focal level of v in the presence of a single complex
is less than the focal level of v in the presence of both complexes). This example illustrates the
fact that multiplexes encode combinations of variables that regulate a given variable, and that
the parameters, by giving a weight to each possible combinations of multiplexes, indicate how
multiplexes regulate a given variable.
As in Thomas’ method [Td90, TK01], it is assumed that variables evolve asynchronously and
by unit steps toward their respective target levels. The dynamics of a gene regulatory network
is then described by the following asynchronous state graph.
Definition 3.4 (State Graph). Let N = (V, M, EV , EM , K) be a grn. The state graph of N
is the directed graph S defined as follows: the set of vertices is the set of states of N , and there
exists an edge (or transition) η → η 0 if one of the following conditions is satisfied:
• there is no v ∈ V such that η(v) 6= Kv,ρ(η,v) and η 0 = η.
• there exists v ∈ V such that η(v) 6= Kv,ρ(η,v) and
(
0
η (v) =
η(v) + 1 if η(v) < Kv,ρ(η,v)
η(v) − 1 if η(v) > Kv,ρ(η,v)
and
∀u 6= v, η 0 (u) = η(u).
Hence a state η is a stable state if and only if it has itself as successor, that is, if and only if
every variable is stable at state η (i.e. η(v) = Kv,ρ(η,v) for every variable v). If η is not a stable
state, then it has at least one outgoing transition. More precisely, for each variable v such that
η(v) 6= Kv,ρ(η,v) , there is a transition allowing v to evolve (±1) toward its focal level Kv,ρ(η,v) .
Every outgoing transition of η is supposed to be possible, so that there is an indeterminism as
soon as η has several outgoing transitions. An example is given in Figure 2 (see also Section 7
for another example).
4
Pre- and post-conditions on path sets
In order to formalize known information about a gene network, we introduce in this section a
language to express properties of states (assertion language) and a language to express properties
of state transitions (path language). Combining properties of state transitions and properties
of states, at the beginning and at the end of a sequences of state transitions, leads to the notion
of Hoare triplet on path programs.
4.1
An assertion language for discrete models of gene networks
To describe properties of states in a meaningful way, we need terms that allow us to check, compare and manipulate variable values while taking parameter values into account. The following
definitions define a language suitable for such needs. It extends [Kha10].
5
Definition 4.1 (Terms of the assertion language) Let N = (V, M, EV , EM , K) be a grn. The
well formed terms of the assertion language of N are inductively defined by:
• Each integer n ∈ IN constitutes a well formed constant term
• For each variable v ∈ V , the name of the variable v, considered as a symbol, constitutes a
well formed constant term.
• Similarly, for each v ∈ V and for each subset ω of N −1 (v), the symbol Kv,ω constitutes a
well formed constant term.
• If t and t0 are well formed terms then (t + t0 ) and (t − t0 ) are also well formed terms.
Definition 4.2 (Assertion language and its semantics) Let N = (V, M, EV , EM , K) be a grn.
The assertion language of N is inductively defined as follows:
• If t and t0 are well formed terms then (t = t0 ), (t < t0 ), (t > t0 ), (t ≤ t0 ) and (t ≥ t0 ) are
atoms of the assertion language.
• If ϕ and ψ belong to the assertion language then ¬ϕ, (ϕ ∧ ψ), (ϕ ∨ ψ) and (ϕ ⇒ ψ) also
belong to the assertion language.
A state η of the network N satisfies an assertion ϕ if and only if its interpretation is valid
in ZZ, after substituting each variable v by η(v) and each symbol Kv,ω by its value according to
the family K. We note η |=N ϕ.
4.2
A path language for discrete models of gene networks
The assertion language introduced above is a subset of first order logic well suited to describe
properties on sets of states. It does not express dynamical aspects, since the dynamics of the
system is encoded in the transitions of the state graph. A description of dynamical properties
equates to a precise formulation of properties of paths. The language proposed here is suitable
for encoding such properties.
Definition 4.3 (Path language and path program) Let N = (V, M, EV , EM , K) be a grn. The
path language of N is the language inductively defined by:
• For each v ∈ V and n ∈ IN the expressions “v+”, “v−” and “v := n” belong to the path
language (respectively increase, decrease or assignment of variable value).
• If e is a formula belonging to the assertion language of N , then “assert(e)” also belongs
to the path language.
• If p1 and p2 belong to the path language then (p1 ; p2 ) also belongs to the path language
(sequential composition). Moreover the sequential composition is associative, so that we
can write (p1 ; p2 ; · · · ; pn ) without intermediate parentheses.
• If p1 and p2 belong to the path language and if e is a formula belonging to the assertion
language of N , then (if e then p1 else p2 ) also belongs to the path language.
• If p belongs to the path language and if e and I are formulas belonging to the assertion
language of N , then (while e with I do p) also belongs to the path language. The assertion
I is called the invariant of the while loop.
• If p1 and p2 belong to the path language then ∀(p1 , p2 ) and ∃(p1 , p2 ) also belong to the path
language (quantifiers). Moreover the quantifiers are associative and commutative, so that
we can write ∀(p1 , p2 , · · · , pn ) and ∃(p1 , p2 , · · · , pn ) as useful abbreviations.
6
For technical purposes, we also consider the empty program “ε” (outside the inductive definition). A well formed expression in the path language is called a path program.
Intuitivelly, “v+” (resp. “v−”, “v := n”) means that the level of variable v is increasing
by one unit (resp. decreasing by one unit, set to a particular value n). “assert(e)” allows one
to express a property of the current state without change of state. The sequential composition
allows one to concatenate two path programs whereas the statement “if ” allows one to choose
between two programs according to the evaluation of the formula e. Finally it becomes possible
to express properties of several paths thanks to the quantifiers ∀ and ∃. Lastly notice that ε
appears in a path program if and only if the path is reduced to the empty program. These
intuitions will be formalized in Section 5.
4.3
Syntax of pre- and post-conditions on path programs
The next step is to combine properties of state transitions (path program) and properties of
states (assertions), at the begining and at the end of the considered path program. This is done
via the notion of Hoare triplet on path programs.
Notation 4.4 A grn N being given, a Hoare triple on path programs is an expression of the
form “{P } p {Q}” where P and Q are well formed assertions, called pre- and post-condition
respectively, and p is a path program.
Intuitively, the precondition P describes a set of states, e.g. all states for which variable v has
value zero (P ≡ v = 0), the path program p describes dynamical processes, e.g. increase of
variable v (p ≡ v+), and the postcondition again describes a set of states, e.g. all states for
which variable v has value one (Q ≡ v = 1). This small example encodes the process of variable
v changing its value from zero to one. Whether or not the expression is satisfied for a given
gene network N depends on its state transition graph, thus it depends on the corresponding
parameter values in K.
5
Semantics of Hoare triples on path programs
We firstly define the semantics of path programs via a binary relation. The general ideas that
motivate the definition below are the following:
• Starting from an initial state η, sequences of instructions without existential or universal
quantifier either transform η into another state η 0 or is not feasible so that η 0 is undefined.
For example, the simple instruction v+ transforms η into η 0 (where ∀u 6= v, η 0 (u) = η(u)
and η 0 (v) = η(v) + 1) if η → η 0 exists. If, on the contrary, this transition does not exist,
the instruction is not feasible.
• Existential quantifiers induce a sort of “non determinism” about η 0 : according to the chosen
path under each existential quantifier one may get differents resulting states. Consequently,
one cannot define the semantics as a partial function that associates a unique η 0 to η; a
binary relation “η ; ...” is a more suited mathematical object.
• Universal quantifiers induce a sort of “solidarity” between all the states η 0 that can be
obtained according to the chosen path under each universal quantifier: all of them will
have to satisfy the postcondition later on. For this reason, we define a binary relation
that associates a set of states E to the initial state η: “η ; E”. Such a set E can be
understood as grouping together the states it contains, under the scope of some universal
quantifier.
7
• When the path program p contains both existential and universal quantifiers, we may
p
consequently get several sets E1 , · · · , En such that “η ; Ei ”, each of the Ei being a
possibility through the existential quantifiers of p and all the states belonging to a given
Ei being together through the universal quantifiers of p. On the contrary, if p is not
p
feasible, then there is no set E at all such that “η ; E”
Notation 5.1 For a state η, a variable v and k ∈ [0, bv ], we define η[v ← k] as the state η 0 such
that η 0 (v) = k and for all u 6= v, η 0 (u) = η(u).
p
Definition 5.2 [Path program relation ;].
Let N = (V, M, EV , EM , K) be a grn, let S be the state graph of N whose set of vertices is
p
denoted S and let p be a path program of N . The binary relation ; is the smallest subset of
S × P(S) such that, for any state η:
1. If p is reduced to the instruction v+ (resp. v−), then let us consider η 0 = η[v ← (η(v) + 1)]
p
(resp. η 0 = η[v ← (η(v) − 1)]): if η → η 0 is a transition of S then η ; {η 0 }
p
2. If p is reduced to the instruction v := k, then η ; {η[v ← k]}
p
3. If p is reduced to the instruction assert(e), if η |=N e, then η ; {η}
p1
p2
p1
p
p
4. If p is of the form ∀(p1 , p2 ): if η ; E1 and η ; E2 then η ; (E1 ∪ E2 )
p2
p
5. If p is of the form ∃(p1 , p2 ): if η ; E1 then η ; E1 , and if η ; E2 then η ; E2
p1
6. If p is of the form (p1 ; p2 ): if η ; F and if {Ee }e∈F is a F -indexed family of state sets
p2
p S
such that e ; Ee , then η ; ( e∈F Ee )
7. If p is of the form (if e then p1 else p2 ):
p1
p
p2
p
• if η |=N e and η ; E then η ; E
• if η 6|=N e and η ; E then η ; E
8. If p is of the form (while e with I do p0 ):
p0 ;p
p
• if η |=N e and η ; E then η ; E
p
• if η 6|=N e then η ; {η}
p
9. If p is the empty program ε, then η ; {η}
This definition calls for several comments.
p
The relation ; exists because (i) the set of all relations that satisfy the properties 1–8
of the definition is not empty (the relation which links all states to all sets of states satisfies
the properties) and (ii) the intersection of all the relations that satisfy the properties 1–8, also
satisfies the properties.
A simple instruction such as v+ can be not feasible from a state η (if η → η 0 is not a
v+
transition of S). In this case, there is no set E such that η ; E. The same situation happens
when the program is an assertion that is evaluated to false at the current state η.
Universal quantifiers “propagate” non feasible paths: if one of the pi is not feasible then
p
∀(p1 , · · · , pn ) is not feasible. It is not the case for existential quantifiers: if η ;i Ei for one of
the pi then η
∃(p1 ···pn )
;
Ei even if one of the pj is not feasible.
while...
When a while loop does not terminate, there does not exist a set E such that η ; E.
p
This is due to the minimality of the binary relation ;. On the contrary, when the while loop
8
terminates, it is equivalent to a program containing a finite number of the sub-program p0 in
sequence, starting from η.
The semantics of sequential composition may seem unclear for whom is not familiar with
p1 ;p2
commutations of quantifiers. We better take an example to explain the construction of ; (see
Figure 1):
p1
p1 ; p2
p2
E1
E1 ∪ E3
ηa
F1 =
E2
ηb
E3
F 2 = ηc
E4
gives:
E1 ∪ E4
η
E2 ∪ E3
η
E2 ∪ E4
Figure 1: An example for the semantics of sequential composition
p1
• Let us assume that starting from state η, two sets of states are reachable via p1 : η ; F1 =
p1
{ηa , ηb } and η ; F2 = {ηc }. It intuitively means that p1 permits a choice between F1 and
F2 through some existential quantifier on paths and that the path leading to F1 contains
a universal quantifier grouping together ηa and ηb .
• Let us also assume that:
p2
p2
– starting from state ηa , two sets of states are reachable via p2 : ηa ; E1 and ηa ; E2 ,
p2
p2
– starting from state ηb , two sets of states are reachable via p2 : ηb ; E3 and ηb ; E4 ,
p2
– there is not any set E such that ηc ; E
When focusing on the paths of (p1 ; p2 ) that encounter F1 after p1 , the paths such that p1 leads
to ηa must be grouped together with the ones that leads to ηb . Nevertheless, for each of them,
p2 permits a choice: between E1 or E2 for ηa and between E3 or E4 for ηb . Consequently,
when grouping together the possible futures of ηa and ηb , one needs to consider the four possible
p1 ;p2
p1 ;p2
p1 ;p2
p1 ;p2
combinations: η ; (E1 ∪ E3 ), η ; (E1 ∪ E4 ) η ; (E2 ∪ E3 ) and η ; (E2 ∪ E4 ).
Lastly, when focusing on the paths of (p1 ; p2 ) that encounter F2 after p1 , since ηc has no future
via p2 , there is no family indexed by F2 as mentioned in the definition and consequently it adds
p1 ;p2
no relation into ; .
p
Lastly, let us remark that, if η ; E then E cannot be empty; it always contains at least one
state. The proof is easy by structural induction of the program p (using the fact that a while loop
which terminates is equivalent to a program containing a finite number of the sub-program p0 ).
Definition 5.3 (Semantics of a Hoare triple). Let N = (V, M, EV , EM , K) be a grn and let S
be the state graph of N whose set of vertices is denoted S. A Hoare triple {P } p {Q} is satisfied
if and only if:
p
for all η ∈ S satisfying P , there exists E such that η ; E and for all η 0 ∈ E, η 0 satisfies Q.
The previous definition implies the consistency of all the paths described by the path program
p with the state graph: if path program p is not feasible from one of the states satisfying precondition P , then the Hoare triplet cannot be satisfied. For instance if some v+ is required by
9
b
µ2
a≥1
1
µ1
a
b
a≥2
(2)
(1)
0
µ3
¬(b ≥ 1)
0
1
a
2
Figure 2: (Left) Graphical representation of the grn N = (V, M, EV , EM , K) with V = {a, b},
the bounds of a and b are respectivelly 2 and 1, M = {µ1 , µ2 , µ3 }, φµ1 is (a ≥ 2), φµ2 is
(a ≥ 1), φµ3 is ¬(b ≥ 1). Finally the family of integers is {Ka,∅ = 1, Ka,{µ1 } = 1, Ka,{µ3 } = 2,
Ka,{µ1 ,µ3 } = 2, Kb,∅ = 1, Kb,{µ2 } = 1}. (Right) Representation of its state graph.
the path program p but the increasing of v is not possible according to the state graph, then
the Hoare triple is not satisfied.
More generally the path language plays the role of the programming language in the classical
Hoare Logic. A Hoare triplet is satisfied iff from the states satisfying the precondition, the
program p is feasible and leads to a set of states where the postcondition is satisfied. The path
program p can then be viewed as the sequence of actions one can use in order to modify the
state (memory) of the system.
Nevertheless, similarly to classical Hoare logic which reflects a partial correctness of imperative programming language, the previous definition does not imply termination of while
loops. Our path language can also define some infinite paths. Notice that if the non terminating
while loop is at the end of the program, then it has a biological meaning: it represents periodic
behaviours (such as the circadian cycle for instance).
Examples.
Let us consider the grn of Figure 2 and its state graph.
1. The Hoare triplet {(a = 0) ∧ (b = 0)} a+; a+; b + {(a = 2) ∧ (b = 1)} is satisfied, because :
• There is a unique state satisfying the precondition (a = 0) ∧ (b = 0) and
• from this state, the path program a+; a+; b+ is possible and leads to the state (2, 1)
and
• the state (2, 1) satisfies the postcondition (a = 2) ∧ (b = 1).
2. On the opposite, the Hoare triplet {(a = 2) ∧ (b = 0)} b+; a−; a − {(a = 0) ∧ (b = 1)} is
not satisfied because from the state satisfying the precondition, the first “instruction” b+
is possible and leads to the state (2, 1) from which the next instruction b− is not consistant
with the state graph.
3. The following Hoare triplet contains two existantial quantifiers and a universal one :
{(a = 0) ∧ (b = 0)} ∀(a+, b+); ∃(a+, b+); ∃(ε, b+) {(b = 1)}
• We have clearly (0, 0)
• Since (1, 0)
∃(a+,b+)
;
have both (0, 0)
∀(a+,b+)
;
{(1, 0), (0, 1)}
{(2, 0)} and (1, 0)
∀(a+,b+);∃(a+,b+)
;
• We have trivially (1, 1)
∃(ε,b+)
;
• Moreover we have both (2, 0)
∃(a+,b+)
;
{(1, 1)} and (0, 1)
{(1, 1), (2, 0)} and (0, 0)
∃(a+,b+)
;
∀(a+,b+);∃(a+,b+)
;
{(1, 1)}
∃(ε,b+)
;
{(2, 0)} and (2, 0)
10
∃(ε,b+)
;
{(2, 1)}
{(1, 1)}, we
{(1, 1)}.
• We deduce that the considered program p can lead to 3 differents set of states :
p
p
p
(0, 0) ; {(1, 1), (2, 0)}, (0, 0) ; {(1, 1)} and (0, 0) ; {(1, 1), (2, 1)}.
Because the postcondition is satisfied in both states (1, 1) and (2, 1), see the last set of
states which is in relation with (0, 0), one can deduce that the Hoare Triplet is satisfied.
6
A Hoare logic for discrete models of gene networks
In this section, we define a “genetically modified” Hoare logic by giving the rules for each
instruction of our path language (definition 4.3). First, let us introduce a few notations for
intensively used formulas.
Notation 6.1 Let N = (V, M, EV , EM , K) be a grn and let v be a variable of V .
1. For each subset ω of N −1 (v) (set of predecessors of v in the network), we denote by Φωv
the following formula:
Φωv
≡
(
^
^
ϕm ) ∧ (
m ∈ ω
m ∈
¬ϕm )
N −1 (v)\ω
where N −1 (v) \ ω stands for the complementary subset of ω in N −1 (v).
From Definition 3.3, for all states η and for all variables v ∈ V , η |=N Φωv if and only if
ω = ρ(η, v), that is, ω is the set of resources of v at state η. Consequently, there exists a
unique ω such that η |=N Φωv .
2. We denote by Φ+
v the following formula:
Φ+
v
≡
^
(Φωv =⇒ Kv,ω > v)
ω⊂G−1 (v)
From Definition 3.4, η |=N Φ+
v if and only if there is a transition (η → η[v ← v + 1]) in
the state graph S, that is, if and only if the variable v can increase.
3. We denote by Φ−
v the following formula:
Φ−
v
≡
^
(Φωv =⇒ Kv,ω < v)
ω⊂G−1 (v)
Similarly, η |=N Φ−
v if and only if the variable v can decrease from the state η in the state
graph S.
By the way, in practice, the assertion assert(Φ=
v ) is often
^useful from the biological point of
=
=
view, where Φv is obviously defined by:
Φv
≡
(Φωv =⇒ Kv,ω = v)
ω⊂G−1 (v)
Our Hoare logic for discrete models of gene networks is then defined by the following rules,
where v is a variable of the grn and k ∈ IN .
1. Rules encoding Thomas’ discrete dynamics.
Incrementation rule:
{ Φ+
v ∧ Q[v←v+1] } v+ {Q}
Decrementation rule:
{ Φ−
v ∧ Q[v←v−1] } v− {Q}
11
2. Rules coming from Hoare Logic. These rules are very similar to the ones given in Section 2.
Obvious rules for the instruction assert(Φ) and the quantifiers are added:
Assert rule:
{ Φ ∧ Q } assert(Φ) { Q }
Universal quantifier rule:
{P1 } p1 {Q}
{P2 } p2 {Q}
{P1 ∧P2 } ∀(p1 ,p2 ) {Q}
Existential quantifier rule:
{P1 } p1 {Q}
{P2 } p2 {Q}
{P1 ∨P2 } ∃(p1 ,p2 ) {Q}
Assignment rule:
{Q[v←k]} v:=k {Q}
Sequential composition rule:
{P2 } p2 {Q}
{P1 } p1 {P2 }
{P1 } p1 ;p2 {Q}
Alternative rule:
{P1 } p1 {Q}
{P2 } p2 {Q}
{(e∧P1 )∨(¬e∧P2 )} if e then p1 else p2 {Q}
Iteration rule:
{e∧I} p {I}
{I} while e with I do p {¬e∧I}
Empty program rule:
P ⇒ Q
{P } ε {Q}
3. Axioms. These axioms assert that all values stay between their bounds, where v is a
variable of the grn N and ω ⊂ N −1 (v):
Boundary axioms:
0≤v
v ≤ bv
0 ≤ Kv,ω
Kv,ω ≤ bv
Remark 6.2
+
• (Φ+
v ⇒ v < bv ) can be derived from the previous rules. Indeed, Φv implies that for ω
corresponding to the current set of resources, Kv,ω > v and, using the boundary axiom
Kv,ω ≤ bv , we get v < bv .
• Similarly, we have Φ−
v ⇒ v > 0.
These implications will be used in section ??.
We will prove in Section 8 that this modified Hoare logic is correct, and that it is complete
provided that the path program under consideration contains the weakest loop invariants for all
while statements. More precisely, the proof strategy called backward strategy, already described
at the end of Section 2, also applies here: It computes the weakest precondition. Before giving
these two proofs, let us show in the next section the usefulness of our genetically modified Hoare
logic via the formal study of the possible biological functions of a very simple network.
7
Example
In [SOMMA02, MSOI+ 02] Uri Alon and co-workers have studied the most common in vivo
patterns involving three genes. Among them, they have enlightened the “incoherent feedforward
loop of type 1”. It is composed by a transcription factor a that activates a second transcription
12
factor c, and both a and c regulate a gene b: a is an activator of b whereas c is an inhibitor of
b. There is a “short” positive action of a on b and a “long” negative action via c: a activates
c which inhibits b. The left hand side of Figure 3 shows such a feedforward loop. Considering
that both thresholds of actions of a are equal leads to a boolean network since, in that case, the
variable a can take only the value 0 (a has no action) or 1 (a activates both b and c). The right
c
c
1
+
1
1
l
a≥1
−
1
a
1 +
b
a
1
λ
¬(c ≥ 1)
σ
a≥1
b
Figure 3: Boolean incoherent feedforward of type 1. At right, graphical representation of the
grn N = (V, M, EV , EM , K) with V = {a, b, c}, the bounds of a, b and c are equal to 1,
M = {l, λ, σ}, φl is (a ≥ 1), φλ is (¬(c ≥ 1)), φσ is (a ≥ 1). Finally the family of integers is
{Ka,∅ , Kc,∅ , Kc,{l} , Kb,∅ , Kb,{σ} , Kb,{λ} , Kb,{σ,λ} }.
hand side of the figure shows the corresponding grn with multiplexes: σ encodes the “short”
action of a on b, whilst l followed by λ constitute the “long” action.
Several authors, like Uri Alon, consider that if a is equal to 0 for a sufficiently long time, both
b and c will also be equal to 0, because b and c need a as a resource in order to reach the state 1.
They also consider that the function of this feedforward loop is to ensure a transitory activity
of b that signals when a has switched from 0 to 1: the idea is that a activates the productions
of b and c, and then c stops the production of b.
Here, we take a look at this question via four different path programs, and we prove formally
that this affirmation is only valid under some constraints on the parameters of the network, and
only under the assumption that b starts its activity before c.
Is a transitory production of b possible? As already stated, the function classically associated with the feedforward loop is to ensure a transitory activity of b that signals when a has
switched from 0 to 1. An interesting question is under which conditions the previous property is
true ? For example the path program
(1)
P1 ≡ (b+; c+; b−)
together with the pre-condition P ≡ (a = 1 ∧ b = 0 ∧ c = 0) and the post-condition
Q0 ≡ {b = 0}, is a possible formalization of the previous property about the behaviour of the
feedforward loop. The backward strategy using our genetically modified Hoare logic on this
example gives the following successive conditions.
• The weakest precondition obtained through the last instruction “b−” is the following
conjunction:
Φ−
b ∧ Q0 [b ← b − 1] ≡
(¬¬(c ≥ 1) ∧ ¬(a ≥ 1)) =⇒ Kb < b
(¬¬(c ≥ 1) ∧ (a ≥ 1)) =⇒ Kb,σ < b
(¬(c ≥ 1) ∧ ¬(a ≥ 1)) =⇒ K
<b
b,λ
(¬(c
≥
1)
∧
(a
≥
1))
=⇒
K
b,σλ < b
b−1=0
13
which simplifies as the conjunction Q1 :
Q1 ≡
b=1
((c ≥ 1) ∧ (a < 1)) =⇒ Kb = 0
((c ≥ 1) ∧ (a ≥ 1)) =⇒ K
=0
b,σ
((c
<
1)
∧
(a
<
1))
=⇒
K
b,λ = 0
((c < 1) ∧ (a ≥ 1)) =⇒ Kb,σλ = 0
• Then, the weakest precondition obtained through the instruction “c+” is:
Φ+
c ∧ Q1 [c ← c + 1] ≡
¬(a ≥ 1) =⇒ Kc > c
a ≥ 1 =⇒ Kc,l > c
b=1
((c + 1 ≥ 1) ∧ (a < 1)) =⇒ K = 0
b
((c
+
1
≥
1)
∧
(a
≥
1))
=⇒
K
b,σ = 0
((c + 1 < 1) ∧ (a < 1)) =⇒ Kb,λ = 0
((c + 1 < 1) ∧ (a ≥ 1)) =⇒ K
b,σλ = 0
which simplifies as Q2 owing to the boundary axioms and remarks 6.2:
Q2 ≡
c=0
a < 1 =⇒ Kc = 1
a ≥ 1 =⇒ K = 1
c,l
b
=
1
a < 1 =⇒ Kb = 0
a ≥ 1 =⇒ Kb,σ = 0
• Lastly, the weakest precondition obtained through the first “b+” of the program is:
Φ+
b ∧ Q2 [b ← b + 1] ≡
which simplifies as Q3 :
Q3 ≡
(¬¬(c ≥ 1) ∧ ¬(a ≥ 1)) =⇒ Kb > b
(¬¬(c ≥ 1) ∧ (a ≥ 1)) =⇒ Kb,σ > b
(¬(c ≥ 1) ∧ ¬(a ≥ 1)) =⇒ Kb,λ > b
(¬(c ≥ 1) ∧ (a ≥ 1)) =⇒ Kb,σλ > b
c=0
a < 1 =⇒ Kc = 1
a ≥ 1 =⇒ Kc,l = 1
b+1=1
a < 1 =⇒ Kb = 0
a ≥ 1 =⇒ Kb,σ = 0
a < 1 =⇒ Kb,λ = 1
a ≥ 1 =⇒ Kb,σλ = 1
c=0
a < 1 =⇒ Kc = 1
a ≥ 1 =⇒ Kc,l = 1
b=0
a < 1 =⇒ Kb = 0
a ≥ 1 =⇒ Kb,σ = 0
Then, using the empty program rule, it comes P =⇒ Q3 i.e. (a = 1 ∧ b = 0 ∧ c = 0) =⇒ Q3
and after simplification we get the correctness if and only if Kb,σλ = 1 and Kc,l = 1 and
Kb,σ = 0. This proves that, whatever the values of the other parameters, the system can exhibit
a transitory production of b in response to a switch of a from 0 to 1.
14
Is a transitory production of b possible without increasing c? The previous program
P1 is not the only one reflecting a transitory production of b, there may be other realisations of
this property. For example one can consider the path program :
P2 ≡ (b+; b−).
(2)
With respect to this path program, the weakest precondition obtained through the last instruction “b−” is of course Q1 as previously. Then, the weakest precondition obtained through “b+”
is:
b=0
((c ≥ 1) ∧ (a < 1)) =⇒ ((Kb = 1) ∧ (Kb = 0))
((c ≥ 1) ∧ (a ≥ 1)) =⇒ ((Kb,σ = 1) ∧ (Kb,σ = 0))
Q4 ≡
((c < 1) ∧ (a < 1)) =⇒ ((Kb,λ = 1) ∧ (Kb,λ = 0))
((c < 1) ∧ (a ≥ 1)) =⇒ ((Kb,σλ = 1) ∧ (Kb,σλ = 0))
Q4 is of course not satisfiable: it implies that each parameter associated with b is both equal to
0 and 1. The path program (b+; b−) is not feasible (inconsistent weakest precondition). Indeed,
we retrieve an obvious property of the Thomas’ approach: if b has no negative action on itself
then the sequence (b+; b−) cannot arise because the resources of b must change in order to
switch its direction of evolution.
Another possible path compatible with the path program (b+, c+, b−). Let us notice
that, when Kb,σλ = 1, Kc,l = 1 and Kb,σ = 0, even if the system can exhibit a transitory
production of b via (b+; c+; b−), this does not prevent from some other paths that do not
exhibit this behaviour. For example the simple path P3 ≡ c+ leaves b constantly equal to 0,
and the Hoare triplet
(
a=1 ∧ b=0 ∧ c=0∧
Kb,σλ = 1 ∧ Kc,l = 1 ∧ Kb,σ = 0
)
c+
n
b=0
o
is satisfied, as the corresponding weakest precondition Q5 is clearly implied by the precondition.
Q5 ≡ Φ+
c ∧ Q0 [c ← c + 1] ≡
c=0
a = 0 =⇒ K = 1
c
a
=
1
=⇒
K
c,l = 1
b=0
When a is constantly equal to 1 and when c = 1, production of b is impossible.
Even worst: when a is constantly equal to 1, once c has reached the level 1, it is impossible for
b to increase to 1. We prove this property by showing that the following triplet is inconsistent,
whatever the loop invariant I:
(
(3)
a=1 ∧ b=0 ∧ c=1∧
Kb,σλ = 1 ∧ Kc,l = 1 ∧ Kb,σ = 0
)
while b < 1 with I do ∃(b+, b−, c+, c−)
|
{z
P4
n
b=1
o
}
The subprogram ∃(b+, b−, c+, c−) reflects the fact that a stays constant but b or c evolves. The
while statement allows b and c to evolve freely until b becomes equal to 1.
Applying the Iteration rule, I has to satisfy:
• ¬(b < 1) ∧ I =⇒ (b = 1)
This property is trivially satisfied whatever the assertion I, due to the boundary axioms.
15
• {b < 1 ∧ I} ∃(b+, b−, c+, c−) {I}
We apply the existential quantifier rule, which gives the following weakest precondition:
−
+
−
Q6 ≡ (Φ+
b ∧I[b ← b + 1]) ∨ (Φb ∧I[b ← b − 1]) ∨ (Φc ∧I[c ← c + 1]) ∨ (Φc ∧I[c ← c − 1])
Consequently I can be any assertion such that
(b = 0 ∧ I) =⇒ Q6
Let us denote P the precondition of the path program P4 . Applying the empty program rule, it
comes that I must also satisfy P =⇒ I. So, because P =⇒ (b = 0), we have P =⇒ (b = 0 ∧ I),
which, in turn implies Q6 . Moreover, let us remark that
−
+
−
Q6 =⇒ (Φ+
b ∨ Φb ∨ Φc ∨ Φc )
Consequently, if the Hoare triple 3 is correct, then:
−
+
−
P =⇒ (Φ+
b ∨ Φb ∨ Φc ∨ Φc )
which is impossible because, if P is satisfied, then:
+
• Φ+
b is false, as a = 1, c = 1 and Kb,σ = 0 (indeed,Φb implies a = 1 ∧ c = 1 ⇒ Kb,σ > 0)
−
• Φ−
b is false, as b = 0 (Φb implies b > 0)
+
• Φ+
c is false, as c = 1 (Φc implies c < 1)
−
• Φ−
c is false, as a = 1, c = 1 and Kc,l = 1 (Φc implies a = 1 ∧ c = 1 ⇒ Kc,l < 1).
So, we have formally proved that when a is constantly equal to 1, once c has reached the level
1, it is impossible for b to increase to 1.
8
Partial Correctness and Completeness
”Partial” has to be understood here as ”Assuming that the while loops terminate”, as usual in
Hoare logic.
8.1
Correctness
Correctness of our modified Hoare logic means that: if ` {P } p {Q} according to the inference
rules of Section 6, then the Hoare triple {P } p {Q} is semantically satisfied according to
Definition 5.3, i.e.: for all η ∈ S (S being the set of states of S) such that η |=N P there exists
p
E ⊂ S such that η ; E and ∀η 0 ∈ E, η 0 |=N Q.
The proof is made as usual by induction on the proof tree of ` {P } p {Q}. Hence, we
have to prove that each rule of Section 6 is correct. Here we develop only the Incrementation
rule and the Sequential composition rule since the correctness of the other inference rules is
either similar (Decrementation rule) or trivial (Assert rule, quantifier rules, Assignment rule
and Empty program rule) or standard in Hoare Logic (Alternative rule and Iteration rule). Let
us note that the correctness of the Sequential composition rule is neither trivial nor standard
because its semantics is enriched to cope with the quantifiers.
Let N be a grn and let η be any state of the associated state space S:
Incrementation rule:
{ Φ+
v ∧ Q[v←v+1] } v+ {Q}
From Definition 5.3, the hypothesis is:
16
(where v is a variable of the grn)
H η |=N Φ+
v
and η |=N Q[v ← v + 1]
and we have to prove the conclusion:
v+
C there exists E ⊂ S such that η ; E and ∀η 0 ∈ E, η 0 |=N Q
Let us choose E = {η 0 } with η 0 = η[v ← η(v) + 1]. From Notation 6.1, the hypotesis
0
η |=N φ+
v is equivalent to (η → η ) ∈ S, which in turn, according to Definition 5.2, implies
v+
η ; {η 0 }. Hence, it only remains to prove that η 0 |=N Q, which results from the hypothesis
η |=N Q[v ← v + 1].
2
Sequential composition rule:
{P2 } p2 {Q}
{P1 } p1 {P2 }
{P1 } p1 ;p2 {Q}
From Definition 5.3, we consider the following three hypotheses:
p1
H1 for all η1 ∈ S such that η1 |=N P1 there exists E1 such that η1 ; E1 and ∀η 0 ∈
E1 , η 0 |=N P2
p2
H2 for all η2 ∈ S such that η2 |=N P2 there exists E2 such that η2 ; E2 and ∀η 00 ∈
E2 , η 00 |=N Q
H3 η |=N P1
and we have to prove the conclusion:
p1 ;p2
C there exists E ⊂ S such that η ; E and ∀η 00 ∈ E, η 00 |=N Q
p1
Let us arbitrarily choose a set E1 such that η ; E1 and ∀η 0 ∈ E1 , η 0 |=N P2 (we know
that E1 exists from H1 and H3 ).
0
For each η 0 ∈ E1 , we similarly choose a set E2η such that:
p2
0
0
0
η 0 ; E2η and ∀η 00 ∈ E2η , η 00 |=N Q (we know that the family {E2η }η0 ∈E1 exists from H2
and the fact that η 0 |=N P2 for all η 0 ∈ E2 )
0
p1 ;p2
Let E = ( η0 ∈E1 E2η ), we have: η ; E from Definition 5.2 and ∀η 00 ∈ E, η 00 |=N Q
(from the way the union is built).
2
S
8.2
Weakest precondition
Completeness of Hoare logic would be of course defined as follows: If the Hoare triple {P } p {Q}
is satisfied (according to Definition 5.3) then ` {P } p {Q} (using the inference rules of Section 6
as well as first order logic and proofs on integers).
Obviously, as such, Hoare logics cannot be complete because, as already mentioned for
classical Hoare logic, finding the weakest loop invariants is undecidable and there is no complete
logic on integers (Gödel). So, following Dijkstra [Dij75], we prove completeness under the
assumptions that the loop invariants of all while statements are weakest invariants and that the
needed properties on integers are admitted. We adopt the strategy that computes the weakest
precondition and we will prove the following theorem:
Theorem 8.1 (Dijkstra theorem on the genetically modified Hoare logic) A grn N and a Hoare
triple {P } p {Q} being given, the backward strategy defined at the end of Section 2, with the
inference rules of Section 6, computes the weakest precondition P0 just before the last inference
that uses the Empty program rule.
It means that: if {P } p {Q} is satisfied, then P ⇒ P0 is satisfied.
17
This theorem has an obvious corrolary:
Corollary 8.2 A grn N being given, our modified Hoare logic is complete under the assumption
that all given loop invariants are the weakest ones and that the needed properties on integers are
established.
Proof of the corollary: if {P } p {Q} is satisfied, then, from the Dijkstra theorem above,
there is a proof tree that infers the Hoare triple if there is a proof tree for the property P ⇒ P0
(which is semantically satisfied because P0 is the weakest precondition). First order logic being
complete and properties on integers being axiomatically assumed, the proof tree for P ⇒ P0
exists.
2
Proof of Dijkstra theorem:
Under the hypotheses that all loop invariants are minimal and that the Hoare triple {P } p {Q}
is satisfied, i.e., under the hypotheses:
p
H1 for all η satisfying P , there exists E such that η ; E and for all η 0 ∈ E, η 0 satisfies Q
H2 for all while statements of p, the corresponding loop invariant I is the weakest one
one has to prove the conclusion:
C P ⇒ P0 is satisfied, where P0 is the precondition computed by the proof of {P } p {Q}
according to the backward strategy with the inference rules of Section 6.
The proof is done by structural induction according to the backward strategy on p.
• If p is of the form while e with I do p0 , then, by construction of the backward strategy,
applying the Iteration rule, we get P0 = I, and the conclusion results immediately from
H2 .
v+
• If p is of the form v+, then the only set E such that η ; E is E = {η[v ← v + 1]}. The
hypothesis H1 becomes:
H1 for all η satisfying P , η 0 = η[v ← v + 1] satisfies Q and η → η 0 is a transition of S
and from the Incrementation rule, the conclusion becomes:
C P ⇒ (Φ+
v ∧ Q[v ← v + 1]) is satisfied.
So, H1 ⇒ C straightforwardly results from the definition of Φv+ (Notation 6.1) and we
do not use H2 .
• If p is of the form p1 ; p2 , then we firstly inherit the two structural induction hypotheses:
H3 for all assertions P 0 and Q0 , if {P 0 } p1 {Q0 } is satisfied then P 0 ⇒ P1 is satisfied,
where P1 is the precondition computed from Q0 via the backward strategy
H4 for all assertions P 00 and Q00 , if {P 00 } p2 {Q00 } is satisfied then P 00 ⇒ P2 is satisfied,
where P2 is the precondition computed from Q00 via the backward strategy
Moreover the hypothesis H1 becomes (Definition 5.2):
18
p1
H1 for all η satisfying P , there exists a family of state sets F = {Ee }e∈F such that η ; F
p2
and e ; Ee for all e ∈ F and for all η 0 ∈ E = (
S
e∈F
Ee ), η 0 satisfies Q
Lastly, from the Sequential composition rule, the conclusion becomes:
C P ⇒ P1 is satisfied, where P1 is the weakest precondition of {· · ·} p1 {P2 }, P2 being
the weakest precondition of {· · ·} p2 {Q}.
From H4 (with Q00 = Q) it results that all the states e ∈ F of hypothesis H1 satisfy
P2 . Consequently {P } p1 {P2 } is satisfied. Thus, from H3 (with Q0 = P2 and P 0 = P ) it
comes P ⇒ P1 , which proves the conclusion.
• Similarly to the correctness proof, we do not develop here the other cases of the structural
induction. They are either similar to already developed cases (Decrementation rule) or
trivial (Assert rule, quantifier rules, and Assignment rule) or standard in Hoare Logic
(Alternative rule).
2
This ends the proof.
9
Discussion
The cornerstone of the modeling process lies, whatever the application domain, in the determination of parameters. In this paper, we proposed an approach for exhibiting constraints on
parameters of gene network models, that relies on the adaptation of the Hoare logic, initially
designed for proofs of imperative programs. It leads to several questions about its usability and
implementations.
9.1
Language issues
The path language is a way to describe formally the specification of correct models of gene
networks. Classically, the specifications can be expressed in temporal logics, like CTL and LTL,
which also allows the modeler to take into account behavioral information. But even if there
exists some links between path language and temporal logics, these formal languages (temporal
logics and path language) are not comparable: some properties can be expressed in the path
language and not in classical temporal logics and converselly.
• On the one hand, in the path program, the assignment instruction allows one to express a
knock-out of a gene (v := 0). Such knock-out of a gene is not expressible in CTL or LTL.
• On the other hand, CTL or LTL is able to express properties on infinite cyclic traces. Such
properties on infinite traces would be expressed in the path language by a program which
does not terminate, and consequently, the post-condition would not make sense.
Nevertheless, a succession of incrementation/decrementation instructions corresponds to a property that can be expressed in the CTL language. For example, if one knows the starting
point, say (v1 = 1 ∧ v2 = 0), the path program v1 +; v2 +; v1 − corresponds to the formula
EX(v1 = 2 ∧ EX(v2 = 1 ∧ EX(v1 = 1 ∧ v2 = 0))). Correctness of this program path
with the precondition (v1 = 1 ∧ v2 = 0) becomes equivalent to verify that the CTL formula
(v1 = 1 ∧ v2 = 0) ⇒ EX(v1 = 2 ∧ EX(v2 = 1 ∧ EX(v1 = 1 ∧ v2 = 0))) is true in all possible
states. More generally, the path language is well suited for sequential properties whereas CTL
can express non sequential ones.
19
# initializing
T3:=1; T4:=1; d3:=1; d2:=0; gi:=0; gp:=0; gt:=0; tr:=0;
# evolutions
gi+; d2+; T3+; tr+; T3+; gp+; d3-; gt+;
(T 3 = 1)
(T 4 = 1)
(d3 = 1)
(d2 = 0)
(gi = 0)
(gp = 0)
(gt = 0)
(tr = 0)
∧
∧
∧
∧
∧
∧
∧
⇒
EX((gi = 1)∧
EX((d2 = 1)∧
EX((T 3 = 2)∧
EX((tr = 1)∧
EX((T 3 = 3)∧
EX((gp = 1)∧
EX((d3 = 0)∧
EX(gt = 1))))))))
Figure 4: A path program (top) with its corresponding CTL formula (bottom)
In the path language, invariants of while loops are mandatory: Hoare logic is able to prove
a program with while statements only if invariants are given. In other words, the entire information that the Hoare logic needs to perform the proof, is given by invariants. Unfortunately,
invariants are difficult to devise. Thus the while statements are often used in proofs by refutation, where the proof is done for each possible invariant, see our example of section ??.
9.2
Plateform issues
The Hoare logic for gene networks has been designed in order to support a software which aims
at helping the determination of parameters of models of gene networks. We have already done
its proof of feasability. Indeed, after having developped a prototype named SMBioNet which
enumerates all possible valuations of parameters and retains only those which are coherent with
a specified temporal logic formula, we developped a new prototype called WP-SMBioNet [Kha10],
which uses a path program and the Weakest Precondition calculus (backward strategy) to produce constraints on parameters.
In order to compare both approaches (CTL formulae versus path programs), we consider
a property which can be expressed in both temporal logic CTL and path program. When
modeling the biological system triggering the tail resorption during the metamorphose of tadpole,
see [TTB+ 07] and references therein, the expression profiles of [LB77] can be translated into
a path program, see Fig 4, which in turn can be translated into an equivalent CTL formula.
For this example, whereas SMBioNet needs more than 3 hours to selects among all possible
parameterizations those which lead to a dynamics coherent with the CTL formula, WP-SMBioNet
needs only 10 seconds (on the same computer) to construct the constraints on the parameters.
If we ask the enumeration of the parameters satisfying the constraints (using Choco [cT10]), the
total search time is about 2 minutes. This example shows that the Hoare logic can speed up the
determination of coherent parameterizations.
We can notice that the complexity of the weakest precondition calculus is linear with the
number of instructions in the path program, and does not depend on the size of gene regulatory
networks: each node of the syntaxic tree of the program is visited only once. At the opposite,
the CTL model checking algorithm depends on the size of the network. Thus, the use of path
program instead of CTL formula leads to postpone the enumeration step which then can use
the constraints on parameters to cut down drastically the set of parameterizations to consider.
A software plateform dedicated to analysis of gene regulatory networks, should have to
20
combine different technics. Indeed constraints solving technics are necessary to enumerate parameters or give counter-examples, theorem prover can be also useful to get strategies for proofs
by refutation, and model checking technics and Hoare logic precondition calculus should be
combined in order to give very efficient algorithm. As already noted, it seems natural to use
Hoare logic when the behavioural specification focuses on a finite time horizon, whereas model
checking is natural when the temporal specification expresses global properties on infinite traces.
It would be interesting to complete this plateform with some improved features. From a
theoretical point of view, one could also develop approaches to help finding loop invariants. To
build them, it seems possible to adapt the iterative approach adopted in ASTREE [CCF+ 05]
but in another context (abstract interpretation [CC04]): Pragmatically one begins with a simple invariant I, then one tries to make the proof and completes iterativelly and partially the
invariant. From an application point of view, specifications often stem from DNA profiles, it
would be valuable to develop a program that automatically produces path programs from DNA
chips data. Two questions emerge: the choice of thresholds on which is based the discretization
of expression levels, and the determination of good time steps. These questions are out of the
scope of this article. They mainly rely on biological expertise and experimental conditions.
Acknowledgment
We are grateful to Alexander Bockmayr and Heike Siebert for fruitful discussions and comments
on this paper. The authors thank the French National Agency for Research (ANR-10-BLANC0218 BioTempo project) for its support. This work has also been supported by the CNRS PEPII
project ”CirClock ”.
References
[BCRG04]
G. Bernot, J.-P. Comet, A. Richard, and J. Guespin. Application of formal methods to biological regulatory networks: Extending Thomas’ asynchronous logical
approach with temporal logic. Journal of Theoretical Biology, 229(3):339–347,
2004.
[BG01]
A. Blass and Y. Gurevich. Inadequacy of computable loop invariants. ACM
Transactions on Computational Logic, 2(1), 2001.
[BPGT01]
C. Boileau, J. Prados, J. Geiselmann, and L. Trilling. Using constraint programming for learning from experiments transcriptional activation and the geometry
of DNA. In T. Schiex L. Duret, C. Gaspin, editor, Actes des Journes Ouvertes
Biologie Informatique Mathmatiques (JOBIM), Toulouse, 2001.
[CC04]
P. Cousot and R. Cousot. Building the Information Society, chapter Basic Concepts of Abstract Interpretation., pages 359–366. Kluwer Academic Publishers,
2004.
[CCF+ 05]
P. Cousot, R. Cousot, J. Feret, L. Mauborgne, A. Min, D. Monniaux, and X Rival.
The ASTRE analyser. In M. Sagiv, editor, ESOP 2005 The European Symposium
on Programming, number 3444 in LNCS, pages 21–30. Springer, 2005.
[Cor08]
Fabien Corblin. Conception et mise en oeuvre d’un outil déclaratif pour l’analyse
des réseaux génétiques discrets. PhD thesis, Université Joseph-Fourier - Grenoble
I, 12 2008.
21
[cT10]
choco Team. choco: an open source java constraint programming library. Research
report 10-02-INFO, Ecole des Mines de Nantes, 2010.
[CTF+ 09]
F. Corblin, S. Tripodi, E. Fanchon, D. Ropers, and L. Trilling. A declarative constraint-based method for analyzing discrete genetic regulatory networks.
Biosystems, 98(2):91–104, 2009.
[Dij75]
Edsger W. Dijkstra. Guarded commands, nondeterminacy and formal derivation
of programs. Commun. ACM, 18:453–457, August 1975.
[FCT+ 04]
E. Fanchon, F. Corblin, L. Trilling, B. Hermant, and D. Gulino. Modeling the
molecular network controlling adhesion between human endothelial cells: Inference
and simulation using constraint logic programming. In CMSB, pages 104–118,
2004.
[Hat74]
W.S. Hatcher. A semantic basis for program verification. J. of Cybernetics,
4(1):61–69, 1974.
[Hoa69]
C.A.R. Hoare. An axiomatic basis for computer programming. Communications
of the ACM, 12(10):576–585, oct 1969.
[KCRB09]
Z. Khalis, J.-P. Comet, A. Richard, and G. Bernot. The SMBioNet method for
discovering models of gene regulatory networks. Genes, Genomes and Genomics,
3(special issue 1):15–22, 2009.
[Kha10]
Z. Khalis. Logique de Hoare et identification formelle des paramtres d’un rseau
gntique. PhD thesis, University of Evry-Val d’Essonne, 2010.
[LB77]
J. Leloup and M. Buscaglia. La triiodothyronine, hormone de la mtamorphose des
amphibiens. CR Acad. Sci., 284:2261–2263, 1977.
[MGCLG07] D. Mateus, J.-P. Gallois, J.-P. Comet, and P. Le Gall. Symbolic modeling of genetic regulatory networks. Journal of Bioinformatics and Computational Biology,
5(2B):627–640, 2007.
[MSOI+ 02]
R. Milo, S. Shen-Orr, S. Itzkovitz, N. Kashtan, D. Chklovskii, and U. Alon. Network motifs: Simple building blocks of complex networks. Science, 298:824–827,
2002.
[SOMMA02] S. Shen-Orr, R. Milo, S. Mangan, and U. Alon. Network motifs in the transcriptional regulation network of escherichia coli. Nature Genetics, 31:64–68, 2002.
[Td90]
R. Thomas and R. d’Ari. Biological Feedback. CRC Press, 1990.
[Tho91]
R. Thomas. Regulatory networks seen as asynchronous automata : A logical
description. J. theor. Biol., 153:1–23, 1991.
[TK01]
R. Thomas and M. Kaufman. Multistationarity, the basis of cell differentiation and
memory. II. logical analysis of regulatory networks in terms of feedback circuits.
Chaos, 11:180–195, 2001.
[TTB+ 07]
S. Troncale, R. Thuret, C. Ben, N. Pollet, J.-P. Comet, and G. Bernot. Modelling
of the TH-dependent regulation of tadpole tail resorption. Journal of Biological
Physics and Chemistry, 7(2):45–50, 2007.
22
| 5cs.CE
|
Parsimonious Random Vector Functional Link Network for Data
Streams
Mahardhika Pratama, Plamen P. Angelov, Edwin Lughofer
Abstract – the theory of random vector functional link network (RVFLN) has provided a breakthrough in
the design of neural networks (NNs) since it conveys solid theoretical justification of randomized
learning. Existing works in RVFLN are hardly scalable for data stream analytics because they are inherent
to the issue of complexity as a result of the absence of structural learning scenarios. A novel class of
RVLFN, namely parsimonious random vector functional link network (pRVFLN), is proposed in this
paper. pRVFLN adopts a fully flexible and adaptive working principle where its network structure can be
configured from scratch and automatically generated in accordance with nonlinearity and time-varying
property of system being modelled. pRVFLN is equipped with complexity reduction scenarios where
inconsequential hidden nodes can be pruned and input features can be dynamically selected. pRVFLN
puts into perspective an online active learning mechanism which expedites the training process and
relieves operator’s labelling efforts. In addition, pRVFLN introduces a non-parametric type of hidden
node, developed using an interval-valued data cloud. The hidden node completely reflects the real data
distribution and is not constrained by a specific shape of the cluster. All learning procedures of pRVFLN
follow a strictly single-pass learning mode, which is applicable for online time-critical applications. The
advantage of pRVFLN was verified through numerous simulations with real-world data streams. It was
benchmarked with recently published algorithms where it demonstrated comparable and even higher
predictive accuracies while imposing the lowest complexities. Furthermore, the robustness of pRVFLN
was investigated and a new conclusion is made to the scope of random parameters where it plays vital
role to the success of randomized learning.
Keywords – Random Vector Functional Link, Evolving Intelligent System, Online Learning, Randomized
Neural Networks
I.
Introduction
For decades, research in artificial neural networks has mainly investigated the best way to determine
network free parameters, which reduces error as close as possible to zero [18]. Various approaches [55],
[56] were proposed, but a large volume of work is based on a first or second-order derivative approach in
respect to the loss function [19], [36]. Due to the rapid technological progress in data storage, capture,
and transmission [34], the machine learning community has encountered an information explosion, which
calls for scalable data analytics. Significant growth of the problem space has led to a scalability issue for
conventional machine learning approaches, which require iterating entire batches of data over multiple
epochs. This phenomenon results in a strong demand for a simple, fast machine learning algorithm to be
well-suited for deployment in numerous data-rich applications [58]. This provides a strong case for
research in the area of randomized neural networks (RNNs) [10], [13], [52], which was very popular in
late 80’s and early 90’s. RNNs offer an algorithmic framework, which allows them to generate most of
the network parameters randomly while still retaining reasonable performance [8], [47]. One of the most
prominent RNNs in the literature is the random vector functional link (RVFL) network which features
solid universal approximation theory under strict conditions [10].
Due to its simple but sound working principle, the RNN has regained its popularity in the current
literature [8], [11], [14], [15], [17], [41], [50], [51], [63], [66], [67]. Nonetheless, vast majority of RNNs
in the literature suffer from the issue of complexity which make their computational complexity and
memory burden prohibitive for data stream analytics since their complexities are manually determined
and rely heavily on expert domain knowledge [12]. They presents a fixed-size structure which lacks of
adaptive mechanism to encounter changing training patterns in the data streams [15]. Random selection
of network parameters often leads the network complexity to go beyond what necessary due to existence
of superfluous hidden nodes which contribute little to the generalization performance [45]. Although the
universal approximation capability of RNN is assured only when a sufficient complexity is selected,
choosing suitable complexity for given problems entails expert-domain knowledge and is problemdependent.
A novel RVFLN, namely parsimonious random vector functional link network (pRVFLN), is
proposed. pRVFLN combines simple and fast working principles of RFVLN where all network
parameters but the output weights are randomly generated with no tuning mechanism of hidden nodes. It
characterises the online and adaptive nature of evolving intelligent systems where network components
are automatically generated on the fly. pRVFLN is capable of following any variations of data streams
no matter how slow, rapid, gradual, sudden or temporal drifts in data streams are because it can initiate
its learning structure from scratch with no initial structure and its structure is self-evolved from data
streams in the one-pass learning mode by automatically adding, pruning and recall its hidden nodes [45].
Furthermore, it is compatible for online real-time deployment because data streams are handled without
revising previously seen samples. pRVFLN is equipped with the hidden node pruning mechanism which
guarantees low structural burdens and the rule recall mechanism which aims to address the cyclic concept
drift. pRVFLN incorporates a dynamic input selection scenario which makes possible activation and
deactivation of input attributes on the fly and an online active learning scenario which rules out
inconsequential samples from training process. Moreover, pRVFLN is a plug-and-play scenario where a
single training process encompasses all learning scenarios and a sample-wise manner without pre-and/or
post-processing steps. Novelties of pRVFLN are elaborated as follows:
•
Network Architecture: Unlike majority of existing RVFLNs, pRVFLN is structured as a locally
recurrent neural network with a self-feedback loop at the hidden node to capture temporal system dynamic
[24]. The recurrent network architecture puts into perspective a spatio-temporal activation degree which
takes into account compatibility of both previous and current data points without compromising the local
learning property because of its local recurrent connection. pRVFLN introduces a new type of hidden
node, namely an interval-valued data cloud inspired by the notion of recursive density estimation and
AnYa by Angelov [1], [2]. This hidden node is parameter-free and requires no parameterization per scalar
variable. It is not constrained by a specific shape and its firing strength is defined as a local density
calculated as accumulated distances between a local mean and all data points in the cloud seen thus far.
Our version in this paper distinguishes itself from its predecessors in [1], [2] in the fact that an interval
uncertainty is incorporated per local region which targets imprecision and uncertainties of sensory data
[50]. Interval-valued data cloud still satisfies the universal approximation condition of the RVFLN since
it is derived using the Cauchy kernel which is asymptotically a Gaussian-like function. Moreover, the
output layer consists of a collection of functional-link output nodes created by the second order
Chebyshev polynomial which increases the degree of freedom of the output weight to improve its
approximation power [38].
•
Online Active Learning Mechanism: pRVFLN possesses an online active learning mechanism
which is meant to extract data samples for training process. This mechanism is capable of finding
important data streams for training process, while discarding inconsequential samples [54]. This strategy
expedites the training mechanism and improves model’s generalization since it prevents redundant
samples to be learned. This scenario is underpinned by sequential entropy method (SEM) which forms a
modified version of neighbourhood probability method [64] for data streams. The SEM quantifies the
entropy of neighbourhood probability recursively and differs from that of [45] because it integrates the
concept of data cloud. The concept of data cloud further simplifies the working principle of SEM since
its activation degree already abstracts a density of a local region – a key attribute of neighbourhood
probability. In realm of classification problems, the SEM does not always call for ground truth which
leads to significant reduction of operator annotation’s efforts.
•
Hidden Node Growing Mechanism: pRVFLN is capable of automatically generating its hidden
nodes on the fly with the help of type-2 self-constructing clustering (T2SCC) mechanism which was
originally designed for an incremental feature clustering mechanism [26], [32]. This method relies on the
correlation measure in the input space and target space, which can be carried out in the single-pass
learning mode with ease. The original version, cannot be directly implemented here because the original
version is not yet designed for the interval-valued data cloud. This rule growing process also differs from
those of other RNNs with a dynamic structure [12], because the hidden nodes are not randomly generated,
rather they are created from the rule growing condition, which considers the locations of data samples in
the input space. We argue that randomly choosing centers or focal points of hidden nodes independently
from the original training data risks on performance deterioration [57], because it may not hold the
completeness principle. pRVFLN relies on the data cloud-based hidden node, which does not require any
parameterization, thus offering a simple and fast working framework as other RNNs.
•
Hidden Node Pruning and Recall Mechanism: a rule pruning scenario is integrated in pRVFLN.
The rule pruning scenario plays vital role to assure modest network structures since it is capable of
detecting superfluous neurons to be removed during the training process [34]. pRVFLN employs the socalled type-2 relative mutual information (T2RMI) method which extends the RMI method [21] to the
working principle of the interval-valued hidden node. The T2RMI method examines relevance of hidden
nodes to target concept and thus captures outdated hidden nodes which is no longer compatible to portray
the current target concept. In addition, the T2RMI method is applied to recall previously pruned rules.
This strategy is important to deal with the so-called recurring drift. The recurring drift refers to a situation
where old concept re-emerges again in the future. The absence of rule recall scenario risks on the
catastrophic forgetting of previously valid knowledge because the cyclic drift imposes an introduction of
a new rule without any memorization of learning history. It differs from that [34] because the rule recall
mechanism is separated from the rule growing scenario.
•
Online Feature Selection Mechanism: pRVFLN is capable of carrying out an online feature
selection process, borrowing several concepts of online feature selection (OFS) [59]. Note that although
feature selection is well-established for an offline situation, online feature selection remains a challenging
and unsolved problem because feature contribution must be measured with the absence of a complete
dataset. Notwithstanding some online feature reduction scenarios in the literature, the issue of stability is
still open, because an input feature is permanently forgotten once pruned. The OFS delivers a flexible
online feature selection scenario, because it makes possible to select or deselect input attributes in every
training observation by assigning crisp weights (0 or 1) to input features. Another prominent characteristic
of the OFS method lies in its capability to deal with partial input features, because the cost of extracting
all input features may be expensive in certain applications. The convergence of the OFS for both full and
partial input attributes have been also proven. Nevertheless, the original version [59] is originally devised
for linear regression and calls for some modification to be perfectly fit to pRVLFN. The generalized OFS
(GOFS) is proposed here and incorporates some refinements to adapt to the working principle of the
pRVFLN.
This paper puts forward the following contributions: 1) it presents a novel random vector functional
link network termed pRVFLN. pRVFLN aims to handle the issue of data streams which remains an
uncharted territory of current RVFLN works; 2) it puts into perspective the interval-valued data cloud
which is not shape-specific and non-parametric. It requires no parameterization per scalar variable and
makes use of local density information of a data cloud. This strategy aims to hinder complete
randomization of hidden node parameters while still maintaining no tuning principle of RVFLN.
Randomization of all hidden node parameters blindly without bearing in mind true data distribution lead
to common pitfall of RVFL where the hidden node matrix tend to be ill-defined and compromises the
completeness of network structure as confirmed in [57]. ; 3) four learning components, namely SEM,
T2SCC, T2RMI, GOFS, are proposed. The efficacy of pRVFLN was thoroughly evaluated using
numerous real-world data streams and benchmark with recently published algorithms in the literature
where pRVFLN demonstrated a highly scalable approach for data stream analytics with trivial cost of its
generalization performance. A supplemental document containing additional numerical studies, analysis
of predefined threshold and the effect of learning components is also provided in 1). Moreover, analysis
of robustness of random intervals was performed. It is concluded that random regions should be carefully
selected and should be chosen close enough to true operating regions of a system being modelled. The
MATLAB codes of pRVFLN is made publicly available in 4) to help further study.
The rest of this paper is structured as follows: Network architecture of pRVFLN is outlined in Section
2; Algorithmic development of pRVFLN is detailed in Section 3; Proof of concepts are outlined in Section
4; Conclusions are drawn in the last section of this paper.
II.
Network Architecture of pRVFLN
pRVFLN utilises a local recurrent connection at the hidden node which generates the spatiotemporal
property. In the literature, there exist at least three types of recurrent network structures referring to its
recurrent connections: global [22], [24], interactive [33], and local [23], but the local recurrent connection
is deemed to be the most compatible recurrent type to our case because it does not harm the local property,
which assures stability when adding, pruning and fine-tuning hidden nodes. pRVFLN utilises the notion
of the functional-link neural network which presents the so-called enhancement layer. Furthermore, the
hidden layer of pRVFLN is built upon an interval-valued data cloud, which does not require any
parametrization and granulation [1]. This hidden node does not have any pre-specified shape and fully
evolves its shape with respect to the real data distribution because it is derived from recursive local density
estimation. We bring the idea of data cloud further by integrating an interval-valued principle.
Suppose that a data tuple streams at t-the time instant, where X t n is an input vector and Tt m is
a target vector while n and m are respectively the numbers of input and output variables. Because
pRVFLN works in a strictly online learning environment, it has no access to previously seen samples,
and a data point is simply discarded after being learned. Due to the pre-requisite of an online learner, the
total number of data N is assumed to be unknown. The output of pRVFLN is defined as follows:
R
~
~
y o i Gi ,temporal ( At X t Bt ), Gtemporal [G , G ]
(1)
i 1
where R denotes the number of hidden nodes. i stands for the i-th output node, produced by weighting
the weight vector with an extended input vector i xe wi , where xe ( 2 n1)1 is an extended input vector
T
resulting from the functional link neural network based on the Chebyshev function and wi ( 2 n1)1 is a
connective weight of the i-th output node. At n is an input weight vector randomly generated here. Bt
~
is the bias fixed at 0 for simplicity. Gi ,temporal is the i-th interval-valued data cloud, triggered by the upper
and lower data cloud G i ,temporal , G i ,temporal . By expanding the interval-valued data cloud, the following is
obtained:
R
R
i 1
i 1
(2)
yo (1 qo ) i G i ,temporal qo i G i ,temporal
where q m is a design factor to reduce an interval-valued function to a crisp one [9]. It is worth noting
that the upper and lower activation functions G i ,temporal , G i ,temporal deliver spatiotemporal characteristics as
a result of a local recurrent connection at the i-th hidden node, which combines the spatial and temporal
firing strength of the i-th hidden node. These temporal activation functions output the following.
t 1
t
t
t 1
G i ,temporal i G i , spatial (1 i )G i ,temporal , G i ,temporal i G i ,spatial (1 i )G i ,temporal
t
t
where R is a weight vector of the recurrent link. The local feedback connection here feeds the
~ 1
spatiotemporal firing strength at the previous time step Git,temporal
back to itself and is consistent with the
local learning principle. This trait happens to be very useful in coping with the temporal system dynamic
because it functions as an internal memory component which memorizes a previously generated
spatiotemporal activation function at t-1. Also, the recurrent network is capable of overcoming overdependency on time-delayed input features and lessens strong temporal dependencies of subsequent
patterns [36]. This trait is desired in practise since it may lower input dimension, because prediction is
done based on recent measurement only. Conversely, the feedforward network assumes a problem as a
function of current and past input and outputs. This strategy at least entails expert knowledge for system
order to determine the number of delayed components.
The hidden node of the pRVFLN is an extension of the cloud-based hidden node, where it embeds an
interval-valued concept to address the problem of uncertainty [30]. Instead of computing an activation
degree of a hidden node to a sample, the cloud-based hidden node enumerates the activation degree of a
sample to all intervals in a local region on-the-fly. This results in local density information, which fully
reflects real data distributions. This concept was defined in AnYa [1], [3] and was patented in [2]. This
concept is also the underlying component of AutoClass and TEDA-Class [4]. All of which come from
Angelov’s sound work of RDE [3]. This paper aims to modify these prominent works to the intervalvalued case. Suppose that N i denotes the support of the i-th data cloud, an activation degree of i-th cloudbased hidden node refers to its local density estimated recursively using the Cauchy function:
~
Gi , spatial
1
~
xk xt 2
1 (
)
Ni
k 1
Ni
~
~
xk ,i [ x k ,i , x k ,i ], Gi ,spatial [G i ,spatial , G i ,spatial ]
(3)
xk is k-th interval in the i-th data cloud and xt is t-th data sample. It is observed that Eq. (3) requires
where ~
the presence of all data points seen so far, which is impossible when dealing with data streams. Its
recursive form is formalised in [3] and is generalized here to the interval-valued problem:
G i , spatial
1
1 At xt i , Ni
T
2
i , Ni i , Ni
2
, G i , spatial
1
1 A xt i , N
T
t
2
i
(4)
i , Ni i , N
2
i
where i , i signify the upper and lower local means of the i-th cloud:
i, N (
i
xi ,k i
xi , N i i
N 1
Ni 1
, i ,1 xi ,1 i
) i , N 1
, i ,1 xi ,1 i , i , Ni ( i ) i , Ni 1
i
Ni
Ni
Ni
N i ,k
(5)
where i is an uncertainty factor of the i-th cloud, which determines the degree of tolerance against
uncertainty. The uncertainty factor creates an interval of the data cloud, which controls the degree of
tolerance for uncertainty. It is worth noting that a data sample is considered as a population of the i-th
cloud when resulting in the highest density. Moreover, i , N i , i , Ni are the upper and lower mean square
lengths of the data vector in the i-th cloud as follows:
2
i , Ni
xi , Ni i
N 1
( i
)i , Ni 1
, i ,1 xi ,1
Ni
Ni
2
2
i , i , Ni
xi , Ni i
N 1
( i ,k
) i , Ni 1
, i ,1 xi ,1
Ni ,k
Ni
2
i
(6)
It can be observed from Eq. (3) that the cloud-based hidden node does not have any specific shape and
evolves naturally according to its supports. Furthermore, it is parameter-free, where no parameters –
centroid, width, etc. as encountered in the conventional hidden node need to be assigned. Although the
concept of the cloud-based hidden node was generalized in TeDaClass [27] by introducing the eccentricity
and typicality criteria, the interval-valued idea is uncharted in [27]. Note that the Cauchy function is
asymptotically a Gaussian-like function, satisfying the integreable requirement of the RVFLN to be a
universal approximator.
Unlike conventional RNNs, pRVFLN puts into perspective a nonlinear mapping of the input vector
through the Chebyshev polynomial up to the second order. This concept realises an enhancement layer
linking the input layer to the output layer - consistent with the original concept of the RVFLN. Note that
recently developed RVFLNs in the literature mostly neglect the direct connection because they are
designed with a zero-order output node [8], [11], [14], [15], [17], [41], [50], [51], [63], [66],[67] . The
direct connection expands the output node to a higher degree of freedom, which aims to improve the local
mapping aptitude of the output node. The direct connection produces the extended input vector xe making
use of the second order Chebyshev polynomial:
p 1 ( x) 2 x j p ( x j ) p 1 ( x j )
(7)
where 0 ( x j ) 1, 1 ( x j ) x j , 2 ( x j ) 2 x 2j 1 . Suppose that three input attributes are given X [ x1 , x2 , x3 ]
, the extended input vector is expressed as the Chebyshev polynomial up to the second order
xe [1, x1 , 2 ( x1 ), x2 , 2 ( x2 ), x3 , ( x3 )] . Note that the term 1 here represents an intercept of the output node.
This avoids an output node from going through origin, which may risk an untypical gradient. There exist
other functions as well in functional-link neural networks: trigonometric and polynomial, power. The
Chebyshev function, however, scatters fewer parameters to be stored into memory than the trigonometric
function, while the Chebyshev function has better mapping capability than other polynomial functions of
the same order. In addition, the polynomial power function is not robust against an extrapolation case.
pRVFLN implements the random learning concept of the RVFLN, in which all parameters, namely the
input weight A , design factor q , recurrent link weight , and uncertainty factor , are randomly
generated. Only the weight vector is left for parameter learning scenario wi . Since the hidden node is
parameter-free, no randomization takes place for hidden node parameters. The network structure of
pRVFLN and the interval-valued data cloud are depicted in Fig. 1 and 2 respectively.
Fig. 1 Network Architecture of pRVFLN
III.
Learning Policy of pRVFLN
This section discusses the learning policy of pRVFLN. Section 3.1 outlines the online active learning
strategy, which deletes inconsequential samples. Samples, selected in the sample selection mechanism,
are fed into the learning process of pRVFLN. Section 3.2 deliberates the hidden node growing strategy
of pRVFLN. Section 3.3 elaborates the hidden node pruning and recall strategy, while Section 3.4
concerns the online feature selection mechanism. Section 3.5 explains the parameter learning scenario of
pRVFLN. Algorithm 1 shows the pRVFLN learning procedure.
3.1
Online Active Learning Strategy
The active learning component of the pRVFLN is built on the extended sequential entropy (ESEM)
method, which is derived from the SEM method [64]. The ESEM method makes use of the entropy of the
neighborhood probability to estimate the sample contribution. There exist at least three salient facets that
distinguish the ESEM from its predecessor [64]:1) it forms an online version of the SEM; 2) it is combined
with the concept of the data cloud, which accurately represents the local density; and 3) it handles
regression as well as classification because the sample contribution is enumerated without the presence
of true class label. One may agree that the vast majority of sample selection variants are designed for
classification problems only, and delve sample’s location in respect to the decision surface. To the best
of our knowledge, only Das et al. [16] address the regression problem, but it still shares the same principle
as its predecessors, exploiting the hinge cost function to evaluate sample contribution [54]. The concept
of neighborhood probability refers to the probability of an incoming data stream sitting in the existing
data clouds, which is written as follows:
Ni
M ( X t , xk )
Ni
P( X t N i ) Rk 1N i
M ( X t , xk )
Ni
i 1 k 1
(8)
Algorithm 1. Learning Architecture of pRVFLN
Algorithm 1: Training Procedure of pRVFLN
( X t , Tt ) ( x1 ,.., xn , t1 ,.., tm )
Predefined Parameters 1 , 2
Define: Training Data
/*Step 1: Online Active Learning Strategy/*
For i=1 to R do
Calculate the neighborhood probability (8) with spatial firing strength (4)
End For
Calculate the entropy of neighborhood probability (8) and the ESEM (10)
IF (34) Then
/*Step 2: Online Feature Selection/*
IF Partial=Yes Then
Execute Algorithm 3
Else IF
Execute Algorithm 2
End IF
/*Step 3: Data Cloud Growing Mechanism/*
For j=1 to n do
Compute ( x j , To )
End For
For i=1 to R do
Calculate input coherence (12)
For o=1 to m do
~ ,T )
Calculate (
i o
Compute the output coherence (13)
End For
IF (16) Then
Associate a sample to a data cloud to the i* data cloud Ni*=Ni*+1
Update the local mean and the square length of i*-th data cloud (5), (6)
Else
Create new data cloud
Take the next sample and Go to Phase 1
End IF
/*Step 4: Data Cloud Pruning Mechanism/*
For i=1 to R do
For o=1 to m do
~
Calculate (Gi ,temp , To )
End For
IF (19) Then
Discard i-th data cloud
End IF
End For
/*Step 5: Adaptation of Output Weight/*
For i=1 to R do
Update output weights using FWGRLS
End For
End For
where XT is a newly arriving data point and xn is a data sample, associated with the i-th rule. M(XT,xk)
stands for a similarity measure, which can be defined as any similarity measure. The bottleneck problem
of (8) is however caused by its requirement to revisit already seen samples. This issue can be tackled by
formulating the recursive expression of (8). Instead of formulating the recursive definition of (8), the
spatial firing strength of the data cloud suffices to be an alternative because it is derived from the idea of
local density and is computed based on the local mean [1] which summarizes the characteristic of data
streams. (8) is then written as follows:
P( X t N i )
i
(9)
R
i 1
i
where i is a type-reduced activation degree
i (1 q)G i ,spatial qG i ,spatial .
is determined, its entropy is formulated as follows:
Once the neighbourhood probability
R
H ( X t ) P( X t N i ) log P( X t N i )
(10)
i 1
The entropy of the neighbourhood probability measures the uncertainty induced by a training pattern. A
sample with high uncertainty should be admitted for the model update, because it cannot be well-covered
by an existing network structure. Learning such a sample is beneficial, because it minimises uncertainty.
A sample is to be accepted for model updates, provided that the following condition is met:
H thres
(11)
where thres is an uncertainty threshold. This parameter is not fixed during the training process, rather it
is dynamically adjusted to suit the learning context. This strategy is necessary to compensate for the
potential increase of training samples to be accepted during the presence of concept drift. The threshold
is set as thresN 1 thresN (1 inc) , where it augments thresN 1 thresN (1 inc) when a sample is admitted
for the training process, whereas it decreases thresN 1 thresN (1 inc) when a sample is ruled out for the
training process. inc here is a step size, set at inc=0.01. This simply follows its default setting in [68].
Fig. 2 Interval Valued Data Cloud
3.2
Hidden Node Growing Strategy
pRVFLN relies on the T2SCC method to grow interval-valued data clouds on demands. This notion is
extended from the so-called SCC method [26], [65] to be well-suited with the type-2 hidden node working
framework. The significance of the hidden nodes in pRVFLN is evaluated by checking its input and
output coherence through analysis of its correlation to existing data clouds and the target concept. Let
~i [ i , i ] 1n is a local mean of the i-th interval-valued data cloud (5), X t
n
is an input vector and Tt m
is a target vector, the input and output coherence are written as follows:
I C ( i , X t ) (1 q) ( i , X t ) q ( i , X t )
(12)
OC ( i , X t ) ( ( X t , Tt ) ( i , Tt )), ( i , Tt ) (1 q) ( i , Tt ) q ( i , Tt )
(13)
where () is the correlation measure. Both linear and non-linear correlation measures are applicable here.
However, the non-linear correlation measure is rather hard to deploy in the online environment, because
it is usually executed using the Discretization or Parzen Window method [49]. This often leads to an
assumption of uniform data distribution as implemented in the differential entropy [28]. The Pearson
correlation measure is a widely used correlation measure
but it is insensitive to the scaling and
translation of variables as well as being sensitive to rotation [35]. The maximal information compression
index (MCI) is one of attempts to tackle these problems and is used in the T2SCC to perform the
correlation measure () [35]. It is defined as follows:
1
( X 1 , X 2 ) (var( X 1 ) var( X 2 ) (var( X 1 ) var( X 2 )) 2 4 var( X 1 ) var( X 2 )(1 ( X 1 , X 2 )2 ))
2
( X1 , X 2 )
where
(X1,X2)
cov( X 1 , X 2 )
var( X 1 ) var( X 2 )
are
to
be
(14)
(15)
substituted
by
( i , X t ), ( i , X t ), ( i , Tt ), ( i , Tt ), ( i , X t ), ( X t , Tt ) .
var( X ), cov( X 1 , X 2 ), ( X 1 , X 2 ) respectively stand for the variance of X, covariance of X1 and X2, and
Pearson correlation index of X1 and X2. The local mean of the interval-valued data cloud is used to
represent a data cloud because it captures the influence of all intervals of a data cloud. In essence, the
MCI method indicates the amount of information compression when ignoring a newly observed sample.
The principal component direction is referred to here, because it signifies the maximum information
compression, resulting in maximum cost to be imposed when ignoring a datum. The MCI method features
the following properties: 1) 0 ( X 1 , Y2 ) 0.5(var( X 1 ) var( X 2 )) , 2) a maximum correlation is given by
( X 1 , X 2 ) 0 , 3) a symmetric property ( X 1 , X 2 ) ( X 2 , X 1 ) , 4) the mean expression is discounted
here. This makes it invariant against the translation of the dataset, and 5) it is also robust against rotation,
which is verifiable from the perpendicular distance of a point to a line, and is unaffected by the rotation
of the input features.
The input coherence explores the similarity between new data and existing data clouds directly, while
the output coherence focusses on their dissimilarity indirectly through a target vector as a reference. The
input and output coherence formulates a test that determines the degree of confidence to the current’s
hypothesis:
I C ( ~i , X t ) 1 , OC ( ~ i , X t ) 2
(16)
where 1 [0.001,0.01], 2 [0.01,0.1] are predefined thresholds. If a hypothesis meets both conditions, a
new training sample is assigned to a data cloud with the highest input coherence i*. Accordingly, the
~
number of intervals Ni*, local mean and square length ~i* , i* are updated respectively with (5) and (6) as
well as Ni*=Ni*+1. A new data cloud is introduced, provided that none of the existing hypotheses pass the
test (16) – one of the conditions is violated. This situation reflects the fact that a new training pattern
conveys significant novelty, which has to be incorporated to enrich the scope of the current hypotheses.
Note that a larger α1 is specified, fewer data clouds are generated and vice versa, whereas a larger α2 is
specified, larger data clouds are added and vice versa. The sensitivity of these two parameters are studied
in the other section of this paper. Because a data cloud is a non-parametric model, no parameterization is
committed for a new data cloud. The output node of a new data cloud is initialised:
WR 1 Wi* , R1 I
(17)
where 10 5 is a large positive constant. The output node is set as the data cloud with the highest input
coherence because this data cloud is the closest one to the new data cloud. Furthermore, the setting of
covariance matrix can produce a close approximation of the global minimum solution of batched learning,
as proven mathematically in [31].
3.3
Hidden Node Pruning and Recall Strategy
pRVFLN incorporates a data cloud pruning scenario, termed the type-2 relative mutual information
(T2RMI) method. This method was firstly developed in [21] for the type-1 fuzzy system and extended in
[41] to adapt to the interval-valued working principle. This method is convenient to use in pRVFLN
because it estimates mutual information between a data cloud and a target concept by analysing their
correlation. Hence, the MCI method (14) - (15) can be applied to measure the correlation. Although this
method has been well-established [21], [41], its effectiveness in handling data clouds and a recurrent
structure as implemented in pRVFLN is to date an open question. Unlike both the RMI method [36] and
T2RMI method [41] that apply the classic symmetrical uncertainty method, the T2RMI method is
formulated using the MCI method as follows:
(Gi ,temp , Tt ) q (G i ,temp , Tt ) (1 q) (G i ,temp , Tt )
(18)
where G i ,temp is a lower temporal activation function of the i-th rule. The temporal activation function is
included in (17) rather than the spatial activation function in order for the inter-temporal dependency of
the recurrent structure to be considered. The MCI method is chosen here, because it offers a good tradeoff between accuracy and simplicity. It possesses significantly lower computational burden than the
symmetrical uncertainty method even when implemented with the differential entropy [28] but is more
robust than a widely used Pearson correlation index. A data cloud is deemed inconsequential and thus is
able to be removed with negligible impact to accuracy, if (19) is met:
i mean( i ) 2std ( i )
(19)
where mean( i ), std ( i ) are respectively the mean and standard deviation of the MCI during its lifespan.
This criterion aims to capture an obsolete data cloud, which does not keep up with current data distribution
due to possible concept drift, because it computes the downtrend of the MCI during its lifespan. It is worth
mentioning that mutual information between hidden nodes and the target variable is a reliable indicator
for changing data distributions, because it is in line with the definition of concept drift. Concept drift
refers to a situation, where the posterior probability changes overtime P(Tt X t ) P(Tt 1 X t 1 ) .
The T2RMI method also functions as a rule recall mechanism, which is capable of coping with cyclic
concept drift. Cyclic concept drifts frequently happen in the weather, customer preference, electricity
power consumption problems, etc. where seasonal change comes into picture. This points to a situation
where previous data distribution reappears again in the current training step. Once pruned by the T2RMI,
a data cloud is not forgotten permanently and is inserted into a list of pruned data clouds R*=R*+1. In
this case, its local mean, square length, population, an output node, and output covariance matrix
~
~R* , R* , N R* , R* , R* are retained in the memory. Such data clouds can be reactivated again in the future,
whenever its validity is confirmed by an up-to-date data trend. It is worth noting that adding a completely
new data cloud when observing previously learned concept violates the notion of an evolving learner and
catastrophically erases learning history. A data cloud is recalled subject to the following condition:
max ( i* ) max( i )
i *1,.., R*
i 1,.., R
(20)
This situation reveals that a previously pruned data cloud is more relevant than any existing ones. This
condition pinpoints that a previously learned concept may reappear again. A previously pruned data cloud
is then regenerated as follows:
~
~
~R 1 ~R* , R 1 R* , N R 1 N R* , R 1 R* , R 1 R*
(21)
Note that although previously pruned data clouds are stored in the memory, the data cloud pruning
module still contributes in lowering the computational load, because all previously pruned data clouds
are excluded from any training scenarios except (17). Unlike its predecessors [39], this rule recall scenario
is not involved in the data cloud growing process (please refer to Algorithm 1) and plays a role as another
data cloud generation mechanism. This mechanism is also developed from the T2RMI method, which can
represent the change of posterior probability – concept drift - more accurately than the density concept.
3.4
Online Feature Selection Strategy
Although feature selection and extraction problems have attracted considerable research attention;
little effort has been paid toward online feature selection. Two common approaches to tackling this issue
are through soft or hard dimensionality reduction techniques [39], [46]. Soft dimensionality reduction
minimizes the effect of inconsequential features by assigning low weights but still retains a complete set
of input attributes in the memory, whereas hard dimensionality reduction lowers the input dimension by
cutting off spurious input features. Nonetheless, the hard dimensionality reduction method undermines
stability, because an input feature cannot be retrieved once pruned [5]. To date, most of existing works
always start the input selection process from a full set of input attributes and gradually reduces the number
as more observation are come across. A prominent work, namely online feature selection (OFS), was
developed in [3] and covers both partial and full input conditions. The appealing trait of OFS lies in its
aptitude for flexible feature selection, where it enables the provision of different combinations of input
attributes in each episode by activating or deactivating input features (1 or 0), which adapts to up-to-date
data trends. Furthermore, this technique is also capable of handling partial input attributes which happens
to be fruitful when the cost of feature extraction is too expensive. OFS was originally devised for the LR
and is generalized here to fit the context of pRVFLN.
We start our discussion from a condition where a learner is provided with full input variables. Suppose
that B input attributes are to be selected in the training process and B<n, the simplest approach is to
R
discard the input features with marginal accumulated output weights
2
i 1 j 1
i, j
and maintain only B input
2
features with the largest output weights. Note that the second term is required because of the extended
j 1
input vector xe (2 n1) . The rule consequent informs a tendency of a rule which can be used as an
alternate of gradient information which changes in each point [35]. Although it is straightforward to use,
it cannot ensure the stability of the pruning process due to a lack of sensitivity analysis of the feature
contribution. To correct this problem, a sparsity property of the L1 norm can be analyzed to examine
whether the values of n input features are concentrated in the L1 ball. This allows the distribution of the
input values to be checked to determine whether they are concentrated in the largest elements and that
pruning the smallest elements won’t harm the model’s accuracy. This concept is actualized by first
inspecting the accuracy of pRVFLN. The input pruning process is carried out when the system error is
large enough Tt yt or, in the realm of the classification problem, misclassification is made.
Nevertheless, the system error is not only large in the case of underfitting, but also in the case of
overfitting. We modify this condition by taking into account the evolution of system error
et t et 1 t 1 . The constant is a predefined parameter and fixed at 1.1 for simplicity. The
output nodes are updated using the gradient descent approach and then projected to the L2 ball to guarantee
a bounded norm. Algorithm 2 details the algorithmic development of pRVFLN.
Algorithm 2. GOFS using full input attributes
Input : α learning rate, χ regularization factor, B the number of features to be retained
Output: selected input features
X t , selected 1 B
For t=1,…., T
Make a prediction yt
IF et t 1.1 et 1 t 1 // for regression or oˆ max( yo ) Tt // for classification
o 1,..,m
i i i
E
i
,
1
i min(1,
i
2
) i
R 2
Prune input attributes X t except those of B largest
i 1 j 1
i, j
Else
i i i
End IF
End FOR
where , are respectively the learning rate and regularization factor. We assign 0.2, 0.01
following the same setting as [3]. The optimization procedure relies on the standard mean square error
(MSE) as the objective function, which leads to the following gradient term:
R
E
R
(Tt yt ) (1 q)G i ,temporal qG i ,temporal
i
i 1
i 1
Furthermore, the system error has been theoretically proven to be bounded in [59] and the upper bound
is also found. One can also notice that the GOFS enables different feature subsets to be elicited in each
training observation t.
A relatively unexplored area of existing online feature selection is a situation where a limited number
of features is accessible for the training process. To actualise this scenario, we assume that at most B input
variables can be extracted during the training process. This strategy, however, cannot be done by simply
acquiring any B input features, because this scenario results in the same subset of input features during
the training process. This problem is addressed using the Bernaoulli distribution with confidence level
to sample B input attributes from n input attributes B<n. Algorithm 3 displays the feature selection
procedure.
Algorithm 3. GOFS using partial input attributes
Input : α learning rate, χ regularization factor, B the number of features to be retained, confidence level
Output: selected input features
X t , selected 1 B
For t=1,…., T
Sample from Bernaoulli distribution with confidence level
IF t 1
~
Randomly select B out of n input attributes X t 1 B
End IF
Make a prediction yt
IF et t 1.1 et 1 t 1 // for regression or oˆ max( yo ) Tt // for classification
o 1,..,m
~
Xˆ t X t ( B / n ) (1 )
i i i
,
E ( Xˆ t )
i min(1,
i
1
i
) i
2
R
2
Prune input attributes X t except those of B largest i , j
i 1 j 1
Else
i ,t i ,t 1
End IF
End FOR
As with Algorithm 2, the convergence of this scenario has been theoretically proven and the upper
bound is derived in [3]. One must bear in mind that pruning process in Algorithm 1, 2 are carried out by
assigning crisp weights (0 or 1), which fully reflects the importance of input features.
3.5
Random Learning Strategy
pRVFLN adopts the random parameter learning scenario of the RVFLN leaving only the output nodes
W to be analytically tuned with an online learning scenario, whereas others, namely At , q, , , can be
randomly generated without any tuning process. To begin the discussion, we recall the output expression
of pRVFLN as follows:
R
~
yo i Gi ,temporal ( X t ; At , q, , )
(22)
i 1
Note that the pRVFLN possesses the enhancement layer expressed as a nonlinear mapping using the
Chebyshev function i xe wi as with the original RVFLN [48]. Referring to the RVFLN theory, the
~
hidden node Gi ,spatial and its derivative must be integrable:
G ( x)dx , or [G( x)] dx
2
R
2
(23)
R
Furthermore, a large number of hidden nodes R is usually selected to ensure adequate coverage of hidden
nodes to a data space because they are chosen at random [41]. Nevertheless, this condition can be relaxed
in the pRVFLN, because the data cloud growing mechanism, namely the T2SCC method, partitions the
input region in respect to real data distributions. The concept of data cloud-based neurons features the
concept of local density, which adapts to any variation of data streams. Furthermore, this concept is
parameter-free and thus does not require any parameterization, which calls for a high-level approximation
or complicated optimization procedure. Other parameters, namely At , q, , , are randomly chosen, and
their region of randomisation should be carefully selected. Referring to Ingelnik and Pao [20], the random
parameters are sampled randomly from the following.
w [0,] [,]n
[0,1]n
[2n, 2n]
b w
(24)
where , , are probability measures. Nevertheless, this strategy is impossible to implement in the
online situations, because it often entails a rigorous trial-error process to determine these parameters.
Most RVFL works simply follow Schmidt et al.’s strategy [53], setting the region of random parameters
in the range of [-1,1].
Assuming that a complete dataset [ X , T ] N ( n m ) is observable, a closed-form solution of (22)
can be defined to determine the output weights W . Although the original RVFLN adjusts the output weight
with the conjugate gradient (CG) method, the closed-form solution can be still utilised with ease [48].
The mere obstacle for the use of pseudo-inversion in the original work was the issue of limited
computational resources in 90’s. Note that the regularization technique needs to be undertaken if the
hidden node matrix is ill-conditioned. Although it is easy to use and ensures a globally optimum solution,
this parameter learning scenario however imposes revisiting preceding training patterns which is
intractable for online learning scenarios. pRVFLN employs the FWGRLS method [35], [39] to adjust the
output weight which is underpinned by a generalized weight decay function. Because the FWGRLS
approach has been detailed in [39], it is not recounted here.
3.6
Robustness of RVFLN
The network parameters are usually sampled uniformly within a range of [-1,1] in the literature. A new
finding of Li and Wang in [29] exhibits that the process of randomly generating network parameters with
a fixed scope [-α, α] does not ensure a theoretically feasible network or often the hidden node matrix is
not full rank. Surprisingly, the hidden node matrix was not invertible in all their cases when randomly
sampling network parameters in the range of [-1,1] and far better numerical results were achieved by
choosing the scope [-200,200]. This trend was consistent with different numbers of hidden nodes. How
to properly select scopes of random parameters and its corresponding distribution still require in-depth
investigation [53]. In practice, a pre-training process is normally required to arrive at decent local
mapping. Note that the range of random parameters by Ingelnik and Pao [20] is still at the theoretical
level and does not touch implementation issue. We study different random regions in Section IV. C to see
how pRVFLN behaves under variations of scope of random parameters.
IV.
Numerical Examples
This section presents numerical validations of our proposed algorithm using case studies and
comparisons with prominent algorithms in the literature. Two numerical examples, namely modelling of
Nox emissions from a car engine, tool condition monitoring in the ball-nose end milling process, are
presented in this section, and the other three numerical examples are placed in the supplemental
document1) to keep the paper compact. Our numerical studies were carried out under two scenarios: the
time-series scenario and the cross-validation scenario. The time-series procedure orderly executes data
streams according to their arrival time and partitions data streams into two parts, namely training and
testing. In the time-series mode, pRVFLN was compared against 10 state-of-the-art evolving algorithms:
eT2Class [40], BARTFIS [37], PANFIS [42], GENEFIS [43], eTS+ [5], eTS [6], simp_eTS [7], DFNN
[61], GDFNN [62], FAOSPFNN [60], ANFIS [25]. The CV scenarios were taken place in our experiment
in order to follow the commonly adopted simulation environment of other RVFLNs in the literature. The
pRVFLN was benchmarked against 8 prominent RVFLNs in the literature: decorelated neural network
ensemble (DNNE) [8], centralized RVFL (A) [51], RLS-consensus-RVFL [51] (B), RLS-local-RVFL (C)
[51], LMS-concensus-RVFL (D) [51]. The MATLAB codes of these algorithms are available online 2,3).
Comparisons were performed against five evaluation criteria: accuracy, data clouds, input attribute,
runtime, network parameters. The scope of random parameters followed Schmidt's suggestion [53] where
the scope of random parameters was [-1,1] but we insert analysis of robustness in part C which provides
additional results with different random regions. The effect of individual learning component to the end
results and the influence of user-defined predefined thresholds are analysed in the supplemental
document1). To allow a fair comparison, all consolidated algorithms were executed in the same
computational resources under the MATLAB’s environment.
Table 1. Prediction of Nox emissions Using Time-Series Mode
Model
pRVFLN (P)
pRVFLN (F)
eT2Class
PANFIS
GENEFIS
Simp_eTS
BARTFIS
DFNN
GDFNN
eTS
FAOS-PFNN
ANFIS
eTS+
RMSE
0.04
0.05
0.045
0.052
0.048
0.14
0.11
0.18
0.48
0.38
0.06
0.15
0.14
Node
2
2
2
5
2
5
4
548
215
27
6
2
4
Input
5
5
170
170
2
170
4
170
170
170
170
170
4
Runtime
0.24
0.56
17.98
3.37
0.41
5.5
2.55
4332.9
2144.1
1098.4
14.8
100.41
a)
Network
22
22
117304
146205
18
1876
52
280198+NS
109865
13797
2216+NS
17178
28
Samples
510
515
667
667
667
667
667
667
667
667
667
667
667
a): the result was obtained under different computer platform
Table 2. Prediction of Nox emissions Using CV Mode
Model
pRVFLN(P)
pRVFLN(F)
DNNE
A-online
B-online
C-online
D-online
A-batch
B-batch
C-batch
NRMSE
0.12±0.04
0.12±0.03
0.14±0
0.52±0.02
0.57±0.01
0.93±0.02
1.7±0.04
0.59±0.05
0.62±0.05
0.83±0.09
Node
2
2
50
100
100
100
100
100
100
100
Input
5
5
170
170
170
170
170
170
170
170
Runtime
1.1
5.98
0.81
0.03
0.41
0.21
0.09
0..003
0.001
0.001
Network
22
22
43600+NS
87200
87200
87200
87200
87200+NS
87200+NS
87200+NS
Samples
699
630.7
744
744
744
744
744
744
744
744
A. Modeling of Nox Emissions from a Car Engine: this section demonstrates the efficacy of the pRVFLN
in modeling Nox emissions from a car engine [32]. This real-world problem is relevant to validate the
learning performance, not only because it features noisy and uncertain characteristics as the nature of a
car engine, but also it characterizes high dimensionality, containing 170 input attributes. That is, 17
physical variables were captured in 10 consecutive measurements. Furthermore, different engine
parameters were applied to induce changing system dynamic to simulate real driving conditions. In the
time-series procedure, 826 data points were streamed to consolidated algorithms, where 667 samples were
1)
https://www.dropbox.com/s/lytpt4huqyoqa6p/supplemental_document.docx?dl=0
http://homepage.cs.latrobe.edu.au/dwang/html/DNNEweb/index.html
3)
http://ispac.ing.uniroma1.it/scardapane/software/lynx/.
2)
set as training samples, and the remainder were fed for testing purposes. In the CV procedure, the
experiment was run under the 10-fold CV scheme, and each fold was repeated five times similar to the
scenario adopted in [22]. This strategy checked the consistency of the RVFLN’s learning performance
because it adopts the random learning scenario and avoids data order dependency. Table 1 and 2 exhibit
consolidated numerical results of benchmarked algorithms.
Table 3. Tool Wear Prediction Using Time Series Mode
Model
pRVFLN (P)
pRVFLN (F)
eT2Class
Simp_eTS
eTS
BARTFIS
PANFIS
GENEFIS
DFNN
GDFNN
FAOS-PFNN
ANFIS
RMSE
0.14
0.19
0.16
0.22
0.15
0.16
0.15
0.13
0.27
0.26
0.38
0.16
Node
2
2
4
17
7
6
3
42
42
7
7
8
Input
8
8
12
12
12
12
12
12
12
12
12
12
Runtime
0.34
0.15
1.1
1.29
0.56
0.43
0.77
0.88
2.41
2.54
3.76
0.52
Network
34
34
1260
437
187
222
507
507
1092+NS
259+ NS
1022+NS
296+ NS
Samples
157
136
320
320
320
320
320
320
320
320
320
320
Table 4. Tool wear prediction using CV Mode
Model
pRVFLN(P)
pRVFLN(F)
DNNE
A-online
B-online
C-online
D-online
A-batch
B-batch
C-batch
NRMSE
0.16±0.01
0.17±0.06
0.11±0
0.16±0.01
0.21±0.01
1.39±0.03
0.25±0.01
0.19±0.04
0.21±0.03
0.23±0.05
Node
1.3±0.2
1.92±0.1
50
100
100
100
100
100
100
100
Input
8
8
12
12
12
12
12
12
12
12
Runtime
0.2
0.3
0.79
0.02
0.05
0.13
0.09
0..002
0.009
0.007
Network
22.1
32.6
3310+NS
1400
1400
1400
1400
1400+NS
1400+NS
1400+NS
Samples
245.8
432.1
571.5
571.5
571.5
571.5
571.5
571.5
571.5
571.5
It is evident that pRVFLN outperformed its counterparts in all the evaluation criteria except GENEFIS
for the number of input attributes and network parameters. It is worth mentioning however that in the
other three criteria: predictive accuracy, execution time, and number of training samples, the GENEFIS
was more inferior than ours. pRVFLN is equipped with the online active learning strategy, which
discarded superfluous samples. This learning module had a significant effect of predictive accuracy.
Furthermore, pRVFLN has the GOFS method, which was capable of coping with the curse of
dimensionality. Note that the unique feature of the GOFS method is that it allows different feature subsets
to be picked up in every training episode. This case avoids the catastrophic forgetting of obsolete input
attributes, temporarily inactive due to changing data distributions. The GOFS can handle partial input
attributes during the training process and resulted in the same level of accuracy as that of the full input
attributes. The use of full input attributes slowed down the execution time because it needed to deal with
170 input variables first, before reducing the input dimension. In this case study, we selected five input
attributes to be kept for the training process. Our experiment showed that the number of selected input
attributes is not problem-dependent and able to be set as any desirable number in most cases. There was
no significant performance difference when using either the full input mode or partial input mode. On the
other hand, consistent numerical results were achieved by pRVFLN, although the pRVFLN is built on
the random vector functional link algorithm, as observed in the CV experimental scenario. In addition,
pRVFLN produced the most encouraging performance in almost all evaluation criteria but computational
speed because the other RVFLNs, (A)-(D), implement less comprehensive training procedure than
pRVFLN by only fine-tuning output weights without any structural learning and feature selection
mechanisms.
B. Tool Condition Monitoring of High-Speed Machining Process: this section presents a real-world
problem, taken from a complex manufacturing problem (Courtesy of Dr. Li Xiang, Singapore) [44]. The
objective of this case study is to perform predictive analytics of the tool wear in the ball-nose end milling
process frequently found in the metal removal process of the aerospace industry. In total, 12 time-domain
features were extracted from the force signal and 630 samples were collected during the experiment.
Concept drift in this case study was resulted from changing surface integrity, tool wear degradation as
well as varying machining configurations. For the time-series experimental procedure, the consolidated
algorithms were trained using data from cutter A, while the testing phase exploited data from cutter B.
For the CV experimental procedure, the 10-fold CV process was undertaken where each fold was
undertaken five times to arrive at consistent finding. Tables 3 and 4 report average numerical results
across all folds. Fig. 3 depicts how many times input attributes are selected during one fold of CV process.
Fig. 3 the frequency of input features
It is observed from Table 3 and 4 that pRVFLN evolved the lowest structural complexities while
retaining a very high level of accuracy. It is worth noting that although the DNNE exceeded pRVFLN in
accuracy, it imposed considerable complexity because it is an offline algorithm revisiting previously seen
data samples and adopts an ensemble learning paradigm. The efficacy of the online sample selection
strategy was seen, as it led to significant reduction of training samples to be learned during the experiment.
Using partial input information led to subtle differences to those with the full input information. It is seen
in Fig. 3 that the GOFS selected different feature subsets in every training episode and it reflected the
importance of input variables to the training process. Additional numerical examples, sensitivity analysis
of predefined thresholds and analysis of learning modules can be found in 1).
C. Analysis of Robustness: this section aims to numerically validate our claim in section III. E that a
range [-1,1] does not always assure a reliable model [29],[53]. Additional numerical results with different
intervals of random parameters are presented. Four intervals, namely [0,0.1],[0,0.5] [0,0.8] and [0,5] were
tried out for two case studies in section IV.A and IV.B. Our experiments were undertaken in the 10-folds
CV procedure where each fold was repeated five times to prevent the effect of random sampling. Our
experiment made use of the feature selection scenario of a full number of input attributes. Table 5 displays
numerical results.
Scope
[0,0.1]
[0,0.5]
[0,0.8]
[0,5]
Table 5. Analysis of Robustness
Criteria
Tool Wear
RMSE
0.13±0.008
Node
1.8±0.25
Input
8
Runtime
0.2±0.1
Network
30.9
Samples
503.1
RMSE
0.14±0.02
Node
1.92±0.2
Input
8
Runtime
0.18±0.008
Network
32.6
Samples
571.5
RMSE
0.47±0.42
Node
1.4±0.05
Input
8
Runtime
0.19±0.13
Network
23.8
Samples
385.1
RMSE
Node
Input
Unstable
Runtime
Network
Samples
Nox emission
1.9±0
1
5
0.1±0.02
11
1
0.1±0.01
1.98±0.14
5
5.7±0.3
21.8
743.4
0.18±0.3
1.96±0.19
5
5.56±0.96
21.6
711.24
Unstable
For the tool wear case study, the best-performing model was generated by the range [0,0.1]. The higher
the range of the model the more inferior the model was. It went up to the point where a model was no
longer stable under the range [0,5]. On the other side, the range [0,0.5] induced the best-performing model
with the highest accuracy while evolving comparable network complexity for the Nox emission case
study. The higher the scope led to the deterioration of numerical results. Moreover, the range [0,0.1] did
not deliver a better accuracy than the range [0,0.5] either since this range did not generate diverse enough
random values. These numerical results are interpreted from the nature of pRVFLN – a clustering-based
algorithm. The success of pRVFLN is mainly determined from the compatibility of zone of influence of
hidden nodes to a real data distribution, and its performance worsens when the scope is remote from the
true data distribution. This finding is complementary to Li and Wang [29] where it relies on a sigmoidbased RVFL network, and the scope of random parameters can be outside the applicable operating
intervals. Its predictive performance is set by its approximation capability in the output space.
V.
Conclusions
A novel random vector functional link network, namely parsimonious random vector functional link
network (pRVFLN), is proposed. pRVFLN aims to provide a concrete solution to the issue of data stream
by putting into perspective a synergy between adaptive and evolving characteristics and fast and easy-touse characteristics of RVFLN. pRVFLN is a fully evolving algorithm where its hidden nodes can be
automatically added, pruned and recalled dynamically while all network parameters except the output
weights are randomly generated with the absence of any tuning mechanism. pRVFLN is fitted by the
online feature selection mechanism and the online active learning scenario which further strengthens its
aptitude in processing data streams. Unlike conventional RVFLNs, the concept of interval-valued data
clouds is introduced. This concept simplifies the working principle of pRVFLN because it neither requires
any parameterization per scalar variables nor follows pre-specified cluster shape. It features an intervalvalued spatiotemporal firing strength, which provides the degree of tolerance for uncertainty. Rigorous
case studies were carried out to numerically validate the efficacy of pRVFLN where pRVFLN delivered
the most encouraging performance. The ensemble version of pRVFLN will be the subject of our future
investigation which aims to further improve the predictive performance of pRVFLN.
VI.
ACKNOWLEDGEMENT
The third author acknowledges the support of the Austrian COMET-K2 program of the Linz Center of Mechatronics (LCM),
funded by the Austrian federal government and the federal state of Upper Austria. We thank Dr. D. Wang for his suggestion
pertaining to robustness issue of RVFLN.
VII.
References
[1] P. Angelov, R. Yager, “ A new type of simplified fuzzy rule-based system”, International Journal of General Systems, pp. 1-21, 2011
[2] P.Angelov,“ Anomalous System State Identification”, Patent gb1208542.9, Priority Date 15May2012
[3] P. Angelov, “ Autonomous Learning Systems: From Data Streams to Knowledge in Real-time”, John Wiley and Sons Ltd., 2012
[4] P. Angelov , D. Kangin , X. Zhou , D. Kolev, “Symbol recognition with a new autonomously evolving classifier autoclass”, in Proceedings of 2014
IEEE Conference on Evolving and Adaptive Intelligent Systems, 2014, pp. 1–7
[5] P. Angelov, “ Evolving Takagi-Sugeno Fuzzy Systems from Data Streams (eTS+)”, In Evolving Intelligent Systems: Methodology and Applications
(Angelov P., D. Filev, N. Kasabov Eds.), John Willey and Sons, IEEE Press Series on Computational Intelligence, pp. 21-50, ISBN: 978-0-470-287194, April 2010
[6] P.Angelov and D. Filev, "An approach to online identification of Takagi-Sugeno fuzzy models," IEEE Transactions on Systems, Man, and
Cybernetics, Part B. vol. 34, pp. 484-498. 2004
[7] P.Angelov and D. Filev, "Simpl_eTS: A simplified method for learning evolving Takagi-Sugeno fuzzy models," in IEEE International Conference on
Fuzzy Systems (FUZZ), pp. 1068-1073. 2005
[8] M. Alhandoosh, D. Wang, “ Fast decorrelated neural network ensembles with random weights”, Information Sciences, Vol.264, pp. 104-117, 2014
[9] R. H. Abiyev and O. Kaynak, “Type-2 fuzzy neural structure for identification and control of time-varying plants,” IEEE Transactions on Industrial
Electronics , vol. 57, no. 12, pp. 4147–4159, 2010
[10] H. Bakırcıo’glu, T. Kocak, “Survey of random neural network applications”, European Journal of Operational Research, Vol. 126(2), pp. 319–330,
2000
[11] P. L. Barlett, T. Downs, “ Using random weights to train multilayer networks of hard-limiting units”, IEEE Transactions on Neural Networks, Vol.
3(2), pp. 202–210, 1992
[12] H. Berry, Quoy,” Structure and dynamics of random recurrent neural networks”, Adaptive Behaviour, Vol 14(2), pp. 129–137, 2006
[13] D. S. Broomhead and D. Lowe, “Multivariable functional interpolation and adaptive networks,” Complex Systems, vol. 2, pp. 321–355, 198
[14] F. Cao, Y. Tan, M. Cai, “ Sparse algorithms of Random Weight Networks and applications”, Expert Systems with Applications, Vol. 41, pp. 24572462, 2014
[15]F. Cao, H. Ye, D. Wang, “ A probabilistic learning algorithm for robust modeling using neural networks with random weights”, Information Sciences,
Vol.313, pp. 62-78, 2015
[16] A. K. Das, K. Subramanian, S. Suresh, “ An Evolving Interval Type-2 Neurofuzzy Inference System and Its Metacognitive Sequential Learning
Algorithm. IEEE Transactions on Fuzzy Systems, 23(6), pp. 2080-2093, 2015
[17] A. Elissee, H. Paugam-Moisy, ” JNN, a randomized algorithm for training multilayer networks in polynomial time”, Neurocomputing, Vol. 29(1),
pp. 3–24, 1999
[18] S. Haykin,“ Neural Networks: A comprehensive foundation. Neural Networks”, Prentice-Hall, 2nd edition, 1998
[19] M. T. Hagan and M. B. Manhaj, “Training feedforward networks with the Marquardt algorithm,” IEEE Transactions on Neural Networks, vol. 5,
pp. 989–993, 1994
[20] B. Igelnik, Y.H. Pao, “Stochastic choice of basis functions in adaptive function approximation and the functional-link net”, IEEE Transactions on
Neural Networks, Vol.6(6), pp.1320-1329, 1995
[21] H. Gan, X. L. Wu, J. F. Qiao, “Nonlinear Systems Modeling Based on Self-Organizing Fuzzy-Neural-Network with Adaptive Computation
Algorithm “, IEEE Transactions on Cybernetics, Vol. 44(4), pp.554-564, 2014
[22] C.F. Juang and C.T. Lin, “A recurrent self-organizing neural fuzzy inference network,” IEEE Transactions on Neural Networks, vol.10, pp.828-845,
1999
[23] C. F. Juang, Y. Y. Lin, and C. C. Tu, “A recurrent self-evolving fuzzy neural network with local feedbacks and its application to dynamic system
processing,” Fuzzy Sets and Systems, vol.161(19), 2010
[24] C.F. Juang, “A TSK-type recurrent fuzzy network for dynamic systems processing by neural network and genetic algorithms,” IEEE Transactions
on Fuzzy Systems, vol.10(2), pp.155-170, 2002
[25] J.-S. R. Jang, “ANFIS: Adaptive-network-based fuzzy inference system”, IEEE Transaction on System. Man. Cybernetic, part b: cybernetics, vol.
23, pp. 665–684, 1993
[26] J-Y. Jiang, R-J. Liao, S-J. Lee, “ A Fuzzy Self-Constructing Feature Clustering Algorithm for Text Classification”, IEEE Transactions on Knowledge
and Data Engineering, Vol. 23(3), pp. 335-349, 2011
[27] D. Kangin, P. Angelov, J. A. Iglesias, “ Autonomously evolving classifier TEDAClass”, Information Sciences, Vol. 366, pp. 1-11, 2016
[28] A. Lazo, P. N. Rathie,” On the entropy of continuous probability distributions”, IEEE Transactions on Information Theory, 24(1), 120–122, 1978
[29] M. Li, D. Wang, “ Insights into randomized algorithms for neural networks: Practical issues and common pitfalls”, Information Sciences, Vol. 382,
pp. 170-178, 2017
[30] Q. Liang and J. M. Mendel, “Interval type-2 fuzzy logic systems: Theory and design,” IEEE Transactions on Fuzzy Systems, vol. 8, no. 5, pp. 535–
550, 2000
[31] E. Lughofer, Evolving Fuzzy Systems --- Methodologies, Advanced Concepts and Applications, Springer, Heidelberg, 2011
[32] E. Lughofer, V. Macian, C. Guardiola and E.P Klement, “ Identifying Static and Dynamic Prediction Models for NOx Emissions with Evolving
Fuzzy Systems”, Applied Soft Computing, vol. 11(2), pp. 2487-2500, 2011
[33] Yang-Yin Lin, Jyh-Yeong Chang, Chin-Teng Lin,” Identification and Prediction of Dynamic Systems Using an Interactively Recurrent Self-Evolving
Fuzzy Neural Network”, IEEE Transactions on Neural Networks and Learning Systems, Vol.24,no.2,pp.310-321, 2013
[34] T.Martin, “Fuzzy sets in the fight against digital obesity,” Fuzzy Sets and Systems, vol. 156(3), pp. 411–417, 2005
[35] P. Mitra, C.A. Murthy, S.K. Pal,” Unsupervised feature selection using feature similarity”, IEEE Transactions on Pattern Analysis and Machine
Intelligence, Vol.24(3), pp.301-312, 2002
[36] D. Mackay, “A practical Bayesian framework for backpropagation networks,” Neural Computing., vol. 4, pp. 448–472, 1992
[37] R. J. Oentaryo, M-J. Er, S. Lin, X. Li, "Online probabilistic learning for fuzzy inference systems," Expert Systems with Applications, vol. 41(11),
pp. 5082-5096, 2014
[38] J.C. Patra, A.C. Kot, “ Nonlinear dynamic system identification using Chebyshev functional link artificial neural networks“, IEEE Transactions on
Systems, Man and Cybernetics-Part B: Cybernetics, Vol.32,no.4, pp.505-511, 2002
[39] M.Pratama, S. Anavatti, M-J. Er, E. Lughofer, “ pClass: An Effective Classifier to Streaming Examples”, IEEE Transactions on Fuzzy Systems,
Vol.23(2), pp.369-386, 2014
[40] M. Pratama, J. Lu, G.Zhang, “ Evolving Type-2 Fuzzy Classifier”, IEEE Transactions on Fuzzy Systems, Vol. 24(3), pp. 574-589, 2016
[41] M. Pratama, G. Zhang, M-J. Er, S. Anavatti, An Incremental Type-2 Meta-cognitive Extreme Learning Machine, IEEE Transactions on Cybernetics,
Vol 47(2), pp. 339-353, 2016
[42] M. Pratama, S. Anavatti, P. Angelov, E. Lughofer, PANFIS: A Novel Incremental Learning, IEEE Transactions on Neural Networks and Learning
Systems, Vol.25(1), pp.55-68, 2014
[43] M.Pratama, S.Anavatti, E.Lughofer, “ GENEFIS:Towards An Effective Localist Network”, IEEE Transactions on Fuzzy Systems, Vol.25(3), pp.
547-562, 2014
[44] M. Pratama, M-J. Er, X. Li, R. J. Oentaryo, E. Lughofer, I. Arifin, "Data driven modeling based on dynamic parsimonious fuzzy neural network,"
Neurocomputing, vol. 110, pp. 18-28, 2013
[45] M. Pratama, J.Lu, E. Lughofer, G. Zhang, S. Anavatti," Scaffolding Type-2 Classifier for Incremental Learning under Concept Drifts",
Neurocomputing, Vol.191, pp. 304-329, 2016
[46] M.Pratama, S.Anavatti, J.Lu, Recurrent Classifier Based on an Incremental Meta-cognitive-based Scaffolding Algorithm, IEEE Transactions on
Fuzzy Systems, Vol.23(6), pp. 2048-2066, 2015
[47] Y.-H. Pao, S.M. Phillips, ” The functional link net and learning optimal control”, Neurocomputing, Vol.9(2), pp. 149–164, 1995
[48] Y-H. Pao, G. H. Park, D. J. Sobajic, “Learning and generalization characteristics of random vector functional-link net,” Neurocomputing, vol. 6, pp.
163-180, 1994
[49] E. Parzen,” On estimation of a probability density function and mode”, The Annals of Mathematical Statistics, Vol.33, 1065–1076, 1962
[50] A. Rubio-Solis, G. Panoutsos, “Interval Type-2 Radial Basis Function Neural Network: A Modeling Framework”, IEEE Transactions on Neural
Networks and Learning Systems, Vol. 23(2), pp.457-473, 2015
[51] S. Scardapane, D. Wang, M. Panella, A. Uncini, “ Distributed learning for Random Vector Functional-Link networks”, Information Sciences, Vol.
301, pp. 271-284, 2015
[52] S. Scardapane, D. Wang, “ Randomness in Neural Networks: An Overview”, WIRE’s Data Mining and Knowledge Discovery, 2017
[53] W. Schmidt, M. Kraaijveld, R. Duin,“ Feedforward neural networks with random weights”, in Proceedings of 11th IAPR International Conference
on Pattern Recognition Methodology and Systems, 1992, pp. 1-4
[54] K. Subramanian, S. Suresh, N. Sundararajan, “A Meta-Cognitive Neuro-Fuzzy Inference System (McFIS) for sequential classification systems”,
IEEE Transactions on Fuzzy Systems, vol.21, no.6, pp.1080-1095, 2013
[55] R. Setiono, L. C. K. Hui, “Use of quasi-Newton method in a feedforward neural network construction algorithm,” IEEE Transactions on Neural
Networks, vol. 6, pp. 273–277, 1995
[56] S. Singhal, L. Wu, “Training feed-forward networks with the extended Kalman filter,” in Advances in Neural Information Processing, D. S. Tourezky,
Ed. San Mateo, CA: Morgan Kaufman, 1989, pp. 133–140
[57] I.Y. Tyukin , D.V. Prokhorov , Feasibility of random basis function approximators for modeling and control, in: Proceedings of IEEE Conference
on Control Applications & Intelligent Control, 2009, pp. 1391–1396
[58] D. Wang, “ Editorial: Randomized algorithms for training neural networks”, Information Sciences, Vol. 364-365, pp. 126-128, 2016
[59] J. Wang, P. Zhao, S. Hoi, R. Jin, “ Online Feature Selection and Its Applications”, IEEE Transactions on Knowledge and Data Engineering, Vol.
26(3), pp. 698-710, 2014
[60] N.Wang, M.J.Er, and M.X, Fast and Accurate Self Organizing Scheme for Parsimonious Fuzzy Neural Network, Neurocomputing, Vol 72, 2009
[61] S-Q. Wu, M-J. Er, Dynamic fuzzy neural networks—a novel approach to function approximation, IEEE Transaction on Systems Man Cybernetics,
part b: Cybernetics, Vol.30, pp. 358–364, 2000
[62] S.-Q. Wu, M-J. Er, Y. Gao, A fast approach for automatic generation dynamic of fuzzy rules by generalized dynamic fuzzy neural networks, IEEE
Transaction on Fuzzy System, Vol.9(4), pp. 578–594, 2003
[63] L. Xie, Y. Yang, Z. Zhou, M. Tao, Z. Man, “Dynamic neural modelling of fatigue crack growth process in ductile alloys”, Information Sciences,
Vol.364-365, pp.167-183, 2016
[64] S. Xiong, J. Azimi, X-Z. Fern,” Active Learning of Constraints for Semi-Supervised Clustering”, IEEE Transactions on Knowledge and Data
Engineering, Vol.26(1), pp. 43-54, 2014
[65] R-F. Xu, S-J. Lee, “ Dimensionality Reduction by Feature Clustering for Regression Problems”, Information Sciences, Vol. 299, pp. 42-57, 2015
[66] L. Zhang, P.N. Suganthan, “ A Survey of Randomized Algorithms for Training Neural Networks”, Information Sciences, Vol.364-365, pp. 146-155,
2016
[67] L. Zhang, P. N. Suganthan, “ A comprehensive evaluation of random vector functional link networks”, Information Sciences, 2016, DOI:
http://dx.doi.org/10.1016/j.ins.2015.09.025
[68] J. Zhao, Z. Wang, F. Cao, D. Wang, “ A local learning algorithm for random weights networks”, Knowledge-based Systems, Vol.74, pp. 159-166,
2015
| 9cs.NE
|
arXiv:1507.06621v2 [math.AG] 16 Jan 2016
CHARACTERS OF EQUIVARIANT D-MODULES ON SPACES OF MATRICES
CLAUDIU RAICU
Dedicated to Jerzy Weyman, on the occasion of his 60th birthday
Abstract. We compute the characters of the simple GL-equivariant holonomic D-modules on the vector
spaces of general, symmetric and skew-symmetric matrices. We realize some of these D-modules explicitly
as subquotients in the pole order filtration associated to the determinant/Pfaffian of a generic matrix, and
others as local cohomology modules. We give a direct proof of a conjecture of Levasseur in the case of general
and skew-symmetric matrices, and provide counterexamples in the case of symmetric matrices. The character
calculations are used in subsequent work with Weyman to describe the D-module composition factors of local
cohomology modules with determinantal and Pfaffian support.
1. Introduction
When G is an algebraic group acting on a smooth algebraic variety X over C, it is a natural problem to
describe the simple G-equivariant holonomic D-modules on X. When G acts with finitely many orbits, all
such D-modules have regular singularities, and they are classified via the Riemann–Hilbert correspondence
by the G-equivariant simple local systems on the orbits of the group action. Describing these D-modules
explicitly is however a difficult problem (see Open Problem 3 in [MV86, Section 6], and [Vil85]). In this
paper we consider the case when X is a vector space of matrices (general, symmetric, or skew-symmetric),
and G is a natural rank preserving group of symmetries. In all these cases G is a reductive group and the
D-modules are G-admissible representations (they decompose into a direct sum of irreducible representations,
each appearing with finite multiplicity). The purpose of this paper is to describe these representations (which
we will refer to as the characters of the equivariant D-modules) and to realize these D-modules explicitly.
The motivation for this work is two-fold:
• Computing local cohomology. In [RW14, RWW14, RW15] we describe the characters, and the
D-module composition factors of the local cohomology modules HY• (X, OX ) in the case when X
is a space of matrices (general, symmetric, or skew-symmetric), and Y is any orbit closure for the
natural group action on X. We expect that the combination of D-module and commutative algebra
techniques that we employ to study local cohomology in the case of matrices will apply to other
cases of interest [Lev09, Appendix]. We note that character calculations in the context of analyzing
local cohomology modules appear also in [Kem78, VdB99]: in both cases, the representations are
T -admissible for T a maximal torus in G; the equivariant D-modules that we study in this paper are
G-admissible, but in general they are too large to be T -admissible.
• Levasseur’s conjecture. For a class of multiplicity-free G-representations X, Levasseur conjectured
[Lev09, Conjecture 5.17] an equivalence between the category C of equivariant holonomic D-modules
whose characteristic variety is a union of conormal varieties to the orbits of the group action, and a
module category admitting a nice quiver description. His formulation is equivalent to the fact that
Date: January 19, 2016.
2010 Mathematics Subject Classification. Primary 14F10, 13D45, 14M12.
Key words and phrases. Equivariant D-modules, rank stratifications, local cohomology.
1
2
CLAUDIU RAICU
any simple D-module M in C contains sections which are invariant under the action of the derived
subgroup G′ = [G, G]. Our character description provides a direct proof of this conjecture for general
and skew-symmetric matrices, and yields counterexamples for symmetric matrices.
Our work complements the existing literature that studies the categories of D-modules on rank stratifications
[Nan08, Nan12] (see also [BG99] for the corresponding categories of perverse sheaves), in that we realize
concretely the simple objects of these categories and discuss some applications, filling some gaps in the
arguments and generally painting a more transparent picture. To give a flavor of the level of concreteness
that we seek, we begin with the following (Zndom denotes the set of dominant weights λ = (λ1 ≥ · · · ≥ λn ) ∈ Zn ,
and Sλ denotes the Schur functor associated to λ; throughout the paper we use the convention λs = ∞ for
s ≤ 0, λs = −∞ for s > n):
Theorem 1.1. Let X = Cn×n be the vector space of n × n matrices, and let S = C[xi,j ] be the coordinate
ring of X. If we write det = det(xi,j ), and let Sdet be the localization of S at det, then we have a filtration
0 ( S ( hdet−1 iD ( · · · ( hdet−n iD = Sdet ,
where Fs = hdet−s iD denotes the D-submodule of Sdet generated by det−s for s = 0, · · · , n (and F−1 = 0).
The successive quotients As = Fs /Fs−1 , s = 0, · · · , n are the simple GLn (C)×GLn (C)-equivariant holonomic
D-modules on X (for the natural action by row and column operations) and their characters are given by
M
As =
Sλ C n ⊗ Sλ C n .
λ∈Zn
dom
λs ≥s≥λs+1
In the case of symmetric matrices, the D-modules obtained as in Theorem 1.1 cover roughly half of the
simple equivariant D-modules. The remaining half are more mysterious, and they provide counterexamples
to [Lev09, Conjecture 5.17]. In the case of m×n matrices with m > n, as well as in the case of skew-symmetric
matrices of odd size, the simple equivariant D-modules arise as local cohomology modules, while in the case
of skew-symmetric matrices of even size the simple equivariant D-modules arise, just as in Theorem 1.1,
from the pole order filtration associated with the Pfaffian of the generic skew-symmetric matrix. Most of
our simple D-modules have irreducible characteristic variety, but for roughly half of the ones arising from
symmetric matrices the characteristic variety has two connected components: this is deduced in Remark 1.5
as a consequence of the character information.
As suggested by Theorem 1.1, one motivation behind our investigation is that the simple D-modules are
the building blocks for many D-modules of interest that one would like to understand. More precisely,
every holonomic D-module M has finite length, i.e. it has a finite filtration (composition series) whose
successive quotients (composition factors) are simple holonomic D-modules. When G is connected and M is
G-equivariant, the composition factors are also G-equivariant [VdB99, Prop. 3.1.2]. We are mainly interested
in two types of G-equivariant holonomic D-modules:
• Local cohomology modules. If Y ⊂ X is a G-invariant subset, then the local cohomology
modules HY• (X, OX ) are G-equivariant D-modules. If Y is smooth and irreducible, and if we write
c = codimX (Y ) for the codimension of Y inside X, then HYc (X, OX ) is the unique non-vanishing
local cohomology module and it is simple. In general, for an irreducible subvariety Y ⊂ X one can
define an intersection homology D-module L(Y, X) which is simple (and it is G-equivariant when Y
is a G-subvariety), and we have an inclusion L(Y, X) ⊂ HYc (X, OX ), whose cokernel is suported on
a proper subset of Y . The case when X = Cn×n and Y is the subvariety of singular matrices is
implicitly described in Theorem 1.1: c = 1, L(Y, X) = A1 , HY1 (X, OX ) = Sdet /S, and the cokernel
HY1 (X, OX )/L(Y, X) has composition factors A2 , · · · , An . In general, the local cohomology modules
HYi (X, OX ) for i 6= c may be non-zero, but they are all supported on proper subsets of Y : it is an
CHARACTERS OF EQUIVARIANT D-MODULES ON SPACES OF MATRICES
3
interesting problem to decide their (non)vanishing, or at a more refined level to understand their
D-module composition factors.
• The D-module (generated by) f α . For a non-zero polynomial f ∈ S = C[x1 , · · · , xN ] and a
complex number α, we can define hf α iD – the (holonomic) D-module generated by f α (see [Wal15]
for a recent survey). A strict inclusion hf α+1 iD ( hf α iD implies that α is a root of the Bernstein-Sato
polynomial of f (this can happen only when α is rational and negative [Kas76]). It is an interesting
question to decide whether each root α gives rise to such a strict inclusion [Wal15, Question 2.1],
[Sai15, Question 1, Section 4]. More generally, one may be interested in the composition factors of
hf α iD . For α ∈ Z and f = det this is completely answered by Theorem 1.1. When α ∈
/ Z, hdetα iD is
a simple D-module (see the proof of Theorem 7.1). Similar conclusions are obtained when f is the
symmetric determinant, or the Pfaffian of a skew-symmetric matrix of even size.
Before stating our results in more detail, we give a simple example to illustrate how character calculations
alone can allow one to determine the D-module composition factors.
Example 1.2. Let X = CN be the N -dimensional affine space, and let G = (C∗ )N be the N -dimensional
torus. The orbits XI of the G-action are indexed by subsets I ⊂ [N ] = {1, · · · , N }, where
XI = {x ∈ CN : xi 6= 0 if and only if i ∈ I}.
The stabilizer of each XI is connected, so there is a one-to-one correspondence between orbits and simple Gequivariant holonomic D-modules DI (Theorem 2.7, Remark 2.8), given by DI = L(YI , X), where YI = XI
is the corresponding orbit closure. Since YI is an affine space of codimension N − |I|, it is in particular
N −|I|
smooth, and therefore the D-module DI is just a local cohomology module DI = HYI (X, OX ). If we
write S = C[x1 , · · · , xN ] for the coordinate ring of X, then each YI is defined by the ideal generated by the
variables xj , j ∈
/ I. Using the Čech complex description of local cohomology we get
M
DI =
C · xλ1 1 · · · xλNN
λ∈ZN
λi ≥0 if and only if i∈I
which is a decomposition into irreducible G-representations. If we take f = x1 · · · xN then we get
M
Sf =
C · xλ1 1 · · · xλNN .
λ∈ZN
The torus weights appearing in the DI ’s form a partition of those appearing in Sf , so each DI appears as
a D-module composition factor of Sf with multiplicity one. Using a similar argument for X = Cn×n we
obtain a proof of Theorem 1.1 (see Section 5).
Symmetric matrices. Our results run in parallel for the three spaces of matrices (general, symmetric, and
skew-symmetric). We have made the effort to apply a uniform strategy to all three cases, but we weren’t
able to treat the combinatorial details uniformly. For the sake of brevity, we have chosen to treat only the
case of symmetric matrices in full detail, and only indicate the changes that are required in the other cases.
Two features that make the case of symmetric matrices more interesting are: (a) the presence of non-trivial
equivariant local systems; (b) the existence of counterexamples to Levasseur’s conjecture.
For each positive integer n and for s = 0, · · · , n, we consider the collections of dominant weights
C 1 (s, n) = {λ ∈ Zndom : λi
(
C 2 (s, n) =
λ ∈ Zndom
(mod 2)
≡
s + 1 for i = 1, · · · , n, λs ≥ s + 1 ≥ λs+2 },
(
)
(mod 2) s + 1 for i = 1, · · · , s
: λi ≡
, λs ≥ s + 1, λs+1 ≤ s .
s
for i = s + 1, · · · , n
(1.1)
4
CLAUDIU RAICU
Note that C 1 (n, n) = C 2 (n, n). For a positive integer n, we identify Sym2 Cn with the vector space M symm
of n × n symmetric matrices, where the squares w2 , w ∈ Cn , correspond to matrices of rank at most one.
We write Misymm for the subvariety of matrices of rank at most i. For s = 0, · · · , n, and j = 1, 2, we define
M
Cjs =
Sλ C n .
λ∈C j (s,n)
Theorem on Equivariant D-modules on Symmetric Matrices (Section 4). There exist (2n+1) simple
GLn (C)-equivariant holonomic D-modules on M symm , whose characters are Cjs , s = 0, · · · , n, j = 1, 2. More
precisely, if we denote by Csj the D-module with character Cjs then Cn1 = Cn2 = L({0}, M symm ) is the simple
holonomic D-module supported at the origin, and for s < n
(
symm
L(Mn−s
, M symm )
if j ≡ s (mod 2),
j
Cs =
symm
symm
; 1/2) if j ≡ s + 1 (mod 2).
L(Mn−s , M
symm
symm
, M symm ; 1/2) is the
, M symm ) is the usual intersection homology D-module, while L(Mn−s
Here L(Mn−s
intersection homology D-module associated to the non-trivial irreducible GLn (C)-equivariant local system on
the orbit of rank (n − s) matrices.
We let S = C[xi,j ] be the coordinate ring of M symm , where xi,j = xj,i . We write sdet = det(xi,j ) for the
determinant of the generic symmetric matrix, and let Ssdet be the localization of S at sdet. We consider
Fs = hsdet−s/2 iD , the D-submodule of Ssdet (or of Ssdet · sdet1/2 ) generated by sdet−s/2 for s = 0, · · · , n + 1
(and F−1 = 0). We have that C02 = F0 = S, and Cs1 = Fs+1 /Fs−1 for s = 0, · · · , n.
Remark 1.3. The D-modules Cs2 for s = 1, · · · , n − 1 contain no SLn (C)-invariant sections, so they provide
counterexamples to [Lev09, Conjecture 5.17]. It may be interesting to note that when n ≥ 3, among these
symm
counterexamples there are the intersection homology D-modules L(Mn−s
, M symm ) with s even, so the
failure of Levasseur’s conjecture can’t be solely explained by the presence of non-trivial local systems!
Remark 1.4. We can now give a quick derivation for the Bernstein-Sato polynomial of sdet [Kim03, Appendix]:
n
Y
1+i
.
(1.2)
bsdet (s) =
s+
2
i=1
Q
It follows from Cayley’s identity that bsdet (s) divides ni=1 s + 1+i
2 , while for each i = 1, · · · , n the strict
inclusion Fi−1 ( Fi+1 shows that − 1+i
2 is a root of bsdet (s). This is enough to conclude the equality (1.2).
Remark 1.5. It is interesting to note that the character calculation allows us to determine the characteristic
varieties for the D-modules Csj . The Fourier transform F (see Section 2.5) permutes the D-modules Csj , and
“rotates” their characteristic varieties by 90◦ (note that “rotating” the conormal variety to the orbit of rank
V
s matrices yields the conormal variety to rank (n − s) matrices). The formula (2.29) where U = 2 Cn ,
2
1
for s = 0, · · · , n.
for s = 0, · · · , n − 1, and F(Cs2 ) = Cn−s
together with (1.1), shows that F(Cs1 ) = Cn−s−1
symm
symm
1
1
Since Cs has support Mn−s and F(Cs ) has support Ms+1 , it follows that the characteristic variety of Cs1
has two components, namely the conormal varieties to the orbits of rank (n−s) and rank (n−s−1) matrices.
symm
Since Cs2 has support Mn−s
and F(Cs2 ) has support Mssymm , it follows that the characteristic variety of Cs2
is irreducible, namely it is the conormal variety to the orbit of rank (n − s) matrices. Similar considerations
show that for general and skew-symmetric matrices, the characteristic varieties of the simple equivariant
D-modules are irreducible. The calculation of characteristic varieties can also be deduced from [BG99].
CHARACTERS OF EQUIVARIANT D-MODULES ON SPACES OF MATRICES
5
Strategy for computing the characters of equivariant D-modules. Our approach to computing characters of equivariant D-modules is based on performing Euler characteristic calculations using the D-module
functoriality together with some combinatorial and
R geometric methods. More precisely, for the inclusion of
an orbit ι : O ֒→ X, the D-module direct image ι OO is an object in the derived category of G-equivariant
Rj
DX -modules, whose cohomology groups ι OO are (in the cases that we study) G-admissible representations.
Analyzing the inclusion ι directly is complicated, so we make use of a resolution of singularities Z of the
orbit closure O. The variety Z is a vector bundle over a Grassmannian G (or a product of Grassmannians),
and the inclusion j : O ֒→ Z is an affine open immersion. The map π : Z → X factors as p ◦ s
O
O
j
s /
/ Z
X ×G
❋❋
❋❋
❋❋ p
π ❋❋❋
#
/ O
/X
where sRis a regular embedding and p is the projection onto the first factor. We compute the Euler characteristic of ι OO as a virtual admissible G-representation, by using the factorization ι = p ◦ s ◦ j. If we pretend
that there is a one-to-one correspondence between simple equivariant D-modules and orbits (which is true
for general and skew-symmetric matrices), and write Xs for the D-module corresponding to matrices of rank
s, then the Euler characteristic calculations together with general considerations regarding the structure of
D-module direct images, allow us to write down an upper-triangular matrix with ones on the diagonal, that
represents the change of coordinates in the Grothendieck group of admissible representations, from (Xs )s to
appropriately defined linearly independent characters (Xs )s . The Fourier transform on one hand preserves
this matrix, and on the other hand it makes it lower-triangular, which allows us to conclude that the matrix
is in fact the identity and therefore Xs = Xs for all s (see Section 2.6).
In the process of computing Euler characteristics, we are led to the following combinatorial problem. Let
X = G(k, Cn ) be the Grassmannian of k-dimensional quotients of Cn , with OX (1) denoting the Plücker line
bundle, and ΩiX denoting the sheaf of differential i-forms on X, and define the virtual GLn (C)-representation
k·(n−k)
pk,r =
X
(−1)i · χ(X, ΩiX (r)).
i=0
Cn
The problem is to compute pk,r ⊗ Sλ
for λ ∈ Zndom . When k = 1, p1,r corresponds to the r-th power sum
symmetric function, and the answer is given in [Mac95, Exercise I.3.11(1)]. The relevance of this formula for
computing Euler characteristics is as follows: if we write E = L({0}, X) for the simple holonomic D-module
supported at the origin, Ok for the orbit of rank k matrices, and ιk for the inclusion of Ok into the ambient
space, then (up to minor adjustments, depending on which space of matrices we analyze)
X
Z
Z j
OOk =
OOk = lim pk,r ⊗ E,
(−1)j
χ
ιk
j∈Z
ιk
r→∞
where the limit is taken in the Grothendieck group of admissible representations (see Section 2.1.1 for a
precise formulation, and Section 3 for the calculations).
Organization. In Section 2 we establish the notation and basic results concerning the representation theory
of general linear groups and D-modules that will be used throughout the rest of the paper. In Section 3 we
compute the relevant Euler characteristics as limits in the Grothendieck group of GL-admissible representations. In Sections 4, 5, and 6 we prove the main results on characters of equivariant D-modules. Finally,
6
CLAUDIU RAICU
in Section 7 we discuss the simple D-modules that arise from non-equivariant local systems on the orbits,
and prove Levasseur’s conjecture for skew-symmetric and general matrices.
2. Preliminaries
2.1. Representation Theory [Wey03, Ch. 2]. Let W be a complex vector space of dimension dim(W ) = n,
and denote by GL(W ) the group of invertible linear transformations of W . The irreducible finite dimensional
GL(W )-representations, denoted Sλ W , are indexed by dominant weights λ = (λ1 ≥ · · · ≥ λn ) ∈ Zn . A
dominant weight λ is said to be a partition if all its parts λ1 , · · · , λn are nonnegative. The size of λ is
|λ| = λ1 + · · · + λn . The conjugate partition λ′ is defined by transposing the associated Young diagram: λ′i
is the number of j’s for which λj ≥ i; for example (5, 2, 1)′ = (3, 2, 1, 1, 1). Write [n] for the set {1, · · · , n},
and for a given a subset I ⊂ [n] and an integer u, let (uI ) be the sequence µ ∈ Zn having µi = u when
i ∈ I, and µi = 0 when i ∈
/ I. When I = [k] for k ≤ n, we simply write (uk ) instead of (uI ). We have that
V
Vk
W is the k-th exterior power of W , and we let det(W ) denote the top exterior power n W .
S(1k ) W =
2.1.1. Admissible representations. Given a reductive algebraic group G, we write Λ for the set of (isomorphism classes of) finite dimensional irreducible G-representations. We will be mainly interested in the case
when G = GL(W ) is a general linear group: we write Γ(G) = Γ(W ), and Λ = {Sλ W : λ ∈ Zndom }. We
also consider G = GL(W1 ) × GL(W2 ), dim(W1 ) = m, dim(W2 ) = n, and write Γ(G) = Γ(W1 , W2 ) and
n
Λ = {Sδ W1 ⊗ Sλ W2 : δ ∈ Zm
dom , λ ∈ Zdom }. An admissible G-representation decomposes as
M
M=
L⊕aL ,
L∈Λ
where each aL ∈ Z≥0 . We say that M is finite if only finitely many of the aL ’s are non-zero. We define the
Grothendieck group Γ(G) of admissible representations to be ZΛ , the direct product of copies of Z, indexed by
the set Λ. We call the elements of Γ(G) virtual representations. We write a typical element U ∈ Γ(G) as
X
aL · L,
U=
L∈Λ
where aL ∈ Z and define hU, Li = aL to be the multiplicity of L inside U . A sequence (Ur )r of virtual
representations is said to be convergent (in Γ(G)) if for every L ∈ Λ, the sequence of integers hUr , Li is
eventually
constant. If (Ur )r is convergent, we write aL = limr→∞ hUr , Li for each L ∈ Λ. We define
P
U = L∈Λ aL · L to be the limit of (Ur )r , and write
lim Ur = U.
r→∞
2.1.2. Combinatorics of weights. It will be convenient to make sense of Sλ W even when λ ∈ Zn is not
dominant. In order to do so we let δ = (n − 1, n − 2, · · · , 1, 0) and consider λ + δ = (λ1 + n − 1, λ2 + n −
2, · · · , λn−1 + 1, λn ). We write sort(λ + δ) for the sequence obtained by rearranging the entries of λ + δ
in non-increasing order. If λ + δ has non-repeated entries, we let sgn(λ) denote the sign of the unique
permutation realizing the sorting of the sequence λ + δ. We define
λ̃ = sort(λ + δ) − δ,
and let Sλ W be the element of Γ(W ) defined by
(
sgn(λ) · Sλ̃ W if λ̃ is dominant (i.e. if λ + δ has non-repeated entries);
Sλ W =
0
otherwise.
(2.1)
CHARACTERS OF EQUIVARIANT D-MODULES ON SPACES OF MATRICES
7
For example, we have S(2,1,4,3) W = 0 and S(1,1,0,7) W = −S(4,2,2,1) W . Note that in particular
Sλ W = 0 if λi+1 = λi + 1 for some i = 1, · · · , n − 1.
(2.2)
[n]
We denote by k the collection of subsets I ⊂ [n] of size |I| = k, and write P (k, n − k) for the set of
partitions µ = (µ1 ≥ · · · ≥ µk ) with µ1 ≤ n − k. There is a one-to-one correspondence between sets I ∈ [n]
k
and partitions µ ∈ P (k, n − k) given by
I = {µk + 1, µk−1 + 2, · · · , µ2 + (k − 1), µ1 + k}.
If we write
µ′
(2.3)
∈ P (n − k, k) for the conjugate partition of µ then the complement of I in [n] is given by
I c = [n] \ I = {k + 1 − µ′1 , k + 2 − µ′2 , · · · , n − µ′n−k }.
(2.4)
n
c
For every λ ∈ Zn , I ∈ [n]
k and r ∈ Z, we define λ(r, I) ∈ Z as follows: we write the elements of I and I
in increasing order
I = {i1 < · · · < ik }, I c = {ik+1 < · · · < in },
(2.5)
and let
(
r + t + λit − it for t = 1, · · · , k;
λ(r, I)t =
(2.6)
t + λit − it
for t = k + 1, · · · , n.
We define λ1 (I) ∈ Zk and λ2 (I) ∈ Zn−k via
λ1 (I)t = t + λit − it , for t = 1, · · · , k,
(2.7)
λ2 (I)t−k = t + λit − it , for t = k + 1, · · · , n,
so that λ(r, I) is the concatenation of λ1 (I) + (r k ) and λ2 (I). In particular
λ1 ([k]) = (λ1 , · · · , λk ) and λ2 ([k]) = (λk+1 , · · · , λn ).
We define the permutation σ(I) of [n] via
σ(I)t = it for t = 1, · · · , n.
(2.8)
Sλ+(rI ) W = sgn(σ(I)) · Sλ(r,I) W = (−1)|µ| · Sλ(r,I) W,
(2.9)
With this notation we obtain
and note that if λ is dominant and r is sufficiently large then λ(r, I) is also dominant.
We define for h, j ∈ Z/2Z the sets of partitions
P h,j (a, b) = {µ ∈ P (a, b) : µi ≡ h (mod 2) for i = 1, · · · , a, µ′i ≡ j (mod 2) for i = 1, · · · , b}.
A quick counting argument yields the following:
Lemma 2.1. The cardinality of P h,j (a, b) is computed by:
|P 0,0 (a, b)| =
⌊ a2 ⌋
+
⌊ 2b ⌋
⌊ 2b ⌋
b−1
a
2 +⌊ 2 ⌋
|P 1,0 (a, b)| =
⌊ b−1
2 ⌋
0
,
|P 0,1 (a, b)| =
a even
a odd
a−1
b
⌊ 2 ⌋+ 2
0
,
|P 1,1 (a, b)| =
b
2
,
b odd
b−1
2
a, b odd
b−1
a−1
2 +
0
b even
2
otherwise
.
(2.10)
8
CLAUDIU RAICU
2.1.3. A generalized Pieri rule. The Grothendieck group Γ(W ) is a module over the representation ring R(W )
V
of finite dimensional GL(W )-representations.
As a ring, R(W ) is generated by the exterior powers k W ,
V
k ≤ n, and by the inverse det(W )−1 = n W ∗ = S(−1n ) W of det(W ). We have Sλ W ⊗ det(W ) = Sλ+(1n ) W .
V
The following lemma generalizes this by describing the multiplicative action of the exterior powers k W on
Γ(W ) (since the multiplication is continuous, i.e. it commutes with limits, it suffices to determine its action
on the indecomposables Sλ W ):
Lemma 2.2 (Pieri’s rule). For every λ ∈ Zn we have the following equality in Γ(W ):
!
k
^
X
W ⊗ Sλ W =
Sλ+(1I ) W.
(2.11)
I∈([n]
k )
Proof. We may assume without loss of generality that λ is dominant. If λi+1 = λi and I is such that i ∈
/I
and (i + 1) ∈ I then it follows from (2.2) that Sλ+(1I ) W = 0. For all the other terms appearing on the right
hand side of (2.11) we have that µ = λ + (1I ) is dominant and µ/λ is a vertical strip (i.e. µi − λi ∈ {0, 1} for
all i) of size k. (2.11) then follows from the usual Pieri formula [Wey03, Corollary 2.3.5].
We define elements pk,r (W ) ∈ R(W ) for every r ∈ Z and 0 ≤ k ≤ n, by
X
pk,r (W ) =
S(rI ) W,
(2.12)
[n]
k
I∈(
and note that pk,1 (W ) =
Vk
)
W . We have the following generalization of Pieri’s rule:
Lemma 2.3. For every λ ∈ Zn we have the following equality in Γ(W ):
X
pk,r (W ) ⊗ Sλ W =
Sλ+(rI ) W.
[n]
k
I∈(
(2.13)
)
Proof. When k = 0, p0,r (W ) = C is the identity element of R(W ), so the conclusion is trivial. We may thus
assume that k > 0. As before, we also assume that λ is dominant. Multiplication by det(W ) is an invertible
operation, so proving (2.13) for λ is equivalent to proving it for λ + (1n ). In particular, we may assume that
λ is a partition and that moreover λn = 0.
We consider the ordering of the partitions λ with at most n parts induced by the graded reverse lexicographic order on their conjugates: more precisely, we say that λ ≻ µ if |λ| > |µ|, or if |λ| = |µ| and for the
largest index i for which λ′i 6= µ′i one has λ′i > µ′i . We prove (2.13) for all partitions λ, by induction with
respect to the said ordering. When λ is the empty partition, (2.13) coincides with (2.12).
Assume now that λ1 > 0 and consider the parititon µ obtained from λ by removing the last column of its
Young diagram: the conjugate µ′ is given by µ′i = λ′i for i < λ1 and µ′i = 0 for i ≥ λ1 . We let l = λ′λ1 denote
the size of the column removed from λ. Using the induction hypothesis for µ ≺ λ and Lemma 2.2 we get
!
l
^
X X
W =
(pk,r (W ) ⊗ Sµ W ) ⊗
Sµ+(rI )+(1J ) W .
(2.14)
J∈([n]
l )
I∈([n]
k )
Consider the collection of partitions P = {α : α/µ is a vertical strip of size l}, so that
!
l
^
X
W =
Sα W,
Sµ W ⊗
α∈P
CHARACTERS OF EQUIVARIANT D-MODULES ON SPACES OF MATRICES
9
and note that λ ∈ P and that α ≺ λ for every λ 6= α ∈ P. We can then rewrite the left hand side of (2.14)
as
X
pk,r (W ) ⊗ Sα W,
α∈P
so in order to prove (2.13) for λ it is sufficient to show that the right hand side of (2.14) is equal to
X X
Sα+(rI ) W .
α∈P
Since P = {µ + (1J ) : J ∈
dominant then
[n]
l ,µ
I∈([n]
k )
+ (1J ) dominant}, we only have to check that when α = µ + (1J ) is not
X
Sα+(rI ) W = 0.
(2.15)
I∈([n]
k )
Note that the only way in which α = µ + (1J ) can fail to be dominant is if for some index j, µj = µj+1 and
j∈
/ J, (j + 1) ∈ J. Fix such an index j, and note that αj+1 = αj + 1. It follows from (2.2) that when I ⊂ [n]
is such that both j, j + 1 ∈ I, or both j, j + 1 ∈
/ I, then Sα+(rI ) W = 0. To show (2.15) it is then enough to
prove that
X
X
Sα+(rI ) W +
Sα+(rI ′ ) W = 0.
(2.16)
I∈([n]
k )
j∈I,(j+1)∈I
/
I ′ ∈([n]
k )
j ∈I
/ ′ ,(j+1)∈I ′
There is a one-to-one correspondence between the collection of subsets I with j ∈ I, (j + 1) ∈
/ I, and subsets
I ′ with j ∈
/ I ′ , (j + 1) ∈ I ′ , given by I ′ = (I ∪ {j + 1}) \ {j}. Moreover, for such a pair I, I ′ it follows from
′
(2.1) that Sα+(rI ) W = −Sα+(rI ′ ) W (because α + (r I ) + δ is obtained from α + (r I ) + δ by switching the
j-th part with the (j + 1)-st part), which proves (2.16) and concludes the proof of the lemma.
2.2. Bott’s theorem for Grassmannians [Wey03, Ch. 4]. We consider X = G(k, V ), the Grassmannian
of k-dimensional quotients of V (or k-dimensional subspaces of W = V ∗ ), with the tautological sequence
0 −→ R −→ V ⊗ OX −→ Q −→ 0,
(2.17)
where Q is the tautological rank k quotient bundle, and R is the tautological rank (n − k) sub-bundle.
Bott’s Theorem for Grassmannians [Wey03, Corollary 4.1.9] computes the cohomology of a large class of
GL-equivariant bundles on X. We only need a weaker version that computes Euler characteristics.
Suppose that M is a quasi-coherent GL(W )-equivariant sheaf on X. We say that M has admissible (resp.
finite) cohomology if its cohomology groups H j (X, M) are admissible (resp. finite) for j = 0, · · · , dim(X).
We can therefore make sense of the Euler characteristic of M as an element of Γ(W ) (resp. R(W )). We
define the Euler characteristic of M to be the virtual representation
k·(n−k)
χ(X, M) =
X
(−1)j H j (X, M).
(2.18)
j=0
n−k
Theorem 2.4 (Bott). Let α ∈ Zkdom and β ∈ Zdom
be dominant weights, and let λ = (α, β) ∈ Zn be their
concatenation. The Euler characteristic of Sα Q ⊗ Sβ R is given (with the convention (2.1)) by
χ(X, Sα Q ⊗ Sβ R) = Sλ V.
We can now give an alternative interpretation of the elements pk,r introduced in (2.12):
10
CLAUDIU RAICU
V
Lemma 2.5. If we let ΩiX = i (R ⊗ Q∗ ) denote the sheaf of i-differential forms on X, and write OX (1) =
det(Q) for the Plücker line bundle on X, then
k·(n−k)
pk,r (V ) =
X
(−1)i · χ(X, ΩiX (r)).
i=0
Proof. Cauchy’s formula [Wey03, Cor. 2.3.3] yields
i
^
(R ⊗ Q∗ ) =
M
Sµ Q∗ ⊗ Sµ′ R.
µ∈P (k,n−k), |µ|=i
Twisting by OX (r) = det(Q)⊗r = S(rk ) Q, and taking Euler characteristics, we get using Theorem 2.4
X
S(r−µk ,r−µk−1 ,··· ,r−µ1 ,µ′1 ,··· ,µ′n−k ) V.
χ(X, ΩiX (r)) =
µ∈P (k,n−k), |µ|=i
Using (2.9) with λ = 0, we get S(r−µk ,r−µk−1 ,··· ,r−µ1 ,µ′1 ,··· ,µ′n−k ) V = (−1)|µ| · S(rI ) V , so
k·(n−k)
X
i=0
(−1)i · χ(X, ΩiX (r)) =
X
[n]
k
I∈(
S(rI ) V = pk,r (V ).
)
2.3. D-modules [BGK+ 87], [HTT08]. For a smooth algebraic variety X over C, we let DX denote the
sheaf of differential operators on X [HTT08, Section 1.1]. A D-module M on X (or a DX -module) is a
quasi-coherent sheaf M on X, with a left module action of DX .
Definition 2.6. Let G be an algebraic group acting on X, and let M be a DX -module. Differentiating the
action of G on X yields a map d : Lie(G) → DerX from the Lie algebra of G to the vector fields on X. The
DX -module operation
DX ⊗ M → M,
(2.19)
composed with d yields an action of Lie(G) on M. The DX -module M is G-equivariant if
(a) M admits an action of G compatible with (2.19) (see [HTT08, Def. 11.5.2] for a precise meaning of
compatibility).
(b) The action of Lie(G) on M obtained by differentiating the action of G on M coincides with the one
induced from d : Lie(G) → DerX and (2.19).
As discussed in the Introduction, examples of G-equivariant holonomic DX -modules are OX , and for a
G-invariant subset Y ⊂ X, the local cohomology modules HY• (X, OX ), as well as the intersection homology
D-modules L(Y, X). When X = U is a vector space, and Y = {0} is the origin, we let
dim(U )
E = L({0}, U ) = H{0}
(U, OU ),
(2.20)
be the unique simple DU -module supported at the origin. As a vector space (and a G-representation)
E = det(U ) ⊗ Sym(U ).
(2.21)
The following theorem gives a classification of the simple equivariant holonomic D-modules, for a group
action with finitely many orbits (see [HTT08, Section 11.6]):
Theorem 2.7. Let G be an algebraic group acting with finitely many orbits on a smooth algebraic variety X.
There is a one-to-one correspondence between:
CHARACTERS OF EQUIVARIANT D-MODULES ON SPACES OF MATRICES
11
(a) Simple G-equivariant holonomic DX -modules.
(b) Pairs (O, L) where O is a G-orbit, and L is an irreducible G-equivariant local system on O.
(b’) Pairs (O, L) where O is a G-orbit, and L is an irreducible representation of the component group of
the isotropy group of O.
Here by the isotropy group of O we mean the stabilizer of any element in O (they are all isomorphic). For an
algebraic group H, we denote by H0 the connected component of the identity, which is a normal subgroup
of H. The quotient H/H0 is called the component group of H.
Remark 2.8. When the representation L in Theorem 2.7(b’) is trivial, the corresponding DX -module in part
(a) is L(O, X), where O is the closure of O. It follows that in the case when the isotropy groups for the Gaction on X are connected, there is a one-to-one correspondence between simple G-equivariant DX -modules
and orbits of the group action.
Let m ≥ n be positive integers and consider the complex vector spaces M of general m × n matrices,
M symm of n × n symmetric, and M skew of n × n skew-symmetric matrices respectively. These spaces admit
a natural action of a group GL via row and column operations: GLm (C) × GLn (C) acts on M , and GLn (C)
acts M symm and M skew . We write Ms (resp. Mssymm ) for the subvariety of M (resp. M symm ) consisting
of matrices of rank at most s, for s = 0, · · · , n, and Msskew for the subvariety of M skew consisting of skewsymmetric matrices of rank at most 2s, for s = 0, · · · , ⌊n/2⌋. We have the following:
Theorem 2.9 (Classification of simple GL-equivariant holonomic D-modules on spaces of matrices).
• (General matrices). There are (n + 1) simple GL-equivariant D-modules on the vector space M of
m × n matrices, namely the intersection homology D-modules L(Ms , M ), s = 0, · · · , n.
• (Symmetric matrices). There are (2n + 1) simple GL-equivariant D-modules on the vector space
M symm of n × n symmetric matrices, (n + 1) of which are the intersection homology D-modules
L(Mssymm , M symm ), s = 0, · · · , n, while the remaining ones are the intersection homology D-modules
L(Mssymm , M symm ; 1/2), s = 1, · · · , n, corresponding to the non-trivial irreducible equivariant local
systems on the orbits.
• (Skew-symmetric matrices). There are (⌊n/2⌋ + 1) simple GL-equivariant D-modules on the vector
space M skew of n × n skew-symmetric matrices, namely L(Msskew , M skew ), s = 0, · · · , ⌊n/2⌋.
Proof. The theorem follows from Theorem 2.7 and Remark 2.8, since the isotropy groups for general and
skew-symmetric matrices are connected, while for symmetric matrices the isotropy groups of the non-zero
orbits have two connected components.
2.4. Computing Euler characteristics. Let X be a smooth complex projective algebraic variety and
denote its dimension by dX . Consider a finite dimensional vector space U , and a short exact sequence
0 −→ ξ −→ U ⊗ OX −→ η −→ 0,
(2.22)
U∗
where ξ, η are locally free sheaves on X. We think of
as an affine space, and of U as linear forms on U ∗ .
∗
We let Y = TotX (η ) denote the total space of the bundle η ∗ , and define a morphism π : Y → U ∗ via the
commutative diagram
/ U∗ × X
(2.23)
Y = TotX (η ∗ )
PPP
PPP
P
π PPPP
PP'
∗
U
where the top map is the inclusion of into the trivial bundle U ∗ , and the vertical map is the projection onto
the U ∗ factor. We will be interested in understanding the (Euler characteristic of the) D-module pushforward
η∗
12
R
CLAUDIU RAICU
M along the map π for certain DY -modules M. For affine morphisms X ′ → X, we will identify freely
quasi-coherent sheaves on X ′ with quasi-coherent OX ′ -modules on X as in [Har77, Exercise II.5.17(e)].
We let S = SymOX (η), so that Y = SpecX (S), and consider a locally free sheaf L of rank one with
L ⊂ Symi (η) for some i > 0. We pull-back L to Y , define L = TotY (L∗ ) to be the total space of the line
bundle L∗ , and write p : L → Y for the natural map. The inclusion L ⊂ Symi (η) defines a section z : Y → L
of p [Har77, Exercise II.5.18(c)], and we define Z to be the zero-locus of z. If X = Spec(C) then Y is an
affine space, S is the ring of polynomial functions on Y , L corresponds to (the vector space spanned by) a
polynomial f ∈ S of degree i, and Z is the vanishing locus of f .
0
0
We consider the
R complement Y = Y \ Z and let j : Y → Y denote the inclusion. Since j is an affine
open immersion, j OY 0 = OY 0 can be thought of as a quasi-coherent sheaf of algebras on Y (or on X):
π
OY 0 = lim L−r ⊗ OY = lim L−r ⊗ S.
−→
−→
r
r
In the case when X = Spec(C), we have OY 0 = Sf is the localization of S at f , which is a D-module on the
affine space Y . We define the quasi-coherent sheaf S ∨ on X (the graded dual of S) by
S ∨ = det(η ∗ ) ⊗ SymOX (η ∗ ).
Proposition 2.10. With the notation above, we assume that X admits an action of a reductive group
G, that U is a finite dimensional G-representation, and that ξ, η, L are G-equivariant locally free sheaves.
Assume further that we have an isomorphism of G-equivariant quasi-coherent sheaves on X
OY 0 ≃ lim Lr ⊗ S ∨ .
−→
(2.24)
r
Let M be a DY -module which is isomorphic, as a quasi-coherent G-equivariant sheaf on X, to OY 0 ⊗OX L′ ,
with L′ a line bundle on X. We denote by ΩiX the sheaf of i-differential forms on X, and assume that for
every i = 0, · · · , dX the sheaves ΩiX ⊗ M ⊗ det(ξ ∗ ) ⊗ SymOX (ξ ∗ ) have G-admissible cohomology. If we define
the sequence Pr (X, L; L′ ) ∈ Γ(G) via
Pr (X, L; L′ ) =
dX
X
(−1)dX −i · χ(X, Lr ⊗ L′ ⊗ ΩiX ),
i=0
then
χ
Z
π
M = lim Pr (X, L; L′ ) ⊗ det(U ∗ ) ⊗ SymC (U ∗ ).
r→∞
(2.25)
Remark 2.11. We will apply Proposition 2.10 in the case when X = G(k, V ) is a Grassmann variety, and
L = OX (1) is the Plücker line bundle (or its square). It follows from Lemma 2.5 that
Pr (X, OX (1); OX ) = (−1)k(n−k) · pk,r (V ).
It follows that if X = G(k, V1 ) × G(k, V2 ), where dim(V1 ) = m, dim(V2 ) = n, and if L = OX (1, 1) then
Pr (X, OX (1); OX ) = (−1)k·(m−n) · pk,r (V1 ) ⊗ pk,r (V2 ).
Proof of Proposition 2.10. Since the sheaves ΩiX ⊗ M ⊗ det(ξ ∗ ) ⊗ SymOX (ξ ∗ ) have admissible cohomology,
it follows from [Rai14, Corollary 2.10] that
X
Z
dX
(2.26)
M =
(−1)dX −i · χ(X, ΩiX ⊗ M ⊗ det(ξ ∗ ) ⊗ SymOX (ξ ∗ )).
χ
π
i=0
Computing Euler characteristics commutes with colimits and associated graded constructions. By (2.22) we
get a filtration of U ∗ ⊗ OX with gr(U ∗ ⊗ OX ) = ξ ∗ ⊕ η ∗ , which yields a filtration of SymOX (U ∗ ⊗ OX ) with
CHARACTERS OF EQUIVARIANT D-MODULES ON SPACES OF MATRICES
13
gr(SymOX (U ∗ ⊗ OX )) = SymOX (ξ ∗ ) ⊗ SymOX (η ∗ ). We also get that det(U ∗ ⊗ OX ) = det(ξ ∗ ) ⊗ det(η ∗ ), and
therefore
χ(X, ΩiX ⊗ L′ ⊗ Lr ⊗ S ∨ ⊗ det(ξ ∗ ) ⊗ SymOX (ξ ∗ )) = χ(X, ΩiX ⊗ L′ ⊗ Lr ⊗ det(U ∗ ) ⊗ SymOX (U ∗ ⊗ OX ))
U ∗ is a trivial bundle
=
χ(X, ΩiX ⊗ L′ ⊗ Lr ) ⊗ det(U ∗ ) ⊗ SymC (U ∗ ).
Multiplying this equality by (−1)dX −i , summing over i = 0, · · · , dX , taking the limit as r → ∞, and using
the identification (2.24) tensored with L′ , we get (2.25).
2.5. The Weyl algebra and the Fourier transform. For a positive integer N , the Weyl algebra
C[x1 , · · · , xN , ∂1 , · · · , ∂N ], ∂i =
∂
∂xi
(2.27)
is the ring of differential operators on CN . In this section we give a coordinate independent description of
the Weyl algebra, and use it to describe the Fourier transform.
Given a finite dimensional C–vector space U of dimension N , we write h, i for the natural pairing U ×U ∗ →
C. We let Ũ = U ⊕ U ∗ and define a non–degenerate skew–symmetric form ω : Ũ ⊗ Ũ → C by
′
if u ∈ U, u′ ∈ U ∗ ,
hu, u i
′
ω(u, u ) = − hu′ , ui if u′ ∈ U, u ∈ U ∗ ,
0
otherwise.
L
We write Tn (Ũ ) for the tensor product Ũ ⊗n , and let T (Ũ ) = n≥0 Tn (Ũ ) denote the tensor algebra on Ũ .
V
We have a natural inclusion 2 Ũ ⊂ T2 (Ũ ), and define the Weyl algebra DU ∗ as the quotient
D
U∗
= T (Ũ )/hx − ω(x) : x ∈
2
^
Ũ i
(2.28)
V
of the tensor algebra by the bilateral ideal generated by differences x − ω(x), with x ∈ 2 Ũ . Note that DU ∗
is the ring of differential operators on the vector space U ∗ . If we choose a basis x1 , · · · , xN of U , and the
dual basis ∂1 , · · · , ∂N of U ∗ , then DU ∗ coincides with (2.27).
Lemma 2.12 (Fourier transform). If M is a (left) DU -module, then det(U ∗ ) ⊗ M has the structure of a
(left) DU ∗ -module.
Example 2.13. The most basic example is when M = Sym(U ∗ ) is the coordinate ring of U . In that case
det(U ∗ ) ⊗ Sym(U ∗ ) is equal to E, the simple holonomic DU ∗ -module supported at the origin (see (2.20)).
Proof of Lemma 2.12. Using the identification of Ũ ∗ with Ũ coming from the natural isomorphism U ∗ ⊕ U ≃
U ⊕ U ∗ , it is easy to see that DU ∗ ≃ DUop , where op denotes the opposite ring. Since M is a left DU -module, it
is also a right DUop -module, i.e. it can be identified with a right DU ∗ -module. The canonical sheaf ωU ∗ on the
vector space U ∗ is a free rank one module generated by det(U ). By [HTT08, Prop. 1.2.12], the association
M 7→ ωU−1∗ ⊗ M = det(U ∗ ) ⊗ M gives an equivalence between the categories of right DU ∗ -modules and left
DU ∗ -modules.
Motivated by Lemma 2.12, we define a Fourier transform relative to U , denoted FU , on the Grothendieck
group Γ(G) of admissible G-representations as follows:
X
X
ai · (det(U ∗ ) ⊗ Mi∗ ).
(2.29)
ai · Mi =
FU
14
CLAUDIU RAICU
The context in which we apply the Fourier transform is as follows: we will have constructions which are
functorial in U for certain DU -modules MU which are admissible representations for some group G, in such
a way that
M ⊕a
M
MU =
Mi i if and only if MU ∗ =
(Mi∗ )⊕ai .
i
i
P
By Lemma 2.12, the Fourier transform of the DU ∗ -module MU ∗ has character equal to FU ( i ai · Mi ). We
will slightly imprecisely refer to this as the character of the Fourier transform of MU .
2.6. A little linear algebra. Consider a finite partially ordered set P, and let A denote the free abelian
group with basis {vp : p ∈ P}. We write p ≻ q to indicate that p is strictly larger than q with respect to the
partial order, and p q when we allow equality. Assume that F : P −→ P is an order reversing bijection,
i.e. p q if and only if F(q) F(p). By abuse of notation, we also write F : A −→ A for the induced
automorphism of A, given by F(vp ) = vF (p) . We have the following:
Lemma 2.14. Suppose that we have a collection of elements vp ∈ A for p ∈ P, for which there exist relations
X
(2.30)
vp = v p +
apq · vq , for some integers apq .
q≻p
If the automorphism F of A permutes the elements vp then vp = vp for all p ∈ P (and hence all apq = 0).
Proof. Write F(vp ) = vσ(p) for some permutation σ : P −→ P. Applying F to (2.30) we get
X
apq · vF (q) ,
vσ(p) = vF (p) +
q≻p
which is necessarily a permutation of the relations (2.30). Since F is order-reversing, it follows that σ(p) =
F(q) for some q p, and if σ(p) = F(p) then one also has vσ(p) = vF (p) , i.e. apq = 0 for all q ≻ p. We get
that F(p) σ(p) for all p ∈ P, and the equality F(p) = σ(p) implies vσ(p) = vF (p) . An easy induction on the
height of F(p), defined by ht(F(p)) = #{q : F(p) ≻ q}, shows that F(p) = σ(p) for all p, which concludes
the proof of the lemma.
3. Some limit calculations in the Grothendieck group of admissible representations
Recall the terminology from Sections 2.1.1–2.1.3 which we will be using freely throughout this section.
In particular recall the notation Γ(G) for the Grothendieck group of admissible G-representations for some
group G, and the definition of pk,r (V ) from (2.12) (also Lemma 2.5). When W is a vector space, we write
V = W ∗ for its dual. In this section we compute in three cases limits in Γ(G) of the type
lim pk,r ⊗ E,
r→∞
(3.1)
where (pk,r )r is a sequence of finite virtual G-representations, E = det(U ) ⊗ Sym(U ) is the (character of
the) simple DU -module supported at the origin (2.20), where U is a finite dimensional G-representation:
• U = Sym2 W , G = GL(W ) (so that Γ(G) = Γ(W )), pk,r = pk,r (V ). The limit (3.1) does not exist if
r is arbitrary, but instead we have to consider the cases when r is even resp. odd separately.
• U = W1 ⊗ W2 , G = GL(W1 ) × GL(W2 ) (so that Γ(G) = Γ(W1 , W2 )), pk,r = pk,r (V1 ) ⊗ pk,r (V2 ).
V
• U = 2 W , G = GL(W ), pk,r = pk,r (V ) with k even.
As mentioned in the Introduction and explained in Section 2.4, the limits (3.1) correspond to Euler characteristic calculations for certain D-module direct images. They are essential to the character calculations
in Sections 4–6 below. The reader who is not interested in the details of the limit calculations may wish to
record the results of Propositions 3.1, 3.5, and 3.6 below, and skip to Section 4.
CHARACTERS OF EQUIVARIANT D-MODULES ON SPACES OF MATRICES
15
3.1. Symmetric matrices. We let W be a vector space of dimension n. For s = 0, · · · , n and j = 1, 2, we
define the elements Cjs ∈ Γ(W ) via
M
Cjs =
Sλ W.
(3.2)
λ∈C j (s,n)
where C j (s, n) is defined in (1.1).
Proposition 3.1. If E = det Sym2 W ⊗ Sym Sym2 W then for k = 0, · · · , n,
n s−2
s−1
n
X
X
2
2
2
· Cs +
· C1s
n−k−2
n−k
2
2
s=n−k
s=n−k+1
!
s even
s odd
pk,r (V ) ⊗ E =
(−1)k(n−k) ·
lim
r→∞
s−1
s−2
n
n
r≡k+1 (mod 2)
X
X
1
2
2
· Cs −
· C2s
n−k−1
n−k−1
2
2
s=n−k
s odd
(−1)k(n−k) ·
lim
pk,r (V
r→∞
r≡k (mod 2)
)⊗E
!
if n − k even,
if n − k odd.
s=n−k+1
s even
=
n s
X
2
· C1s
n−k
2
s=n−k
s even
s−1
n
X
2
· C2s
n−k−1
s=n−k
2
if n − k even,
if n − k odd.
s odd
When k = 0 the above equalities are easy to verify: pk,r (V ) = C is the trivial representation, so the left
hand side reduces to E, regardless of the parity of r; the right hand side is either C1n or C2n , but E = C1n = C2n .
We therefore fix 1 ≤ k ≤ n for the rest of this section. We begin with some notation and preliminary results
before proving the proposition. For j ∈ Z/2Z we let
C j = {λ ∈ Zkdom : λi ≡ n + 1 + j (mod 2) for i = 1, · · · , k},
j
n−k
: λi ≡ n + 1 + j (mod 2) for i = 1, · · · , n − k, and λn−k ≥ n + 1}.
= {λ ∈ Zdom
C≥n+1
(3.3)
With the convention λ0 = ∞, λn+1 = −∞, we define for s = 0, · · · , n,
Z(s) = {λ ∈ Zndom : λs ≥ s + 1 ≥ λs+1 },
(3.4)
Zndom .
For h, j ∈ Z/2Z we let
and note that the sets Z(s), s = 0, · · · , n form a partition of
(
(
)
X
(mod 2) h, for i = 1, · · · , s,
h,j
C (s) = λ ∈ Z(s) : λi ≡
Sλ W.
, Ch,j
s =
j, for i = s + 1, · · · , n.
h,j
λ∈C
(3.5)
(s)
Comparing with (1.1) we get that C 1 (s, n) = C s+1,s+1 (s) ∪ C s+1,s+1 (s + 1) and C 2 (s, n) = C s+1,s (s) so
C1s = Cs+1,s+1
+ Cs+1,s+1
,
s
s+1
[n]
and C2s = Css+1,s .
0
Lemma 3.2. If I ∈ k , λ1 (I) ∈ C k+1+j and λ2 (I) ∈ C≥n+1
then
• λ ∈ Z(s) for some s = n − k, · · · , n.
• {s + 1, · · · , n} ⊂ I.
• λs+1 ≡ · · · ≡ λn ≡ j (mod 2).
(3.6)
16
CLAUDIU RAICU
Proof. Consider the unique s for which λ ∈ Z(s). Let s′ be the maximal element of I c and assume that
s′ > s. We have (using (2.5)) that in = s′ and therefore λs′ ≤ λs+1 ≤ s + 1 and
(2.7)
λ2 (I)n−k = n + λin − in = n + λs′ − s′ ≤ n + s + 1 − s′ < n + 1,
0
which contradicts λ2 (I) ∈ C≥n+1
. It follows that s′ ≤ s and hence {s+1, · · · , n} ⊂ I, which implies n−s ≤ k,
or s ≥ n − k. From (2.5) we get
it = t + n − k for t = k − n + s + 1, · · · , k,
which using the fact that λ1 (I) ∈ C k+1+j yields for t = k − n + s + 1, · · · , k
n+1+k+1+j
(mod 2)
≡
λ1 (I)t = t + λit − it = t + λt+n−k − (t + n − k) = λt+n−k + k − n
so λt+n−k ≡ j (mod 2), concluding the proof of the lemma.
Lemma 3.3. Assume that λ ∈ Z(s) and that there exists an index 1 ≤ i < s such that λi 6≡ λi+1 (mod 2).
For any j ∈ Z/2Z, consider the collection
[n]
1
k+1+j
2
0
Pλ (j) = I ∈
: λ (I) ∈ C
, λ (I) ∈ C≥n+1 .
(3.7)
k
We have (using notation (2.8))
X
sgn(σ(I)) = 0.
(3.8)
I∈Pλ (j)
Proof. We show that if I ∈ Pλ (j) then exactly one of i, i + 1 is contained in I. Moreover, we show that the
assignment I ′ = I \ {i} ∪ {i + 1} establishes a bijection between
{I ∈ Pλ (j) : i ∈ I}
and {I ′ ∈ Pλ (j) : i + 1 ∈ I ′ }.
(3.9)
sgn(σ(I ′ ))
Since
= − sgn(σ(I)), the conclusion (3.8) follows.
Assume that I is such that i, i + 1 are both in I, or both in I c . We can then find t < k or t > k such that
it = i and it+1 = i + 1. If t < k then λ1 (I)t 6≡ λ1 (I)t+1 (mod 2), contradicting λ1 (I) ∈ C k+1+j . If t > k
0
then λ2 (I)t−k 6≡ λ2 (I)t−k+1 (mod 2), contradicting λ2 (I) ∈ C≥n+1
.
c
Choose now a set I with i ∈ I, i + 1 ∈ I , and choose t0 ≤ k, t1 ≥ k + 1, such that it0 = i, it1 = i + 1. If
we let I ′ = I \ {i} ∪ {i + 1} then λ1 (I)t = λ1 (I ′ )t for t 6= t0 , and λ2 (I)t = λ2 (I ′ )t for t 6= t1 − k. We have
λ1 (I)t0 = t0 + λi − i,
λ1 (I ′ )t0 = t0 + λi+1 − (i + 1),
λ2 (I)t1 −k = t1 + λi+1 − (i + 1), λ2 (I ′ )t1 −k = t1 + λi − i,
and since λi 6≡ λi+1 (mod 2), we get λ1 (I)t ≡ λ1 (I ′ )t (mod 2) and λ2 (I)t ≡ λ2 (I ′ )t (mod 2) for all t. Since
λ2 (I)t1 −k ≤ λ2 (I ′ )t1 −k , the only way in which the correspondence I ↔ I ′ could fail to induce a bijection
0
(3.9) is if for some I, I ′ we get t1 = n and λ2 (I)n−k ≤ n < n + 1 ≤ λ2 (I ′ )n−k , in which case λ2 (I) 6∈ C≥n+1
,
2
′
0
2
but λ (I ) ∈ C≥n+1 . However, the inequality λ (I)n−k ≤ n would imply
λ2 (I)t1 −k = t1 + λit1 − it1 = n + λi+1 − (i + 1) ≤ n or equivalently λi+1 ≤ i + 1.
Since i < s by hypothesis, we get λs ≤ λi+1 ≤ (i + 1) ≤ s, contradicting the fact that λ ∈ Z(s).
Lemma 3.4. If λ ∈ C h,j (s), s ≥ n − k, then there is a one-to-one correspondence between elements Pλ (j)
and the set P n−k+j−h,s+1−h(k − n + s, n − k) (defined in (2.10)). Moreover, for every I ∈ Pλ (j) we have
sgn(σ(I)) = (−1)(n−k)·(k+h) ,
and Pλ (j) is empty if h ≡ s ≡ j + 1 (mod 2).
CHARACTERS OF EQUIVARIANT D-MODULES ON SPACES OF MATRICES
17
c
Proof. The correspondence between sets I ∈ [n]
k (resp. their complements I ) and partitions µ ∈ P (k, n−k)
(resp. their conjugates µ′ ) is given in (2.3) (resp. (2.4)). If I ∈ Pλ (j) then it follows from Lemma 3.2 that
s + 1, · · · , n are the largest elements of I, namely ik−n+s+1 , · · · , ik , so µ1 = · · · = µn−s = n − k. The set
I is then determined by µ = (µn−s+1 , · · · , µk ) ∈ P (k − n + s, n − k). Since λ ∈ C h,j (s), the condition
0
λ1 (I) ∈ C k+1+j is equivalent to µi ≡ n − k + j − h (mod 2). The condition λ2 (I) ∈ C≥n+1
is equivalent to
′
′
µi ≡ n + 1 − h (mod 2), which in turn is equivalent to µi ≡ s + 1 − h (mod 2). It follows that I ∈ Pλ (j) if
and only if µ ∈ P n−k+j−h,s+1−h(k − n + s, n − k), which establishes the desired bijection. Moreover
′
sgn(σ(I)) = (−1)|µ| = (−1)|µ | = (−1)(n−k)·(n+1−h) = (−1)(n−k)·(k+h) ,
where the last equality follows from the fact that (n − k) · (n + 1 − k) is even. If h ≡ s ≡ j + 1 (mod 2) then
|Pλ (j)| = |P n−k+1,1 (k − n + s, n − k)| = 0 by Lemma 2.1.
Proof of Proposition 3.1. We have
(2.13),(2.9)
hSλ W, pk,r (V ) ⊗ Ei = hSλ W ⊗ pk,r (W ), Ei
=
X
[n]
k
I∈(
sgn(σ(I)) · Sλ(r,I) W, E .
)
Since det(Sym2 W ) = det(W )⊗(n+1) = S(n+1)n W , we get using Cauchy’s formula [Wey03, Prop. 2.3.8] that
E = det Sym2 W ⊗ Sym Sym2 W =
M
Sλ W.
λ∈Zn
dom ,λn ≥n+1
λi ≡n+1 (mod 2)
Using notation (2.6–2.7) and (3.3) we obtain for r ≫ 0
(
0
1, if λ1 (I) ∈ C r and λ2 (I) ∈ C≥n+1
,
Sλ(r,I) W, E =
0, otherwise.
It follows (using notation (3.7)) that for j ∈ Z/2Z
pk,r (V
lim
r→∞
r≡k+1+j (mod 2)
)⊗E =
X
λ∈Zn
dom
X
I∈Pλ (j)
sgn(σ(I)) · Sλ W,
(3.10)
and by Lemmas 3.2 and 3.3 we only need to consider λ ∈ Z(s) for s ≥ n − k such that (for some h ∈ Z/2Z)
λ1 ≡ · · · ≡ λs ≡ h (mod 2)
and
λs+1 ≡ · · · ≡ λn ≡ j (mod 2),
i.e. λ ∈ C h,j (s). Multiplying both sides of (3.10) by (−1)k·(n−k) and using Lemma 3.4 we get
X
pk,r (V ) ⊗ E =
(−1)k·(n−k) ·
lim
(−1)(n−k)·h · |P n−k+j−h,s+1−h(k − n + s, n − k)| · Ch,j
s .
r→∞
r
(mod 2)
≡
k+1+j
n−k≤s≤n
h=j,j+1
We separate the contributions of the right hand side according to two cases:
18
CLAUDIU RAICU
Terms with h = j + 1: By Lemma 3.4 we can consider only the terms with s ≡ j (mod 2), in which case
s+1,s
we get from (3.6) that Ch,j
= C2s . We have
s = Cs
s−1
⌊ 2 ⌋
n − k odd,
n−k−1
2
Lemma
2.1
s−2
n−k+j−h,s+1−h
n−k+1,0
|P
(k − n + s, n − k)| = |P
(k − n + s, n − k)|
=
2
n − k and s even,
n−k−2
2
0
otherwise.
Comparing the coefficient of C2s in Proposition 3.1 with (−1)(n−k)·h · |P n−k+j−h,s+1−h(k − n + s, n − k)| in
each of the cases j = 0, 1, and (n − k) even and odd, we see that they agree.
Terms with h = j: The terms with s ≡ j + 1 (mod 2) contribute Ch,j
= Cs+1,s+1
with coefficient
s
s
h,j
(n−k)·h
n−k,0
(−1)
· |P
(k − n + s, n − k)|. The terms with s ≡ j (mod 2) contribute Cs = Cs,s
s with co(n−k)·h
n−k,1
n−k,1
efficient (−1)
· |P
(k − n + s, n − k)|. For s = n − k we get |P
(k − n + s, n − k)| = 0 so Cs,s
s
only appears for s > n − k. Observing that |P n−k,0 (k − n + s, n − k)| = |P n−k,1 (k − n + s + 1, n − k)| for
s ≥ n − k, and using C1s = Cs+1,s+1
+ Cs+1,s+1
in (3.6), we conclude that the terms with h = j contribute
s
s+1
X
(−1)(n−k)·h · |P n−k,0 (k − n + s, n − k)| · C1s ,
s≡j+1 (mod 2)
where
s
⌊2⌋
n−k
2
Lemma 2.1
s−1
|P n−k,0 (k − n + s, n − k)|
=
2
n−k−1
2
0
n − k even,
n − k and s odd,
otherwise.
Comparing with the coefficient of C1s in Proposition 3.1 we conclude the proof of the proposition.
3.2. General matrices. For positive integers m ≥ n and for s = 0, · · · , n, we let
A(s; m, n) = {λ ∈ Zndom : λs ≥ s + m − n, λs+1 ≤ s}.
(3.11)
If λ ∈ A(s; m, n) then we define a dominant weight λ(s) ∈ Zm
dom by
λ(s) = (λ1 − (m − n), · · · , λs − (m − n), s, · · · , s, λs+1 , · · · , λn )
| {z }
(3.12)
m−n
For vector spaces W1 , W2 , with dim(W1 ) = m, dim(W2 ) = n, and s = 0, · · · , n we define As ∈ Γ(W1 , W2 ) by
M
As =
Sλ(s) W1 ⊗ Sλ W2 .
(3.13)
λ∈A(s;m,n)
CHARACTERS OF EQUIVARIANT D-MODULES ON SPACES OF MATRICES
19
Proposition 3.5. We write W = W1 ⊗ W2 . If E = det(W ) ⊗ Sym(W ) then for k = 0, · · · , n,
n
X
s
(m−n)·(n−k−s)
k·(m−n)
(−1)
·
(−1)
· lim pk,r (V1 ) ⊗ pk,r (V2 ) ⊗ E =
· As .
r→∞
s−n+k
s=n−k
n
Proof of Proposition 3.5. Consider dominant weights δ̃ ∈ Zm
dom and λ̃ ∈ Zdom , and let
δ = δ̃ − (nm ),
λ = λ̃ − (mn ).
(3.14)
We obtain using (2.9), (2.13), and easy manipulations that Sδ̃ W1 ⊗ Sλ̃ W2 , pk,r (V1 ) ⊗ pk,r (V2 ) ⊗ E equals
X
sgn(σ(I)) · sgn(σ(J)) · Sδ(r,I) W1 ⊗ Sλ(r,J) W2 , Sym(W ) .
I∈([m]
, J∈([n]
k )
k )
Using (2.7) and writing µ|(0m−n ) for the sequence obtained by appending m − n zeros to µ, we get for r ≫ 0
(
1, if δ1 (I) = λ1 (J), δ2 (I) = λ2 (J)|(0m−n ), and δ2 (I) ∈ Zm−k
≥0 ,
Sδ(r,I) W1 ⊗ Sλ(r,J) W2 , Sym(W ) =
0, otherwise.
is
Let u ∈ {0, · · · , m} be the unique index such that δu ≥ u − m ≥ δu+1 . The condition δ2 (I) ∈ Zm−k
≥0
equivalent to the inclusion {u+ 1, · · · , m} ⊂ I, which implies u ≥ m − k. When m > n, the last m − n entries
of δ2 (I) being 0 forces δu = δu−1 = · · · = δu−m+n+1 = u − m, and all the elements u, u − 1, · · · , u − m + n + 1
to be contained in I c = [m] \ I. We modify δ and I as follows: we consider δ ∈ Zndom and I ∈ [n]
k defined by
δ = (δ1 , · · · , δu−m+n , δu+1 − (m − n), · · · , δm − (m − n)),
I = {i1 , · · · , ik−m+u , u + 1 − (m − n), u + 2 − (m − n), · · ·
c
I \{u, u−1, · · · , u−m+n+1}. The conditions δ1 (I) = λ1 (J)
, n},
so that I = [n]\I =
and δ2 (I) = λ2 (J)|(0m−n )
1
2
1
2
are then equivalent to δ (I) = λ (J) and δ (I) = λ (J). Since both λ, δ are dominant weights, these
equalities can only hold for δ = λ and I = J. Note that the freedom in choosing I (orI = J) is in the choice
of an increasing sequence i1 < · · · < ik−m+u inside {1, · · · , u}, i.e. there are u−m+n
k−m+u choices for I once we
fix δ. Writing s = u − m + n we get
c
s ≥ (m − k) − m + n = n − k,
λs+1
λs = δs = δu−m+n ≥ δu ≥ u − m = s − n,
= δu+1 − (m − n) ≤ (u − m) − (m − n) = (s − n) − (m − n) = s − m,
and moreover
δ = (λ1 , · · · , λs , s − n, · · · , s − n, λs+1 + (m − n), · · · , λn + (m − n)).
|
{z
}
m−n
It follows (using (3.12), (3.14)) that δ̃ = λ̃(s). Since sgn(σ(I)) = (−1)(m−n)·(m−u) · sgn(σ(I)), it follows that
if I = J and m − u = n − s then
sgn(σ(I)) · sgn(σ(J)) = (−1)(m−n)·(n−s) .
s
Putting everything together, and using u−m+n
k−m+u = s−n+k , we obtain for r
s
(m−n)·(n−s)
·
(−1)
s−n+k
Sδ̃ W1 ⊗ Sλ̃ W2 , pk,r (V1 ) ⊗ pk,r (V2 ) ⊗ E =
0
≫0
if λ̃ ∈ A(s; m, n) and δ̃ = λ̃(s)
for some s ≥ n − k,
otherwise.
20
CLAUDIU RAICU
Multiplying by (−1)k·(m−n) and taking the limit r → ∞ yields the desired conclusion.
3.3. Skew-symmetric matrices. For a positive integer m and for s = 0, · · · , m, we let
B(s, 2m) = {λ ∈ Z2m
dom : λ2s ≥ (2s − 1), λ2s+1 ≤ 2s, λ2i−1 = λ2i for all i},
B(s, 2m + 1) = {λ ∈ Z2m+1
dom : λ2s+1 = 2s, λ2i−1 = λ2i for i ≤ s, λ2i = λ2i+1 for i > s}.
For a vector space W with dim(W ) = n, and for s = 0, · · · , m = ⌊n/2⌋ we define Bs ∈ Γ(W ) via
M
Bs =
Sλ W.
(3.15)
(3.16)
λ∈B(s,n)
Proposition 3.6. If E = det
V
2
W ⊗ Sym
V
2
W
lim p2k,r (V ) ⊗ E =
r→∞
and k = 0, · · · , m = ⌊n/2⌋, then
m
X
s=m−k
s
· Bs .
m−k
With m = ⌊n/2⌋ and 1 ≤ k ≤ m, we define the following collections of dominant weights
B = {λ ∈ Z2k
dom : λ2i−1 = λ2i for i = 1, · · · , k},
n−2k
B≥n−1 = {λ ∈ Zdom
: λ2i−1 = λ2i ≥ n − 1 for i = 1, · · · , m − k, λn−2k = n − 1 if n is odd}.
(3.17)
We partition Zndom into the following collections of dominat weights Y(u), u = 0, · · · , n, defined by
Y(u) = {λ ∈ Zndom : λu ≥ u − 1 ≥ λu+1 },
(3.18)
In analogy with Lemma 3.2 one can prove:
1
2
Lemma 3.7. If I ∈ [n]
2k , then the conditions λ (I) ∈ B and λ (I) ∈ B≥n−1 are equivalent to
• λ ∈ Y(u) for some u = n − 2k, · · · , n.
• {u + 1, · · · , n} ⊂ I.
• λi2t−1 = λi2t and i2t = i2t−1 + 1 for all t = 1, · · · , m. If n is odd then u ∈ I c is odd and λu = u − 1.
Lemma 3.8. Assume that λ, I satisfy the equivalent conditions in Lemma 3.7. If n = 2m is even then
{(i1 , i2 ), (i3 , i4 ), · · · , (i2m−1 , i2m )} = {(1, 2), (3, 4), · · · , (2m − 1, 2m)}.
(3.19)
If n = 2m + 1 is odd and if we write u = 2s + 1, then we have
{(i1 , i2 ), (i3 , i4 ), · · · , (i2m−1 , i2m )} = {(1, 2), · · · , (2s − 1, 2s), (2s + 2, 2s + 3), · · · , (2m, 2m + 1)}.
(3.20)
Moreover, we have that λ ∈ B(s, n) for some s = m − k, · · · , m.
Proof. The conclusions (3.19–3.20) follow from the fact that i1 , · · · , in give a permutation of [n] with i2t =
i2t−1 + 1, and in = u is odd when n is odd. If n is odd and u = 2s + 1, it follows from u ≥ n − 2k that
s ≥ m − k. Moreover, we have λ2s+1 = λu = u − 1 = 2s, and it follows from (3.20) that λ ∈ B(s, n).
Assume now that n = 2m is even. It follows from (3.19) that in is even, so we can write in = 2s′ . Since
in + 1, · · · , n ∈ I, we get n − in ≤ 2k, i.e. s′ ≥ m − k. Since in ≤ u, we have λ2s′ ≥ λu ≥ u − 1 ≥ 2s′ − 1. We
have by (3.19) that λ2i−1 = λ2i for i = 1, · · · , m, so taking s to be the maximal index for which λ2s ≥ 2s − 1
we find that s ≥ s′ and λ2s+1 = λ2s+2 < 2s + 1, i.e. λ ∈ B(s, n).
CHARACTERS OF EQUIVARIANT D-MODULES ON SPACES OF MATRICES
21
Lemma 3.9. Let m = ⌊n/2⌋, and for m − k ≤ s ≤ m define the collection of partitions
B(k, n/2 − k, s) = {µ ∈ P (2k, n − 2k) :µ′i even for i = 1, · · · , n − 2k, µ′n−2k = 2m − 2s,
µi even for i = 2m − 2s + 1, · · · , 2k}.
(3.21)
Every partition µ ∈ B(k, n/2 − k, s) has even size, and the cardinality of the set B(k, n/2 − k, s) is given by
(
s
if n = 2m + 1 is odd;
m−k
|B(k, n/2 − k, s)| =
s−1
if n = 2m is even.
m−1−k
Proof. Since each µ′i is even, |µ| = |µ′ | is even. To compute the size of B(k, n/2 − k, s) we first note that
the condition µ′n−2k = 2m − 2s implies that µ1 = · · · = µ2m−2s = n − 2k, so any µ ∈ B(k, n/2 − k, s) is
determined by µ = (µ2m−2s+1 , · · · , µ2k ) ∈ P (2(k + s − m), n − 2k). Since µ′n−2k = 2m − 2s, we must have
µ1 < n − 2k. The condition µ ∈ B(k, n/2 − k, s) is then equivalent (using (2.10)) to
(
P 0,0 (2(k + s − m), 2(m − k))
if n is odd,
µ∈
0,0
P (2(k + s − m), 2(m − 1 − k)) if n is even.
s
s−1
By Lemma 2.1, the number of choices for µ is m−k
if n is odd, respectively m−1−k
if n is even.
Lemma 3.10. Assume that λ ∈ B(s, n) for some s ≥ m − k. The collection of subsets I ∈ [n]
2k for which
′
1
2
λ (I) ∈ B and λ (I) ∈ B≥n−1 corresponds via (2.3) to B (k, n/2 − k, s), where
n/2 − k, s)
if n is odd,
B(k,
s
′
[
B (k, n/2 − k, s) =
B(k, n/2 − k, s′ ) if n is even.
s′ =m−k
Proof. Consider λ ∈ B(s, n) for s ≥ m−k, and I ∈ [n]
2k satisfying the conditions of Lemma 3.7. If n = 2m+1
is odd then in = 2s + 1 and I contains 2s + 2, · · · , n, i.e. the corresponding µ ∈ P (2k, n − 2k) has
µ1 = · · · = µ2m−2s = n − 2k,
µ2m−2s+1 < n − 2k,
= 2m − 2s. For 2k < t < n we have that it ≤ 2s, so µ′t−2k = t − it is even by (3.20). The set of µi
so
with 2m − 2s < i ≤ 2k coincides with that of differences it − t for 1 ≤ t ≤ 2(k − m + s), which are all even
again by (3.20) and the fact that it ≤ 2s for t ≤ 2(k − m + s).
Assume next that n = 2m is even, and use (3.19) to write in = 2s′ . As in the previous paragraph,
this implies µ′n−2k = 2m − 2s′ . By (3.19) all the differences t − it are even, so all µi , µ′i are even. This
shows that I ∈ B(k, n/2 − k, s′ ). Since in + 1, · · · , n ∈ I we get as before s′ ≥ m − k. If s′ > s then
λ2 (I)n−2k = λin + n − in = λ2s′ + n − 2s′ ≤ λ2s+1 + n − 2s′ ≤ 2s + n − 2s′ ≤ n − 2, a contradiction.
The verification that µ ∈ B ′ (k, n/2 − k, s) yields a subset I with λ1 (I) ∈ B and λ2 (I) ∈ B≥n−1 follows
easily by tracing back the arguments.
µ′n−2k
Proof of Proposition 3.6. We have
hSλ W, p2k,r (V ) ⊗ Ei = hSλ W ⊗ p2k,r (W ), Ei
(2.13),(2.9)
=
X
I∈(
sgn(σ(I)) Sλ(r,I) W, E .
[n]
2k
)
Using notation (2.7) and (3.17) we get that for r ≫ 0
(
1, if λ1 (I) ∈ B and λ2 (I) ∈ B≥n−1 ,
Sλ(r,I) W, E =
0, otherwise.
22
CLAUDIU RAICU
It follows that
X
lim p2k,r (V ) ⊗ E =
r→∞
m
X
Lemmas 3.8−3.10
=
s=m−k
sgn(σ(I)) · Sλ W
[n]
2k
λ∈Zn
dom , I∈( )
λ1 (I)∈B, λ2 (I)∈B≥n−1
m
|µ| even X
X
(−1)|µ| Sλ W =
X
s=m−k
λ∈B(s,n)
µ∈B′ (k,n/2−k,s)
λ∈B(s,n)
|B ′ (k, n/2 − k, s)| · Sλ W
s
, and the desired equality follows.
If n = 2m + 1 is odd, then |B ′ (k, n/2 − k, s)| = |B(k, n/2 − k, s)| = m−k
Similarly, when n = 2m is even, we get
′
s
s
X
X
s −1
s
|B ′ (k, n/2 − k, s)| =
|B(k, n/2 − k, s′ )| =
=
.
m−1−k
m−k
′
′
s =m−k
s =m−k
4. Equivariant D-modules on symmetric matrices
In this section we compute the characters of the GL-equivariant D-modules on the vector space M symm of
symmetric n×n matrices. We let W denote a complex vector space of dimension n, V = W ∗ , and we identify
Sym2 W with M symm , where squares w2 correspond to matrices of rank one. If we write GL = GL(W ), and
let Mssymm denote the subvariety of matrices of rank at most s then the main result of this section is:
Theorem 4.1. There exist (2n + 1) simple GL-equivariant holonomic D-modules on M symm , namely
(
symm
L(Mn−s
, M symm )
if j ≡ s (mod 2),
Csj =
for s = 0, · · · , n − 1, j = 1, 2,
symm
symm
; 1/2) if j ≡ s + 1 (mod 2),
L(Mn−s , M
and Cn1 = Cn2 = L({0}, M symm ). For all s, j, the character of Csj is Cjs (as defined in (3.2)).
The remaining assertion of the Theorem on Equivariant D-modules on Symmetric Matrices described in
the Introduction is the identification Cs1 = Fs+1 /Fs−1 for s = 0, · · · , n: its proof follows closely the proof
of Theorem 1.1 in the next section, so we leave the details to the interested reader. The classification of
GL-equivariant holonomic simple D-modules is explained in Theorem 2.9, so we only need to check that Cjs
is the character of Csj . For k = 1, · · · , n, we consider the situation of Section 2.4, with X = Xk = G(k, V )
and R, Q as in (2.17). We let U = Sym2 V , η = Sym2 Q. If we write Y = Yk , π = πk , then (2.23) becomes
/ Sym2 W × G(k, V )
❚❚❚❚
❚❚❚❚
❚❚
πk ❚❚❚❚❚❚
❚*
Yk = TotXk (Sym2 Q∗ )
(4.1)
Sym2 W
Locally on Xk , Q∗ trivializes to a vector space of dimension k, and Yk gets identified with the space of k × k
symmetric matrices. We take L = (det Q)⊗2 , consider its GL-equivariant inclusion L ⊂ Symk η and note
that L is locally generated by the symmetric determinant. If we let Yk0 ⊂ Yk be the open set defined locally
by the non-vanishing of the determinant, M0k = OY 0 is a DYk -module. Note that Yk0 maps isomorphically
k
via πk to the orbit of symmetric matrices of rank k. As a GL-equivariant quasi-coherent sheaf on Xk
M
(det Q)⊗r ⊗ det Sym2 Q∗ ⊗ Sym Sym2 Q∗ ,
(4.2)
Sλ Q =
lim
M0k =
−→
λ∈Zkdom ,λi even
r≡k+1 (mod 2)
CHARACTERS OF EQUIVARIANT D-MODULES ON SPACES OF MATRICES
so condition (2.24) is satisfied in our context. The Euler characteristic of the D-module pushforward
is now easily computed as a consequence of Proposition 2.10 and of Remark 2.11:
!
Z
M0k = (−1)k·(n−k) ·
pk,r (V ) ⊗ det Sym2 W ⊗ Sym Sym2 W
,
lim
χ
r→∞
πk
23
R
πk
M0k
(4.3)
r≡k+1 (mod 2)
which is evaluated explicitly in Proposition 3.1.
We next explain why M0k ⊗ det(Q) also has the structure of a DYk -module. Consider the étale double
1/2
cover Yk of Yk0 defined locally by the square-root of the symmetric determinant. The structure sheaf OY 1/2
k
is naturally a DY 0 -module [CL01] and hence also a DYk -module. It contains M0k so we can define M1k as
k
the cokernel of the inclusion M0k ⊂ OY 1/2 . As a GL-equivariant quasi-coherent sheaf on Xk , M1k is given by
k
M
M1k =
λ∈Zkdom ,λi
Sλ Q = OY 0 ⊗ det(Q).
k
(4.4)
odd
It follows that M1k satisfies the setting of Proposition 2.10 with L′ = det(Q) so we can compute the Euler
characteristic of its direct image via πk as
!
Z
1
k·(n−k)
2
2
Mk = (−1)
·
lim
,
(4.5)
χ
pk,r (V ) ⊗ det Sym W ⊗ Sym Sym W
r→∞
πk
r≡k (mod 2)
which is evaluated in Proposition 3.1. We are now ready to prove the main result of this section:
Proof of Theorem 4.1. The classification of simple D-modules follows from Theorem 2.9, so it remains to
check that in Γ(W ) we have the equalities Csj = Cjs for s = 0, · · · , n and j = 1, 2. The equalities (4.3–4.5)
together with Proposition 3.1 yield for s = 1, · · · , n and j = 1, 2,
j
Cn−s
= Cjn−s +
n
X
(asi · C1i + bsi · C2i ), for some integers asi , bsi .
i=n−s+1
Since Cnj = det(Sym2 W ) ⊗ Sym(Sym2 W ) has character Cjn (by Cauchy’s formula [Wey03, Prop. 2.3.8]), the
equation above is also satisfied for s = 0. The Fourier transform F permutes the modules Csj , and it takes
F(C1s ) = C1n−s for s = 0, · · · , n and
F(C2s ) = C2n−s−1 for s = 0, · · · , n − 1.
We can then apply Lemma 2.14 to the poset P = {(s, j) : s = 0, · · · , n − 1, j = 1, 2} ∪ {(n, 1)} with the
lexicographic ordering given by (s, j) < (s′ , j ′ ) if and only if s < s′ , or s = s′ and j < j ′ . We let v(s,j) = Cjs
and v(s,j) = Csj , and conclude using Lemma 2.14 that Csj = Cjs for all s = 0, · · · , n and j = 1, 2.
5. Equivariant D-modules on m × n matrices
In this section we compute the characters of the GL-equivariant D-modules on the vector space M of
m × n matrices, for m ≥ n. We consider W1 , W2 vector spaces of dimension dim(W1 ) = m, dim(W2 ) = n,
let Vi = Wi∗ , and identify W = W1 ⊗ W2 with M , where tensor products w1 ⊗ w2 correspond to matrices of
rank one. If we write GL = GL(W1 ) × GL(W2 ), let Ms denote the subvariety of matrices of rank at most s,
and recall the notation (3.13) for the characters As , then the main result of this section is:
24
CLAUDIU RAICU
Theorem on Equivariant D-modules on General Matrices. The simple GL-equivariant holonomic
D-modules on M are As = L(Mn−s , M ), s = 0, · · · , n, and for each s the character of As is As . When
m = n, As is as described in Theorem 1.1, while for m > n it can be expressed in terms of local cohomology:
1+s·(m−n)
As = HMn−1
codim(Mn−s )
(M, OM ) = HMn−s
(M, OM ).
(5.1)
We only need to show that As is the character of As , and to prove Theorem 1.1. The classification of
GL-equivariant holonomic simple D-modules is explained in Theorem 2.9, while (5.1) follows by comparing
As with the characters of local cohomology modules from [RWW14, Thm. 4.5] and [RW14, Thm. 6.1].
Proof of Theorem 1.1. Let’s assume for now that As is the character of As , and write W1 = W2 = Cn . Using
Cauchy’s formula [Wey03, Cor. 2.3.3], we get an equality of GL-representations
Sdet =
M
S λ W1 ⊗ S λ W2 =
λ∈Zn
dom
n
M
Ai .
i=0
As in Example 1.2, this shows that A0 , · · · , An are the D-module composition factors of Sdet , each appearing
with multiplicity one. It remains to check that As = Fs /Fs−1 where Fs = hdet−s iD .
We prove by induction on s that the D-module composition factors of Fs are A0 , · · · , As , which is clearly
true
Ln for s = 0. Assume that s > 0 and that the induction hypothesis is valid for Fs−1 , so that Sdet /Fs−1 =
i=s Ai as GL-representations. We must then have for some i ≥ s an inclusion of D-modules Ai ⊂ Sdet /Fs−1 .
Using the character description, Ai must contain the class of det−i inside the quotient Sdet /Fs−1 , and
therefore it must also contain the classes of det−i+1 , det−i+2 , · · · . If i > s this contradicts the formula for
the character of Ai . We conclude that i = s and that we have an inclusion As ⊂ Sdet /Fs−1 . Since As is
simple, it is generated by the class of det−s , so the image of As is Fs /Fs−1 .
We note that, just as in Remark 1.4, the strict inclusions Fi−1 ( Fi , i = 1, · · · , n, in Theorem
1.1 combined
Qn
with Cayley’s identity show that the b-function of the generic determinant is bdet (s) = i=1 (s + i).
We conclude by showing that As is the character of As . For k = 1, · · · , n, we consider the situation of
Section 2.4, with X = Xk = G(k, V1 ) × G(k, V2 ) and R1 , Q1 , R2 , Q2 as in (2.17). We let U = V1 ⊗ V2 ,
η = Q1 ⊗ Q2 , and write Y = Yk , π = πk in (2.23). We note that locally on Xk , Q∗1 , Q∗2 trivialize to vector
spaces of dimension k, and Yk gets identified with the space of k × k matrices. We take the line bundle
L = det Q1 ⊗ det Q2 , consider its GL-equivariant inclusion L ⊂ Symk η, and note that L is locally generated
by the function that assigns to a matrix its determinant. If we let Yk0 ⊂ Yk be the open set defined locally by
the non-vanishing of the determinant, then as a GL-equivariant quasi-coherent sheaf on Xk , OY 0 is given by
k
OY 0 =
k
M
λ∈Zkdom
Sλ Q1 ⊗ Sλ Q2 = lim L⊗r ⊗ det (Q∗1 ⊗ Q∗2 ) ⊗ Sym (Q∗1 ⊗ Q∗2 ) ,
−→
r
so condition (2.24) is satisfied in our context. The Euler characteristic of the D-module pushforward
is now easily computed as a consequence of Propositions 2.10 and 3.5, and of Remark 2.11:
Z
OY 0 = (−1)k·(m−n) · lim pk,r (V1 ) ⊗ pk,r (V2 ) ⊗ det(W ) ⊗ Sym(W )
χ
πk
R
πk
OY 0
k
r→∞
k
=
n
X
s=n−k
(m−n)·(n−k−s)
(−1)
s
·
· As .
s−n+k
Since OY 0 maps isomorphically via πk to the orbit of rank k matrices in M , the conclusion that As is the
k
character of As follows as in the proof of Theorem 4.1 by the linear algebra trick in Section 2.6.
CHARACTERS OF EQUIVARIANT D-MODULES ON SPACES OF MATRICES
25
6. Equivariant D-modules on skew-symmetric matrices
In this section we compute the characters of the GL-equivariant D-modules on the vector space of skewsymmetric
a complex vector space of dimension n, V = W ∗ , and we
V2n × n matrices. We let W denote
skew
W with the vector space M
of n × n skew-symmetric matrices, where exterior products
identify
w1 ∧ w2 correspond to matrices of rank two. If we write GL = GL(W ), m = ⌊n/2⌋, let Msskew denote the
subvariety of matrices of rank at most 2s, and recall the notation (3.16) for the characters Bs then we have:
Theorem on Equivariant D-modules on Skew-symmetric Matrices. The simple GL-equivariant
skew , M skew ), s = 0, · · · , m, and for each s the character of
holonomic D-modules on M skew are Bs = L(Mm−s
Bs is Bs . If n = 2m + 1 is odd then for s = 1, · · · , m, Bs can be described in terms of local cohomology:
skew )
codim(Mm−s
2s+1
skew
Bs = H M
, OM skew ) = HM skew
skew (M
m−1
m−s
(M skew , OM skew ).
(6.1)
skew . We let S denote the coordinate
If n = 2m is even, we let Pf be an equation defining the hypersurface Mm−1
ring of M skew , and consider Fs = hPf −2s iD , the D-submodule of the localization SPf generated by Pf −2s for
s = 0, · · · , m (and F−1 = 0). We have that Bs = Fs /Fs−1 for s = 0, · · · , m.
The classification of GL-equivariant holonomic simple D-modules is explained in Theorem 2.9, while
the equality (6.1) follows from [RWW14, Theorem 5.5] and [RW15, (1.4)]. When n = 2m, we get that
Bs = Fs /Fs−1 just as in the proof of Theorem 1.1. Note that Cayley’s identity shows that bPf (s) divides
Q
m
−2i
iD = hPf −2i+1 iD . The
i=1 (s + 2 · i − 1), which in turn implies that hPf
Q strict inclusions Fi−1 ( Fi then
force 2 · i − 1 to be a root of bPf (s) for i = 1, · · · , m, so in fact bPf (s) = m
i=1 (s + 2 · i − 1).
To prove the theorem, it remains to check that Bs is the character of Bs . For k = 1, · · ·V, m, we consider
V
the situation of Section 2.4, with X = Xk = G(2k, V ) and R, Q as in (2.17). We let U = 2 V , η = 2 Q,
and write Y = Yk , π = πk in (2.23). Locally on Xk , Q∗ trivializes to a vector space of dimension 2k, and
Yk gets identified with the space of 2k × 2k skew-symmetric matrices. We take the line bundle L = det Q
to be the Plücker line bundle on X, consider its GL-equivariant inclusion L ⊂ Symk η, and note that L is
locally generated by the function that assigns to a skew-symmetric matrix its Pfaffian. If we let Yk0 ⊂ Yk
be the open set defined locally by the non-vanishing of the Pfaffian, then we get using Cauchy’s formula
[Wey03, Prop. 2.3.8] that condition (2.24) is satisfied. As a consequence of Propositions 2.10 and 3.6, and
of Remark 2.11 we obtain
!
!
Z
m
2
2
X
^
^
s
W =
OY 0 = lim p2k,r (V ) ⊗ det
W ⊗ Sym
χ
· Bs .
(6.2)
k
r→∞
m−k
πk
s=m−k
Since OY 0 maps isomorphically via πk to the orbit of rank 2k matrices in M skew , we conclude as in the proof
k
of Theorem 4.1 that Bs is the character of Bs for all s.
7. The simple regular holonomic D-modules on rank stratifications
We let X denote any of the vector spaces of general, symmetric, or skew-symmetric matrices, with the
natural group action by row and column operations of the corresponding group G as considered in the
previous sections. We denote by Λ the union of conormal varieties to the orbits of G, and consider the
category C = modrh
Λ (DX ) of regular holonomic DX -modules whose characteristic variety is contained in Λ.
The goal of this section is to describe explicitly the simple objects in C and obtain as a corollary a direct
proof of Levasseur’s conjecture [Lev09, Conj. 5.17] in the case of general and skew-symmetric matrices.
Via the Riemann–Hilbert correspondence, the simple objects in C are classified by irreducible local systems
on the G-orbits. When the local systems are G-equivariant, the corresponding DX -modules have been
26
CLAUDIU RAICU
described in the previous sections. The only orbits with irreducible non-equivariant local systems are the
orbits O ⊂ X of rank n matrices, when X is the vector space of n × n general or symmetric matrices, or
when X is the vector space of 2n × 2n skew-symmetric matrices. In each of these cases, the complement of
O in X is defined by a single polynomial f which is the determinant of the generic (symmetric) n × n matrix
in the first two cases, and it is the Pfaffian of the generic 2n × 2n skew-symmetric matrix in the last case.
The fundamental group of O is equal to Z, so the monodromy of the corresponding local system is given
by a non-zero complex number λ = e2πiα with α ∈ C/Z. We let S denote the coordinate ring of X and for
α ∈ C we consider the DX -module Fα = Sf · f α (which only depends on the class of α in C/Z).
Theorem 7.1. With notation as above, consider the irreducible local system Lα on O whose monodromy is
given by λ = e2πiα . If Lα is not G-equivariant then the corresponding simple object in modrh
Λ (DX ) is Fα .
Proof. The restriction of Fα to O is a rank one integrable connection whose corresponding local system has
monodromy given by λ = e2πiα . It follows that in order to prove the theorem we need to check that Fα is a
simple DX -module. The condition that Lα is not G-equivariant is equivalent to (see Theorems 2.7 and 2.9)
• α∈
/ Z if X is the space of general or skew-symmetric matrices.
• α∈
/ 21 Z if X is the space of symmetric matrices.
From now on we assume that Lα is not G-equivariant. It follows from the Cayley’s identity (and its symmetric
and skew-symmetric versions) that Fα is generated as a DX -module by f α (or by f r+α for any r ∈ Z). In
order to prove that Fα is simple, it is then sufficient to show that any non-zero DX -submodule F ⊂ Fα
contains f r+α for r ≫ 0. Fix any such F .
We write g for the Lie algebra of G, and note that any DX -module is a g-representation. In particular this
is true about F ⊂ Fα . Since Fα has a multiplicity free decomposition into irreducible g-representations of
the form M · f α , where M ⊂ Sf is an irreducible integral g-representation, we may assume that F contains
one such M · f α . Replacing α by α − r and M by M · f r for r ∈ Z, we may assume that M ⊂ S. Since
M generates a non-zero ideal which is invariant under the action of G, it defines set-theoretically a proper
closed G-invariant subset of X, which is necessarily contained in the zero locus of f (the complement of f
is a dense orbit for the G-action). We obtain that the ideal in S generated by M contains all large enough
powers of f , and therefore that F contains f r+α for r ≫ 0, which concludes the proof of the theorem.
We end by remarking that Theorem 7.1 yields a proof of Levasseur’s conjecture in the case of general
and skew-symmetric matrices. We have already seen that the irreducible G-equivariant local systems on
the orbits of the group action give rise to simple DX -modules containing (and hence generated by) non-zero
sections invariant under the action of the derived subgroup G′ . By Theorem 7.1, the remaining simple
objects of C are all of the form Fα = Sf · f α . Since f is a G′ -invariant, the same is true about f α , so Fα
contains non-zero G′ -invariant sections.
Acknowledgments
I am grateful to Nero Budur, David Eisenbud, Sam Evens, Mircea Mustaţă, Uli Walther and Jerzy
Weyman for interesting conversations and helpful advice, as well as to the anonymous referee for suggesting
many improvements to the presentation. Experiments with the computer algebra software Macaulay2 [GS]
have provided numerous valuable insights. This work was supported by the National Science Foundation
Grant No. 1458715.
References
+
[BGK 87] A. Borel, P.-P. Grivel, B. Kaup, A. Haefliger, B. Malgrange, and F. Ehlers, Algebraic D-modules, Perspectives in
Mathematics, vol. 2, Academic Press, Inc., Boston, MA, 1987. MR882000 (89g:32014)
CHARACTERS OF EQUIVARIANT D-MODULES ON SPACES OF MATRICES
27
[BG99] Tom Braden and Mikhail Grinberg, Perverse sheaves on rank stratifications, Duke Math. J. 96 (1999), no. 2, 317–362,
DOI 10.1215/S0012-7094-99-09609-6. MR1666554 (2000f:14031)
[CL01] S. C. Coutinho and D. Levcovitz, D-modules and étale morphisms, Comm. Algebra 29 (2001), no. 4, 1487–1497,
DOI 10.1081/AGB-100002114. MR1853107 (2002m:14014)
[GS] Daniel R. Grayson and Michael E. Stillman, Macaulay 2, a software system for research in algebraic geometry,
Available at http://www.math.uiuc.edu/Macaulay2/.
[Har77] Robin Hartshorne, Algebraic geometry, Springer-Verlag, New York, 1977. Graduate Texts in Mathematics, No. 52.
MR0463157 (57 #3116)
[HTT08] Ryoshi Hotta, Kiyoshi Takeuchi, and Toshiyuki Tanisaki, D-modules, perverse sheaves, and representation theory,
Progress in Mathematics, vol. 236, Birkhäuser Boston, Inc., Boston, MA, 2008. Translated from the 1995 Japanese
edition by Takeuchi. MR2357361 (2008k:32022)
[Kas76] Masaki Kashiwara, B-functions and holonomic systems. Rationality of roots of B-functions, Invent. Math. 38
(1976/77), no. 1, 33–53. MR0430304 (55 #3309)
[Kem78] George Kempf, The Grothendieck-Cousin complex of an induced representation, Adv. in Math. 29 (1978), no. 3,
310–396, DOI 10.1016/0001-8708(78)90021-X. MR509802 (80g:14042)
[Kim03] Tatsuo Kimura, Introduction to prehomogeneous vector spaces, Translations of Mathematical Monographs, vol. 215,
American Mathematical Society, Providence, RI, 2003. Translated from the 1998 Japanese original by Makoto Nagura
and Tsuyoshi Niitani and revised by the author. MR1944442 (2003k:11180)
[Lev09] Thierry Levasseur, Radial components, prehomogeneous vector spaces, and rational Cherednik algebras, Int. Math.
Res. Not. IMRN 3 (2009), 462–511, DOI 10.1093/imrn/rnn137. MR2482122 (2010g:22028)
[Mac95] I. G. Macdonald, Symmetric functions and Hall polynomials, 2nd ed., Oxford Mathematical Monographs, The Clarendon Press Oxford University Press, New York, 1995. With contributions by A. Zelevinsky; Oxford Science Publications. MR1354144 (96h:05207)
[MV86] Robert MacPherson and Kari Vilonen, Elementary construction of perverse sheaves, Invent. Math. 84 (1986), no. 2,
403–435, DOI 10.1007/BF01388812. MR833195 (87m:32028)
[Nan08] Philibert Nang, On a class of holonomic D-modules on Mn (C) related to the action of GLn (C) × GLn (C), Adv.
Math. 218 (2008), no. 3, 635–648, DOI 10.1016/j.aim.2008.01.001. MR2414315 (2009c:14028)
[Nan12]
, On the classification of regular holonomic D-modules on skew-symmetric matrices, J. Algebra 356 (2012),
115–132, DOI 10.1016/j.jalgebra.2012.01.021. MR2891125
[Rai14] Claudiu Raicu, Characters of equivariant D-modules on Veronese cones, arXiv 1412.8148 (2014). To appear in
Trans. AMS.
[RW14] Claudiu Raicu and Jerzy Weyman, Local cohomology with support in generic determinantal ideals, Algebra & Number
Theory 8 (2014), no. 5, 1231–1257, DOI 10.2140/ant.2014.8.1231. MR3263142
[RWW14] Claudiu Raicu, Jerzy Weyman, and Emily E. Witt, Local cohomology with support in ideals of maximal minors and
sub-maximal Pfaffians, Adv. Math. 250 (2014), 596–610, DOI 10.1016/j.aim.2013.10.005. MR3122178
[RW15] Claudiu Raicu and Jerzy Weyman, Local cohomology with support in ideals of symmetric minors and Pfaffians, arXiv
1509.03954 (2015).
[Sai15] Morihiko Saito, D-modules generated by rational powers of holomorphic functions, arXiv 1507.01877 (2015).
[VdB99] Michel Van den Bergh, Local cohomology of modules of covariants, Adv. Math. 144 (1999), no. 2, 161–220, DOI
10.1006/aima.1998.1809. MR1695237 (2000d:14051)
[Vil85] K. Vilonen, Intersection homology D-module on local complete intersections with isolated singularities, Invent. Math.
81 (1985), no. 1, 107–114, DOI 10.1007/BF01388774. MR796193 (87f:32023)
[Wal15] Uli Walther, Survey on the D-module f s , arXiv 1504.07516 (2015).
[Wey03] Jerzy Weyman, Cohomology of vector bundles and syzygies, Cambridge Tracts in Mathematics, vol. 149, Cambridge
University Press, Cambridge, 2003. MR1988690 (2004d:13020)
Department of Mathematics, University of Notre Dame, 255 Hurley, Notre Dame, IN 46556
Institute of Mathematics “Simion Stoilow” of the Romanian Academy
E-mail address: [email protected]
| 0math.AC
|
LE CAM THEORY ON THE COMPARISON OF STATISTICAL
MODELS
arXiv:1605.03301v1 [math.ST] 11 May 2016
ESTER MARIUCCI
Abstract. We recall the main concepts of the Le Cam theory of statistical experiments, especially the notion of Le Cam distance and its properties. We also review
classical tools for bounding such a distance before presenting some examples. A
proof of the classical equivalence result between density estimation problems and
Gaussian white noise models will be analyzed.
Keywords: Statistical experiments, Le Cam distance, deficiency, density estimation model.
AMS Classification: Primary 62B15; secondary 62G20, 62G07.
1. Introduction
The theory of Mathematical Statistics is based on the notion of statistical model, also
called statistical experiment or just experiment. A statistical model, as in its original
formulation due to Blackwell (1951), is a triple
P = (Ω, T , (Pθ : θ ∈ Θ)),
where (Ω, T ) is a sample space, Θ is a set called the parameter space and (Pθ : θ ∈ Θ) is a
family of probability measures on (Ω, T ). This definition is a mathematical abstraction
intended to represent a concrete experiment; consider for example the following situation
taken from the book of Le Cam and Yang (2000). A physicist decides to estimate the
half life of Carbon 14, C 14 . He supposes that the life of a C 14 atom has an exponential
distribution with parameter θ and, in order to develop his investigation, he takes a
sample of n atoms of C 14 . The physicist fixes in advance the duration of the experiment,
say 2 hours, and then he counts the number of disintegrations. Formally, this leads to the
definition of the statistical model P1 = (N, P(N), (Pθ : θ ∈ (0, ∞))) where Pθ represents
the law of the random variable X counting the number of disintegrations observed in
2 hours. This is not the only way to proceed if we want to estimate the half life of
Carbon 14. Indeed, the physicist could choose to consider the first random time Y after
which a fixed number of disintegrations, say 106 , have occurred. In this case he will
represent the experiment via the statistical model P2 = (R+ , B(R+ ), (Qθ : θ ∈ (0, ∞)))
where Qθ is the law of the random variable Y . A natural question is then how much
“statistical information” the considered experiments contain or, more precisely, when the
experiment P1 will be more informative than P2 and conversely.
The quest for comparison of statistical experiments was initiated by the paper of
Bohnenblust, Shapley and Sherman (1949) followed by the papers of Blackwell (1951,
1953) where the following definition was introduced: “P1 is more informative than P2 ”
if for any bounded loss function L, kLk∞ ≤ 1, and any decision procedure ρ2 in the
experiment P2 there exists a decision procedure ρ1 in the experiment P1 such that
Rθ (P1 , ρ1 , L) ≤ Rθ (P2 , ρ2 , L),
∀θ ∈ Θ.
Date: 12th May 2016.
The research leading to these results has received funding from the European Research Council
under ERC Grant Agreement 320637.
1
2
ESTER MARIUCCI
Here we denote by Rθ (P1 , ρ1 , L) and Rθ (P2 , ρ2 , L) the statistical risk for the experiments P1 and P2 , respectively.
However, this can lead to two models being non-comparable. This issue was solved
by Le Cam who introduced the notion of deficiency δ(P1 , P2 ). We will give a precise
definition in the forthcoming sections. Here, we only remark two interesting properties:
• δ(P1 , P2 ) is a well defined non-negative real number for every two given statistical models P1 and P2 sharing the same parameter space.
• For every loss function L with 0 ≤ L ≤ 1 and every decision procedure ρ2
available on Θ using P2 , there exists a decision procedure ρ1 in P1 such that
for all θ ∈ Θ,
Rθ (P1 , ρ1 , L) ≤ Rθ (P2 , ρ2 , L) + δ(P1 , P2 ).
This solves the issue mentioned above: It could be that both δ(P1 , P2 ) and δ(P2 , P1 )
are strictly positive, in which case they will not be comparable according to the first
definition; nevertheless, we can still say “how much information” we lose when passing
from one model to the other one. Le Cam’s theory has found applications in several
problem in statistical decision theory and it has been developed, for example, for nonparametric regression, nonparametric density estimation problems, generalized linear
models, diffusion models, Lévy models, spectral density estimation problem. Historically, the first results of asymptotic equivalence in a nonparametric context date from
1996 and are due to Brown and Low (1996) and Nussbaum (1996). The first two authors have shown the asymptotic equivalence of nonparametric regression and a Gaussian
white noise model while the third one those of density estimation problems and Gaussian white noise models. Over the years many generalizations of these results have been
proposed such as Brown et al. (2002); Carter (2006, 2007, 2009); Grama and Nussbaum
(2002); Meister and Reiß (2013); Reiß (2008); Rohde (2004); Schmidt-Hieber (2014) for
nonparametric regression or Brown et al. (2004); Carter (2002); Jähnisch and Nussbaum
(2003); Mariucci (To appear) for nonparametric density estimation models. Another
very active field of study is that of diffusion experiments. The first result of equivalence
between diffusion models and Euler scheme was established in 1998, see Milstein and
Nussbaum (1998). In later papers generalizations of this result have been considered (see
Genon-Catalot and Laredo (2014); Mariucci (2016b)) as well as different statistical problems always linked with diffusion processes (see, e.g., Dalalyan and Reiß (2006, 2007);
Delattre and Hoffmann (2002); Genon-Catalot, Laredo and Nussbaum (2002)). Among
others we can also cite equivalence results for generalized linear models (see, e.g., Grama
and Nussbaum (1998)), time series (see, e.g., Grama and Neumann (2006); Milstein and
Nussbaum (1998)), GARCH model (see, e.g., Buchmann and Müller (2012)), functional
linear regression (see, e.g.,Meister (2011)), spectral density estimation (see, e.g. Golubev, Nussbaum and Zhou (2010)), volatility estimation (see, e.g. Reiß (2011)) and
jump models (see, e.g., Mariucci (2015, 2016a)). Negative results are somewhat harder
to come by; the most notable among them are Brown and Zhang (1998); Efromovich and
Samarov (1996); Wang (2002). Another new research direction that has been explored
involves quantum statistical experiments (see, e.g., Buscemi (2012)).
The aim of this survey paper is to present some basic concepts of the Le Cam theory
of asymptotic equivalences between statistical models. Our aim in this review is to
give an accessible introduction to the subject. Therefore, we will not follow the most
general approach to the theory, also because such an approach is already available in
the literature, see e.g., Le Cam (1986); Le Cam and Yang (2000); van der Vaart (2002).
In order to achieve such a goal, the paper has been organized as follows. In Section 2
we recall the definition of the Le Cam distance and its statistical meaning. Particular
attention has been payed to the interpretation of the Le Cam distance in terms of
decision theory. In Section 3 we collect some classical tools to control the Le Cam
LE CAM THEORY ON THE COMPARISON OF STATISTICAL MODELS
3
distance before passing to some examples described in Section 4. Section 5 is devoted
to show in details a proof of a classical result in Le Cam theory, namely the asymptotic
equivalence between density estimation problems and Gaussian white noise models.
2. Deficiency and Le Cam distance
As we have already pointed out, a possible way to compare two given statistical
models (having the same parameter space) could be to compare the corresponding risk
functions or to ask “how much information” we lose when passing from one model to
the other one, saying that there is no loss if we have at our disposal a mechanism able
to convert the observations from the distribution P1,θ to observations from P2,θ . If we
adopt the latter point of view a natural formalization for such a mechanism is the notion
of Markov kernel.
Definition 2.1. Let (Xi , Ti ), i = 1, 2, be two measurable spaces. A Markov kernel
K with source (X1 , T1 ) and target (X2 , T2 ) is a map K : X1 × T2 → [0, 1] with the
following properties:
• The map x 7→ K(x, A) is T1 -measurable for every A ∈ T2 .
• The map A 7→ K(x, A) is a probability measure on (X2 , T2 ) for every x ∈ X1 .
We will denote by K : (X1 , T1 ) → (X2 , T2 ) a Markov kernel with source (X1 , T1 )
and target (X2 , T2 ).
Starting from a Markov kernel K : (X1 , T1 ) → (X2 , T2 ) and a probability measure
P1 on (X1 , T1 ) one can construct a probability measure on (X2 , T2 ) in the following
way:
Z
KP1 (A) = K(x, A)P1 (dx), ∀A ∈ T2 .
Roughly speaking we can think that two models P1 and P2 contain “the same amount
of information about θ” if there exist two Markov kernels, K1 and K2 , not depending
on θ, such that K1 P1,θ = P2,θ and K2 P2,θ = P1,θ . This idea has been formalized in
the sixties by Lucien Le Cam and led to the notion of the deficiency, hence to the
introduction of a pseudo-metric on the class of all statistical experiments having the
same parameter space.
The definition of the deficiency in its most general form involves the notion of “transition” which is a generalization of the concept of Markov kernel. In this paper, however,
we prefer to keep things simpler and only focus on the case in which one has to deal
with dominated statistical models having Polish sample spaces (see below for a definition). The advantage is that in this case the definition of deficiency simplifies and the
abstract concept of transition coincides with that of Markov kernel (see Proposition 9.2
in Nussbaum (1996)).
Definition 2.2. A statistical model P1 = (X1 , T1 , (P1,θ : θ ∈ Θ)) is called Polish if its
sample space (X1 , T1 ) is a separable completely metrizable topological space.
P1 is said to be dominated if there exists a σ-finite measure µ on (X1 , T1 ) such that,
for all θ ∈ Θ, P1,θ is absolutely continuous with respect to µ. The measure µ is called
the dominating measure.
Example 2.3. Typical examples of Polish spaces in probability theory are the spaces
R, Rn , R∞ , the space CT of continuous functions on [0, T ] equipped with the supremum
norm d(x, y) = sup0≤t≤T |xt − yt |, the space D of càdlàg functions equipped with the
Skorokhod metric.
Definition 2.4. Let Q1 and Q2 be two probability measures defined on a measurable
space Ω. The total variation distance between Q1 and Q2 is defined as the quantity:
1
kQ1 − Q2 kT V = sup |Q1 (A) − Q2 (A)| = L1 (Q1 , Q2 ),
2
A⊆Ω
4
ESTER MARIUCCI
where L1 (Q1 , Q2 ) denotes the L1 norm between Q1 and Q2 .
Definition 2.5. Let Pi = (Xi , Ti , (Pi,θ : θ ∈ Θ)), i = 1, 2, be two experiments. The
deficiency δ(P1 , P2 ) of P1 with respect to P2 is the number
δ(P1 , P2 ) = inf sup kT P1,θ − P2,θ kT V ,
T θ∈Θ
for an infimum taken over all Markov kernels T : (X1 , T1 ) → (X2 , T2 ) and k · kT V
denotes the total variation distance.
Definition 2.6. The Le Cam distance or ∆-distance between P1 and P2 is defined as
∆(P1 , P2 ) = max(δ(P1 , P2 ), δ(P2 , P1 )).
The ∆-distance is a pseudo-metric on the space of all statistical models: It satisfies the triangle inequality ∆(P1 , P3 ) ≤ ∆(P1 , P2 ) + ∆(P2 , P3 ) but the equality
∆(P1 , P2 ) = 0 does not imply that P1 and P2 actually coincide.
Concerning the glossary, when δ(P1 , P2 ) = 0 (i.e. if the experiment P2 can be
reconstructed from the experiment P1 by a Markov kernel), we will say that P2 is less
informative than P1 , or that P1 is better than P2 , or that P1 is more informative
than P2 . When ∆(P1 , P2 ) = 0 the models P1 and P2 are said to be equivalent and
two sequences of statistical models (P1,n )n∈N and (P2,n )n∈N are called asymptotically
equivalent when ∆(P1,n , P2,n ) → 0 as n → ∞.
A way to interpret the Le Cam distance between experiments is to see it as a numerical
indicator of the cost needed to reconstruct one model from the other one and viceversa, via Markov kernels. But, as we said in the introduction, a way to compare
statistical models that seems just as natural is to compare the respective risk functions.
Let us then highlight how the definition of the deficiency has a clear interpretation in
terms of statistical decision theory. To that aim, we will start by recalling the standard
framework:
• A statistical model, which is just an indexed set P = (X , T , (Pθ : θ ∈ Θ))
of probability measures all defined on the same measurable space (X , T ), for
some set X equipped with a σ-field T . The elements of Θ are sometimes called
the states of Nature.
• A space A of possible actions or decisions that the statistician can take after
observing x ∈ X . For example, in estimation problems we can take A = Θ.
To make sense of the notion of integral on A we need it to be equipped with a
σ-field A.
• A loss function L : Θ × A 7→ (−∞, ∞], with the interpretation that action z ∈ A
incurs a loss L(θ, z) when θ is the true state of Nature.
• A (randomized) decision rule ρ in P is a Markov kernel ρ : (X , T ) → (A, A ).
• The risk is:
Z Z
L(z, θ)ρ(y, dz) Pθ (dy).
Rθ (P, ρ, L) =
X
A
More precisely, the standard interpretation of risk is as follows. The statistician observes a value x ∈ X obtained from a probability measure Pθ . He
does not know the value of θ and must take a decision z ∈ A. He does so by
choosing a probability measure ρ(x, ·) on A and picking a point in A at random according to ρ(x, ·). If he has chosen z when the true distribution of x
is
R Pθ , he suffers a loss L(θ, z). His average loss when x is observed is then
L(θ, z)ρ(x,Rdz).
loss when x is picked according to Pθ is
R His all over average
the integral
L(θ, z)ρ(x, dz) Pθ (dx).
A very important result allowing to translate the notion of deficiency as described
above in a decision theory language is the following:
LE CAM THEORY ON THE COMPARISON OF STATISTICAL MODELS
5
Theorem 2.7 (See Le Cam (1964) or Theorem 2, page 20 in Le Cam (1986)). Let ε > 0
be fixed. δ(P1 , P2 ) < ε if and only if: ∀ decision rule ρ2 on P2 and for all bounded
loss function L, kLk∞ ≤ 1, there exists a decision rule ρ1 on P1 such that
Rθ (P1 , ρ1 , L) < Rθ (P2 , ρ2 , L) + ε,
In other words we have that
∀θ ∈ Θ.
δ(P1 , P2 ) = inf sup sup sup |R(P1 , ρ1 , L, θ) − R(P2 , ρ2 , L, θ)|,
ρ1
ρ2
θ
L
where the last supremum is taken on the set of all loss functions L s.t. 0 ≤ L(θ, z) ≤ 1,
∀z ∈ A, ∀θ ∈ Θ and ρi belongs to the set of all randomised decision procedures in the
experiment Pi , i = 1, 2.
Remark 2.8. An important consequence of the previous theorem is that if two sequences
of experiments (P1,n )n∈N and (P2,n )n∈N are asymptotically equivalent in the Le Cam’s
sense then asymptotic properties of any inference problem are the same for these experiments. This means that when two sequences of statistical experiments are proven
to be asymptotically equivalent it is enough to choose the simplest one, to study there
the inference problems one is interested in and to transfer the knowledge about such
inference problems to the more complicated sequence, via Markov kernels.
2.1. How to transfer decision rules via randomisations. p
Let Pi,n = (Xi,n , Ti,n , (Pi,n,θ : θ ∈ Θ)), i = 1, 2, be two sequences of statistical models sharing the same parameter space Θ and having Polish sample spaces (Xi,n , Ti,n ).
Suppose that there exist Markov kernels Kn such that kKn P1,n,θ − P2,n,θ kT V → 0 uniformly on the parameter space. Then, given a decision rule (or an estimator) π2,n on
P2,n we can define a decision rule π1,n on P1,n that, asymptotically, has the same
statistical risk as π2,n . To show that let us start by considering the easier case in which
both Kn and π2,n are deterministic. More precisely, we suppose that Kn is of the form
Kn (A) = IA Sn (x) for all A ∈ T2,n for some functions Sn .
Then, we have (suppressing the index n to shorten notations):
Z
Z
L(θ, π2 (y))P2,θ (dy) ≤
L(θ, π1 (y))P1,θ (dy) −
X2
X1
Z
L(θ, π1 (y))P1,θ (dy) −
X1
+
Z
X2
≤
Z
X1
Z
L(θ, π2 (y))KP1,θ (dy)
X2
L(θ, π2 (y)) KP1,θ (dy) − P2,θ (dy)
L(θ, π1 (y))P1,θ (dy) −
Z
X1
L(θ, π2 (S(y)))P1,θ (dy) + kLk∞ kKP1 − P2 kT V
In particular, assuming that the loss function L is bounded by 1 and defining
π1 (y) := π2 (S(y))
one finds that
Z
Z
L(θ, π1 (y))P1,θ (dy) −
X2
X1
L(θ, π2 (y))P2,θ (dy) ≤ kKP1 − P2 kT V → 0,
that is, the decision rule π1,n (y) = π2,n (Sn (y)) has asymptotically the same risk as
π2,n . The same kind of computations work in the general case in which the Kn ’s are
not deterministic and (π2,n ) is a sequence of decision rule having (An , An ) as action’s
spaces. In this case one can show that the randomized sequence of decision rules
Z
π2,n (x, C)K(y, dx), ∀y ∈ X1,n , ∀C ∈ An
π1,n (y, C) :=
X2,n
has asymptotically the same risk as π2,n .
6
ESTER MARIUCCI
Remark 2.9. Let Pi be a probability measure on (EiQ
, Ei ) and Ki a Markov kernel on
n
(Gi , Gi ). One can then define a Markov kernel K on ( i=1 Ei , ⊗ni=1 Gi ) in the following
way:
n
Y
Ki (xi , Ai ), ∀xi ∈ Ei , ∀Ai ∈ Gi .
K(x1 , . . . , xn ; A1 × · · · × An ) :=
i=1
Clearly K ⊗ni=1 Pi = ⊗ni=1 Ki Pi .
3. How to control the Le Cam distance
Even if the definition of deficiency has a perfectly reasonable statistical meaning, it
is not easy to compute: Explicit computations have appeared but they are rare (see
Hansen and Torgersen (1974); Torgersen (1972, 1974) and Section 1.9 in Shiryaev and
Spokoiny (2000)). More generally, one may hope to find more easily some upper bounds
for the ∆-distance. We collect below some useful techniques for this purpose.
Property 3.1. Let Pj = (X , T , (Pj,θ ; θ ∈ Θ)), j = 1, 2, be two statistical models
having the same sample space and define ∆0 (P1 , P2 ) := supθ∈Θ kP1,θ − P2,θ kT V . Then,
∆(P1 , P2 ) ≤ ∆0 (P1 , P2 ).
In particular, Property 3.1 allows us to bound the ∆-distance between statistical
models sharing the same sample space by means of classical bounds for the total variation
distance. To that aim, we collect below some useful (and classical) results.
Fact 3.2 (see Le Cam 1969, p. 35). Let P1 and P2 be two probability measures on X ,
i
dominated by a common measure ξ, with densities gi = dP
dξ , i = 1, 2. Define
Z
|g1 (x) − g2 (x)|ξ(dx),
L1 (P1 , P2 ) =
X
H(P1 , P2 ) =
Z
X
Then,
1/2
2
p
p
g1 (x) − g2 (x) ξ(dx)
.
1
H 2 (P1 , P2 )
≤ kP1 − P2 kT V = L1 (P1 , P2 ) ≤ H(P1 , P2 ).
2
2
An important property is the following:
Property
ν are product measures defined on the same measurable space,
Nm 3.3. If µ and
Nm
µ = j=1 µj and ν = j=1 νj , then
m
Y
H 2 (µj , νj )
1−
.
H 2 (µ, ν) = 2 1 −
2
j=1
Proof. See, e.g., Zolotarev (1983), p. 279.
Thus one can express the distance between distributions of vectors with independent
components in terms of the component-wise distances. A consequence of Property 3.3
is:
Property
ν are product measures defined on the same measurable space,
Nm 3.4. If µ and
Nm
µ = j=1 µj and ν = j=1 νj , then
H 2 (µ, ν) ≤
m
X
H 2 (µi , νi ).
i=1
Proof. See, e.g., Strasser (1985), Lemma 2.19.
LE CAM THEORY ON THE COMPARISON OF STATISTICAL MODELS
7
Property 3.5. The Hellinger distance between two normal distributions µ ∼ N (m1 , σ12 )
and ν ∼ N (m2 , σ22 ) is:
1/2
2σ1 σ2
(m1 − m2 )2
H 2 (µ, ν) = 2 1 − 2
exp
−
σ1 + σ22
4(σ12 + σ22 )
σ2
(m1 − m2 )2
≤ 2 1 − 12 +
.
σ2
2σ22
Proof. See, e.g., Mariucci (2015), Fact 1.5.
3.1. The likelihood process. Another way to control the Le Cam distance lies in
the deep relation linking the equivalence between experiments to the proximity of the
distributions of the related likelihood ratios.
Let Pj = (Xj , Tj , (Pj,θ : θ ∈ Θ)) be a statistical model dominated by Pj,θ0 , θ0 ∈ Θ,
dPj,θ
and let Λj (θ) = dPj,θ
be the density of Pj,θ with respect to Pj,θ0 . In particular, one can
0
see Λj (θ) as a real random variable defined on the probability space (Xj , Tj ), i.e. one
can see (Λj (θ))θ∈Θ as a stochastic process. For that reason we introduce the notation
ΛPj := (Λj (θ), θ ∈ Θ) and we call ΛPj the likelihood process.
A key result of the theory of Le Cam is the following.
Property 3.6. Let Pj = (Xj , Tj , (Pj,θ : θ ∈ Θ)), j = 1, 2, be two experiments. If the
family (Pj,θ : θ ∈ Θ) is dominated by Pj,θ0 , then P1 and P2 are equivalent if and only
if their likelihood processes under the dominating measures P1,θ0 and P2,θ0 coincide.
Proof. see Strasser (1985), Corollary 25.9.
Let us now suppose that there are two processes (Λn,∗
j (θ))θ∈Θ , j = 1, 2 defined on a
same probability space (X ∗ , T ∗ , Π∗ ) and such that the law of (Λnj (θ))θ∈Θ under Pj,θ0
∗
is equal to the law of (Λn,∗
j (θ))θ∈Θ under Π , j = 1, 2. Then, the following holds (see
Le Cam and Yang (2000), Lemma 6).
Property 3.7. If ΛP1 and ΛP2 are the likelihood processes associated with the experiments P1 and P2 , then
n,∗
∆(P1n , P2n ) ≤ sup EΠ∗ Λn,∗
1 (θ) − Λ2 (θ) .
θ∈Θ
3.2. Sufficiency and Le Cam distance. A very useful tool, when comparing statistical models having different sample spaces, is to look for a sufficient statistic. The
introduction of the term sufficient statistic is usually attributed to R.A. Fisher who gave
several definitions of the concept. We cite here the presentation of the subject from Le
Cam (1964). Fisher’s most relevant statement seems to be the requirement “...that the
statistic chosen should summarize the whole of the relevant information supplied by the
sample.” Such a requirement may be made precise in various ways, the following three
interpretations are the most common.
(i) The classical, or operational definition of sufficiency, claims that a statistic S is
sufficient if, given the value of S, one can proceed to a post-experimental randomization reproducing variables which have the same distributions as the originally
observable variables.
(ii) The Bayesian interpretation. A statistic S is sufficient if for every a priori distribution of the parameter the a posteriori distributions of the parameter given S is
the same as if the entire result of the experiment was given.
(iii) The decision theoretical concept. A statistic S is sufficient if for every decision
problem and every decision procedure made available by the experiment there
is a decision procedure, depending on S only, which has the same performance
characteristics.
8
ESTER MARIUCCI
The study of sufficiency in an abstract way can be found in Halmos and Savage (1949).
The last section of such a work is named “The value of sufficient statistics in statistical
methodology" and starts with the following observation:
We gather from conversations with some able and prominent mathematical statisticians that there is doubt and disagreement about just what
a sufficient statistic is sufficient to do, and in particular about in what
sense if any it contains “all the information in a sample".
In Bahadur (1954) a continuation of the work of Halmos and Savage (1949) can be
found. A particular effort was done to highlight the interest of using sufficient statistics
in statistical methodology. One of the main results in Bahadur (1954) is Theorem 7.1
establishing the equivalence of the decision theoretical concept of sufficiency and the
operational concept in terms of conditional probabilities. We mention this fact because
of its similarity with the result of Le Cam, here stated as Theorem 2.7, that is the core
of the theory of comparison of statistical experiments.
Formally, let P = (X , T , (Pθ : θ ∈ Θ)) be a statistical model. A statistic is a
measurable map from a measurable space (X , T ) to another measurable one (X2 , T2 ).
We denote by S# Pθ the image law of S under Pθ , i.e S# Pθ (B) = Pθ (S −1 (B)), for all
B ∈ T2 .
Definition 3.8. S is a sufficient statistic for (Pθ : θ ∈ Θ) if for any A ∈ T there exists
a function φA , with φA ◦ S T -measurable, such that
Z
Pθ (A ∩ S −1 (B)) =
φA (y)S# Pθ (dy), ∀A ∈ T , ∀B ∈ T2 , ∀θ ∈ Θ.
B
An arbitrary subalgebra T0 of T is said to be sufficient for (Pθ : θ ∈ Θ) if for all A ∈ T
there exists a T0 -measurable function φA such that
Z
φA (x)Pθ (dx), ∀A0 ∈ T0 , ∀θ ∈ Θ.
Pθ (A ∩ A0 ) =
A0
The set {S −1 (B) : B ∈ T2 } is called the subalgebra induced by the statistic S.
Property 3.9 (See, e.g. Bahadur (1954)). A statistic S is sufficient for (Pθ : θ ∈ Θ) if
the subalgebra induced by S is sufficient for (Pθ : θ ∈ Θ).
In accordance with the notation introduced in Section 2, we will state Theorem 7.1
in Bahadur (1954) as follows (recall that (A, A ) denotes the action/decision space.)
Theorem 3.10 (See Theorem 7.1, Bahadur (1954)). If the subalgebra T0 of T is
sufficient for (Pθ : θ ∈ Θ), then for every decision rule ρ : (X , T ) 7→ (A, A ) there
exists a decision rule π : (X , T0 ) 7→ (A, A ) such that
Pθ ρ(C) = Pθ π(C),
∀C ∈ A , ∀θ ∈ Θ.
Before focusing on the relation between the notion of sufficient statistic and the one
of equivalence between statistical models, let us recall the Neyman-Fisher factorization
theorem, a powerful tool for identifying sufficient statistics for a given dominated family
of probabilities. Let (Pθ : θ ∈ Θ) be a family of probabilities on (Ω, T ), absolutely
θ
continuous with respect to a σ-finite measure µ, and denote by pθ := dP
dµ the density.
Theorem 3.11. A statistic S : (Ω, T ) → (X , B) is sufficient for (Pθ : θ ∈ Θ) if and
only if there exists a B-measurable function gθ ∀θ ∈ Θ and a T -measurable function
h 6= 0 such that
pθ (x) = gθ (S(x))h(x), µ-a.s. ∀x ∈ Ω.
An important result linking the Le Cam distance with the existence of a sufficient
statistic is the following:
LE CAM THEORY ON THE COMPARISON OF STATISTICAL MODELS
9
Property 3.12. Let Pi = (Xi , Ti , (Pi,θ : θ ∈ Θ)), i = 1, 2, be two statistical models.
Let S : X1 → X2 be a sufficient statistic such that the distribution of S under P1,θ is
equal to P2,θ . Then ∆(P1 , P2 ) = 0.
Proof. In order to prove that δ(P1 , P2 ) = 0 it is enough to consider the Markov
kernel M : (X1 , T1 ) → (X2 , T2 ) defined as M (x, B) := IB (S(x)) ∀x ∈ X1 and
∀B ∈ T2 . Conversely, to show that δ(P2 , P1 ) = 0 one can consider the Markov kernel
K : (X2 , T2 ) → (X1 , T1 ) defined as K(y, A) = EP2,θ (IA |S = y), ∀y ∈ X2 , ∀A ∈ T1 .
Since S is a sufficient statistic, the Markov kernel K does not depend on θ. Denoting
by S# P1,θ the distribution of S under P1,θ , one has:
Z
Z
KP2,θ (A) = K(y, A)P2,θ (dy) = EP2,θ (IA |S = y)S# P1,θ (dy) = P1,θ (A).
For asymptotic arguments, one also needs an appropriate version of the notion of
sufficiency.
Definition 3.13. Let Pn = (Xn , Tn , (Pn,θ : θ ∈ Θ)) be a sequence of statistical models.
The sequence of subalgebras T˜n of Tn is asymptotically sufficient for (Pn,θ : θ ∈ Θ) if
∆(Pn , Pn|T˜ ) → 0, where Pn|T˜ denotes the restriction of the experiment Pn to T˜n ,
n
n
i.e. Pn| = (Xn , T˜n , (P̃n,θ : θ ∈ Θ)), where P̃n,θ (A) = Pn,θ (A), for all A ∈ T˜n .
T˜n
This is a stronger notion than asymptotic equivalence; indeed, let P1,n and P2,n be
two sequences of experiments having the same parameter space. Then, by the triangle
inequality, it is clear that if there exist two sequences S1,n and S2,n of asymptotically
sufficient statistics in P1,n and P2,n respectively, taking values in the same measurable
space, and such that
sup kS1,n # P1,θ − S2,n # P2,θ kT V → 0
θ∈Θ
as n → ∞,
then the sequences P1,n and P2,n are asymptotically equivalent. We also recall that an
important generalization of the notion of sufficiency is the notion of insufficiency. The
discussion of this concept is beyond the purposes of this paper, the reader is referred
to Le Cam (1974) or Chapter 5 in Le Cam (1986) for an exhaustive treatment of the
subject.
4. Examples
To better understand what is the typical form of an asymptotic equivalence result let
us analyze some examples. As a toy example let us start by considering the following
parametric case.
Example 4.1. Let P1,n be the statistical model associated with the observation of
a vector X of n independent Gaussian random variables N (θ, 1). Here the inference
concerns θ and the parameter space Θ will be an interval of R. Formally
P1,n = (Rn , B(Rn ), (P1,θ : θ ∈ Θ)),
where Pθ is the law of X.
Then, let us denote by P2,n the experiment associated with the observation of the
empirical mean relative to the previous random variables, i.e.
P2,n = (R, B(R), (P2,θ : θ ∈ Θ)),
where P2,θ is the law of a Gaussian random variable N (θ, 1/n). By means of the
Neyman-Fisher factorization
theorem it is easy to see that the application S : Rn → R,
Pn
i=1 xi
S(x1 , . . . , xn ) =
is a sufficient statistic. An immediate application of Property
n
3.12 implies that ∆(P1,n , P2,n ) = 0 for all n.
10
ESTER MARIUCCI
Before passing to some examples in a nonparametric framework, let us recall a result
due to Carter and concerning the asymptotic equivalence between a multinomial and
a Gaussian multivariate experiment. The parameter space will be a subset of Rm and
the reason for which we focus on such a result lies on its being a very useful tool in
establishing global asymptotic equivalence results for density estimation problems.
Example 4.2. Let X = (X1 , . . . , Xm ) be a random vector having a multinomial
distriP
bution of parameters n and (p1 , . . . , pm ) with pi ≥ 0 for i = 1, . . . , m and m
p
i=1 i = 1.
Denote by P the statistical model associated with a multinomial distribution Pθ =
M(n; (θ1 , . . . , θm )) with parameters θ = (θ1 , . . . , θm ) that belong to ΘR ⊂ Rm , a set
consisting of all vectors of probabilities such that
maxi θi
≤ R.
mini θi
The main result in Carter (2002) is a bound of the Le Cam distance between statistical
models associated with multinomial distributions and multivariate normal distributions
with the same means and covariances as the multinomial ones. More precisely, let
us denote by Q the statistical model associated with a family of multivariate normal
distributions Qθ = N (µ, Σ), θ ∈ ΘR , where
µ = (nθ1 , . . . , nθm ), Σ = (σi,j )i,j=1,...,m with σi,j = nθi (1 − θi )δi=j − nθi θj δi6=j .
Theorem 4.3 (see Carter (2002), p. 709). With the notations above,
m ln m
∆(P, Q) ≤ CR √
n
for a constant CR that depends only on R.
Another interesting result contained in Carter (2002) is the approximation of Q by a
Gaussian experiment with independent coordinates. Let us denote by √
Q˜ the statistical
model associated with m independent Gaussian random variables N ( θi , 1/(4n)), i =
1, . . . , m.
Theorem 4.4 (see Carter (2002), p. 717–719). With the notations above,
m
˜ ≤ CR √
∆(Q, Q)
n
for a constant CR that depends only on R.
Let us now consider some examples in a nonparametric framework. More precisely,
we will recall the results of Brown and Low (1996) and Nussbaum (1996) that are the
first asymptotic equivalence results for nonparametric experiments.
Example 4.5. In Brown and Low (1996), the authors consider the problem of estimating
the function f from a continuously observed Gaussian process y(t), t ∈ [0, 1], which
satisfies the SDE
σ(t)
dyt = f (t)dt + √ dWt , t ∈ [0, 1],
n
where dWt is a Gaussian white noise. They find that the statistical model associated
with the continuous observation of (yt ) is asymptotically equivalent to the statistical
model associated with its discrete counterpart, i.e. the nonparametric regression:
yi = f (ti ) + σ(ti )ξi ,
i−1
n ,
i = 1, . . . , n.
The time grid is uniform, ti =
and the ξi ’s are standard normal variables; they
assume that f varies in a nonparametric subset F of L2 [0, 1] defined by some smoothness
conditions and n tends to infinity not too slowly. More precisely, the drift function f (·)
is unknown and such that, for B a positive constant, one has:
sup |f (t)| : t ∈ [0, 1], f ∈ F = B < ∞.
LE CAM THEORY ON THE COMPARISON OF STATISTICAL MODELS
Moreover, defining
f¯n (t) =
one asks:
if i−1
f ni
n ≤t<
f (1) if t = 1;
lim sup n
n→∞ f ∈F
Z
0
1
i
n,
11
i = 1, . . . , n;
(f (t) − f¯n (t))2
dt = 0.
σ 2 (t)
The diffusion coefficient σ 2 (·) > 0 is supposed to be a known absolutely continuous
function on [0, 1] such that
d
ln σ(t) ≤ C,
dt
t ∈ [0, 1],
for some positive constant C.
Example 4.6. In Nussbaum (1996) the author establishes a global asymptotic equivalence between the problem of density estimation from an i.i.d. sample and a Gaussian
white noise model. More precisely, let (Yi )ni=1 be i.i.d. random variables with density
f (·) on [0, 1] with respect to the Lebesgue measure. The densities f (·) are the unknown
parameters and they are supposed to belong to a certain nonparametric class F subject
to a Hölder restriction: |f (x) − f (y)| ≤ C|x − y|α with α > 12 and a positivity restriction: f (x) ≥ ε > 0. Let us denote by P1,n the statistical model associated with the
observation of the Yi ’s. Furthermore, let P2,n be the experiment in which one observes
a stochastic process (yt )t∈[0,1] such that
p
1
dyt = f (t)dt + √ dWt , t ∈ [0, 1],
2 n
where (Wt )t∈[0,1] is a standard Brownian motion. Then the main result in Nussbaum
(1996) is that ∆(P1,n , P2,n ) → 0 as n → ∞. This is done by first showing that the
result holds for certain subsets Fn (f0 ) of the class F described above. Then it is shown
that one can estimate the f0 rapidly enough to fit the various pieces together. Without
entering into any detail, let us just mention that the key steps are a Poissonization
technique and the use of a functional KMT inequality.
In the last years, asymptotic equivalence results have also been established for discretely observed stochastic processes. As an example, let us present the result in
Mariucci (2015), very close in spirit to the one of Brown and Low (1996).
Example 4.7. Let {Xt }t≥0 be a sequence of one-dimensional time inhomogeneous
jump-diffusion processes defined by
Z t
Z t
Nt
X
Yi , t ∈ [0, T ],
f (s)ds + εn
Xt = η +
σ(s)dWs +
0
0
i=1
where:
• η is some random initial condition;
• W = {Wt }t≥0 is a standard Brownian motion;
• N = {Nt }t≥0 is an inhomogeneous Poisson process with intensity function λ(·),
independent of W ;
• (Yi )i≥1 is a sequence of i.i.d. real random variables with distribution G, independent of W and N ;
• σ 2 (·) is supposed to be known. The horizon of observation T is finite and εn → 0
as n → ∞.
• f (·) belongs to some non-parametric class F .
• λ(·) and G are also unknown and belong to non-parametric classes Λ and G ,
respectively.
12
ESTER MARIUCCI
In Mariucci (2015), the problem of estimating f from high frequency observations of
{Xt }t≥0 is considered. More precisely, we suppose to observe
{Xt }t≥0 at discrete times
0 = t0 < t1 < · · · < tn = T such that ∆n = max1≤i≤n |ti − ti−1 | ↓ 0 as n goes to
infinity. Let Pn be the statistical model associated with the continuous observation of
{Xt }t∈[0,T ] and Qn the one associated with the observations (Xti )ni=0 . Finally, let Wn
be the Gaussian white noise model associated with the continuous observation of the
Gaussian process
dyt = f (t)dt + εn σ(t)dWt ,
y0 = η,
t ∈ [0, T ].
Suppose that F is a subclass of α-Hölder, uniformly bounded functions on R and the
nuisance parameters σ(·) and λ(·) satisfy the following conditions:
• There exist two constants m and M such that 0 < m ≤ σ(·) ≤ M < ∞ and σ(·)
is derivable with derivative σ ′ (·) in L∞ (R).
• There exists a constant L < ∞ such that for all λ ∈ Λ, kλkL2 ([0,T ]) < L.
−2
Then, under the assumption that ∆2α
n εn → 0 as n → ∞, the three models Pn , Qn
and Wn are asymptotically equivalent. A bound for ∆(Qn , Wn ) and ∆(Pn , Qn ) is given
by
−2
+ T ∆2α
∆β/2
n εn + T ∆n ,
n
where β = 1 if G is a subclass of discrete distributions with support on Z and β = 1/2 if G
is a subclass of absolutely continuous distributions with respect to the Lebesgue measure
on R with uniformly bounded densities on a fixed neighborhood of 0. In particular, this
result tells us that the jumps of the process {Xt }t≥0 can be ignored when the goal is
the estimation of the drift function f (·). Moreover, the proof is constructive: an explicit
Markov kernel is constructed to filter the jumps out.
5. Density estimation problems and Gaussian white noise models: A
constructive proof
In this Section, following Carter (2002) (see p. 720-725), we will detail how one can
prove, in a constructive way, the asymptotic equivalence between a density estimation
problem and a Gaussian white noise model, as presented in Example 4.6. However, with
respect to the work of Nussbaum (1996), we will ask some stronger hypotheses on the
parameter space in order to simplify the proofs. More precisely, for fixed γ ∈ (0, 1] and
K, ε, M strictly positive constants, we will consider a functional parameter space of the
form
n
o
F(γ,K,ε,M) = f ∈ C 1 (I) : ε ≤ f (x) ≤ M, |f ′ (x) − f ′ (y)| ≤ K|x − y|γ , ∀x, y ∈ [0, 1] .
As in Example 4.6, P1,n will be a density estimation problem:
(1)
(Yi )1≤i≤n i.i.d. r.v. with density f ∈ F(γ,K,ε,M)
and P2,n a Gaussian with noise model:
p
1
dyt = f (t)dt + √ dWt ,
2 n
t ∈ [0, 1],
f ∈ F(γ,K,ε,M) .
The idea of Carter was to use the bound on the distance between multinomial and
Gaussian normal variables as presented in Example 4.2 to make assertions about density estimation experiments. The intuition is to see the multinomial experiment as the
result of grouping independent observations from a continuous density into m subsets,
say Ji , i = 1, . . . , m. Using the square root as a variance-stabilizing transformation,
these multinomial variables can be asymptotically approximated by m normal variables
with constant variances. These normal variables, in turn, are approximations to the
increments of the process (yt ) over the sets Ji . In Subsection 5.1 we will analyze how
LE CAM THEORY ON THE COMPARISON OF STATISTICAL MODELS
13
to obtain a asymptotically equivalent multinomial experiment starting from P1,n . Assuming the results of Carter stated here as Theorems 4.3 and 4.4 we will then obtain
a bound of the ∆-distance between such a multinomial experiment and one associated
with independent Gaussian random variables. In Subsection 5.2 we will explain how
to show the asymptotic equivalence between an adequate normal approximation with
independent coordinates and P2,n .
5.1. Density estimation problems and multinomial experiments. Let us consider a partition of [0, 1] in m intervals Ji = [(i − 1)/m, i/m] and denote by S :
[0, 1]n → {1, . . . , n}m the application mapping the n-tuple (x1 , . . . , xn ) to the m-tuple
(#{j : xj ∈ J1 }, . . . , #{j : xj ∈ Jm }), where the writing #{j : xj ∈ Ji } stands for the
number of xj belonging to the interval Ji . Let Pf⊗n be the law of (Y1 , . . . , Yn ) as in (1).
R
The law of S under Pf⊗n is a multinomial distribution M (n; θ1 , . . . , θm ), θi = Ji f (x)dx,
i = 1, . . . , m. In particular this means that an appropriate multinomial experiment is
more informative than P1,n . More precisely, we have proven that the statistical model
associated with the multinomial distribution (M (n; θ1 , . . . , θm ) : f ∈ F ), denoted by
Pm , is such that δ(P1,n , Pm ) = 0.
Let us now investigate the quantity δ(Pm , P1,n ). A trivial observation is that the
total variation distance between the multinomial distribution M (n; θ1 , . . . , θm ) and the
law Pf⊗n is always 1, hence, in order to prove that δ(Pm , P1,n ) → 0 we need to construct
a non trivial Markov kernel. We will divide the proof in three main steps.
Step 1: We denote by x∗i the midpoints of the intervals Ji , i.e. x∗i = 2i−1
2m , and
we introduce a discrete random variable X ∗ concentrated at the points x∗i with masses
θi . Let us then denote by P ∗ the statistical model associated with the observation of
n independent realizations of X ∗ . Then, by means of a “sufficient statistic” argument
we can get ∆(Pm , P ∗ ) = 0. Indeed, consider the application S : {x∗1 , . . . , x∗m }n →
{1, . . . , n}m mapping (y1 , . . . , yn ) to (#{j : yj = x∗1 }, . . . , #{j : yj = x∗m }) and observe
that the density h of the law of n independent realizations of X ∗ with respect to the
counting measure is given by
h(y1 , . . . , yn ) =
n
Y
#{j:yj =x∗
1}
P(X ∗ = yi ) = θ1
i=1
#{j:yj =x∗
m}
· · · · · θm
.
By means of the Neyman-Fisher factorization theorem, we conclude that S is a sufficient
statistic, thus ∆(Pm , P ∗ ) = 0.
Step 2: Starting from n realizations of X ∗ we want to obtain something close to
n independent realizations of Pf , the law of Y1 as in (1). To that aim we define an
approximation of f as follows:
fˆm (x) =
m
X
Vj (x)θj ,
j=1
where Vj ’s are piecewise linear functions interpolating the values in the points x∗j as in
Figure 1.
Figure 1. The definition of the Vj functions.
V1
m
0 x∗1
x∗2
Vj
m
1
0 x∗j−1 x∗j
x∗j+1
Vm
m
1
0
x∗m−1 x∗m 1
14
ESTER MARIUCCI
In particular, fˆm is a piecewise linear function that
mθ1 I[0,x∗1 ] (x),
ˆ
fm (x) =
m − m2 |x − x∗i | I[x∗i ,x∗i+1 ] (x)
mθm I[x∗m ,1] (x),
We then consider the Markov kernel
Z
m
X
I{x∗j } (k)
Vj (y)dy,
M (k, A) =
A
j=1
can be written as
if i = 1,
if i ∈ {2, . . . , m − 1},
if i = m.
∀k ∈ N, ∀A ∈ B([0, 1]).
Denoting by P ∗ the law of the random variable X ∗ , we have
∗
M P (A) =
X
∗
M (k, A)P(X = k) =
m
X
θi M (x∗i , A)
=
i=1
k∈N
m
X
θi
i=1
Z
Vi (y)dy =
Z
fˆm (y)dy.
A
A
Let P̂m be the statistical model associated with the observation of n i.i.d. random
variables (Ŷi )1≤i≤n having fˆm as a density with respect to the Lebesgue measure on
[0, 1]. Applying Remark 2.9 we get δ(P ∗ , P̂m ) = 0.
Step 3: We are only left to check that δ(P̂m , P1,n ) → 0. This is actually the case
and we can show that
√
∆(P1,n , P̂m ) = O n m−3/2 + m−1−γ .
Indeed, the total variation distance between the family of probabilities associated with
√
the experiments P1,n and P̂m is bounded by nH(f, fˆm ). Since f (x) ≥ ε for all
x ∈ [0, 1] one can write:
2
Z 1
f (x) − fˆm (x)
q
dx
H 2 (f, fˆm ) =
p
0
f (x) + fˆm (x)
Z 1
2
1
≤
f (x) − fˆm (x) dx.
4ε 0
R1
In order to control the L2 distance between f and fˆm we will split 0 f (x) −
2
fˆm (x) dx as follows:
Z
0
1
2
f (x) − fˆm (x) dx =
Z
1/2m
0
+
Z
1
2
f (x) − mθ1 dx +
1−1/2m
Z
1−1/2m
1/2m
2
f (x) − mθm dx.
2
f (x) − fˆm (x) dx+
2
R
An application of the mean theorem gives Ji f (x) − mθi dx = O(m−3 ), i = 1, . . . , m.
R 1−1/2m
2
To control the term 1/2m
f (x) − fˆm (x) dx, let us consider the Taylor expansion of
f at points x∗i , where x denotes a point in Ji , i = 2, . . . , m − 1:
(2)
f (x) = f (x∗i ) + f ′ (x∗i )(x − x∗i ) + Ri (x).
The smoothness condition on f allows us to bound the error Ri as follows:
|Ri (x)| = f (x) − f (x∗i ) − f ′ (x∗i )(x − x∗i )
(3)
= f ′ (ξi ) − f ′ (x∗i ) |ξi − x∗i | ≤ Km−1−γ ,
where ξi is a certain point in Ji .
By the linear character of fˆm , we can write:
′
fˆm (x) = fˆm (x∗i ) + fˆm
(x∗i )(x − x∗i )
LE CAM THEORY ON THE COMPARISON OF STATISTICAL MODELS
15
′
where fˆm
denotes the left or right derivative of fˆm in x∗i depending whether x < x∗i
′
∗
(x∗i ) = f ′ (χi ) for some χi ∈ Ji ∪ Ji+1 (here, we are
or x > xi . Let us observe that fˆm
considering right derivatives; for left ones, this would be Ji−1 ∪ Ji ), indeed:
′
fˆm
(x∗i ) = −m(fˆm (x∗i ) − fˆm (x∗i+1 )) = −m2
=m
2
Z
i
m
i−1
m
Z
i
m
i−1
m
f (s + 1/m) − f (s) ds = m
Z
f (s)ds −
Z
i+1
m
f (s)ds
i
m
f ′ (ξs )ds
Ji
for some ξs ∈ [s, s + 1/m]. Applying the mean theorem to the function g(s) = f ′ (ξs )
R
1 ′
′
f (t) for some t ∈ Ji ∪ Ji+1 . The fact that fˆm
(x∗i ) = f ′ (t),
we get that Ji f ′ (ξs )ds = m
allows us to exploit the Hölder condition. Indeed, if x ∈ Ji , i = 1, . . . , m, then there
exists t ∈ Ji ∪ Ji+1 such that:
|f (x) − fˆm (x)| ≤ |f (x∗i ) − fˆm (x∗i )| + |f ′ (x∗i ) − f ′ (t)||x − x∗i | + |Ri (x)|
≤ |f (x∗i ) − fˆm (x∗i )| + K|t − x∗i ||x − x∗i |γ + Km−1−γ
≤ |f (x∗i ) − fˆm (x∗i )| + 3Km−1−γ .
R
Using (2) and the fact that Ji (x − x∗i )dx = 0, we get:
Z
∗
∗
ˆ
f (xi ) − fm (xi ) = m
f (x∗i ) − f (x) dx ≤ Km−1−γ .
Ji
Collecting all the pieces together we find
Z 1
2
f (x) − fˆm (x) dx = O m−3 + m−2γ−2 ,
0
hence we can conclude that ∆(P1,n , P̂m ) = O
√
n(m−3/2 + m−1−γ ) .
5.2. Independent Gaussian random variables and Gaussian white noise experiments. In Subsection 5.1 we have seen how to reduce a density estimation problem
to an adequate multinomial experiment. An application of the results of Carter (2002)
recalled in Example 4.2 allows us to obtain an asymptotic equivalence between the statistical model associated with the observation of n i.i.d. random variables of density
f : [0, 1] → R with respect to the Lebesgue measure and an experiment√ in which
one observes m = mn Gaussian and independent random variables N ( θi , 1/4n),
i = 1, . . . , m. Of course, such a Gaussian experiment is equivalent to Nmq
, the stat
θi
1
istical model associated with independent Gaussian random variables N
m , 4nm ,
i = 1, . . . , m. We claim that Nm is asymptotically equivalent to the white noise model
P2,n associated with the continuous observation of a trajectory of a Gaussian process
(yt )t∈[0,1] solution of the SDE:
(4)
dyt =
p
1
f (t)dt + √ dWt ,
2 n
t ∈ [0, 1],
where (Wt )t is a standard Brownian motion. We will divide the proof in two steps.
Denote by Nm∗ the statistical model associated with the observation of (yt )t over the
∗
intervals Ji , i = 1, . . . , m, i.e. N
associated with m independent
R mpis the experiment
1
, i = 1, . . . , m. Firstly, we will show
Gaussian random variables N Ji f (s)ds, 4nm
that Nm is asymptotically equivalent to Nm∗ , then that observing (yt )t is asymptotically
equivalent to observing its increments.
16
ESTER MARIUCCI
Step 1: By means of Property 3.5 we get
v
r 2
u m Z
p
uX
√
θi
∗
t
f (t)dt −
∆(Nm , Nm ) ≤ 2mn
m
Ji
i=1
v
um
Z
2
p
p
√ uX
1
t
m
= 2n
f (t) − mθi dt
m
Ji
i=1
R p
√
f (t) − mθi dt|. By the triangular inequality, we bound
Denote by Ei = |m Ji
Ei by Fi + Gi where:
Z p
q
q
p
and Gi =
Fi =
mθi − f (x∗i )
f (x∗i ) − m
f (y)dy .
Ji
Using the same trick as in Step 3 of Subsection 5.1, we can bound:
Z
|mθi − f (x∗i )|
m
1
|mθi − f (x∗i )|
∗
√
√
p
≤
=
f
(s)
−
f
(x
)
ds = √ kRi k∞ ,
Fi = √
i
2 ε
2 ε Ji
2 ε
mθi + f (x∗i )
R
where we have used the fact that Ji (x − x∗i ) = 0 and Ri denotes the remainder in the
Taylor expansion of f in x∗i , as in (2). On the other hand,
Z q
p
Gi = m
f (x∗i ) − f (y) dy
Jj
f ′ (x∗i )
∗
p
(x − xi ) + R̃i (y) dy ≤ kR̃i k∞ ,
=m
2 f (x∗i )
Ji
√
∗
where R̃i is the remainder in the Taylor expansion
√ of f in xi . We observe that if f
belongs to the functional class F(γ,K,ε,M) then f is still bounded away from zero and
√
from infinity with a Hölder continuous derivative, more precisely f ∈ F(γ,K/√ε,√ε,√M) .
q
In particular, we deduce that kR̃i k∞ has the same magnitude as M
ε kRi k∞ . Thanks to
Z
(3) we know that kRi k∞ ≤ Km−1−γ for any i = 2, . . . , m − 1 and kRi k∞ = O(m−3/2 )
for i ∈ {1, m}. Hence the quantities Fi and Gi are of the same order and we find that
√
3
∆(Nm , Nm∗ ) = O n m−1−γ + m− 2 .
Step 2: Since Nm∗ is the model associated with the observation of the increments
(Ȳi )1≤i≤n of the process (yt )t defined as in (4) it is clear that δ(P2,n , Nm∗ ) = 0. Let
us now discuss how to bound δ(Nm∗ , P2,n ). We start by introducing a new stochastic
process:
Z t
m
m
X
1 X
Ȳi
Vi (y)dy + √
yt∗ =
Bi (t), t ∈ [0, 1],
2 nm i=1
0
i=1
where the functions Vi are defined as in Figure 1 and Bi (t) are independent centered
Gaussian processes independent of (Wt ) and with variances
Z t
2
Z t
Var(Bi (t)) =
Vi (y)dy −
Vi (y)dy .
0
0
These processes can be constructed from a standard Brownian bridge B(t), independent
of (Wt ), via
Z t
Bi (t) = B
Vi (y)dy .
0
REFERENCES
17
By construction, (yt∗ ) is a Gaussian process with mean and variance given by, respectively:
Z t
Z t
m Z p
m
X
X
∗
E[Ȳi ]
Vi (y)dy =
E[yt ] =
f (y)dy
Vi (y)dy,
Var[yt∗ ] =
i=1
m
X
0
Var[Ȳi ]
i=1
=
1
4nm
Z
Z tX
m
i=1
2
t
Vi (y)dy
0
Vi (y)dy =
0 i=1
Ji
+
t
.
4n
1
4nm
0
m
X
Var(Bi (t))
i=1
One can compute in the same way the covariance of (yt∗ ) and deduce that
Z tp
Z t
1
d
∗
√ dWs∗ , t ∈ [0, 1],
Yt =
f m (y)dy +
2
n
0
0
where (Wt∗ ) is a standard Brownian motion and
m Z p
X
p
d
f m (x) :=
f (y)dy Vi (x).
i=1
Ji
Applying Fact 3.5, we get that the total variation distance between the process
(yt∗ )t∈[0,1] constructed from the random variables Ȳi , i = 1, . . . , m and the Gaussian
process (yt )t∈[0,1] is bounded by
s
Z 1 p
p
2
d
4n
f m (y) − f (y) dy.
0
√
Since f ∈ F(γ,K,ε,M) implies f ∈ F(γ,K √M/√ε,√ε,√M ) , the same kind of computations
made in Step 3 of Subsection 5.1 allows us to conclude that
√
∆(Nm∗ , P2,n ) = δ(Nm∗ , P2,n ) = O n m−3/2 + m−1−γ .
5.3. The choice of m. In Subsection 5.1 we have proven that the cost needed to pass
from the model associated to the observation of n i.i.d. random variables with unknown
density f ∈ F(γ,K,ε,M) to an adequate multinomial approximation M(n; θ1 , . . . , θm )
√
is of the order of n m−3/2 + m−1−γ . Using Theorem 4.3 we can take a further
step obtaining a Gaussian approximation (with independent coordinates) starting from
the multinomial one. This comes to the price of m√lnnm . Finally, in Subsection 5.2
we have found that for appropriate choices of m there is an asymptotic equivalence
between such a Gaussian approximation and the Gaussian with noise model P2,n . A
bound
for the rate of convergence of the ∆-distance up to constants is, again, given by
√
n m−3/2 + m−1−γ . In particular we deduce that
(
γ
O n− 2(γ+2) log n ,
if 0 < γ ≤ 21 ,
∆(P1,n , P2,n ) =
1
if 12 < γ ≤ 1.
0(n− 10 log n)
after the choice m = n1/(2+γ) .
References
Bahadur, R. R. (1954). ‘Sufficiency and statistical decision functions’. In: Ann. Math.
Statistics 25, pp. 423–462.
Blackwell, D. (1951). ‘Comparison of experiments’. In: Proceedings of the Second Berkeley Symposium on Mathematical Statistics and Probability, 1950. Berkeley and Los
Angeles: University of California Press, pp. 93–102.
18
REFERENCES
Blackwell, D. (1953). ‘Equivalent comparisons of experiments’. In: Ann. Math. Statistics
24, pp. 265–272.
Bohnenblust, A., L. Shapley and S. Sherman (1949). ‘Reconnaissance in game theory’.
In: Rand Research Memorandum 1949/208.
Brown, L. D. and M. G. Low (1996). ‘Asymptotic equivalence of nonparametric regression and white noise’. In: Ann. Statist. 24.6, pp. 2384–2398.
Brown, L. D. and C.-H. Zhang (1998). ‘Asymptotic nonequivalence of nonparametric
experiments when the smoothness index is 1/2’. In: Ann. Statist. 26.1, pp. 279–287.
Brown, L. D., T. T. Cai, M. G. Low and C.-H. Zhang (2002). ‘Asymptotic equivalence theory for nonparametric regression with random design’. In: Ann. Statist. 30.3.
Dedicated to the memory of Lucien Le Cam, pp. 688–707.
Brown, L. D., A. V. Carter, M. G. Low and C.-H. Zhang (2004). ‘Equivalence theory for
density estimation, Poisson processes and Gaussian white noise with drift’. In: Ann.
Statist. 32.5, pp. 2074–2097.
Buchmann, B. and G. Müller (2012). ‘Limit experiments of GARCH’. In: Bernoulli 18.1,
pp. 64–99.
Buscemi, F. (2012). ‘Comparison of quantum statistical models: equivalent conditions
for sufficiency’. In: Comm. Math. Phys. 310.3, pp. 625–647.
Carter, A. V. (2002). ‘Deficiency distance between multinomial and multivariate normal
experiments’. In: Ann. Statist. 30.3. Dedicated to the memory of Lucien Le Cam,
pp. 708–730.
– (2006). ‘A continuous Gaussian approximation to a nonparametric regression in two
dimensions’. In: Bernoulli 12.1, pp. 143–156.
– (2007). ‘Asymptotic approximation of nonparametric regression experiments with unknown variances’. In: Ann. Statist. 35.4, pp. 1644–1673.
– (2009). ‘Asymptotically sufficient statistics in nonparametric regression experiments
with correlated noise’. In: J. Probab. Stat. Art. ID 275308, 19.
Dalalyan, A. and M. Reiß (2006). ‘Asymptotic statistical equivalence for scalar ergodic
diffusions’. In: Probab. Theory Related Fields 134.2, pp. 248–282.
– (2007). ‘Asymptotic statistical equivalence for ergodic diffusions: the multidimensional case’. In: Probab. Theory Related Fields 137.1-2, pp. 25–47.
Delattre, S. and M. Hoffmann (2002). ‘Asymptotic equivalence for a null recurrent diffusion’. In: Bernoulli 8.2, pp. 139–174.
Efromovich, S. and A. Samarov (1996). ‘Asymptotic equivalence of nonparametric regression and white noise model has its limits’. In: Statist. Probab. Lett. 28.2, pp. 143–
145.
Genon-Catalot, V. and C. Laredo (2014). ‘Asymptotic equivalence of nonparametric
diffusion and Euler scheme experiments’. In: The Annals of Statistics 42.3, pp. 1145–
1165.
Genon-Catalot, V., C. Laredo and M. Nussbaum (2002). ‘Asymptotic equivalence of
estimating a Poisson intensity and a positive diffusion drift’. In: Ann. Statist. 30.3.
Dedicated to the memory of Lucien Le Cam, pp. 731–753.
Golubev, G. K., M. Nussbaum and H. H. Zhou (2010). ‘Asymptotic equivalence of
spectral density estimation and Gaussian white noise’. In: Ann. Statist. 38.1, pp. 181–
214.
Grama, I. and M. Nussbaum (1998). ‘Asymptotic equivalence for nonparametric generalized linear models’. In: Probab. Theory Related Fields 111.2, pp. 167–214.
– (2002). ‘Asymptotic equivalence for nonparametric regression’. In: Math. Methods
Statist. 11.1, pp. 1–36.
Grama, I. G. and M. H. Neumann (2006). ‘Asymptotic equivalence of nonparametric
autoregression and nonparametric regression’. In: Ann. Statist. 34.4, pp. 1701–1732.
REFERENCES
19
Halmos, P. R. and L. J. Savage (1949). ‘Application of the Radon-Nikodym theorem to
the theory of sufficient statistics’. In: Ann. Math. Statistics 20, pp. 225–241.
Hansen, O. H. and E. N. Torgersen (1974). ‘Comparison of linear normal experiments’.
In: The Annals of Statistics, pp. 367–373.
Jähnisch, M. and M. Nussbaum (2003). ‘Asymptotic equivalence for a model of independent non identically distributed observations’. In: Statist. Decisions 21.3, pp. 197–
218.
Le Cam, L. (1964). ‘Sufficiency and approximate sufficiency’. In: Ann. Math. Statist. 35,
pp. 1419–1455.
Le Cam, L. (1969). Théorie asymptotique de la décision statistique. Séminaire de Mathématiques Supérieures, No 33 (Été, 1968). Les Presses de l’Université de Montréal,
Montreal, Que., p. 140.
Le Cam, L. (1974). ‘On the information contained in additional observations’. In: Ann.
Statist. 2, pp. 630–649.
Le Cam, L. (1986). Asymptotic methods in statistical decision theory. Springer Series in
Statistics. New York: Springer-Verlag, pp. xxvi+742.
Le Cam, L. and G. L. Yang (2000). Asymptotics in statistics. Second. Springer Series in
Statistics. Some basic concepts. New York: Springer-Verlag, pp. xiv+285.
Mariucci, E. (2015). ‘Asymptotic equivalence for inhomogeneous jump diffusion processes and white noise’. In: ESAIM Probab. Stat. 19, pp. 560–577.
– (2016a). ‘Asymptotic equivalence for pure jump Lévy processes with unknown Lévy
density and Gaussian white noise’. In: Stochastic Process. Appl. 126.2, pp. 503–541.
– (2016b). ‘Asymptotic equivalence of discretely observed diffusion processes and their
Euler scheme: small variance case’. In: Stat. Inference Stoch. Process. 19.1, pp. 71–91.
– (To appear). ‘Asymptotic equivalence for density estimation and Gaussian white
noise: an extension’. In: Annales de l’ISUP. ArXiv:1503.05019.
Meister, A. (2011). ‘Asymptotic equivalence of functional linear regression and a white
noise inverse problem’. In: Ann. Statist. 39.3, pp. 1471–1495.
Meister, A. and M. Reiß (2013). ‘Asymptotic equivalence for nonparametric regression
with non-regular errors’. In: Probab. Theory Related Fields 155.1-2, pp. 201–229.
Milstein, G. and M. Nussbaum (1998). ‘Diffusion approximation for nonparametric
autoregression’. In: Probab. Theory Related Fields 112.4, pp. 535–543.
Nussbaum, M. (1996). ‘Asymptotic equivalence of density estimation and Gaussian white
noise’. In: Ann. Statist. 24.6, pp. 2399–2430.
Reiß, M. (2008). ‘Asymptotic equivalence for nonparametric regression with multivariate
and random design’. In: Ann. Statist. 36.4, pp. 1957–1982.
– (2011). ‘Asymptotic equivalence for inference on the volatility from noisy observations’. In: Ann. Statist. 39.2, pp. 772–802.
Rohde, A. (2004). ‘On the asymptotic equivalence and rate of convergence of nonparametric regression and Gaussian white noise’. In: Statist. Decisions 22.3, pp. 235–243.
Schmidt-Hieber, J. (2014). ‘Asymptotic equivalence for regression under fractional noise’.
In: The Annals of Statistics 42.6, pp. 2557–2585.
Shiryaev, A. N. and V. G. Spokoiny (2000). Statistical experiments and decisions. Vol. 8.
Advanced Series on Statistical Science & Applied Probability. Asymptotic theory.
River Edge, NJ: World Scientific Publishing Co. Inc., pp. xvi+281.
Strasser, H. (1985). Mathematical theory of statistics. Vol. 7. de Gruyter Studies in
Mathematics. Statistical experiments and asymptotic decision theory. Berlin: Walter
de Gruyter & Co., pp. xii+492.
Torgersen, E. N. (1972). ‘Comparison of translation experiments’. In: The Annals of
Mathematical Statistics 43.5, pp. 1383–1399.
– (1974). ‘Comparison of experiments by factorization’. In: Stat.Res.Report, Univ. of
Oslo.
20
REFERENCES
van der Vaart, A. (2002). ‘The statistical work of Lucien Le Cam’. In: Ann. Statist. 30.3.
Dedicated to the memory of Lucien Le Cam, pp. 631–682.
Wang, Y. (2002). ‘Asymptotic nonequivalence of Garch models and diffusions’. In: Ann.
Statist. 30.3. Dedicated to the memory of Lucien Le Cam, pp. 754–783.
Zolotarev, V. M. (1983). ‘Probability metrics’. In: Teoriya Veroyatnostei i ee Primeneniya 28.2, pp. 264–287.
Leiden University.
E-mail address: [email protected]
| 10math.ST
|
An Adaptive Algorithm for Precise Pupil Boundary Detection
using Entropy of Contour GradientsI
Cihan Topala , Halil Ibrahim Cakirb , Cuneyt Akinlarb
a Department
of Electrical and Electronics Engineering, Anadolu University, 26470 Eskisehir, Turkey
of Computer Engineering, Anadolu University, 26470 Eskisehir, Turkey,II
arXiv:1709.06366v1 [cs.CV] 19 Sep 2017
b Department
Abstract
Eye tracking spreads through a vast area of applications from ophthalmology, assistive technologies to
gaming and virtual reality. Detection of pupil is the most critical step in many of these tasks hence needs
to be performed accurately. Although detection of pupil is a smooth task in clear sight, possible occlusions
and odd viewpoints complicate the problem. We present an adaptive pupil boundary detection method that
is able to infer whether entire pupil is in clearly visible by a modal heuristic. Thus, a faster detection is
performed with the assumption of no occlusions. If the heuristic fails, a deeper search among extracted
image features is executed to maintain accuracy. Furthermore, the algorithm can find out if there is no
pupil as an aidful information for many applications. We prepare a dataset containing 1509 high resolution
eye images collected from five subjects and perform an extensive set of experiments to obtain quantitative
results in terms of accuracy, localization and timing. The proposed method outperforms three other state
of the art algorithms and can run up to 140 Hz in single-thread on a standard laptop computer.
Keywords:
Pupil detection, eye tracking, elliptical arc detection, ellipse detection, gaze estimation, shape recognition.
1. Introduction
Eye tracking (ET) has emerged as an important
research area with a diverse set of applications including human computer interaction, diagnosis of
psychological, neurological and ophthalmologic individuals, assistive systems for drivers and disabled
people, marketing research, and biometrics. In addition, there are efforts to integrate ET technology
in virtual reality (VR) studies to increase the feeling of immersion via rendering virtual environment
with a depth of field effect similar to human vision.
Pupil boundary detection and center estimation
is an essential step in all eye tracking systems and
I This work is supported by The Scientific and Technological Research Council of Turkey (TUBITAK) and
Anadolu University Commission of Scientific Research
Projects (BAP) under the grant numbers 111E053 and
1207F113, respectively.
II Cuneyt Akinlar was with Department of Computer Engineering, Anadolu University during the time of this study.
Email addresses: [email protected] (Cihan
Topal), [email protected] (Halil Ibrahim
Cakir), [email protected] (Cuneyt Akinlar)
Preprint submitted to Elsevier
has to be performed precisely. In point-of -gaze
(PoG) detection, the extraction of pupil center is
required to estimate the location of gaze. In such
applications, even loss of a single pixel precision in
pupil center may cause an error of a few degrees
in the gaze direction vector, which would result
in a significant drift in the estimated gaze point.
Pupil boundary detection is a more difficult problem which has to be performed accurately for biometric applications and medical studies.
Another emerging application area of eye tracking is virtual reality (VR) technologies which recently have a significant leap in popularity. VR
technology renders a 3D scene from two different
point of views, ie. from the views of left and right
eyes of a user. To prevent problems like motions
sickness, these rendering 3D locations should match
the interpupillary distance of the user. Moreover,
developers integrate eye tracking into VR systems
to better simulate the human visual system. They
render locations where the user focuses sharper and
blur other regions to boost immersion effect and enhance the VR experience.
September 20, 2017
In this study, we propose a novel adaptive pupil
boundary detection method from eye images, which
works by extracting arcs from the edge segments
of the image and joining them to find the pupil
boundary, hence the center. Organization of the
paper is as follows; we give a comprehensive related work on pupil boundary and center detection
in section 2, we explain our method with fine detail
in section 3, analysis of the method in terms of accuracy (F-measure), localization and running time
are presented in section 4, and we finalize the paper
with concluding remarks.
boundaries aiming to maximize the contour integral
value on the smoothed image derivative (Eq. 1).
max(r,x0 ,y0 ) Gσ (r) ∗
∂
∂r
I
(r,x0 ,y0 )
I(x, y)
ds
2πr
(1)
Arvacheh and Tizhoosh [16] developed an iterative
algorithm based on an active counter model which
is also capable of detecting near-circular shapes.
These methods work fine; however, they require
a full search of the image plane in order to find
r, x0 , y0 parameters that maximize the response to
the given model. Model-based search approach is
computationally expensive and therefore, cannot be
employed in real-time eye tracking applications.
In [17] authors use curvature of pupil contour to
sort out boundary pixels which belong to prospective occlusions. They detect blobs in the binarized
image and extract contour of the biggest blob. Finally, edge pixels of the pupil boundary are selected
by employing a set of heuristics (e.g. eyelids have
positive curvature, etc.) and ellipse fit applied to
chosen pixels.
Another interesting approach on pupil detection
is proposed and utilized in EyeSeeCam project [18].
The algorithm extracts edge segments, then removes glints and other unfavourable artefacts by a
sequence of morphological operations based on several assumptions. Finally, Delaunay Triangulation
is applied to remaining pixels and pupil boundary
is detected assuming it is a convex hull.
Starburst algorithm [19] estimates the pupil center by an iterative radial feature detection technique instead of finding all edges. It starts by locating and removing glints if any exists. Then, rays are
cast from an initial point within a 20◦ of radial step.
Each ray stops where image derivative is greater
than a threshold value, i.e., when a sharp intensity
change occurs. This operation is iterated with an
updated starting point and a set of feature points
are collected at each step. Finally, ellipse fit is applied to the collected points with RANSAC [11].
In another study, authors aim to adapt the Starburst algorithm to elliptical iris segmentation problem [20].
In [21], Świrski et al. approximately detects
pupil region by a Haar-like feature [23, 11]. Secondly, they apply a k -means segmentation to determine a proper pupil threshold. Then a modified
RANSAC-based ellipse fitting method is employed
which utilizes gradient information as well as the
2. Related Work
The literature on pupil detection is very rich, and
many different techniques have been proposed. In
this section, our goal is to give a high-level picture of
the proposed solutions for pupil boundary detection
and/or pupil center estimation.
Many early methods in the literature utilize discriminative visual structure of human eye to detect the pupil. Dark intensity of pupil region and
its high contrast between bright sclera region offers a relatively easy way to solve the problem.
In this manner, many algorithms extracts pupil
(or iris in some studies) center with combinations
of several methods like thresholding, morphological operations, connected component analysis and
center of mass algorithms with various additional
steps [1, 2, 3, 4, 5]. In addition, there are methods which also benefit model fitting approaches
to find pupil or iris boundary as a circle or ellipse [6, 7, 8, 9]. In these studies, edge and contour extraction is employed and followed by Hough
transform or RANSAC algorithm to accurately estimate the boundary [10, 11].
Dark/bright pupil differencing is another approach to roughly detect eye locations in a remotely
taken image [12, 13, 14]. It works by differencing
two successive frames that captured with on-axis
and off-axis illumination, respectively. Due to the
physical structure of human eye, on-axis illumination causes a significant brightness inside the pupil.
Therefore, pupil regions become more salient in the
difference image.
Along with the feature-based methods, there
are also purely model-based approaches which are
mostly utilized in iris recognition studies in the
literature. Daugman [15] proposes an integrodifferential operator for detecting the pupil and iris
2
Lin et al. [5]
•
Wang and Sun [6]
Ma et al. [7]
Dey and Samanta [8]
•
•
4
, ∅
•
•
•
•
•
•
vertical
ellipse
•
Hough
•
for edges
•
1
ellipse
•
•
•
Hiley et al. [13]
•
•
iterative
•
Morimoto et al. [14]
•
•
•
•
•
Fuhl et al [22]
•
∅
∅
2
ellipse
,3
for edges
radial
ellipse
4
•
•
iterative
,5
•
•
ellipse
6
Li et al. [19]
k-means
∅
•
•
Świrski et al. [21]
, ∅
circle
•
Kumar et al. [18]
1
4
Ebisawa [12]
Zhu et al. [17]
, ∅,
parallelogram
•
∅
symmetric
for glint
adaptive
Agustin et al. [9]
modified
Other /
Comments
•
Use of
Temporal
Information
Keil et al. [4]
•
Center of Mass
Algorithm
•
•
Circle /
Ellipse Fitting
•
Long et al. [3]
Blob Detection/
Connected
Comp. Analysis
Mäenpää [2]
Gõni et al. [1]
Edge Detection
Morphological
Operations
•
Bright/Dark
Pupil
Differencing
adaptive
Algorithms
Downsampling
Image
Thresholding /
Binarization
Table 1: A brief taxonomy for pupil boundary detection / center estimation algorithms.
∅ does not detect the pupil boundary, only estimates its center.
4 performs iris detection.
performs ROI detection.
1
applies histogram back-projection or non-linear power transform on the image to make the pupil more salient.
2
before ellipse fitting, tries to determine the false pupil contour pixels w.r.t. their curvature values by a set of heuristics.
4
requires removal of glints. Assumes that the initial point for ray casting is inside the pupil. Iterative algorithm.
3
performs Fast Radial Symmetry detection and Delaunay Triangulation. Removes glints and artefacts by a set of morphological assumptions.
tries to find an ellipse that matches the edge image points and is orthogonal to the gradients of the image.
5
6
detects and filter edges. Uses two different approaches, i.e. algorithmic and morphological. Rescales image if it fails in the first attempt.
spatial coordinates to find the pupil boundary.
In a more recent study Fuhl et al. detect edges on
the eye image and filter them with respect to several
morphological criteria [22]. Later, edge segments
are constructed from the remaining edge pixels and
some of the edge segments (i.e. straight lines) are
eliminated by various heuristics. Finally remaining
contours evaluated by ellipse fitting and the best
ellipse is selected by a cost function which utilize
inner gray value and the roundest shape.
Table 1 gives a brief taxonomy of the abovementioned pupil detection algorithms. As seen in
the table, thresholding is a common technique in
the literature. Despite thresholding can quickly discriminate image regions having different intensity
values, it is highly vulnerable to lighting conditions
and parameter configuration. Consequently, it fails
on finding the exact location where intensity change
occurs and can easily causes a decrease on the ac-
curacy. Another frequently employed technique is
morphological operations which are applied on the
thresholded binary image to suppress remaining undesired pixel sets and improve modal structure of
the image. However, morphological operations may
also degrade the actual information on the image
and cause significant errors on the result. Similarly, algorithms that utilize thresholding and blob
detection to find a center point for pupil are obviously not capable of detecting the boundary. Hence,
they cannot be applied on most biometrics or medical studies which requires precise detection of the
boundaries of pupil and iris.
Downsampling the image to save computational
time has an obvious cost as it decreases the accuracy by wasting spatial resolution. Bright/dark
pupil differencing requires a little amount of computation and eases roughly locating the pupil, however
it has important drawbacks. First, it needs addi3
Near-circular
segment
YES
Input
Frame
ROI
Detection
Edge
Segment
Detection
Near-circular
Segment
Search
Corner
Detection
?
Near-Circular
Segment
NO
Ellipse
Fitting
Pupil
Candidate
Generation
Pupil
Detection
Elliptical Arc
All segments Extraction
Figure 1: Processing pipeline of the proposed algorithm.
in the image as a very aidful information for many
applications.
The proposed method has a simple flow and consists of the processing pipeline shown in Fig. 1. The
processing starts by detection of region of interest
(ROI) by convolving eye image with a Haar-like feature. Then, we extract edge segments each of which
is a contiguous array of pixels. The next step is
to determine whether a near-circular segment exists that traces the entire boundary of the pupil.
Such an edge segment would be found if the pupil
is clearly visible with no or very little occlusion. To
find whether an edge segment has a circular geometry, we devise a fast heuristic method which utilizes
gradient distribution of an edge segments. On the
condition that a near-circular segment is found, we
extract elliptical arcs from only that segment. If no
near-circular segment is found, which would be the
case if the pupil is severely occluded by the eyelids
or eyelashes, then arcs from all edge segments in
the ROI are extracted.
Following the extraction of the elliptical arcs, we
join them in every possible combination to generate a set of ellipse candidates that at least one of
them traces the pupil boundary. Candidates are finally evaluated for their relevance to be the actual
pupil contour and the best one, if it exists, is chosen among the candidate ellipses. In the following
subsections we elaborate each step of the proposed
algorithm in fine detail to make the discussion clear.
tional hardware to obtain bright and dark pupil images consecutively in a synchronous manner. Furthermore, it reduces temporal resolution since it
needs two frames to perform a single detection. Due
to the same reason it is very sensitive to motion and
it fails if a large pupil displacement occurs between
two consecutive images. Ebisawa specifically addresses this problem in [24], and proposes various
methods for positional compensation.
In this section we presented an overview of related studies covering both biometrics and eye
tracking areas from the viewpoint of pupil detection problem. For interested readers, there are also
comprehensive surveys that review the gaze estimation literature; in particular [25, 26].
3. Proposed Method
In this study, we propose an adaptive method
for pupil boundary detection which is able to save
computation by inferring whether an occlusion is
the case or not. In this manner, the computation
takes very little time if the pupil is in clear sight by
the camera. On the contrary, algorithm infers if the
pupil is severely occluded, and spends more effort
to detect the pupil without compromising real-time
applicability. The main strategy which improves
the algorithm against occlusions is extracting the
elliptical arcs from input image and finding one arc
or a group of arcs representing the pupil contour.
In this way, relevant features from a partially visible pupil can be extracted and detection can be
performed by fusion of separate features. Besides
detecting the pupil boundary and center precisely,
the algorithm can also identify if there is no pupil
3.1. ROI Estimation
In the first step of the proposed method we
roughly estimate the pupil area in the entire eye
image. For this purpose, we utilize pupil’s geomet4
3R
factor:
3x
=
*
5R
(a)
external r = 4x
internal r = 3x
Figure 3: Sample eye image in which detected ROI is indicated (left), edge segments obtained by EDPF algorithm
within the ROI are indicated in different colors (right).
(b)
edge localization. ED outputs not only a binary
edge map similar to those output by conventional
edge detectors, but it also outputs the result as a
set of edge segments each of which is a contiguous
and connected pixel chain [29]. This property of ED
extremely eases the application of the algorithm to
further detection and recognition problems.
Similar to other edge detectors, ED has several
parameters that must be tuned by the user for different tasks. Ideally, one would want to have an
edge detector which runs with a fixed set of parameters for any type of image. To achieve this goal,
we have incorporated ED with the a contrario edge
validation mechanism due to the Helmholtz principle [30, 31], and obtained a fast parameter-free edge
segment detector, i.e. EDPF2 [32].
EDPF works by running ED with all ED’s parameters at their extremes, which detects all possible edge segments in a given image with many false
positives. It then validates the extracted edge segments by the Helmholtz principle, which eliminates
false detections leaving only perceptually meaningful segments with respect to the a contrario approach with very little overhead of computation.
Fig. 3 illustrates detected edge segments for an example eye image. In the figure, each color represents a different edge segment, which is one-pixel
width, contiguous array of pixels.
Figure 2: (a) ROI detection by means of a convolution operation with a Haar-like feature. (b) Detected ROIs of two
eye images having different size of pupils.
ric and intensity attributes in a similar vein as the
method [21]. A pupil can be described as a dark
and compact blob since it consists of darker intensity levels than its surrounding iris, and it has usually a low-eccentric elliptical shape. To locate pupil
region we convolve input image with a Haar-like feature and find the maximum response on the image
as shown in Fig. 2(a) [23].
Pupil size can vary in real life due to the reasons
like physiological differences and pupil dilation in
light changes. For this reason, the Haar-like feature kernel is applied in various aperture sizes and
the maximum response per unit is picked at the
end. In Fig. 2(b) results of ROI estimation process
for two eye images from two individuals having different pupil sizes are presented.
3.2. Edge Segment Detection
To detect all edge segments inside the ROI, we
employ Edge Drawing (ED) edge segment detector1 [27, 28]. Unlike traditional edge detectors
which work by identifying a set of potential edge
pixels in an image and eliminating non-edge pixels through operations such as morphological operations, non-maximal suppression, hysteresis thresholding etc., ED follows a proactive approach. The
ED algorithm works by first identifying a set of
points in the image, called the anchors, and then
joins these anchors in a way which maximizes the
gradient response of edge paths, hence ensures good
1 Online:
3.3. Near-Circular Segment Search
The main goal of this step is detecting the pupil
in an easy and computation efficient way when its
circumference is entirely visible in case of no occlusion. Once we have the edge segments detected,
we need to find the one that traverses the pupil
2 Demo
http://c-viz.anadolu.edu.tr/EdgeDrawing
5
page: http://c-viz.anadolu.edu.tr/EDPF
|G|
Gy
a few values dominate. Thus, we can distinguish
edge segments in circular shapes by picking the
ones which result in plain gradient distribution. To
achieve this, we use the entropy function (Eq. 2)
on the quantized gradient distributions of the segments.
n
X
E=−
pi . log(pi )
(2)
y
67.5°
α
45°
Gx
22.5°
0°
Gradient operator
x
-22.5°
G
α = tan −1 y
Gx
(a)
-90°
i
-45°
-67.5°
Since entropy function maximizes for flat distributions where the frequency of each symbol is equal,
we compute the entropy of gradient distribution for
each separate edge segments as follows:
(b)
Figure 4: (a) Gradient computation with image derivatives,
(b) Quantization of computed gradient directions.
arg max
8
X
fGi . log(fGi )
(3)
i
boundary. The most intuitive solution is to apply a
brute force search as follows: fit an ellipse to each
edge segment, compute the fitting error, and pick
the edge segment that has the smallest fitting error. This method might work only when the pupil
is clearly visible, i.e., when it is not occluded by
IR-LED glints or eyelashes, however, fitting an ellipse and calculating fitting error for each segment
requires too much computation.
To reduce this computational burden, we devise a
faster method based on the analysis of gradient directions to find the near-circular segment, if one exists. Gradients of the segments contain substantial
information about the geometrical structure and is
used in shape matching, retrieval and recognition
problems [33, 34]. Since we already have the vertical and horizontal derivatives of eye image computed during the edge detection scheme, we can find
the gradient directions by very little amount of computation (see Fig. 4.a). The arctan function obviously results angle values in an interval [−π/2, π/2],
providing an angle range of 180◦ . Before examining
the distribution of gradients, we quantize the angles
with 22.5◦ to obtain discrete symbols, thus we divide the unit circle into 16 regions into 8 different
directions (see Fig. 4.b).
Once we get quantized gradient directions for all
pixels in a segment, we infer the modal characteristics of that segment by analyzing their gradient distributions. It is easy to observe that any segment in
near-circular form would have an even gradient distribution if the tangential gradients on its perimeter
is sampled with a fixed angular step. Intuitively,
circular edge segments would have relatively uniform gradient distribution; whereas, straight edge
segments have an unbalanced distribution where
where fGi is the frequency of the ith gradient direction. The entropy values for edge segments are
maximized for a perfect circle and gets lower as the
segment shape differs from being a circle (elliptic,
etc.), and finally entropy becomes zero for straight
lines since a straight line has only one gradient direction along its trajectory. Since we quantize the
unit circle into 8 direction (see Fig. 4.b), the number
of different symbols is 8 and the maximum entropy
value is log2 8 = 3 in our case. Fig. 5 shows edge
segments of an input eye image and gradient distributions, lengths and entropy values for 10 sample
segments. It is easy to observe that circular edge
segments have higher gradient entropy values regardless their size, whereas straight edge segments
have lower values as expected. With this heuristic, we can discard the segments producing small
entropy values than a certain threshold in an extremely fast manner. When examining speed of this
method, we measure that computing the gradient
entropy of an edge segment with available image
derivatives is faster than ellipse fitting and error
computation up to 200 times. In this way, we avoid
spending computation time on the segments which
have irrelevant geometries than ellipticals.
Following the computation of gradient entropies
of edge segments, one segment is chosen to be the
near-circular segment to extract elliptical arcs if it
satisfies the following three criteria:
i) must have a high gradient entropy. The theoretical entropy upper-bound for 8 different gradient directions is log2 8 = 3. Accordingly, we
choose the segments which have 2.9 or more
gradient entropy.
6
due to the small occlusions such as glints, we
consider a 15 pixels threshold for the distance
between the start and end points of the segment.
Edge Segment
Gradient Distr.
Length
Entropy
512
2.96
348
0.95
186
1.35
246
2.89
314
2.86
88
2.03
63
2.82
71
2.61
73
1.38
86
1.04
Along with the second criterion, ellipse fitting
is an essential tool employed in various steps of
the proposed method. Through a crowd literature, there are two renowned ellipse fitting methods
which are known to be fast and robust [35], [36].
Among these two algorithms, Taubin’s method [35]
results a better ellipse contour with slightly lower
error, however, it does not guarantee that the
resulted conic is an ellipse, rather it can return
a hyperbola as well. In addition, Fitzgibbon’s
method [36] always ensures that the resulted conic
is an ellipse, but it tends to extract more eccentric ellipses with higher fitting errors. To benefit
advantages of both methods, we follow a simple
Taubin-prior procedure as the following. First, we
use Taubin’s method and examine the coefficients
of resulted conic to understand whether its an ellipse or hyperbola. If it turns out that we get a
hyperbola, then we use Fitzgibbon’s method and
get an ellipse. Due to the fact that we apply ellipse fit to consecutive edge elements rather than
scattered pixel data, we usually end up with a valid
ellipse with Taubin’s method.
In both ellipse fitting methods, we need to compute a fitting error to quantitatively evaluate the
success. For this purpose, there is no straightforward method in the literature except numerical approximations [37]. Since inaccurate approximations
easily cause misjudgements of elliptical features, we
developed a more quantitative fitting error computation method based on [38]. We estimate the distance between the ellipse and each point by solving the equations described in [38] with NewtonRaphson method in a couple of iterations. Once
we estimate all distance values between each point
and the ellipse, we calculate the normalized rms distance to obtain a single scalar to represent the fitting error. For ellipse perimeter calculation, which
also does not have an exact solution, hence we employ Ramanujan’s second approximation [39]. In
the event that more than one edge segment satisfies all three conditions given above, the one having the minimum ellipse fitting error is chosen to
be the near-circular segment. While existence of a
near-circular segment speeds up the computation,
its not compulsory for the detection of pupil.
It is important to note that shape of a segment
Figure 5: Two sample eye ROI with and without occlusion
and detected edge segments (top), List of gradient distributions, lengths ans entropy value of several edge segments
selected from above sample images (bottom).
ii) must have a small ellipse fitting error, e.g., 2
pixels, when an ellipse is fit to the pixels which
form the segment.
iii) must be a closed segment. To avoid problems
7
Figure 6: ROI detection, arc extraction, pupil candidate generation and pupil detection steps. The first two rows consist of
examples where the pupil is completely in clear sight; there are examples with occlusions in the rows 2-to-5, and there is no
pupil in the last one. (a) Input image with detected ROI, (b) Detected edge segments within the ROI. Near-circular segment
is indicated in red, if exists. High entropy segments which are subjected to arc extraction if a near-circular segment could
be found indicated in green. Short and low gradient entropy segments which are not omitted indicated in blue. (c) Detected
corners (green boxes) and ellipses which are fit to pixels lying in between two consecutive corners. Only successful (i.e. having
low fitting error) ellipses are indicated. (d) Extracted elliptical arcs. (e) 2n − 1 pupil candidates are generated by joining all
possible arc combinations, (f) the selected ellipse representing the pupil contour using the argument in Eq. 5 (Best viewed in
color)
does not necessarily have to be near-circular to
give high entropy values. In addition to segments
with near-circular geometry, gradient distributions
of segments which have concave shapes or follow
complex trajectories can also end up with high entropy values. Therefore, we use entropy test as a
prerequisite to accelerate the algorithm and make
final decision about a segment by ellipse fitting.
then all segments which have high gradient entropy
(i.e., > 2) are subjected to arc extraction process.
In this manner, the algorithm adapts itself and requires less computation when there is no occlusions
and pupil contour is entirely visible. Due to the
fact that their straight geometry rarely contains elliptical arcs, we omit segments having low gradient
entropy and short segments (i.e., < 25 pixels) to
save further computation time.
In a previous work, we extract circular arcs by
combining consecutive line segments to detect circles in an image3 [27]. However, pupil’s projection
onto the camera plane can be more elliptic, hence
we need to detect the elliptical arcs in this study.
3.4. Elliptical Arc Extraction
The next step of the algorithm is extracting the
elliptical arcs (which will be referred to as arc hereafter) from edge segments obtained in the previous step. If a near-circular segment could be found
at the previous step, arcs are extracted only from
that segment. If no near-circular segment is found,
3 Demo
8
page: http://c-viz.anadolu.edu.tr/EDCircles
Cost Function vs. Frames
1000
Cost Function
Threshold
100
10
1
0.1
0.01
Figure 8: Images of apparatus that we collect database video
sequences.
Figure 7: Output of our cost function for a set of example
frames. It is seen that the result of cost function increases
proportional to the pupil occlusion. If the occlusion is dramatic or there is no pupil in the image, the value of the
function overshoots. (Note that the plot is in logarithmic
scale.)
we try to fit ellipse to each subset of all extracted
arcs. Excluding the empty set, there are 2n − 1 different arc combinations for n arcs. Fig. 6.e shows
all generated pupil candidates generated from extracted arcs in Fig. 6.d.
Since the pupil candidate generation process considers all subsets of selected arcs, groups of unrelated arcs which do not form a valid elliptic structure are also subjected to be eliminated after ellipse fit. Therefore, we eliminate those candidates
which result in high fitting error due to the fact hat
they cannot belong to the pupil boundary. After
we eliminate candidates which result high fitting
error, one of the remaining candidates is going to
be selected as the final pupil by utilization of a cost
function in the final step. Or, the algorithm ends
up with the decision that there is no pupil in the
image, if the output of the cost function diverges.
To solve this problem, we devise another strategy
that finds the start and end points of a potential elliptical arc within an edge segment by locating the
corners along the segment [40]. We detect corners
on the segments with a fast curvature scale-space
(CSS) method which utilize image gradient information to compute turning angle curvature [41].
Afterwards, we apply ellipse fit to the points lying in between two consecutive corners along each
segment and obtain elliptical arcs.
In Fig. 6.b-d we present the results of arc extraction process for several test images. In the first
two rows, the pupil is completely visible; hence, the
near-circular segment (indicated in red) is detected.
Therefore, arc extraction is applied to only this segment. When no near-circular segment is found due
to occlusions, arcs are extracted from all segments
having high gradient entropy to avoid missing any
critical information (see 3rd , 4th and 5th rows of
Fig. 6).
We should note that even if the pupil is in clear
sight, it may appear highly elliptical because of the
view angle. In such cases, pupil’s segment may result a low gradient entropy and near-circular segment would not be detected. As a consequence,
elliptical arc extraction would be applied to all segments even though there is no occlusion.
3.6. Detection of The Pupil
In the previous step, we get a number of pupil
candidates each of which is a subset of elliptical arcs
consisting several arcs. Accordingly, we still need
to select one candidate ellipse to be the final pupil
contour. To make the decision, we define a cost
function Jc which considers the following properties
of a candidate ellipse:
i. the ellipse fitting error (ε),
ii. the eccentricity (e),
iii. the ratio of the arc pixels to the perimeter of
resulting ellipse (φ).
3.5. Generation of the Pupil Candidates
In this step, we generate candidate ellipses by
grouping the extracted arcs. Thus we aim to detect
the pupil boundary completely even if its boundary
is partially visible. To generate pupil candidates,
Each of the pupil candidates is formed by one or
more arcs. If the pupil boundary is detected from
multiple arcs, the fitting error should be reasonable
because we expect the arcs to be parts of the same
9
elliptic contour. Thus we need to minimize the fitting error ε.
The eccentricity (e) indicates the compactness of
an ellipse, or in other words, diversity of an ellipse
from a circle and it is computed as
r
b2
0<e<1
(4)
e= 1− 2 ,
a
where a and b are semi-major and semi-minor axes,
respectively. The eccentricity is 0 for a circle and
1 for a parabola. Among the pupil candidates each
of which is a subset of elliptical arcs, there are also
diverse ellipses whose eccentricities can get close to
1. However, pupil’s projection onto the image plane
is usually closer to a circle rather than a skewed
ellipse in majority of the applications. Therefore,
we tend to select a candidate having an eccentricity
close to 0.
The parameter φ is the ratio of the number of pixels involved in ellipse fitting to the perimeter of the
resulting ellipse. In some circumstances, one single
and short arc may result in a large pupil candidate
ellipse that may lead to inconsistency. Therefore,
we look for the pupil candidates which are formed
by consensus of more arc pixels and have a greater
φ.
During our experiments, we observed that the effect of the eccentricity (e) is less than the effect of
ε and φ due to the possibility of true pupil not being the most compact ellipse among the candidates.
Accordingly, we take squares of ε and φ to increase
their effect on the cost function. Finally, we need
to minimize ε and e and maximize φ in our formulation, and select the candidate that minimizes the
following argument:
Jc (pi ) = arg min(ε,e,φ)
ε2i . πie
φ2i
Figure 9: A snapshot of our pupil annotation tool. Two
different ellipse fitting algorithms are utilized to find the best
conic to represent the pupil GT. After clicking a location
inside the pupil, a guide is displayed in order to help user to
equally sample contour points.
gorithm, it is still possible to obtain arcs and pupil
candidates although there is actually no pupil. We
observe that the cost function overshoots in these
circumstances due to large ε and small φ values.
Therefore, we can easily find if there is no pupil by
quantifying output of Jc . In Fig. 7 we present a
plot of the cost function versus a number of frames
sampled from an eye blink operation.
It is clearly seen that the Jc ’s output rapidly increases as the visible part of the pupil periphery
gets smaller due to occlusions. Similarly, the algorithm ends up that there is no pupil in the image
because the cost function overshoots for all of pupil
candidates in the last row of Fig. 6. By examining
several frames, we find out that a stable threshold value can provide promising results on deciding
whether there is no pupil. We provide more detail
on this topic in the next section with quantitative
experimental results.
(5)
4. Experimental Results
where pi is the ith pupil candidate and π is constant.
Fig. 6.f shows the pupil detection results for sample images. Among the pupil candidates shown in
Fig. 6.e, the one that minimizes the Jc in Eq. 5 is
selected as the pupil.
In this section, we present results of a comprehensive set of experiments in both quantitative and
qualitative manner. We compare the proposed algorithm to three state of the art pupil detection
algorithms, i.e. Starburst [19], Świrski et al. [21]
and ElSe [22]. We quantitatively assess algorithms
in terms of pupil detection accuracy (by means of Fmeasure), pupil localization and running time. We
also provide qualitative results which provide useful
insight to readers about the performance of algorithms. In addition to the content that we present
in the paper, we also provide more supplementary
3.7. Detection of True Negatives
In many applications, having the information
that there is no pupil in the image is important
as much as detecting it. This information can provide very useful extensions to eye tracking applications such as blink detection. In the proposed al10
(a)
(b)
(c)
(d)
Figure 10: Illustration of localization test. (a) Input image, (b) Pupil ground truth pupil (GT), (c) Detected pupil (Det), (d)
Subject4_0022.png
(Swirski)
Overlapping (indicated in blue) and non-overlapping
(indicated in red
and green) pixels.
material (e.g. codes, videos, etc.) on our webEven though having 5 points is sufficient to fit
Other Candidates
an ellipse hence its degree of freedom, we picked 10
site [42].
points in average from each pupil’s boundary to better reduce the effect of perspective distortion. Note
4.1. Pupil Detection Dataset
that a circle’s projection onto image may not be a
perfect ellipse due to the perspective distortion and
In order to perform experiments, we first prelens distortion. Once all points are set along the
pare a dataset containing 1509 high resolution
pupil’s boundary, then we fit ellipse to them with
(1280×720) eye frames collected from 5 Subject1_0074.png
subjects.
(Starburst)
two
different algorithms [36, 35] and select the paWe used a simple head-mounted eye tracking apparameters
which provides lower fitting error. Thereratus (see Fig. 8) consisting of two HD cameras (for
fore,
we
obtain
the best possible conic to represent
scene & eye) that we built for 3D gaze estimation
the
pupil
GT
in
each eye image.
study. To our knowledge, this is the only available
pupil detection dataset in resolution higher than
VGA (640×480) in the literature.
4.2. Localization Assessment
During collection of the frames, we ask all subThe first quantitative test we perform is localiza(Starburst)
jects to move their eyes different directionsSubject1_0075.png
in a certion assessment of pupil detection algorithms. In
tain order. In this way, we obtain eye images as
this evaluation, we quantify success of algorithms
pupil is viewed in diverse angles with and without
for how precisely they detect pupils with respect
occlusions by camera. Furthermore, we also want
to ground truth data. We apply this test only the
users to blink several times to obtain negative imframes in which the pupil is truly detected by each
age samples that pupil does not exist. Eventually,
algorithm in the dataset. Source codes of all algoin 57% of the frames pupil is in clear sight, in 16%
rithms were downloaded from websites that authors
of them there are severe occlusions and in 27% of
provided in corresponding papers [43, 44, 45]. We
them there is no pupil in the dataset. We count a
set all parameters of all algorithms according to corpupil positive sample if more than half of its peresponding publications, or use the best performing
riphery is visible, otherwise it is considered as a
values if it is not explicitly indicated in the paper or
negative sample.
code. For each algorithm, we used a single parameAfter we collect the test frames, we implement an
ter set for all images in the dataset which provides
efficient annotation tool which eases rigorous annothe best overall result.
tation procedure (see Fig. 9). Our annotation tool
In order to quantify the localization performance,
overlays a grid in polar form to ease the selection
we compute overlap ratio OR between the detected
of pixels from pupil boundary in an equal angular
pupil and ground truth by counting number of corresolution. In addition, to ensure localization of
responding pixels as follows:
ground truth (GT) conics, it does not collect the
exact pixel coordinates that users click. Instead, it
Area(EDet ) ∩ Area(EGT )
(6)
OR(EDet , EGT ) =
searches a local pixel neighborhood of clicked locaArea(EDet ) ∪ Area(EGT )
tion to find the maximum image gradient response
where EDet and EGT are ellipse of detected pupil
and picks that location. In this way, we guarantee
the selection of exact edge pixels in between pupil
and ground truth ellipse, respectively [46]. In this
manner we calculate the ratio of the number of overand iris in high resolution images precisely.
11
1.00
0.98
0.97
0.95
0.98
0.97
0.96
0.96
0.94
0.92
0.97
0.95
0.95
0.92 0.93
0.92
0.90
0.90
0.86
0.85
Localization
0.83
0.80
0.77
0.76
0.75
0.70
0.70
0.67
0.60
Subject 1
Subject 2
Starburst
Subject 5
Subject 4
Subject 3
Swirski
ElSe
Overall
Proposed
Figure 11: Localization test results of each algorithm for each subject.
lapping pixels to total number of overlapping and
non-overlapping pixels as seen in Fig. 10.
Once the number of overlapping and nonoverlapping pixels are determined, we calculate OR
and take the average for all images for each algorithm. Higher OR indicates better localization,
hence provides higher accuracy in the application
which pupil detection is utilized, obviously. In
Fig. 11 we present average localization results for
individual subjects and overall for each algorithm.
Although ElSe and Świrski also give prospering results, the proposed algorithm performs the best
with a 4% improvement over the runner up in overall results.
we count TP, FP, TN ve FN samples, we calculate
Precision (Eq. 8) and Recall (Eq. 9) values in order
to compute F-Measure (Eq. 10).
We present F-Measure results in Fig. 12 with respect to a range of εO in between 0.0 to 0.2 which
corresponds to OR varying from a perfectly aligned
ellipses at 80% overlap. We do not evaluate pupils
as TP if they are detected with a OR lower than
80%. From the sketches in Fig. 12 it is clearly seen
that accuracy tests are less contentious than localization experiments where the competition among
the algorithms are tighter. In this experiment, proposed method outperforms others as its accuracy
rapidly increases even in very small εO errors and
follows a very stable path regardless the subject.
We also see that Świrski and ElSe algorithms performs very closely with a significant success over
Starburst algorithm.
4.3. Accuracy Assessment
In the previous experiment we assess the localization performance of algorithms by considering
only the images that they detect pupil correctly. In
this step, we evaluate the accuracy of algorithms
by counting the number of images that the pupil
is correctly detected in the entire frame sequences.
To consider a pupil image as a correct detection
(TP), we calculate overlap error (εO ) as in Eq. 7
and compare the result with a threshold value [47].
εO (EDet , EGT ) = 1 − OR
Precision =
Recall =
count(TP Pupils)
cellipsest(TP Pupils + FP Pupils)
(8)
count(TP Pupils)
count(TP Pupils + FN Pupils)
(7)
F-Measure =
The range of εO is in between 0 and 1, and its value
obviously increases as the intersection area between
detected ellipse and GT decreases. We compare εO
with a threshold value to make a decision on the
detected pupil on whether it is a true positive (TP)
or a false positive (FP). Likewise, we also evaluate
images in which algorithms do not detect a pupil
as true negative (TN) if there is no actual pupil in
the image; or false negative (FN) vice versa. After
2 × Precision × Recall
Precision + Recall
(9)
(10)
4.4. Qualitative Results
Along with quantitative accuracy and localization results, we also present qualitative results in
Fig. 13. In the figure, we provide two results from
each of five subjects from top to bottom. It is also
clearly shown that the algorithm can successfully
12
Subject 2
0.9
0.9
0.9
0.8
0.8
0.8
0.7
0.7
0.7
0.6
0.5
0.4
F-Measure
1.0
0.6
0.5
0.4
0.6
0.5
0.4
0.3
0.3
0.3
0.2
0.2
0.2
0.1
0.1
0.1
0.0
0.0
0.00
0.02
0.04
0.06
0.08
0.10
0.12
0.14
0.16
0.18
0.0
0.00
0.20
0.02
0.04
0.06 0.08
0.10
0.12
0.14
0.16 0.18
0.20
0.00
Subject 4
0.8
0.8
0.7
0.7
0.7
0.6
0.5
0.4
0.6
0.5
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0.0
0.08
0.10
0.12
0.14
0.16
0.18
0.20
F-Measure
0.9
0.8
F-Measure
0.9
0.06
0.08
0.12
0.14
0.16
0.18
0.20
0.14
0.16
0.18
0.20
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0.0
Overlap Error
0.10
Overall
1.0
0.9
0.04
0.06
Subject 5
1.0
0.02
0.04
Overlap Error
1.0
0.00
0.02
Overlap Error
Overlap Error
F-Measure
Subject 3
1.0
F-Measure
F-Measure
Subject 1
1.0
0.00
0.02
0.04
0.06
0.08
0.10
0.12
0.14
0.16
0.18
0.20
0.00
0.02
Overlap Error
Starburst
Swirski
ElSe
0.04
0.06
0.08
0.10
0.12
Overlap Error
Proposed
Figure 12: F-Measure results of each algorithm for each subject.
determine true negatives, i.e. in 2nd and 9th rows.
In Fig. 14 we present several examples where the
proposed algorithm fails. The most common reason for fail cases is motion blur where the algorithm cannot extract edges from the pupil contour.
Therefore, elliptical arcs, hence the pupil contour
cannot be detected. Besides the images presented
here, we also provide video sequences of all algorithms in [42] for interested readers.
into consideration. All algorithm implementations
are in C++ except Starburst which is in MATLAB.
According to a study [48], a typical execution in
MATLAB is 50 times slower than a C++ based
application, therefore we divide timing results of
Starburst by 50. Świrski’s implementation was implemented in order to benefit from parallel computing libraries in order to utilize multi-core CPUs. To
able to make a fair comparison, we assign the application to a specific core and measure running times.
The running times of all algorithms in average for
all images are summarized in Table 2.
According to average running times in Table 2,
proposed method is the fastest one among all algorithms. It can be seen that the proposed algorithm
can run up to 140 Hz in single thread for HD images in 1280×720 resolution. In per subject analysis
we see that the proposed method is slightly slower
than the ElSe algorithm for subject 4. When we
investigate the reason behind the longer execution
of subject 4, we see that there are too many occlusions which cause algorithm to fail on detecting
a near-circular segment and extract arcs from all
edge segments.
Table 3 gives a dissection of running times of the
proposed algorithm for separate steps. ROI detection is obviously the most computation demanding
4.5. Running Time Assessment
We run all experiments on a laptop computer
with Intel i7 2.40 GHz CPU. To be able to make a
fair comparison, we take implementation platforms
Table 2: Average running times of algorithms for each subject in milliseconds. Best timings are indicated in bold.
Algorithm
Subject
Subject 1
Subject 2
Subject 3
Subject 4
Subject 5
Average
Starburst
Świrski
ElSe
Proposed
113.36
56.36
55.89
66.89
183.71
93.79
81.34
79.63
74.62
70.34
92.35
79.48
9.79
9.51
9.72
10.14
9.88
9.81
5.58
6.39
5.04
10.77
6.51
7.25
13
Figure 13: Qualitative results from all algorithms and all subjects (images in every two-rows from the beginning belong to a
different subject).
step of the algorithm which takes roughly half of
the entire execution due to the computation of in-
tegral images and convolution of Haar-like features
at several scales. Another time-consuming step 14
(a)
(b)
(c)
(d)
(e)
Figure 14: Several examples in which the proposed algorithm fails. The algorithm could detect no pupil in the last two images.
especially when a near-circular segment could not
be detected- is the last step, i.e. pupil detection,
which is the main reason behind the algorithm’s
fall back at subject 4. Since this step contains too
many computationally expensive ellipse fitting and
error calculation routines, it significantly stretchs
out the execution time if absence of a near-circular
segment is the case.
analysis, we boost the execution of algorithm and
pave the way for real-time applications for high resolution images.
We performed a comprehensive set of experiments by comparing the proposed method with
three state of the art algorithms and provided both
quantitative and qualitative results. Experimental
evaluations show that the proposed algorithm can
detect the pupil even in tough occlusive cases without compromising the real-time applicability constraints.
5. Conclusions
Pupil detection is an indispensable step in many
eye tracking applications and have to be performed
precisely. In most studies, pupil detection is handled with straightforward methods which lack accuracy and fail in occlusive cases. In this study we
focused on developing an efficient feature-based algorithm for pupil boundary detection by using the
entropy of edge segments. We basically find elliptical arcs in an input image and try to obtain a final
ellipse encircling the pupil with the consensus of all
obtained features.
Because the edge segment detection method we
employed provide optimum localization, elliptical
arcs we extract from edge segments accurately
encircle pupil boundary and estimates its center.
Moreover, by means of the gradient distribution
References
References
[1] S. Goni, J. Echeto, A. Villanueva, R. Cabeza, Robust
algorithm for pupil-glint vector detection in a videooculography eyetracking system, in: Int’l Conf. Pattern
Recognition (ICPR), 2004, pp. 941–944.
[2] T. Mäenpää, An iterative algorithm for fast iris detection, in: Advances in Biometric Person Authentication,
Springer, Berlin, Heidelberg, 2005, pp. 127–134.
[3] X. Long, O. K. Tonguz, A. Kiderman, A high speed
eye tracking system with robust pupil center estimation
algorithm, in: IEEE Int’l Conf. Engineering in Medicine
and Biology, 2007, pp. 3331–3334.
[4] A. Keil, G. Albuquerque, K. Berger, M. A. Magnor,
Real-time gaze tracking with a consumer-grade video
camera, Václav Skala-UNION Agency, 2010.
[5] L. Lin, L. Pan, L. Wei, L. Yu, A robust and accurate
detection of pupil images, in: Int’l Conf. Biomedical
Engineering and Informatics, Vol. 1, 2010, pp. 70–74.
[6] J.-G. Wang, E. Sung, Study on eye gaze estimation,
IEEE Trans. Systems, Man, and Cybernetics, Part B
(Cybernetics) 32 (3) (2002) 332–350.
[7] L. Ma, T. Tan, Y. Wang, D. Zhang, Efficient iris
recognition by characterizing key local variations, IEEE
Transactions on Image Processing 13 (6) (2004) 739–
750.
[8] S. Dey, D. Samanta, An efficient approach for pupil
detection in iris images, in: Int’l Conf. Advanced Computing and Communications, 2007, pp. 382–389.
[9] J. San Agustin, H. Skovsgaard, E. Mollenbach, M. Barret, M. Tall, D. W. Hansen, J. P. Hansen, Evaluation
of a low-cost open-source gaze tracker, in: ACM Int’l
Symposium on Eye-Tracking Research & Applications
(ETRA), New York, NY, USA, 2010, pp. 77–80.
[10] D. Ballard, Generalizing the hough transform to detect
arbitrary shapes, Pattern Recognition 13 (2) (1981) 111
– 122.
Table 3: Detailed average timing results of proposed algorithm for different subjects
Dataset
ROI
Detection
Edge Segment
Detection
Gradient
Entropy Comp.
Corner
Detection
Arc
Extraction
Pupil
Detection
TOTAL
(ms)
Step
Subject 1
Subject 2
Subject 3
Subject 4
Subject 5
Overall
3.47
3.31
2.81
3.91
3.52
3.49
1.46
1.51
1.28
1.96
1.71
1.68
0.07
0.21
0.20
0.17
0.27
0.23
0.06
0.09
0.06
0.14
0.08
0.09
0.25
0.43
0.29
0.62
0.41
0.46
0.27
0.83
0.39
3.97
0.52
1.30
5.58
6.39
5.04
10.77
6.51
7.25
15
[11] M. A. Fischler, R. C. Bolles, Random sample consensus:
A paradigm for model fitting with applications to image
analysis and automated cartography, Communications
of the ACM 24 (6) (1981) 381–395.
[12] Y. Ebisawa, Improved video-based eye-gaze detection
method, IEEE Transactions on Instrumentation and
Measurement 47 (4) (1998) 948–955.
[13] J. B. Hiley, A. H. Redekopp, R. Fazel-Rezai, A low cost
human computer interface based on eye tracking, in:
Int’l Conf. IEEE Engineering in Medicine and Biology
Society (EMBS), 2006, pp. 3226–3229.
[14] C. Morimoto, D. Koons, A. Amir, M. Flickner, Pupil
detection and tracking using multiple light sources, Image and Vision Computing 18 (4) (2000) 331 – 335.
[15] J. Daugman, How iris recognition works, in: IEEE Int’l
Conf. Image Processing (ICIP), Vol. 1, 2002, pp. I–33–
I–36.
[16] E. M. Arvacheh, H. R. Tizhoosh, Iris segmentation: Detecting pupil, limbus and eyelids, in: IEEE Int’ Conf.
Image Processing (ICIP), 2006, pp. 2453–2456.
[17] D. Zhu, S. T. Moore, T. Raphan, Robust pupil center
detection using a curvature algorithm, Computer Methods and Programs in Biomedicine 59 (3) (1999) 145 –
157.
[18] N. Kumar, S. Kohlbecher, E. Schneider, A novel approach to video-based pupil tracking, in: IEEE Int’l
Conf. Systems, Man and Cybernetics (SMC 2009),
2009, pp. 1255–1262.
[19] D. Li, D. Winfield, D. J. Parkhurst, Starburst: A
hybrid algorithm for video-based eye tracking combining feature-based and model-based approaches, in:
IEEE Conf. Computer Vision and Pattern Recognition
(CVPR) Workshops, 2005, pp. 79–79.
[20] W. J. Ryan, D. L. Woodard, A. T. Duchowski, S. T.
Birchfield, Adapting starburst for elliptical iris segmentation, in: IEEE Int’l Conf. Biometrics: Theory, Applications and Systems,, 2008, pp. 1–7.
[21] L. Świrski, A. Bulling, N. Dodgson, Robust real-time
pupil tracking in highly off-axis images, in: ACM Int’l
Symposium on Eye Tracking Research and Applications
(ETRA), 2012, pp. 173–176.
[22] W. Fuhl, T. C. Santini, T. Kuebler, E. Kasneci, Else:
Ellipse selection for robust pupil detection in real-world
environments, ACM Int’l Symp. Eye Tracking Research
and Applications (ETRA).
[23] P. Viola, M. Jones, Rapid object detection using a
boosted cascade of simple features, in: IEEE Conf.
Computer Vision and Pattern Recognition (CVPR),
Vol. 1, 2001, pp. 511–518.
[24] Y. Ebisawa, Robust pupil detection by image difference
with positional compensation, in: IEEE Int’l Conf. Virtual Environments, Human-Computer Interfaces and
Measurements Systems, 2009, pp. 143–148.
[25] D. W. Hansen, Q. Ji, In the eye of the beholder: A
survey of models for eyes and gaze, IEEE Trans. Pattern
Analysis and Machine Intelligence (PAMI) 32 (3) (2010)
478–500.
[26] C. H. Morimoto, M. R. Mimica, Eye gaze tracking techniques for interactive applications, Computer Vision
and Image Understanding 98 (1) (2005) 4 – 24.
[27] C. Akinlar, C. Topal, EDCircles: A real-time circle detector with a false detection control, Pattern Recognition 46 (3) (2013) 725 – 740.
[28] C. Topal, C. Akinlar, Edge Drawing: A combined realtime edge and segment detector, Journal of Visual Com-
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
16
munication and Image Representation 23 (6) (2012) 862
– 872.
C. Topal, O. Ozsen, C. Akinlar, Real-time edge segment
detection with edge drawing algorithm, in: Int’l Symp.
Image and Signal Processing and Analysis (ISPA), 2011,
pp. 313–318.
A. Desolneux, L. Moisan, J.-M. Morel, Edge detection
by helmholtz principle, Journal of Mathematical Imaging and Vision 14 (3) (2001) 271–284.
A. Desolneux, L. Moisan, J. M. Morel, From Gestalt
Theory to Image Analysis: A Probabilistic Approach,
Springer Publishing Company, Incorporated, 2007.
C. Akinlar, C. Topal, EDPF: A Real-time Parameterfree Edge Segment Detector with a False Detection Control, Int’l Journal of Pattern Recognition and Artificial
Intelligence 26 (1).
L. Jia, L. Kitchen, Object-based image similarity computation using inductive learning of contour-segment
relations, IEEE Transactions on Image Processing 9 (1)
(2000) 80–87.
C. Martinez-Ortiz, J. Žunić, Curvature weighted gradient based shape orientation, Pattern Recognition 43 (9)
(2010) 3035 – 3041.
G. Taubin, Estimation of planar curves, surfaces, and
nonplanar space curves defined by implicit equations
with applications to edge and range image segmentation, IEEE Trans. Pattern Anal. Mach. Intell. 13 (11)
(1991) 1115–1138.
A. Fitzgibbon, M. Pilu, R. B. Fisher, Direct least square
fitting of ellipses, IEEE Trans. Pattern Anal. Mach. Intell. (PAMI) 21 (5) (1999) 476–480.
P. L. Rosin, Assessing error of fit functions for ellipses,
Graphical models and image processing 58 (5) (1996)
494–502.
R. Nürnberg, Distance from a point to an ellipse,
wwwf.imperial.ac.uk/~rn/distance2ellipse.pdf, accessed September 15, 2017 (2006).
S. Ramanujan, Collected papers of srinivasa ramanujan,
Chelsea Publishing, New York.
H. I. Cakir, C. Topal, C. Akinlar, An occlusion-resistant
ellipse detection method by joining coelliptic arcs, in:
European Conference on Computer Vision (ECCV),
2016, pp. 492–507.
C. Topal, K. Özkan, B. Benligiray, C. Akinlar, A robust CSS corner detector based on the turning angle
curvature of image gradients, in: ICASSP, 2013, pp.
1444–1448.
Pupil detector supplementary webpage, http://c-viz.
anadolu.edu.tr/PupilDetector, accessed on September 15, 2017.
Starburst source codes page., https://github.com/
thirtysixthspan/Starburst, accessed on September
15, 2017.
Swirski’s source codes page., https://github.com/
LeszekSwirski/pupiltracker/, accessed on September
15, 2017.
ElSe source codes page., ftp://messor.informatik.
uni-tuebingen.de/, accessed on September 15, 2017.
D. K. Prasad, M. K. H. Leung, S. Y. Cho, Edge curvature and convexity based ellipse detection method,
Pattern Recognition 45 (9) (2012) 3204–3221.
A. Y. S. Chia, S. Rahardja, D. Rajan, M. K. Leung, A split and merge based ellipse detector with selfcorrecting capability, IEEE Transactions on Image Processing 20 (7) (2011) 1991–2006.
[48] M. Fornaciari, A. Prati, R. Cucchiara, A fast and effective ellipse detector for embedded vision applications,
Pattern Recognition 47 (2014) 3693–3708.
17
| 1cs.CV
|
Received: 00 Month 0000
Revised: 00 Month 0000
Accepted: 00 Month 0000
DOI: xxx/xxxx
LEARNING-BASED ADAPTIVE CONTROL : THEORY AND APPLICATIONS
Observer-based Adaptive Optimal Output Containment Control
problem of Linear Heterogeneous Multi-agent Systems with
Relative Output Measurements
arXiv:1803.11411v1 [cs.SY] 30 Mar 2018
Majid Mazouchi1 | Mohammad Bagher Naghibi-Sistani1 | Seyed Kamal Hosseini Sani1 | Farzaneh
Tatari2 | Hamidreza Modares3
1 Department
of Electrical Engineering,
Ferdowsi University of Mashhad, Mashhad,
Iran
2 Department of Electrical Engineering,
University of Semnan, Semnan, Iran
3 Missouri University of Science and
Technology, Rolla, MO 65401, USA
Summary
This paper develops an optimal relative output-feedback based solution to the containment control problem of linear heterogeneous multi-agent systems. A distributed
optimal control protocol is presented for the followers to not only assure that their
outputs fall into the convex hull of the leaders’ output (i.e., the desired or safe
region), but also optimizes their transient performance. The proposed optimal control solution is composed of a feedback part, depending of the followers’ state, and a
feed-forward part, depending on the convex hull of the leaders’ state. To comply with
most real-world applications, the feedback and feed-forward states are assumed to be
unavailable and are estimated using two distributed observers. That is, since the followers cannot directly sense their absolute states, a distributed observer is designed
that uses only relative output measurements with respect to their neighbors (measured for example by using range sensors in robotic) and the information which is
broadcasted by their neighbors to estimate their states. Moreover, another adaptive
distributed observer is designed that uses exchange of information between followers over a communication network to estimate the convex hull of the leaders’ state.
The proposed observer relaxes the restrictive requirement of knowing the complete
knowledge of the leaders’ dynamics by all followers. An off-policy reinforcement
learning algorithm on an actor-critic structure is next developed to solve the optimal containment control problem online, using relative output measurements and
without requirement of knowing the leaders’ dynamics by all followers. Finally, the
theoretical results are verified by numerical simulations.
KEYWORDS:
adaptive distributed observer, cooperative output regulation, output containment control, optimal control,
reinforcement learning
1
INTRODUCTION
Distributed control of multi-agent systems has attracted a surge of interest in variety of disciplines, due to its broad applications, including cooperation of multiple robot systems 1,2 , satellite formation flying 3 , vehicles formation control 4 , transportation
2
systems 5 , cooperative surveillance 6 , distributed sensor networks 7 and so forth. Distributed cooperative control offers many
advantages such as less communication requirement, more flexibility, enhanced reliability, and scalability, compared with its
centralized counterpart. A fundamental problem in distributed cooperative control of multi-agent systems is consensus or synchronization, in which the goal is to design distributed control policies for agents to ensure that they reach an agreement on
certain quantities of interest on their states or outputs, using only the local state or output information available to each agent. A
comprehensive review of consensus and synchronization problems is provided in Olfati-Saber et al. 8 . Based on the number of
leaders, consensus or synchronization problems can be categorized into three classes, namely, leaderless 6,9 , leader-following 10,11
and containment control 2,12,13 . In the latest problem, which is the problem of interest in this paper, there exist multiple leaders, and the objective is to drive the followers into a convex geometric space spanned by the leaders. The containment control
problem has been extensively investigated in recent years 2,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27 , due to its numerous potential
applications in practical engineering, for example, in stellar observation for satellite formation 28 , removing hazardous materials
for autonomous robots 2 , and so forth.
In most practical situations, the full state information of agents is unavailable for measurement and/or expensive to measure.
For instance, for a group of mobile agents navigating in environments that global navigation satellite systems signals are rather
attenuated, such as forests, urban canyons, and even some building interiors, no position measurement might be possible using
ordinary global positioning system (GPS) receivers. In this situation, one obvious solution might be attained by installing more
precise and powerful GPS receivers on all the agents. However, not only a more precise and powerful GPS receiver is costly, but
also it uses more electrical power due to the fact that it requires more amplifying of attenuated signal and possibly more burden
of weight. Therefore, in some real world scenarios, this solution may be unfeasible or too costly. Motivating by the concept of
anchor agents 29 in localization problem 30 in the context of wireless sensor networks, another solution to the aforementioned
situation and scenario is to equip only small fraction of agents (leaders) with more precise and powerful GPS receivers to measure
their absolute states. However, the rest of agents (followers), which are just equipped with ordinary GPS receivers, do not have
access to their absolute position measurements, and they just have access to relative output measurements with respect to their
neighbors and the information which is broadcasted to them through the communication network by their neighbors.
Most of existing containment control protocols focus on the case of homogeneous agents, in which all agents have identical
dynamics 15,16,19,24,31 . Some recent results on the containment control problem 22,25 have considered the case of heterogeneous
followers with non-identical dynamics, but assumed that the dimensions of all agents are the same. However, in many real-world
applications, for which there are different types of agents performing different tasks, not only the agents’ dynamics but also
their dimensions are different. This requires designing distributed control protocols to drive the followers’ output into a convex
hull spanned by the leaders’ output. Nevertheless, existing results based on relative state measurements in zheng et al. 22 and
Haghshenas et al. 25 cannot be used, as the relative state does not make sense anymore for followers with different dimensions.
Although the design of distributed relative output-feedback based control protocols is considered in Li et al. 21 and Wen et al. 31 ,
these results are still limited to homogeneous multi-agent systems. Output containment control of heterogeneous multi-agent
systems is considered in Zuo et al. 32 . However, in their method, all the followers require their absolute state or output, as well
as complete knowledge of the leaders’ dynamics, which may not be available to the followers in many applications. Moreover,
their approach requires the restrictive assumption of requiring a strongly connected communication graph.
Besides the above mentioned shortcomings of existing results, another shortcoming is that they do not take into account the
transient behavior of the followers and give importance only to the steady-state response of the followers, i.e., they only assure
that the followers’ states or outputs eventually converge to a convex combination of the leaders’ states or outputs. However, it is
desired to find optimal solutions that not only guarantee steady-state convergence, but also minimize the transient containment
error over time. Another important issue which is not considered in the existing results for containment control is designing online
solutions that do not require complete knowledge of the leaders. Reinforcement learning (RL) 33,34 has been successfully used
to design adaptive optimal controllers for single-agent systems 35,36,37,38,39,40,41,42,43 and multi-agent systems 44,45,46,47,48 online in
real time. However, to our knowledge, there is no RL-based solution to the optimal containment control problem.
To overcome the aforementioned shortcomings of the existing work, this paper presents an observer-based adaptive optimal solution to the output containment control problem of linear heterogeneous multi-agent systems, where two distributed
observers are used. A distributed adaptive observer is designed to estimate followers’ state, and another distributed adaptive
observer is developed to estimate the convex hull of the leaders’ state. The proposed distributed adaptive observer relaxes the
restrictive requirement of knowing the complete knowledge of the leaders’ dynamics by all followers. Then, an off-policy reinforcement learning algorithm on an actor-critic structure is developed to solve the optimal output containment control problem
3
online in real time. The proposed algorithm does not require any knowledge of leaders’ dynamics and uses only the relative output measured data of the followers and the information which is broadcasted through the communication network by neighbors.
The main contributions of the paper are as follows :
1. A novel distributed dynamic relative output feedback control protocol is developed based on cooperative output regulation
framework to solve the output containment control problem of linear fully heterogeneous multi-agent systems.
2. An adaptive distributed observer is presented to estimate the leaders’ dynamics, as well as their outputs, and a convex combination of the leaders’ states for each follower. In contrast to the existing work, this observer relaxes the restrictive requirement
of knowing the complete knowledge of the leaders’ dynamics by all followers.
3. An optimal solution to the distributed containment control problem is presented to optimize the transient output containment
error of followers as well as their control efforts, while assuring a zero-steady state containment error.
4. An off-policy RL algorithm is developed to solve the formulated optimal output containment control problem online in real
time, using relative output measurements of followers with respect to their neighbors and the information which is broadcasted
by neighbors, and without requirement of knowing the complete knowledge of the leaders’ dynamics by all followers.
The subsequent sections are organized as follows: some basic concepts of graph theory, definitions and notations are presented
in Section 2. Section 3 states the output containment control problem in output regulation framework. Moreover, analysis is
provided to find containment control problem offline and non-optimal solution. Distributed adaptive observer is designed in
Section 4. The optimality is explicitly imposed in solving the containment control problem in Section 5, which enables us to use
RL techniques to learn solution online in real time. Numerical simulation is given to validate the effectiveness of the theoretical
results in Section 6. Finally, in Section 7 conclusions are drawn.
2
2.1
PRELIMINARIES
Notations
The following notations are used throughout the paper. Let ℜ𝑛 and ℜ𝑛×𝑚 represent the 𝑛 dimensional real vector space and the
𝑛 × 𝑚 real matrix space, respectively. 0𝑚×𝑛 denotes the 𝑚 × 𝑛 matrix with all zeros. Let 1𝑛 be a column vector with all entries
(
)
equal to 1. 𝐼𝑛 represents an 𝑛 × 𝑛 identity matrix. 𝑑𝑖𝑎𝑔 𝑑1 , ..., 𝑑𝑛 represents a block-diagonal matrix with matrices 𝑑1 , ..., 𝑑𝑛 on
its diagonal. ‖.‖2 denotes the Euclidean norm. For any matrix 𝐻𝑖 ∈ ℜ𝑛×𝑞 , 𝑖 = 1, ..., 𝑚, 𝑐𝑜𝑙(𝐻1 , ..., 𝐻𝑚 ) = [𝐻1𝑇 , ..., 𝐻𝑚𝑇 ]𝑇 and
[ ]
𝑉 𝑒𝑐(𝐻𝑖 ) = 𝑐𝑜𝑙(𝐻𝑖1 , ..., 𝐻𝑖𝑞 ) where 𝐻𝑖𝑗 ∈ ℜ𝑛 is the 𝑗-th column of 𝐻𝑖 . 𝐻𝑖 𝑗 denotes the 𝑗-th row of the matrix 𝐻𝑖 . The symbol
⊗ represents the Kronecker product. The distance from 𝑥 ∈ ℜ𝑁 to the set ⊆ ℜ𝑁 is denoted by 𝑑𝑖𝑠𝑡(𝑥, ) = inf ‖𝑥 − 𝑦‖2 .
𝑦∈
2.2
Graph Theory
In this subsection, some basic concepts on algebraic graph theory are briefly reviewed. Let the communication topology among
𝑛 + 𝑚 agents be presented by a weighted directed acyclic graph = (𝑉 , , 𝐴) with a set of nodes 𝑉 = {𝜈1 , … , 𝜈𝑛+𝑚 }, a set of
edges ⊆ 𝑉 ×𝑉 , and a weighted adjacency matrix 𝐴 = [𝑎𝑖𝑗 ] with non-negative adjacency elements 𝑎𝑖𝑗 , i.e., an edge (𝜈𝑗 , 𝜈𝑖 ) ∈
if and only if 𝑎𝑖𝑗 > 0. Node 𝜈𝑗 is called the parent node, node 𝜈𝑖 is the child node, and 𝜈𝑗 is a neighbor of 𝜈𝑖 . We assume that there
are no self-connections, i.e. (𝜈𝑖 , 𝜈𝑖 ) ∉ . A directed graph is acyclic if graph does not have any directed cycle. The set of node
𝜈𝑖 neighbors is denoted by 𝑁̄ 𝑖 = {𝜈𝑗 ∈ 𝑉 ∶ (𝜈𝑗 , 𝜈𝑖 ) ∈ , 𝑗 ≠ 𝑖}. A directed path from node 𝜈𝑖 to node 𝜈𝑗 is a sequence of edges
(𝜈𝑖 , 𝜈𝑘1 ), (𝜈𝑘1 , 𝜈𝑘2 ), ..., (𝜈𝑘𝓁 , 𝜈𝑗 ) with distinct nodes 𝜈𝑘𝑚 , 𝑚 = 1, ..., 𝓁 in a directed graph. A directed graph is strongly connected if
there is a directed path between every ordered pair of nodes. A directed graph is said to have a spanning forest if there exists
at least one node such that there is a directed path from this node to all the other nodes. An agent is called leader if it does not
receive any information from others, i.e., it has no neighbor. Otherwise, it is called a follower. We assume that agents 1 to 𝑛 are
Δ
Δ
followers, and agents 𝑛 + 1 to 𝑛 + 𝑚 are leaders. For notational convenience, = {1, ..., 𝑛} and = {𝑛 + 1, ..., 𝑛 + 𝑚} are used
to denote, the set of followers and the set of leaders, respectively. The Laplacian matrix 𝐿 = [𝓁𝑖𝑗 ] ∈ ℜ(𝑛+𝑚)×(𝑛+𝑚) associated
4
with 𝐴 is defined as 𝓁𝑖𝑖 =
∑
𝑗∈𝑁𝑖
𝑎𝑖𝑗 and 𝓁𝑖𝑗 = −𝑎𝑖𝑗 where 𝑖 ≠ 𝑗. The Laplacian matrix 𝐿 associated with can be partitioned as
[
]
𝐿1 𝐿2
𝐿=
(1)
0𝑚×𝑛 0𝑚×𝑚
where 𝐿1 ∈ ℜ𝑛×𝑛 and 𝐿2 ∈ ℜ𝑛×𝑚 . Note that since the last 𝑚 agents are the leaders, the last 𝑚 rows of 𝐿 are all equal to zero.
In the sequel, we assume that the communication graph satisfies the following assumption.
Assumption 1. The directed graph is acyclic, and for each follower, there exists at least one leader that has a directed path to it.
Lemma 1 (Meng et al. 14 ). Under Assumption 1, 𝐿1 is invertible, all the eigenvalues of 𝐿1 have positive real parts, each entry
of −𝐿−1
𝐿2 is non-negative, and each row of −𝐿−1
𝐿2 has a sum equal to one.
1
1
Proposition 1 (Qin et al. 49 ). A directed acyclic graph can be relabeled such that its Laplacian matrix is lower triangular
matrix.
Definition 1 ( Rockafellar 50 ). A set ⊆ ℜ is said to be convex if (1 − 𝜆)𝑥 + 𝜆𝑦 ∈ , whenever 𝑥, 𝑦 ∈ and 0 ≤ 𝜆 ≤ 1.
{
}
The convex hull 𝐶𝑜(𝑋) of a finite set of points 𝑋 = 𝑥1 , ..., 𝑥𝑛 is the minimal convex set containing all points in 𝑋. That is
{∑𝑛
}
∑
𝑛
𝐶𝑜(𝑋) =
𝛼 𝑥 |𝑥 ∈ 𝑋, 𝛼𝑖 ∈ ℜ, 𝛼𝑖 ≥ 0, 𝑖=1 𝛼𝑖 = 1 .
𝑖=1 𝑖 𝑖 | 𝑖
3
OUTPUT CONTAINMENT CONTROL: AN OFFLINE AND NON-OPIMAL SOLUTION
In this section, the output containment control problem is first introduced, and some standard assumptions are listed. Then, a
distributed dynamic output feedback control protocol is introduced for each follower that uses only the relative output measured
data of the followers and the information which is broadcasted through the communication network by neighbors. The output
containment control problem is then formulated into a linear cooperative output regulation problem and a non-optimal offline
solution is provided.
Consider a set of agents with 𝑛 heterogeneous followers whose models are described as
𝑥̇ 𝑖 = 𝐴𝑖 𝑥𝑖 + 𝐵𝑖 𝑢𝑖 ,
(2)
𝑦𝑖 = 𝐶𝑖 𝑥𝑖
(3)
and a set of 𝑚 homogeneous leaders as
𝜔̇ 𝑘 = 𝑆𝜔𝑘 ,
(4)
𝑦𝑘 = 𝐷𝜔𝑘
(5)
where 𝑖 ∈ is the set of followers, 𝑘 ∈ is the set of leaders, 𝑥𝑖 ∈ ℜ𝑁𝑖 is the state of agent 𝑖, 𝑢𝑖 ∈ ℜ𝑝𝑖 is its control input, and
𝑦𝑖 ∈ ℜ𝑞 is its output, 𝜔𝑘 ∈ ℜ𝑞̄ is the state of leader 𝑘 and 𝑦𝑘 ∈ ℜ𝑞 is its output.
The following standard assumptions are made on the dynamics of agents 51,52 .
Assumption 2. The pairs (𝐴𝑖 , 𝐵𝑖 ) for 𝑖 ∈ are stabilizable.
Assumption 3. The pairs (𝐴𝑖 , 𝐶𝑖 ) for 𝑖 ∈ are detectable, and 𝐶𝑖 for 𝑖 ∈ , are full row rank.
Assumption 4. The leaders’ dynamics in (4)-(5) are marginally stable, and 𝐷 is full row rank.
Assumption 5. The linear matrix equations
Π𝑖 𝑆 = 𝐴𝑖 Π𝑖 + 𝐵𝑖 Γ𝑖 ,
𝐷 = 𝐶𝑖 Π𝑖
(6)
have unique solutions Π𝑖 ∈ ℜ𝑁𝑖 ×𝑞̄ and Γ𝑖 ∈ ℜ𝑝𝑖 ×𝑞̄ for all 𝑖 ∈ .
Assumption 6. Each follower has access only to the relative output measured data between itself and its neighbors, and the
information which is gotten by its neighbors, i.e., the absolute output measurements of followers are not available.
Assuming 𝐶𝑖 to be full row rank in Assumption 3, is a standard assumption for output feedback (See Gadewadikar et al. 53 )
which is made to avoid redundant measurements. Furthermore, the transmission zeros condition (Huang 51 , Assumption 1.4) is
a standard assumption which is satisfied if 𝐶𝑖 is required to be full row rank. Therefore, Assumption 3 is a standard assumption.
5
Regarding Assumption 4, it is worth to mention that, in containment control problem, the leaders’ dynamic cannot be unstable,
due to the fact that convex hull should be bounded, and it is not reasonable to have unbounded convex hull. Assumption 6 is
made here to relax the need for the full or relative state measurements in the existing literature, which might not be feasible in
some scenarios.
The output containment control problem is now defined as follows.
Definition 2. The multi-agent system (2)-(5) achieves the output containment if the followers outputs, converge to the convex
hull spanned by the leaders outputs. That is,
(
(
))
𝑙𝑖𝑚 𝑑𝑖𝑠𝑡 𝑦𝑖 (𝑡) , 𝐶𝑜 𝑦𝑘 (𝑡) , 𝑘 ∈ = 0, ∀𝑖 ∈ .
(7)
𝑡→∞
Now, define the containment error of the follower 𝑖 as
𝑛
∑
∑
𝑛+𝑚
𝛿𝑖𝑘 (𝑦𝑖 − 𝑦𝑘 )
(8)
𝑛+𝑚
∑ (
)
(
)
Δ𝑘 1𝑛 ⊗ 𝑦𝑘
𝑒 = 𝐿1 ⊗ 𝐼𝑞 𝑌𝐹 −
(9)
𝑒𝑖 =
𝑎𝑖𝑗 (𝑦𝑖 − 𝑦𝑗 ) +
𝑘=𝑛+1
𝑗=1
which, in compact form, yields
𝑘=𝑛+1
where 𝑒 and 𝑌𝐹 are the stack column vectors of 𝑒𝑖 and 𝑦𝑖 for 𝑖 ∈ , respectively, 𝐿1 is defined in (1) and Δ𝑘 denotes the
connection weight of leader 𝑘, which is a diagonal matrix with diagonal elements 𝛿𝑖𝑘 = 𝑎𝑖𝑘 where 𝑖 ∈ and 𝑘 ∈ .
Remark 1. lim𝑡→∞ 𝑒(𝑡) = 0 with 𝑒 defined in (9) implies that the followers’ outputs converge to the convex hull spanned by the
leaders’ outputs
∑ (
𝑛+𝑚
𝐿−1
Δ𝑘 1𝑛 ⊗ 𝑦𝑘
1
)
(10)
𝑘=𝑛+1
where 𝐿1 is defined in (1). Note that, 𝐿2 in (1) can be expressed as
[
]
𝐿2 = − Δ𝑛+1 1𝑛 , ..., Δ𝑛+𝑚 1𝑛
(11)
then, one has
∑ (
𝑛+𝑚
)
𝐿−1
Δ𝑘 1𝑛 ⊗ 𝑦𝑘 = −(𝐿−1
𝐿2 ⊗ 𝐼𝑞 )𝑌𝑅 .
1
1
(12)
𝑘=𝑛+1
where 𝑌𝑅 is the stack column vector of 𝑦𝑘 for 𝑘 ∈ .
To solve the containment control problem in Definition 2, the following distributed dynamic measurement relative output
feedback control protocol is introduced in this paper.
𝑢𝑖 =𝐾𝑖1 𝜉𝑖 + 𝐾𝑖2 𝜂𝑖 ,
(13)
𝑛
𝑛+𝑚
∑
∑
𝜉̇ 𝑖 =𝐴𝑖 𝜉𝑖 + 𝐵𝑖 𝑢𝑖 − 𝜇𝑖 𝐹𝑖 ( 𝑎𝑖𝑗 ((𝑦𝑗 − 𝑦𝑖 ) + 𝑦̂𝑖 − 𝑦̂𝑗 ) +
𝛿𝑖𝑘 ((𝑦𝑘 − 𝑦𝑖 ) + 𝑦̂𝑖 − 𝐷𝜔𝑘 ),
𝑗=1
𝑛
∑
𝜂̇ 𝑖 =𝑆 𝜂𝑖 + 𝛽(
𝑗=1
(14)
𝑘=𝑛+1
∑
𝑛+𝑚
𝑎𝑖𝑗 (𝜂𝑗 − 𝜂𝑖 ) +
𝛿𝑖𝑘 (𝜔𝑘 − 𝜂𝑖 ))
(15)
𝑘=𝑛+1
for 𝑖 ∈ , where 𝜉𝑖 ∈ ℜ𝑁𝑖 denotes the estimate of the state of agent 𝑖, 𝑦̂𝑖 = 𝐶𝑖 𝜉𝑖 denotes the estimate of the output 𝑦𝑖 , 𝜂𝑖 ∈ ℜ𝑞̄
denotes the estimate of the convex combination of the leaders’ states. Moreover, 𝜇𝑖 > 0 and 𝐹𝑖 ∈ ℜ𝑁𝑖 ×𝑞 are coupling gains and
gain matrices, respectively, to be designed in Lemma 3, and 𝛽 > 0 is a coupling gain. Finally, 𝐾𝑖1 ∈ ℜ𝑝𝑖 ×𝑁𝑖 and 𝐾𝑖2 ∈ ℜ𝑝𝑖 ×𝑞̄ are
design feedback and feed-forward gain matrices, respectively, for 𝑖 ∈ .
Note that the (15) is a dynamic compensator which its compact form can be regarded as an observer for the convex combination
of the leaders’ states 25 , and (14) is a distributed observer for the state of 𝑖th follower, i.e., 𝑥𝑖 . Note also that absolute output
measurements of followers are not used in this control protocol.
6
Remark 2. Since the leaders act as command generators and are usually equipped with more powerful sensors, we reasonably
assume that they know their own states and can broadcast them to their neighbors. This assumption is a standard assumption
in cooperative control with relative measurements literature (For instant see (Zhang et al. 54 , Section IV) and (Wu et al. 55 , Eqn.
(6)). However, the followers are assumed to have not access to their absolute output measurements.
Remark 3. Note that the leaders are autonomous agents in the sense that they do not receive information from other agents and
acts autonomously to guide other followers. That is, leaders are not being influenced by followers and, therefore, the observer
design in (14) and as well as the control design in (13) are only designed for the followers to assure that they converge into a
convex hull of the leaders’ output trajectory, which is the ultimate goal of the output containment control problem.
Consider the distributed state observer (14) and let the global state estimation error be
𝑋̃ = 𝑋 − 𝜉
(16)
where 𝑋 and 𝜉 are the stack column vectors of 𝑥𝑖 and 𝜉𝑖 , for 𝑖 ∈ , respectively. Using (14) and (2), the dynamics of the global
state estimation error yields
(
(
) )
𝑋̃̇ = 𝐴 − 𝜇𝐹 𝐿 ⊗ 𝐼 𝐶̄ 𝑋̃
(17)
1
𝑞
where 𝐴 = 𝑑𝑖𝑎𝑔(𝐴1 , ..., 𝐴𝑛 ), 𝐶̄ = 𝑑𝑖𝑎𝑔(𝐶1 , ..., 𝐶𝑛 ), 𝜇 = 𝑑𝑖𝑎𝑔(𝐼𝑁1 ⊗ 𝜇1 , ..., 𝐼𝑁𝑛 ⊗ 𝜇𝑛 ), and 𝐹 = 𝑑𝑖𝑎𝑔(𝐹1 , ..., 𝐹𝑛 ).
Before proceeding, the following technical results are required.
Lemma 2. Let Assumptions 1 and 3 be satisfied. Let 𝜆𝑖 be the 𝑖-th eigenvalue of 𝐿1 . Design the observer gain 𝐹𝑖 in (14) for
𝑖 ∈ as
𝐹𝑖 = Φ𝑖 𝐶𝑖𝑇 𝑅−1
𝑖
(18)
where Φ𝑖 is the unique positive definite solution of the observer ARE
𝐴𝑖 Φ𝑖 + Φ𝑖 𝐴𝑇𝑖 + 𝐸𝑖 − Φ𝑖 𝐶𝑖𝑇 𝑅−1
𝐶𝑖 Φ𝑖 = 0
𝑖
(19)
with 𝐸𝑖 = 𝐸𝑖𝑇 ∈ ℜ𝑁𝑖 ×𝑁𝑖 and 𝑅𝑖 = 𝑅𝑇𝑖 ∈ ℜ𝑞×𝑞 positive definite design matrices. Then, the global state estimation error dynamic
(17) is asymptotically stable if the coupling gain satisfies
1
(20)
𝜇𝑖 ≥
2𝜆𝑖 min
where 𝜆𝑖 min = min𝑖∈𝑁̄ 𝑖 Re(𝜆𝑖 ).
Proof. One can show that, using Proposition 1 and the same procedure as Theorem 3 in Fax et al. 4 , if all the matrices 𝐴𝑖 −
𝜇𝑖 𝜆𝑖 𝐹𝑖 𝐶𝑖 , ∀𝑖 ∈ are Hurwitz, then, the global state estimation error dynamic (17) is asymptotically stable. Under Assumption
1 and based on Lemma 1, all the eigenvalues of 𝐿1 have positive real parts. The rest of proof is similar to that of Theorem 1 in
Zhang et al. 54 and is omitted.
Remark 4. It is worth noting that, the requirement for acycle assumption in Lemma 2 can be obviated if all agents have identical
dynamics (See Zhang et al. 54 for more details).
The composition of (2)-(5), the containment error (9), and the control laws (13)-(15) result in the following closed-loop system
𝑥̇ 𝑐 = 𝐴𝑐 𝑥𝑐 + 𝐵𝑐 Ω
̄
Ω̇ = 𝑆Ω
(21)
𝑒 = 𝐶𝑐 𝑥𝑐 + 𝐷𝑐 Ω
(23)
(22)
where
̄ 1 𝐵𝐾
̄ 2⎤
0𝑁×𝑚𝑞̄
⎡ 𝐴 𝐵𝐾
⎡
⎤
[
]
⎥ , 𝐶𝑐 = (𝐿 ⊗ 𝐼 )𝐶̄ 0
̄ 2 ⎥ , 𝐵𝑐 = ⎢
𝐴𝑐 = ⎢ 𝐴𝑐21 𝐴𝑐22 𝐵𝐾
(24)
0𝑁×𝑚𝑞̄
1
𝑞
𝑛𝑞×𝑁 0𝑛𝑞×𝑛
̄ 𝑞̄ , 𝐷𝑐 = 𝐿2 ⊗ 𝐷,
⎢
⎥
⎢
⎥
0
0
𝐴
−(𝐿
⊗
𝛽𝐼
)
⎣ 𝑛𝑞×𝑁
⎣
̄
𝑛𝑞×𝑁
̄
𝑐33 ⎦
2
𝑞̄ ⎦
[ 𝑇 𝑇 𝑇 ]𝑇
𝑥𝑐 = 𝑋 𝜉 𝜂
is the closed-loop state, 𝜂 is the stack column vectors of 𝜂𝑖 , Ω is the stack column vectors of 𝜔𝑘 , 𝐴𝑐21 =
̄ 𝐴𝑐22 = 𝐴 + 𝐵𝐾
̄ 1 − 𝜇𝐹 (𝐿1 ⊗ 𝐼𝑞 )𝐶,
̄ 𝐴𝑐33 = 𝑆̄ − (𝐿1 ⊗ 𝛽𝐼𝑞̄ ) , 𝑆̄ = 𝐼𝑚 ⊗ 𝑆, 𝑆̄ = 𝐼𝑛 ⊗ 𝑆, 𝐴 = 𝑑𝑖𝑎𝑔(𝐴1 , ..., 𝐴𝑛 ),
𝜇𝐹 (𝐿1 ⊗ 𝐼𝑞 )𝐶,
𝑛
∑
𝐵̄ = 𝑑𝑖𝑎𝑔(𝐵1 , ..., 𝐵𝑛 ), 𝐶̄ = 𝑑𝑖𝑎𝑔(𝐶1 , ..., 𝐶𝑛 ), 𝑁 = 𝑁𝑖 , 𝐾1 = 𝑑𝑖𝑎𝑔(𝐾11 , ..., 𝐾𝑛1 ) and 𝐾2 = 𝑑𝑖𝑎𝑔(𝐾12 , ..., 𝐾𝑛2 ).
𝑖=1
7
Now, by using (21)-(24), we are ready to describe the output containment control problem, defined in Definition 2, as a
cooperative output regulation problem as follows.
Problem 1 (Cooperative output regulation problem). Given the multi-agent system (2)-(5) and digraph , find the control
protocol of the form (13) such that the closed-loop system (21)-(23) has the following properties:
Property 1 The origin of the closed-loop system (21)-(23) with Ω being set to zero is asymptotically stable, i.e., the matrix 𝐴𝑐
in (24) is Hurwitz.
Property 2 For any initial conditions 𝑥𝑖 (0), 𝜉𝑖 (0), 𝜂𝑖 (0), 𝑖 ∈ and 𝜔𝑘 (0), 𝑘 ∈ , lim 𝑒 (𝑡) = 0.
𝑡→∞
Remark 5. The Property 1 indicates that all solutions of the closed-loop system (21)-(23) forget their initial conditions and
converge to zero when the exosystems (i.e., leaders) are disconnected. Moreover, Properties 1 and 2 together indicate that all
solutions of the closed-loop system (21)-(23) forget their initial conditions and converge to some solutions, depending only on
the exosignals (i.e., 𝜔𝑘 for 𝑘 ∈ ). (See (Haghshenas et al. 25 , Problem 3) and (Huang 52 , Problem 1 and 2) for more details)
Remark 6. Note that, the error given in (23) is actually the output containment error of (9). Therefore, based on Remark 1, by
solving Problem 1, the output containment control problem, described in Definition 2, is also solved.
In order to solve Problem 1, we present following lemma on the closed-loop system (21)-(23).
Lemma 3. Suppose that the closed-loop system (21)-(23) satisfies Property 1, under the distributed control laws (13)-(15).
Then, lim 𝑒(𝑡) = 0, if there exists a matrix 𝑋𝑐 that satisfies the following linear matrix equations:
𝑡→∞
{
𝐴𝑐 𝑋𝑐 + 𝐵𝑐 = 𝑋𝑐 𝑆̄
(25)
𝐶𝑐 𝑋𝑐 + 𝐷𝑐 = 0𝑛𝑞×𝑚𝑞̄
Proof. Let 𝑥̃ 𝑐 = 𝑥𝑐 − 𝑋𝑐 Ω. Using (25), one has
𝑥̃̇ 𝑐 = 𝐴𝑐 𝑥̃ 𝑐
(26)
Since 𝐴𝑐 is Hurwitz, lim 𝑥̃ 𝑐 (𝑡) = 0. Using (25) and some manipulation, (23) can be rewritten as
𝑡→∞
𝑒 = 𝐶𝑐 𝑥̃ 𝑐
(27)
then, it follows from lim 𝑥̃ 𝑐 (𝑡) = 0, that lim 𝑒(𝑡) = 0.
𝑡→∞
𝑡→∞
Now the following theorem shows that Problem 1 can be solved using the distributed control laws (13)-(15).
Theorem 1. Consider the multi-agent system (2)-(5). Let Assumptions 1 and 5 be satisfied and 𝜇𝑖 and 𝐹𝑖 in the control protocol
(13) be designed as of Lemma 2. Design 𝐾𝑖1 such that 𝐴𝑖 + 𝐵𝑖 𝐾𝑖1 is Hurwitz and 𝐾𝑖2 using
𝐾𝑖2 = Γ𝑖 − 𝐾𝑖1 Π𝑖 .
(28)
where Γ𝑖 and Π𝑖 are solutions of (6). Then, Problem 1 is solved using the distributed control laws (13)-(15), for any positive
constant 𝛽.
Proof. For nonsingular matrix
⎡ 𝐼𝑁 0𝑁×𝑁 0𝑁×𝑛𝑞̄ ⎤
𝑇 = ⎢ 𝐼𝑁
𝐼𝑁 0𝑁×𝑛𝑞̄ ⎥ ,
⎢
⎥
0𝑛𝑞×𝑁
𝐼𝑛𝑞̄ ⎦
⎣ 0𝑛𝑞×𝑁
̄
̄
one can verify that
𝑇
−1
̄ 1
̄ 1
̄ 2
𝐵𝐾
𝐵𝐾
⎡ 𝐴 + 𝐵𝐾
⎤
̄
⎢
⎥
0𝑁×𝑁 𝐴 − 𝜇𝐹 (𝐿1 ⊗ 𝐼𝑞 )𝐶
0𝑁×𝑛𝑞̄
𝐴𝑐 𝑇 =
⎢
⎥
⎣ 0𝑛𝑞×𝑁
0𝑛𝑞×𝑁
𝑆̄ − (𝐿1 ⊗ 𝛽𝐼𝑞̄ ) ⎦
̄
̄
(29)
which has the block-triangular structure. Under Assumption 2, there exists 𝐾𝑖1 , ∀𝑖 ∈ such that 𝐴𝑖 + 𝐵𝑖 𝐾𝑖1 is Hurwitz. Based
on Lemma 2, 𝐴 − 𝜇𝐹 (𝐿1 ⊗ 𝐼𝑞 )𝐶̄ is Hurwitz, if positive constants 𝜇𝑖 and 𝐹𝑖 , ∀𝑖 ∈ are chosen as (20) and (18), respectively.
Based on Lemma 1, under Assumption 1, all the eigenvalues of 𝐿1 have positive real parts. Therefore, under Assumption 4, one
8
can see that 𝑆̄ − (𝐿1 ⊗ 𝛽𝐼𝑞̄ ) is Hurwitz for any positive constant 𝛽. Thus, under the distributed control laws (13)-(15), Property
1 in Problem 1 is satisfied.
Next we verify the Property 2 in Problem 1. Let 𝐾𝑖2 , ∀𝑖 ∈ be given by (28). Then, under Assumption 5, one has
(
)
̄ 1 )Π + 𝐵𝐾
̄ 2 = Π 𝐼𝑛 ⊗ 𝑆
(𝐴 + 𝐵𝐾
(30)
where Π = 𝑑𝑖𝑎𝑔(Π𝑖 ) for 𝑖 ∈ , Let
𝐿2 ⊗ 𝐷) ⎤
⎡ −𝐶̄ −1 (𝐿−1
1
−1
−1
⎢
̄
𝑋𝑐 = −𝐶 (𝐿1 𝐿2 ⊗ 𝐷) ⎥ .
⎢
⎥
𝐿2 ⊗ 𝐼𝑞̄ ) ⎦
⎣ −(𝐿−1
1
(31)
̄ 1 )𝐶̄ −1 (𝐿−1 𝐿2 ⊗ 𝐷) + 𝐵𝐾
̄ 2 (𝐿−1 𝐿2 ⊗ 𝐼𝑞̄ ) ⎤
⎡ (𝐴 + 𝐵𝐾
1
1
−1
−1
̄ 1 )𝐶̄ (𝐿 𝐿2 ⊗ 𝐷) + 𝐵𝐾
̄ 2 (𝐿−1 𝐿2 ⊗ 𝐼𝑞̄ ) ⎥
𝐴𝑐 𝑋𝑐 + 𝐵𝑐 = − ⎢ (𝐴 + 𝐵𝐾
1
1
⎢
⎥
⎣ (𝑆̄ − (𝐿1 ⊗ 𝛽𝐼𝑞̄ ))(𝐿−1
𝐿2 ⊗ 𝐼𝑞̄ ) + (𝐿2 ⊗ 𝛽𝐼𝑞̄ ) ⎦
1
𝐿2 ⊗ 𝐷)𝑆̄ ⎤
⎡ 𝐶̄ −1 (𝐿−1
1
−1
−1
̄
= − ⎢ 𝐶̄ (𝐿1 𝐿2 ⊗ 𝐷)𝑆̄ ⎥ = 𝑋𝑐 𝑆.
⎢
⎥
−1
̄
⎣ (𝐿 𝐿2 ⊗ 𝐼𝑞̄ )𝑆 ⎦
(32)
Δ
Then, using (21)-(23) and (30), yields
1
Note that
̄ −1 𝐿 ⊗ 𝐼 ) = (𝐼 ⊗ 𝑆)(𝐿−1 𝐿 ⊗ 𝐼 )
𝑆(𝐿
2
𝑞̄
𝑛
2
𝑞̄
1
1
̄
= (𝐿−1 𝐿2 ⊗ 𝐼𝑞̄ )𝑆.
1
(33)
Moreover,
𝐿2 ⊗ 𝐷) + (𝐿2 ⊗ 𝐷)
𝐶𝑐 𝑋𝑐 + 𝐷𝑐 = −(𝐿1 ⊗ 𝐼𝑞 )𝐶̄ 𝐶̄ −1 (𝐿−1
1
= 0.
(34)
Hence, 𝑋𝑐 satisfies the linear matrix equations (25) and it follows from Lemma 3 that property 2 is also satisfied, i.e.,
lim 𝑒(𝑡) = 0. This completes the proof.
𝑡→∞
Remark 7. It is seen from (31) in the proof of Theorem 1 that, Assumption 1-5 guarantees the feasibility of the solution in
Lemma 3, i.e., 𝑋𝑐 .
Remark 8. The graph condition in Haghshenas et al. 25 and Zuo et al. 32 is unnecessarily strong. In this paper, by defining the
output containment error as (9), the restrictive required assumption of strongly connected communication graph in (Haghshenas
et al. 25 , Assumption 6) and (Zuo et al. 32 , Assumption 1) is relaxed to a milder assumption of communication graph with spanning
forest.
Remark 9. The solutions to the output regulator equations (6) as well as the distributed observer (15) need complete knowledge
of leaders’ dynamics. This knowledge, however, is not available to the followers in many applications. In order to obviate the
requirement of the leaders’ dynamics in (15), a distributed adaptive observer is designed for convex combination of the leaders’
states in Section 4. Optimality is next implicitly incorporated in the design of the containment control problem in Section 5
to optimize the transient containment errors of agents. An off-policy RL algorithm is developed to solve the optimal output
containment control problem online in real time and without requiring the knowledge of the leaders’ dynamics.
4
DISTRIBUTED ADAPTIVE OBSERVER FOR LEADERS CONVEX HULL
In this section, a distributed adaptive observer is developed to estimate convex combination of the leaders’ states and outputs
for each follower, as well as the leaders’ dynamics, simultaneously. To estimate the convex combination of the leaders’ states
and outputs for each follower, consider the following distributed adaptive observer
𝑛
∑
𝜂̂̇ 𝑖 = 𝑆𝑖 𝜂̂𝑖 + 𝛽2 (
𝑗=1
∑
𝑛+𝑚
𝑎𝑖𝑗 (𝜂̂𝑗 − 𝜂̂𝑖 ) +
𝑘=𝑛+1
𝛿𝑖𝑘 (𝜔𝑘 − 𝜂̂𝑖 ))
(35)
9
where
𝑛
∑
𝑆̇ 𝑖 = 𝛽1 (
∑
𝛿𝑖𝑘 (𝑆 − 𝑆𝑖 ))
(36)
𝑘=𝑛+1
𝑛+𝑚
𝑗=1
𝑛
𝐷̇ 𝑖 = 𝛽3 (
∑
𝑛+𝑚
𝑎𝑖𝑗 (𝑆𝑗 − 𝑆𝑖 ) +
𝑎𝑖𝑗 (𝐷𝑗 − 𝐷𝑖 ) +
∑
𝛿𝑖𝑘 (𝐷 − 𝐷𝑖 ))
(37)
𝑘=𝑛+1
𝑗=1
𝑦0𝑖 = 𝐷𝑖 𝜂̂𝑖
(38)
̄ 𝑞̄
where 𝑆𝑖 ∈ ℜ𝑞×
and 𝐷𝑖 ∈ ℜ𝑞×𝑞̄ , for 𝑖 ∈ , are the estimation of the leaders’ dynamics 𝑆 and 𝐷 respectively, 𝜂̂𝑖 ∈ ℜ𝑞̄
𝑞
and 𝑦0𝑖 ∈ ℜ are the estimation of convex combination of the leaders’ states and outputs for 𝑖th follower, respectively, and
𝛽1 , 𝛽2 , 𝛽3 > 0. The following lemma is used in the proof of Theorem 2 for the observer design.
Lemma 4 (Cai et al. 56 ). Consider the following system
𝜐̇ = 𝜀Z𝜐 + Z1 (𝑡)𝜐 + Z2 (𝑡)
𝑞̄
(39)
𝑞×
̄ 𝑞̄
𝑞×
̄ 𝑞̄
𝑞̄
where 𝜐 ∈ ℜ , Z ∈ ℜ
is asymptotically stable, 𝜀 is a positive constant, Z1 (𝑡) ∈ ℜ
and Z2 (𝑡) ∈ ℜ are bounded and
continuous for all 𝑡 ≥ 𝑡𝑜 . If Z1 (𝑡) and Z2 (𝑡) decay to zero exponentially as time go to infinity, then, for any 𝜐(𝑡0 ) and any 𝜀 > 0,
𝜐(𝑡) decays to zero exponentially as time go to infinity.
Theorem 2. Consider the leader dynamics (4)-(5), and the adaptive observer (35)-(37). Let 𝑆̃𝑖 = 𝑆𝑖 − 𝑆, 𝐷̃ 𝑖 = 𝐷𝑖 − 𝐷 be
leaders’ dynamics estimation errors and 𝜂̃𝑖 = 𝜂̂𝑖 − 𝜔∗𝑖 and 𝑦̃0𝑖 (𝑡) = 𝐷𝑖 𝜂̂𝑖 − 𝐷𝜔∗𝑖 be 𝑖th follower state estimation error of the convex
]
[
𝐿2 𝑖 ⊗ 𝐼𝑞̄ )𝜔 with 𝜔 = 𝑐𝑜𝑙(𝜔1 , ..., 𝜔𝑚 ). Then,
combination of the leaders’ states and outputs, respectively, where 𝜔∗𝑖 is −( 𝐿−1
1
for any initial conditions 𝑆̃𝑖 (0), 𝐷̃ 𝑖 (0), 𝑦̃0𝑖 (0) and 𝜂̃𝑖 (0), one obtains
1. for any positive constant 𝛽1 , ∀𝑖 ∈ , lim𝑡→∞ 𝑆̃𝑖 (𝑡) = 0 exponentially;
2. for any positive constant 𝛽3 , ∀𝑖 ∈ , lim𝑡→∞ 𝐷̃ 𝑖 (𝑡) = 0 exponentially;
3. for any positive constant 𝛽1 and 𝛽2 , ∀𝑖 ∈ , lim𝑡→∞ 𝜂̃𝑖 (𝑡) = 0 exponentially;
4. for any positive constant 𝛽1 , 𝛽2 and 𝛽3 , ∀𝑖 ∈ , lim𝑡→∞ 𝑦̃0𝑖 (𝑡) = 0 exponentially.
Proof. The dynamics of the global leaders’ dynamics estimation error can be written as
𝑆̃̇ = −𝛽 (𝐿 ⊗ 𝐼 )(𝑆 + (𝐿−1 𝐿 ⊗ 𝐼 )(1 ⊗ 𝑆))
1
1
𝑞
𝑐
1
2
𝑞̄
𝑚
(40)
where 𝑆𝑐 = 𝑐𝑜𝑙(𝑆1 , ..., 𝑆𝑛 ). Using Lemma 1, it is easy to show that
−(𝐿−1
𝐿2 ⊗ 𝐼𝑞̄ )(1𝑚 ⊗ 𝑆) = (1𝑛 ⊗ 𝑆).
1
(41)
𝑆̃̇ = −𝛽1 (𝐿1 ⊗ 𝐼𝑞̄ )𝑆̃
(42)
̃̇ = −𝛽1 (𝐼𝑞̄ ⊗ 𝐿1 ⊗ 𝐼𝑞̄ )𝑉 𝑒𝑐(𝑆).
̃
𝑉 𝑒𝑐(𝑆)
(43)
Then,
or equivalently
̃
According to Assumption 1 and Lemma 1, all the eigenvalues of 𝐿1 have positive real parts. Hence, lim𝑡→∞ 𝑉 𝑒𝑐(𝑆(𝑡))
=0
̃
̃
exponentially, which yields lim𝑡→∞ 𝑆𝑖 (𝑡) = 0, ∀𝑖 ∈ . Similar to part (1), one can see that lim𝑡→∞ 𝐷𝑖 (𝑡) = 0, ∀𝑖 ∈ exponentially
for any positive constant 𝛽3 .
Next, it remains to prove part (3) and part (4), i.e. lim𝑡→∞ 𝜂̃𝑖 (𝑡) = 0 and lim𝑡→∞ 𝑦̃0𝑖 (𝑡) = 0. To this end, using (35) and some
manipulations yields
𝑛
∑
𝜂̃̇ 𝑖 =𝑆𝑖 𝜂̂𝑖 + 𝛽2 (
∑
𝑛+𝑚
𝑎𝑖𝑗 (𝜂̂𝑗 − 𝜂̂𝑖 ) +
𝑗=1
[
]
𝛿𝑖𝑘 (𝜔𝑘 − 𝜂̂𝑖 )) + ( 𝐿−1
𝐿2 𝑖 ⊗ 𝐼𝑞̄ )(𝐼𝑚 ⊗ 𝑆)𝜔
1
𝑘=𝑛+1
𝑛
𝑛+𝑚
∑
∑
𝛿𝑖𝑘 (𝜔𝑘 − 𝜂̂𝑖 )) + 𝑆 𝜂̃𝑖 + 𝑆̃𝑖 𝜂̃𝑖 + 𝑆̃𝑖 𝜔∗𝑖 .
=𝛽2 ( 𝑎𝑖𝑗 (𝜂̂𝑗 − 𝜂̂𝑖 ) +
𝑗=1
𝑘=𝑛+1
(44)
10
The global form of (44) can be written as
𝜂̃̇ =((𝐼𝑛 ⊗ 𝑆))𝜂̃ + 𝑆̃𝑑 𝜂̃ + 𝑆̃𝑑 𝜔∗ − 𝛽2 (𝐿2 ⊗ 𝐼𝑞̄ )𝜔 − 𝛽2 (𝐿1 ⊗ 𝐼𝑞̄ )(𝜂̃ − (𝐿−1
𝐿2 ⊗ 𝐼𝑞̄ )𝜔)
1
∗
=((𝐼𝑛 ⊗ 𝑆) − 𝛽2 (𝐿1 ⊗ 𝐼𝑞̄ ))𝜂̃ + 𝑆̃𝑑 𝜂̃ + 𝑆̃𝑑 𝜔
(45)
where 𝜂̃ = 𝑐𝑜𝑙(𝜂̃1 , ..., 𝜂̃𝑛 ), 𝜔∗ = 𝑐𝑜𝑙(𝜔∗1 , ..., 𝜔∗𝑛 ) and 𝑆̃𝑑 = 𝑑𝑖𝑎𝑔(𝑆̃1 , ..., 𝑆̃𝑛 ). According to Assumption 4, for any 𝛽1 , 𝛽2 > 0, the
matrix ((𝐼𝑛 ⊗ 𝑆) − 𝛽2 (𝐿1 ⊗ 𝐼𝑞̄ )) is Hurwitz, and 𝑆̃𝑑 𝜔∗ decays to zero exponentially. Hence, applying Lemma 4 we conclude
that lim𝑡→∞ 𝜂̃𝑖 (𝑡) = 0 exponentially. Adding and subtracting 𝐷𝑖 𝜔∗𝑖 to the right hand side of 𝑦̃0𝑖 (𝑡), we have
𝑦̃0𝑖 (𝑡) = 𝐷𝑖 (𝜂̂𝑖 − 𝜔∗𝑖 ) + (𝐷𝑖 − 𝐷)𝜔∗𝑖 .
(46)
Therefore, if lim𝑡→∞ 𝜂̃𝑖 (𝑡) = 0 and lim𝑡→∞ 𝐷̃ 𝑖 (𝑡) = 0, then lim𝑡→∞ 𝑦̃0𝑖 (𝑡) = 0, which completes the proof.
5
OPTIMAL OUTPUT CONTAINMENT CONTROL PROBLEM
In this section, first an optimal formulation for the output containment control problem is introduced. Then, an off-policy reinforcement learning (RL) algorithm is proposed to drive the followers optimally into a convex hull spanned by the leaders’ outputs.
At first, it is assumed that followers can have access to their states and the leaders’ states and dynamics. Then, this restrictive
assumption is relaxed by combining the RL based optimal control presented in this section with distributed adaptive observer
designed in Section 4 and the distributed observer (14).
5.1
Problem Formulation and Its Offline Solution
The aim of optimal output containment control problem is to design the distributed control laws (13)-(15) such that the followers
converge into a convex hull spanned by the leaders’ outputs, and at the same time, the transient output containment error captured
by (9) as well as the followers’ control efforts are minimized. To do so, an optimal output containment control problem is defined
as follows.
Problem 2 (Optimal output containment control problem). Consider the control protocol (13) and let the distributed state
observer (14) be designed as of Lemma 2. Moreover, let the distributed observers (35)-(37) be designed as of Theorem 2. Design
the gain matrices 𝐾𝑖1 and 𝐾𝑖2 , ∀𝑖 ∈ in the control protocol (13), so that not only properties 1 and 2 in Problem 1 are satisfied,
but also the following discounted performance function is minimized.
∞
𝑉𝑖 (𝑡) =
∫
𝑒−𝛾𝑖 (𝜏−𝑡) (𝑒̄𝑇𝑖 𝑄𝑖 𝑒̄𝑖 + 𝑢𝑖 𝑇 𝑊𝑖 𝑢𝑖 )𝑑𝜏
(47)
𝑡
𝑝
for 𝑖 ∈ ,where 𝑄𝑖 ∈ ℜ𝑞×𝑞 and 𝑊𝑖 = 𝑑𝑖𝑎𝑔(𝑤1𝑖 , ..., 𝑤𝑖 𝑖 ) ∈ ℜ𝑝𝑖 ×𝑝𝑖 are symmetric positive definite weight matrices, 𝑤𝑗𝑖 > 0 for
𝑗 = 1, ..., 𝑝𝑖 , 𝛾𝑖 > 0 is the discount factor, and
𝑒̄𝑖 = 𝑦𝑖 − 𝐷𝜔∗𝑖
(48)
(
)
𝑒 = 𝐿1 ⊗ 𝐼𝑞 𝑒̄
(49)
(
)
𝑒̄ = 𝑌𝐹 + 𝐿−1
𝐿2 ⊗ 𝐼𝑞 𝑌𝑅 .
1
(50)
[
]
is the containment error with 𝜔∗𝑖 = −( 𝐿−1
𝐿2 𝑖 ⊗ 𝐼𝑞̄ )𝜔, 𝜔 = 𝑐𝑜𝑙(𝜔1 , ..., 𝜔𝑚 ).
1
Remark 10. Note that (9) can be rewritten as follows
where
Therefore, according to Lemma 1, one can see that minimizing 𝑒̄ results in minimizing the output containment error (9). It
was shown that if the state observer gains in (14) are designed as of Lemma 2, then 𝜉𝑖 → 𝑥𝑖 . Moreover, it was shown that if the
distributed observers (35)-(37) are designed as of Theorem 2, then 𝑆𝑖 → 𝑆, 𝐷𝑖 → 𝐷, and 𝜂̂𝑖 → 𝜔∗𝑖 . Therefore, in steady-state,
the control protocol (13) becomes
𝑢𝑖 = 𝐾𝑖1 𝑥𝑖 + 𝐾𝑖2 𝜔∗𝑖 .
(51)
11
It is shown in the following that if this control protocol is designed to minimize (47), then the output containment control
Problem 2 is solved. One can write (51) in the following form
𝑢𝑖 = 𝐾̄ 𝑖 𝑋̄ 𝑖
(52)
where 𝐾̄ 𝑖 = [𝐾𝑖1 , 𝐾𝑖2 ] and 𝑋̄ 𝑖 (𝑡) = [𝑥𝑖 𝑇 , 𝜔∗𝑖 𝑇 ]𝑇 denotes an augmented state.
The augmented dynamics is given by
𝑋̄̇ = 𝑃 𝑋̄ + 𝐵̄ 𝑢
𝑖
𝑖
𝑖
𝑖 𝑖
𝑒̄𝑖 = 𝐶̄𝑖 𝑋̄ 𝑖
with
(53)
(54)
[
]
[ ]
𝐴𝑖 0
𝐵𝑖
̄
𝑃𝑖 =
, 𝐵𝑖 =
0 𝑆
0
[
]
̄
𝐶𝑖 = 𝐶𝑖 −𝐷 .
(55)
(56)
Now, the value function (47) can be written as the following quadratic form
∞
𝑉 (𝑋𝑖 , 𝑢𝑖 ) =
∫
𝑇
𝑒−𝛾𝑖 (𝜏−𝑡) 𝑋̄ 𝑖𝑇 (𝐶̄ 𝑖 𝑄𝑖 𝐶̄ 𝑖 + 𝐾̄ 𝑖𝑇 𝑊𝑖 𝐾̄ 𝑖 )𝑋̄ 𝑖 𝑑𝜏
𝑡
= 𝑋̄ 𝑖𝑇 Ψ𝑖 𝑋̄ 𝑖 .
(57)
Hamiltonian is defined as follows
𝑑𝑉 (𝑋𝑖 )
𝑇
− 𝛾𝑖 𝑉 (𝑋𝑖 ) + 𝑋̄ 𝑖𝑇 (𝐶̄𝑖 𝑄𝑖 𝐶̄𝑖 + 𝐾̄ 𝑖𝑇 𝑊𝑖 𝐾̄ 𝑖 )𝑋̄ 𝑖
𝐻(𝑋̄ 𝑖 , 𝑢𝑖 , 𝑉 (𝑋𝑖 )) =
𝑑𝑡
/
Then, using stationary condition 57 , i.e., 𝜕𝐻 𝜕𝑢𝑖 = 0, the optimal control gain in (52) is derived as follows
(58)
𝑢∗𝑖 = 𝐾̄ 𝑖∗ 𝑋̄ 𝑖
(59)
∗
∗
𝐾̄ 𝑖∗ = [𝐾 1𝑖 , 𝐾 2𝑖 ]
= −𝑊 −1 𝐵̄ 𝑇 Ψ𝑖
(60)
with
𝑖
𝑖
where Ψ𝑖 satisfies the discounted algebraic Riccati equation (ARE) (61).
Substituting (59), (60), and (57) into (58), yields the discounted ARE as follows
𝑇
𝑃𝑖𝑇 Ψ𝑖 + 𝑃𝑖 Ψ𝑖 − 𝛾𝑖 Ψ𝑖 + 𝐶̄𝑖 𝑄𝑖 𝐶̄𝑖 − Ψ𝑖 𝐵̄ 𝑖 𝑊𝑖−1 𝐵̄ 𝑖𝑇 Ψ𝑖 = 0.
(61)
Remark 11. Note that the ARE (61) and consequently the control gain (60) requires complete knowledge of the leaders’ dynamics. To learn the gains (60) in an online fashion without requiring complete knowledge of the leaders’ dynamics, an off-policy
RL algorithm is designed in Subsection 5.2.
Remark 12. It is shown in Modares et al. 58 that if discount factor 𝛾𝑖 in (47) satisfies
1∕2 ‖
‖
𝛾𝑖 ≤ 𝛾𝑖∗ = 2 ‖(𝐵̄ 𝑖 𝑊𝑖−1 𝐵̄ 𝑖𝑇 𝑄𝑖 ) ‖ ,
(62)
‖
‖
then, the value function (47) is bounded for the control policy (59) and the closed-loop agents’ dynamics are stable. This is
equivalent to Property 1 in Problem 1. The following results show that solving Problem 2 actually solves Problem 1 in an optimal
manner.
Lemma 5. Consider the multi-agent system (2)-(5) and the control policy (59) with gain given by (60) and (61). If the dis(
)
count factors 𝛾𝑖 satisfies (62), then 𝑒̄ = 𝑌𝐹 + 𝐿−1
𝐿2 ⊗ 𝐼𝑞 𝑌𝑅 and consequently output containment error (9) goes to zero
1
asymptotically.
Proof. First, let the ARE (61) be written in the compact form as follows
̄ −1 𝐵̄ 𝑇 Ψ = 0
𝑃 𝑇 Ψ + 𝑃 Ψ − 𝛾̄ Ψ + 𝐶̄ 𝑇 𝑄̄ 𝐶̄ − Ψ𝐵𝑊
(63)
12
where 𝑃 = 𝑑𝑖𝑎𝑔(𝑃1 , ..., 𝑃𝑛 ), Ψ = 𝑑𝑖𝑎𝑔(Ψ1 , ..., Ψ𝑛 ), 𝛾̄ = 𝑑𝑖𝑎𝑔(𝛾1 , ..., 𝛾𝑛 ), 𝑊 = 𝑑𝑖𝑎𝑔(𝑊1 , ..., 𝑊𝑛 ), 𝐶̄ = 𝑑𝑖𝑎𝑔(𝐶̄1 , ..., 𝐶̄𝑛 ), 𝑄̄ =
𝑑𝑖𝑎𝑔(𝑄1 , ...𝑄𝑛 ), and 𝐵̄ = 𝑑𝑖𝑎𝑔(𝐵1 , ..., 𝐵𝑛 ). Multiplying 𝑋̄ 𝑇 and 𝑋̄ to right and left sides of (63), respectively, one has
(
)𝑇
̄ −1 𝐵̄ 𝑇 Ψ𝑋̄ = 0.
(64)
2𝑋̄ 𝑇 𝑃 𝑇 Ψ𝑋̄ − 𝑋̄ 𝑇 𝛾̄ Ψ𝑋̄ + 𝑋̄ 𝑇 𝐶̄ 𝑇 𝑄̄ 𝐶̄ 𝑋̄ − Ψ𝑋̄ 𝐵𝑊
̄ Using this observation, it can be
Observing (64), one can see that, the null space of Ψ is a subspace of the null space of 𝐶̄ 𝑇 𝑄̄ 𝐶.
̄
̄
−1
𝑇
𝑇
𝑇
𝑇
̄ 𝐹 − (𝐿−1 𝐿2 ⊗ 𝐼𝑞 )𝑌𝑅 ) = 0 which
seen that if 𝑋̄ Ψ𝑋̄ = 0, then 𝑋̄ 𝐶 𝑄̄ 𝐶 𝑋̄ = 0 and consequently (𝑌𝐹 − (𝐿1 𝐿2 ⊗ 𝐼𝑞 )𝑌𝑅 ) 𝑄(𝑌
1
yields 𝑌𝐹 + (𝐿−1
𝐿
⊗
𝐼
)𝑌
=
0
and
consequently
the
containment
error
𝑒
=
0,
which
is
defined
in
(9). Therefore, one can
2
𝑞
𝑅
1
conclude that the null space of Ψ is a subspace in which the containment error 𝑒 is zero. It remains to show that providing the
discount factors 𝛾𝑖 satisfies the (62), then the null space of Ψ is attractive. To this end, choose the following Lyapunov function
Taking the derivative of (65) gives
where 𝐴̄ 𝑐 = 𝑑𝑖𝑎𝑔(𝐴𝑐1 , ..., 𝐴𝑐𝑛 ) and
̄ = 𝑋̄ 𝑇 Ψ𝑋.
̄
𝑉 (𝑋)
(65)
(
)
̄ = ∑ 𝑋̄ 𝑇 Ψ𝑖 𝐴𝑐𝑖 + 𝐴𝑇 Ψ𝑖 𝑋̄ 𝑖
𝑉̇ (𝑋)
𝑖
𝑐𝑖
𝑖∈ (
)
= 𝑋̄ 𝑇 Ψ𝐴̄ 𝑐 + 𝐴̄ 𝑇𝑐 Ψ 𝑋̄
(66)
[
∗
𝐴𝑖 + 𝐵𝑖 𝐾 1𝑖 𝐵𝑖 𝐾 2𝑖
𝐴𝑐𝑖 =
0
𝑆
∗
]
(67)
.
It was mentioned in Remark 12 that 𝐴𝑖 + 𝐵𝑖 𝐾𝑖1 is Hurwitz, providing that the discount factors 𝛾𝑖 satisfies the (62). Therefore,
under Assumption 4, one can conclude ∀𝑖 ∈ , 𝐴𝑐𝑖 and consequently 𝐴̄ 𝑐 are marginally stable. Therefore, there exist a positive
(
)
̄ = −𝑋̄ 𝑇 Υ𝑋̄ ≤ 0. Based on LaSalle’s invariance principle, 𝑋̄
semi definite matrix Υ = − Ψ𝐴̄ 𝑐 + 𝐴̄ 𝑇𝑐 Ψ such that 𝑉̇ (𝑋)
̄ = 0. As mentioned above, the null space of Ψ is a subspace of where
converges to the largest invariance subspace where 𝑉̇ (𝑋)
̄ = 0 if Ψ𝑋 = 0 and consequently the output
the output containment error 𝑒 is equal to zero, therefore based on (66), 𝑉̇ (𝑋)
containment error 𝑒 is equal to zero, which completes the proof.
Note that 𝑋̄ 𝑖 (𝑡) = [𝑥𝑖 𝑇 , 𝜔∗𝑖 𝑇 ]𝑇 in (59) is depending on absolute state of the 𝑖th follower, and requiring the knowledge of the
leaders’ dynamics and the graph topology , which are not available for 𝑖th follower, therefore 𝑋̂̄ 𝑖 (𝑡) = [𝜉𝑖 𝑇 , 𝜂̂𝑖 𝑇 ]𝑇 should be
used in place of 𝑋̄ 𝑖 (𝑡) to implement the optimal control (59) without requiring the knowledge of the leaders’ dynamics, graph
topology , and absolute state of the 𝑖th follower. By doing so, optimal control (59) becomes
𝑢∗ = 𝐾̄ ∗ 𝑋̄̂
(68)
𝑖
𝑖
𝑖
where 𝑋̂̄ 𝑖 (𝑡) = [𝜉𝑖 𝑇 , 𝜂̂𝑖 𝑇 ]𝑇 and 𝐾̄ 𝑖∗ is obtained by (60)-(61).
Theorem 3. Consider the multi-agent system (2)-(5) and the distributed adaptive observer (35) along with adaptation laws (36)
and (37). Under Assumptions 1 - 6, Problem 2 and consequently Problem 1 are solved using optimal control policy (68) with
𝐾̄ 𝑖∗ given by (60) and (61). As long as, 𝜇𝑖 and 𝐹𝑖 be designed as of Lemma 2, the discount factors 𝛾𝑖 satisfy (62), and 𝛽1 , 𝛽2 , and
𝛽3 in (35)-(37) be any positive constant.
Proof. Using (2)-(5), (14), (35), and (68), we obtain
̄ ∗
̄ ∗
𝐴
𝐵𝐾
𝐵𝐾
0𝑁×𝑚𝑞̄
⎡ 𝑋̇ ⎤ ⎡
⎤⎡𝑋 ⎤ ⎡
⎤
1
2
∗
∗
̄
̄
̄
⎢ 𝜉̇ ⎥ = ⎢ 𝜇𝐹 (𝐿1 ⊗ 𝐼𝑞 )𝐶̄ 𝐴 + 𝐵𝐾
⎥
⎢
⎥
⎢
⎥Ω
−
𝜇𝐹
(𝐿
⊗
𝐼
)
𝐶
𝐵𝐾
+
𝜉
(𝐿
⊗
𝐷
)
1
𝑞
2
𝑖
1
2
⎢ ̇ ⎥ ⎢
⎥⎢ ⎥ ⎢
⎥
̄
𝜂
̂
𝜂
̂
−(𝐿
⊗
𝛽𝐼
)
⎣ ⎦ ⎣
0𝑛𝑞×𝑁
0𝑛𝑞×𝑁
𝑆 𝑖 − (𝐿1 ⊗ 𝛽𝐼𝑞̄ ) ⎦ ⎣ ⎦ ⎣
2
𝑞̄ ⎦
̄
̄
(69)
𝑛
∑
where 𝜂̂ = 𝑐𝑜𝑙(𝜂̂1 , ..., 𝜂̂𝑛 ), 𝐴 = 𝑑𝑖𝑎𝑔(𝐴1 , ..., 𝐴𝑛 ), 𝐵̄ = 𝑑𝑖𝑎𝑔(𝐵1 , ..., 𝐵𝑛 ), 𝐶̄ = 𝑑𝑖𝑎𝑔(𝐶1 , ..., 𝐶𝑛 ), 𝑁 =
𝑁𝑖 , 𝐾1∗ =
∗
∗
∗
∗
𝑑𝑖𝑎𝑔(𝐾11 , ..., 𝐾𝑛1 ), 𝐾2∗ = 𝑑𝑖𝑎𝑔(𝐾12 , ..., 𝐾𝑛2 ) and 𝑆̄𝑖 = 𝐼𝑛 ⊗ 𝑆𝑖 . For nonsingular matrix
𝑖=1
⎡ 𝐼𝑁 0𝑁×𝑁 0𝑁×𝑛𝑞̄ ⎤
𝑇 = ⎢ 𝐼𝑁
𝐼𝑁 0𝑁×𝑛𝑞̄ ⎥ ,
⎢
⎥
0𝑛𝑞×𝑁
𝐼𝑛𝑞̄ ⎦
⎣ 0𝑛𝑞×𝑁
̄
̄
one can verify that (69) has a block-triangular structure and according to separation principle, the distributed observers and
control gains in (69) can be designed separately. It is shown in Theorem 2 that, for any positive constant 𝛽1 , 𝛽2 , and 𝛽3 , 𝑆𝑖 →
13
𝑆, 𝐷𝑖 → 𝐷, 𝜂̂𝑖 → 𝜔∗𝑖 , ∀𝑖 ∈ , 𝐷𝑖 𝜂̂𝑖 → 𝐷𝜔∗𝑖 and thus for any full row rank matrix 𝐷, (𝐼𝑛 ⊗ 𝐷)𝜂̂ → −(𝐿−1
𝐿2 ⊗ 𝐼𝑞 )𝑌𝑅
1
asymptotically. Besides, it is shown in Lemma 2 that 𝜉𝑖 → 𝑥𝑖 , 𝐶𝑖 𝜉𝑖 → 𝑦𝑖 , ∀𝑖 ∈ asymptotically. Finally, Lemma 5 shows that
𝑌𝐹 → (𝐼𝑛 ⊗ 𝐷)𝜂̂ asymptotically, therefore 𝑌𝐹 → −(𝐿−1
𝐿2 ⊗ 𝐼𝑞 )𝑌𝑅 asymptotically, which completes the proof.
1
Remark 13. Theorem 3 shows that one can design the observers gains appear in (14), (35)-(37) and the control gains 𝐾̄ 𝑖∗ in (68)
independently. As stated in Theorem 3, for 𝑖 ∈ , 𝜇𝑖 should be chosen sufficiently large, 𝐹𝑖 should be designed as (18) according
to Lemma 2, and 𝛽1 , 𝛽2 , and 𝛽3 in (35)-(37) are any positive constants. It can be seen from (43) and (45) that the larger 𝛽1 , 𝛽2 ,
and 𝛽3 are , the faster the estimation errors of the distributed observer (35)-(37) decays. It can also be seen from (17) that the
larger 𝜇𝑖 for 𝑖 ∈ are, the faster the estimation errors decays for the observer (14). Choosing sufficiently large 𝜇𝑖 for 𝑖 ∈ 𝐹 , 𝛽1 ,
𝛽2 , and 𝛽3 makes the convergence of the observers sufficiently fast, and therefore, their effects on the control performance are
negligible. The weight matrices 𝑄𝑖 and 𝑊𝑖 for 𝑖 ∈ are design parameters and they should be chosen as symmetric positive
definite matrices. The discount factors 𝛾𝑖 for 𝑖 ∈ are used to guarantee that the performance functions is bounded for given
control policies, and to do so, they should be chosen to satisfy condition (62) in Remark 12.
5.2
Distributed Optimal Output Containment Control: An Online and Optimal Solution
In this subsection, an off-policy RL method 58 is combined with actor-critic algorithm to solve discounted ARE (61) and learn
the optimal gain (60) online in real time, using relative output measurements of followers with respect to their neighbors and
the information which is broadcasted by neighbors, and without requiring complete knowledge of the leaders’ dynamics.
Since Value function (57) is in the form of a quadratic polynomial, the quadratic polynomial basis vector for the critic neural
network (NN) of each follower is chosen as
𝜎𝑐
𝜎𝑐
2
2
𝜎𝑖𝑐 (𝑋̄ 𝑖 ) = [ 𝑋̄ 𝑖1 𝑋̄ 𝑖1 𝑋̄ 𝑖2 … 𝑋̄ 𝑖𝑁 𝜎𝑐 ]𝑇 ∈ ℜ𝑁𝑖 (𝑁𝑖 +1)∕2
(70)
𝑖
𝜎
where 𝑋̄ 𝑖 = [𝑋̄ 𝑖1 , ..., 𝑋̄ 𝑖𝑁𝑖𝜎𝑐 ]𝑇 and 𝑁𝑖 𝑐 = 𝑁𝑖 + 𝑞.
̄ Then, the optimal value function 𝑉𝑖 (𝑋̄ 𝑖 ) can be perfectly approximated by
(
)𝑇
(71)
𝑉𝑖 (𝑋̄ 𝑖 ) = 𝑊̄ 𝑖𝑐 𝜎𝑖𝑐 (𝑋̄ 𝑖 )
/
𝜎𝑐
𝜎
𝜎
𝜎
where 𝑊̄ 𝑖𝑐 ∈ ℜ𝑀𝑖 is the optimal critic NN weight of 𝑖th follower and 𝑀𝑖 𝑐 = 𝑁𝑖 𝑐 (𝑁𝑖 𝑐 + 1) 2 is the number of hidden-layer
neurons of the critic NN of 𝑖th follower.
By the same token, the optimal policy in (59) can be perfectly approximated by an actor NN for each follower in the form of
(
)𝑇
(72)
𝑢𝑖 ∗ (𝑋̄ 𝑖 ) = 𝑊̄ 𝑖𝑎 𝜎𝑖𝑎 (𝑋̄ 𝑖 )
𝜎
𝑎
𝜎
𝜎
where 𝜎𝑎 (𝑋̄ 𝑖 ) = [𝑋̄ 𝑖1 , ..., 𝑋̄ 𝑖𝑀𝑖𝜎𝑎 ]𝑇 , 𝑊̄ 𝑖𝑎 ∈ ℜ𝑀𝑖 ×𝑝𝑖 is the optimal actor NN weight of 𝑖th follower and 𝑀𝑖 𝑎 = 𝑁𝑖 𝑐 is the number
of hidden-layer neurons of the actor NN of 𝑖th follower, for 𝑖 ∈ . The ideal weights of the critic NNs, i.e., 𝑊̄ 𝑖𝑐 , and the ideal
weights of the actor NNs, i.e., 𝑊̄ 𝑖𝑎 , are unknown and must be estimated.
Let the value function corresponding to 𝑢̂ 𝑘𝑖 be written as follows
∞
𝑇
𝑇
𝑉̂𝑖𝑘 (𝑋̄ 𝑖 ) = ∫ 𝑒−𝛾𝑖 (𝜏−𝑡) 𝑋̄ 𝑖𝑇 (𝐶̄ 𝑖 𝑄𝑖 𝐶̄ 𝑖 + 𝐾̄ 𝑘𝑖 𝑊𝑖 𝐾̄ 𝑖𝑘 )𝑋̄ 𝑖 𝑑𝜏 = 𝑋̄ 𝑖𝑇 Ψ𝑘𝑖 𝑋̄ 𝑖
(73)
𝑢̂ 𝑘𝑖 = 𝐾̄ 𝑖𝑘 𝑋̄ 𝑖
(74)
𝐾̄ 𝑖𝑘 = −𝑊𝑖−1 𝐵̄ 𝑖𝑇 Ψ𝑖 𝑘
(75)
𝑡
where
with
∗
𝑘
is the estimation of 𝑢𝑖 in the 𝑘th iteration, and Ψ𝑖 satisfies the discounted ARE (80).
Let the estimation of 𝑊̄ 𝑖𝑐 in the 𝑘th iteration be denoted by 𝑊̂̄ 𝑐𝑖 𝑘 . Then, the value function (71) and its gradient ∇𝑉𝑖 (𝑋̄ 𝑖 ) =
/
𝜕𝑉𝑖 𝜕 𝑋̄ 𝑖 are approximated as
𝑉̂𝑖𝑘 (𝑋̄ 𝑖 ) = (𝑊̄̂ 𝑐𝑖 𝑘 )𝑇 𝜎𝑖𝑐 (𝑋̄ 𝑖 )
(76)
14
∇𝑉̂𝑖𝑘 (𝑋̄ 𝑖 ) = (∇𝜎𝑖𝑐 (𝑋̄ 𝑖 ))𝑇 𝑊̂̄ 𝑐𝑖 𝑘
(77)
Correspondingly, let 𝑊̂̄ 𝑎𝑖 𝑘 denote the estimation of 𝑊̄ 𝑖𝑎 in the 𝑘th iteration. Then, the estimation of u∗𝑖 (𝑋̄ 𝑖 ) in the 𝑘th iteration
is
𝑢̂ 𝑘𝑖 (𝑋̄ 𝑖 ) = (𝑊̂̄ 𝑎𝑖 𝑘 )𝑇 𝜎𝑖𝑎 (𝑋̄ 𝑖 )
(78)
One can now write (53) in following augmented form as
𝑋̄̇ 𝑖 = 𝑃 𝑘 𝑋̄ 𝑖 + 𝐵̄ 𝑖 (𝑢𝑖 − 𝑢̂ 𝑘 )
(79)
𝑖
𝑖
where 𝑃𝑖𝑘 = 𝑃𝑖 + 𝐵̄ 𝑖 𝐾̄ 𝑖𝑘 and 𝑢𝑖 is a behavior policy, which is an admissible policy, applied to the 𝑖th follower and 𝑢̂ 𝑘𝑖 is the
estimation of target policy, which is the optimal policy, in the 𝑘th iteration.
Using (73)-(75), (79) and some manipulation, the discounted ARE (61) becomes
𝑇
Ψ𝑘𝑖 𝑃𝑖𝑘 + (𝑃𝑖𝑘 )𝑇 Ψ𝑘𝑖 − 𝛾𝑖 Ψ𝑘𝑖 + 𝐶̄𝑖 𝑄𝑖 𝐶̄𝑖 − (𝐾̄ 𝑖𝑘 )𝑇 𝑊𝑖 𝐾̄ 𝑖𝑘 = 0
(80)
Consider the control policy update law as follows
𝑢𝑘+1
= −𝑊𝑖−1 𝐵̄ 𝑖𝑇 Ψ𝑘𝑖 𝑋̄ 𝑖
𝑖
(81)
The derivative of (73) with respect to time along the system dynamics (79) can be derived as
𝑉̂̇ 𝑘 (𝑋̄ ) = 𝑋̄ 𝑇 (Ψ𝑘 𝑃 𝑘 + (𝑃 𝑘 )𝑇 Ψ𝑘 )𝑋̄ + 2𝑋̄ 𝑇 Ψ𝑘 𝐵̄ (𝑢 − 𝑢̂ 𝑘 )
𝑖
𝑖
𝑖
𝑖
𝑖
𝑖
𝑖
𝑖
𝑖
𝑖
𝑖
𝑖
𝑖
(82)
𝑇
𝑇
= 𝛾𝑖 𝑉̂𝑖𝑘 (𝑋̄ 𝑖 ) − 𝑋̄ 𝑖𝑇 𝐶̄ 𝑖 𝑄𝑖 𝐶̄ 𝑖 𝑋̄ 𝑖 − (𝑢̂ 𝑘𝑖 )𝑇 𝑊𝑖 𝑢̂ 𝑘𝑖 − 2(𝑢̂ 𝑘+1
̂ 𝑘𝑖 )
𝑖 ) 𝑊𝑖 (𝑢𝑖 − 𝑢
Multiplying 𝑒−𝛾𝑖 (𝜏−𝑡) to the both sides of (82), and integrating both sides on the time interval of [𝑡; 𝑡 + 𝑇 ], one can obtain the
off-policy Bellman equation as follows
𝑡+𝑇
𝑒−𝛾𝑖 𝑇 𝑉̂𝑖𝑘 (𝑋̄ 𝑖 (𝑡 + 𝑇 )) − 𝑉̂𝑖𝑘 (𝑋̄ 𝑖 (𝑡)) = ∫
𝑡
𝑡+𝑇
𝑑
( 𝑒−𝛾𝑖 (𝜏−𝑡) 𝑉̂𝑖𝑘 (𝑋̄ 𝑖 ))𝑑𝜏
𝑑𝜏
=
(83)
𝑡+𝑇
−𝛾𝑖 (𝜏−𝑡)
∫ 𝑒
𝑡
𝑇
𝑇
𝑇
(−𝑋̄ 𝑖𝑇 𝐶̄ 𝑖 𝑄𝑖 𝐶̄ 𝑖 𝑋̄ 𝑖 − (𝑢̂ 𝑘𝑖 ) 𝑊𝑖 𝑢̂ 𝑘𝑖 )𝑑𝜏 + ∫ 𝑒−𝛾𝑖 (𝜏−𝑡) (−2(𝑢̂ 𝑘+1
̂ 𝑘𝑖 ))𝑑𝜏
𝑖 ) 𝑊𝑖 (𝑢𝑖 − 𝑢
𝑡
Remark 14. Note that to calculate the first integrand in the right-hand-side of (83), the knowledge of the leaders’ dynamic 𝐷 is
(
)𝑇
𝑇
𝑇
required due to dependency of 𝐶̄𝑖 𝑄𝑖 𝐶̄𝑖 to it, i.e., 𝑋̄ 𝑖𝑇 𝐶̄𝑖 𝑄𝑖 𝐶̄𝑖 𝑋̄ 𝑖 = 𝑦𝑖 − 𝐷𝜔∗𝑖 𝑄𝑖 (𝑦𝑖 − 𝐷𝜔∗𝑖 ). To obviate this requirement, the
(
)𝑇
𝑇
term of 𝑋̄ 𝑇 𝐶̄𝑖 𝑄𝑖 𝐶̄𝑖 𝑋̄ 𝑖 , in first integrand in the right-hand-side of (83), is replaced by 𝑦̂𝑖 − 𝑦0𝑖 𝑄𝑖 (𝑦̂𝑖 − 𝑦0𝑖 ) , where 𝑦0𝑖 is the
𝑖
estimation of convex combination of the leaders’ outputs, defined in (38) and 𝑦̂𝑖 = 𝐶𝑖 𝜉𝑖 is the output estimation of 𝑖th follower.
𝑝
Define 𝜐𝑘𝑖 = 𝑢𝑖 − 𝑢̂ 𝑘𝑖 , exploiting the critic NN (76), actor NN (78), weighted matrices of 𝑊𝑖 = 𝑑𝑖𝑎𝑔(𝑤1𝑖 , ..., 𝑤𝑖 𝑖 ), and Remark
14, (83) can be put into the following form
𝑇
∑
𝑒𝑘𝑖 = (𝑊̂̄ 𝑐𝑖 𝑘 )𝑇 [𝑒−𝛾𝑖 𝑇 𝜎𝑖𝑐 (𝑋̄ 𝑖 (𝑡 + 𝑇 )) − 𝜎𝑖𝑐 (𝑋̄ 𝑖 (𝑡))] + 2 (𝑊̂̄ 𝑎𝑖,𝑗 𝑘+1 ) 𝑤𝑗𝑖 ∫ 𝑒−𝛾𝑖 (𝜏−𝑡) 𝜎𝑖𝑎 (𝑋̄ 𝑖 )𝜐𝑘𝑖,𝑗 𝑑𝜏
𝑡+𝑇
𝑝𝑖
𝑗=1
𝑡+𝑇
−𝛾𝑖 (𝜏−𝑡)
−∫ 𝑒
𝑡
𝑇
(−(𝐶𝑖 𝜉𝑖 − 𝐷𝑖 𝜂̂𝑖 ) 𝑄𝑖 (𝐶𝑖 𝜉𝑖 − 𝐷𝑖 𝜂̂𝑖 ) −
𝑡
𝑇
𝑇
𝜎𝑖𝑎 (𝑋̄ 𝑖 ) 𝑊̂̄ 𝑎𝑖 𝑘 𝑊𝑖 (𝑊̂̄ 𝑎𝑖 𝑘 )
(84)
𝜎𝑖𝑎 (𝑋̄ 𝑖 ))𝑑𝜏
where 𝑊̂̄ 𝑎𝑖 𝑘+1 = [ 𝑊̂̄ 𝑎𝑖,1 𝑘+1 ⋯ 𝑊̂̄ 𝑎𝑖,𝑝 𝑘+1 ], 𝜐𝑘𝑖 = [ 𝜐𝑘𝑖,1 ⋯ 𝜐𝑘𝑖,𝑝𝑖 ]𝑇 , and 𝑒𝑘𝑖 is the Bellman approximation error which should be
𝑖
minimized in order to drive the critic NN weights and actor NN weights toward their ideal values, i.e., 𝑊̂̄ 𝑐𝑖 𝑘 → 𝑊̄ 𝑖𝑐 and
𝑊̂̄ 𝑎 𝑘+1 → 𝑊̄ 𝑎 .
𝑖
𝑖
Rearranging (84), the Bellman equation (83) can be reformulated as follows
𝑌𝑖 𝑘 (𝑡) = (𝑊̂ 𝑖𝑘 )𝑇 Ξ𝑘𝑖 (𝑡) − 𝑒𝑘𝑖
(85)
where
𝑡+𝑇
𝑘
𝑌𝑖 (𝑡) =
∫
𝑡
𝑇
𝑇
𝑒−𝛾𝑖 (𝜏−𝑡) (−(𝐶𝑖 𝜉𝑖 − 𝐷𝑖 𝜂̂𝑖 )𝑇 𝑄𝑖 (𝐶𝑖 𝜉𝑖 − 𝐷𝑖 𝜂̂𝑖 ) − 𝜎𝑖𝑎 (𝑋̄ 𝑖 ) 𝑊̂̄ 𝑎𝑖 𝑘 𝑊𝑖 (𝑊̂̄ 𝑎𝑖 𝑘 ) 𝜎𝑖𝑎 (𝑋̄ 𝑖 ))𝑑𝜏
(86)
15
𝜎
𝜎
𝑐
𝑎
𝑊̂ 𝑖𝑘 = [(𝑊̂̄ 𝑐𝑖 𝑘 )𝑇 , (𝑊̂̄ 𝑎𝑖,1 𝑘+1 )𝑇 , ..., (𝑊̂̄ 𝑎𝑖,𝑝 𝑘+1 )𝑇 ]𝑇 ∈ ℜ𝑀𝑖 +𝑝𝑖 ×𝑀𝑖
(87)
⎡ 𝑒−𝛾𝑖 𝑇 𝜎𝑖𝑐 (𝑋̄ 𝑖 (𝑡 + 𝑇 )) − 𝜎𝑖𝑐 (𝑋̄ 𝑖 (𝑡)) ⎤
⎢
⎥
𝑡+𝑇
⎢ 2𝑤1 ∫ 𝑒−𝛾𝑖 (𝜏−𝑡) 𝜎 𝑎 (𝑋̄ )𝜐𝑘 𝑑𝜏 ⎥
𝑖
𝑖
𝑖
𝑖,1
𝜎𝑐
𝜎𝑎
⎢
⎥
𝑡
Ξ𝑘𝑖 (𝑡) = ⎢
∈ ℜ𝑀𝑖 +𝑝𝑖 ×𝑀𝑖
⎥
⋮
⎢
⎥
𝑡+𝑇
⎢
⎥
𝑝𝑖
𝑎 ̄
𝑘
−𝛾
(𝜏−𝑡)
𝑖
𝜎𝑖 (𝑋𝑖 )𝜐𝑖,𝑝 𝑑𝜏 ⎥
⎢ 2𝑤𝑖 ∫ 𝑒
𝑖
⎣
⎦
𝑡
(88)
𝑖
𝜎
𝜎
Assume that 𝑌𝑖 𝑘 (𝑡) and Ξ𝑘𝑖 (𝑡) are collected at 𝕟𝑖 ≥ 𝑀𝑖 𝑐 + 𝑝𝑖 × 𝑀𝑖 𝑎 points 𝑡1 to 𝑡𝕟𝑖 , over the same time interval [𝑡; 𝑡 + 𝑇 ].
Using least square method in average sense, one has
𝑊̂ 𝑖𝑘 = (Ξ̄ 𝑘𝑖 (Ξ̄ 𝑘𝑖 )𝑇 )−1 Ξ̄ 𝑘𝑖 𝑌̄𝑖𝑘
(89)
Ξ̄ 𝑘𝑖 = [Ξ𝑘𝑖 (𝑡1 ), ..., Ξ𝑘𝑖 (𝑡𝕟𝑖 )]
(90)
𝑌̄𝑖𝑘 = [𝑌𝑖 𝑘 (𝑡1 ), ..., 𝑌𝑖 𝑘 (𝑡𝕟𝑖 )]𝑇
(91)
where
Remark 15. Note that to calculate the information in (86) and (88) for each follower, the knowledge of graph topology and
absolute states of followers are required due to the dependency of 𝑋̄ 𝑖 (𝑡) = [𝑥𝑖 𝑇 , 𝜔∗𝑖 𝑇 ]𝑇 in (86) and (88) to them. To obviate these
requirements, in Algorithm 2, the term of 𝑋̂̄ 𝑖 (𝑡) = [𝜉𝑖 𝑇 , 𝜂̂𝑖 𝑇 ]𝑇 will be used in place of 𝑋̄ 𝑖 (𝑡) = [𝑥𝑖 𝑇 , 𝜔∗𝑖 𝑇 ]𝑇 in (86) and (88).
To this end, by using 𝑋̂̄ 𝑖 (𝑡) = [𝜉𝑖 𝑇 , 𝜂̂𝑖 𝑇 ]𝑇 in place of 𝑋̄ 𝑖 (𝑡) = [𝑥𝑖 𝑇 , 𝜔∗𝑖 𝑇 ]𝑇 , (89) can be rewritten as follows
𝑊̂ 𝑖𝑘 = (Ξ̂̄ 𝑘𝑖 (Ξ̂̄ 𝑘𝑖 )𝑇 )−1 Ξ̂̄ 𝑘𝑖 𝑌̂̄𝑖𝑘
(92)
Ξ̂̄ 𝑘𝑖 = [Ξ̂ 𝑘𝑖 (𝑡1 ), ..., Ξ̂ 𝑘𝑖 (𝑡𝕟𝑖 )]
(93)
where 𝑊̂ 𝑖𝑘 is given in (87), and
𝑘
𝑘
𝑘
𝑌̂̄𝑖 = [𝑌̂𝑖 (𝑡1 ), ..., 𝑌̂𝑖 (𝑡𝕟𝑖 )]𝑇
(94)
𝑡+𝑇
𝑌̂𝑖𝑘 (𝑡) =
∫
𝑒−𝛾𝑖 (𝜏−𝑡) (−(𝐶𝑖 𝜉𝑖 − 𝐷𝑖 𝜂̂𝑖 )𝑇 𝑄𝑖 (𝐶𝑖 𝜉𝑖 − 𝐷𝑖 𝜂̂𝑖 ) − 𝜎𝑖𝑎 (𝑋̂̄ 𝑖 )𝑇 𝑊̂̄ 𝑎𝑖 𝑘 𝑊𝑖 (𝑊̂̄ 𝑎𝑖 𝑘 )𝑇 𝜎𝑖𝑎 (𝑋̂̄ 𝑖 ))𝑑𝜏
(95)
𝑡
⎡ 𝑒−𝛾𝑖 𝑇 𝜎 𝑐 (𝑋̂̄ 𝑖 (𝑡 + 𝑇 )) − 𝜎 𝑐 (𝑋̂̄ 𝑖 (𝑡)) ⎤
𝑖
𝑖
⎢
⎥
𝑡+𝑇
⎢ 2𝑤1 ∫ 𝑒−𝛾𝑖 (𝜏−𝑡) 𝜎 𝑎 (𝑋̂̄ )𝜐𝑘 𝑑𝜏 ⎥
𝑖 𝑖,1
𝑖
𝑖
⎥
⎢
𝑡
Ξ̂ 𝑘𝑖 (𝑡) = ⎢
(96)
⎥
⋮
⎢
⎥
𝑡+𝑇
⎥
⎢
𝑝
⎢ 2𝑤𝑖 𝑖 ∫ 𝑒−𝛾𝑖 (𝜏−𝑡) 𝜎𝑖𝑎 (𝑋̂̄ 𝑖 )𝜐𝑘𝑖,𝑝𝑖 𝑑𝜏 ⎥
⎦
⎣
𝑡
To simultaneously solve the ARE (83) and find the optimal policy (72) online, i.e., solve Problem 2, the off-policy RL
algorithm is given in Algorithm 1. Note that, Algorithm 1 solve Problem 2 without requiring any knowledge about graph topology , and absolute states or outputs of followers. Moreover, Algorithm 1 solve problem 2 without the restrictive assumption
of which all followers should have knowledge about the leaders’ dynamics 𝑆 and 𝐷.
16
Algorithm 1: Online Off-policy RL Algorithm
1) Start with an admissible control policy 𝑢𝜅𝑖 = 𝐾𝑖𝜅 𝑋̂̄ 𝑖 + 𝜀, where 𝜀 is the exploration noise and 𝜅 = 0, and collect required
information, i.e., Ξ̂̄ 𝑘𝑖 and 𝑌̂̄𝑖𝑘 .
2) Solve the least square problem (92) to obtain 𝑊̂̄ 𝑐𝑖 𝑘 and 𝑊̂̄ 𝑎𝑖 𝑘+1 simultaneously.
3) Let 𝜅 = 𝜅 + 1, and repeat step 2 until ‖𝑊̂ 𝑖𝑘 − 𝑊̂ 𝑖𝑘−1 ‖ ≤ 𝜏, where 𝜏 is a small predefined positive constant.
4) On convergence set 𝑢∗ = 𝑢̂ 𝑘 = (𝑊̂̄ 𝑎 𝑘 )𝑇 𝜎 𝑎 (𝑋̂̄ 𝑖 ) as the optimal control policy.
𝑖
𝑖
𝑖
𝑖
Remark 16. Note that, due to the terms 𝜂̂𝑖 and 𝜉𝑖 in (94) and (96), information about Ξ̂̄ 𝑘𝑖 and 𝑌̂̄𝑖𝑘 , which are required in Algorithm
1, cannot be collected and used from the beginning of the learning process. However, once the distributed observers in (14) and
(35)-(37) converge, information about Ξ̂̄ 𝑘𝑖 and 𝑌̂̄𝑖𝑘 is collected, and then, it is used by Algorithm 1. Note also that, as mentioned
in Remark 13, the observers gains appear in (14), (35)-(37), and the control gains 𝐾̄ 𝑖∗ in (68) can be designed independently, and
choosing sufficiently large 𝜇𝑖 for 𝑖 ∈ , 𝛽1 , 𝛽2 , and 𝛽3 makes the convergence of the observers sufficiently fast, and therefore,
the effect of this issue on Algorithm 1 is negligible.
6
SIMULATION RESULTS
In this section, an output containment control example is given to validate the proposed approach. Consider the multi-agent
system consists of four heterogeneous followers and three leaders. The leaders’ dynamics are given as
[
]
[
]
1 −3
10
𝑆=
,𝐷 =
(97)
1 −1
01
The initial leaders’ state vectors are chosen as 𝜔5 (0) = [ 2 1 ]𝑇 , 𝜔6 (0) = [ −1 1 ]𝑇 , and 𝜔7 (0) = [ 0.4 0.4 ]𝑇 .
The dynamics of heterogeneous followers are given as
⎡0 0⎤
⎡4⎤
⎡ −1 0 0 ⎤
𝑇
⎥
⎥
⎢
⎢
𝐴1 = 0 3 0 , 𝐵1 = 1 , 𝐶1 = ⎢ 1 0 ⎥
⎥
⎢
⎥
⎢ ⎥
⎢
⎣0 1⎦
⎣1⎦
⎣ 0 3 2⎦
[
[
]
]
[
]
10
1 −1
−2
𝑇
𝐴2 =
, 𝐵2 =
, 𝐶2 =
01
1 0
−1
[
]
]
[
]
10
20
−1
𝑇
𝐴3 =
, 𝐵3 =
, 𝐶3 =
01
22
−1
[
(98)
⎡ −1 0 0 ⎤
⎡5⎤
⎡0 0⎤
𝐴4 = ⎢ 0 2 −1 ⎥ , 𝐵4 = ⎢ 1 ⎥ , 𝐶4𝑇 = ⎢ 1 0 ⎥ .
⎢
⎥
⎢ ⎥
⎢
⎥
⎣ 0 3 3 ⎦
⎣2⎦
⎣0 1⎦
The communication graph among the agents is given in Fig. 1, where nodes 5, 6, and 7 represent the leaders and other nodes
represent four heterogeneous agents. Moreover, all the communication weights, in Fig. 1, are chosen to be one. It can be verified
that Assumptions 1-5 are satisfied.
The distributed observer (14) is implemented for 𝑖 = 1, 2, 3, 4. The observer matrices and gains are set as
[
]
⎡2 0 0⎤
0.19 −0.11
⎢
⎥
Φ1 = 𝐼3 , 𝐸1 = 0 1 0 , 𝑅1 =
⎢
⎥
−0.11 0.26
⎣0 0 1⎦
17
FIGURE 1 The multi-agent systems communication graph.
x1 − ξ1
x2 − ξ2
x3 − ξ3
x4 − ξ4
2.5
2
2
1.5
1
X̃
1
0
0.5
−1
0
0.5
1
1.5
2
0
−0.5
−1
0
2
4
6
8
10
12
14
16
18
20
Time (Sec)
FIGURE 2 Error between state observer (14) and followers’ states.
[
Φ2 = 𝐼2 , 𝐸2 = 𝐼2 , 𝑅2 =
[
0.33 0
0 1
]
0.23 −0.09
Φ3 = 𝐼2 , 𝐸3 = 𝐼2 , 𝑅3 =
−0.09 0.23
]
(99)
[
]
⎡2 0 0⎤
0.22 −0.06
,
Φ4 = 𝐼3 , 𝐸4 = ⎢ 0 1 0 ⎥ , 𝑅4 =
⎥
⎢
−0.06 0.16
⎣0 0 1⎦
and 𝜇𝑖 = 1, for 𝑖 = 1, 2, 3, 4. The error between state observer (14) and followers’ states is given in Fig. 2. It can be observed
from Fig. 2 that the error between state observers and the followers’ states converges to zero.
The distributed adaptive observer (35) along with adaptation laws (36) and (37) are also implemented for 𝑖 = 1, 2, 3, 4. The
observer and adaptive laws gains are set as 𝛽1 = 3, 𝛽2 = 10, and 𝛽3 = 3. It can be seen from Fig. 3 that the error between
observers (35) and the convex combination of the leaders’ states converges to zero, i.e., the adaptive observers (35) converge to
the convex combination of the leaders’ states.
18
0.2
η̃1
η̃2
η̃3
η̃4
0.15
0.1
0.05
η̃ i
0
−0.05
0
−0.1
−0.15
−0.5
−0.2
−1
−0.25
0
−0.3
0
0.5
1
5
1.5
10
15
20
Time (Sec)
FIGURE 3 Error between the adaptive observer (35) and the convex combination of the leaders’ states.
By solving the output regulator equation (6), one can obtain
⎡ 4 −16 ⎤
[
]
Π1 = ⎢ 1 0 ⎥ , Γ1 = −2 −3
⎥
⎢
⎣0 1 ⎦
[
]
Π2 = 𝐼2 , Γ2 = 0 1
(100)
[
]
Π3 = 𝐼2 , Γ3 = 1 3
⎡ 3.33 −11.66 ⎤
[
]
Π4 = ⎢ 1
0 ⎥ , Γ4 = −1 −2 .
⎥
⎢
1 ⎦
⎣ 0
The weight matrices 𝑄𝑖 , 𝑊𝑖 , and the discount factors 𝛾𝑖 for 𝑖 = 1, 2, 3, 4 are chosen as
𝑄1 = 10𝐼3 , 𝑊1 = 10, 𝛾1 = 0.01
𝑄2 = 10𝐼2 , 𝑊2 = 10, 𝛾2 = 0.01
𝑄3 = 10𝐼2 , 𝑊3 = 10, 𝛾3 = 0.01
𝑄4 = 10𝐼3 , 𝑊4 = 10, 𝛾4 = 0.01.
(101)
Using (98), (101), and LQR method, the optimal dynamic output feedback gains are
[
]
𝐾11 = −0.13 7.91 −20.65
[
]
𝐾21 = 1 1.78
[
]
𝐾31 = −0.23 8.61
[
]
𝐾41 = −0.29 7.02 −10.1 .
(102)
We now use Algorithm 1 to find the optimal dynamic output feedback control online in real-time. To do so, time interval is
𝜎
𝜎
𝜎
𝜎
𝜎
𝜎
𝜎
𝜎
𝜎
set as 𝑇 = 0.5 Sec, and 𝑁1 𝑐 = 5, 𝑁2 𝑐 = 4, 𝑁3 𝑐 = 4, 𝑁4 𝑐 = 5, 𝑀1 𝑐 = 15, 𝑀2 𝑐 = 10, 𝑀3 𝑐 = 10, 𝑀4 𝑐 = 15, 𝑀1 𝑎 = 5,
𝜎
𝜎
𝜎
2 ̂̄
𝑋 𝑖1 𝑋̂̄ 𝑖2 … 𝑋̂̄ 2 𝜎𝑐 ]𝑇 , 𝜎 𝑎 (𝑋̂̄ 𝑖 ) = [ 𝑋̂̄ 𝑖1 ⋯ 𝑋̂̄ 𝑖𝑀 𝜎𝑎 ]𝑇 , for 𝑖 = 1, 2, 3, 4.
𝑀 𝑎 = 4, 𝑀 𝑎 = 4, 𝑀 𝑎 = 5, 𝜎 𝑐 (𝑋̂̄ 𝑖 ) = [ 𝑋̂̄ 𝑖1
2
3
4
𝑖
𝑖𝑁𝑖
𝑖
𝑖
Figs. 4 and 5 show the evaluation of learned optimal dynamic output feedback controls along with the state observer (14), and
the adaptive observer (35) with adaptation laws (36) and (37) for the heterogeneous multi-agent systems (98). Fig. 4 confirms
that the followers move into the convex hull formed by the leaders’ outputs. Note that, there exist directed paths from the leaders
5 and 6 to followers 1 and 2 and there is no directed path from the leader 7 to these followers, i.e., the leader 7 is disconnected
19
Leaders
Follower
Follower
Follower
Follower
2
1.5
1
2
3
4
t= 4 Sec
1
yi2
0.5
0
t= 7.2 Sec
−0.5
−1
−1.5
t= 15 Sec
−2
−3
−2
−1
0
yi1
1
2
3
FIGURE 4 The outputs of the agents using Algorithm 1, along with adaptive observer (35) and state observer (14).
4
Leaders
Envelope
Followers
yi1
2
0
−2
0
5
10
15
yi2
4
20
Leaders
Envelope
Followers
2
0
−2
0
5
10
15
20
Time (Sec)
FIGURE 5 The outputs of all four heterogeneous agents.
from the followers 1 and 2. In this case, the convex hull is a line segment which is spanned by the outputs of these two leaders.
Therefore, these two followers converge into the line segment which is spanned by the outputs of the leaders 5 and 6. Note also
that, there exists a directed path from the leaders 5, 6, and 7 to the followers 3 and 4. Therefore, these two followers converge
into the triangular area which is spanned by the outputs of the leaders 5, 6, and 7. The followers and the leaders’ outputs along
with the envelope of the leaders’ outputs are shown in Fig. 5. It can be seen in Fig. 5 that the followers’ outputs move into the
envelopes formed by the leaders’ outputs and stay in them. The time history of output containment errors of followers is shown
in Fig. 6. It can be observed that the containment errors decay to zero. Therefore, it can be seen that the followers’ outputs
converge to the convex hull formed by leaders’ outputs, i.e., the containment control problem 1 is solved.
These results show that the introduced approach solves Problem 2 without requiring all followers to have knowledge of
leaders’ dynamics or states and based on only relative output measurements of followers with respect to their neighbors and the
information which is broadcasted to them through the communication network.
20
i=1
i=2
i=3
i=4
4
ei1
2
0
−2
0
5
10
15
i=1
i=2
i=3
i=4
2
ei2
20
0
−2
−4
0
5
10
15
20
Time (Sec)
FIGURE 6 Time history of output containment errors.
7
CONCLUSIONS
In this paper, an online optimal relative output-feedback based solution for the output containment control problem of linear
multi-agent systems was presented. The followers were assumed heterogeneous in dynamics and dimensions. First, a distributed
dynamic relative output-feedback control protocol was developed based on cooperative output regulation framework, which was
provided the offline and non-optimal solution to the output containment control problem in hand. The proposed control protocol
was composed of a feedback part and a feed-forward part. However, the feedback and feed-forward states were assumed to be
unavailable to followers and were estimated using two distributed observers. A distributed adaptive observer was designed to
estimate followers’ states using only relative output measurements of followers with respect to their neighbors and the information which is broadcasted to them through the communication network. Another distributed observer was developed to estimate
the convex hull of the leaders’ states. To relax the restrictive assumption in existent work that each follower should have the
knowledge of the leader’s dynamics, an adaptive distributed observer was next designed to estimate both the leaders’ dynamics
and convex combination of the leaders’ states. Optimality was explicitly imposed in finding the feedback and feed-forward control gains to not only assure convergence of followers’ outputs to a convex combination of the leaders’ outputs, but also optimize
their transient output containment errors. To do this, augmented AREs were employed to solve the optimal output containment
control problem in hand. An off-policy RL algorithm on an actor-critic structure was next developed to solve these AREs online
in real time, based on using only relative output measurements of followers with respect to their neighbors and the information
which is broadcasted to them through the communication network, and without requirement of knowing the complete knowledge
of the leaders’ dynamics by all followers. Finally, a simulation example verified the effectiveness of the proposed algorithm.
References
1. Jadbabaie A, Jie Lin, Morse AS. Coordination of groups of mobile autonomous agents using nearest neighbor rules. IEEE
Transactions on Automatic Control. 2003;48(6):988-1001.
2. Ji M, Ferrari-Trecate G, Egerstedt M, Buffa A. Containment Control in Mobile Networks. IEEE Transactions on Automatic
Control. 2008;53(8):1972-1975.
3. Russell Carpenter J. Decentralized control of satellite formations. International Journal of Robust and Nonlinear Control.
2002;12(2-3):141-161.
4. Fax JA, Murray RM. Information Flow and Cooperative Control of Vehicle Formations. IEEE Transactions on Automatic
Control. 2004;49(9):1465-1476.
21
5. Tomlin C, Pappas GJ, Sastry S. Conflict resolution for air traffic management: a study in multiagent hybrid systems. IEEE
Transactions on Automatic Control. 1998;43(4):509-521.
6. Wei Ren, Beard RW, Atkins EM. Information consensus in multivehicle cooperative control. IEEE Control Systems
Magazine. 2007;27(2):71-82.
7. Olfati-Saber R, Shamma JS. Consensus Filters for Sensor Networks and Distributed Sensor Fusion. In: Proceedings of the
44th IEEE Conference on Decision and Control. IEEE.
8. Olfati-Saber R, Fax JA, Murray RM. Consensus and Cooperation in Networked Multi-Agent Systems. Proceedings of the
IEEE. 2007;95(1):215-233.
9. Wei Ren, Beard RW. Consensus seeking in multiagent systems under dynamically changing interaction topologies. IEEE
Transactions on Automatic Control. 2005;50(5):655-661.
10. Hong Y, Hu J, Gao L. Tracking control for multi-agent consensus with an active leader and variable topology. Automatica.
2006;42(7):1177-1182.
11. Hong Y, Chen G, Bushnell L. Distributed observers design for leader-following control of multi-agent networks. Automatica. 2008;44(3):846-850.
12. Lou Y, Hong Y. Target containment control of multi-agent systems with random switching interconnection topologies.
Automatica. 2012;48(5):879-885.
13. Notarstefano G, Egerstedt M, Haque M. Containment in leader-follower networks with switching communication topologies. Automatica. 2011;47(5):1035-1040.
14. Meng Z, Ren W, You Z. Distributed finite-time attitude containment control for multiple rigid bodies. Automatica.
2010;46(12):2092-2099.
15. Cao Y, Ren W, Egerstedt M. Distributed containment control with multiple stationary or dynamic leaders in fixed and
switching directed networks. Automatica. 2012;48(8):1586-1597.
16. Liu H, Xie G, Wang L. Necessary and sufficient conditions for containment control of networked multi-agent systems.
Automatica. 2012;48(7):1415-1422.
17. Liu H, Xie G, Wang L. Containment of linear multi-agent systems under general interaction topologies. Systems and Control
Letters. 2012;61(4):528-534.
18. Mei J, Ren W, Ma G. Distributed containment control for Lagrangian networks with parametric uncertainties under a
directed graph. Automatica. 2012;48(4):653-659.
19. Li J, Ren W, Xu S. Distributed Containment Control with Multiple Dynamic Leaders for Double-Integrator
Dynamics Using Only Position Measurements. IEEE Transactions on Automatic Control. 2012;57(6):1553-1559.
doi:10.1109/tac.2011.2174680.
20. Yoo SJ. Distributed adaptive containment control of uncertain nonlinear multi-agent systems in strict-feedback form.
Automatica. 2013;49(7):2145-2153.
21. Li Z, Ren W, Liu X, Fu M. Distributed containment control of multi-agent systems with general linear dynamics in the
presence of multiple leaders. International Journal of Robust and Nonlinear Control. 2011;23(5):534-547.
22. Zheng Y, Wang L. Containment control of heterogeneous multi-agent systems. International Journal of Control.
2013;87(1):1-8.
23. Wang Y, Cheng L, Hou Z-G, Tan M, Wang M. Containment control of multi-agent systems in a noisy communication
environment. Automatica. 2014;50(7):1922-1928.
24. Liu H, Cheng L, Tan M, Hou Z-G. Containment control of continuous-time linear multi-agent systems with aperiodic
sampling. Automatica. 2015;57:78-84.
22
25. Haghshenas H, Badamchizadeh MA, Baradarannia M. Containment control of heterogeneous linear multi-agent systems.
Automatica. 2015;54:210-216.
26. Kan Z, Klotz JR, Pasiliao EL Jr, Dixon WE. Containment control for a social network with state-dependent connectivity.
Automatica. 2015;56:86-92.
27. 1. Kan Z, Shea JM, Dixon WE. Leader-follower containment control over directed random graphs. Automatica. 2016;66:5662. doi:10.1016/j.automatica.2015.12.016.
28. Dimarogonas DV, Tsiotras P, Kyriakopoulos KJ. Leader-follower cooperative attitude control of multiple rigid bodies.
Systems and Control Letters. 2009;58(6):429-435.
29. Mahalik NP, ed. Chapter 13 - Sensor Networks and Configuration. Springer Berlin Heidelberg; 2007.
30. Shames I, Fidan B, Anderson BDO. Minimization of the effect of noisy measurements on localization of multi-agent
autonomous formations. Automatica. 2009;45(4):1058-1065.
31. Wen G, Zhao Y, Duan Z, Yu W, Chen G. Containment of Higher-Order Multi-Leader Multi-Agent Systems: A Dynamic
Output Approach. IEEE Transactions on Automatic Control. 2016;61(4):1135-1140.
32. Zuo S, Song Y, Lewis FL, Davoudi A. Output Containment Control of Linear Heterogeneous Multi-Agent Systems Using
Internal Model Principle. IEEE Transactions on Cybernetics. 2017;47(8):2099-2109.
33. Sutton RS, Barto AG. Reinforcement Learning-An Introduction. Cambridge, MA, USA: MIT Press, 1998.
34. Powell WB. Approximate Dynamic Programming: Solving the Curses of Dimensionality. Hoboken, NJ, USA: Wiley, 2007.
35. Cui X, Zhang H, Luo Y, Jiang H. Adaptive dynamic programming for Hinf tracking design of uncertain nonlinear systems
with disturbances and input constraints. International Journal of Adaptive Control and Signal Processing. 2017;31(11):15671583.
36. Moghadam R, Lewis FL. Output-feedback Hinf quadratic tracking control of linear systems using reinforcement learning.
International Journal of Adaptive Control and Signal Processing. October 2017.
37. Vamvoudakis KG, Lewis FL, Dixon WE. Open-loop Stackelberg learning solution for hierarchical control problems.
International Journal of Adaptive Control and Signal Processing. October 2017.
38. Yasini S, Karimpour A, Naghibi Sistani M-B, Modares H. Online concurrent reinforcement learning algorithm to solve twoplayer zero-sum games for partially unknown nonlinear continuous-time systems. International Journal of Adaptive Control
and Signal Processing. 2014;29(4):473-493.
39. Modares H, Lewis FL, Sistani M-BN. Online solution of nonquadratic two-player zero-sum games arising in the Hinf control
of constrained input systems. International Journal of Adaptive Control and Signal Processing. 2012;28(3-5):232-254.
40. Song R, Lewis FL, Wei Q, Zhang H. Off-Policy Actor-Critic Structure for Optimal Control of Unknown Systems With
Disturbances. IEEE Transactions on Cybernetics. 2016;46(5):1041-1050.
41. Luo B, Liu D, Huang T, Wang D. Model-Free Optimal Tracking Control via Critic-Only Q-Learning. IEEE Transactions
on Neural Networks and Learning Systems. 2016;27(10):2134-2144.
42. Yang X, He H, Liu D. Event-Triggered Optimal Neuro-Controller Design With Reinforcement Learning for Unknown
Nonlinear Systems. IEEE Transactions on Systems, Man, and Cybernetics: Systems. 2017:1-13.
43. Luo B, Liu D, Wu H-N. Adaptive Constrained Optimal Control Design for Data-Based Nonlinear Discrete-Time Systems
With Critic-Only Structure. IEEE Transactions on Neural Networks and Learning Systems. 2017:1-13.
44. Modares H, Nageshrao SP, Lopes GAD, Babuska R, Lewis FL. Optimal model-free output synchronization of heterogeneous
systems using off-policy reinforcement learning. Automatica. 2016;71:334-341.
23
45. Adib Yaghmaie F, Lewis FL, Su R. Output regulation of heterogeneous linear multi-agent systems with differential graphical
game. International Journal of Robust and Nonlinear Control. 2015;26(10):2256-2278.
46. Tatari F, Naghibi-Sistani M-B, Vamvoudakis KG. Distributed learning algorithm for non-linear differential graphical games.
Transactions of the Institute of Measurement and Control. 2016;39(2):173-182.
47. Tatari F, Naghibi-Sistani M-B, Vamvoudakis KG. Distributed optimal synchronization control of linear networked systems
under unknown dynamics. In: 2017 American Control Conference (ACC). IEEE; 2017.
48. Mazouchi M, Naghibi-Sistani MB, Sani SKH. A novel distributed optimal adaptive control algorithm for nonlinear multiagent differential graphical games. IEEE/CAA Journal of Automatica Sinica. 2018;5(1):331-341.
49. Qin J, Yu C. Cluster consensus control of generic linear multi-agent systems under directed topology with acyclic partition.
Automatica. 2013;49(9):2898-2905.
50. Rockafellar RT. Convex analysis. New Jersey: Princeton University Press, 1972.
51. Huang J. Nonlinear Output Regulation: Theory and Applications. SIAM: Philadelphia, 2004.
52. Huang J. Chapter Fourteen - Certainty Equivalence, Separation Principle, and Cooperative Output Regulation of Multi-agent
Systems by the Distributed Observer Approach. In: Control of Complex Systems. Elsevier; 2016:421-449.
53. Gadewadikar J, Lewis FL, Xie L, Kucera V, Abu-Khalaf M. Parameterization of all stabilizing Hinf static state-feedback
gains: Application to output-feedback design. Automatica, 2007; 43(9):1597-1604.
54. Hongwei Zhang, Lewis FL, Das A. Optimal Design for Synchronization of Cooperative Systems: State Feedback, Observer
and Output Feedback. IEEE Transactions on Automatic Control. 2011;56(8):1948-1952.
55. Wu J, Ugrinovskii V, Allgower F. Observer-based synchronization with relative measurements and unknown neighbour
models. In: 2016 Australian Control Conference (AuCC). IEEE; 2016.
56. Cai H, Lewis FL, Hu G, Huang J. The adaptive distributed observer approach to the cooperative output regulation of linear
multi-agent systems. Automatica. 2017;75:299-305.
57. Lewis FL, Syrmos VL. Optimal Control. John Wiley and Sons: 1995.
58. Modares H, Lewis FL, Jiang Z-P. 𝐻∞ Tracking Control of Completely Unknown Continuous-Time Systems via Off-Policy
Reinforcement Learning. IEEE Transactions on Neural Networks and Learning Systems. 2015;26(10):2550-2562.
| 3cs.SY
|
arXiv:1604.08603v1 [cs.DS] 28 Apr 2016
Decomposing Cubic Graphs into Connected
Subgraphs of Size Three
Laurent Bulteau1 , Guillaume Fertin2 , Anthony Labarre1, Romeo Rizzi3 , and
Irena Rusu2
1
2
Université Paris-Est, LIGM (UMR 8049), CNRS, ENPC, ESIEE Paris, UPEM,
F-77454, Marne-la-Vallée, France
Laboratoire d’Informatique de Nantes-Atlantique, UMR CNRS 6241, Université de
Nantes, 2 rue de la Houssinière, 44322 Nantes Cedex 3, France
3
Department of Computer Science, University of Verona, Italy
Abstract. Let S = {K1,3 , K3 , P4 } be the set of connected graphs of size
3. We study the problem of partitioning the edge set of a graph G into
graphs taken from any non-empty S ′ ⊆ S. The problem is known to be
NP-complete for any possible choice of S ′ in general graphs. In this paper,
we assume that the input graph is cubic, and study the computational
complexity of the problem of partitioning its edge set for any choice
of S ′ . We identify all polynomial and NP-complete problems in that
setting, and give graph-theoretic characterisations of S ′ -decomposable
cubic graphs in some cases.
1
Introduction
General context. Given a connected graph G and a set S of graphs, the Sdecomposition problem asks whether G can be represented as an edge-disjoint
union of subgraphs, each of which is isomorphic to a graph in S. The problem has
a long history that can be traced back to Kirkman [7] and has been intensively
studied ever since, both from pure mathematical and algorithmic point of views.
One of the most notable results in the area is the proof by Dor and Tarsi [3] of the
long-standing “Holyer conjecture” [6], which stated that the S-decomposition
problem is NP-complete when S contains a single graph with at least three edges.
Many variants of the S-decomposition problem have been studied while
attempting to prove Holyer’s conjecture or to obtain polynomial-time algorithms
in restricted cases [12], and applications arise in such diverse fields as traffic
grooming [10] and graph drawing [5]. In particular, Dyer and Frieze [4] studied
a variant where S is the set of connected graphs with k edges for some natural
k, and proved the NP-completeness of the S-decomposition problem for any
k ≥ 3, even under the assumption that the input graph is planar and bipartite
(see Theorem 3.1 in [4]). They further claimed that the problem remains NPcomplete under the additional constraint that all vertices of the input graph
have degree either 2 or 3. Interestingly, if one looks at the special case where
k = 3 and G is a bipartite cubic graph (i.e., each vertex has degree 3), then G
can clearly be decomposed in polynomial time, using K1,3 ’s only, by selecting
1
either part of the bipartition and making each vertex in that set the center of a
K1,3 . This shows that focusing on the case k = 3 and on cubic graphs can lead
to tractable results — as opposed to general graphs, for which when k = 3, and
for any non empty S ′ ⊆ S, the S ′ -decomposition problems all turn out to be
NP-complete [4, 6].
In this paper, we study the S-decomposition problem on cubic graphs in
the case k = 3 — i.e., S = {K1,3 , K3 , P4 }. For any non-empty S ′ ⊆ S, we settle
the computational complexity of the S ′ -decomposition problem by showing
that the problem is NP-complete when S ′ = {K1,3 , P4 } and S ′ = S, while all
the other cases are in P. Table 1 summarises the state of knowledge regarding the complexity of decomposing cubic and arbitrary graphs using connected
subgraphs of size three, and puts our results into perspective.
Allowed subgraphs
Complexity according to graph class
K1,3 K3
P4
cubic
arbitrary
X
in P (Proposition 3)
NP-complete [4, Theorem 3.5]
X
O(1) (impossible)
NP-complete [6]
X
in P [8]
NP-complete [4, Theorem 3.4]
X X
in P (Proposition 6)
NP-complete [4, Theorem 3.5]
X
X
NP-complete (Theorem 2) NP-complete [4, Theorem 3.1]
X
X
in P (Proposition 2)
NP-complete [4, Theorem 3.4]
X X
X
NP-complete (Theorem 3) NP-complete [4, Theorem 3.1]
Table 1. Known complexity results on decomposing graphs using subsets of
{K1,3 , K3 , P4 }.
Terminology. We follow Brandstädt et al. [2] for notation and terminology. All
graphs we consider are simple, connected and nontrivial (i.e. |V (G)| ≥ 2 and
|E(G)| ≥ 1). Given a set S of graphs, a graph G admits an S-decomposition, or
is S-decomposable, if E(G) can be partitioned into subgraphs, each of which is
isomorphic to a graph in S. Throughout the paper, S denotes the set of connected
graphs of size 3, i.e. S = {K3 , K1,3 , P4 }. We study the following problem:
S ′ -decomposition
Input:
a cubic graph G = (V, E), a non-empty set S ′ ⊆ S.
Question: does G admit a S ′ -decomposition?
We let G[U ] denote the subgraph of G induced by U ⊆ V (G). Given a graph
G = (V, E), removing a subgraph H = (V ′ ⊆ V, E ′ ⊆ E) of G consists in
removing edges E ′ from G as well as the possibly resulting isolated vertices.
Finally, let G and G′ be two graphs. Then:
– subdividing an edge {u, v} ∈ E(G) consists in inserting a new vertex w into
that edge, so that V (G) becomes V (G) ∪ {w} and E(G) is replaced with
E(G) \ {u, v} ∪ {u, w} ∪ {w, v};
– attaching G′ to a vertex u ∈ V (G) means building a new graph H by identifying u and some v ∈ V (G′ );
2
– attaching G′ to an edge e ∈ E(G) consists in subdividing e using a new
vertex w, then attaching G′ to w.
Figure 1 illustrates the process of attaching an edge to an edge of the cube
graph, and shows other small graphs that we will occasionally use in this paper.
u
v
u
(a)
v
(b)
(c)
(d)
Fig. 1. (a) Attaching a new edge to {u, v}; (b) the diamond graph; (c) the co-fish
graph; (d) the net graph.
2
Decompositions Without a K1,3
In this section, we study decompositions of cubic graphs that use only P4 ’s or
K3 ’s. Note that no cubic graph is {K3 }-decomposable, since all its vertices have
odd degree. According to Bouchet and Fouquet [1], Kotzig [8] proved that a cubic
graph admits a {P4 }-decomposition iff it has a perfect matching. However, the
proof of the forward direction as presented in [1] is incomplete, as it requires the
use of Proposition 1.(b) below, which is missing from their paper. Therefore, we
provide the following proposition for completeness, together with another result
which will also be useful for the case where S ′ = {K3 , P4 }.
Proposition 1. Let G be a cubic graph that admits a {K3 , P4 }-decomposition
D. Then, in D, (a) no K3 is used, and (b) no three P4 ’s are incident to the
same vertex.
Proof. Partition V (G) into three sets V1 , V2 and V3 , where V1 (resp. V2 , V3 ) is
the set of vertices that are incident to exactly one P4 (resp. two, three P4 ’s) in
D. Note that V1 is exactly the set of vertices involved in K3 ’s in D. Let ni = |Vi |,
1 ≤ i ≤ 3. Our goal is to show that n1 = n3 = 0, i.e. V1 = V3 = ∅. For this, note
that (1) each vertex in V3 is the extremity of three different P4 ’s, (2) each vertex
in V2 is simultaneously the extremity of one P4 and an inner vertex of another
P4 , while (3) each vertex in V1 is the extremity of one P4 . Since each P4 has two
extremities and two inner vertices, if p is the number of P4 ’s in D, we have:
• p=
• p=
3n3 +n2 +n1
2
n2
2 (by (2)
(by (1), (2) and (3) above, counting extremities);
above, counting inner vertices).
Putting together the above two equalities yields n1 = n3 = 0, which completes
the proof.
⊓
⊔
3
Since K3 ’s cannot be used in cubic graphs for {K3 , P4 }-decompositions by
Proposition 1 above, we directly obtain the following result, which implies that
{K3 , P4 }-decomposition is in P.
Proposition 2. A cubic graph admits a {K3 , P4 }-decomposition iff it has a
perfect matching.
3
Decompositions Without a P4
In this section, we study decompositions of cubic graphs that use only K1,3 ’s or
K3 ’s.
Proposition 3. A cubic graph G admits a {K1,3 }-decomposition iff it is bipartite.
Proof. For the reverse direction, select either set of the bipartition, and make
each vertex in that set the center of a K1,3 . For the forward direction, let D be
a {K1,3 }-decomposition of G, and let C and L be the sets of vertices containing,
respectively, all the centers and all the leaves of K1,3 ’s in D. We show that this
is a bipartition of V (G). First, C ∪ L = V since D covers all edges and therefore
all vertices. Second, C ∩ L = ∅ since a vertex in C ∩ L would have degree at least
4. Finally, each edge in D connects the center of a K1,3 and a leaf of another
K1,3 in D, which belong respectively to C and L. Therefore, G is bipartite. ⊓
⊔
We now prove that {K1,3 , K3 }-decompositions can be computed in polynomial time. Recall that a graph is H-free if it does not contain an induced
subgraph isomorphic to a given graph H. Since bipartite graphs admit a {K1,3 }decomposition (by Proposition 3), we can restrict our attention to non-bipartite
graphs that contain K3 ’s (indeed, if they were K3 -free, then only K1,3 ’s would be
allowed and Proposition 3 would imply that they admit no decomposition). Our
strategy consists in iteratively removing subgraphs from G and adding them to
an initially empty {K1,3 , K3 }-decomposition until G is empty, in which case we
have an actual decomposition, or no further removal operations are possible, in
which case no decomposition exists. Our analysis relies on the following notion:
a K3 induced by vertices {u, v, w} in a graph G is isolated if V (G) contains no
vertex x such that {u, v, x}, {u, x, w} or {x, v, w} induces a K3 .
Lemma 1. If a cubic graph G admits a {K1,3 , K3 }-decomposition D, then every
isolated K3 in G belongs to D.
Proof (contradiction). If an isolated K3 were not part of the decomposition,
then exactly one vertex of that K3 would be the center of a K1,3 , leaving the
remaining edge uncovered and uncoverable.
⊓
⊔
C6 is a minimal example of a cubic non-bipartite graph with K3 ’s that admits no {K1,3 , K3 }-decomposition: both K3 ’s in that graph must belong to the
decomposition (by Lemma 1), but their removal yields a perfect matching.
4
Observation 1. Let G be a connected cubic graph. Then no sequence of at least
one edge or vertex removal from G yields a cubic graph.
Proof (contradiction). If after applying at least one removal from G we obtain a
cubic graph G′ , then the graph that precedes G′ in this removal sequence must
have had a vertex of degree at least four, since G is connected.
⊓
⊔
Proposition 4. For any non-bipartite cubic graph G whose K3 ’s are all isolated, one can decide in polynomial time whether G is {K1,3 , K3 }-decomposable.
Proof. We build a {K1,3 , K3 }-decomposition by iteratively removing K1,3 ’s and
K3 ’s from G, which we add as we go to an initially empty set D. By Lemma 1,
all isolated K3 ’s must belong to D, so we start by adding them all to D and
removing them from G; therefore, G admits a {K1,3 , K3 }-decomposition iff the
resulting subcubic graph G′ admits a {K1,3 }-decomposition. Observe that G′
contains vertices of degree 1 and 2; we note that:
1. each vertex of degree 1 must be the leaf of some K1,3 in D;
2. each vertex of degree 2 must be the meeting point of two K1,3 ’s in D.
The only ambiguity arises for vertices of degree 3, which may either be the center
of a K1,3 in D or the meeting point of three K1,3 ’s in D; however, there will
always exist at least one other vertex of degree 1 or 2 until the graph is empty
(by Observation 1.). Therefore, we can safely remove K1,3 ’s from our graph and
add them to D by following the above rules in the stated order; if we succeed in
deleting the whole graph in this way, then D is a {K1,3 , K3 }-decomposition of
G, otherwise no such decomposition exists.
⊓
⊔
We conclude with the case where the graph may contain non-isolated K3 ’s.
Proposition 5. If a cubic graph G contains a diamond, then one can decide in
polynomial time whether G is {K1,3 , K3 }-decomposable.
Proof. The only cubic graph on 4 vertices is K4 , which is diamond-free and
{K1,3 , K3 }-decomposable, so we assume |V (G)| ≥ 6. Let D be a diamond in
G induced by vertices {u, v, w, x} and such that {u, x} 6∈ E(G), as shown in
Figure 2(a). D is connected to two other vertices u′ and x′ of G, which are
respectively adjacent to u and x, and there are only two ways to use the edges
of D in a {K1,3 , K3 }-decomposition, as shown in Figure 2(b) and (c). If u′ = x′ ,
regardless of the decomposition we choose for D, u′ and its neighbourhood induce
a P3 in the graph obtained from G by removing the parts added to D. But
then that P3 cannot be covered, so no {K1,3 , K3 }-decomposition exists for G.
Therefore, we assume that u′ 6= x′ .
As Figure 2(b) and (c) show, either {u, v, w} or {v, w, x} must form a K3 in
D, thereby forcing either {v, w, x, x′ } or {u′ , u, v, w} to form a K1,3 in D. In both
cases, removing the K3 and the K1,3 yields a graph G′ which contains vertices
of degree 1, 2 or 3. As in the proof of Proposition 4, Observation 1 allows us to
make the following helpful observations:
5
1. every leaf in G′ must be the leaf of some K1,3 in D;
2. every vertex y of degree two in G′ must either belong to a K3 or be a leaf
of two distinct K1,3 ’s in D, which can be decided as follows:
(a) if y belongs to a K3 in G′ , then it must also belong to a K3 in D;
otherwise, it would be the leaf of a K1,3 and the graph obtained by
removing that K1,3 would contain a P3 , which we cannot cover;
(b) otherwise, y must be a leaf of two K1,3 ’s in D.
We therefore iteratively remove subgraphs from our graph and add them to
D according to the above rules, which we follow in the stated order; if we succeed
in deleting the whole graph in this way using either decomposition in Figure 2(b)
or (c) as a starting point, then D is a {K1,3 , K3 }-decomposition of G, otherwise
no such decomposition exists.
⊓
⊔
v
u
v
x
u′
u
x′
v
x
u′
u
x′
x
u′
x′
w
w
w
(a)
(b)
(c)
Fig. 2. (a) A diamond in a cubic graph, and (b), (c) the only two ways to decompose
it in a {K1,3 , K3 }-decomposition.
All the arguments developed in this section lead to the following result.
Proposition 6. The {K1,3 , K3 }-decomposition problem on cubic graphs is
in P.
4
Decompositions That Use Both K1,3 ’s and P4 ’s
In this section, we show that problems {K1,3 , P4 }-decomposition and {K1,3 ,
K3 , P4 }-decomposition are NP-complete. Our hardness proof relies on two
intermediate problems that we define below and is structured as follows:
cubic planar monotone 1-in-3 satisfiability
≤P degree-2,3 {K1,3 , K3 , P4 }-decomposition with marked edges
≤P {K1,3 , K3 , P4 }-decomposition with marked edges
≤P {K1,3 , P4 }-decomposition
(Theorem 1 page 10)
(Lemma 4 page 9)
(Lemma 3 page 7)
We start by introducing the following intermediate problem:
{K1,3 , K3 , P4 }-decomposition with marked edges
Input:
a cubic graph G = (V, E) and a subset M ⊆ E of edges.
Question: does G admit a {K1,3 , K3 , P4 }-decomposition D such that no
edge in M is the middle edge of a P4 in D and such that every
K3 in D has either one or two edges in M ?
6
The drawings that illustrate our proofs in this section show marked edges as
dotted edges. The proof of Lemma 3 uses the following result.
Lemma 2. Let e be a bridge in a cubic graph G which admits a {K1,3 , K3 , P4 }decomposition D. Then e must be the middle edge of a P4 in D.
Proof (contradiction). First note that e cannot belong to a K3 in D. Now suppose e is part of a K1,3 in D. The situation is as shown below (without loss of
generality):
e
bank A
bank B
If we remove from G the K1,3 in D that contains e, then summing the terms of
the degree sequence of G[V (B)] yields 2 + 3(|V (B)| − 1) = 2|E(B)|, which means
that 2|E(B)| ≡ 2 (mod 3), so |E(B)| 6≡ 0 (mod 3) and therefore B admits no
decomposition into components of size three. The very same argument shows
that if e belongs to a P4 in D, then it must be its middle edge, which completes
the proof.
⊓
⊔
Lemma 3. Let (G, M ) be an instance of {K1,3 , K3 , P4 }-decomposition with
marked edges, and G′ be the graph obtained by attaching a co-fish to every
edge in M . Then G can be decomposed iff G′ admits a {K1,3 , P4 }-decomposition.
Proof. We prove each direction separately.
⇒: we show how to transform a decomposition D of (G, M ) into a decomposition
D′ of G′ . The subgraphs in D that have no edge in M are not modified. For
the other subgraphs, we distinguish between four cases:
(a) if an edge of M belongs to a K1,3 in D, then attaching a co-fish does not
prevent us from adapting the decomposition of G in G′ :
(b) if an edge of M belongs to a P4 in D, then it is an extremity of that P4
and attaching a co-fish does not prevent us from adapting that part of
the decomposition:
(c) if a K3 in D has one edge in M , we can adapt the partition as follows:
7
(d) if a K3 in D has two edges in M , we can adapt the partition as follows:
⇐: we now show how to transform any {K1,3 , P4 }-decomposition D′ of G′ into a
decomposition of (G, M ). Again, the only parts of D′ that will need adapting
are those connected to the co-fishes that we inserted when transforming G
into G′ . Since the leaf u of the co-fish we inserted has a neighbour x such that
{u, x} is a bridge in G′ , {u, x} is the middle edge of a P4 in D′ (Lemma 2)
and we may therefore assume without loss of generality that our starting
point in G′ is as follows:
v
G′ :
v
x
u w
G:
w
with {v, w} 6∈ E(G′ ) since G is simple; therefore {u, w} cannot belong to a
K3 in G′ , and we have two cases to consider:
(a) if {u, w} belongs to a K1,3 in D′ , that K1,3 can be mapped onto a K1,3
in D by replacing {u, w} with {v, w};
(b) otherwise, {u, w} is an extremal edge of a P4 in D′ ; since {u, w} 6∈ E(G),
either that edge will remain in a P4 when removing the co-fish and
replacing {u, w} with {v, w}, or it will end up in a K3 with either one or
two marked edges. Either way, the part can be added as such to D. ⊓
⊔
We now show that we can restrict our attention to the following variant
of {K1,3 , K3 , P4 }-decomposition with marked edges. We say a graph is
degree-2,3 if its vertices have degree only 2 or 3.
degree-2,3 {K1,3 , K3 , P4 }-decomposition with marked edges
Input:
a degree-2,3 graph G = (V, E) and a subset M ⊆ E of edges.
Question: does G admit a {K1,3 , K3 , P4 }-decomposition D such that no
edge in M is the middle edge of a P4 in D and such that every
K3 in D has either one or two edges in M ?
The following observation will help.
Observation 2. Let G be a degree-2,3 graph with |V2 | degree-2 vertices. If G is
{K1,3 , K3 , P4 }-decomposable, then |V2 | ≡ 0 (mod 3).
8
t1
t1
v1
v2
v1
v3
t2
v2
t3
v3
t2
(a)
t3
(b)
Fig. 3. Adding a net (a) to a graph with degree-2 vertices t1 , t2 , t3 (dotted edges belong
to M ′ ), and (b) its only possible decomposition (up to symmetry).
Proof. If G = (V, E) admits a {K1,3 , K3 , P4 }-decomposition, then |E| ≡ 0
(mod 3). Let V2 and V3 be the subsets of vertices of degree 2 and 3 in G. Then
2|V2 | + 3|V3 | = 2|E|, so 2|V2 | ≡ 0 (mod 3).
⊓
⊔
We prove that allowing degree-2 vertices does not make the problem substantially
more difficult, by adding the following gadgets until all vertices have degree 3. Let
(G, M ) be an instance of degree-2,3 {K1,3 , K3 , P4 }-decomposition with
marked edges, where G has at least three degree-2 vertices t1 , t2 , t3 ; by adding
a net over {t1 , t2 , t3 }, we mean attaching a net by its leaves to v1 , v2 and v3 and
adding the edges incident to the net’s leaves to M (see Figure 3(a)).
Proposition 7. Let (G, M ) be an instance of degree-2,3 {K1,3 , K3 , P4 }-decomposition with marked edges, where G has at least three degree-2 vertices
t1 , t2 , t3 , and let (G′ , M ′ ) be the instance obtained by adding a net to (G, M ).
Then (G′ , M ′ ) has three degree-2 vertices less than (G, M ), and (G, M ) can be
decomposed iff (G′ , M ′ ) can be decomposed.
Proof. By construction, G′ has fewer degree-2 vertices, since t1 , t2 , t3 now have
degree 3 instead of 2, other vertices of G are unchanged, and new vertices
{v1 , v2 , v3 } have degree 3. We now prove the equivalence.
⇒: given a decomposition D for (G, M ), we only need to add the K1,3 induced
by {v1 , t1 , v2 , v3 } and the P4 induced by {t2 , v2 , v3 , t3 } to cover the edges
of the added net in order to obtain a decomposition D′ for (G′ , M ′ ) (see
Figure 3(b)).
⇐: we show that the only valid decompositions must include the choice we made
in the proof of the forward direction. Indeed, the marked edges cannot be
middle edges in a P4 , and the K3 induced by v1 , v2 and v3 cannot appear as a
K3 in a decomposition. Moreover, no marked edge can be the extremity of a
P4 with two edges lying in the K3 , since this would force another marked edge
to be the middle edge of a P4 . Therefore the only possible decomposition of
the net is the one defined above (up to symmetry), and we can safely remove
the P4 and the K1,3 from D′ while preserving the rest of the decomposition.
⊓
⊔
Lemma 4. degree-2,3 {K1,3 , K3 , P4 }-decomposition with marked edges
≤P {K1,3 , K3 , P4 }-decomposition with marked edges.
9
Proof. Given an instance (G, M ) of degree-2,3 {K1,3 , K3 , P4 }-decomposition with marked edges, create an instance (G′ , M ′ ) by successively adding
a net to any triple of degree-2 vertices, until no such triple remains. By Proposition 7, (G, M ) is decomposable iff (G′ , M ′ ) is decomposable. Moreover, either
G′ is cubic (hence (G′ , M ′ ) is an instance of {K1,3 , K3 , P4 }-decomposition
with marked edges), or G is trivially a no-instance by Observation 2.
⊓
⊔
Finally, we show that degree-2,3 {K1,3 , K3 , P4 }-decomposition with
marked edges is NP-complete. Our reduction relies on the cubic planar
monotone 1-in-3 satisfiability problem [9]:
cubic planar monotone 1-in-3 satisfiability
a Boolean formula φ = C1 ∧ C2 ∧ · · · ∧ Cn without negations over
a set Σ = {x1 , x2 , . . . , xm }, with exactly three distinct variables
per clause and where each literal appears in exactly three clauses;
moreover, the graph with clauses and variables as vertices and
edges joining clauses and the variables they contain is planar.
Question: does there exist an assignment of truth values f : Σ → {true,
false} such that exactly one literal is true in every clause of φ?
Input:
Theorem 1. degree-2,3 {K1,3 , K3 , P4 }-decomposition with marked edges is NP-complete.
Proof. We first show how to transform an instance φ = C1 ∧ C2 ∧ · · · ∧ Cn
of cubic planar monotone 1-in-3 satisfiability into an instance (G, M )
of degree-2,3 {K1,3 , K3 , P4 }-decomposition with marked edges. The
transformation proceeds by:
1. mapping each variable xi onto a K1,3 denoted by K(xi ) and whose edges all
belong to M ;
2. mapping each clause C = {xi , xj , xk } onto a cycle with five vertices in such
a way that K(xi ), K(xj ) and K(xk ) each have a leaf that coincides with a
vertex of the cycle and exactly two such leaves are adjacent in the cycle.
Figure 4 illustrates the construction, which yields a degree-2,3 graph. We now
show that φ is satisfiable iff (G, M ) admits a decomposition.
⇒: we apply the following rules for transforming a satisfying assignment for φ
into a decomposition D for (G, M ):
– if variable xi is set to false, then the corresponding K(xi ) is added as
such to D;
– otherwise, the three edges of K(xi ) will be the meeting points of three
different K1,3 ’s in the decomposition, one of which will have two edges
in the current clause gadget.
Two cases can be distinguished based on whether or not a leaf of K(xi ) is
adjacent to a leaf of K(xj ) or K(xk ), but in both cases the rest of the clause
gadget yields a P4 that we add as such to the decomposition (see Figure 4(b)
and (c)).
10
⇐: we now show how to convert a decomposition D for (G, M ) into a satisfying
truth assignment for φ. First, we observe that D must satisfy the following
crucial structural property:
For each clause C = (xi ∨ xj ∨ xk ), exactly two subgraphs out of
K(xi ), K(xj ) and K(xk ) appear as K1,3 ’s in D.
Indeed, G is K3 -free by construction, and:
(a) if all of them appear as K1,3 ’s in D, then the remaining five edges of the
clause gadget cannot be decomposed;
(b) if only K(xi ) appear as a K1,3 in D, then xj — without loss of generality
— must be a leaf either of a K1,3 in D with a center in the clause gadget
or of a P4 in D with two edges in the clause gadget (the P4 cannot connect
xj and xk , otherwise the rest of the gadget cannot be decomposed); in
both cases, the remaining three edges of the clause gadget must form a
P4 , thereby causing K(xk ) to appear as a K1,3 in D, a contradiction (a
similar argument allows us to handle K(xj ) and K(xk ));
(c) finally, if none of them appear as K1,3 ’s in D, then xi must be the leaf
either of a K1,3 in D with a center in the clause gadget, or of a P4
with two edges in the clause gadget; in both cases, the remaining three
edges of the clause gadget must form a P4 in D, which in turn makes it
impossible to decompose the rest of the graph.
Therefore, D yields a satisfying assignment for φ in the following simple way:
if K(xi ) appears as a K1,3 in D, set it to false, otherwise set it to true. ⊓
⊔
Theorem 2. {K1,3 , P4 }-decomposition is NP-complete.
⊓
⊔
Proof. Immediate from Lemmas 3 and 4 and Theorem 1.
xi
xj
xi
xk
(a)
xj
xi
xk
(b)
xj
xk
(c)
Fig. 4. (a) Connecting clause and variable gadgets in the proof of Theorem 1; dotted
edges belong to M . (b), (c) Converting truth assignments into decompositions in the
proof of Theorem 1; the only variable set to true is mapped onto a K1,3 in the
decomposition; (b) shows the case where the only variable set to true — namely, xi
— is such that K(xi ) has no leaf adjacent to a leaf of K(xj ) nor K(xk ); (c) shows the
other case, where xj is set to true and K(xi ) and K(xk ) have leaves made adjacent
by the clause gadget.
11
A like-minded reduction1 allows us to prove the hardness of {K1,3 , K3 , P4 }decomposition.
Theorem 3. {K1,3 , K3 , P4 }-decomposition is NP-complete, even on K3 -free
graphs.
5
Conclusions and Future Work
We provided in this paper a complete complexity landscape of {K1,3 , K3 , P4 }decomposition for cubic graphs. A natural generalisation, already studied by
other authors, is to study decompositions of k-regular graphs into connected
components with k edges for k > 3. We would like to determine whether our
positive results generalise in any way in that setting. It would also be interesting
to identify tractable classes of graphs in the cases where those decomposition
problems are hard, and to refine our characterisation of hard instances; for instance, does there exist a planarity-preserving reduction for Theorem 3? Finally,
we note that some applications relax the size constraint by allowing the use of
graphs with at most k edges in the decomposition [10]; we would like to know
how that impacts the complexity of the problems we study in this paper.
References
[1] A. Bouchet and J.-L. Fouquet, Trois types de décompositions d’un
graphe en chaı̂nes, in Combinatorial Mathematics: Proceedings of the International Colloquium on Graph Theory and Combinatorics, C. Berge,
D. Bresson, P. Camion, J. F. Maurras, and F. Sterboul, eds., vol. 75 of
North-Holland Mathematics Studies, North-Holland, 1983, pp. 131–141.
[2] A. Brandstädt, V. B. Le, and J. P. Spinrad, Graph classes: a survey,
SIAM Monographs on Discrete Mathematics and Applications, Society for
Industrial Mathematics, 1987.
[3] D. Dor and M. Tarsi, Graph decomposition is NP-complete: A complete
proof of Holyer’s conjecture, SIAM J. Comput., 26 (1997), pp. 1166–1187.
[4] M. E. Dyer and A. M. Frieze, On the complexity of partitioning graphs
into connected subgraphs, Discrete Appl. Math., 10 (1985), pp. 139–153.
[5] É. Fusy, Transversal structures on triangulations: A combinatorial study
and straight-line drawings, Discrete Math., 309 (2009), pp. 1870–1894.
[6] I. Holyer, The NP-completeness of some edge-partition problems, SIAM
J. Comput., 10 (1981), pp. 713–717.
[7] T. P. Kirkman, On a problem in combinatorics, Cambridge Dublin Mathematical Journal, 2 (1847), pp. 191–204.
[8] A. Kotzig, Z teorie konečných pravidelných grafov tretieho a štvrtého
stupňa, Časopis pro pěstovánı́ matematiky, (1957), pp. 76–92.
[9] C. Moore and J. M. Robson, Hard tiling problems with simple tiles,
Discrete Comput. Geom., 26 (2001), pp. 573–590.
1
See Appendix for details.
12
[10] X. Muñoz, Z. Li, and I. Sau, Edge-partitioning regular graphs for ring
traffic grooming with a priori placement of the ADMs, SIAM J. Discrete
Math., 25 (2011), pp. 1490–1505.
[11] T. J. Schaefer, The complexity of satisfiability problems, in Proc. 10th
STOC, San Diego, California, USA, May 1978, ACM, pp. 216–226.
[12] R. Yuster, Combinatorial and computational aspects of graph packing and
graph decomposition, Computer Science Review, 1 (2007), pp. 12–26.
13
A
Appendix: Omitted Proofs
Our hardness proof uses ideas similar to those used for {K1,3 , P4 }-decomposition, and is based on a slightly different intermediate problem. The structure is
as follows:
monotone not-all-equal 3-satisfiability
≤P K3 -free {K1,3 , P4 }-decomposition with marked edges
≤P {K1,3 , K3 , P4 }-decomposition
(Theorem 4 page 14)
(Lemma 5 page 14)
We use the following intermediate problem.
K3 -free {K1,3 , P4 }-decomposition with marked edges
Input:
a cubic, K3 -free graph G = (V, E) and a subset M ⊆ E of edges.
Question: does G admit a {K1,3 , P4 }-decomposition D such that no edge
in M is the middle edge of a P4 in D?
Lemma 5. Let (G, M ) be an instance of K3 -free {K1,3 , P4 }-decomposition
with marked edges, and G′ be the graph obtained by attaching a co-fish to
every edge in M . Then G can be decomposed iff G′ admits a {K1,3 , K3 , P4 }-decomposition.
Proof. The proof of the forward direction is exactly the same as that of the
forward direction of Lemma 3. For the reverse direction, let D′ be a {K1,3 ,
K3 , P4 }-decomposition of G′ . The only K3 ’s in G′ are those that belong to the
co-fishes we inserted, so we only need to show that removing those co-fishes
does not prevent us from adapting the decomposition of G′ in order to obtain
a {K1,3 , P4 }-decomposition D of G. The proof is similar to that of the reverse
direction of Lemma 3, with the following modification: since G is K3 -free, we
have NG (u) ∩ NG (w) = ∅, so if {u, w} is the extremal edge of a P4 in D′ , then
it will map onto {v, w} in G, where it will become the extremal edge of a P4 in
D (as opposed to, possibly, a K3 in the proof of Lemma 3).
⊓
⊔
We give a reduction from the following NP-complete variant of sat [11]:
monotone not-all-equal 3-satisfiability
a Boolean formula φ = C1 ∧ C2 ∧ · · · ∧ Cn without negations over
a set Σ = {x1 , x2 , . . . , xm }, with exactly three distinct variables
per clause.
Question: does there exist an assignment of truth values f : Σ → {true,
false} such that exactly one or two literals are true in every
clause of φ?
Input:
Theorem 4. K3 -free {K1,3 , P4 }-decomposition with marked edges is
NP-complete.
14
Proof. Given an instance φ of monotone not-all-equal 3-satisfiability,
we build an instance (G, M ) of K3 -free {K1,3 , P4 }-decomposition with
marked edges as follows:
1. For each variable xi with k occurrences (we can assume k ≥ 2), we create a
tree T (xi ) with 3k leaves, called a variable tree, whose edges are all marked
(see Figure 5(a)). Edges incident to a leaf are called border edges, the others
are called internal edges.
2. For each clause xi ∨ xj ∨ xk , we create a P7 (v1 , v2 , . . . , v7 ), called a clause
path, with marked edges {v2 , v3 } and {v5 , v6 }, to which we join three border
edges of each variable tree as follows (see Figure 6):
• one leaf of T (x1 ) is joined to v1 , another to v3 , and another to v7 ;
• one leaf of T (x2 ) is joined to v1 , another to v5 , and another to v7 ;
• one leaf of T (x2 ) is joined to v2 , another to v4 , and another to v6 .
Note that the resulting graph is indeed cubic: each inner vertex of a variable
tree has degree 3, and each leaf is also part of a clause path. Furthermore, the
inner vertices of a clause path are adjacent to two other vertices in the path and
one other vertex in a variable tree, and each endpoint of a clause path is adjacent
to another vertex in the path and two vertices in different variable trees.
We further observe that G is K3 -free. First, any cycle included only in variable
trees has length at least 4 (since it must be included in at least 2 such tree, and
in each tree the path between any pair of leaves has length at least 2). Therefore,
a K3 would have to use 1 or 2 edges in a clause path. If it uses only one edge
{vi , vi+1 }, then both vi and vi+1 must be joined to leaves of the same variable
tree, which is impossible since each clause consists of three different variables. If
two edges are used, then the last edge of the K3 would be joining two leaves in
some variable tree, which is also impossible. Therefore, G is K3 -free.
We now prove that φ is satisfiable iff (G, M ) admits a decomposition.
⇒: From a satisfying truth assignment, we create a {K1,3 , P4 }-decomposition
of (G, M ) as described in Figures 5(b − c) and 6(b − e). Specifically, for each
variable xi with k occurrences, if xi = true, then we cover all edges of T (xi )
with 2k − 1 K1,3 ’s (Figure 5(b)). If xi = false, then we cover all internal
edges of T (xi ) with k − 1 K1,3 ’s(Figure 5(c)).
Now for any clause xi ∨ xj ∨ xk , at least 1 and at most 2 variables among
{xi , xj , xk } are set to false. For these two variables and the corresponding
variable trees, the border edges are still uncovered (for variables assigned
true, border edges are covered with the rest of the variable tree). Each tree
has 3 border edges coming to the clause path, so there are either 9 or 12
edges to cover (6 in the clause path and 3 or 6 in border variable trees). As
shown in Figures 5(b − e), there always exist a decomposition of these edges
into K1,3 ’s and P4 ’s (with the constraint that no marked edge is the middle
of a P4 ).
Overall, all edges in all variable trees and clause paths are covered by a K1,3
or a P4 , therefore (G, M ) admits a {K1,3 , P4 }-decomposition.
15
A
...
B
leaves
(a)
...
(b)
...
(c)
Fig. 5. (a) A variable tree T (xi ) for a variable xi with k occurrences: all its edges are
marked, and it has 3k leaves. Internal vertices are partitioned into two sets A and B.
(b) A decomposition of T (xi ), corresponding to xi = true. (c) A decomposition of the
internal edges of T (xi ), corresponding to xi = false.
⇐: We first consider any variable tree, and show that the decompositions used
for true and false assignments above are in fact the only two possible
decompositions of the internal edges of this tree. Indeed, consider internal
vertices and their partition into A and B (see Figure 5(a)): because of marked
edges, all edges between any two internal vertices must be part of a K1,3 ,
linking a leaf to a center. Since internal vertices form a path, they must
alternate along this path between leaves and centers of K1,3 ’s, so either all
vertices of B are centers, or all vertices of A are centers. Each case yields
only one possible decomposition of adjacent edges, as described respectively
in Figures 5(b) and 5(c). Naturally, we assign true to any variable whose
tree is decomposed as in the first case, and false to other variables. We
further make the following observation for the false case: consider any leaf
x of the tree, and its parent y. Due to marked edges, x can only be the center
of a K1,3 , or a middle node of a P4 , of which y is an endpoint.
We now consider a clause xi ∨xj ∨xk , and show that its variables can neither
be all set to true nor all set to false. Aiming at a contradiction, assume
first that all variables are set to true. Then all border edges of their trees
are already covered, and the K1,3 ’s and P4 ’s covering the clause path may
only use the 6 edges of the path. The only possibility to decompose the path
in such a way is to use two P4 ’s, however, such P4 ’s would have marked
16
T (xi )
v1
T (xj )
v3
v5
v2
v4
v7
v6
T (xk )
(a)
T (xi )
v1
T (xj )
v3
v2
T (xi )
v5
v4
v7
v1
v6
T (xj )
v3
v2
v5
v4
v7
v6
T (xk )
T (xk )
(b) xi = true, xj = true, xk = false
(c) xi = true, xj = false, xk = true
T (xi )
v1
T (xj )
v3
v2
T (xi )
v5
v4
v7
v6
v1
T (xj )
v3
v2
v5
v4
v7
v6
T (xk )
T (xk )
(d) xi = false, xj = true, xk = false
(e) xi = false, xj = false, xk = true
Fig. 6. (a) A clause path and its connections to the variable trees. (b − e) For each
truth assignment of xi , xj , xk (up to symmetry), a decomposition of the path edges
and the neighboring uncovered edges of variable trees for false variables.
17
edges as middle edges, which is forbidden. We now assume that all variables
of a clause are set to false, i.e. it remains to cover the clause path and all
border edges of the variable trees. Thanks to the observation made on leaves
of the tree in the false case, vertices v1 and v2 are either centers of K1,3 ’s,
or middle nodes in P4 ’s. They cannot both be centers of K1,3 ’s, and due to
marked edges, they must both be middle nodes of the same P4 . However,
as noted above, the parents of v1 in both trees T (xi ) and T (xj ) should be
endpoints of this P4 , which is impossible.
Finally, each variable has been assigned a truth value, and for each clause
there must be at least one true and one false variable: therefore, we have
a satisfying assignment for our instance of monotone not-all-equal 3satisfiability.
⊓
⊔
18
| 8cs.DS
|
1
Fuel-Efficient En Route Formation
of Truck Platoons
arXiv:1704.08836v1 [cs.SY] 28 Apr 2017
Sebastian van de Hoef, Karl H. Johansson, Fellow, IEEE and Dimos V. Dimarogonas, Member, IEEE
Abstract—The problem of how to coordinate a large fleet of
trucks with given itinerary to enable fuel-efficient platooning is
considered. Platooning is a promising technology that enables
trucks to save significant amounts of fuel by driving close together
and thus reducing air drag. A setting is considered in which each
truck in a fleet is provided with a start location, a destination, a
departure time, and an arrival deadline from a higher planning
level. Fuel-efficient plans should be computed. The plans consist
of routes and speed profiles that allow trucks to arrive by
their arrival deadlines. Hereby, trucks can meet on common
parts of their routes and form platoons, resulting in decreased
fuel consumption. We formulate a combinatorial optimization
problem that combines plans involving only two vehicles. We
show that this problem is hard to solve for large problem
instances. Hence a heuristic algorithm is proposed. The resulting
plans are further optimized using convex optimization techniques.
The method is evaluated with Monte Carlo simulations in a
realistic setting. We demonstrate that the proposed algorithm
can compute plans for thousands of trucks and that significant
fuel savings can be achieved.
I. I NTRODUCTION
LATOONING is foreseen to become a common element
in intelligent transportation systems. The term refers to a
group of vehicles forming a road train without any physical
coupling between them. A short inter-vehicle distance is
maintained by automatic control and vehicle-to-vehicle communication. Platooning has received a lot of attention due to
its potential to increase road throughput by reducing the intervehicle gaps. It can also help facilitate the (semi-)automatic
operation of vehicles [1]. This paper focuses on the potential of platooning in reducing fuel consumption. Similar to
what racing cyclists exploit, the follower vehicles, and to a
lesser degree the lead vehicle, experience a reduction in air
drag, which translates into reduced fuel consumption [2]–[4].
Advances in wireless communication, satellite-based position,
and advanced driver support systems have made the wide
deployment of platooning systems feasible and have attracted
the attention of major truck manufacturers. Increased fuel costs
and environmental awareness make the implementation of such
systems likely in the near future.
Using platooning to reduce fuel consumption for a large
number of trucks leads to a challenging coordination problem.
Consider two trucks that travel between the same two regions
but from different locations and at approximately the same
time. Then the trucks can adjust their speeds slightly at the
beginning of their journeys, form a platoon at the start of the
P
ACCESS Linnaeus Center and the School of Electrical Engineering, KTH
Royal Institute of Technology, SE-100 44, Stockholm, Sweden (e-mail:
{shvdh,kallej,dimos}@kth.se).
common part of their route and thus save fuel during part
of their trips. This approach might involve one of the trucks
having to drive slightly faster before the two merge, which
increases air-drag and consequently fuel consumption during
the initial phase. One truck might instead slow down to let
the other truck catch up but then travel at an increased speed
later on to arrive at its destination on time. If many trucks are
involved, it is not straightforward to compute an optimal plan
for all trucks.
The main contribution of the paper is to derive an efficient
and scalable method to coordinate platooning of a large number of trucks in a fuel-efficient way explicitly considering the
effect of speed and platooning on the fuel consumption. The
core novelty is that the computation of platoon plans is done
in three computationally and mathematically tractable stages.
The first stage involves the computation of platoon plans only
taking into account two vehicles at a time. The second stage
selects one such plan for each vehicle. Since the problem to
be solved in the second stage is shown to be NP-hard, an
iterative algorithm to compute heuristic solutions is proposed.
In the third stage, the resulting plans are optimized further
using convex optimization techniques. The potential of the
method is demonstrated using Monte Carlo simulations. The
simulations demonstrate that the method is able to significantly
reduce the fuel consumption of a fleet of vehicles as well as
to handle a realistic number of transport assignments for the
example of Sweden.
The efficient operation of transport systems is a widely studied field due to its large economic and environmental impact.
Planning conducted by transportation operators ranges all the
way from strategic over tactical to operational planning [5].
The planning on the latter stage typically happens on the
level of departure and arrival times which is the input to
the problem considered in this paper. Research on eco-routing
aims to reduce fuel consumption by appropriate choice of route
and travel speeds for individual vehicles [6]. Furthermore,
operators of road infrastructure use variable speed limits, ramp
metering, variable route recommendations, etc. to improve
safety and efficiency of the road transportation system [7].
Various aspects of platoon coordination have been considered in the literature. In [8] the authors formulate a mixed
integer linear programming problem, without considering the
speed dependency of fuel consumption, and prove that the
problem is NP-hard. In [9] the authors consider a simple
catch-up coordination scheme and evaluate it on real fleet
data. In [10] local controllers for coordinating the formation of
platoons are proposed. In [11] the authors use data-mining to
identify economic platoons based on various criteria. Unlike
2
this paper the method presented in [11] allows that trucks wait
for other trucks to form the platoon. Preliminary material used
in this paper has been presented in [12], [13].
One of the key elements to make the problem tractable for
realistic numbers of trucks, is to select a subset of vehicles
called coordination leaders to which other vehicles adapt. The
way we select this subset of vehicles is inspired by a clustering
algorithm called partitioning around medoids [14]. Clustering
is present in a variety of different contexts. A large body of
research focuses on clustering methods as an analysis tool to
structure, understand, and classify large data sets [15], [16].
Examples include the clustering of graphs in the scope of
community detection [17]–[19]. Closely related to our problem
of choosing coordination leaders is leader election, where a
group of agents has to jointly determine a leader [20], [21].
The approach developed in this paper can be seen as local
leader election, where pairwise fuel savings are interpreted as
preferences for trucks being coordination leaders.
The paper is organized as follows. We start in Section II
by formulating the coordination problem and introduce the
structure of the proposed solution in Section III. In Section IV
the problem is broken down to a purely combinatorial problem
of selecting coordination leaders and the problem is shown
to be NP-hard. In Section V an iterative algorithm is developed to find heuristic solutions to the combinatorial problem.
Section VI discusses how to jointly optimize the plans when
the trucks are constrained to the platoons as proposed by
the algorithm from Section V. In Section VII the method
is demonstrated by Monte Carlo simulations in a realistic
scenario with up to 5 000 trucks on the Swedish road network.
II. P ROBLEM F ORMULATION
In this section we formulate the problem and introduce the
notation. We have an index set Nc of finitely many transport
assignments, each tied to a specific truck. A transport assignment A = (P S , P D , tS , tD ) consists of a start position P S ,
a destination P D , a start time tS , and an arrival deadline tD .
We model the road network as a directed graph Gr = (Nr , Er )
with nodes Nr and edges Er . Nodes correspond to intersections
or endpoints in the road network and edges correspond to
road segments connecting these intersections. The function
L : Er → R+ maps each edge in Er to the length of the
corresponding road segment. A vehicle position is a pair
(e, x) ∈ Er × [0, L(e)] where e indicates the current road
segment and x how far the vehicle has traveled along that
segment.
The goal is to compute fuel-efficient plans for the trucks
that ensure arrival before each trucks’ individual deadline.
Each plan includes a route in the road network from start
to destination and encodes a piecewise constant speed trajectory. The speed is constrained to a range of feasible speeds
[vmin , vmax ], which is supposed to be the same for all vehicles
and road segments.1 For the sake of this high-level planning,
it is reasonable to assume that trucks change their speed
instantaneously.
1 The approach developed in this paper can be generalized in order to relax
this assumption.
Definition 1 (Vehicle Plan). A vehicle plan P = (e, v, t̂)
consists of a route e, a speed sequence v, and a time sequence
t̂. The route is a sequence of Ne edges in the road network:
e = (e[1], . . . , e[Ne ]), e[i] ∈ Er . The speed sequence is a
sequence of Nv speeds v = (v[1], . . . , v[Nv ]), where speeds
are within the feasible speed range 0 < vmin ≤ v[i] ≤ vmax .
The time sequence t̂ = (t̂[1], . . . , t̂[Nv + 1]) defines when the
speed changes. Speed v[i] is selected from t̂[i] until t̂[i + 1].
Note that changes in speed can happen, in principle, everywhere and not only at the beginning of route segments. The
symbols Ne and Nv are introduced for notational convenience
and their value may be different for different vehicle plans.
We want to compute a vehicle plan for each truck. A
valid vehicle plan brings the truck from its start position
P S = (eS , xS ), where it is at time tS , to its destination
P D = (eD , xD ) before its deadline tD .
Vehicle plans are constrained by two conditions. The first
condition requires the trip to start at the start time t̂[1] = tS
and ends before the deadline t̂[Nv +1] = tA ≤ tD . The second
condition ensures that the truck arrives at its destination when
the trip ends, i.e., the distance traveled is
D :=
NX
e −1
i=1
L(e[i]) + xD − xS =
Nv
X
i=1
v[i](t̂[i + 1] − t̂[i]).
We introduce the notion of trajectories as functions of continuous time. A vehicle trajectory consists of an edge trajectory
and a linear position trajectory ξ. The edge trajectory for
t ∈ [tS , tA ) is given by (t) = e[j] where j depends on t
j−1
P
L(e[i]) − xS <
and is the largest integer that satisfies
i=1
Rt
φ(τ )dτ , and where the speed trajectory φ(t) = v[i] for
tS
t ∈ [t̂[i], t̂[i + 1]), i ∈ {1, . . . , Nv }. The linear position,
i.e., the second element of the position, at time t is given
j−1
Rt
P
by ξ(t) = φ(τ )dτ −
L(e[i]) + xS .
tS
i=1
When trucks platoon, their positions coincide in our model
and hence we neglect the physical dimension of the trucks.
Each platoon consists of a platoon leader and a number
of platoon followers. We introduce the platoon trajectory
πn : [tSn , tA
n ) → {0, 1} for truck n ∈ Nc . A platoon trajectory
equals 1 when truck n is a platoon follower and 0 when
it is a platoon leader or traveling alone. Thus πn (t) = 1
implies that there is another truck m ∈ Nc with m 6= n and
(n (t), ξn (t)) = (m (t), ξm (t)).
We model the fuel consumption per distance traveled as a
function of the speed and of whether the truck is a platoon
follower or not. A platoon leader is assumed to have the
same fuel consumption as a truck that travels alone while
a platoon follower has a reduced fuel consumption. Hereby,
we neglect the relatively small reduction in fuel consumption
when a truck is a platoon leader compared to traveling by
itself. We denote the fuel consumption per distance traveled
as f : [vmin , vmax ] × {0, 1} → R+ where
f0 (v) if p = 0
f (v, p) =
(1)
fp (v) if p = 1.
3
The function f0 models the fuel consumption when the truck
is a platoon leader or when it travels solo, and fp the fuel
consumption when the truck is a platoon follower. These
functions can either be derived from an analytical model or
fitted to data [22]. We purposely omit that fuel consumption
depends on road and vehicle parameters in order to keep
the presentation concise. All the presented results can be
augmented to handle those additional parameters.
The problem that we want to solve is to find a vehicle
plan for each vehicle, and we want to minimize the combined
fuel consumption of these plans. The total fuel consumption
F (φn , πn ) associated with vehicle n’s plan is given by integrating the fuel consumption according to (1) over the duration
of the trip
Platoon Coordinator
Assignments
D
An = (PnS , PnD , tS
n , tn )
n=1
n=2
Vehicle Plans
Pn = (en , vn , t̂n )
...
Fig. 1. Schematic of the platoon coordinator. Trucks provide their assignment
data and the platoon coordinator computes fuel efficient vehicle plans.
A
Ztn
F (φn , πn ) =
f (φn (t), πn (t))φn (t)dt,
(2)
tS
n
where φn is the speed trajectory, πn the platoon trajectory, tSn
the start time, and tA
n the arrival time of truck n. The combined
fuel consumption Fc is given by
X
Fc =
F (φn , πn ).
(3)
n∈Nc
Our primary goal is to compute vehicle plans that minimize
Fc .
III. P LATOON C OORDINATOR
Consider the centralized platoon coordinator in Fig. 1.
Trucks connect to the coordinator via vehicle-to-infrastructure
communication and share their assignment data. The coordinator then computes fuel-efficient vehicle plans for the trucks.
These plans are sent to the trucks and executed. This process
is repeated whenever there is updated information, such as
deviations from the plans and new assignments. The current
vehicle position is then the new start position of an assignment
that is already being executed.
The computation of the vehicle plans happens in four stages:
1) Computation of the routes en , n ∈ Nc : routes are
calculated using an algorithm for route calculation in
road networks.
2) Computation of pairwise vehicle plans: many plans
involving two vehicles are computed. The fuel savings
of these plans are recorded as the coordination graph Gc
introduced in the following.
3) Selection of pairwise plans: a consistent subset of the
plans computed in the previous stage is combined by
selecting a subset Nl ⊂ Nc , so-called called coordination leaders.
4) Joint speed-profile optimization: the selected pairwise
plans are jointly optimized for lower fuel consumption
while platoon partners as well as merge and split locations are kept as computed in step 3).
Stage 1) computes the routes en , n ∈ Nc and stages 2)-4)
compute the speed sequences vn and time sequences t̂n for
n ∈ Nc making use of the ability of the trucks to form platoons
in order to achieve low fuel consumption. Algorithms for route
calculation in road networks are well developed [6], [23] and
not further discussed in this paper. We discuss stages 2)-4) in
the following sections.
IV. S ELECTING PAIRWISE V EHICLE P LANS
In this section, we formulate a combinatorial optimization
problem corresponding to the second and third computation
stage introduced in Section III. The problem is proven to be
NP-hard which motivates the heuristic algorithm developed in
Section V.
To begin with, we need to be able to compute what we
call a default plan. This is a valid vehicle plan according
to Definition 1 with either the lowest possible or most fuel
optimal constant speed.
Definition 2 (Default Plan). The default plan is a vehicle
plan P = (e, v, t̂) with speed sequence v = (vcd ) and
time sequence t̂ = (tS , D/vcd ). The most fuel optimal speed
without platooning vcd is computed as
vcd =
argmin
f0 (v),
v∈(vcm ,vmax ]
where vcm is the lowest constant speed to arrive before the
deadline:
D
vcm = max vmin , D
.
t − tS
An adapted plan, as introduced next, is such that the speed
sequence vn and time sequence t̂n of a follower truck n is
adapted in a way that allows the follower to platoon during
part of its journey with a leader m. The leader sticks to its
default plan, which is important in order to be able to compose
these plans. The plan is computed in a way that minimizes the
fuel consumption of n.
Definition 3 (Adapted Plan). An adapted plan is a vehicle
plan Pn = (en , vn , t̂n ) adapted to vehicle plan Pm =
(em , vm , t̂m ), such that (n (t), ξn (t)) = (m (t), ξm (t)) for
t ∈ [t̂n [2], t̂n [Nv,n ]).
We denote the merge time as tM = t̂n [2] and the split time
as tSp = t̂n [Nv,n ]. Truck n becomes the platoon follower of
truck m at time tM , stays platoon follower until tSp , when
the two trucks separate. This sequence of events occurs only
4
(n (tM ), ξn (tM ))
(n (tSp ), ξn (tSp ))
(n (tS ), ξn (tS ))
(n (tA ), ξn (tA ))
Fig. 2. Overview of the relevant time instances of the adapted plan. The solid
line illustrates the route of the adapted plan n and the dashed line the one of
the plan that it is adapted to and has index m. The parallel sections of the
line indicate that the trucks share the route and the section where the lines
are on top of each other indicates that the trucks platoon there.
once. Fig. 2 illustrates the adapted plan. We denote the speed
trajectory φ corresponding to the speed sequence v and the
time sequence t̂ of the adapted vehicle plan of truck n adapted
to truck m as φn,m .
The fuel consumption of truck n with its plan adapted to
truck m is modeled as in (2). We denote the platoon trajectory
of the adapted plan πn,m (t). We have that πn,m (t) = 1 for
t ∈ [tM , tSp ) and πn,m (t) = 0 for t ∈ [tS , tM ) ∪ [tSp , tA ). The
fuel consumption of m is not altered by the fact that n and m
platoon, since m’s speed trajectory does not change and since
m takes the role of a platoon leader. The reduction in fuel
consumption that results from n implementing the adapted
plan and not its default plan is ∆F (n, m) = F (φn , πn ) −
F (φn,m , πn,m ) where πn ≡ 0, which is positive if n adapting
to m saves fuel. If no plan that is adapted to m exists for n,
we define ∆F (n, m) = 0. There might exist no adapted plan
because the routes do not overlap or because the constraint on
the maximum speed in conjunction with the arrival deadline
makes it impossible for the trucks to form a platoon.
We now compute ∆F for all 2-permutations in Nc . We are
only interested in adapted plans that save fuel, i.e., for which
∆F is positive. We can conveniently collect this information
in a weighted graph that we call the coordination graph.
Definition 4 (Coordination Graph). The coordination graph
is a weighted directed graph Gc = (Nc , Ec , ∆F ). Recall that
the elements of Nc represent the trucks. Ec ⊆ Nc × Nc is a
set of edges, and ∆F : Ec → R+ are edge weights, such that
there is an edge (n, m) ∈ Ec , if the adapted plan of n to m
saves fuel compared to i’s default plan, i.e., Ec = {(i, j) ∈
Nc × Nc : ∆F (i, j) > 0, i 6= j}.
Furthermore, we introduce the set of in-neighbors of a node
n ∈ Nc as Nni = {i ∈ Nc : (i, n) ∈ Ec } and the set of outneighbors n as Nno = {i ∈ Nc : (n, i) ∈ Ec }. We define the
the maximum over an empty set to be zero, i.e., max(·) = 0.
i∈∅
With these definitions, we are ready to formulate the problem of finding a fuel optimal set of coordination leaders Nl .
Problem 1. Given as input a coordination graph Gc =
(Nc , Ec , ∆F ) find a subset Nl ⊂ Nc of nodes that maximizes
fce (Nl ) =
X
i∈Nc \Nl
max ∆F (i, j).
j∈Nio ∩Nl
(4)
The coordination leaders select their default plans. The
remaining assignments, called coordination followers, select
Input: Gc
Output: Nl
Nl ← ∅
while {n̄ ∈ Nc : ∆u(n̄, Nl ) > 0} =
6 ∅ do
Select n ∈ {n̄ ∈ Nc : ∆u(n̄, Nl ) > 0}
if n ∈ Nl then
Nl ← Nl \ {n}
else
Nl ← Nl ∪ {n}
end if
end while
Fig. 3. The clustering algorithm, which is an iterative algorithm to compute
the set of coordination leaders Nl .
their plans adapted to the coordination leader that yields the
largest fuel savings ∆F (n, m). Since the selection of adapted
plans does not alter the speed trajectories of the coordination
leaders, several coordination followers can select the same
coordination leader without affecting the fuel savings that
result from this adaptation, potentially resulting in platoons
of more than two vehicles. The objective function fce (Nl )
equals the sum of all these fuel savings. If (n, m) ∈ Ec
with n ∈ Nc \ Nl and m = arg max
∆F (n, m), we
o
m∈Nn ∩Nl
say that n is the coordination follower of m and m is the
coordination leader of n. If m has no out-neighbor in Nl ,
then maxm∈(Nno ∩Nl ) ∆F (n, m) = maxm∈∅ ∆F (n, m) = 0.
Problem 1 is a combinatorial optimization problem. We
can compute an optimal solution in finite time by using
exhaustive search or a branch and bound technique. However,
the computational complexity of such an exact computation
might be too high. In fact, we show in the following that
this problem is an NP-hard problem. This a strong indicator
that searching for an algorithm that computes solutions for
every coordination graph and scales well in the size of the
coordination graph is futile [24].
Proposition 1. Problem 1 is NP-hard, where the size of the
input Gc is measured as |Nc | + |Ec |.
The proof can be found in the appendix.
One disadvantage of the approach presented in this section
is that each truck can only join one platoon. This can however
be somewhat mitigated by frequent re-planning. For instance,
at some later point in time, it might turn out more beneficial
for a truck to leave its current platoon and join another one.
V. I TERATIVE S ELECTION OF C OORDINATION L EADERS
In this section we present an algorithm that computes
heuristic solutions to Problem 1. Motivated by the result that
Problem 1 is NP-hard, we apply an iterative strategy that
converges to a local maximum.
Consider the algorithm in Fig. 3, which we call the clustering algorithm. The input is a coordination graph Gc and the
output is a set of coordination leaders Nl . Initially Nl is an
empty set. In each iteration, a node n ∈ Nc is selected for
which the objective function fce is increased if it is added to
Nl or removed from Nl , and Nl is updated accordingly. The
5
difference in fce when adding or removing a node in Nc to or
from the set of coordination leaders Nl is given by a function
∆u. The algorithm iterates until no further increase of fce is
possible.
The function ∆u that measures how much is gained from
switching whether n belongs to Nl is defined as follows:
fce (Nl \ {n}) − fce (Nl )
if n ∈ Nl
∆u(n, Nl ) =
fce (Nl ∪ {n}) − fce (Nl ) otherwise .
(5)
If n ∈
/ Nl , we get
fce (Nl ∪ {n}) − fce (Nl ) =
X
∆F (i, j) −
max
o
i∈Nni \Nl
−
j∈Ni ∩(Nl ∪{n})
max
∆F (i, j)
o
j∈Ni ∩Nl
max ∆F (n, i).
i∈Nno ∩Nl
The sum over i covers nodes that can select n as their new
coordination leader. The last summand accounts for n possibly
not being a coordination follower any longer.
If n ∈ Nl , we get
fce (Nl \ {n}) − fce (Nl ) =
X
max
∆F (i, j) −
o
i∈Nni \Nl
+
j∈Ni ∩(Nl \{n})
max
i∈Nno ∩(Nl \{n})
max
∆F
(i,
j)
o
j∈Ni ∩Nl
∆F (n, i).
The sum over i covers nodes that can have n as their coordination leader before the change. The last summand accounts
for n possibly becoming a coordination follower.
In this paper, we consider two methods to select n from the
set {n̄ ∈ Nc : ∆u(n̄, Nl ) > 0}. The first method is to select n
in a greedy manner according to n = arg max ∆u(n̄, Nl ). The
n̄∈Nc
second method is to choose n randomly with equal probability
from the set {n̄ ∈ Nc : ∆u(n̄, Nl ) > 0}.
The clustering algorithm is guaranteed to converge in finite
time. This is due to the number of possible subsets of Nc being
finite and thus the number of possible assignments of Nl is
finite. In every iteration fce (Nl ) strictly increases which means
that Nl changes in every iteration and the same assignment
for Nl never reoccurs. So in the worst case the clustering
algorithm iterates over all subsets of Nc before termination.
The clustering algorithm can be efficient. Note for instance
that the function ∆u can be computed based on the sub-graph
induced by the one- and two-hop neighbors of n only. This
means that the average complexity of computing ∆u is a
function of the average node degree and not of the number
of nodes in the coordination graph. Furthermore, if a node
is added to or removed from Nl , then only the ∆u for the
two-hop neighbors needs to be recomputed.
Simulations suggest that selecting n in a greedy or a random
manner makes little difference for the quality of the computed
solution. However, greedy node selection tends to lead to less
iterations of the algorithm and is thus better suited for a serial
implementation. Random node selection might be preferable
for a parallel implementation due to the reduced need for
synchronization.
Having computed the set of coordination leaders, there is
immediately a vehicle plan for each truck. These plans are
jointly optimized as discussed in the following section.
VI. J OINT V EHICLE P LAN O PTIMIZATION
In this section, we derive how to jointly optimize the
vehicle plans that are selected by the clustering algorithm.
We do this by formulating a convex optimization problem
with linear constraints for a group consisting of a coordination
leader and its coordination followers. Hereby, the timing when
platoons are assembled and broken apart is adjusted while the
locations where this happens is not changed. Trucks that are
not matched to any coordination leader or are not coordination
leaders themselves just follow their default plans and are not
considered in this section.
Consider a coordination leader nl ∈ Nl and its followers
Nfl,nl = {n ∈ Nc \ Nl : nl = arg max o ∆F (n, i)}. This
i∈Nl ∩Nn
group of agents is denoted Ng = {nl } ∪ Nfl,nl . We construct
an ordered set of time instances t = (t[1], t[2], . . . ). This
set contains the start time and the arrival deadline of the
coordination leader, and the merge times and the split times
of its followers. We divide the distance traveled by the leader
from start to destination according to these time instances and
get the distances Wnl [i] = (t[i + 1] − t[i])/vcd between these
points, where vcd is the speed of the leader according to its
default plan. These are the distances between the points where
coordination followers join or leave the platoon. Similarly,
for a coordination follower n ∈ Nfl,nl , we have Wn =
Sp
A
Sp
M
S
(vn [1](tM
n −tn ), Wnl [in ], . . . , Wnl [in ], vn [Nv,n ](tn −tn )).
S M Sp A
The variables tn , tn , tn , tn denote the start time, merge time,
split time, and arrival time of follower n according to its
adapted plan. The first element of Wn is the distance along
the route from start to the merge point. For the part of the
route the follower platoons with the coordination leader, the
entries are the same as for the coordination leader. The indices
Sp
iM
n , in are defined accordingly. The last element of Wn is the
distance from the split point to the destination of the follower.
Fig. 4 illustrates the definition of Wn .
We introduce sequences pn = (pn [1], . . . , pn [|Wn |]) that
indicate on which segments of the journey the coordination
follower is a platoon follower. If truck n is a platoon follower
on the segment that corresponds to Wn [i] for some i, then
pn [i] = 1. Otherwise we have pn [i] = 0. For the coordination
leader nl , we have pnl = (0, . . . , 0) and for a coordination
follower n ∈ Nfl,nl , we have that pn = (0, 1, 1, . . . , 1, 0).
We express the speed and time sequence of truck n ∈
{nl } ∪ Nfl,nl as traversal times Tn = (Tn [1], . . . , Tn [|Wn |])
of the segments Wn . The speed on each such segment remains
constant and can be computed as vn [i] = Wn [i]/Tn [i].
The traversal times of the segments in all trucks’ routes
are the optimization variables. Working with traversal times
rather than the sequence of speeds v allows us to state the
optimization problem with linear constraints. The times when
i−1
P
the speed changes t̂n , are computed as t̂n [i] = tSn +
Tn [j]
j=1
for i = 1, . . . , Nv,n + 1.
With these definitions, we are ready to state the following
optimization problem:
6
vn [1](tM − tS
n)
Wnl [1] Wnl [2] Wnl [3]
Wnl [5]
Wnl [4]
Wnl [8]
Wnl [6] Wnl [7]
Fig. 4. Illustration of how the sequences Wn are defined. The red, dotted
line represents the route of the coordination leader and the black, solid lines
with arrows represent the routes of the coordination followers. The thin lines
indicate the distances that the elements of Wnl correspond to.
Problem 2.
min
{Tn :n∈Ng }
v,n
X NX
Wn [i]
f
, pn [i] Wn [i]
Tn [i]
i=1
(6a)
n∈Ng
s.t.
for n ∈ Ng :
Wn [i]
≤ Tn [i], i ∈ {1, . . . , Nv,n }
vmax
Wn [i]
≥ Tn [i], i ∈ {1, . . . , Nv,n }
vmin
Nv,n
X
tSn +
Tn [i] ≤ tD
n
(6b)
(6c)
(6d)
i=1
and for n ∈ Nfl,nl :
tSn
+ Tn [1] =
tSnl
+
When f0 , fp are such that f0 (T −1 ) and fp (T −1 ) are convex
in T for T > 0, then the objective (6a) is a sum of convex
functions and hence convex. For instance, polynomials with
arbitrary constant part and non-negative coefficients fulfill
this requirement. Furthermore, all constraints are linear. Thus,
Problem 2 is a convex optimization problem for which well
developed numerical solvers are readily available [25], [26].
The optimization is initialized with the pairwise plans.
iM
n −1
X
Tnl [i]
(6e)
i=1
Sp
M
Tn [1 + i] = Tnl [iM
n + i − 1], i ∈ {1, . . . , in − in + 1}.
(6f)
The objective
P function (6a) equals the combined fuel consumption
F (φn , πn ) for the assignments Ng which is
n∈Ng
Fig. 5. Population density map from which the start and goal locations are
sampled. The brighter the pixel, the larger the population density in that area.
Areas not belonging to mainland Sweden are shown in blue.
part of the sum that defines the combined fuel consumption
of all assignment Fc defined in (3). It is composed of the fuel
consumption of the coordination leader and the coordination
followers. The coordination leader is considered to travel alone
or take the role as the platoon leader throughout its journey.
The coordination followers travel alone on the first and the
last segment of their journey. They become platoon followers
in-between these segments.
There are two sets of constraints. The first set applies to
all trucks and ensures that the sequences Tn correspond to
valid vehicle plans. In particular, the constraints (6b) and (6c)
express that the trajectories stay within the allowed range
of speed. The constraints (6d) express that all trucks arrive
before their deadline. The second set of constraints ensures
that platooning happens as specified in the original pairwise
plans. The constraints (6e) ensure that the coordination leader
and each of its followers arrive at the same time at their
respective merge point. The constraints (6f) ensure that the
speed of the leader and the speed of the follower are the same
when they are supposed to platoon.
VII. S IMULATIONS
In this section, we evaluate the coordination method outlined in the previous sections with Monte Carlo simulations.
We show that coordination of truck platooning can lead to
significant reductions in fuel consumption compared to the
current situation where trucks do not platoon, as well as
compared to spontaneous platooning where trucks only form
platoons if they happen to be in the vicinity of another.
We generate transport assignments randomly. The start and
goal locations are sampled within mainland Sweden. The
probability of an assignment starting or ending at a particular
location is proportional to the population density [27], see
Fig. 5. The resolution is 0.1 degrees in longitude and latitude
and the road network node that is closest to the sampled
coordinate is chosen. We calculate the routes with the Open
Source Routing Machine [28]. Assignments for which no
route can be found are disregarded. If the route is longer
than 400 kilometers, a 400 kilometers long subsection of the
route is randomly selected. This is to take into account that
merge points too far from the current position should not be
considered for coordination since the uncertainty becomes too
large due to traffic, new assignments, and rest periods of the
driver. Start locations along the route are considered since we
believe that platoon coordination systems will frequently replan for assignments that are already en route and suspended
for the driver to rest.
7
any kind of coordination effort which would be present for
time gaps up to one minute.
In order to assess the the quality of the solution computed
by the clustering algorithm, we establish an upper bound on
the solution of Problem 1. This upper bound is based on
the intuition to assign every truck its best coordination leader
and ignore that coordination leaders do not contribute to the
objective. We have that
X
fce (Nl ) =
max
∆F (i, j)
o
i∈Nc \Nl
≤
≤
Fig. 6. The routes of a platoon coordinator with four coordination followers.
The route of the coordination leader in shown in black, the routes of the
coordination followers are dashed. The beginning of a route is marked with
a star. The merge point of a follower is indicated with an upwards-facing
triangle and the split point with a downwards-facing triangle.
The fuel model is an affine approximation around 80 km/h
of the analytical fuel model in [29]. We have for the fuel per
distance traveled in kilograms diesel per meter
f0 (v) = 8.4159 · 10−6 v + 4.8021 · 10−5
fp (v) = 5.0495 · 10−6 v + 8.5426 · 10−5 .
According to this model, the relative reduction in fuel consumption of a platoon follower is 15.9 percent at a speed of
80 km/h.
We consider a default speed of 80 km/h and we assume that
the speed can be freely chosen between vmin = 70 km/h and
vmax = 90 km/h throughout the entire journey. We sample
the start time of the assignments uniformly in an interval of
2 hours and compute the arrival deadlines according to the
default speed.
The pairwise plans are such that trucks platoon as long
as possible. Once a coordination follower splits up from the
coordination leader, it drives fast enough to arrive in time at
its destination and at least at default speed. The split points
are such that arriving in time is feasible. Thus, trucks are
guaranteed to meet their deadlines and the initial value for the
joint vehicle plan optimization fulfills the constraints. Fig. 6
shows an example of the routes of a coordination leader and its
coordination followers and where the coordination followers
join and leave the platoon.
We compare our proposed platoon coordinator to fuel savings that arise from spontaneous platooning, i.e., that trucks
happen to get into each others vicinity and then spontaneously
form platoons. To this end, we collect all the link arrival times
according to the default plans for each link in the scenario. We
sort these times and collect them in ascending order in groups
of at most one minute difference in their edge arrival time. We
assume that each of these groups forms a platoon driving at
default speed and that the default trajectory is not altered by
the platooning. This is a generous estimate since it neglects
X
i∈Nc \Nl
X
i∈Nc
j∈Ni ∩Nl
max ∆F (i, j)
j∈Nio
(7)
max ∆F (i, j),
j∈Nio
where the second inequality holds since ∆F (i, j) > 0 for all
(i, j) ∈ E.
This bound can only be tight when there is an optimal
solution where no coordination leader has an out-neighbor.
Otherwise the coordination leaders cannot contribute to the
sum. Nevertheless, the bound helps us assess how far a
heuristic solution can be away from the optimum.
We implemented platoon coordination in Python and used
CVXOPT [26] for convex optimization. The execution of the
clustering algorithm takes less than a second for 2000 transport
assignments. Even faster computation times could be achieved
by optimizing the implementation.
Each simulation consists of the following steps:
1) Random generation of transport assignments
2) Computation of routes and default plans
3) Computation of the coordination graph
4) Computation of coordination leaders according to Section V
5) Joint vehicle plan optimization according to Section VI
We evaluate how different numbers of assignments affect
the amount of platooning and the fuel savings relative to the
default plans. For comparison we compute the fuel savings
of spontaneous platooning. We run the clustering algorithm
with greedy and random node selection and compute the upper
bound of the objective function fce . The results are averaged
over 150 simulation runs.
Fig. 7 visualizes an example coordination graph. In addition
it shows which assignments are selected in step 4). We can
see that only a small fraction of assignment pairs can save fuel
by forming a platoon. As the number of assignments grows,
more opportunities are available for each assignment which
can translate into larger fuel savings [30].
Fig. 8 shows the effect on the fuel savings when the numbers
of transport assignments that are coordinated is varied. It is
possible to make a number of observations based on these
data. First of all, the fuel savings increase rapidly with the
number of transport assignments when the absolute number
of assignments is small. As more and more assignments are
added, this trend stagnates and the relative fuel savings increase only slowly. Ideally this should approach asymptotically
the maximum fuel savings of 15.9 % as the number of transport
assignments goes to infinity, since then virtually every truck
8
Upper Bound
Greedy, Before
Random, Before
Greedy, After
Random, After
Spontaneous
14
Relative Fuel Savings [%]
12
Fig. 7. This plots visualizes the adjacency matrix of a coordination graphs
with 100 assignments. Nonzero entries are indicated with a black or a red
dot, each corresponding to an edge in the coordination graph. Edges whose
corresponding plans are selected by the clustering algorithm correspond to
the red dots.
10
8
6
4
2
0
is a platoon follower for its entire journey. There is only a
small difference between greedy and random node selection,
however, with the greedy node selection outperforming the
random node selection consistently. For a parallel or even a
distributed implementation of the clustering algorithm, random
node selection would be preferable due to the reduced need for
synchronization whereas greedy node selection is faster in a
centralized setting. Furthermore, the results after selecting the
coordination leaders and before the joint convex optimization
are less than the upper bound but only about 30 % worse. Since
the upper bound is not tight, this indicates that the clustering
algorithm performs well. We can see a clear improvement
in the fuel savings by the joint optimization of the vehicle
plans. Spontaneous platooning gives fuel savings that are less
than half of what can be achieved by coordination. Also
bear in mind that this is a generous estimate of fuel savings
by spontaneous platooning so that the real difference would
probably be even larger.
We conclude that coordinated platooning can yield significant fuel savings and that coordination is crucial leveraging
these savings. For 2000 transport assignments starting over
the course of two hours, we get 7.6 % reduction in fuel
consumption. A number of 2000 trucks starting in that time
interval on an area like Sweden is a realistic number. The
total distance traveled in the simulated scenario is in the same
order of magnitude as the total distance traveled by domestic
road freight transport in Sweden within two hours, assuming
that traffic volume is equally spread over the year [31]. The
density of the road freight traffic that was simulated is only
a fraction of the total road freight traffic in countries with
high population density. The small fuel savings for the platoon
leaders, which have been neglected, would further increase the
platooning benefit.
Fig. 9 shows how the distribution of platoon sizes changes
with the number of transport assignments. We can see that the
larger the number of transport assignments, the more distance
is traveled in large platoons. For 2000 assignments over half
the distance traveled is in a platoon. Most of the distance is
0
1000
2000
3000
4000
5000
Number of Transport Assignments
Fig. 8. The relative fuel savings due to platooning compared to the default
plans with varying numbers of assignments. “Greedy” indicates that greedy
node selection was used in the clustering algorithm, whereas “Random”
indicates random node selection. The keywords “Before”/“After” refer to the
relative fuel savings before/after the joint optimization of the vehicle plans.
“Spontaneous” are the relative fuel savings based on the estimate of fuel
savings due to spontaneous platooning. “Upper Bound” refers to the upper
bound the fce as stated in (7).
traveled in platoons with ten or less vehicles. This is promising
since large platoons might be difficult to control and thus
the platoon coordinator would have to prevent planning for
larger platoons. Since these large platoons only account for a
small fraction of the distance traveled, this would not have too
large an impact on the total fuel savings. The largest platoon
formed has 28 vehicles. A noticeable effect occurs at a number
of 200 transport assignments when more distance is traveled
in relatively large platoons compared to the distribution with
a number of 300 transport assignments. It seems that some
kind of phase transition occurs at these points, where enough
assignments are in the system to go from one coordination
leader with many followers to having several coordination
leaders that are better suited for their followers. To understand
this phenomenon is subject of future work.
The simulations show that computing plans for a large
number of vehicles to form platoons is feasible with the
methods outlined in this paper. It motivates that real-time
platoon coordination enables significant reductions in fuel
consumption and might be the key to leveraging the full
potential of truck platooning.
VIII. C ONCLUSION
A centralized truck platoon coordinator was proposed. The
system provides trucks with vehicle plans that lead to reduced
fuel consumption by making use of platooning. As time
evolves, plans are updated to account for deviations and new
9
Greedy
Platoon Size Distribution [%]
100
10
9
8
7
6
5
4
80
60
3
2
40
20
0
1
1000
2000
3000
4000
5000
Number of Transport Assignments
Random
Platoon Size Distribution [%]
100
10
9
8
7
6
5
4
80
60
3
2
40
20
0
network relate to fuel savings achieved by this method. Furthermore, we want to study a receding horizon implementation
of the platoon coordinator under the presence of disturbances.
Another direction is to study the system in a setting where
more practical details such as speed limits, traffic, driver
rest times, different vehicle types etc. are taken into account.
Finally, similar coordination strategies might be relevant for
other types of multi-agent systems. This work was carried
out in the scope of the COMPANION EU project [32] and
the proposed platoon coordinator is being implemented in a
demonstrator featuring simulated and real trucks.
1
1000
2000
3000
4000
5000
Number of Transport Assignments
Fig. 9. This figure shows the distribution of platoon sizes per distance traveled
over the number of assignments in percent. The upper plot shows the results
of greedy node selection whereas the lower plot shows those of random
node selection in the clustering algorithm. To the right, the size of platoon is
indicated for a platoon size up to ten. So, when the difference between the
first and the second boundary from below is for instance at 20 %, it means
that 20 % of the distance was traveled as member of a platoon of size 2.
assignments. In order to handle the complexity of such coordination, the problem was formulated in a way that for each truck
a number of plans adapted to the default plans of other vehicles
are computed. Each adapted plan involves platooning for some
distance as platoon follower and thus saving fuel. It was
derived how these plans should be systematically combined
in order to maximize the total fuel savings. The NP-hardness
of this problem motivates the proposed heuristic solution
method. Furthermore, we derived how to jointly optimize
the vehicle plans resulting from the combination of default
plans and adapted plans. The effectiveness of the method was
demonstrated in a realistic simulation study. The simulations
motivate that such systems should be deployed once trucks
with the ability to platoon are commercially available.
There are various directions of future work. One direction
is to understand how the transport assignments and the road
A PPENDIX
P ROOF OF P ROPOSITION 1
We show the result by reduction of the optimization version
of the set covering problem to Problem 1. The optimization
version of the set covering problem is well known to be
NP-hard. Reduction to a known hard problem is a common
proof technique for this kind of result [24]. We do this by
constructing a coordination graph Gc for which there is a
one-to-one correspondence between coordination leaders and
selected sets for the cover. Then we show that the minimum
number of leaders that corresponds to a set cover gives the
maximum value for fce .
Consider the following set covering problem. We have a
finite set
S U. Furthermore, let Su be a family of subsets of U
with
S = U. The problem is to find the smallest number
S∈Su
of subsets in Su whose union is U.
We construct the coordination graph as the one shown in
Fig. 10. We introduce a node for each element in U. We denote
the set of these nodes with N3 and let µ3 : U → N3 be a
bijective mapping from the elements in U to the nodes in N3 .
We introduce a node for each element in Su . We denote the set
of these nodes with N2 and let µ2 : Su → N2 be a bijective
mapping from elements in Su to nodes in N2 . Consider a
node n2 ∈ N2 that corresponds to the element S ∈ Su .
The in-neighbors of n2 are Nni 2 = {µ3 (S) : S ∈ µ−1
2 (n2 )}.
The weight of the corresponding edges is 1. We introduce an
additional node N1 . There is an edge from each node in N2
to N1 with weight 0.5. Clearly, this reduction is linear in the
size of the input U, Su .
Since N1 has no out-neighbors, its membership in Nl can
only increase fce (Nl ). Since all nodes in N3 have no inneighbors, adding a node in N3 to Nl can only decrease
fce (Nl ). Thus, the problem of finding the optimal Nl reduces
to finding which nodes in N2 belong to Nl . In the optimal
solution, each node in N3 has at least one out-neighbor in
Nl . Otherwise we could add any out-neighbor of that node to
Nl and increase fce (Nl ) by at least 0.5. Therefore, {µ−1
2 (n) :
n ∈ Nl ∩ N2 } is a set cover of U. Otherwise there would be
u ∈ U such that there is no S ∈ {µ−1
2 (n) : n ∈ Nl ∩ N2 }
with u ∈ S. If such a u existed, µ3 (u) would be a node with
no out-neighbor in Nl ∩ N2 . Furthermore let S¯u ⊆ Su be a
set cover
{µ2 (S) : S ∈ S¯u } has the property that
S of U. Then
i
Nn = N3 , so any set cover has the property
n∈{µ2 (S):S∈S¯u }
that all nodes in N3 have at least one out-neighbor in Nl .
10
N1
.5
N2
1
N3
Fig. 10. Illustration of the graph used to prove that Problem 1 is NP-hard.
Each node in N2 contributes with 0.5 to the objective if it
is not in Nl . Therefore, the optimal Nl contains a minimum
number of nodes from N2 such that every node in N3 has at
least one out-neighbor in Nl ∩ N2 . Since any Nl ∩ N2 that
fulfills this property maps to a set cover S¯u and vice versa, and
since |Nl ∩ N2 | = |S¯u |, we have that S¯u is the solution to the
set covering problem. Thus the NP-hard set-covering problem
can be reduced to Problem 1, which shows that Problem 1 is
NP-hard.
ACKNOWLEDGMENTS
This work was supported by the COMPANION EU project,
the Knut and Alice Wallenberg Foundation, the Swedish
Strategic Research Foundation, and the Swedish Research
Council.
R EFERENCES
[1] R. Horowitz and P. Varaiya, “Control design of an automated highway
system,” Proc. IEEE, vol. 88, no. 7, pp. 913–925, Jul. 2000.
[2] C. Bonnet and H. Fritz, “Fuel consumption reduction in a platoon:
Experimental results with two electronically coupled trucks at close
spacing,” SAE Tech. Paper 2000-01-3056, 2000.
[3] M. P. Lammert et al., “Effect of platooning on fuel consumption of class
8 vehicles over a range of speeds, following distances, and mass,” SAE
Int. J. Commer. Veh., vol. 7, pp. 626–639, Sep. 2014.
[4] S. Tsugawa, “An overview on an automated truck platoon within the
Energy ITS Project,” Advances in Automotive Control, vol. 7, pp. 41–
46, 2013.
[5] T. G. Crainic and G. Laporte, “Planning models for freight transportation,” European J. Operational Research, vol. 97, no. 3, pp. 409–438,
1997.
[6] K. Boriboonsomsin et al., “Eco-routing navigation system based on
multisource historical and real-time traffic information,” IEEE Trans.
Intell. Transp. Syst., vol. 13, no. 4, pp. 1694–1704, Dec. 2012.
[7] M. Papageorgiou et al., “Review of road traffic control strategies,” Proc.
IEEE, vol. 91, no. 12, pp. 2043–2067, Dec. 2003.
[8] E. Larsson et al., “The vehicle platooning problem: Computational
complexity and heuristics,” Transportation Research Part C: Emerging
Technologies, vol. 60, pp. 258–277, 2015.
[9] K.-Y. Liang et al., “When is it fuel efficient for a heavy duty vehicle to
catch up with a platoon?” in 7th IFAC Symp. Advances in Automotive
Control, 2013.
[10] J. Larson et al., “A distributed framework for coordinated heavy-duty
vehicle platooning,” IEEE Trans. Intell. Transp. Syst., vol. 16, no. 1, pp.
419–429, Feb. 2015.
[11] P. Meisen et al., “A data-mining technique for the planning and organization of truck platoons,” in Int. Conf. Heavy Vehicles, Heavy Vehicle
Transport Technology, 2008, pp. 389–402.
[12] S. van de Hoef et al., “Fuel-optimal coordination of truck platooning
based on shortest paths,” in American Control Conf., 2015.
[13] ——, “Coordinating truck platooning by clustering pairwise fuel-optimal
plans,” in 18th IEEE Int. Conf. Intelligent Transportation Syst., 2015.
[14] L. Kaufman and P. J. Rousseeuw, Finding Groups in Data: An introduction to Cluster Analysis. John Wiley & Sons, Inc., 2008.
[15] A. K. Jain and R. C. Dubes, Algorithms for Clustering Data. Upper
Saddle River, NJ, USA: Prentice-Hall, Inc., 1988.
[16] A. K. Jain, “Data clustering: 50 years beyond K-means,” Pattern
Recognition Lett., vol. 31, no. 8, pp. 651–666, 2010.
[17] V. Blondel et al., “Fast unfolding of communities in large networks,” J.
Stat. Mech., P10008, 2008.
[18] S. Harenberg et al., “Community detection in large-scale networks:
a survey and empirical evaluation,” Wiley Interdisciplinary Reviews:
Computational Stat., vol. 6, no. 6, pp. 426–439, 2014.
[19] S. Fortunato, “Community detection in graphs,” Physics Rep., vol. 486,
no. 3–5, pp. 75–174, 2010.
[20] H. Garcia-Molina, “Elections in a distributed computing system,” IEEE
Trans. Comput., vol. C-31, no. 1, pp. 48–59, 1982.
[21] S. Singh and J. Kurose, “Electing “good” leaders,” J. Parallel and
Distributed Computing, vol. 21, no. 2, pp. 184–201, 1994.
[22] E. Demir et al., “A review of recent research on green road freight
transportation,” European J. Operational Research, vol. 237, no. 3, pp.
775–793, 2014.
[23] P. Sanders and D. Schultes, “Engineering fast route planning algorithms,”
in Proc. Experimental Algorithms: 6th International Workshop, Rome,
Italy, Jun. 2007, pp. 23–36.
[24] T. H. Cormen et al., Introduction to Algorithms, 3rd ed. MIT Press,
2009.
[25] S. Boyd and L. Vandenberghe, Convex Optimization. New York, NY:
Cambridge University Press, 2004.
[26] M. S. Andersen et al. (2013) CVXOPT: A Python package for convex
optimization. [Online]. Available: cvxopt.org
[27] Socioeconomic Data and Application Center, “Population density grid,
v3, 2000,” 2015.
[28] D. Luxen and C. Vetter, “Real-time routing with openstreetmap data,”
in Proc. 19th ACM SIGSPATIAL Int. Conf. Advances in Geographic
Inform. Syst., ser. GIS ’11, New York, NY, USA, 2011, pp. 513–516.
[29] B. Besselink et al., “Cyber-physical control of road freight transport,”
Proc. IEEE, 2016.
[30] J. Larson et al., “A distributed framework for coordinated heavy-duty
vehicle platooning,” IEEE Trans. Intell. Transp. Syst., vol. 16, no. 1, pp.
419–429, Feb. 2015.
[31] “Swedish national and international road goods transport 2014,” Transport Analysis Stockholm, Tech. Rep., 2015.
[32] S. Eilers et al., “COMPANION – Towards Co-operative Platoon Management of Heavy-Duty Vehicles,” in 18th IEEE Int. Conf. Intelligent
Transportation Syst., 2015.
| 3cs.SY
|
Optimal Dynamic Sensor Subset Selection for
Tracking a Time-Varying Stochastic Process
arXiv:1711.10610v1 [cs.SY] 28 Nov 2017
Arpan Chattopadhyay, Urbashi Mitra
Abstract—Motivated by the Internet-of-things and sensor
networks for cyberphysical systems, the problem of dynamic
sensor activation for the tracking of a time-varying process
is examined. The tradeoff is between energy efficiency, which
decreases with the number of active sensors, and fidelity, which
increases with the number of active sensors. The problem of
minimizing the time-averaged mean-squared error over infinite
horizon is examined under the constraint of the mean number
of active sensors. The proposed methods artfully combine three
key ingredients: Gibbs sampling, stochastic approximation for
learning, and modifications to consensus algorithms to create
a high performance, energy efficient tracking mechanisms with
active sensor selection. The following progression of scenarios are
considered: centralized tracking of an i.i.d. process; distributed
tracking of an i.i.d. process and finally distributed tracking of
a Markov chain. The challenge of the i.i.d. case is that the
process has a distribution parameterized by a known or unknown
parameter which must be learned. The key theoretical results
prove that the proposed algorithms converge to local optima for
the two i.i.d process cases; numerical results suggest that global
optimality is in fact achieved. The proposed distributed tracking
algorithm for a Markov chain, based on Kalman-consensus
filtering and stochastic approximation, is seen to offer an error
performance comparable to that of a competetive centralized
Kalman filter.
Index Terms—Wireless sensor networks, active sensing, sensor
subset selection, distributed tracking, data estimation, Gibbs
sampling, stochastic approximation, Kalman-consensus filter.
I. I NTRODUCTION
Controlling and monitoring physical processes via sensed
data are integral parts of internet-of-things (IOT), cyberphysical systems, and defense with applications to industrial
process monitoring and control, localization, tracking of mobile objects, environmental monitoring, system identification
and disaster management. In such applications, sensors are
simultaneously resource constrained (power and/or bandwdith)
and tasked to achieve high performance sensing, control,
communication, and tracking. Wireless sensor networks must
further contend with interference and fading. One strategy
for balancing resource use with performance is to activate a
subset of the total possible number of sensors to limit both
computation as well as bandwidth use.
Parts of this paper have been published in previous conferences; see [1],
[2].
Arpan Chattopadhyay and Urbashi Mitra are with Ming Hsieh Department
of Electrical Engineering, University of Southern California, Los Angeles,
USA. Email: {achattop,ubli}@usc.edu
This work was funded by the following grants: ONR N00014-15-1-2550,
NSF CNS-1213128, NSF CCF-1410009, AFOSR FA9550-12-1-0215, and
NSF CPS-1446901.
Herein, we address the fundamental problem of optimal
dynamic sensor subset selection for tracking a time-varying
stochastic process. We first examine the centralized tracking of
an iid process with an unknown, parametric distribution which
serves as a benchmark for the first extension to decentralized
tracking of this process. For both proposed algorithms, almost
sure convergence to local optima can be proven. Next a distributed algorithm for tracking a Markov process with a known
probability transition matrix is developed. All algorithms are
numerically validated.
A. Related Literature
Optimal sensor subset selection problems can be broadly
classified into two categories: (i) optimal sensor subset selection for static data with known prior distribution, but
unknown realization, and (ii) dynamic sensor subset selection
to track a time-varying stochastic process. There have been
several recent attempts to solve the first problem; see [3],
[1]. This problem poses two major challenges: (i) computing
the estimation error given the observations from a subset
of sensors, and (ii) finding the optimal sensor subset from
exponentially many number of subsets. In [3], a tractable lower
bound on performance addressed the first challenge and a
greedy algorithm addressed the second. In [1] only the second
challenge is addressed via Gibbs sampling approach.
Dynamic sensor subset selection for a time-varying stochastic process is considered in e.g. [4], [5], [6], [7], [8], [9].
Markov process tracking with centralized controllers for single
([6]) or multiple sensor selection with energy constraints and
sequential decision making ([4], [5]) have been previously
studied. The optimal policy and its structural properties for
a special case of dynamic sensor selection over an infinite
horizon was examined in [7]. Single sensor selection with the
broadcast of sensor data is studied in [8]. To our knowledge,
the combination of Gibbs sampling (e.g. [10]) and stochastic
approximation (e.g. [11]) has not been previously applied to
iid process tracking as we do herein. The paper [12], using
Thompson sampling, has solved the problem of centralized
tracking of a linear Gaussian process (with unknown noise
statistics) via active sensing.
Given our consideration of dynamic sensor subset selection for distributed tracking of a Markov process, traditional
Kalman filtering is not applicable; however, there is much
prior, recent art on the developmemt of consensus-based
Kalman filtering [13], [14], [15], [16], [17] which we further
adapt to work with our Gibbs sampling/stochastic approxi-
mation approach for the iid case. This distributed tracking
problem does not appear to be heavily studied. In [8] perfect
information sharing between sensors is assumed, which is
impractical and [18] assumes that the estimation is done
in a centralized fashion, while sensors make decentralized
decisions about whether to sense or communicate to the fusion
center. In [18], the sparsity of the Markov process is also
exploited.
iid process having a parametric distribution with unknown
parameters. Distributed tracking of iid process is discussed in
Section V. Distributed tracking of a Markov chain is described
in Section VI. Numerical results are presented in Section VII,
followed by the conclusion in Section VIII.
B. Our Contribution
We consider a connected single or multi-hop wireless sensor
network. The sensor nodes are denoted by the set N =
{1, 2, · · · , N }. It might also be possible that the connectivity
in the network is maintained via a few relay nodes, but
we ignore this possibility for the ease of analysis. There
can be a fusion center (connected to all sensors via single
hop wireless links) responsible for all control or estimation
operations in the network, or, alternatively, the sensors can
operate autonomously in a multihop mesh network.
The physical process under measurement is denoted by
{X(t)}t≥0 , where t is a discrete time index and X(t) ∈ Rq×1 .
We consider two models for the evolution of {X(t)}t≥0 :
In this paper, we make the following contributions:
1) A centralized tracking and learning algorithm for an iid
process with an unknown, but parametric distribution
is developed. In particular, Gibbs sampling minimizes
computational complexity with stochastic approximation
to achieve the mean number of activated sensors constraint. Furthermore, simultaneous perturbation stochastic
approximation (SPSA) is employed for parameter estimation obviating the need for the expectation-maximization
algorithm. A challenge we overcome in the analysis, is
handling updates at different time scales. As a precursor
to the algorithm for unknown parameters, algorithms have
been developed for a simpler version of the problem when
the parameter of the distribution is known.
2) The centralized algorithm, which serves as a benchmark,
is adapted to the distributed case by exploiting partial
consensus. Our partial consensus is novel in that SPSA
is again employed to learn the optimal consensus gains
adaptively.
3) A trick for ensuring that all sensors employ similar
sampling strategies is to have each sensor use the same
seed in a random number generator.
4) For both the centralized and distributed algorithms, we
can prove almost sure convergence to local optima.
Furthermore, we an prove that the resources needed
for communication and learning can be made arbitrarily
small, by exploiting properties of the multi-scale updates.
5) Our final generalization to is to develop an algorithm
for sensor subset selection for the decentralized tracking
of a Markov chain with known probability transition
matrix. We adapt methods for Kalman consensus filtering
to our framework with Gibbs sampling and stochastic
approximation.
6) Numerical results show that the decentralized scheme for
a Markov chain performs close to that of the centralized
scheme. Numerical results also show a tradeoff between
performance and message exchange for learning. Furthermore, the numerical results show that global (not local)
optima are achieved in tracking iid process.
C. Organization
The rest of the paper is organized as follows. The system model is described in Section II. Centralized tracking
of an iid process with known distribution is described in
Section III. Section IV deals with centralized tracking of an
II. S YSTEM M ODEL
A. Network, data and sensing model
•
•
IID model: {X(t)}t≥0 is an i.i.d. process with a parametric distribution pθ0 (X), where the unknown parameter θ0
needs to be be learnt via the measurements. θ0 lies inside
the interior of a compact subset Θ ⊂ Rd .
Markov model: {X(t)}t≥0 is a finite-state ergodic
Markov chain with known transition probability matrix.
At time t, if a sensor k is used to sense the process, then
the observation at sensor k is provided by a r-dimensional
column vector zk (t) = Hk X(t) + vk (t) where Hk is an
observation matrix of appropriate dimension, and vk (t) is a
Gaussian vector; the observation noise vk (t) is assumed to be
independent across k and i.i.d. across t.
Let B(t) ∈ {0, 1}1×N be a vector where Bk (t) is the
indicator that the sensor k is activated at time t; Bk (t) = 1 if
sensor k is active at time t, or else Bk (t) = 0. The decision to
activate any sensor for sensing and communicating the observation is taken either by the fusion center or by the sensor itself
in the absence of a fusion center. We denote by B := {0, 1}N
the set of all possible configurations in the network, and by B
a generic configuration. Clearly, B(t) ∈ B. Each configuration
represents a set of activated sensors. B−j ∈ {0, 1}N −1 is used
to represent the configuration B with its j-th entry removed.
The observation made by sensor k at time t is Zk (t) =
Bk (t)zk (t). We define Z(t) := {Zk (t) : 1 ≤ k ≤ N }}.
B. Centralized estimation problem
The estimate of X(t) at the fusion center (connected to
all sensors via direct wireless links) is denoted by X̂(t). We
denote by Hp (t) the information available at time t at the
fusion center about the history of observations, activations
and estimates up to time (t − 1), before the fusion center
determines B(t). On the other hand, we define Hc (t) :=
{B(t); Z(t); θ(t)} where θ(t) is the current estimate of θ0 ;
Hc (t) denotes the information used by the fusion center at
time t to estimate X̂(t). For i.i.d. time varying process, Hc (t)
is sufficient to estimate X(t) and obtain X̂(t), and Hc (t) will
be available only after deciding the activation vector B(t) and
collecting all observations. In order to optimally decide B(t),
the fusion center needs knowledge about the performance of
all configurations in the past. Hence, Hp (t) and Hc (t) have
two different information structures. However, we will see that,
our Gibbs sampling algorithm determines B(t) by using only
a sufficient statistic calculated iteratively in each slot.
The information structure Hc (t) used to track a Markov
chain will be different, which we will see in Section VI.
We define a policy µ = {(µ1 , µ2 } as a pair of mappings,
where µ1 (Hp (t)) = B(t) and µ2 (Hc (t)) = X̂(t). Our first
goal is to solve the following centralized problem of minimizing the time-average mean squared error (MSE) subject to a
constraint on the mean number of sensors active per unit time:
µ
t→∞
(P1)
t
N
1 X 1 X
Eµ ||X(τ ) − X̂ (k) (τ )||2
t τ =1 N k=1
s.t. lim sup
t→∞
t
1 X
Eµ ||B(τ )||1 ≤ N̄
t τ =1
(P4)
:=f (B)
In the absence of a fusion center in a multi-hop network, the
estimate of X(t) at sensor k is denoted by X̂ (k) (t). We denote
(k)
by Hp (t) the information available at time t at sensor k
about the history before the sensor determines Bk (t), and by
(k)
Hc (t) the information available at sensor k at time t just
before X̂ (k) (t) is estimated.
(k)
(k)
We define a policy µ = {(µ1 , µ2 )}1≤k≤N , where
(k)
(k)
(k)
(k)
µ1 (Hp (t)) = Bk (t) and µ2 (Hc (t)) = X̂ (k) (t).
We seek to solve the following distributed problem:
arg min lim sup
t
1X
2
Eµ ||X(τ )−X̂(τ )|| +λ||B(τ )||1
µ = arg min lim sup
µ
t→∞ t
τ =1
(P3)
The multiplier λ ≥ 0 can be viewed as the cost incurred for
activating a sensor at any time instant. Now, since (P3) is an
unconstrained problem and X(t) is i.i.d. across t, there exists
one optimizer B ∗ ∈ B (not necessarily unique) for the problem
(P3); if the configuration B ∗ is chosen at each t, the minimum
cost of (P3) can be achieved by the law of large numbers.
Hence, for known θ0 , the problem (P3) can be equivalently
written as:
∗
{z
|
C. Distributed estimation problem
µ∗ =
We relax (P1) by using a Lagrance multiplier λ and obtain
the following unconstrained problem:
arg min Eµ2 ,B ||X(τ ) − X̂(τ )||2 +λ||B||1
B∈B |
{z
}
t
1 X
Eµ ||X(τ ) − X̂(τ )||2
µ∗ = arg min lim sup
µ
t→∞ t
τ =1
t
1 X
s.t. lim sup
Eµ ||B(τ )||1 ≤ N̄
t→∞ t
τ =1
A. The relaxed version of the constrained problem
(P2)
III. C ENTRALIZED T RACKING OF IID PROCES : KNOWN θ0
In this section, we provide an algorithm for solving the
centralized problem (P1) when X(t) ∼ pθ0 (x) i.i.d. This
is done by relaxing (P1) by a Lagrange multiplier. Though
our final goal is to track an i.i.d. process with unknown θ0 ,
we discuss algorithms for known θ0 in this section, as a
precursor to the algorithms developed in subsequent sections
for unknown θ0 and also tracking a Markov chain. Also,
extension to distributed tracking will be discussed in Section V
for unknown θ0 , and hence will be omitted in this section.
:=h(B)
}
The following result tells us how to choose the optimal λ∗ to
solve (P1).
Theorem 1: Consider problem (P1) and its relaxed version
(P4). If there exists a Lagrange multiplier λ∗ ≥ 0 and a
B ∗ ∈ B, such that an optimal configuration for (P4) under
λ = λ∗ is B ∗ , and the constraint in (P1) is satisfied with
equality under the pair (B ∗ , λ∗ ), then B ∗ is an optimal
configuration for (P1).
In
case
there
exist
multiple
configurations
∗
, a multiplier λ∗ ≥ 0, and a probability
B1∗ , B2∗ , · · · , Bm
mass function (p1 , p2 , · · · , pm ) such that (i) each of
∗
is optimal for problem (P4) under λ∗ ,
· · · , Bm
B1∗ , B2∗ , P
m
and (ii) i=1 pi ||Bi∗ ||1 = N̄ , then an optimal solution for
∗
(P1) is to choose one configuration from B1∗ , B2∗ , · · · , Bm
with probability mass function (p1 , p2 , · · · , pm ).
Proof: See Appendix A.
Remark 1: Theorem 1 allows us to obtain a solution for
(P1) from the solution of by choosing an appropriate λ∗ ; this
will be elaborated upon in Section III-D.
B. Basics of Gibbs sampling for known pθ0 (·)
Finding the optimal solution of (P4) requires us to search
over 2N possible configurations and to compute MMSE for
each of these configurations. Hence, we propose Gibbs sampling based algorithms to avoid this O(2N ) computation.
Let us define the probability distribution πβ (·) over B as
(with β > 0):
e−βh(B)
e−βh(B)
:=
.
0
−βh(B )
Zβ
B 0 ∈B e
πβ (B) := P
(1)
Following the terminology in statistical physics, we call
β the inverse temperature, and Zβ the partition function.
h(B) is P
viewed as the energy under configuration B. Now,
limβ↑∞ B∈arg minA∈B h(A) πβ (B) = 1. Hence, if a con-
figuration B(t) is selected at each time t with probability
distribution πβ (·) for sufficiently large β > 0, then B(t) will
belong to the set of minimizers of (P3) with high probability.
However, computing Zβ requires 2N addition operations;
hence, we use a sequential subset selection algorithm based
on Gibbs sampling (see [10, Chapter 7]) in order to avoid
explicit computation of Zβ while picking X(t) ∼ pθ0 (·).
Algorithm 1: Start with an initial configuration B(0). At
time t, pick a random sensor jt uniformly from the set of
all sensors. Choose Bjt (t) = 1 with probability p(t) :=
−βh(B−j (t−1),1)
t
e
−βh(B−j (t−1),1)
−βh(B−j (t−1),0) and choose Bjt (t) = 0
t
t
e
+e
with probability (1 − p(t)). For k 6= jt , choose Bk (t) =
Bk (t − 1). Activate the sensors according to B(t).
Theorem 2: Under Algorithm 1, {B(t)}t≥0 is a reversible,
ergodic, time-homogeneous Markov chain with stationary
distribution πβ (·).
Proof: Follows from the theory in [10, Chapter 7]).
Remark 2: Theorem 2 tells us that if the fusion center runs
Algorithm 1 and reaches the steady state distribution of the
Markov chain {B(t)}t≥0 , then the configuration chosen by the
algorithm will have distribution πβ (·). For very large β > 0,
if one runs {B(t)}t≥0 for a sufficiently long, finite time T ,
then the terminal state BT will belong to arg minB∈B h(B)
with high probability. Also, by the ergodicity of {B(t)}t≥0 ,
the time-average occurence rates of all configurations match
the distribution πβ (·) almost surely.
C. The exact solution
Algorithm 1 is operated with a fixed β; but the optimal soultion of the unconstrained problem (P3) can only be obtained
with β ↑ ∞; this is done by updating β at a slower time-scale
than the iterates of Algorithm 1.
Algorithm 2: This algorithm is same as Algorithm 1 except that at time t, we use β(t) := β(0) log(1 + t)
to compute the update probabilities, where β(0) > 0,
β(0)N ∆ < 1, and ∆ := maxB∈B,A∈B |h(B) − h(A)|.
Theorem 3: Under
Algorithm
2,
the
Markov
chain
{B(t)}t≥0
is
strongly
ergodic,
and
the
limiting
probability
distribution
satisfies
P
limt→∞ A∈arg minC∈B h(C) P(B(t) = A) = 1.
Proof: See Appendix C. We have used the notion of
weak and strong ergodicity of time-inhomogeneous Markov
chains from [10, Chapter 6, Section 8]), which is provided in
Appendix B. The proof is similar to the proof of one theorem
in [19], but is given here for completeness.
Remark 3: Theorem 3 shows that we can solve (P3) exactly
if we run Algorithm 2 for infinite time, in contrast with
Algorithm 1 which provides an approximate solution.
Remark 4: For i.i.d. time varying {X(t)}t≥0 with known
joint distribution, we can either: (i) find the optimal configuration B ∗ using Algorithm 2 and use B ∗ for ever, or (ii)
run Algorithm 2 at the same timescale as t, and use the
running configuration B(t) for sensor activation; both schemes
will minimize the cost in (P3). By the strong ergodicity
of {B(t)}t≥0 , optimal cost will be achieved for (P3) under
Algorithm 2.
1) Convergence rate of Algorithm 1: Let π (t) denote the
probability distribution of B(t) under Algorithm 1. Let us
consider the transition probability matrix P of the Markov
chain {Y (l)}l≥0 with Y (l) = B(lN ), under Algorithm 1. Let
us recall the definition of the Dobrushin’s ergodic coefficient
δ(P ) from [10, Chapter 6, Section 7] for the matrix P ; using
a method similar to that of the proof of Theorem 3, we can
−βN ∆
show that δ(P ) ≤ (1 − e N N ). Then, by [10, Chapter 6,
Theorem 7.2], we can say that under Algorithm
1, we have
dV (π (lN ) , πβ ) ≤ dV (π (0) , πβ ) 1 −
e−βN ∆
NN
l
. We can prove
similar bounds for any t = lN + k, where 0 ≤ k ≤ N − 1.
Unfortunately, we are not aware of such a closed-form
bound for Algorithm 2.
Remark 5: Clearly, under Algorithm 1, the convergence
rate decreases as β increases. Hence, there is a trade-off
between convergence rate and accuracy of the solution in this
case. Also, the rate of convergence decreases with N . For
Algorithm 2, the convergence rate is expected to decrease with
time.
D. Gibbs sampling and stochastic approximation based approach to solve the constrained problem
In Section III-B and Section III-C, we presented Gibbs
sampling based algorithms for (P3). Now we provide an
algorithm that updates λ with time in order to meet the
constraint in (P1) with equality, and thereby solves (P1) (via
Theorem 1).
Lemma 1: For the unconstrained problem (P4), the optimal mean number of active sensors, Eµ2 ||B ∗ ||1 , decreases
with λ. Similarly, the optimal error, Eµ2 f (B ∗ ), increases
with λ.
Proof: See Appendix D.
Remark 6: The optimal mean number of active sensors,
Eµ2 ||B ∗ ||1 , for the unconstrained problem (P4) is a decreasing
staircase function of λ, where each point of discontinuity is
associated with a change in the optimizer B ∗ (λ).
Lemma 1 provides an intuition about how to update λ in
Algorithm 1 or in Algorithm 2 in order to solve (P1). We
seek to provide one algorithm which updates λ(t) at each
time instant, based on the number of active sensors in the
previous time instant. In order to maintain the necessary
timescale difference between the {B(t)}t≥0 process and the
λ(t) update process, we use stochastic approximation ([11])
based update rules for λ(t). But the above remark tells us that
the optimal solution of the constrained problem (P1) requires
us to randomize between two values of λ in case the optimal
λ∗ as in Theorem 1 belongs to the set of such discontinuities.
However, this randomization will require us to update a randomization probability at another timescale; having stochastic
approximations running in multiple timescales leads to slow
convergence. Hence, instead of using a varying β(t), we use a
fixed, but large β and update λ(t) in an iterative fashion using
stochastic approximation. Our proposed Algorithm 3 updates
λ(t) iteratively in order to solve (P1).
Algorithm 3: Choose any initial B(0) ∈ {0, 1}N and
λ(0) ≥ 0. At each discrete time instant t = 0, 1, 2, · · · ,
pick a random sensor jt ∈ N independently and uniformly. For sensor jt , choose Bjt (t) = 1 with probability
−βhλ(t) (B−j (t−1),1)
t
e
p := −βhλ(t) (B−j
(t−1),1)
−βhλ(t) (B−j (t−1),0) and choose
t
t
e
+e
Bjt (t) = 0 with probability (1−p). For k 6= jt , we choose
Bk (t) = Bk (t − 1).
After this operation, before the (t + 1) decision instant,
update λ(t) at each node as follows.
λ(t + 1) = [λ(t) + a(t)(||B(t − 1)||1 − N̄ )]cb
The stepsize
sequence
P∞{a(t)}t≥1 constitutes
P∞a positive
2
such that
t=1 a(t) = ∞ and
t=1 a (t) = ∞. The
nonnegative projection boundaries b and c for the λ(t)
iterates are such that λ∗ ∈ (b, c) where λ∗ is defined in
Assumption 1.
The update of λ(t) in Algorithm 3 is inspired by the
following result which is crucial in the convergence proof.
Lemma 2: Under Algorithm 1, Eµ2 ||B(t)||1 is a Lipschitz
continuous and decreasing function of λ.
Proof: See Appendix E.
Discussion of Algorithm 3:
•
•
If ||B(t−1)||1 is more than N̄ , then λ(t) is increased with
the hope that this will reduce the number of active sensors
in subsequent iterations, as suggested by Lemma 2.
The B(t) and λ(t) processes run on two different
timescales; B(t) runs in the faster timescale whereas λ(t)
runs in a slower timescale. This can be understood from
the fact that the stepsize in the λ(t) update process decreases with time t. Here the faster timescale iterate will
view the slower timescale iterate as quasi-static, while the
slower timescale iterate will view the faster timescale as
almost equilibriated. This is reminiscent of two-timescale
stochastic approximation (see [11, Chapter 6]).
We make the following feasibility assumption for (P1),
under the chosen β > 0.
Assumption 1: There exists λ∗ ≥ 0 such that the constraint
in (P1) under λ∗ and Algorithm 1 is met with equality.
Remark 7: By Lemma 2, E||B||1 continuously decreases in
λ. Hence, if N̄ is feasible, then such a λ∗ must exist by the
intermediate value theorem.
Let us define: hλ(t) (B) := f (B) + λ(t)||B||1 . Let πβ|λ∗ (·)
denote πβ (·) under λ = λ∗ .
Theorem 4: Under Algorithm 3 and Assumption 1, we
have λ(t) → λ∗ almost surely, and the limiting distribution
of {B(t)}t≥0 is πβ|λ∗ (·).
Proof: See Appendix F.
This theorem says that Algorithm 3 produces a configuration
from the distribution πβ|λ∗ (·) under steady state.
E. A hard constraint on the number of activated sensors
Let us consider the following modified constrained problem:
min f (B) s.t. ||B||1 ≤ N̄
B∈B
(MCP)
It is easy to see that (MCP) can be easily solved using similar
Gibbs sampling algorithms as in Section III, where the Gibbs
sampling algorithm runs only on the set of configurations
which activate N̄ number of sensors. Thus, as a by-product,
we have also proposed a methodology for the problem in [3],
though our framework is more general than [3].
Remark 8: The constraint in (P1) is weaker than (MCP).
Remark 9: If we choose β very large, then the number of
sensors activated by GIBBSLEARNING will have very small
variance. This allows us to solve (MCP) with high probability.
IV. C ENTRALIZED T RACKING OF IID PROCESS :
U NKNOWN θ0
In Section III, we described algorithms for centralized
tracking of an i.i.d. process where pθ0 (·) is known. In this
section, we will deal with the centralized tracking of an i.i.d.
process {X(t)}t≥0 where X(t) ∼ pθ0 (·) with θ0 unknown; in
this case, θ0 has to be learnt over time through observations,
which creates many nontrivial issues that need to be addressed
before using Gibbs sampling for sensor subset selection.
A. The proposed algorithm for unknown θ0
Since θ0 is unknown, its estimate θ(t) has to be updated over
time using the sensor observations. On the other hand, to solve
the constrained problem (P1), we need to update λ(t) over time
so as to attain the optimal λ∗ of Theorem 1 iteratively. Further,
f (B) (MSE under configuration B) in (P4) is unknown since
θ0 is unknown, and f (t) (B) has to be learnt over time using the
sensor observations. Hence, we combine the Gibbs sampling
algorithm with update schemes for f (t) (B), λ(t) and θ(t)
using stochastic approximation (see [11]).
The algorithm also requires a sufficiently large positive
number A0 and a large integer T as input.
Let J (t) denote the indicator
Pt that time t is an integer
multiple of T . Define ν(t) := τ =0 J (τ ).
We first describe some key features and steps of the algorithm and then provide a brief summary of the algorithm.
1) Step size: For the stochastic approximation updates, the
algorithm uses step sizes which are nonnegative sequences
{a(t)}t≥0 , {b(t)}t≥0 , {c(t)}t≥0 , {d(t)}t≥0 such that:
P∞
a(t)
2 (t)
a
<
t=0
(i)
P∞
(iii)
t=0
limt→∞ d(t)
limt→∞
b(t)
a(t)
P∞
=
b(t)
P∞t=02
∞, t=0 b (t)
=
= limt→∞
0,
t c)
c(b T
b(t)
P∞
=
=
t=0 c(t)
P
2 (t)
<
∞, ∞
c
t=0
P∞ c2 (t)
<
(iv)
t=0 d2 (t)
= 0.
∞, (ii)
<
∞,
∞,
(v)
2) Gibbs sampling step: The algorithm also maintains
a running estimate h(t) (B) of h(B) for all B ∈ B. At
time t, it selects a random sensor jt ∈ N with uniformly
and independently, and sets Bjt (t) = 1 with probability
p(t) :=
−βh(t) (B−j (t−1),1)
t
e
−βh(t) (B−j (t−1),1)
−βh(t) (B−j (t−1),0)
t
t
e
+e
and Bjt (t) = 0
with probability (1−p(t)) (similar to Algorithm 1). For k 6= jt ,
it sets Bk (t) = Bk (t−1). This operation can even be repeated
multiple times. The sensors are activated according to B(t),
and the observations ZB(t) (t) := {zk (t) : Bk (t) = 1} are
collected. Then the algorithm declares X̂(t) = µ2 (Hc (t)).
3) Occasional reading of all the sensors: If J (t) = 1,
the fusion center reads all sensors and obtains Z(t). This
is required primarily because we seek to update θ(t) iteratively and reach a local maximum of the function g(θ) =
EX(t)∼pθ0 (·),B(t)=[1,1,··· ,1] log p(Z(t)|θ).
4) θ(t) update when J (t)
=
1: Since we
seek to reach a local maximum of g(θ)
=
EX(t)∼pθ0 (·),B(t)=[1,1,··· ,1] log p(Z(t)|θ), a gradient ascent
scheme needs to be used. The gradient of g(θ) along any
coordinate can be computed by perturbing θ in two opposite
directions along that coordinate and evaluating the difference
of g(·) at those two perturbed values. However, if θ0 is
high-dimensional, then estimating this gradient along each
coordinate is computationally intensive. Moreover, evaluating
g(θ) for any θ requires us to compute an expectation, which
might also be expensive. Hence, we perform a noisy gradient
estimation for g(θ) by simultaneous perturbation stochastic
approximation (SPSA) as in [20]. Our algorithm generates
∆(t) ∈ {1, −1}d uniformly over all sequences, and perturbs
the current estimate θ(t) by a random vector d(ν(t))∆(t) in
two opposite directions, and estimates each component of the
gradient from the difference log p(Z(t)|θ(t) + d(ν(t))∆(t)) −
log p(Z(t)|θ(t) − d(ν(t))∆(t)); this estimate is noisy because
(i) Z(t) and ∆(t) are random, and (ii) d(ν(t)) > 0.
The k-th component of θ(t) is updated as follows:
θk (t + 1)
=
log p(Z(t)|θ(t) + d(ν(t))∆(t))
θk (t) + c(ν(t))J (t)
2d(ν(t))∆k (t)
log p(Z(t)|θ(t) − d(ν(t))∆(t))
−
(2)
2d(ν(t))∆k (t)
Θ
The iterates are projected onto the compact set Θ to ensure
boundedness. The diminishing sequence {d(t)}t≥0 ensures
that the gradient estimate becomes more accurate with time.
5) λ(t) update: λ(t) is updated as follows:
0
λ(t + 1) = [λ(t) + b(t)(||B(t)||1 − N̄ )]A
0 .
(3)
The intuition here is that, if ||B(t)||1 > N̄ , the sensor activation cost λ(t) needs to be increased to prohibit activating large
number of sensors in future; this is motivated by Lemma 2.
The goal is to converge to λ∗ as defined in Theorem 1.
6) f (t) (B) update: Since pθ0 (X) is not known initially,
the true value of f (B) is not known; hence, the algorithm
updates an estimate f (t) (B) using the sensor observations.
If J (t) = 1, the fusion center obtains Z(t) by reading all
sensors. The goal is to obtain an estimate YB (t) of the MSE
under a configuration B, by using these observations, and
update f (t) (B) using YB (t). However, since θ0 is unknown
and only θ(t) is available, as an alternative to the MSE
under configuration B, the fusion center uses the trace of the
conditional covariance matrix of X(t) given ZB (t), assuming
that X(t) ∼ pθ(t) (·). Hence, we define a random variable
YB (t) := EX(t)∼p(·|θ(t),ZB (t)) (||X(t) − X̂B (t)||2 |ZB (t), θ(t))
for all B ∈ B, where X̂B (t) is the MMSE estimate declared by
µ2 under configuration B, and ZB (t) is the observation made
by active sensors determined by B. Clearly, YB (t) is a random
variable with the randomness coming from two sources: (i)
randomness of θ(t), and (ii) randomness of ZB (t) which has
a distribution p(ZB (t)|θ0 ) since the original X(t) process that
yields ZB (t) has a distribution pθ0 (·). Computation of YB (t)
is simple for Gaussian X(t) and the MMSE estimator, since
closed form expressions are available to compute YB (t).
Using YB (t), the following update is made for all B ∈ B:
0
f (t+1) (B) = [f (t) (B) + J (t)a(ν(t))(YB (t) − f (t) (B))]A
0
(4)
The iterates are projected onto [0, A0 ] to ensure boundedness. The goal here is that, if θ(t) → θ∗ , then f (t) (B)
will converge to EZB (t)∼p(·|θ0 ) EX(t)∼p(·|θ∗ ,ZB (t)) (||X(t) −
X̂B (t)||2 |ZB (t), θ∗ ), which is equal to f (B) under θ∗ = θ0 .
We will later argue that this occasional O(2N ) computation
for all B ∈ B can be avoided, but convergence will be slow.
7) The algorithm: A summary of all the steps of our
scheme is provided in Algorithm 4. We will show in Theorem 5
that this algorithm almost surely converges to the set of locally
optimum solutions for (P1).
Algorithm 4: Initialize all iterates arbitrarily.
For any time t = 0, 1, 2, · · · :
1. Perform the Gibbs sampling step as in Section IV-A2,
obtain the observations ZB(t) (t), and estimate X̂(t). Update λ(t) according to (3).
2. If J (t) = 1, compute h(t) (B) = f (t) (B) + λ(t)||B||1
for all B ∈ B. Read all sensors and obtain Z(t). Update
f (t) (B) for all B ∈ B using (4) and θ(t) using (2).
8) Multiple timescales in Algorithm 4: Algorithm 4 has
multiple iterations running in multiple timescales (see [11,
Chapter 6]). The {B(t)}t≥0 process runs ar the fastest
timescale, whereas the {θ(t)}t≥0 update scheme runs at the
slowest timescale. The basic idea is that a faster timescale
iterate views a slower timescale iterate as quasi-static, whereas
a slower timescale iterate views a faster timescale iterate as
c(t)
= 0,
almost equilibriated. For example, since limt→∞ a(t)
the θ(t) iterates will vary very slowly compared to f (t) (B)
iterates; as a result, f (t) (B) iterates will view quasi-static θ(t).
B. Complexity of Algorithm 4
1) Sampling and communication complexity: Since all sensors are activated when J (t) = 1, the mean number of additional active sensors per unit time is O( N
T ); these observations
need to be communicated to the fusion center. O( N
T ) can be
made arbitrarily small by choosing T large enough.
2) Computational complexity: The computation of YB (t)
in (4) for all B ∈ B requires O(2N ) computations whenever
J (t) = 1. However, if one chooses large T (e.g., O(4N )),
then this additional computation per unit time will be small.
However, if one wants to avoid that computation also, then one
can simply compute YB(t) (t) and update f (t) (B(t)) instead of
doing it for all configurations B ∈ B. However, the stepsize
sequence a(ν(t)) cannot be used; instead, a stepsize a(νB (t))
has to be used when P
B(t) = B and f (t) (B) is updated using
t
(4), where νB (t) := τ =0 J (τ )I(B(τ ) = B) is the number
of times configuration B was chosen till time t. In this case,
the convergence result (Theorem 5) on Algorithm 4 will still
hold; however, the proof will require a technical condition
lim inf t→∞ νBt(t) > 0 almost surely for all B ∈ B, which will
be satisfied by the Gibbs sampler using finite β and bounded
h(t) (B). However, we discuss only (4) update in this paper for
the sake of simplicity in the convergence proof, since technical
details of asynchrounous stochastic approximation required in
the variant is not the main theme of this paper.
When J (t) = 1, one can avoid computation of h(t+1) (B)
for all B ∈ B in Step 2 of Algorithm 4. Instead, the fusion
center can update only h(t) (B(t)), h(t) (B−jt (t − 1), 1) and
h(t) (B−jt (t − 1), 0) at time t, since only these iterates are
required in the Gibbs sampling.
C. Convergence of Algorithm 4
1) List of assumptions:
Assumption 2: The distribution pθ (·) and the mapping
(k)
µ2 (·; ·; θ) (or µ2 for distributed case) as defined before are
Lipschitz continuous in θ ∈ Θ.
Assumption 3: µ2 is known to the fusion center (centralized
(k)
case), and {µ2 }k≥1 are known to all sensors (distributed).
Remark 10: Assumption 3 allows us to focus only on the
sensor subset selection problem rather than the problem of estimating the process given the sensor observations. For optimal
MMSE estimators, µ2 (Hc (t)) = E(X(t)|Hc (t)). Computation
of µ2 (·) will depend on the exact functional form of pθ (X),
and it can be done by using Bayes’ theorem.
Assumption 4: Let us consider YB (t) with θ(t) = θ fixed in
Algorithm 4. Suppose that, one uses Algorithm 1 to solve (P3)
for a given λ but with the MSE ||X(t) − X̂(t)||2 replaced by
YB(t) (t) in the objective function of (P3), and then finds the
λ∗ (θ) as in Theorem 1 to meet the constraint N̄ . We assume
that, for the given β and N̄ , and for each θ ∈ Θ, there exists
λ∗ (θ) ∈ [0, A0 ) such that, the optimal Lagrange multiplier to
relax this new unconstrained problem is λ∗ (θ) (Theorem 1).
Also, λ∗ (θ) is Lipschitz continuous in θ ∈ Θ.
Remark 11: Assumption 4 makes sure that the λ(t) iteration (3) converges, and the constraint is met with equality.
Θ −θ
.
Let us define the function Γ̄θ (φ) := limδ↓0 [θ+δφ]
δ
Assumption 5: Consider
the
function
g(θ)
=
EX(t)∼pθ0 (·),B(t)=[1,1,··· ,1] log p(Z(t)|θ); this is the expected
conditional log-likelihood function of Z(t) conditioned on θ,
given that X(t) ∼ pθ0 (·) and B(t) = [1, 1, · · · , 1]. We assume
that the ordinary differential equation θ̇(t) = Γ̄θ(t) (∇g(θ(t)))
has a globally asymptotically stable solution θ∗ in the interior
of Θ. Also, ∇g(θ) is Lipschitz continuous in θ.
Remark 12: One can show that the θ(t) iteration (2) asymptotically tracks the ordinary differential equation θ̇(τ ) =
∇g(θ(τ )) inside the interior of Θ. In fact, Γ̄θ(τ ) (∇g(θ(τ )) =
∇g(θ(τ )) when θ(τ ) lies inside the interior of Θ. The condition on θ̇(τ ) = Γ̄θ(τ ) (∇g(θ(τ ))) is required to make sure that
the iteration does not converge to some unwanted point on the
boundary of Θ due to the forced projection. The assumption
on θ∗ makes sure that the θ(t) iteration converges to θ∗ .
2) The main result: Now we present convergence result for
Algorithm 4. This result tells us that the iterates of Algorithm 4
almost surely converge to the set of local optima for (P1).
Theorem 5: Under Assumptions 2, 3, 4, 5 and Algorithm 4, we have limt→∞ θ(t) = θ∗ almost surely.
Correspondingly, λ(t) → λ∗ (θ∗ ) almost surely. Also,
f (t) (B) → EZB (t)∼p(·|θ0 ) EX(t)∼p(·|θ∗ ,ZB (t)) (||X(t) −
X̂B (t)||2 |ZB (t), θ∗ ) =: fθ∗ (B) almost surely for all B ∈
B. The B(t) process reaches the steady-state distribution
πβ,fθ∗ ,λ∗ (θ∗ ),θ∗ (·) which can be obtained by replacing
h(B) in (1) by fθ∗ (B) + λ∗ ||B||1 .
Proof: See Appendix G.
Remark 13: If θ(t) → θ∗ , but the constraint in (P1) is
satisfied with λ = 0 and policy µ2 (·; ·; θ∗ ), then λ(t) → 0,
i.e., λ∗ (θ∗ ) = 0, and the constraint becomes redundant. If
θ∗ = θ0 , then the algorithm reaches the global optimum.
Remark 14: If all sensors are not read when J (t) =
1, then one has to update θ(t) based on the observations
ZB(t) (t) collected from the sensors determined by B(t).
In that case, θ(t) will converge to a local maximum θ1
of limt→∞ Eπβ,fθ ,λ∗ (θ),θ log p(ZB(t) (t)|θ), which will be different from θ∗ of Theorem 5 in general. However, in the
numerical example in Section VII, we observe that θ1 = θ∗ .
V. D ISTRIBUTED TRACKING OF THE I . I . D . PROCESS
We next seek to solve the constrained problem (P2). This
problem brings in additional challenges compared to (P1):
(i) each sensor k has access only to its local measurement,
i.e., zk (t), if Bk (t) = 1, (ii) sharing measurements across
the network will consume a large amount of energy and
bandwidth, and (iii) ideally, the iterates λ(t) and θ(t) should
be known at all sensors. To resolve these issues, we propose an
algorithm that combines Algorithm 4 with consensus among
sensor nodes (see [16]). However, our approach is different
from traditional consensus schemes in the following aspects:
(i) traditional consensus schemes run many steps of a consensus iteration, thus requiring many rounds of message exchange
among neighbouring nodes, and (ii) traditional consensus
schemes do not care about the correctness of the data at any
particular sensor node. In contrast, our proposed algorithm
allows each sensor to broadcast its local information only once
to its neighbours in a time slot. Also, since many of the sensors
may use outdated estimates, we propose an on-line learning
scheme based on stochastic approximation in order to optimize
the coefficients of the linear combination used in consensus.
the following updates are done for all links (i, j):
(t+1)
KB
Y (t)
A0
Y (t)
KB +d(νB (t))Γt
KB −d(νB (t))Γt
I(B(t) = B)
−
(6)
2d(νB (t))Γt (i, j)
2d(νB (t))Γt (i, j)
−A0
A. The proposed algorithm
Note that, if the Gibbs sampling step of Algorithm 4 is run
at all sensors to make their individual activation decisions,
but all sensors are supplied with the same initial seed for
randomization, then all sensors will sample the same B(t) at
each time t. We will exploit this fact in the next algorithm.
However, depending on the current configuration B(t) = B,
each node uses a linear combination of its own estimate
and the estimates made by its neighbours. Let us denote
the initial estimate made by node k at time t by X̄ (k) (t);
this estimation is done at node k based on Hc (t). The
actual estimate X̂(t) := [X̂ (1) (t), X̂ (2) (t), · · · , X̂ (N ) (t)]0 is
obtained from X̄(t) := [X̄ (1) (t), X̄ (2) (t), · · · , X̄ (N ) (t)]0 by
(t)
X̂(t) = KB(t) X̄(t); this method is motivated by the Kalman
(t)
consensus filter proposed in [15]. Here KB ∈ RN ×N is the
weight matrix to be used at time t under configuration B;
this matrix has (i, j)-th entry zero if nodes i and j are not
connected in the wireless network, it can induce a consensus
(t)
such as in [16]. The matrices KB for all B ∈ B are updated
for all t when J (t) = 1, and are broadcast to the sensors. As
with the θ(t) update in (2), we use SPSA to find optimal KB
in order to minimize the error.
First we
the algorithm. Define
Ptdescribe some special steps
Pof
t
ν(t) = τ =1 J (τ ), and νB (t) = τ =1 J (τ )I(B(τ ) = B)
(the number of times configuration B is sampled till time t).
1) f (t) (B) update when J (t) = 1: If J (t) = 1, all sensors
are read to obtain Z(t); this Z(t) is either supplied to all
sensors, or sent to some specific node which does centralized
computation (this is done only when J (t) = 1) and broadcasts
the results to all sensors. For each B ∈ B, all estimates
(k)
(k)
{X̂B (t)}1≤k≤N are computed using ZB (t), where X̂B (t)
denotes the estimate at node k at time t if B(t) P
= B. Then, the
N
quantity YK (t) := EX(t)∼p(·|θ(t),ZB (t),K (t) ) ( N1 k=1 ||X(t) −
B
(k)
(t)
B
X̂B (t)||2 |θ(t), ZB (t), KB ) is computed, and the following
update is done for all B ∈ B:
f (t+1) (B) = f (t) (B) + J (t)a(ν(t)) Y
(t)
(i, j) = KB (i, j) − b(νB (t))J (t)
(t)
KB
− f (t) (B)
Remark 15: Algorithm 5 is similar to Algorithm 4 except
that (i) consensus is used for deciding the estimates, and (ii)
an additional SPSA algorithm has been used to optimize the
consensus gains. However, this scheme does not achieve perfect consensus, and is optimal only when one round message
exchange among neighbouring nodes is allowed per slot.
(t)
Remark 16: Since KB iteration does not depend on λ(t),
we can run (6) at the same timescale as λ(t) iteration.
B. Performance of Algorithm 5
1) Complexity and distributed nature: The mean number
of additional sensors activated per slotis O( N
T ), which can
be made small by taking T large enough. The same argument
applies to the computation of YK (t) . Moreover, one can deB
cide only to compute YK (t) (t) and update f (t) (B(t)) when
B(t)
(5)
We will see in Section V-B1 that this O(2N ) computation can
be avoided without sacrificing convergence.
(t)
2) KB update when J (t) = 1: The algorithm requires
P∞ b2 (t)
another condition
t=0 d2 (t) < ∞ (apart from those in
(t)
(6) is a gradient descent scheme, with the goal to converge
∗
to some KB
so that limt→∞ E(YK (t) (t)) (if it exists) is
B
minimized. ∇E(YK (t) (t)) is estimated by SPSA.
B
3) Outline of the proposed algorithm: The entire scheme
is described in Algorithm 5.
Algorithm 5: The same seed is supplied to all sensors for
Gibbs sampling. All iterates are initialized arbitrarily.
For any time t = 0, 1, 2, · · · , do at each sensor k:
1. Run Gibbs sampling as in Section IV-A2, once or
multiple times. All sensors will have the same B(t); they
will be activated according to B(t) and make observations.
(k)
(t)
Compute X̄ (k) (t) = µ2 (Hc (t)) and X̂(t) = KB(t) X̄(t)
locally. Update λ(t) using (3).
2. If J (t) = 1, read all sensors and obtain Z(t). For
(k)
each B ∈ B, compute all estimates {X̂B (t)}1≤k≤N using
(k)
ZB (t), where X̂B (t) denotes the estimate at node k at
(t)
time t if B(t) = B. Update f (t) (B) using (5), and KB
using (6) for all B ∈ B.
3. Do the same θ(t) update as in (2).
(t+1)
4. Broadcast KB(t) , f (t) (B) and θ(t + 1) if J (t) = 1.
Section IV-A1) to ensure convergence of the KB iterates
which are updated via the following SPSA algorithm whenever
J (t) = 1 (i.e., t is an integer multiple of T ).
A random matrix Γt ∈ RN ×N is generated such that
Γt (i, j) = 0 if i and j are not neighbours, otherwise
Γt (i, j) ∈ {−1, 1} independently with equal probability. Now,
J (t) = 1 as discussed in Section IV-B2; the algorithm will
converge slowly but O(2N ) computation will be avoided.
Gibbs sampling is run at all nodes; they will yield the same
B(t) since all sensors have the same seed. All sensors will
have the same λ(t), and can update h(t) (B(t)) for each t. The
consensus gains and f (t) (B) updates need to be sent to all
sensors if J (t) = 1; however, a bounded delay in broadcast
does not affect convergence. Since nodes use only local
consensus and periodic broadcast, and Gibbs sampling step
is distributed, Algorithm 5 is distributed.
2) Convergence of Algorithm 5:
Assumption 6: For given λ ∈ [0, A0 ], θ ∈ Θ, for all B ∈
B, the function E(YKB ) of KB (Section V-A1) is Lipschitz
continuous in KB . The set of ordinary differential equations
K̇B (τ ) = −∇E(YKB (τ ) (τ )) (vectorized) for any fixed θ ∈ Θ
∗
has a globally asymptotically stable equilibrium KB
(θ) which
is Lipschitz continuous in θ.
Now we present the main result related to Algorithm 5,
which shows that the iterates of Algorithm 5 almost surely
converge to a set of locally optimal solutions of (P2).
Theorem 6: Under Assumptions 2-6 and Algorithm 5,
we have limt→∞ θ(t)
=
θ∗ almost surely.
Correspondingly,
λ(t)
→
λ∗ (θ∗ )
almost
surely. As a result, for all B ∈ B, we have
∗
KB (t) → KB
(θ∗ ) almost surely. Also, f (t) (B) →
PN
1
−
k=1 EZB (t)∼p(·|θ0 ) Eµ(k) (·;·;θ ∗ ),K ∗ (θ ∗ ),θ ∗ (||X(τ )
N
2
to combine the estimates made by its neighbouring sensors.
Kalman filtering operation is required since the dynamical
system can be expressed as a linear stochastic system (for
which Kalman filter is the best linear MMSE estimator):
X(t + 1)
=
AX(t) + (X(t + 1) − AX(t))
{z
}
|
zk (t + 1)
=
[mk,1 , mk,2 , · · · , mk,|X | ] X(t) +
{z
}
|
Zk (t)
=
Bk (t)zk (t)
:=w(t)
:=Hk
vk (t)
| {z }
∼N (mk,i ,Σk,i ) if X(t)=ei
Since A is known, the conditional covariance matrix Q(i) of
w(t) given X(t) = ei is known to all sensors.
B
X̂ (k) (τ )||2 |ZB (t)) := fθ∗ (B) almost surely for all B ∈ B.
The B(t) process reaches the steady-state distribution
πβ,λ∗ (θ∗ ),KB∗ (θ∗ ),θ∗ (·) which can be obtained by replacing
h(B) in (1) by fθ∗ (B) + λ∗ ||B||1 .
Proof: The proof is similar to that of Theorem 5.
If θ∗ = θ0 , then global optimum for (P2) is reached.
VI. D ISTRIBUTED TRACKING OF A M ARKOV CHAIN
In this section, we seek to track a Markov chain {X(t)}t≥0
with transition probability matrix AT and finite state space X .
In order to have a meaningful problem, we enumerate each
state and denote the i-th state by ei which is an |X |-length
0 − 1 column vector with 1 only at the i-th location. Thus,
the state space becomes X = {ei : 1 ≤ i ≤ |X |}. We also
consider a measurement process where, given X(t) = ei and
a configuration B(t) = B of active sensors, any sensor k
with Bk = 1 makes an observation zk (t) ∼ N (mk,i , Σk,i )
where the mean mk,i and covariance matrix Σk,i depend on
the state ei . Under this model, a centralized, finite horizon
version of the dynamic sensor subset selection problem has
been solved in [4], where it is shown that a sufficient statistic
for decision making is a belief vector on the state space
conditioned on the history. The authors of [4] formulated a
partially observable Markov decision process (POMDP) with
this belief vector working as a proxy to the (hidden) state,
and also proposed a Kalman-like estimator to make on-line
update to the belief vector using new observations made by
chosen sensors. Hence, in this section, we skip the centralized
problem (P1) and directly solve the distributed problem (P2).
The centralized problem in [4] itself leads to intractability in
the sequential subset selection problem; the POMDP formulation in [4] does not provide any structural result on the optimal
policy. Hence, for solving the distributed problem (P2), we
restrict ourselves to the class of myopic policies which seek
to minimize the cost at the current time instant.
The estimation scheme yields X̂ (k) (t) at each node k;
X̂ (k) (t) is the belief vector on the state space, and is generated
by the use of a Kalman-consensus filter (KCF) as in [15].
Consensus is required since all nodes do not have access to
the complete observation set; consensus requires each sensor
A. Kalman consensus filtering
The KCF that we use is adapted from [15] with the
additional consideration of B(t). Here each sensor k maintains
an estimate X̄ (k) (t) at time t before making any observation.
Once the observations are made, sensor k computes X̂ (k) (t)
by using KCF. The estimates evolve as follows:
X̂ (k) (t)
=
[X̄(t) + Kk (t)Bk (t)(zk (t) − Hk X̄ (k) (t))
X
+Ck
(X̄ (j) (t) − X̄ (k) (t))]D
X̄ (k) (t + 1)
=
AX̂ (k) (t)
j∈nbr(k)
(7)
Here Ck and Kk (t) are called consensus gain and Kalman
gain matrices for sensor k, and nbr(k) is the set of neighbours
of sensor k. Projection on the probability simple D is done to
ensure that X̂ (k) (t) is a valid probability belief vector.
In [15, Theorem 1], the Kalman gains Kk (t) at all nodes
areP
optimized for given consensus gains Ck , so that the MSE
N
1
(k)
(t) − X(t)||2 at the current time step is mink=1 E||X̂
N
imized; but the computational and communication complexity
per node for its implementation grows rapidly with N . Hence,
[15, Section IV] also provides an alternative suboptimal KCF
algorithm to compute Kk (t) at each sensor k, which has low
complexity and is easily implementable. Hence, we adapt this
suboptimal algorithm from [15, Section IV] to our problem.
The KCF gain update scheme from [15, Section IV] maintains two matrices Mk (t) and Pk (t) for each k, which are
viewed as proxies for the covariance matrices of the two
errors X̂ (k) (t) − X(t) and X̄ (k) (t) − X(t). It also requires
system noise covariance matrix; since system noise w(t) is
dependent on X(t) whose
PNexact value is unknown to the
sensors, we use Qk (t) = i=1 X̄ (k) (t + 1)(i)Q(i) at node k
as an estimate of the covariance matrix of w(t + 1). Similarly,
P|X |
Rk (t + 1) = Bk (t) i=1 X̄ (k) (t + 1)(i)Σk,i is used as an
alternative to the covariance matrix vk (t + 1). The KCF filter
also maintains an abstract iterate Fk (t). The overall KCF gain
update equations from [15, Section IV] are as follows:
Kk (t)
=
Rk (t + 1)
=
Qk (t)
=
small enough by making T arbitrarily large but finite, and the
broadcast can even be done over T slots to avoid network
congestion at any particular slot. Interestingly, Mk (t) and
Pk (t) can be updated using local iterates, and do not need any
communication. Computing Mk (t) and Pk (t) involves simple
matrix operations which have polynomial complexity.
Bk (t)Pk (t)HkT (Rk (t) + Hk Pk (t)HkT )−1
|X |
X
Bk (t)
X̄ (k) (t + 1)(i)Σk,i
i=1
N
X
X̄ (k) (t + 1)(i)Q(i)
i=1
Fk (t)
=
I − Kk (t)Hk Bk (t)
Mk (t)
=
Fk (t)Pk (t)FkT (t) + Kk (t)Rk (t + 1)KkT (t)
Pk (t + 1)
=
AMk (t)AT + Qk (t)
VII. N UMERICAL R ESULTS
(8)
B. The proposed algorithm
The sensor subset selection is done via Gibbs sampling
run at all nodes supplied with the same seed; all nodes
generate the same configuration B(t) at each time t. The
quantity f (t) (B) is updated via stochastic approximation to
converge to the MSE under configuration B; but since the
MSE
PNunder B cannot be computed directly, sensor k uses
1
k=1 T r(Mk (t − T )) of a past slot (t − T ) to update
N
f (t) (B). λ(t) is varied at a slower timescale.
The proposed scheme is provided in Algorithm 6.
Algorithm 6: Input: A0 , T > 0 stepsize sequences
{a(t)}t≥0 and {b(t)}t≥0 as in Section IV, consensus gain
matrices Ck for all k ∈ {1, 2, · · · , N }, the same seed
for randomization to all sensors, initial covariance matrix
P (0) of X(0).
Pt All iterates are initialized arbitrarily. Define
νB (t) := τ =0 I(B(τ ) = B)J (t).
For any time t = 0, 1, 2, · · · , do at each sensor k:
1. Select B(t) at each sensor by running the Gibbs
sampling step as in Section IV-A2. Activate sensors autonomously according to the common B(t) selected by
them, and make observations accordingly.
2. At sensor k, perform state estimation with (7) and gain
update with (8). Update λ(t) at all sensors using (3).
Compute h(t+1) (B) = f (t+1) (B) + λ(t)||B||1 either for
all B ∈ B, or for B(t) if computation is a bottleneck.
3. If J (t) = 1, update:
f (t+1) (B)
=
[f (t) (B) + J (t)I(B(t − T ) = B)a(νB (t − T ))
(
N
1 X
T r(Mk (t − T )) − f (t) (B))]0A0
N k=1
4. If J (t) = 1, broadcast T r(Mk (t)) to all sensors, so that
step 3 can be performed at time (t + T ).
Remark 17: Algorithm 6 is suboptimal for (P2) because: (i)
it greedily chooses B(t) via Gibbs sampling without caring
about the future cost, and (ii) KCF update is suboptimal. But
it has low complexity, and it performs well (see Section VII).
C. Complexity of Algorithm 6
At each time t, a sensor k needs to obtain X̄ (j) (t)
from all its neighbours j ∈ nbr(k) for consensus. Also,
{T r(Mk (t))}1≤k≤N needs to be broadcast to all nodes when
J (t) = 1 so that f (t+T ) (B) update can be done at each node;
but the per slot communication for this broadcast can be made
Now we demonstrate the performance of Algorithm 4
(centralized) and Algorithm 6 (distributed). We consider the
1
,
following parameter values: N = 5, N̄ = 2, a(t) = t0.6
2
1
0.2
b(t) = t0.8 , c(t) = t , d(t) = t0.1 , T = 20, λ(0) = 0.1,
β = 150. Gibbs sampling is run 10 times per slot.
1) Performance of Algorithm 4: For illustration purpose,
we assume that X(t) ∼ N (θ0 , (1 − θ0 )2 ) scalar, and zk (t) =
X(t) + wk (t), where θ0 = 0.5 and wk (t) is zero mean i.i.d.
Gaussian noise independent across k. Standard deviation of
wk (t) is chosen uniformly and independently from the interval
[0, 0.5], for each k ∈ {1, 2, · · · , N }. Initial estimate θ(0) =
0.2, Θ = [0, 0.8].
We consider three possible algorithms: (i) Algorithm 4 in
its basic form, which we call GIBBS, (ii) a variation of
Algorithm 4 called LOWCOMPLEXGIBBS where all sensors
are not read when J (t) = 1, and the relatively expensive
f (t) (B(t)) and θ(t) updates are done every T slots, and (iii)
an algorithm GREEDY where N̄ sensors are picked arbitrarily
and used for ever with the wrong estimate θ(0) = 0.2 without
any update. The MSE per slot, mean number of active sensors
per slot, λ(t) and θ(t) are plotted against t in Figure 1.
MSE of all these three algorithms are much smaller than
V ar(X(t)) = (1 − θ0 )2 (this is MMSE without any observation). We notice that GIBBS and LOWCOMPLEXGIBBS
significantly outperform GREEDY in terms of time-average
MSE; this shows the power of Gibbs sampling and learning
θ(t) over time. We have plotted only one sample path since
GIBBS and LOWCOMPLEXGIBBS converge almost surely
to the global optimum. We also observe that GIBBS converges faster than LOWCOMPLEXGIBBS, since it uses more
computational
and communication resources. We observe that
Pt
1
||B(τ
)||
1 → N and θ(t) → θ0 almost surely for
τ =1
t
both GIBBS and LOWCOMPLEXGIBBS (verified by simulating multiple sample paths). It is interesting to note that
θ∗ = θ1 = θ0 in this numerical example (recall Theorem 5 and
Remark 14), i.e., both GIBBS and LOWCOMPLEXGIBBS
converge to the true parameter. Convergence rate will vary
with stepsize and other parameters.
2) Performance of Algorithm 6: We consider |X | = 4
states, and assume that the sensors form a line topology.
Transition probability matrix AT is chosen randomly, and
Ck = 0.1 is set in Algorithm 6 (which we call GIBBSKCF
here) for all k. The values mk,i are chosen uniformly from
[0, 1], and Σk,i = 0.05 ∗ (1 + |k − i|) (scalar values in this
case) are set for all 1 ≤ i ≤ |X |, 1 ≤ k ≤ N .
We compared the MSE performance of three algorithms:
(i) GIBBSKCF, i.e., Algorithm 6, (ii) CENTRALKALMAN,
0.15
0.1
0.09
GIBBS
GREEDY
LOWCOMPLEXGIBBS
0.1
0.08
0.07
(t)
MSE
0.06
GIBBS
0.05
0.04
0.05
0.03
0.02
0.01
0
0
2000
4000
6000
8000
0
0
10000
1
2
t
3
4
2.5
5
104
t
0.7
0.5
2
0.4
(t)
Mean number of
active sensors per slot
0.6
GIBBS
LOWCOMPLEXGIBBS
1.5
0.3
GIBBS
LOWCOMPLEXGIBBS
0.2
0.1
1
0
2000
4000
6000
8000
10000
0
0
1
2
Fig. 1.
3
t
t
4
5
104
Performance of Algorithm 4 for centralized tracking of the iid process.
where a centralized Kalman filter tracks X(t) using observations only from two arbitrary sensors, and (iii) PERFECTBLIND, where at each time t, the state X(t − 1) is known
perfectly to all sensors, but no observation is allowed to
compute X̂ (k) (t). The MSE of PERFECTBLIND will be
limt→∞ E(V ar(X(t)|X(t − 1))).
In Figure 2, we plot the time-average P
MSE for all three
t
algorithms along one sample path, and 1t τ =1 ||B(τ )||1 for
Pt
1
GIBBSKCF. We observe that, t τ =1 ||B(τ )||1 converges to
N̄ . For the given sample path, GIBBSKCF provides better
MSE than PERFECTBLIND, but the MSE was seen to be
slightly worse along some other sample paths. We also observe
that the MSE of GIBBSKCF is slightly worse than CENTRALKALMAN for the given instance, but they are very close
in many other problem instances (verified numerically); this
establishes the efficacy of GIBBSKCF and the power of Gibbs
sampling based sensor subset selection, despite using only
one round of consensus per slot. Basically, dynamic subset
selection compensates for the performance loss due to lack of
a fusion center.
VIII. C ONCLUSIONS
We have proposed low-complexity centralized and distributed learning algorithms for dynamic sensor subset selection for tracking i.i.d. time-varying as well as Markovian processes. We first provided algorithms based on Gibbs
sampling and stochastic approximation for i.i.d. time-varying
data with unknown parametric distribution, and proved almost
sure convergence. Next, we provided an algorthm based on
Kalman consensus filtering, Gibbs sampling and stochastic
approximation for distributed tracking of a Markov chain.
Numerical results demonstrate the efficacy of the algorithms
against simple algorithms without learning.
R EFERENCES
[1] A. Chattopadhyay and U. Mitra. Optimal sensing and data estimation in a large sensor network. Technical Report, available in
https://arxiv.org/abs/1707.08074, a shorter version has been accepted
in IEEE Globecom 2017, 2017.
[2] A. Chattopadhyay and U. Mitra.
Optimal dynamic
sensor
subset
selection
for
tracking
a
time-varying
stochastic process.
submitted to IEEE INFOCOM 2018,
https://drive.google.com/open?id=0B458nrWAgTUEbjd4R3Z5NS1wSlk.
[3] D. Wang, J. Fisher III, and Q. Liu. Efficient observation selection
in probabilistic graphical models using bayesian lower bounds. In
0.9
0.8
0.7
MSE
0.6
0.5
0.4
0.3
0.2
0.1
0
0
PERFECTBLIND
CENTRALKALMAN
GIBBSKCF
2000
4000
6000
8000
10000
8000
10000
t
2.1
Mean number of
active sensors per slot
2
1.9
GIBBSKCF
1.8
1.7
[13] S. Das and J.M.F. Moura. Consensus+innovations distributed kalman
filter with optimized gains. IEEE Transactions on Signal Processing,
65(2):467–481, 2017.
[14] S. Kar and J.M.F. Moura. Gossip and distributed kalman filtering:
Weak consensus under weak detectability. IEEE Transactions on Signal
Processing, 59(4):1766–1784, 2011.
[15] R. Olfati-Saber. Kalman-consensus filter : Optimality, stability, and
performance. In Conference on Decision and Control, pages 7036–7042.
IEEE, 2009.
[16] L. Xiao, S. Boyd, and S. Lall. A scheme for robust distributed sensor
fusion based on average consensus. In International Symposium on
Information Processing in Sensor Networks (IPSN), pages 63–70. IEEE,
2005.
[17] V. Tzoumas, A. Jadbabaie, and G.J. Pappas. Near-optimal sensor
scheduling for batch state estimation: Complexity, algorithms, and limits.
In Conference on Decision and Control, pages 2695–2702. IEEE, 2016.
[18] N. Michelusi and U. Mitra. Cross-layer estimation and control for
cognitive radio: Exploiting sparse network dynamics. IEEE Transactions
on Cognitive Communications and Networking, 1(1):128–145, 2015.
[19] A. Chattopadhyay and B. Baszczyszyn. Gibbsian on-line distributed
content caching strategy for cellular networks. https:// arxiv.org/ abs/
1610.02318, 2016.
[20] J.C. Spall. Multivariate stochastic approximation using a simultaneous
perturbation gradient approximation. IEEE Transactions on Automatic
Control, 37(3):332–341, 1992.
[21] A. Chattopadhyay, M. Coupechoux, and A. Kumar. Sequential decision algorithms for measurement-based impromptu deployment of a
wireless relay network along a line. IEEE/ACM Transactions on Networking, longer version available in http:// arxiv.org/ abs/ 1502.06878,
24(5):2954–2968, 2016.
1.6
1.5
1.4
0
Fig. 2.
chain.
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
2000
4000
6000
Performance of Algorithm 6 for distributed tracking of a Markov
Proceedings of the Thirty-Second Conference on Uncertainty in Artificial
Intelligence (UAI), pages 755–764. ACM, 2016.
D.S. Zois, M. Levorato, and U. Mitra. Active classification for pomdps:
A kalman-like state estimator. IEEE Transactions on Signal Processing,
62(23):6209–6224, 2014.
D.S. Zois, M. Levorato, and U. Mitra. Energy-efficient, heterogeneous
sensor selection for physical activity detection in wireless body area
networks. IEEE Transactions on Signal Processing, 61(7):1581–1594,
2013.
V. Krishnamurthy and D.V. Djonin. Structured threshold policies for
dynamic sensor schedulinga partially observed markov decision process
approach. IEEE Transactions on Signal Processing, 55(10):4938–4957,
2007.
W. Wu and A. Arapostathis. Optimal sensor querying: General markovian and lqg models with controlled observations. IEEE Transactions
on Automatic Control, 53(6):1392–1405, 2008.
V. Gupta, T.H. Chung, B. Hassibi, and R.M. Murray. On a stochastic
sensor selection algorithm with applications in sensor scheduling and
sensor coverage. Automatica, 42:251–260, 2006.
A. Bertrand and M. Moonen. Efficient sensor subset selection and link
failure response for linear mmse signal estimation in wireless sensor
networks. In European Signal Processing Conference (EUSIPCO), pages
1092–1096. EURASIP, 2010.
P. Bremaud. Markov Chains, Gibbs Fields, Monte Carlo Simulation,
and Queues. Springer, 1999.
Vivek S. Borkar. Stochastic approximation: a dynamical systems
viewpoint. Cambridge University Press, 2008.
F. Schnitzler, J.Y. Yu, and S. Mannor. Sensor selection for crowdsensing
dynamical systems. In International Conference on Artificial Intelligence
and Statistics (AISTATS), pages 829–837, 2015.
Arpan Chattopadhyay obtained his B.E. in Electronics and Telecommunication Engineering from
Jadavpur University, Kolkata, India in the year 2008,
and M.E. and Ph.D in Telecommunication Engineering from Indian Institute of Science, Bangalore,
India in the year 2010 and 2015, respectively. He
is currently working in the Ming Hsieh Department
of Electrical Engineering, University of Southern
California, Los Angeles as a postdoctoral researcher.
Previously he worked as a postdoc in INRIA/ENS
Paris. His research interests include wireless networks, cyber-physical systems, machine learning and control.
Urbashi Mitra is a Deans Professor in the Departments of Electrical Engineering and Computer
Science, University of Southern California, Los
Angeles, CA, USA. Previous appointments inPLACE
clude Bellcore and the Ohio State University. Her
PHOTO
honors include: U.S.-U.K. Fulbright, Leverhulme
HERE
Trust Visiting Professorship, Editor-in-Chief IEEE
TRANSACTIONS ON MOLECULAR, BIOLOGICAL AND MULTISCALE COMMUNICATIONS,
IEEE Communications Society Distinguished Lecturer, U.S. NAE Galbreth Lectureship, Okawa Foundation Award, and an NSF CAREER Award. Her research is in wireless
communications.
A PPENDIX A
P ROOF OF T HEOREM 1
We will prove only the first part of the theorem where there
exists one B ∗ . The second part of the theorem can be proved
similarly.
Let us denote the optimizer for (P1) by B, which is possibly
different from B ∗ . Then, by the definition of B ∗ , we have
f (B ∗ ) + λ∗ ||B ∗ ||1 ≤ f (B) + λ∗ ||B||1 . But ||B||1 ≤ K
(since B is a feasible solution to the constrained problem)
and ||B ∗ ||1 = K (by assumption). Hence, f (B ∗ ) ≤ f (B).
This completes the proof.
Hence,
∞
X
(1 − δ(Pl ))
l=0
=
≥
≥
=
A PPENDIX B
W EAK AND S TRONG E RGODICITY
Consider a discrete-time Markov chain (possibly not timehomogeneous) {B(t)}t≥0 with transition probability matrix
(t.p.m.) P (m; n) between t = m and t = n. We denote
by D the collection of all possible probasbility distributions
on the state space. Let dV (·, ·) denote the total variation
distance between two distributions in D. Then {B(t)}t≥0
is called weakly ergodic if, for all m ≥ 0, we have
limn↑∞ supµ,ν∈D dV (µP (m; n), νP (m; n)) = 0.
The Markov chain {B(t)}t≥0 is called strongly
ergodic if there exists π
∈
D such that,
limn↑∞ supµ∈D dV (µT P (m; n), π) = 0 for all m ≥ 0.
A PPENDIX C
P ROOF OF T HEOREM 3
We will first show that the Markov chain {B(t)}t≥0 in
weakly ergodic.
Let us define ∆ := maxB∈B,A∈B |h(B) − h(A)|.
Consider the transition probability matrix (t.p.m.) Pl for
the inhomogeneous Markov chain {Y (l)}l≥0 (where Y (l) :=
B(lN )). The Dobrushin’s ergodic coefficient δ(Pl ) is given
by (see [10, Chapter
6, Section 7] for definition) δ(Pl ) =
P
0
00
1 − inf B 0 ,B 00 ∈B B∈B min{Pl (B , B), Pl (B , B)}. A sufficient condition
P∞for the Markov chain {B(t)}t≥0 to be weakly
ergodic is
l=1 (1 − δ(Pl )) = ∞ (by [10, Chapter 6,
Theorem 8.2]).
Now, with positive probability, activation states for all nodes
0
are updated over a period of N slots. Hence, Pl (B , B) > 0 for
0
all B , B ∈ B. Also, once a node jt for t = lN + k is chosen
in MODIFIEDGIBBS algorithm, the sampling probability for
−β(lN +k)∆
. Hence,
any activation state in a slot is greater than e
2
for independent sampling over N slots, we have, for all pairs
0
B , B:
N
−1 −β(lN +k)∆
Y
0
e
Pl (B , B) >
>0
2N
k=0
≥
∞
X
0
l=0
∞
X
X
inf00
B ,B ∈B
2N
0
00
min{Pl (B , B), Pl (B , B)}
B∈B
N
−1
−β(0) log(1+lN +k)×∆
Y
e
2N
l=0
k=0
∞ N
−1 −β(0) log(1+lN +N )×∆
X
Y
e
l=0 k=0
∞
X
1
NN
l=1
1
N N +1
N
1
(1 + lN )β(0)N ∆
∞
X
1
β(0)N ∆
(1
+
i)
i=N +1
= ∞
(9)
Here the first inequality uses the fact that the cardinality of
B is 2N . The second inequality follows from replacing k by
N in the numerator. The thirdP
inequality follows from lowerlN +N −1
1
bounding (1+lN )1β(0)N ∆ by N1 i=lN
β(0)N ∆ . The last
P∞ (1+i)
equality follows from the fact that i=1 i1a diverges for 0 <
a < 1.
Hence, the Markov chain {B(t)}t≥0 is weakly ergodic.
In order to prove strong ergodicity of {B(t)}t≥0 , we
invoke [10, Chapter 6, Theorem 8.3]. We denote the t.p.m.
of {B(t)}t≥0 at a specific time t = T0 by Q(T0 ) , which is
a given specific matrix. If {B(t)}t≥0 evolves up to infinite
time with fixed t.p.m. Q(T0 ) , then it will reach the stationary
−βT h(B)
distribution πβT0 (B) = e Zβ0
. Hence, we can claim that
T0
Condition 8.9 of [10, Chapter 6, Theorem 8.3] is satisfied.
Next, we check Condition 8.10 of [10, Chapter 6, Theo0
rem 8.3]. For any B ∈ arg minB 0 ∈B h(B ), we can argue
that πβT0 (B) increases with T0 for sufficiently large T0 ; this
can be verified by considering the derivative of πβ (B) w.r.t.
0
β. For B ∈
/ arg minB 0 ∈B h(B ), the probability πβT0 (B)
decreases with T0 for large T0 . Now, using the fact that
any
bounded sequence converges, we can write
P∞ monotone,
P
T0 =0
B∈B |πβT0 +1 (B) − πβT0 (B)| < ∞.
Hence, by [10, Chapter 6, Theorem 8.3], the Markov chain
{B(t)}t≥0 is strongly ergodic. It is straightforward to verify
the claim regarding the limiting distribution.
A PPENDIX D
P ROOF OF L EMMA 1
Let λ1 > λ2 > 0, and the corresponding optimal error and
mean number of active sensors under these multiplier values
be (f1 , n1 ) and (f2 , n2 ), respectively. Then, by definition, f1 +
λ1 n1 ≤ f2 + λ1 n2 and f2 + λ2 n2 ≤ f1 + λ2 n1 . Adding these
two inequalities, we obtain λ1 n1 + λ2 n2 ≤ λ1 n2 + λ2 n1 ,
i.e., (λ1 − λ2 )n1 ≤ (λ1 − λ2 )n2 . Since λ1 > λ2 , we obtain
n1 ≤ n2 . This completes the first part of the proof. The second
part of the proof follows using similar arguments.
1) Convergence in the fastest timescale: Let us denote the
probability distribution of B(t) under Algorithm 4 by π (t) (a
column vector indexed by the cofigurations from B), and the
A PPENDIX E
corresponding
transition probability matrix (TPM) by A(t);
P ROOF OF L EMMA 2
(t+1) T
i.e.,
(π
)
= (π (t) )T A(t) = (1 − 1) × (π (t) )T + 1 ×
P
−βh(B)
(t) T
B∈B ||B||1 e
Let us denote Eµ2 ||B||1 =: g(λ) =
. (π ) A(t). This form is similar to a standard stochastic
Zβ
It is straightforward to see that Eµ2 ||B||1 is continuously approximation scheme as in [11, Chapter 2] except that the
differentiable in λ. Let us denote Zβ by Z for simplicity. The step size sequence for π (t) iteration is a constant sequence.
derivative of g(λ) w.r.t. λ is given by:
Also, if f (t) (B), λ(t) and θ(t) are constant with time t, then
0
A(t) = A will also be constant with time t, and the stationary
g (λ)
P
−β(f (B)+λ||B||1 ) − P
−β(f (B)+λ||B||1 ) dZ
−Zβ B∈B ||B||2
distribution for the TPM A will exist and will be Lipschitz
B∈B ||B||1 e
1e
dλ
=
Z2
continuous in all (constant) slower timescale iterates. Hence,
by using similar argument as in [11, Chapter 6, Lemma 1],
Now, it is straightforward to verify that dZ
dλ = −βZg(λ).
Hence,
one can show the following for all B ∈ B:
0
g (λ)
−Zβ
=
P
−β(f (B)+λ||B||1 ) βZg(λ)
2 −β(f (B)+λ||B||1 ) + P
B∈B ||B||1 e
B∈B ||B||1 e
Z2
Now, g 0 (λ) ≤ 0 is equivalent to
P
||B||21 e−β(f (B)+λ||B||1 )
g(λ) ≤ PB∈B
−β(f (B)+λ||B||1 )
B∈B ||B||1 e
Noting that E||B||1 =: g(λ) and dividing the numerator and
denominator of R.H.S. by Z, the condition is reduced to
E||B||2
E||B||1 ≤ E||B||11 , which is true since E||B||21 ≥ (E||B||1 )2 .
Hence, E||B||1 is decreasing in λ for any β > 0. Also, it
is easy to verify that |g 0 (λ)| ≤ (β + 1)N 2 . Hence, g(λ) is
Lipschitz continuous in λ.
A PPENDIX F
P ROOF OF T HEOREM 4
Let the distribution of B(t) under Algorithm 3
be π (t) (·). Since limt→∞ a(t) = 0, it follows that
limt→∞ dV (π (t−1) , πβ|λ(t−1) ) = 0 (where dV (·, ·) is
the total variation distance), and limt→∞ (Eπ(t−1) ||B||1 −
Eπβ |λ(t−1) ||B||1 ) := limt→∞ e(t) = 0. Now, we can rewrite
the λ(t) update equation as follows:
λ(t + 1) = [λ(t) + a(t)(Eπβ |λ(t−1) ||B||1 − N̄ + Mt + et )]cb (10)
Here Mt := ||B(t − 1)||1 − Eπ(t−1) ||B(t − 1)||1 is a
Martingale difference noise sequence, and limt→∞ et = 0. It is
easy to see that the derivative of Eπβ |λ ||B||1 w.r.t. λ is bouned
for λ ∈ [b, c]; hence, Eπβ |λ ||B||1 is a Lipschitz continuous
function of λ. It is also easy to see that the sequence {Mt }t≥0
is bounded. Hence, by the theory presented in [11, Chapter 2]
and [11, Chapter 5, Section 5.4], λ(t) converges to the unique
zero of Eπβ |λ ||B||1 − N̄ almost surely. Hence, λ(t) → λ∗
almost surely. Since limt→∞ dV (π (t−1) , πβ|λ(t−1) ) = 0 and
πβ|λ is continuous in λ, the limiting distribution of B(t)
becomes πβ|λ∗ .
A PPENDIX G
P ROOF OF T HEOREM 5
The proof involves several steps, and the brief outline of
these steps are provided one by one.
lim |π (t) (B) − πβ,f (t) ,λ(t),θ(t) (B)| = 0 a.s.
t→∞
(11)
2) Convergence of iteration (4): Note that, (4) depends on
θ(t) and not on B(t) and λ(t); the iteration (4) depends on
θ(t) through the estimation function µ2 (·; ·; ·). Now, f (t) (B)
is updated at a faster timescale compared to θ(t). Let us
consider the iterations (4) and (2); they constitute a twotimescale stochastic approximation.
Note that, for a given θ, the iteration (4) remains bounded
inside a compact set independent of θ; hence, using [11,
Chapter 2, Theorem 2] with additional modification as suggested in [11, Chapter 5, Section 5.4] for projected stochastic
approximation, we can claim that limt→∞ f (t) (B) → fθ (B)
almost surely for all B ∈ B, if θ(t) is kept fixed at a value
θ. Also, since µ2 (·; ·; θ) is Lipschitz continuous in θ, we can
claim that fθ (B) is Lipschitz continuous in θ for all B ∈ B.
c(ν(t))
= 0.
We also have limt→∞ a(ν(t))
Hence, by using an analysis similar to that in [21, Appendix E, Section C.2] (which uses [11, Chapter 6, Lemma 1]),
one can claim that:
lim |f (t) (B) − fθ(t) (B)| = 0 a.s. ∀B ∈ B
t→∞
(12)
This proves the desired convergence of the iteration (4).
3) Convergence of λ(t) iteration: The λ(t) iteration will
view θ(t) as quasi-static and B(t), f (t) (·) iterations as equilibriated.
Let us assume that θ(t) is kept fixed at θ. Then, by (11)
and (12), we can work with πβ,fθ ,λ(t) ,θ in this timescale.
Under this situation, (3) P
asymptotically tracks the iteration
λ(t + 1) = [λ(t) + b(t)( B∈B πβ,fθ ,λ(t),θ (B)||B||1 − N̄ +
Mt )]A
0 where {Mt }t≥0 is a Martingale differenece sequence.
Now, πβ,fθ ,λ(t),θ (B) is Lipschitz continuous in θ and λ(t)
(using Assumption 2, Assumption 4 and a little algebra on
the expression (1)). If A0 is large enough, then, by the theory
of [11, Chapter 2, Theorem 2] and [11, Chapter 5, Section 5.4],
one can claim that λ(t) → λ∗ (θ) almost surely, and λ∗ (θ) is
Lipschitz continuous in θ (by Assumption 4).
Hence, by using similar analysis as in [21, Appendix E,
Section C.2] (which uses [11, Chapter 6, Lemma 1]), we can
say that, under iteration (3):
lim |λ(t) − λ∗ (θ(t))| = 0 a.s.
t→∞
(13)
4) Convergence of the θ(t) iteration: Note that, (2) is the
slowest timescale iteration and hence it will view all other
there iterations (at three different timescales) as equilibriated.
However, this iteration is not affected by other iterations.
Hence, this iteration is an example of simultaneous perturbation stochastic approximation as in [20] (since θ∗ lies inside
the interior of Θ), but with a projection operation applied on
the iterates. Hence, by combining [20, Proposition 1] and the
discussion in [11, Chapter 5, Section 5.4], we can say that
limt→∞ θ(t) = θ∗ almost surely.
5) Completing the proof: We have seen that limt→∞ θ(t) =
θ∗ almost surely. Hence, by (13), limt→∞ λ(t) = λ∗ (θ∗ )
almost surely. By (12), limt→∞ f (t) (B) = fθ∗ (B) almost
surely for all B ∈ B. Then, by (11), limt→∞ π (t) (B) =
πβ,fθ∗ ,λ∗ (θ∗ ),θ∗ (B) almost surely. Hence, Theorem 5 is
proved.
| 3cs.SY
|
Priv’IT: Private and Sample Efficient Identity Testing
Bryan Cai
EECS, MIT
[email protected]
Constantinos Daskalakis
EECS and CSAIL, MIT
[email protected]
Gautam Kamath
EECS and CSAIL, MIT
[email protected]
arXiv:1703.10127v3 [cs.DS] 7 Jun 2017
June 8, 2017
Abstract
We develop differentially private hypothesis testing methods for the small sample regime. Given a
sample D from a categorical distribution p over some domain Σ, an explicitly described distribution q
over Σ, some privacy parameter ε, accuracy parameter α, and requirements βI and βII for the type I and
type II errors of our test, the goal is to distinguish between p = q and dTV (p, q) ≥ α.
We provide theoretical bounds for the sample size |D| so that our method both satisfies (ε, 0)differential privacy, and guarantees βI and βII type I and type II errors. We show that differential
privacy may come for free in some regimes of parameters, and we always beat the sample complexity
resulting from running the χ2 -test with noisy counts, or standard approaches such as repetition for
endowing non-private χ2 -style statistics with differential privacy guarantees. We experimentally compare the sample complexity of our method to that of recently proposed methods for private hypothesis
testing [GLRV16, KR17].
1
Introduction
Hypothesis testing is the age-old problem of deciding whether observations from an unknown phenomenon
p conform to a model q. Often p can be viewed as a distribution over some alphabet Σ, and the goal is
to determine, using samples from p, whether it is equal to some model distribution q or not. This type of
test is the lifeblood of the scientific method and has received tremendous study in statistics since its very
beginnings. Naturally, the focus has been on minimizing the number of observations from the unknown
distribution p that are needed to determine, with confidence, whether p = q or p 6= q.
In several fields of research and application, however, samples may contain sensitive information about
individuals; consider for example, individuals participating in some clinical study of a disease that carries
social stigma. It may thus be crucial to guarantee that operating on the samples needed to test a statistical
hypothesis protects sensitive information about the samples. This is not at odds with the goal of hypothesis
testing itself, since the latter is about verifying a property of the population p from which the samples are
drawn, and not of the samples themselves.
Without care, however, sensitive information about the sample might actually be divulged by statistical
processing that is improperly designed. As recently exhibited, for example, it may be possible to determine
whether individuals participated in a study from data that would typically be published in genome-wide
association studies [HSR+ 08]. Motivated in part by this realization, there has been increased recent interest
in developing data sharing techniques which are private [JS13, USF13, YFSU14, SSB16].
Protecting privacy when computing on data has been extensively studied in several fields ranging from
statistics to diverse branches of computer science including algorithms, cryptography, database theory, and
machine learning; see, e.g., [Dal77, AW89, AA01, DN03, Dwo08, DR14] and their references. A notion of
privacy proposed by theoretical computer scientists which has found a lot of traction is that of differential
privacy [DMNS06]. Roughly speaking, it requires that the output of an algorithm on two neighboring
datasets D and D0 that differ in the value of one element be statistically close. For a formal definition see
Section 2.
Our goal in this paper is to develop tools for privately performing statistical hypothesis testing. In
particular, we are interested in studying the tradeoffs between statistical accuracy, power, significance, and
1
privacy in the sample size. To be precise, given samples from a categorical distribution p over some domain
Σ, an explicitly described distribution q over Σ, some privacy parameter ε, accuracy parameter α, and
requirements βI and βII for the type I and type II errors of our test, the goal is to distinguish between
p = q and dTV (p, q) ≥ α. We want that the output of our test be (ε, 0)-differentially private, and that the
probability we make a type I or type II error be βI and βII respectively. Treating these as hard constraints,
we want to minimize the number of samples that we draw from p.
Notice that the correctness constraint on our test pertains to whether we draw the right conclusion about
how p compares to q, while the privacy constraint pertains to whether we respect the privacy of the samples
that we draw from p. The pertinent question is how much the privacy constraint increases the number of
samples that are needed to guarantee correctness. Our main result is that privacy may come for free in
certain regimes of parameters, and has a mild cost for all regimes of parameters.
To be√ precise, without privacy constraints, it is well known that identity testing can be performed
from O( α2n · log β1 ) samples, where n is the size of Σ and β = min{βI , βII }, and that this is tight [BFF+ 01,
Pan08, VV14, ADK15]. Our main theoretical result is that, with privacy constraints, the number of samples
that are needed is
√
√
n
n
n1/3
,
,
Õ max
·
log(1/β)
.
(1)
α2 α3/2 ε α5/3 ε2/3
Our statistical test is provided in Section 5 where the above upper bound on the number of samples that
it requires
is proven as Theorem 3. Notice that privacy comes for free when the privacy requirement ε is
√
Ω( α) – for example when ε = 10% and the required statistical accuracy is 3%.
The precise constants sitting in the O(·) notation of Eq. (1) are given in the proof of Theorem 3. We
experimentally verify the sample efficiency of our tests by comparing them to recently proposed private
statistical tests [GLRV16, KR17], discussed in more detail shortly. Fixing a differential privacy and type I,
type II error constraints, we compare how many samples are required by our and their methods to distinguish
between hypotheses that are α = 0.1 apart in total variation distance. We find that different algorithms are
more efficient depending on the regime and properties desired by the analyst. Our experiments and further
discussion of the tradeoffs are presented in Section 6.
Approach. A standard approach to turn an algorithm differentially private is to use repetition. As already
mentioned above,√ absent differential privacy constraints, statistical tests have been provided that use an
optimal m = O( α2n · log β1 ) number of samples. A trivial way to get (ε, 0)-differential privacy using such a
non-private test is to create O(1/ε) datasets, each comprising m samples from p, and run the non-private
test on one of these datasets, chosen randomly. It is clear that changing the value of a single element in
the combined dataset may only affect the output of the test with probability at most ε. Thus the output is
(ε, 0)-differentially private; see Section√3 for a proof. The issue with this approach is that the total number
of samples that it draws is m/ε = O( εαn2 · log β1 ), which is higher than our target. See Corollary 1.
A different approach towards private hypothesis testing is to look deeper into the non-private tests and try
to “privatize” them. The most sample-efficient tests are variations of the classical χ2 -test. They compute
the number of times, Ni , that element i ∈ Σ appears in the sample and aggregate those counts using a
statistic that equals, or is close to, the χ2 -divergence between the empirical distribution defined by these
counts and the hypothesis distribution q. They accept q if the statistic is low and reject q if it is high, using
some threshold.
A reasonable approach to privatize such a test is to add noise, e.g. Laplace(1/ε) noise, to each count
Ni , before running the test. It is well known that adding Laplace(1/ε) noise to a set of counts makes them
differentially private, see Theorem 1. However, it also increases the variance of the statistic. This has been
noticed empirically in recent work of [GLRV16] for the χ2 -test. We show that the variance of the optimal χ2 style test statistic significantly increases
if we add Laplace noise to the counts, in Section 4.1, thus increasing
√
the sample complexity from O( n) to Ω(n3/4 ). So this route, too, seems problematic.
A last approach towards designing differentially private tests is to exploit the distance beween the null
and the alternative hypotheses. A correct test should accept the null with probability close to 1, and reject
an alternative that is α-far from the null with probability close to 1, but there are no requirements for
correctness when the alternative is very close to the null. We could thus try to interpolate smoothly between
2
datasets that we expect to see when sampling the null and datasets that we expect to see when sampling an
alternative that is far from the null. Rather than outputting “accept” or “reject” by merely thresholding our
statistic, we would like to tune the probability that we output “reject” based on the value of our statistic, and
make it so that the “reject” probability is ε-Lipschitz as a function of the dataset. Moreover, the probability
should be close to 0 on datasets that we expect to see under the null and close to 1 on datasets that we
expect to see under an √alternative that is α-far. As we show in Section 4.2, χ2 -style statistics have high
sensitivity, requiring ω( n) samples to be made appropriately Lipschitz.
While both the approach of adding noise to the counts, and that of turning the output of the test Lipschitz
fail in isolation, our test actually goes through by intricately combining these two approaches. It has two
steps:
1. A filtering step, whose goal is to “reject” when p is blatantly far from q. This step is performed by
comparing the counts Ni with their expectations under q, after having added Laplace(1/ε) noise to
these counts. If the noisy counts deviate from their expectation, taking into account the extra variance
introduced by the noise, then we can safely “reject.” Moreover, because noise was added, this step is
differentially private.
2. If the filtering step fails to reject, we perform a statistical step. This step just computes the χ2 -style
statistic from [ADK15], without adding noise to the counts. The crucial observation is that if the
filtering step does not reject, then the statistic is actually ε-Lipschitz with respect to the counts, and
thus the value of the statistic is still differentially private. We use the value of the statistic to determine
the bias of a coin that outputs “reject.”
Details of our test are given in Section 5.
Related Work. Identity testing is one of the most classical problems in statistics, where it is traditionally
called hypothesis or goodness-of-fit testing, see [Pea00, Fis35, RS81, Agr12] for some classical and contemporary references. In this field, the focus is often on asymptotic analysis, where the number of samples goes to
infinity, and we wish to get a grasp on their asymptotic distributions and error exponents [Agr12, TAW10].
In the past twenty years, this problem has enjoyed significant interest in the theoretical computer science
community (see, i.e., [BFF+ 01, Pan08, LRR13, VV14, ADK15, CDGR16, DK16, DDK16], and [Can15] for a
survey), where the focus has instead been on the finite sample regime, rather than asymptotics. Specifically,
the goal is to minimize the number of samples required, while still remaining computationally tractable.
A number of recent works [WLK15, GLRV16, KR17] (and a simultaneous work, focused on independence
testing [KSF17]) investigate differential privacy with the former set of goals. In particular, their algorithms
focus on fixing a desired significance (type I error) and privacy requirement, and study the asymptotic
distribution of the test statistics. On the other hand, we are the first work to apply differential privacy to
the latter line of inquiry, where our goal is to minimize the number of samples required to ensure the desired
significance, power and privacy. As a point of comparison between these two worlds, we provide an empirical
evaluation of our method versus their methods.
The problem of distribution estimation (rather than testing) has also recently been studied under the
lens of differential privacy [DHS15]. This is another classical statistics problem which has recently piqued
the interest of the theoretical computer science community. We note that the techniques required for this
setting are quite different from ours, as we must deal with issues that arise from very sparsely sampled data.
2
Preliminaries
In this paper, we will focus on discrete probability distributions over [n]. For a distribution p, we will use
the notation pi to denote the mass p places on symbol i.
Definition 1. The total variation distance between p and q is defined as
dTV (p, q) =
1 X
|pi − qi | .
2
i∈[n]
3
Definition 2. A randomized algorithm M with domain Nn is (ε, δ)-differentially private if for all S ⊆
Range(M ) and for all pairs of inputs D, D0 such that kD − D0 k1 ≤ 1:
Pr [M (D) ∈ S] ≤ eε Pr [M (D0 ) ∈ S] + δ.
If δ = 0, the guarantee is called pure differential privacy.
In the context of distribution testing, the neighboring dataset definition corresponds to two datasets
where one dataset is generated from the other by removing one sample. Up to a factor of 2, this is equivalent
to the alternative definition where one dataset is generated from the other by arbitrarily changing one sample.
Definition 3. An algorithm for the (α, βI , βII )-identity testing problem with respect to a (known) distribution
q takes m samples from an (unknown) distribution p and has the following guarantees:
• If p = q, then with probability at least 1 − βI it outputs “p = q;”
• If dTV (p, q) ≥ α, then with probability at least 1 − βII it outputs “p 6= q.”
In particular, βI and βII are the type I and type II errors of the test. Parameter α is the radius of distinguishing accuracy. Notice that, when p satisfies neither of cases above, the algorithm’s output may be
arbitrary.
We note that if an algorithm is to satisfy both these definitions, the latter condition (the correctness
property) need only be satisfied when p falls into one of the two cases, while the former condition (the
privacy property) must be satisfied for all realizations of the samples from p (and in particular, for p which
do not fall into the two cases above).
We recall the classical Laplace mechanism, which states that applying independent Laplace noise to a
set of counts is differentially private.
Theorem 1 (Theorem 3.6 of [DR14]). Given a set of counts N1 , . . . , Nn , the noised counts (N1 +Y1 , . . . , Nn +
Yn ) are (ε, 0)-differentially private when the Yi ’s are i.i.d. random variables drawn from Laplace(1/ε).
Finally, we recall the definition of zero-concentrated differential privacy from [BS16] and its relationship
to differential privacy.
Definition 4. A randomized algorithm M with domain Nn is ρ-zero-concentrated differentially private
(ρ-zCDP) if for all pairs of inputs D, D0 such that kD − D0 k1 ≤ 1 and all α ∈ (1, ∞):
Dα (M (D)||M (D0 )) ≤ ρα,
where Dα is the α-Rényi divergence between the distribution of M (D) and M (D0 ).
Proposition 1 (Propositions 1.3 and 1.4 of [BS16]). If a mechanism M1 satisfies (ε, 0)-differential privacy,
p
2
then M1 satisfies ε2 -zCDP. If a mechanism M2 satisfies ρ-zCDP, then M2 satisfies (ρ + 2 ρ log(1/δ), δ)differential privacy for any δ > 0.
3
A Simple Upper Bound
√
In this section, we provide an O α2nε upper bound for the differentially private identity testing problem.
More generally, we show that if an algorithm requires a dataset of size m for a decision problem, then it
can be made (ε, 0)-differentially private at a multiplicative cost of 1/ε in the sample size. This is a folklore
result, but we include and prove it here for completeness.
Theorem 2. Suppose there exists an algorithm for a decision problem P which succeeds with probability at
least 1 − β and requires a dataset of size m. Then there exists an (ε, 0)-differentially private algorithm for
P which succeeds with probability at least 45 (1 − β) + 1/10 and requires a dataset of size O(m/ε).
4
Proof. First, with probability 1/5, we flip a coin and output yes or no with equal probability. This guarantees
that we have probability at least 1/10 of either outcome, which will allow us to satisfy the multiplicative
guarantee of differential privacy.
We then draw 10/ε datasets of size m, and solve the decision problem (non-privately) for each of them.
Finally, we select a random one of these computations and output its outcome.
The correctness follows, since we randomly choose the right answer with probability 1/10, or with probability 4/5, we solve the problem correctly with probability 1 − β. As for privacy, we note that, if we remove
a single element of the dataset, we may only change the outcome of one of these computations. Since we
pick a random computation, this is selected with probability ε/10, and thus the probability of any outcome
is additively shifted by at most ε/10. Since we know the minimum probability of any output is 1/10, this
gives the desired multiplicative guarantee required for (ε, 0)-differential privacy.
√
We obtain the following corollary by noting that the tester of [ADK15] (among others) requires O( n/α2 )
samples for identity testing.
Corollary 1. There exists an (ε, 0)-differentially private testing algorithm for the (α, βI , βII )-identity testing
problem for any distribution q which requires
√
n
· log(1/β)
m=O
εα2
samples, where β = min (βI , βII ).
4
Roadblocks to Differentially Private Testing
In this section, we describe roadblocks which prevent two natural approaches to differentially private testing
from working.
In Section 4.1, we show that if one simply adds Laplace noise to the empirical counts of a dataset (i.e.,
runs the Laplace mechanism of Theorem 1) and then attempts to run an optimal identity tester, the variance
of the statistic increases dramatically, and thus results in a much larger sample complexity, even for the case
of uniformity testing. The intuition behind this phenomenon is as follows. When performing uniformity
testing in the small sample regime (when the number
√ of samples m is the square root of the domain size
n), we will see a (1 − o(1))n elements 0 times, O( n) elements 1 time, and O(1) elements 2 times. If we
add Laplace(10) noise to guarantee (0.1, 0)-differential privacy, this obliterates the signal provided by these
collision statistics, and thus many more samples are required before the signal prevails.
In Section 4.2, we demonstrate that χ2 statistics have high sensitivity, and thus are not naturally differentially private. In other words, if we consider a χ2 statistic Z on two datasets D and D0 which differ in one
record, |Z(D) − Z(D0 )| may be quite large. This implies that methods such as rescaling this statistic and
interpreting it as a probability, or applying noise to the statistic, will not be differentially private until we
have taken a large number of samples.
4.1
A Laplaced χ2 -statistic has large variance
Proposition 2. Applying the Laplace mechanism to a dataset before applying the identity tester of [ADK15]
results in a significant increase in the variance, even when considering the case of uniformity. More precisely,
if we consider the statistic
X (Ni + Yi − m/n)2 − (Ni + Yi )
Z 0 (D) =
m/n
i∈[n]
where Ni is the number of occurrences of symbol i in the dataset D (which is of size P oisson(m)) and
Yi ∼ Laplace(1/ε), then
• If p is uniform, then E[Z 0 ] =
2n2
ε2 m
and Var[Z 0 ] ≥
20n3
ε4 m2 .
• If p is a particular distribution which is α-far in total variation distance from uniform, then E[Z 0 ] =
2
4mα2 + ε2n
2m .
5
The variance of the statistic can be compared to that of the unnoised statistic, which is upper bounded by
m2 α4 . We can see that the noised statistic has larger variance until m = Ω(n3/4 ).
Proof. First, we compute the mean of Z 0 . Note that since |D| ∼ P oisson(m), the Ni ’s will be independently
distributed as P oisson(mpi ) (see, i.e., [ADK15] for additional discussion).
X
(Ni + Yi − m/n)2 − (Ni + Yi )
0
E[Z ] = E
m/n
i∈[n]
X
(Ni − m/n)2 − Ni
=E
m/n
i∈[n]
X Y 2 + 2Yi (Ni − m/n) − Yi
i
+
m/n
i∈[n]
X
= m · χ2 (p, q) +
i∈[n]
2
ε2
m/n
2n2
ε2 m
= m · χ2 (p, q) +
In other words, the mean is a rescaling of the χ2 distance between p and q, shifted by some constant amount.
When p = q, the χ2 -distance between p and q is 0, and the expectation is just the second term. Focus on
the case where n is even, and consider p such that pi = (1 + 2α)/n if i is even, and (1 − 2α)/n otherwise.
This is α-far from uniform in total variation distance. Furthermore, by direct calculation, χ2 (p, q) = 4α2 ,
2
and thus the expectation of Z 0 in this case is 4mα2 + ε2n
2m .
Next, we examine the variance of Z 0 . Let λi = mpi and λ0i = mqi = m/n. By a similar computation as
before, we have that
X 1
Var[Z 0 ] =
2λ2i + 4λi (λi − λ0i )2
λ02
i∈[n] i
20
1
0
2
+ 2 (8λi + 2(2λi − 2λi − 1) ) + 4 .
ε
ε
Since all four summands of this expression are non-negative, we have that
Var[Z 0 ] ≥
20n3
20 X 1
=
.
ε4
λ02
ε4 m 2
i
i∈[n]
If we wish to use Chebyshev’s inequality to separate these two cases, we require that Var[Z 0 ] is at most
the square of the mean separation. In other words, we require that
20n3
≤ m2 α 4 ,
ε4 m2
or that
m=Ω
4.2
n3/4
εα
.
A χ2 -statistic has high sensitivity
Consider the primary statistic which we use in Algorithm 1:
Z(D) =
1 X (Ni − mqi )2 − Ni
.
mα2
mqi
i∈[n]
6
As shown in Section 5, E[Z] = 0 if p = q and E[Z] ≥ 1 if dTV (p, q) ≥ α, and the variance of Z is such that
these two cases can be separated with constant probability. A natural approach is to truncate this statistic
to the range [0, 1], interpret it as a probability and output the result of Bernoulli(Z) – if p = q, the result is
likely to be 0, and if dTV (p, q) ≥ α, the result is likely to be 1. One might hope that this statistic is naturally
private. More specifically, we would like that the statistic Z has low sensitivity, and does not change much
if we remove a single individual. Unfortunately, this is not the case. We consider datasets D, D0 , where D0
is identical to D, but with one fewer occurrence of symbol i. It can be shown that the difference in Z is
|Z(D) − Z(D0 )| =
2|Ni − mqi − 1|
m2 α2 qi
Letting q be the uniform distribution and requiring that this is at most ε (for the sake of privacy), we have
a constraint which is roughly of the form
2Ni n
≤ ε,
m2 α2
or that
√ √
Ni n
m=Ω
.
ε0.5 α
√
In particular, if Ni = nc for any c > 0, this does not achieve the desired O( n) sample complexity. One
may observe that, if Ni is this large, looking at symbol i alone is sufficient to conclude p is not uniform, even
if the count Ni had Laplace noise added. Indeed, our main algorithm of Section 5 works in part due to our
formalization and quantification of this intuition.
5
Priv’IT: A Differentially Private Identity Tester
In this section, we prove our main testing upper bound:
Theorem 3. There exists an (ε, 0)-differentially private testing algorithm for the (α, βI , βII )-identity testing
problem for any distribution q which requires
√
√
n
n
n1/3
m = Õ max
,
,
· log(1/β)
α2 α3/2 ε α5/3 ε2/3
samples, where β = min (βI , βII ).
The pseudocode for this algorithm is provided in Algorithm 1. We fix the constants c1 = 1/4 and
c2 = 3/40. For a high-level overview of our algorithm’s approach, we refer the reader to the Approach
paragraph in Section 1.
Proof of Theorem 3: We will prove the theorem for the case where β = 1/3, the general case follows at
the cost of a multiplicative log(1/β) in the sample complexity from a standard amplification argument. To
be more precise, we can consider splitting our dataset into O(log(1/β)) sub-datasets and run the β = 1/3
test on each one independently. We return the majority result – since each test is correct with probability
≥ 2/3, correctness of the overall test follows by Chernoff bound. It remains to argue privacy – note that a
neighboring dataset will only result in a single sub-dataset being changed. Since we take the majority result,
conditioning on the result of the other sub-tests, the result on this sub-dataset will either be irrelvant to or
equal to the overall output. In the former case, any test is private, and in the latter case, we know that the
individual test is ε-differentially private. Overall privacy follows by applying the law of total probability.
We require the following two claims, which give bounds on the random variables Ni and Yi . Note that,
due to the fact that we draw P oisson(m) samples, each Ni ∼ P oisson(mpi ) independently.
Claim 1. |Yi | ≤ c22ε log 1−(1−c12 )1/|A| simultaneously for all i ∈ A with probability exactly 1 − c2 .
Proof. The survival function of the folded Laplace distribution with
parameter 2/c2 ε is exp (−c2 εx/2), and
the probability that a sample from it exceeding the value c22ε log 1−(1−c12 )1/|A| is equal to 1 − (1 − c2 )1/|A| .
The probability that probability that it does not exceed this value is (1 − c2 )1/|A| , and since the Yi ’s are
independent, the probability that none exceeds this value is 1 − c2 , as desired.
7
Algorithm 1 Priv’IT: A differentially private identity tester
1: Input: ε; an explicit distribution q; sample access to a distribution p
2: Define A ← {i : qi ≥ c1 α/n}, Ā ← [n] \ A
3: Sample Yi ∼ Laplace(2/c2 ε) for all i ∈ A
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
if there exists i ∈ A such that |Yi | ≥ c22ε log 1−(1−c12 )1/|A| then
return either “p 6= q” or “p = q” with equal probability
end if
Draw a multiset S of P oisson(m) samples from p
Let Ni be the number of occurrences of the ith domain element in S
for i ∈ A do
√
if |Ni + Yi − mqi | ≥ c22ε log 1−(1−c12 )1/|A| + max 4 mqi log n, log n then
return “p 6= q”
end if
end for P
(Ni −mqi )2 −Ni
2
Z ← mα
2
i∈A
mqi
Let T be the closest value to Z which is contained in the interval [0, 1]
Sample b ∼ Bernoulli(T )
if b = 1 then
return “p 6= q”
else
return “p = q”
end if
√
Claim 2. |Ni − mpi | ≤ max 4 mpi log n, log n
2
− 1.1
1 − n0.84
n .
simultaneously for all i ∈ A with probability at least
Proof. We consider this in two cases. Let X be a P oisson(λ) random variable. First, assume that λ ≥
e−3 log n. By Bennett’s inequality, we have the following tail bound [Pol15, Can17]:
x2 x
Pr [|X − λ| ≥ x] ≤ 2 exp − ψ
,
2λ
λ
where
ψ(t) =
(1 + t) log(1 + t) − t
.
t2 /2
√
Consider x = 4 λ log n. At this point, we have
p
ψ(x/λ) = ψ(4 log n/λ) ≥ ψ(4e3/2 ) ≥ 0.23.
Thus,
i
h
p
Pr |X − λ| ≥ 4 λ log n ≤ 2 exp (−0.23 · 8 log n)
≤ 2n−1.84 .
Now, we focus on the other case, where λ ≤ e−3 log n. Here, we appeal to Proposition 1 of [Kla00], which
implies the following via Stirling’s approximation:
Pr [|X − λ| ≥ kλ] ≤
k
exp(−λ + kλ − kλ log k).
k−1
We set kλ = log n, giving the upper bound
k
n1−log k ≤ 1.1 · n−2 .
k−1
We conclude by taking a union bound over [n], with the argument for each i ∈ [n] depending on whether
λ = mpi is large or small.
We proceed with proving the two desiderata of this algorithm, correctness and privacy.
8
Correctness.
We use the following two properties of the statistic Z(D), which rely on the condition that
√
m = Ω( n/α2 ). The proofs of these properties are identical to the proofs of Lemma 2 and 3 in [ADK15],
and are omitted.
Claim 3. If p = q, then E[Z] = 0. If dTV (p, q) ≥ α, then E[Z] ≥ 1.
Claim 4. If p = q, then Var[Z] ≤ 1/1000. If dTV (p, q) ≥ α, then Var[Z] ≤ 1/1000 · E[Z]2 .
First, we note that, by Claim 1, the probability that we return in line 5 is exactly c2 . We now consider
the case where p = q. We note that by Claim 2, the probability that we output “p 6= q” in line 10 is o(1),
and thus negligible. By Chebyshev’s inequality, we get that Z ≤ 1/10 with probability at least 9/10, and we
output “p = q” with probability at least c2 /2 + (1 − c2 ) · (9/10 − c2 )2 ≥ 2/3 (note that we subtract c2 from
9/10 since we are conditioning on an event with probability 1 − c2 , and by union bound). Similarly, when
dTV (p, q) ≥ α, Chebyshev’s inequality gives that Z ≥ 9/10 with probability at least 9/10, and therefore we
output “p 6= q” with probability at least 2/3.
Privacy. We will prove (0, c2 ε/2)-differential privacy. By Claim 1, the probability that we return in line 5
is exactly c2 . Thus the minimum probability of any output of the algorithm is at least c2 /2, and therefore
(0, c2 ε/2)-differential privacy implies (ε, 0)-differential privacy.
We first consider the possibility of rejecting in line 11. Consider two neighboring datasets D and D0 ,
which differ by 1 in the frequency of symbol i. Coupling the randomness of the Yj ’s on these two datasets,
the only case in which the output differs is when Yi is such that the value of |Ni + Yi − mqi | lies on opposite
sides of the threshold for the two datasets. Since Ni differs by 1 in the two datasets, and the probability
mass assigned by the PDF of Yi to any interval of length 1 is at most c2 ε/4, the probability that the outputs
differ is at most c2 ε/4. Therefore, this step is (0, c2 ε/4)-differentially private.
We next consider the value of Z for two neighboring datasets D and D0 , where D0 has one fewer occurrence
of symbol i. We only consider the case where we have not already returned in line 11, as otherwise the value
of Z is irrelevant for determining the output of the algorithm.
Z(D) − Z(D0 )
(Ni − mqi )2 − Ni
(Ni − 1 − mqi )2 − (Ni − 1)
1
−
=
mα2
mqi
mqi
2
2
1
(Ni − mqi ) − Ni
(Ni − mqi ) − 2(Ni − mqi ) + 1 − Ni + 1
=
−
mα2
mqi
mqi
2(Ni − mqi − 1)
.
=
m2 α2 qi
Since we did not return in line 11,
n p
o
1
4
log
+
max
4
mq
log
n,
log
n
|Ni − mqi | ≤
i
c2 ε
1 − (1 − c2 )1/n
n p
o
4 log(n/c2 )
≤
+ max 4 mqi log n, log n
c2 ε
This implies that
2|Ni − mqi − 1|
m2 α 2 qi
p
2
6 log(n/c2 )
≤ 2 2
+ 4 mqi log n .
m α qi
c2 ε
|Z(D) − Z(D0 )| =
We will enforce that each of these terms are at most c2 ε/8.
s
p
n log(n/c2 )
12 log(n/c2 )
c2 ε
96
≤
⇒m≥
m2 α2 qi c2 ε
8
c22 c1
α1.5 ε
√
2/3
8 log n
c2 ε
64
(n log n)1/3
≤
⇒
m
≥
√
√
m1.5 α2 qi
8
c2 c1
α5/3 ε2/3
9
Since both terms are at most c2 ε/8, this step is (0, c2 ε/4)-differentially private. Combining with the
previous step gives the desired (0, c2 ε/2)-differential privacy, and thus (as argued at the beginning of the
privacy section of this proof) ε-pure differential privacy.
6
Experiments
We performed an empirical evaluation of our algorithm, Priv’IT, on synthetic datasets. All experiments
were performed on a laptop computer with a 2.6 GHz Intel Core i7-6700HQ CPU and 8 GB of RAM.
Significant discussion is required to provide a full comparison with prior work in this area, since performance
of the algorithms varies depending on the regime.
We compared our algorithm with two recent algorithms for differentially private hypothesis testing:
1. The Monte Carlo Goodness of fit test with Laplace noise from [GLRV16], MCGOF;
2. The projected Goodness of Fit test from [KR17], zCDP-GOF.
We note that we implemented a modified version of Priv’IT, which differs from Algorithm 1 in lines 14
to 21. In particular, we instead consider a statistic
Z=
X (Ni − mqi )2 − Ni
mqi
i∈A
.
We add Laplace noise to Z, with scale parameter Θ(∆/ε), where ∆ is the sensitivity of Z, which guarantees
(ε/2, 0)-differential privacy. Then, similar to the other algorithms, we choose a threshold for this noised
statistic such that we have the desired type I error. This algorithm can be analyzed to provide identical
theoretical guarantees as Algorithm 1, but with the practical advantage that there are fewer parameters to
tune.
To begin our experimental evaluation, we started with uniformity testing. Our experimental setup was
as follows. The algorithms were provided q as the uniform distribution over [n]. The algorithms were also
provided with samples from some distribution p. This (unknown) p was q for the case p = q, or a distribution
which we call the “Paninski construction” for the case dTV (p, q) ≥ α. The Paninski construction is a
distribution where half the elements of the support have mass (1 + α)/n and half have mass (1 − α)/n. We
use this name for the construction
√ as [Pan08] showed that this example is one of the hardest to distinguish
from uniform: one requires Ω( n/α2 ) samples to (non-privately) distinguish a random permutation of this
construction from the uniform distribution. We fixed parameters ε = 0.1 and α = 0.1. In addition, recall
that Proposition 1 implies that pure differential privacy (the privacy guaranteed by Priv’IT) is stronger
than zCDP (the privacy guaranteed by zCDP-GOF). In particular, our guarantee of ε-pure differential privacy
implies ε2 /2-zCDP. As a result, we ran zCDP-GOF with a privacy parameter of 0.005-zCDP, which is equivalent
to the amount of zCDP our algorithm provides. Our experiments were conducted on a number of different
support sizes n, ranging from 10 to 10600. For each n, we ran the testing algorithms with increasing sample
sizes m in order to discover the minimum sample size when the type I and type II errors were both empirically
below 1/3. To determine these empirical error rates, we ran all algorithms 1000 times for each n and m, and
recorded the fraction of the time each algorithm was correct. As the other algorithms take a parameter βI
as a target type I error, we input 1/3 as this parameter.
The results of our first test are provided in Figure 1. The x-axis indicates the support size, and the y-axis
indicates the minimum number of samples required. We plot three lines, which demonstrate the empirical
number of samples required to obtain 1/3 type I and type II error for the different algorithms. We can see
that in this case, zCDP-GOF is the most statistically efficient, followed by MCGOF and Priv’IT.
To explain this difference in statistical efficiency, we note that the theoretical guarantees of Priv’IT
imply that it performs well even when data is sparsely sampled. More precisely, one of the benefits of our
tester is that it can reduce the variance induced by elements whose expected number of occurrences is less
than 1. Since none of these testers reach this regime (i.e., even zCDP-GOF at n = 10000 expects to see each
element 10 times), we do not reap the benefits of Priv’IT. Ideally, we would run these algorithms on the
uniform distribution at sufficiently large support sizes. However, since this is prohibitively expensive to do
10
Uniformity Testing
Priv’IT
zCDP-GOF
MCGOF
250000
Sample Complexity (m)
200000
150000
100000
50000
0
0
2000
4000
6000
Support Size (n)
8000
10000
Figure 1: The sample complexities of Priv’IT, MCGOF, and zCDP-GOF for uniformity testing
with thousands of repetitions (for any of these methods), we instead demonstrate the advantages of our
tester on a different distribution.
Our second test is conducted with q being a 2-histogram1 , where all but a vanishing fraction of the
probability mass is concentrated on a small, constant fraction of the support2 . This serves as our proxy
for a very large support, since now we will have elements which have a sub-constant expected number of
occurrences. The algorithms are provided with samples from a distribution p, which is either q or a similar
Paninski construction as before, where the total variation distance from q is placed on the support elements
containing non-negligible mass. We ran the test on support sizes n ranging from 10 to 6800. All other
parameters are the same as in the previous test.
The results of our second test are provided in Figure 2. In this case, we compare Priv’IT and zCDP-GOF,
and note that our test is slightly better for all support sizes n, though the difference can be pronounced
or diminished depending on the construction of the distribution q. We found that MCGOF was incredibly
inefficient on this construction – even for n = 400 it required 130000 samples, which is a factor of 10 worse
than zCDP-GOF on a support of size n = 6800. To explain this phenomenon, we can inspect the contribution
of a single domain element i to their statistic:
(Ni + Yi − mqi )2
.
mqi
Y2
In the case where mqi 1 and p = q, this is approximately equal to mqi i . The standard deviation of this
term will be of the order mq1i ε2 , which can be made arbitrarily large as mqi → 0. While zCDP-GOF may
naively seem susceptible to this same pitfall, their projection method appears to elegantly avoid it.
As a final test, we note that zCDP-GOF guarantees zCDP, while Priv’IT guarantees (vanilla) differen2
tial privacy. In our previous tests, our guarantee was ε-differential privacy, while theirs was ε2 -zCDP: by
Proposition 1, our guarantees imply theirs. In the third test, we revisit uniformity testing, but when their
2
guarantees imply ours. More specifically, again with ε = 0.1, we ran zCDP-GOF with the guarantee of ε2 p
2
zCDP and Priv’IT with the guarantee of ( ε2 + ε 2 log(1/δ), δ) for various δ > 0. We note that δ is often
thought in theory to be “cryptographically small” (such as 2−100 ), but we compare with a wide range of δ,
both large and small: δ = 1/et for t ∈ {1, 2, 4, 8, 16}. This test was conducted on support sizes n ranging
from 10 to 6000.
The results of our third test are provided in Figure 3. We found that, for all δ tested, Priv’IT required
fewer samples than zCDP-GOF. This is unsurprising for δ very large and small, since the differential privacy
guarantees become very easy to satisfy, but we found it to be true for even “moderate” values of δ. This
1 A k-histogram is a distribution where the domain can be partitioned into k intervals such that the distribution is uniform
over each interval.
2 In particular, in Figure 3, n/200 support elements contained 1 − 10/n probability mass, but similar trends hold with
modifications of these parameters.
11
Identity Testing on a 2-Histogram
14000
Priv’IT
zCDP-GOF
12000
Sample Complexity (m)
10000
8000
6000
4000
2000
0
0
1000
2000
3000
4000
Support Size (n)
5000
6000
7000
Figure 2: The sample complexities of Priv’IT and zCDP-GOF for identity testing on a 2-histogram
Uniformity Testing, Revisited
zCDP-GOF
Priv’IT (δ = 1/e)
80000
Priv’IT (δ = 1/e2 )
Priv’IT (δ = 1/e4 )
Priv’IT (δ = 1/e8 )
Sample Complexity (m)
60000
Priv’IT (δ = 1/e16 )
40000
20000
0
0
1000
2000
3000
Support Size (n)
4000
5000
6000
Figure 3: The sample complexities of Priv’IT and zCDP-GOF for uniformity testing, with approximate
differential privacy
implies that if an analyst is satisfied with approximate differential privacy, she might be better off using
Priv’IT, rather than an algorithm which guarantees zCDP.
While the main focus of our evaluation was statistical in nature, we will note that Priv’IT was more
efficient in runtime than our implementation of MCGOF, and more efficient in memory usage than our implementation of zCDP-GOF. The former point was observed by noting that, in the same amount of time, Priv’IT
was able to reach a trial corresponding to a support size of 20000, while MCGOF was only able to reach 10000.
The latter point was observed by noting that zCDP-GOF ran out of memory at a support size of 11800. This
is likely because zCDP-GOF requires matrix computations on a matrix of size O(n2 ). It is plausible that all of
these implementations could be made more time and memory efficient, but we found our implementations
to be sufficient for the sake of our comparison.
Acknowledgments
The authors would like to thank Jon Ullman for helpful discussions in the early stages of this work. The authors were supported by NSF CCF-1551875, CCF-1617730, CCF-1650733, and ONR N00014-12-1-0999.
12
References
[AA01]
Dakshi Agrawal and Charu C. Aggarwal. On the design and quantification of privacy preserving
data mining algorithms. In Proceedings of the 20th ACM SIGMOD-SIGACT-SIGART Symposium on Principles of Database Systems, PODS ’01, pages 247–255, New York, NY, USA, 2001.
ACM.
[ADK15]
Jayadev Acharya, Constantinos Daskalakis, and Gautam Kamath. Optimal testing for properties
of distributions. In Advances in Neural Information Processing Systems 28, NIPS ’15, pages
3577–3598. Curran Associates, Inc., 2015.
[Agr12]
Alan Agresti. Categorical Data Analysis. Wiley, 2012.
[AW89]
Nabil R. Adam and John C. Worthmann. Security-control methods for statistical databases: A
comparative study. ACM Computing Surveys (CSUR), 21(4):515–556, 1989.
[BFF+ 01] Tugkan Batu, Eldar Fischer, Lance Fortnow, Ravi Kumar, Ronitt Rubinfeld, and Patrick White.
Testing random variables for independence and identity. In Proceedings of the 42nd Annual IEEE
Symposium on Foundations of Computer Science, FOCS ’01, pages 442–451, Washington, DC,
USA, 2001. IEEE Computer Society.
[BS16]
Mark Bun and Thomas Steinke. Concentrated differential privacy: Simplifications, extensions,
and lower bounds. In Proceedings of the 14th Conference on Theory of Cryptography, TCC ’16-B,
pages 635–658, Berlin, Heidelberg, 2016. Springer.
[Can15]
Clément L. Canonne. A survey on distribution testing: Your data is big. but is it blue? Electronic
Colloquium on Computational Complexity (ECCC), 22(63), 2015.
[Can17]
Clément L. Canonne. A short note on Poisson tail bounds. http://www.cs.columbia.edu/
~ccanonne/files/misc/2017-poissonconcentration.pdf, 2017.
[CDGR16] Clément L. Canonne, Ilias Diakonikolas, Themis Gouleakis, and Ronitt Rubinfeld. Testing shape
restrictions of discrete distributions. In Proceedings of the 33rd Symposium on Theoretical Aspects
of Computer Science, STACS ’16, pages 25:1–25:14, 2016.
[Dal77]
Tore Dalenius. Towards a methodology for statistical disclosure control. Statistisk Tidskrift,
15:429–444, 1977.
[DDK16]
Constantinos Daskalakis, Nishanth Dikkala, and Gautam Kamath. Testing Ising models. arXiv
preprint arXiv:1612.03147, 2016.
[DHS15]
Ilias Diakonikolas, Moritz Hardt, and Ludwig Schmidt. Differentially private learning of structured discrete distributions. In Advances in Neural Information Processing Systems 28, NIPS
’15, pages 2566–2574. Curran Associates, Inc., 2015.
[DK16]
Ilias Diakonikolas and Daniel M. Kane. A new approach for testing properties of discrete distributions. In Proceedings of the 57th Annual IEEE Symposium on Foundations of Computer
Science, FOCS ’16, pages 685–694, Washington, DC, USA, 2016. IEEE Computer Society.
[DMNS06] Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith. Calibrating noise to sensitivity
in private data analysis. In Proceedings of the 3rd Conference on Theory of Cryptography, TCC
’06, pages 265–284, Berlin, Heidelberg, 2006. Springer.
[DN03]
Irit Dinur and Kobbi Nissim. Revealing information while preserving privacy. In Proceedings
of the 22nd ACM SIGMOD-SIGACT-SIGART Symposium on Principles of Database Systems,
PODS ’03, pages 202–210, New York, NY, USA, 2003. ACM.
[DR14]
Cynthia Dwork and Aaron Roth. The Algorithmic Foundations of Differential Privacy. Now
Publishers, Inc., 2014.
13
[Dwo08]
Cynthia Dwork. Differential privacy: A survey of results. In Proceedings of the 5th International
Conference on Theory and Applications of Models of Computation, TAMC ’08, pages 1–19, Berlin,
Heidelberg, 2008. Springer.
[Fis35]
Ronald A. Fisher. The Design of Experiments. Macmillan, 1935.
[GLRV16] Marco Gaboardi, Hyun-Woo Lim, Ryan M. Rogers, and Salil P. Vadhan. Differentially private
chi-squared hypothesis testing: Goodness of fit and independence testing. In Proceedings of the
33rd International Conference on Machine Learning, ICML ’16, pages 1395–1403. JMLR, Inc.,
2016.
[HSR+ 08] Nils Homer, Szabolcs Szelinger, Margot Redman, David Duggan, Waibhav Tembe, Jill Muehling,
John V. Pearson, Dietrich A. Stephan, Stanley F. Nelson, and David W. Craig. Resolving
individuals contributing trace amounts of dna to highly complex mixtures using high-density snp
genotyping microarrays. PLoS Genetics, 4(8):1–9, 2008.
[JS13]
Aaron Johnson and Vitaly Shmatikov. Privacy-preserving data exploration in genome-wide association studies. In Proceedings of the 19th ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, KDD ’13, pages 1079–1087, New York, NY, USA, 2013. ACM.
[Kla00]
Bernhard Klar. Bounds on tail probabilities of discrete distributions. Probability in the Engineering and Informational Sciences, 14(02):161–171, 2000.
[KR17]
Daniel Kifer and Ryan M. Rogers. A new class of private chi-square tests. In Proceedings of
the 20th International Conference on Artificial Intelligence and Statistics, AISTATS ’17, pages
991–1000. JMLR, Inc., 2017.
[KSF17]
Kazuya Kakizaki, Jun Sakuma, and Kazuto Fukuchi. Differentially private chi-squared test by
unit circle mechanism. In Proceedings of the 34th International Conference on Machine Learning,
ICML ’17. JMLR, Inc., 2017.
[LRR13]
Reut Levi, Dana Ron, and Ronitt Rubinfeld. Testing properties of collections of distributions.
Theory of Computing, 9(8):295–347, 2013.
[Pan08]
Liam Paninski. A coincidence-based test for uniformity given very sparsely sampled discrete
data. IEEE Transactions on Information Theory, 54(10):4750–4755, 2008.
[Pea00]
Karl Pearson. On the criterion that a given system of deviations from the probable in the case
of a correlated system of variables is such that it can be reasonably supposed to have arisen from
random sampling. Philosophical Magazine Series 5, 50(302):157–175, 1900.
[Pol15]
David Pollard. A few good inequalities. http://www.stat.yale.edu/~pollard/Books/Mini/
Basic.pdf, 2015.
[RS81]
Jon N.K. Rao and Alastair J. Scott. The analysis of categorical data from complex sample
surveys: Chi-squared tests for goodness of fit and independence in two-way tables. Journal of
the Americal Statistical Association, 76(374):221–230, 1981.
[SSB16]
Sean Simmons, Cenk Sahinalp, and Bonnie Berger. Enabling privacy-preserving gwass in heterogeneous human populations. Cell Systems, 3(1):54–61, 2016.
[TAW10]
Vincent Y.F. Tan, Animashree Anandkumar, and Alan S. Willsky. Error exponents for composite
hypothesis testing of Markov forest distributions. In Proceedings of the 2010 IEEE International
Symposium on Information Theory, ISIT ’10, pages 1613–1617, Washington, DC, USA, 2010.
IEEE Computer Society.
[USF13]
Caroline Uhler, Aleksandra Slavković, and Stephen E. Fienberg. Privacy-preserving data sharing
for genome-wide association studies. The Journal of Privacy and Confidentiality, 5(1):137–166,
2013.
14
[VV14]
Gregory Valiant and Paul Valiant. An automatic inequality prover and instance optimal identity
testing. In Proceedings of the 55th Annual IEEE Symposium on Foundations of Computer Science,
FOCS ’14, pages 51–60, Washington, DC, USA, 2014. IEEE Computer Society.
[WLK15]
Yue Wang, Jaewoo Lee, and Daniel Kifer. Differentially private hypothesis testing, revisited.
arXiv preprint arXiv:1511.03376, 2015.
[YFSU14] Fei Yu, Stephen E. Fienberg, Aleksandra B. Slavković, and Caroline Uhler. Scalable privacypreserving data sharing methodology for genome-wide association studies. Journal of Biomedical
Informatics, 50:133–141, 2014.
15
| 8cs.DS
|
Under review as a conference paper at ICLR 2016
I DENTIFYING
L EARNING
THE
A BSORPTION B UMP
WITH
D EEP
arXiv:1511.05607v2 [cs.CV] 20 Nov 2015
Min Li, Sudeep Gaddam, Xiaolin Li
Department of Electrical and Computer Engineering
University of Florida
Gainesville, FL 32611, USA
{minli, sudeepgaddam}@ufl.edu, [email protected]
Yinan Zhao, Jingzhe Ma, Jian Ge
Department of Astronomy
University of Florida
Gainesville, FL 32611, USA
{yinanzhao, jingzhema}@ufl.edu, [email protected]
A BSTRACT
The pervasive interstellar dust grains provide significant insights to understand the
formation and evolution of the stars, planetary systems, and the galaxies, and may
harbor the building blocks of life. One of the most effective way to analyze the
dust is via their interaction with the light from background sources. The observed
extinction curves and spectral features carry the size and composition information
of dust. The broad absorption bump at 2175 Å is the most prominent feature in the
extinction curves. Traditionally, statistical methods are applied to detect the existence of the absorption bump. These methods require heavy preprocessing and the
co-existence of other reference features to alleviate the influence from the noises.
In this paper, we apply Deep Learning techniques to detect the broad absorption
bump. We demonstrate the key steps for training the selected models and their
results. The success of Deep Learning based method inspires us to generalize a
common methodology for broader science discovery problems. We present our
on-going work to build the DeepDis system for such kind of applications.
1
I NTRODUCTION
The pervasive interstellar dust contains footprints of how the cosmos has been evolving. Dust grains
come from condensation of heavy elements produced by stars. They are considered the primary reaction sites for molecules to form and essentially the source for all the H2 in the interstellar medium
(Gould & Salpeter (1963); Hollenbach & Salpeter (1971)). Dust grains also play an important role
in controlling the gas-phase metal abundances and the thermodynamic evolution of the interstellar
medium. Besides, coagulation of interstellar dust grains in a protostellar disk, along with their catalyzed complex organic molecules, eventually leads to planets. Dust is thought to harbor the secrets
of planetary formation, and even life.
Research on dust and gas is conducted by analyzing the reddening and extinction effects on the
spectra of background light sources and re-radiated emission. In the Local Universe, it is feasible
to compare stellar spectra and explore the extinction curves to investigate the dust grains. However,
more luminous background light sources, such as quasars, are necessary to probe the dust and gas
ingredients in the galaxies at larger distances.
Among the research topics on cosmic dust grains, the 2175 Å broad absorption bump 1 stands out
with significant values. Although the precise characteristics have not yet been established, the 2175
Å broad absorption bump is believed to be tightly bounded with some types of aromatic carbonaceous materials (Draine (2003); Li & Greenberg (2003)). The most promising carriers of the bump
1
We use the 2175 Å broad absorption bump and absorption bump for short in the rest of this paper.
1
Under review as a conference paper at ICLR 2016
are the Polycyclic Aromatic Hydrocarbon (PAH) molecules. PAHs are recognized as the most abundant organic molecules in our Milk Way and other neighboring galaxies (Peeters et al. (2003)), and
are believed to be the building blocks of organic life (Bernstein et al. (2002); Caro et al. (2002)).
The traditional way of discovering the absorption bump mainly depends on statistic techniques
(Jiang et al. (2011)). It generally involves three steps. First, a composite quasar spectrum is constructed using median combining (Jiang et al. (2010a;b)). The composite spectrum is then reddened
and used to fit every spectrum from the candidate absorption bump spectra with a parameterized
absorption profile. Finally, by applying constrains on peak position, bump width and bump height,
a large portion of spectra without bump features are filtered out. In order to further exclude the
false positives caused by some noises, a simulation technique (Jiang et al. (2010a;b)) is applied to
determine the detection significance. Among the steps, curve fitting occupies considerable amount
of time. In order to convolve the absorption bump with the composite spectrum, extra information is
needed to determine the absorption redshift. This significantly constrains the available observations
due to the lack of this extra information in some cases. While the composite quasar spectrum could
potentially be reused for other related computations, the fitting procedure evolves iterative error
minimization to get the best bump profile using curve fitting method, and is required for each and
every new observation. The fitting is restricted to relatively smooth continuum of the quasar spectrum. However, the complicated types of emission and absorption in the candidate spectra could
potentially disturb the fitting process, reducing the effectiveness of the method. All these issues are
tightly coupled with the method itself, and cannot be easily resolved.
In this paper, we propose to apply a Deep Learning based method to detect the absorption bump
and try to alleviate the aforementioned issues. Deep Learning is recently recognized with its ability
to automatically extract high level features and accurately recognize or classify the target objects
(Hinton & Salakhutdinov (2006); LeCun et al. (2015)). Deep Learning is most effective when
the model is trained with sufficient data and the features are complex. For the absorption bump
application, a normal observed spectrum consists of over 4,000 raw features. Each of them is a flux
value at a certain wavelength. Due to the complex cosmos environment, the relation space among
the features could be complicated. We also observe that in the traditional method, the curving fitting
process is actually generating various spectra with bumps. These two facts make Deep Learning
model a natural fit for such an application.
For the rest of this paper, we present the details of our Deep Learning based method for the absorption bump detection. Specifically, we first give more background information about the problem
and the traditional method. We then show how we generate the raw training and testing data sets.
Next, we present how we transform the raw data into different formats for models including fully
connected neural networks and two convolutional neural networks. The details and the results are
also reported. Finally, we present a generalization from our methodology to simplify and improve
the efficiency for similar astronomy problems, and also resembling problems in other science fields.
We discuss the workflow and the design details of our ongoing work, the DeepDis system, which
aims to provide such services.
2
T RADITIONAL METHOD FOR DISCOVERING ABSORPTION BUMP
Previous research (Fitzpatrick & Massa (1990)) showed that the extinction curve and the variation
of the intrinsic spectral slopes of a quasar can be represented by a parameterized linear component. A Drude component can be added to approximate the possible absorption bump. Overall, the
representation for an extinction spectrum with an absorption bump can be expressed as:
A(λ) = c1 + c2 x + c3 D(x, x0 , γ)
(1)
, where λ is the wavelength, and x = λ−1 . The component D(x, x0 , γ) is a Drude profile, and its
definition is:
x2
D(x, x0 , γ) =
(2)
(x2 −
+ x2 γ 2
, where x0 and γ is the peak position and full width at half maximum (FWHM) of the Drude profile,
respectively. The strength can be measured by the area of the bump, i.e., Abump = πc3 /(2γ).
x20 )2
2
Under review as a conference paper at ICLR 2016
Rest Wavelength in Absorber Frame
2000
1500
Flux (10-17 erg s-1 cm-2
-1
)
SDSS J1211+0833
Best-fit model with bump
Reddened by linear component only
2500
Bump
3000
MgII
10
1
4000
5000
6000
7000
Observed Wavelength
8000
9000
Figure 1: Absorption Bump Example
The most popular method for discovering the 2175 Å absorption bump is rule-based filtering after fitting. The main idea is to get the best fitted parameters for each observed quasar spectrum,
following Equation 1 and Equation 2. The whole procedure consists of three major steps. First,
obtaining the median quasar composite spectrum. The basic assumption is that the quasars under
similar condition will have similar observed spectrum. A median quasar composite spectrum is regarded as the base spectrum. It is created by combining the spectra from a set of observed quasars
at their rest frame (e.g., Fitzpatrick & Massa (1990); Schneider et al. (2010)). Second, curve fitting
based on method such as Chi-Square Testing. The median quasar spectrum is first reddened to the
quasar’s emission redshift, and an absorption bump profile is convolved at the absorption redshift.
The absorption redshift is obtained by referring to some unique absorption lines, such as the Mg II
absorption lines. The parameters in Equation 1 and Equation 2 are selected to minimize the fitting
error. With the parameters, the last step is conducted by applying a set of rules to filter out low confidence candidates. One fitting example is shown in Figure 1, where the black curve is the observed
spectrum, the green curve is the reddened normal extinction spectrum continuum, and the red line is
the continuum with absorption bump.
There are a couple of issues in this method. First, hard dependency on Mg II absorption lines. The
absorption redshift is obtained by referring to the Mg II absorption lines. The Mg II absorption
lines are identified using some extra step, and is itself a research topic (Quider et al. (2011); Zhu
& Ménard (2013)). However, not all observations contain this absorption feature, due to different
emission and absorption redshift ranges. Second, the error minimizing curve fitting method is not
effective in some cases, e.g., broader Fe emission. This sometimes requires human identification.
Third, the fitting procedure for each observation is time consuming. Multiple iterations of error
minimizing steps are required to produce the best profile.
3
I DENTIFYING THE A BSORPTION B UMP WITH D EEP L EARNING
For the several aforementioned drawbacks of the traditional method, we propose a Deep Learning
based method that could alleviate these issues and provide more flexible usage. Specifically, the
proposed method exhibits the following merits.
• Large data set empowers accurate representation learning. The curve fitting step in the
traditional method is repeatedly generating training samples. By carefully designing the
generation, we are able to train an accurate model for the absorption bump detection.
• No hard dependency on Mg II absorption lines. When generating the samples, we are in
control of selecting the emission and absorption redshifts. After the model is trained, no
extra information is needed for new observations.
• Easy to use and share. With a trained model, each prediction on newly added observations
is just one feed forward process. Trained models could also be easily shared among researchers. Beyond trained models, simulation parameters and simulated data sets are all
sharing-friendly.
3
Under review as a conference paper at ICLR 2016
• Easy to extend. Extensions of already trained models happen at two levels. First, additional
training samples could be included to cover more cases. Second, multiple targets could also
be added based on the same feature set. These extensions could be achieved conveniently
by fine tuning of a trained model.
In order to validate the feasibility, we present our work on applying Deep Learning based method
for the absorption bump detection in the following subsections. Specifically, we show the details of
training data generation, and our experience on model selection and data transformation.
3.1
R AW DATA G ENERATION
We applied two methods to generate the data set. The first method follows the corresponding part
from the curve fitting procedure. We refer to Berk et al. (2001) and obtain the SDSS data release 7
composite quasar spectrum as the intrinsic spectrum. We then change the emission and absorption
redshifts and add the absorption bump profile according to Equation 2. We choose a typical bump
profile: x0 = 4.59, γ = 1.0 and Abump = 2.0.
In order to increase the diversity and closely simulate the real-world observations, we apply the
second method. We selected a subset of the spectra from the SDSS data release 7 according to the
Mg II catalog from Zhu & Ménard (2013). We follow the fitting method described in Jiang et al.
(2010a; 2011), and convolve them with the absorption bump profile same as the above one. These
generated samples are extremely close to real-world cases. Note that we refer to existing Mg II
catalog to generate data. However, no dependency is required for predicting with new observations.
A total of 30K samples are generated with the two methods described above. Half of them contains
the absorption bump and half does not. Each generated sample is stored in a file following the
<wavelength, flux> pattern. There are around 4600 pairs in each file. We divide the data into two
parts, 22K as the training data set and 8K as the testing data set.
3.2
C ASE O NE : F ULLY C ONNECTED N EURAL N ETWORK
Table 1: Configurations and Test Results for Fully Connected Neural Network
One-Layer
neurons accuracy
10000
71.603%
1000
71.677%
600
71.727%
Four-Layer
neurons accuracy
800
95.826%
600
96.460%
400
95.603%
Two-Layer
neurons accuracy
800
96.658%
600
96.485%
400
96.758%
Five-Layer
neurons accuracy
800
96.298%
600
96.236%
400
96.124%
Three-Layer
neurons accuracy
800
95.540%
600
95.888%
400
96.211%
The first straightforward choice for the model is the fully connected neural network. Its effectiveness
in capturing the non-linearity relationships among the input features has been tested through the past
decades. In order to feed the training data into the fully connected neural network, we extracted the
flux value as a vector, and use this vector to train the model. We perform this data transformation
based on the observation that the absorption bump is not related to the absolute wavelength value due
to redshifts, but rather the neighboring flux value relationship. In addition, the wavelength interval
between two consecutive wavelength-flux pair is relative stable in the data (both simulated and realworld). Due to the data collection mechanism, the number of pairs in one sample could be slightly
different. We padded the samples with zeros to round up to a total of 4761 flux values.
We changed the configurations of the fully connected neural network and vary the number of hidden
layers and the number of neurons per layer. The output layer is softmax with two classes, with or
without absorption bump. The initial learning rate is set to 0.01 and step decreasing police is applied.
The testing results after converged training are shown in Table 1. The performance for one hidden
4
Under review as a conference paper at ICLR 2016
layer is significantly worse than multiple-hidden-layer neural networks, due to its incapability to
capture the non-linearity. The best performance is observed in a two-hidden-layer network with 400
neurons per layer. Its total number of trainable parameters is around 2 million.
3.3
C ASE T WO : I MAGE AND C ONVOLUTIONAL N EURAL N ETWORK
A quasar spectrum is plotted as the black curve in Figure 1 when inspected by researchers. This
inspires us to transform raw data into images. We first plot all the training and testing samples.
The coverage span for the wavelength is 8000 Å and around 50 for the flux. The raw image drawn
can be as large as hundreds of KB. The images are down scaled to 256*256 gray images and used
for training and testing. Convolutional Neural Network (CNN) (LeCun & Bengio (1995)) recently
has proved its effectiveness in image related recognition and classification, due to its capability of
parameter sharing and capturing localized patterns. We adapted two popular CNN based models,
the AlexNet (Krizhevsky et al. (2012)) and the GoogLeNet (Szegedy et al. (2014)). The results we
get for modified AlexNet is 98.52% and for GoogLeNet is 98.55%.
(b)
(d)
(c)
(e)
1
(a)
Figure 2: Visualization for AlexNet. (a) shows the filters for the first convolutional layer. (b) and
(d) plot two test images with absorption bump, marked with red dashed circle. (c) and (e) are the
last convolutional layer feature maps for the two test images. The maps with effective activation are
marked with red dashed rectangles. Better viewed in electronic version.
In order to understand the learned representation, we select the trained AlexNet model and visualize
the first convolutional layer filters and also the last convolutional layer feature maps, as shown
in Figure 2. The first convolutional layer filters are typical edge detection filters. The effective
activation area in the feature maps (Figure 2c and 2e, marked with red dash rectangles) corresponds
to the absorption bump locations in the test images. This proves that the representation is sensitive
to the absorption bump. We also reconstructed the input images from a randomly generated image
with respect to the given classes (with or without bump) using the gradient ascend method. The
reconstructed images are shown in Figure3. The image for with bump clearly shows that the model
is sensitive to the absorption bump in the input spectrum.
3.4
C ASE T HREE : M ATRIX AND C ONVOLUTIONAL N EURAL N ETWORK
The 256*256 image plus CNN model improves the accuracy. However, the input image is too sparse
considering the large white background. In order to condense the input as well as empower the CNN
model, we perform another data transformation. This transformation is also based on the assumption
used for fully connected neural network. Different from the previous one, where the input is a 1D
vector, we pad and fold the vector into a 69*69 matrix. These folded matrices are fed into the CNN
models. This transformation makes it possible for filters in the convolutional layer to capture the
localized information with neighboring flux values. We use different configurations of the CNN and
5
Under review as a conference paper at ICLR 2016
(a) Without Bump
(b) With Bump
Figure 3: Reconstructed Input Image for with and without bump
post the best results we find in Table 2. We fix the CNN models with two fully connected layers after
the convolutional layers, and alter the number of convolutional layers, the number of filters in each
convolutional layer, and the kernel size. The best model we get achieves 99.404% test accuracy, and
is a model with 4 convolutional layers, each with 50 filters, and the filter size is 5*5, 3*3, 3*3, 3*3
for the convolutional layers, respectively. The trainable parameters is around 900,000, instead of
several or tens of millions in the previous two cases.
Table 2: Configurations and results for CNN on 69*69 Matrix
Configurations
Conv Layers
2, 3, 4
Filters
40, 50, 60, 70
Kernel Size 7*7, 5*5, 3*3
3.5
Best Results
CNN2 98.696%
CNN3 99.155%
CNN4 99.404%
D ISCUSSION
In previous sections, we present three kinds of models and the corresponding data transformations.
We select some of the trained models and plot their ROC curves, shown in Figure 4, and the corresponding AUCs are given in Table 3. The ROC curves for the selected models are consistent with
the previously reported testing accuracy. For this absorption bump detection application, we could
improve the sensitivity of the trained models by slightly decreasing the decision threshold. This
could result in slightly more false positive signals, but is more confident in capturing the potential
absorption bump events we are interested in.
true positive rate
1.0
Table 3: AUC of the ROC curves
0.9
alexnet
googlenet
convnet2
convnet3
convnet4_1
convnet4_2
convnet4_3
fc2
0.8
0.7
0.6
0.0
0.1
0.2
0.3
false positive rate
0.4
Figure 4: ROC curves for various models
6
Type
AlexNet
GoogLeNet
ConvNet2
ConvNet3
ConvNet4-1
ConvNet4-2
ConvNet4-3
FC2
AUC
0.99915
0.99898
0.99802
0.99873
0.99943
0.99947
0.99875
0.99455
Under review as a conference paper at ICLR 2016
Based on our experience with absorption bump detection application, there are a couple of things
worth notice for data preparation and model selection. When generating raw data, in order to eliminate the bias in the trained model, carefully designed criteria should be followed to cover the cases
as much as possible. Another necessary step before feeding the data into training is shuffling. The
raw data are generated in a systematical manner. Using non-shuffled data, localized characteristics
in the training batches could impede the convergence of the model. It could also cause diverged
performance on training and testing data set. We have experienced these problems at early stage.
The accuracy for the best model we get from non-shuffled data is as low as 88%.
From the presented details, we believe that Deep Learning based method for absorption bump detection is effective. With a trained model, Mg II information is not required to filter new observations for
candidates. The specifications for data generation and model configuration, and the trained model
are easy to share and reuse. We observe from our experience that the data generation could also be
modeled into a MapReduce-like programming model, significantly relieving users from heavy programming, but still benefiting from the power of distributed computing. In the following section, we
present our generalization and current work on providing such a Deep Learning based framework
for similar problems.
4
4.1
G ENERALIZATION : B OOSTING S CIENCE D ISCOVERY WITH D EEP
L EARNING
S CIENCE D ISCOVERY P ROBLEM
Similar as the absorption bump detection, a category of problems in other science field involves
the detection of a certain phenomenon among the background events. We call them the Science
Discovery Problem. These discoveries are building blocks for further scientific research. Scientists
rely on them to propose ideas, validate hypotheses, and prove theories. A typical science discovery
problem consists of two main components, the feature set, F = {fi |0 < i <= N }, and the relation
among the features, R. A science discovery problem can be expressed as:
D = {F|F ∼ Rtarget (P, C), Rtarget ⊂ R}
(3)
D is the targeted discoveries, P and C are parameters and constants. F is one event in F. Rtarget
defines the relations among the features that depict the targeted phenomenon. Rtarget and its complementary set Rctarget compose R.
With the F, R, Rtarget , scientists aim to pick the discoveries out of the observations. The mainstream method to perform the detection is rule-based filtering. The filtering rules are constructed
upon raw features or extracted high level features. High level feature extraction can be done by
numerically combining raw features, or by some fitting method, similar as the absorption bump
application. While scientists are concerned with the accuracy of the approximation theories and
continue improving their effectiveness, there are some issues with such methods.
Feature Engineering. Feature engineering is a time-consuming procedure. It also depends on a lot
of experience and numerous trial-and-error experiments. The complex and mysterious nature of the
science discovery problems also increase the difficulty of extracting meaningful high level features
to effectively capture the target event characteristics. Even with a carefully developed feature extraction technique with excellent theoretical explanation, real-world data could also offset the expected
effectiveness considering the existence of various noise and uncertainty during the data collection
phase.
Extra Dependency. In some science discovery cases, in order to get a better approximated representation for the targeted event, extra effort is required to obtain additional information. For example,
the approximated absorption bump representation requires emission and absorption redshifts. While
the emission redshift is included in the public data set, the absorption redshift is calculated based on
Mg II absorption lines. Similar requirements occur in other science fields. Two potential drawbacks
in such scenarios impede the science exploration. First, these extra dependencies incur considerable
amount of time and efforts. Second, they may not co-exist with the target discovery, thus, restricting
the exploration scope due to incomplete data sets.
7
Under review as a conference paper at ICLR 2016
Sharing and Collaboration. Information exchange and collaboration are two significant factors to
speed up the science discoveries. The theories and findings are well shared by researchers, but heavy
data processing are required to generate results with existing methods for each new observation.
Even though distributed computing technologies are pervasive and performance-ascendant, crossdiscipline knowledge requirement impedes its wide adoption.
4.2
D EEP L EARNING BASED M ETHOD
Inspired by the advance and practical effectiveness of Deep Learning, we propose a data-driven,
simulation-based method for science discovery problems. The proposed method is rooted in three
observations. First, Deep Learning has proven itself with practical success on classification and
recognition problems. Second, Deep Learning is more powerful when the relations among features
are complicated. This is often the case in science discovery problems. Finally, Deep Learning
requires a large amount of training data to be effective. Despite of the fact that real-world data
might not be adequate and they are sometimes biased (i.e., the targeted discoveries are normally
rare compared with the background events), we observe that it is relatively easy to simulate enough
eligible training samples with desired information embedded in. The control over the training data
sets ensures us to accurately capture the targeted phenomena. This both solves the biased training
issue and satisfies Deep Learning requirement for large-scale training samples.
Existing Knowledge
,
, ,
1
Raw
Data Set
2
Training
Samples
3
Train
Model
4
Trained Model
Additional
Knowledge
Raw
Data Set
6
Fine
Tuning
Trained Model
Real Data
Decision
5
Training
Samples
Model
Figure 5: Workflow for DeepDis. The mark on the top left corner means able-to-share.
The workflow of the proposed method is shown in Figure 5. There are three main steps: data preparation, model training, and model service. During data preparation, two operations are performed:
data generation (1) and data transformation (2). The approximation theory Rtarget and Rctarget ,
the constant variables C, and a set of range parameters P are used to generate enough labeled raw
data. The raw data can be represented by a series of key-value pairs. The data transformation phase
maps the raw data sets into desired formats for training, e.g., images or multi-dimensional matrix.
The transformed data sets are used in model training step to train a pre-defined model (3). After the
converge criteria are satisfied, the trained model will be published and ready for use (4). In model
service step, the trained model is used to give decisions for real-world data (5). The real-world data
set go through similar transformations as the simulated raw data sets. In science discovery problems,
various phenomena could be detected upon the same feature set. This inspires us to introduce the
fine tuning and transfer learning techniques into the workflow (6). Specifically, additional data could
be generated to further tune an already trained model. By substituting some layers in the previous
model and apply the fine tuning using extra data sets, a new model could be trained to identify the
new phenomenon.
The architecture of the proposed DeepDis system is shown in Figure 6. There are four layers in
DeepDis, namely user interface layer, control layer, runtime layer, and compute and storage resource
layer. From the beginning of the design, several main considerations are planted in the gene of
DeepDis. The first one is flexibility and easy-to-use. The second one is harnessing the power of
various computing and storage resources. The third one is sharing. We show how we emphasize on
these aspects in the following descriptions.
User Interface Layer. Two kinds of information are required from the users. The first part is the
data generation function and parameters. Users are requested to implement their own generation
functions. The function will take a dictionary of parameters, including constants and range parameters, and output all possible key-value pairs constrained by the ranges. It resembles the flat map
function in many programming languages. Data specifications also contains instructions on how to
8
Under review as a conference paper at ICLR 2016
Generation
Function
Parameter
Specs
Data
Specs
Model
Specs
User Interface
DeepDis Controller
Data
Coordinator
Model
Coordinator
Generation
Training
Service
Coordinator
Control
Detection
Transformation
Sharing
DeepDis
Caffe
Runtime
Compute & Storage
Resources
Figure 6: Architecture Overview for DeepDis
transform raw data sets into ready-to-process data format. The second part is the model configuration. The specifications are passed to DeepDis by json/prototxt. Only these configurations are
required from the users. This makes DeepDis easy-to-use even for scientists with less knowledge
about distributed computing or Deep Learning.
Control Layer. This layer is responsible for understanding users requests and translate them into
computing tasks. Specifically, DeepDis has a controller and three coordinators, Data Coordinator,
Model Coordinator and Service Coordinator, to fulfill this. The controller divides the requests into
several inter-dependent tasks and use DAG to schedule. The coordinators are responsible for the
designated tasks: preparing, dividing, and distributing them to the underlying runtime. This layer
also opens pluggable and extensible interfaces for advanced users to provide new data format, transformation functions, etc.
Runtime Layer and Compute and Storage Resources Layer. DeepDis integrates state-of-the-art
distributed in-memory computing framework, Spark (Zaharia et al. (2012)), and the Caffe (Jia et al.
(2014)) Deep Learning toolbox to efficiently execute the data generation and model training tasks.
Spark will be extended to incorporate adaptations from the control layer and to collaborate with
Caffe for distributed model training. For the Compute and Storage Resources Layer, different types
of resources will be utilized to achieve efficient data processing, model training, and sharing.
5
C ONCLUSION
In this paper, we presented a Deep Learning based method for detecting the absorption bump. We
discussed in detail about how we generated and transformed the training data according to the selected models. For different models, we presented the results. In order to get a sense of what the
models have learned, we chose the CNN based AlexNet model and provided the visualizations of
filters, feature maps and reconstructed maximum activation input images. The resulted images prove
that the model is identifying the absorption bump, rather than any other background noise. With the
success in applying Deep Learning based method for the absorption bump application, we generalized the methodology to the broad science discovery problems, where Deep Learning models can
be effectively trained upon sufficient amount of data. We also showed our ongoing work in building
a specialized system, DeepDis, to support the proposed method. Distributed data processing techniques are used to automatically handle data generation, data transformation, model training and
sharing. DeepDis is designed to provide science discovery as a service for researchers without too
much knowledge about distributed computing and Deep Learning. With DeepDis, we hope to boost
the science discovery process.
R EFERENCES
Berk, Daniel E Vanden, Richards, Gordon T, Bauer, Amanda, Strauss, Michael A, Schneider, Donald P, Heckman, Timothy M, York, Donald G, Hall, Patrick B, Fan, Xiaohui, Knapp, GR, et al.
9
Under review as a conference paper at ICLR 2016
Composite quasar spectra from the sloan digital sky survey. The Astronomical Journal, 122(2):
549, 2001.
Bernstein, Max P, Elsila, Jamie E, Dworkin, Jason P, Sandford, Scott A, Allamandola, Louis J,
and Zare, Richard N. Side group addition to the polycyclic aromatic hydrocarbon coronene by
ultraviolet photolysis in cosmic ice analogs. The Astrophysical Journal, 576(2):1115, 2002.
Caro, GM Munoz, Meierhenrich, UJ, Schutte, WA, Barbier, B, Segovia, A Arcones, Rosenbauer,
H, Thiemann, WH-P, Brack, A, and Greenberg, JM. Amino acids from ultraviolet irradiation of
interstellar ice analogues. Nature, 416(6879):403–406, 2002.
Draine, BT. Interstellar dust grains. Annu. Rev. Astron. Astrophys, 41:241–89, 2003.
Fitzpatrick, Edward L and Massa, Derck. An analysis of the shapes of ultraviolet extinction curves.
iii-an atlas of ultraviolet extinction curves. The Astrophysical Journal Supplement Series, 72:
163–189, 1990.
Gould, Robert J and Salpeter, Edwin E. The interstellar abundance of the hydrogen molecule. i.
basic processes. The Astrophysical Journal, 138:393, 1963.
Hinton, Geoffrey E and Salakhutdinov, Ruslan R. Reducing the dimensionality of data with neural
networks. Science, 313(5786):504–507, 2006.
Hollenbach, David and Salpeter, EE. Surface recombination of hydrogen molecules. The Astrophysical Journal, 163:155, 1971.
Jia, Yangqing, Shelhamer, Evan, Donahue, Jeff, Karayev, Sergey, Long, Jonathan, Girshick, Ross,
Guadarrama, Sergio, and Darrell, Trevor. Caffe: Convolutional architecture for fast feature embedding. In Proceedings of the ACM International Conference on Multimedia, pp. 675–678.
ACM, 2014.
Jiang, P, Ge, J, Prochaska, JX, Kulkarni, VP, Lu, HL, and Zhou, HY. A Dusty Mg II Absorber
Associated with the Quasar SDSS J003545. 13+ 011441.2. The Astrophysical Journal, 720(1):
328, 2010a.
Jiang, Peng, Ge, Jian, Prochaska, J Xavier, Wang, Junfeng, Zhou, Hongyan, and Wang, Tinggui.
High Dust Depletion in two Intervening Quasar Absorption Line Systems with the 2175 Å Extinction Bump at z˜ 1.4. The Astrophysical Journal, 724(2):1325, 2010b.
Jiang, Peng, Ge, Jian, Zhou, Hongyan, Wang, Junxian, and Wang, Tinggui. Toward Detecting
the 2175 Å Dust Feature Associated with Strong High-redshift Mg II Absorption Lines. The
Astrophysical Journal, 732(2):110, 2011.
Krizhevsky, Alex, Sutskever, Ilya, and Hinton, Geoffrey E. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems, pp. 1097–1105,
2012.
LeCun, Yann and Bengio, Yoshua. Convolutional networks for images, speech, and time series. The
handbook of brain theory and neural networks, 3361(10), 1995.
LeCun, Yann, Bengio, Yoshua, and Hinton, Geoffrey. Deep learning. Nature, 521(7553):436–444,
2015.
Li, Aigen and Greenberg, J Mayo. In dust we trust: an overview of observations and theories of
interstellar dust. In Solid State Astrochemistry, pp. 37–84. Springer, 2003.
Peeters, Els, Allamandola, LJ, Hudgins, DM, Hony, S, and Tielens, AGGM. The unidentified infrared features after iso. arXiv preprint astro-ph/0312184, 2003.
Quider, Anna M, Nestor, Daniel B, Turnshek, David A, Rao, Sandhya M, Monier, Eric M, Weyant,
Anja N, and Busche, Joseph R. The pittsburgh sloan digital sky survey mg ii quasar absorptionline survey catalog. The Astronomical Journal, 141(4):137, 2011.
10
Under review as a conference paper at ICLR 2016
Schneider, Donald P, Richards, Gordon T, Hall, Patrick B, Strauss, Michael A, Anderson, Scott F,
Boroson, Todd A, Ross, Nicholas P, Shen, Yue, Brandt, WN, Fan, Xiaohui, et al. The sloan digital
sky survey quasar catalog. v. seventh data release. The Astronomical Journal, 139(6):2360, 2010.
Szegedy, Christian, Liu, Wei, Jia, Yangqing, Sermanet, Pierre, Reed, Scott, Anguelov, Dragomir,
Erhan, Dumitru, Vanhoucke, Vincent, and Rabinovich, Andrew. Going deeper with convolutions.
arXiv preprint arXiv:1409.4842, 2014.
Zaharia, Matei, Chowdhury, Mosharaf, Das, Tathagata, Dave, Ankur, Ma, Justin, McCauley, Murphy, Franklin, Michael J, Shenker, Scott, and Stoica, Ion. Resilient distributed datasets: A faulttolerant abstraction for in-memory cluster computing. In Proceedings of the 9th USENIX conference on Networked Systems Design and Implementation, pp. 2–2. USENIX Association, 2012.
Zhu, Guangtun and Ménard, Brice. The jhu-sdss metal absorption line catalog: redshift evolution
and properties of mg ii absorbers. The Astrophysical Journal, 770(2):130, 2013.
11
| 9cs.NE
|
1
Adaptive-Robust Control of a Class of Uncertain
Nonlinear Systems Utilizing Time-Delayed Input
and Position Feedback
arXiv:1509.07314v3 [cs.SY] 11 Feb 2016
Spandan Roy and Indra Narayan Kar, Senior Member, IEEE
Abstract
In this paper, the tracking control problem of a class of Euler-Lagrange systems subjected to unknown uncertainties is addressed
and an adaptive-robust control strategy, christened as Time-Delayed Adaptive Robust Control (TARC) is presented. The proposed
control strategy approximates the unknown dynamics through time-delayed logic, and the switching logic provides robustness
against the approximation error. The novel adaptation law for the switching gain, in contrast to the conventional adaptive-robust
control methodologies, does not require either nominal modelling or predefined bounds of the uncertainties. Also, the proposed
adaptive law circumvents the overestimation-underestimation problem of switching gain. The state derivatives in the proposed
control law is estimated from past data of the state to alleviate the measurement error when state derivatives are not available
directly. Moreover, a new stability notion for time-delayed control is proposed which in turn provides a selection criterion for
controller gain and sampling interval. Experimental result of the proposed methodology using a nonholonomic wheeled mobile
robot (WMR) is presented and improved tracking accuracy of the proposed control law is noted compared to time-delayed control
and adaptive sliding mode control.
Index Terms
Adaptive-robust control, Euler-Lagrange system, time-delayed control, state derivative estimation, wheeled mobile robot.
I. I NTRODUCTION
A. Background and Motivation
ESIGN of an efficient controller for nonlinear systems subjected to parametric and nonparametric uncertainties has always
been a challenging task. Among many other approaches, Adaptive control and Robust control are the two popular control
strategies that researchers have extensively employed while dealing with uncertain nonlinear systems. In general, adaptive
control uses predefined parameter adaptation laws and equivalence principle based control law which adjusts the parameters of
the controller on the fly according to the pertaining uncertainties [1]. However, this approach has poor transient performance and
online calculation of the unknown system parameters and controller gains for complex systems is computationally intensive [2].
Whereas, robust control aims at tackling the uncertainties of the system within an uncertainty bound defined a priori. It reduces
computation complexity to a great extent for complex systems compared to adaptive control as exclusive online estimation
of uncertain parameters is not required [3]. However, nominal modelling of the uncertainties is necessary to decide upon
their bounds, which is not always possible. Again, to increase the operating region of the controller, often higher uncertainty
bounds are assumed. This in turn leads to problems like higher controller gain and consequent possibility of chattering for the
switching law based robust controller like Sliding Mode Control (SMC). This in effect reduces controller accuracy [4]. Higher
order sliding mode [5] can alleviate the chattering problem but prerequisite of uncertainty bound still exists.
Time-Delayed Control (TDC) is utilized in [6] to implement state derivative feedback for enhancing stability margin of SISO
linear time invariant (LTI) systems. In [7], [8], [26]-[27], [34], [35], TDC is used to provide robustness against uncertainties.
In this process, all the uncertain terms are represented by a single function which is then approximated using control input and
state information of the immediate past time instant. The advantage of this robust control approach in uncertain systems is that
it reduces the burden of tedious modelling of complex system to a great extent. In spite of this, the unattended approximation
error, commonly termed as time-delayed error (TDE) causes detrimental effect to the performance of the closed system and its
stability. In this front, a few work have been carried out to tackle TDE which includes internal model [9], gradient estimator
[10], ideal velocity feedback [11], nonlinear damping [12] and sliding mode based approach [13]-[14]. The stability of the
closed loop system [9]-[11], [26]-[27], depends on the boundedness of TDE as shown in [7]. This method approximates the
continuous time closed loop system in a discrete form without considering the effect of discretization error. Again, the stability
criterion mentioned in [7] restricts the allowable range of perturbation and thus limits controller working range. Stability of the
system in [13] is established in frequency domain, which makes the approach inapplicable to the nonlinear systems. Moreover,
the controllers designed in [12], and [14], [35] require nominal modelling and upper bound of the TDE respectively which is
not always possible in practical circumstances. Also, to the best knowledge of the authors, controller design issues such as
D
2
selection of controller gains and sampling interval to achieve efficient performance is still an open problem. In contrast to TDC,
works reported in [28]-[30] use low pass filter to approximate the unknown uncertainties and disturbances. However, frequency
range of system dynamics and external disturbances are required to determine the time constant of the filter. Furthermore, the
order of the low pass filter needs to be adjusted according to order of the disturbance to maintain stability of the controller.
Considering the individual limitations of adaptive and robust control, recently global research is reoriented towards adaptiverobust control (ARC) where switching gain of the controller is adjusted online. The series of publications [2], [15]-[20] regarding
ARC, estimates the uncertain terms online based on predefined projection function, but predefined bound on uncertainties is
still a requirement. The work reported in [21], [33] attempts to estimate the maximum uncertainty bound but the integral
adaptive law makes the controller susceptible to very high switching gain and consequent chattering [34]. The adaptive sliding
mode control (ASMC) as presented in [22]-[23] proposed two laws for the switching gain to adapt itself online according
to the incurred error. In the first adaptive law, the switching gain decreases or increases depending on a predefined threshold
value. However, until the threshold value is achieved, the switching gain may still be increasing (resp. decreasing) even if
tracking error decreases (resp. increases) and thus creates overestimation (resp. underestimation) problem of switching gain
[24]. Moreover, to decide the threshold value the maximum bound of the uncertainty is required. For the second adaptive law,
the threshold value changes online according to switching gain. Yet, nominal model of the uncertainties is needed for defining
the control law. This limits the adaptive nature of the control law and applicability of the controller.
B. Problem Definitions and Contributions
In this paper, three specific related problems on TDC have been dealt with and the corresponding solutions to the same
which are also the contributions of this paper are summarized below:
• Problem 1: The stability analysis of TDC, as provided in [7], [26]-[27], [34], [35], approximates the continuous time
system in discrete time domain without considering the effects of discretization error. Again, choice of the delay time and
its relation with the controller gains is still an open problem.
In this paper, a new stability analysis for TDC, based on the Lyapunov-Krasvoskii method, is provided in continuous time
domain. Furthermore, through the proposed stability approach, a relation between the sampling interval and controller
gain is established.
• Problem 2: The TDC reported in ([7], [26]-[27], [34], [35]), velocity and acceleration feedback are necessary to compute
the control law. While in [6], only velocity feedback is required and acceleration term is approximated numerically using
time delay. However, in many applications velocity and acceleration feedback are not available explicitly and numerical
approximation of these terms invokes measurement error.
As a second contribution of this paper, Filtered Time-Delayed Control (F-TDC) control law is formulated where only
position feedback is sufficient while velocity and acceleration terms are estimated using past and present position
information to curb the effect measurement error. Stability analysis of the proposed F-TDC is provided which also
maintains the relation between controller gains and sampling interval.
• Problem 3: Robustness property against TDE is essential to achieve good tracking accuracy. The robust controllers
reported in literature, either requires nominal model of the uncertainties ([22]-[23]) or its predefined bound ([2], [12],
[14], [15]-[20]). So, it is required to devise a control law which would avoid any prior knowledge of the uncertainties
while providing robustness against TDE.
Towards the last contribution of this article, an adaptive-robust control strategy, Time-Delayed Adaptive Robust Control
(TARC) has been formulated for a class of uncertain Euler-Lagrange systems. The proposed control law approximates
uncertainties by time-delayed logic and provides robustness against the TDE, arising from time-delayed logic based estimation, by switching control. The novel adaptive law, presented here, aims at overcoming the overestimation-underestimation
problem of the switching gain without any prior knowledge of uncertainties. The proposed adaptive law provides flexibility
to the control designer to select any suitable error function according to the application requirement while maintaining
similar system stability notion.
As a proof of concept, experimental validation of the proposed control methodology is provided using the "PIONEER-3"
nonholonomic WMR in comparison to TDC [7] and ASMC [22]-[23].
C. Organization
The article is organized as follows: a new stability analysis of TDC along with its design issues is first discussed in Section
II. This is followed by the proposed adaptive-robust control methodology and its detail analysis. Section III presents the
experimental results of the proposed controller and its comparison with TDC and ASMC. Section IV concludes the entire
work.
D. Notations
The following notations are assumed for the entirety of the paper: any variable µ delayed by an amount h as µ(t − h), is
denoted as µh ; λmin (·) and ||·|| represent minimum eigen value and Euclidean norm of the argument respectively; I represents
identity matrix.
3
II. C ONTROLLER D ESIGN
A. Time-Delayed Control: Revisited
In general, an Euler-Lagrange system with second order dynamics, devoid of any delay, can be written as,
M (q)q̈ + N (q, q̇) = τ (t),
(1)
where, q(t) ∈ Rn is the system state, τ (t) ∈ Rn is the control input, M (q) ∈ Rn×n is the mass/inertia matrix and N (q, q̇) ∈ Rn
denotes combination of other system dynamics terms based on system properties. In practice, it can be assumed that unmodelled
dynamics and disturbances is subsumed by N . The control input is defined to be,
τ = M̂ u + N̂ ,
(2)
where, u is the auxiliary control input, M̂ and N̂ are the nominal values of M and N respectively. To reduce the modelling
effort of the complex systems, N̂ can be approximated from the input-output data of previous instances using the time-delayed
logic ([7], [26]-[27]) and the system definition (1) as,
N̂ (q, q̇) ∼
= N (qh , q̇h ) = τh − M̂ (qh )q̈h ,
(3)
where, h > 0 is a fixed small delay time. Substituting (2) and (3) in (1), the system dynamics is converted into an input as
well state delayed dynamics as,
M̂ (q)q̈ + N̄ (q, q̇, q̈, q̈h ) = τh .
(4)
where N̄ = (M − M̂ )q̈ + M̂h q̈h − M̂ u + N .
Let, q d (t) be the desired trajectory to be tracked and e1 (t) = q(t) − q d (t) is the tracking error. The auxiliary control input
u is defined in the following way,
u(t) = q̈ d (t) − K2 ė1 (t) − K1 e1 (t),
(5)
where, K1 and K2 are two positive definite matrices with appropriate dimensions. Putting (5) and (2) in (4), following error
dynamics is obtained,
ë1 = −K2 ė1h − K1 e1h + σ1 ,
(6)
where, σ1 = (M̂ −1 M̂h − I)uh + M̂ −1 (N̂h − N̄ ) + q̈hd − q̈ d and can be treated as overall uncertainty. Further, (6) can be written
in state space form as,
ė = A1 e + B1 eh + Bσ1 ,
(7)
R0
0
e
0 I
0
0
ė(t + θ)dθ, where the
,B =
where, e = 1 , A1 =
. Noting that, e(t − h) = e(t) −
, B1 =
I
0 0
−K1 −K2
ė1
−h
derivative inside the integral is with respect to θ, the error dynamics (7) is modified as,
ė(t) = Ae(t) − B1
Z0
ė(t + θ)dθ + Bσ1 ,
(8)
−h
where, A = A1 + B1 . It is assumed that the choice controller gains K1 and K2 makes A Hurwitz which is always
possible. Also, it is assumed that the unknown uncertainties are bounded. In this paper, a new stability criterion, based on the
Lyapunov-Krasvoskii method, is presented through Theorem 1 which addresses the issues defined in Problem 1.
Theorem 1. The system (4) employing the control input (2), having auxiliary control input (5) is UUB if the controller gains
and delay time is selected such that the following condition holds:
#
"
2
0
Q − E − (1 + ξ) hβ D
>0
(9)
Ψ=
2
0
(ξ − 1) hβ D
where, E = βP B1 (A1 D−1 AT1 + B1 D−1 B1T + D−1 )B1T P , ξ > 1 and β > 0 are scalar, and P > 0 is the solution of the
Lyapunov equation AT P + P A = −Q for some Q > 0.
Proof. Let us consider the following Lyapunov function:
V (e) = V1 (e) + V2 (e) + V3 (e) + V4 (e),
(10)
4
where,
V1 (e) = eT P e
Z Z
h 0 t T
V2 (e) =
e (ψ)De(ψ)dψdθ
β −h t+θ
Z Z
h 0 t T
V3 (e) =
e (ψ − h)De(ψ − h)dψdθ
β −h t+θ
Z
h2 t T
e (ψ)De(ψ)dψ
V4 (e) = ξ
β t−h
(11)
(12)
(13)
(14)
Using (8), the time derivative of V1 (e) yields,
V̇1 (e) = −eT Qe − 2eT P B1
Z
0
ė(t + θ)dθ + 2ŝT σ1
(15)
−h
where, ŝ = B T P e. Again using (7),
Z 0
Z
T
T
−2e P B1
ė(t + θ)dθ = −2e P B1
−h
0
[A1 e(t + θ) + B1 e(t − h + θ) + Bσ1 (t + θ)]dθ,
(16)
−h
For any two non zero vectors z1 and z2 , there exists a scalar β > 0 and matrix D > 0 such that the following inequality holds,
± 2z1T z2 ≤ βz1T D−1 z1 + (1/β)z2T Dz2 .
Again, using Jensen’s inequality the following inequality holds [32],
Z
Z 0
Z 0
1 0 T
e (ψ)dψD
e(ψ)dψ.
eT (ψ)De(ψ)dψ ≥
h −h
−h
−h
(17)
(18)
Applying (17) and (18) to (16) the following inequalities are obtained,
Z 0
Z
Z 0
1 0 T
e(t + θ)dθ
e (t + θ)dθD
− 2eT P B1 A1
e(t + θ)dθ ≤ βeT P B1 A1 D−1 AT1 B1T P e +
β −h
−h
−h
Z
h 0 T
≤ βeT [P B1 A1 D−1 AT1 B1T P ]e +
e (t + θ)De(t + θ)dθ
(19)
β −h
Z
Z 0
Z 0
1 0 T
e (t − h + θ)dθD
e(t − h + θ)dθ
− 2eT P B1 B1
e(t − h + θ)dθ ≤ βeT P B1 B1 D−1 B1T B1T P e +
β −h
−h
−h
Z
h 0 T
≤ βeT [P B1 B1 D−1 B1T B1T P ]e +
e (t − h + θ)De(t − h + θ)dθ
(20)
β −h
Z 0
Z
Z 0
1 0
− 2eT P B1
[Bσ1 (t + θ)]dθ ≤ βeT P B1 D−1 B1T P e +
(Bσ1 (t + θ))T dθD
Bσ1 (t + θ)]dθ
β −h
−h
−h
Z
h 0
−1 T
T
(Bσ1 (t + θ))T Dσ1 (t + θ)dθ
(21)
≤ βe P B1 D B1 P e +
β −h
Since D > 0, we can write D = D̄T D̄ for some D̄ > 0. Then, assuming the uncertainties to be square integrable within the
delay, let there exists a scalar Γ1 > 0 such that the following inequality holds:
Z 0
h
(Bσ1 (t + θ))T D̄T D̄Bσ1 (t + θ) dθ ≤ Γ1 .
(22)
β −h
Again,
Z
h 0 T
h2 T
e De −
e (t + θ)De(t + θ)dθ
V̇2 (e) =
β
β −h
Z
h2 T
h 0 T
V̇3 (e) =
eh Deh −
e (t − h + θ)De(t − h + θ)dθ
β
β −h
h2
V̇4 (e) = ξ (eT De − eTh Deh )
β
(23)
(24)
(25)
Substituting (19)-(22) into (15) and adding it with (23)-(25) yields,
V̇ (e) ≤ −ēT Ψē + Γ1 + 2ŝT σ1 ,
(26)
5
T
where, ē = eT eTh . Let controller gains K1 , K2 and delay time h are selected to make Ψ > 0. One can find a positive
scalar ι such that ||ŝ|| ≤ ι||||ē||. Then, V̇ (e) < 0 would be established if λmin (Ψ)||ē||2 > Γ1 + 2ι||σ1 ||||ē||. Thus (4) would
be UUB with the ultimate bound,
s
Γ1
(27)
+ γ12 = ̟0 .
||ē|| = γ1 +
λmin (Ψ)
ι||σ1 ||
where, γ1 = λmin
(Ψ) . Let Ξ denote the smallest level surface of V containing the ball B̟0 with radius ̟0 centred at ē = 0.
For initial time t0 , if ē(t0 ) ∈ Ξ then the solution remains in Ξ. If ē(t0 ) ∈
/ Ξ then V decreases as long as ē(t) ∈
/ Ξ. The time
required to reach ̟0 is zero when ē(t0 ) ∈ Ξ, otherwise, while ē(t0 ) ∈
/ Ξ the finite time tr0 to reach ̟0 , for some c0 > 0, is
given by [25],
tr0 − t0 ≤ (||ē(t0 )|| − ̟0 )/c0 where V̇ (t) ≤ −c0
Remark 1. : Since E depends on the controller gains, (9) provides a selection criterion for the choice of delay h for given
controller gains and Q. This design issue was previously unaddressed in the literature. Moreover, the approximation error
(N̂ − N ), as in (3), would reduce for small values of h. However, h cannot be selected smaller than the sampling interval
because, the input output data is only available at sampling intervals. So, the lowest possible selection of h is the sampling
interval. Again, choice of sampling interval is governed by the corresponding hardware response time, computation time etc.
Hence, the proposed stability approach provides a necessary step for the selection of sampling interval for given controller
gains or vice-versa.
B. Filtered Time-Delayed Control (F-TDC)
It can be noticed from (3) and (5) that state derivatives are necessary to compute the control law of TDC. However, in
many circumstances, only q is available amongst q, q̇, q̈. Under this scenario, a new control strategy F-TDC is proposed, which
estimates the state derivatives from the state information of past instances [31]. Before proposing the control structure of
F-TDC, the following two Lemmas are stated which are instrumental for formulation as well as stability analysis of F-TDC.
Lemma 1 ([31]). : For time t ≥ ς, the j-th order time derivative of the Λ-th degree polynomial q in (4) can be computed in
the following way,
Z 0
q̂ (j) (t) =
Ωj (ς, ψ)q(t + ψ)dψ
(28)
−ς
where, ς > 0 is a prespecified scalar and
Ωj (ς, ψ) =
k
Λ
−ψ
(−1)k (Λ + 1 + k)!
(Λ + 1 + j)! X
.
ς
ς (j+1) j!(Λ − j)! k=0 (j + k + 1)(Λ − k)!(k!)2
(29)
Lemma 2. For any non zero vector ϑ(ψ), constant matrix F > 0 the following relation holds,
Z 0 Z 0
Z 0 Z 0
Z 0 Z 0
1
T
T
ϑ (ψ)F ϑ(ψ)dψdθ ≥
ϑ (ψ)dψdθ F
ϑ(ψ)dψdθ
hς
−h −ς
−h −ς
−h −ς
(30)
The structure of F-TDC is similar to (2), except, the auxiliary control input u and N̂ in (2) selected in the following way,
u(t) = q̈ d (t) − K1 e1 (t) − K2 ê˙ 1 (t)
∼ Nh = uh − M̂h q̂¨h ,
N̂ (t) =
(31)
(32)
where, ê˙ 1 = q̂˙ − q̇ d . q̂˙ and q̂¨ are evaluated from (28) and (29). The stability of the system (4) employing F-TDC is derived in
the sense of Uniformly Ultimately Bounded (UUB) notion as stated in Theorem 2.
Theorem 2. The system (4) employing the control input (2), having the auxiliary input (31) and (32) is UUB if K1 , K2 , h and
ς are selected such that the following condition holds:
2
Q − Ē − (1 + ξ) hβ D
P B̆
P B̄
2
0
B̆ T P
(ξ − 1) hβ D − F̄
B̄ T P
0
L
=Θ>0
where, Ē = βP B1 (A1 D−1 AT1 + B1 D−1 B1T + D−1 + B̄D−1 B̄ T )B1T P , F̄ =
Ω1 (ς, ψ), B̄ = B K2 0 , B̆ = B 0 K2 .
Proof. The proof is provided in Appendix B.
(33)
2
( hβ D
+ L)ς
R0
−ς
A2d (ψ)dψ, L > 0, Ad (ψ) =
6
C. Adaptive-Robust Control: Related Work
It can be observed that TDE degrades tracking performance of both the TDC and F-TDC in the face of uncertainties. The
control methods that attempt to counter uncertainties, as reported in [2], [12], [14], [15]-[20], requires predefined bound on
the uncertainties which is not always possible in practical circumstances. To circumvent this situation Adaptive Sliding Mode
Control (ASMC) was proposed in [22]-[23]. The control input of ASMC is given by,
τ = Σ−1
n (−κn + ∆us ),
(34)
where, Σn and κn is the nominal values of Σ and κ, and ∆us is the switching control input. For a choice of sliding surface
s̄, Σ and κ is defined as follows:
s̄˙ = Σ + κ∆us ,
(35)
The switching control ∆us is calculated as
∆us = −ĉ
ĉ˙ =
(
s̄
||s̄||
c̄||s̄||sgn(||s̄|| − ρ)
γ
(36)
ĉ > γ,
ĉ ≤ γ,
(37)
where, ĉ is the switching gain, c̄ > 0 is a scalar adaptive gain, ρ > 0 is a threshold value, γ > 0 is small scalar to always
keep ĉ positive. Evaluation of ρ can be done in two ways [22]:
ρ=̺
or,
ρ(t) = 4ĉ(t)ts ,
(38)
(39)
where, ̺ > 0 is a scalar, ts is the sampling interval. The choice (38) requires predefined bound of uncertainties. It can be
noted from (37) that even if ||s̄|| decreases (resp. increases), unless it falls below (resp. goes above) ρ switching gain does
not decrease (resp. increase). This causes overestimation (resp. underestimation) of switching gain and controller accuracy is
compromised. Again, improper and low choice of ρ may lead to very high switching gain and consequent chattering. On the
other hand, method (39) assumes that the nominal value of the uncertainties are always greater than the perturbations. This
assumption may not hold due to the effect of unmodelled dynamics and thus, necessitates rigorous nominal modelling of the
uncertainties in N to design the control law. Either of the two situations, i.e. bound estimation or uncertainty modelling, is
not always feasible in practical circumstances and consequently compromises the adaptive nature of the controller.
D. Time-Delayed Adaptive Robust Control
Considering the limitations of the existing controllers that aim at negotiating the uncertainties, as discussed earlier, a novel
adaptive-robust control law, named Time-Delayed Adaptive Robust Control (TARC) is proposed in this endeavour, which
neither requires the nominal model nor any predefined bound of the uncertainties as well as eliminates the overestimationunderestimation problem of switching gain. The structure of the control input of TARC is similar to (2) and N̂ is also evaluated
according to (32). However, the auxiliary control input u is selected as below,
u = û + ∆u.
(40)
û is the nominal control input and selected as similar to (31). ∆u is the switching control law which is responsible for
negotiating the TDE and it is defined as below,
(
s
if k s k≥ ǫ,
−αĉ(e, t) ksk
(41)
∆u =
s
if k s k< ǫ,
−αĉ(e, t) ǫ
T
where, s = B T P e1 ê˙ 1
and ǫ > 0 is a small scalar. The following novel adaptive control law for evaluation of ĉ is
proposed:
ĉ > γ, f (e) > 0
||s||
˙ĉ = −||s||
(42)
ĉ > γ, f (e) ≤ 0
γ
ĉ ≤ γ,
where, α > 0 is a scalar adaptive gain and ǫ > 0 represents a small scalar, f (e) is a suitable function of error defined by
the designer. Here, it is selected as f (e) = ||s(t)|| − ||sh ||. According to the adaptive law (42) and present choice of f (e),
ĉ increases (resp. decreases) whenever error trajectories move away from (resp. close to ) ||s|| = 0 The advantages of the
proposed TARC can be summarized as follows:
7
TARC reduces complex system modelling effort as only the knowledge of M̂ suffices the controller design since N along
with the uncertainties is approximated using the time-delayed logic as in (32). This in turn reduces the tedious modelling
effort of complex nonlinear systems.
• Evaluation of switching gain does not require either of the nominal model or predefined bound of the uncertainties and
also removes the overestimation-underestimation problem.
• State derivatives are not required to compute the control law explicitly, as they are evaluated from the past state information
using (28) and (29).
The stability of the system (4) employing TARC is analysed in the sense of UUB as stated in Theorem 3.
•
Assumption 1. Let, ||σ1 || ≤ c. Here, c is an unknown scalar quantity. Knowledge of c, however, is only required for stability
analysis but not to compute control law.
Theorem 3. The system (4) employing (2), (40) and having the adaptive law (42) is UUB, provided the selection of K1 , K2 , h
and ς holds condition (33).
Proof. : Let us define the Lyapunov functional as,
Vr (e) = Vf (e) + (ĉ − c)2 ,
where, Vf (e) is defined in (53). Again, putting (40) in (4) the the error dynamics becomes,
Z 0
ė = A1 e + B1 eh − B̄
Ad (ψ)e(t − h + ψ)dψ + Bσ,
(43)
(44)
−ς
where, σ = ∆uh + σ1 . Also following similar steps while proving Theorem 2 (provided in Appendix B) we have,
V̇f (e) ≤ −eTf Θef + Γ + 2ŝT (∆u + σ1 ) + 2ŝT Υ,
(45)
R
0
where, ef is defined in Appendix B, Γ ≥ βh −h (Bσ(t + θ))T D̄T D̄Bσ(t + θ) dθ is a positive scalar, Υ = ∆uh − ∆u.
Let us define the following,
T
ŝ = s + ∆s where ∆s = B T P 0 (ė1 − ê˙ 1 )T .
(46)
Evaluating the structure of s and ∆s one can find two positive scalars ι2 , ι3 such that ||s|| ≤ ι2 ||ef ||, ||∆s|| ≤ ι3 ||ef ||. Using
(43) the stability analysis for (4) employing TARC is carried out for the following various cases.
Case (i): f (e) > 0, ĉ > γ, ||s|| ≥ ǫ.
Utilizing (41), (42) and (45) we have,
s
+ σ1 ) + 2ŝT Υ + 2(ĉ − c)||s||
V̇r (e) ≤ −eTf Θef + Γ + 2ŝT (−αĉ
||s||
sT s
∆sT s
= −eTf Θef + Γ − 2αĉ
− 2αĉ
+ 2ŝT σ1 + 2ŝT Υ + 2(ĉ − c)||s||
||s||
||s||
≤ −λmin (Θ)||ef ||2 − 2(α − 1)ĉ||s|| + Γ + 2||s||||Υ|| + 2(αĉ + c + ||Υ||)||∆s||
(47)
So, for α > 1, V̇r (e) < 0 would be established if λmin (Θ)||ef ||2 > Γ + 2||s||||Υ|| + 2(αĉ + c + ||Υ||)||∆s||. Thus, using the
relation ||s|| ≤ ι2 ||ef ||, ||∆s|| ≤ ι3 ||ef ||, the system would be UUB with the following ultimate bound
s
Γ
+ µ21 = ̟1 .
(48)
||ef || = µ1 +
λmin (Θ)
3 (αĉ+c+||Υ||)
where, µ1 = ι2 ||Υ||+ι
.
λmin (Θ)
Case (ii): f (e) ≤ 0, ĉ > γ, ksk ≥ ǫ.
Again, utilizing (42) for Case (ii),
s
+ σ1 ) + 2ŝT Υ − 2(ĉ − c)||s||
||s||
≤ −λmin (Θ)||ef ||2 + (4c − 2(α + 1)ĉ + 2||Υ||)||s|| + Γ + 2(αĉ + c + ||Υ||)||∆s||.
V̇r (e) ≤ −eTf Θef + Γ + 2ŝT (−αĉ
(49)
V̇r (e) < 0 would be achieved if λmin (Θ)||ef ||2 > Γ + (4c − 2(α + 1)ĉ + 2||Υ||)||s|| + 2(αĉ + c + ||Υ||)||∆s|| and system
would be UUB having following ultimate bound,
s
Γ
+ µ22 = ̟2 .
(50)
||ef || = µ2 +
λmin (Θ)
where, µ2 =
ι2 (2c−(α+1)ĉ+||Υ||)+ι3 (αĉ+c+||Υ||)
.
λmin (Θ)
8
Case (iii): ĉ ≤ γ, ksk ≥ ǫ.
Since ĉ ≤ γ we have (ĉ − c)γ ≤ γ 2 − cγ ≤ γ 2 . Using the adaptive law (42), for Case (iii) we have,
s
V̇r (e) ≤ −eTf Θef + Γ + 2ŝT (−αĉ
+ σ1 ) + 2(ĉ − c)γ + 2ŝT Υ
||s||
≤ −λmin (Θ)||ef ||2 + Γ + 2(c − αĉ + ||Υ||)||s|| + 2(αĉ + c + ||Υ||)||∆s|| + 2γ 2 .
Similarly, as argued earlier the system would be UUB with the following ultimate bound,
s
(Γ + 2γ 2 )
||ef || = µ3 +
+ µ23 = ̟3 .
λmin (Θ)
where, µ3 =
(51)
(52)
ι2 (c−αĉ+||Υ||)+ι3 (αĉ+c+||Υ||)
.
λmin (Θ)
Remark 2. The performance of TARC can be characterized by the various error bounds under various conditions. It can be
noticed that low value of h and high value of α would result in better accuracy. However, too large α may result in high
control input. Also, one may choose different values of α for ||s|| > ||sh || and ||s|| ≤ ||sh ||. Moreover, it is to be noticed that
the stability notion of TARC is invariant to the choice of f (e) and thus provides the designer the flexibility to select a suitable
f (e) according to the application requirement.
III. C ONCLUSION
Selection of the controller gain and sampling interval is crucial for the performance of TDC and this design issue is
addressed in this paper through a new stability approach. A bound on the delay is derived to select a suitable sampling interval.
A new control approach, F-TDC is devised where the state derivatives are estimated from the previous state information.
Moreover, a novel adaptive-robust control law, TARC has been proposed for a class of uncertain nonlinear systems subjected
to unknown uncertainties. The proposed controller approximates unknown dynamics through time-delayed law and negotiates
the approximation error, that surfaces due to the time-delayed approximation of uncertainties and state derivatives, by switching
logic. The adaptive law eliminates the overestimation-underestimation problem for online evaluation of switching gain without
any prior knowledge of uncertainties. Experimentation with a WMR shows improved path tracking performance of TARC
compared to TDC and conventional ASMC. The proposed framework can also be extended for other systems such as
Autonomous Underwater Vehicle, Unmanned Aerial Vehicle, Robotic manipulator etc.
A PPENDIX A
P ROOF OF T HEOREM 2
Let us define the Lyapunov functional as,
Vf (e) = V (e) + Vf 1 (e) + Vf 2 (e) + Vf 3 (e),
Z Z Z
hς 0 0 t−h T
Vf 1 (e) =
e (η + θ)D×
β −h −ς t−h+ψ
(53)
× A2d (ψ)e(η + θ)dηdψdθ
Vf 2 (e) =
hς
β
Z
0
Z
0
Z
Vf 3 (e) = ς
−h
−ς
Z
0
−ς
Z
t
eT (η − h)D×
t+θ
× A2d (ψ)e(η − h)dηdψdθ
t
eT (η − h)(ψ)RA2d (ψ)e(η − h)dηdψ.
t+ψ
R EFERENCES
[1] M. Kristic, I. Kanellakopoulos and P. V. Kokotovic, Nonlinear and Adaptive Control Design, Wiley, New York, 1995.
[2] X. Liu, H. Su, B. Yao and J.Chu, Adaptive robust control of a class of uncertain nonlinear systems with unknown sinusoidal disturbances, 47th IEEE
Conference on Decision and Control, pp. 2594-2599, 2008.
[3] M. J. Corless and G. Leitmann, Continuous state feedback guaranteeing uniform ultimate boundness for uncertain dynamic system, IEEE Transactions
on Automatic Control, vol. 26(10), pp. 1139-1144, 1981.
[4] H. Lee and V. I. Utkin, Chattering suppression methods in sliding mode control systems, Annual Reviews in Control, vol. 31, pp. 179-188, 2007.
[5] A. Levant, Higher-order sliding modes, differentiation and output-feedback control, International Journal of Control, vol. 76(9/10), pp. 924-941, 2003.
[6] A. G. Ulsoy, Time-delayed control of siso systems for improved stability margins, ASME Journal of Dynamic Systems, Measurement, and Control, vol.
137(4), pp. 1-12, 2015.
[7] T. C. Hsia and L. S. Gao, Robot manipulator control using decentralized linear time-invariant time-delayed joint controllers, IEEE International Conference
on Robotics and Automation, pp. 2070-2075, 1990.
9
[8] K. Youcef-Toumi and O. Ito, A time delay controller for systems with unknown dynamics, ASME Journal of Dynamic Systems, Measurement, and Control,
vol. 112(1), pp. 133-142, 1990.
[9] G. R. Cho, P. H. Chang, S. H. Park and M. Jin, Robust tracking under nonlinear friction using time delay control with internal model, IEEE Transactions
on Control System Technology, vol. 17(6), pp. 1406-1414, 2009.
[10] D. K.Han and P. H. Chang, Robust tracking of robot manipulator with nonlinear friction using time delay control with gradient estimator, Journal of
Mechanical Science and Technology, vol. 24(8), pp. 1743-1752, 2010.
[11] M. Jin, S. H. Kang and P. H. Chang, Robust compliant motion control of robot with nonlinear friction using time-delay estimation, IEEE Transactions
on Industrial Electronics, vol. 55(1), pp. 258-269, 2008.
[12] Y. Jin, P. H. Chang, M. Jin and D. G. Gweon, Stability guaranteed time-delay control of manipulators using nonlinear damping and terminal sliding
mode, IEEE Transactions on Industrial Electronics, vol. 60(8), pp. 3304-3318, 2013.
[13] P. H. Chang and S. H. Park, On improving time-delay control under certain hard nonlinearities, Mechatronics, vol. 13(4), pp. 393-412, 2003.
[14] S. Roy, S. Nandy, R. Ray and S. N. Shome, Time delay sliding mode control nonholonomic wheeled mobile robot: experimental validation, IEEE
International Conference on Robotics and Automation, pp. 2886-2892, 2014.
[15] X. Zhu, G. Tao, B. Yao and J. Cao, Adaptive robust posture control of parallel manipulator driven by pneumatic muscles with redundancy, IEEE/ASME
Transactions on Mechatronics, vol. 13(4), pp. 441-450, 2008.
[16] X. Zhu, G. Tao, B. Yao and J. Cao, Integrated direct/indirect adaptive robust posture control of parallel manipulator driven by pneumatic muscles, IEEE
Transactions on Control System Technology, vol. 17(3), pp. 576-588, 2009.
[17] W. Sun, Z. Zhao and H. Gao, Saturated adaptive robust control for active suspension systems, IEEE Transactions on Industrial Electronics, vol. 60(9),
pp. 3889-3896, 2013.
[18] S. Islam, P. X. Liu and A. E. Saddik, Robust control of four rotor unmanned aerial vehicle with disturbance uncertainty, IEEE Transactions on Industrial
Electronics, vol. 63(3), pp. 1563-1571, 2015.
[19] Z. Chen, B. Yao and Q. Wang, µ-synthesis based adaptive robust posture control of linear motor driven stages with high frequency dynamics: a case
study, IEEE/ASME Transactions on Mechatronics, vol. 20(3), pp. 1482-1490, 2015.
[20] Z. Liu, H. Su and S. Pan, A new adaptive sliding mode control of uncertain nonlinear dynamics, Asian Journal of Control, vol. 16(1), pp. 198-208,
2014.
[21] C. Chen, T. S. Li, Y. Yeh and C. Chang, Design and implementation of an adaptive sliding-mode dynamic controller for wheeled mobile robots,
Mechatronics, vol. 19, pp. 156-166, 2009.
[22] F. Plestan, Y. Shtessel, V. Bregeault and A. Poznyak, New methodologies for adaptive sliding mode control, International Journal of Control, vol. 83(9),
pp. 1907-1919, 2010.
[23] F. Plestan, Y. Shtessel, V. Bregeault and A. Poznyak, Sliding mode control with gain adaptation - application to an electropneumatic actuator state,
Control Engineering Practice, vol. 21, pp. 679-688, 2013.
[24] B. Bandyopadhayay, S. Janardhanan and S. K. Spurgeon, Advances in Sliding Mode Control, Springer-Verlag, New York, 2013.
[25] G. Leitmann, On the efficiency of nonlinear control in uncertain linear systems, ASME Journal of Dynamic Systems, Measurement and Control, vol.
103, pp. 95-102, 1981.
[26] P. H. Chang and J. W. Lee, A model reference observer for time-delay control and its application to robot trajectory control, IEEE Transactions on
Control System Technology, vol. 4(1), pp. 2-10, 1996.
[27] Y. H. Shin and K. J. Kim, Performance enhancement of pneumatic vibration isolation tables in low frequency range by time delay control, Journal of
Sound and Vibration, vol. 321, pp. 537-553, 2009.
[28] A. Kuperman and Q. C. Zhong, Robust control of uncertain nonlinear systems based on an uncertainty and disturbance estimator, International Journal
of Robust and Nonlinear Control, vol. 21, pp. 79-92, 2011.
[29] S. E. Talole and S. B. Phadke, Robust input-output linearization using uncertainty and disturbance estimator, International Journal of Control, vol.
82(10), pp. 1794-1803, 2009.
[30] P. V. Suryawanshi, P. D. Shengde, and S. B. Phadke, Robust sliding mode control for a class of nonlinear systems using inertial delay control, Nonlinear
Dynamics, vol. 78(3), pp. 1921-1932, 2014.
[31] J. Reger and J. Jouffroy, On algebraic time-derivative estimation and deadbeat state reconstruction, Joint 47th IEEE Conference on Decision and Control
and 28th Chinese Control Conference, pp. 1740-1745, 2009.
[32] K. Gu, V. Khartionov and J. Chen, Stability of time-delay systems, Birkhäuser, Boston (2003).
[33] A. Nasiri, S. K. Nguang and A. Swain, Adaptive sliding mode control for a class of MIMO nonlinear systems with uncertainty, Journal of The Franklin
Institute, vol. 351, pp. 2048-2061, 2014.
[34] S. Roy, S. Nandy, R. Ray and S. N. Shome, Robust path tracking control of nonholonomic wheeled mobile robot, International Journal of Control,
Automation and Systems, vol. 13(4), pp. 897-905, 2015.
[35] J. Kim, H. Joe, S-C Yu, J. S. Lee and M. Kim, Time delay controller design for position control of autonomous underwater vehicle under disturbances,
IEEE Transactions on Industrial Electronics, DOI 10.1109/TIE.2015.2477270, 2015.
| 3cs.SY
|
arXiv:cs/0408045v1 [cs.PL] 19 Aug 2004
On computing the fixpoint of a set of boolean equations
Viktor Kuncak
K. Rustan M. Leino
MIT
[email protected]
Microsoft Research
[email protected]
30 December 2003
Technical Report
MSR-TR-2003-08
This paper presents a method for computing a least fixpoint of a system of equations over booleans. The resulting computation can be
significantly shorter than the result of iteratively evaluating the entire
system until a fixpoint is reached.
Microsoft Research
Microsoft Corporation
One Microsoft Way
Redmond, WA 98052
http://www.research.microsoft.com
KRML 118 / VK0099 / MSR-TR-2003-08, p.1
0 Introduction
Many problems in computer science, in particular those arising in the context of
program analysis, involve the computation of a least (or, dually, greatest) fixpoint
of a system of equations. In this paper, we consider a way to compute a least
fixpoint when the equations involved are over the booleans. In some important
cases, the resulting computation can be significantly shorter than the computation
that iteratively evaluates the entire system until a fixpoint is reached.
Let us begin with an overview of our result. We restrict our attention to a
finite lattice. A finite lattice is a complete lattice and has no infinite ascending
chains, and any monotonic function on such a lattice is also continuous. Hence,
the Kleene Fixpoint Theorem [2] states that the least fixpoint of any monotonic
function F is the lattice join of the sequence of elements
F 0 (⊥), F 1 (⊥), F 2 (⊥), . . .
where exponentiation denotes successive function applications and ⊥ denotes the
bottom element of the lattice. Because this sequence is ascending and because the
lattice is finite, there exists a natural number K such that
F K (⊥)
is the least fixpoint of F . We call the least such K the fixpoint depth of F .
If we are able to evaluate function F and if we are able to determine whether
two given lattice elements are equal, then we can compute the least fixpoint of
F : starting from the value ⊥ , repeatedly apply F until the application of F
leaves the value unchanged. The existence of a fixpoint depth guarantees that
this process terminates. In this paper, we consider the problem of computing an
expression for the least fixpoint, without computing the value of the expression.
By first computing a small expression for the least fixpoint, we can relegate the
computation of the value of the expression to an external tool such as a SAT solver
[3]. In the sequel we therefore do not assume that we are able to compute the value
of an expression into a particular lattice element.
The fixpoint depth of a function F on a lattice is bounded by the height of the
lattice. Therefore, for the 2-element lattice B of the booleans (which has height
1), the least fixpoint of F is given by F (⊥) , and for the 2n -element lattice Bn
that is the Cartesian product space of n booleans (which has height n ), the least
fixpoint of F is given by F n (⊥) .
KRML 118 / VK0099 / MSR-TR-2003-08, p.2
Any function F : Bn → Bn can be represented isomorphically by n functions
fi : Bn → B . We write
F = (f1 , . . . , fn )
where the tuple of functions is itself defined to be a function, as follows, for any
n -tuple X of booleans:
(f1 , . . . , fn )(X ) = (f1 (X ), . . . , fn (X ))
For example, let n = 3 and let F = (f , g, h) . Then, the least fixpoint of F
equals F 3 (⊥, ⊥, ⊥) , as we have argued above. In terms of the functions f , g, h ,
this expands to:
(f (f (f (⊥, ⊥, ⊥),
g(f (⊥, ⊥, ⊥),
h(f (⊥, ⊥, ⊥),
g(f (f (⊥, ⊥, ⊥),
g(f (⊥, ⊥, ⊥),
h(f (⊥, ⊥, ⊥),
h(f (f (⊥, ⊥, ⊥),
g(f (⊥, ⊥, ⊥),
h(f (⊥, ⊥, ⊥),
g(⊥, ⊥, ⊥),
g(⊥, ⊥, ⊥),
g(⊥, ⊥, ⊥),
g(⊥, ⊥, ⊥),
g(⊥, ⊥, ⊥),
g(⊥, ⊥, ⊥),
g(⊥, ⊥, ⊥),
g(⊥, ⊥, ⊥),
g(⊥, ⊥, ⊥),
h(⊥, ⊥, ⊥)),
h(⊥, ⊥, ⊥)),
h(⊥, ⊥, ⊥))),
h(⊥, ⊥, ⊥)),
h(⊥, ⊥, ⊥)),
h(⊥, ⊥, ⊥))),
h(⊥, ⊥, ⊥)),
h(⊥, ⊥, ⊥)),
h(⊥, ⊥, ⊥))))
We refer to this closed form of the fixpoint as the Expanded Closed Form. A
different way to write down the Expanded Closed Form, which shares common
subexpressions, is:
let a1 = ⊥,
a2 = ⊥,
a3 = ⊥
let b1 = f (a1 , a2 , a3 ), b2 = g(a1 , a2 , a3 ), b3 = h(a1 , a2 , a3 )
let c1 = f (b1 , b2 , b3 ), c2 = g(b1 , b2 , b3 ), c3 = h(b1 , b2 , b3 )
let d1 = f (c1 , c2 , c3 ), d2 = g(c1 , c2 , c3 ), d3 = h(c1 , c2 , c3 )
(d1 , d2 , d3 )
in
in
in
in
This representation is cubic in n , which means that computing it may take time
and space that is cubic in n .∗
If we allow ourselves to write functions of n arguments as functions over n -tuples, then we
can obtain a quadratic representation. For example, with n = 3 , we have let a1 = . . . , a2 =
. . . , a3 = . . . in let a = (a1 , a2 , a3 ) in let b1 = f (a), b2 = g(a), b3 = h(a) in let b =
(b1 , b2 , b3 ) in . . . .
∗
KRML 118 / VK0099 / MSR-TR-2003-08, p.3
Let us consider another closed form, which we call the Pruned Closed Form.
In the Pruned Closed Form, an application of a function fi is replaced by ⊥ if
it occurs in another application of the same function fi . For the example above,
where n = 3 , the Pruned Closed Form is:
(f (⊥,
g(⊥,
h(⊥,
g(f (⊥,
⊥,
h(⊥,
h(f (⊥,
g(⊥,
⊥
⊥,
h(⊥, ⊥, ⊥)),
g(⊥, ⊥, ⊥), ⊥
)),
⊥,
h(⊥, ⊥, ⊥)),
⊥,
⊥
g(⊥, ⊥, ⊥), ⊥
⊥,
⊥
)),
),
),
))
If we do not have any interpretation for the functions fi —in other words, if
each fi is just a symbolic name for an uninterpreted function—then the cubicsized Expanded Closed Form may be a reasonably small closed-form representation of the fixpoint. The Pruned Closed Form is generally much larger than cubic
in n : for every subset S of f2 , . . . , fn , function f1 appears expanded in a context where the set of enclosing functions is S . (A smaller Pruned Closed Form
can be obtained by taking advantage of common subexpressions.) However, there
are cases where the Pruned Closed Form can be significantly smaller than the Expanded Closed Form, for example when the fixpoint computation is dominated
by the computation of local fixpoints, meaning fixpoints that involve only a small
number of the functions. An important situation in program analysis where this
case applies is when each function represents a control point in a given program,
a function is defined in terms of the functions corresponding to the successor (or
predecessor) control points, and the given program contains many local loops.
For example, suppose
f (x , y, z ) = f(x , y)
g(x , y, z ) = g(x )
h(x , y, z ) = h(y, z )
for some functions f , g , and h . Then the Expanded Closed Form is
let a1 = ⊥,
a2 = ⊥,
a3 = ⊥
let b1 = f(a1 , a2 ), b2 = g(a1 ), b3 = h(a2 , a3 )
let c1 = f(b1 , b2 ), c2 = g(b1 ), c3 = h(b2 , b3 )
let d1 = f(c1 , c2 ), d2 = g(c1 ), d3 = h(c2 , c3 )
(d1 , d2 , d3 )
in
in
in
in
KRML 118 / VK0099 / MSR-TR-2003-08, p.4
In contrast, the Pruned Closed Form yields the much shorter expression
(f( f(⊥, g(⊥)),
g(⊥)),
g(f(⊥, ⊥)),
h(g(f(⊥, ⊥)),
⊥))
More generally, for an even n , suppose fi (x1 , . . . , xn ) is fi (xi , xi+1 ) when i is
odd and fi (xi−1 , xi ) when i is even. Then the Expanded Closed Form is still
cubic, whereas the Pruned Closed Form is the linear-sized expression
( f1 (⊥, f2 (⊥, ⊥)), f2 (f1 (⊥, ⊥), ⊥),
...
fi (⊥, fi+1 (⊥, ⊥)), fi+1 (fi (⊥, ⊥), ⊥),
...
fn−1 (⊥, fn (⊥, ⊥)), fn (fn−1 (⊥, ⊥), ⊥)
)
In the rest of this paper, we define the Pruned Closed Form more precisely and
prove that it yields the same value as the Expanded Closed Form.
1 Using the Bekić-Leszczyłowski Theorem
In this section, we sketch how to obtain the Pruned Closed Form by applications
of the Bekić-Leszczyłowski Theorem [1, 4].
We write
( ↓ x • R(x ) )
for the lattice meet of all values for x that satisfy the predicate R(x ) . For any
monotonic function F , we then write
( ↓ x • x = F (x ) )
(0)
to denote the least fixpoint of F , because the Tarski Fixpoint Theorem [5] says
that the meet of all fixpoints is itself a fixpoint. Using for a function F : Bn →
Bn the isomorphic representation of n functions fi : Bn → B , we can write (0)
equivalently as:
( ↓ x1 , . . . , xn •
x1 = f1 (x1 , . . . , xn )
..
.
xn = fn (x1 , . . . , xn ) )
∧
∧
KRML 118 / VK0099 / MSR-TR-2003-08, p.5
We can now state the Bekić-Leszczyłowski Theorem [1, 4], for any monotonic
functions F and G (possibly over different lattices):
( ↓ a, b • a = F (a, b) ∧ b = G(a, b) )
=
( ↓ a, b • a = F (a, b) ∧ b = ( ↓ b • b = G(a, b) ))
Note that each side of the equality expresses a fixpoint in the lattice Bn if F and
G are functions of types Bp × Bq → Bp and Bp × Bq → Bq , respectively, for p
and q such that p + q = n .
A consequence of the Bekić-Leszczyłowski Theorem and the Kleene Fixpoint
Theorem for a known fixpoint depth is the following lemma:
Lemma 0 For any lattice domain A and monotonic functions F : A × B → A
and G: A × B → B ,
( ↓ a, b • a = F (a, b) ∧ b = G(a, b) )
=
( ↓ a, b • a = F (a, b) ∧ b = G(a, ⊥) )
Proof.
=
=
( ↓ a, b • a = F (a, b) ∧ b = G(a, b) )
{ Bekić-Leszczyłowski Theorem }
( ↓ a, b • a = F (a, b) ∧ b = ( ↓ b • b = G(a, b) ))
{ ( λ b • G(a, b) ) is a function on B , and
therefore its fixpoint depth is at most 1, and
therefore ( ↓ b • b = G(a, b) ) = G(a, ⊥) }
( ↓ a, b • a = F (a, b) ∧ b = G(a, ⊥) )
Using Lemma 0, we now show that the Pruned Closed Form is indeed the least
fixpoint in B2 . For any monotonic boolean functions f and g :
=
=
=
( ↓ a, b • a = f (a, b) ∧ b = g(a, b) )
{ Lemma 0 with F , G := f , g }
( ↓ a, b • a = f (a, b) ∧ b = g(a, ⊥) )
{ substitute equals for equals }
( ↓ a, b • a = f (a, g(a, ⊥)) ∧ b = g(a, ⊥) )
{ Lemma 0 with
G, F := ( λ a, b • f (a, g(a, ⊥)) ), ( λ a, b • g(a, ⊥) ) }
( ↓ a, b • a = f (⊥, g(⊥, ⊥)) ∧ b = g(a, ⊥) )
(1)
KRML 118 / VK0099 / MSR-TR-2003-08, p.6
This calculation shows that an expression for the least solution of a in equation
(1) is
f (⊥, g(⊥, ⊥))
By a symmetric argument, an expression for the least solution of b in equation
(1) is
g(f (⊥, ⊥), ⊥)
That is, an expression for (1) is
( f (⊥, g(⊥, ⊥)), g(f (⊥, ⊥), ⊥) )
which is the Pruned Closed Form.
Using the result for B2 , we can show that the Pruned Closed Form is also the
least fixpoint in B3 . For any monotonic boolean functions f , g , and h :
=
=
=
=
=
=
=
=
( ↓ a, b, c • a = f (a, b, c) ∧ b = g(a, b, c) ∧ c = h(a, b, c) )
(2)
{ Lemma 0 with G := h (and with F as the isomorphic
representation of functions f and g ) }
( ↓ a, b, c • a = f (a, b, c) ∧ b = g(a, b, c) ∧ c = h(a, b, ⊥) )
{ substitute equals for equals }
( ↓ a, b, c • a = f (a, b, c) ∧ b = g(a, b, h(a, b, ⊥)) ∧ c = h(a, b, ⊥) )
{ Lemma 0 with G := ( λ a, b, c • g(a, b, h(a, b, ⊥)) ) }
( ↓ a, b, c • a = f (a, b, c) ∧ b = g(a, ⊥, h(a, ⊥, ⊥)) ∧ c = h(a, b, ⊥) )
{ substitute equals for equals }
( ↓ a, b, c • a = f (a, g(a, ⊥, h(a, ⊥, ⊥)), c) ∧
b = g(a, ⊥, h(a, ⊥, ⊥)) ∧ c = h(a, b, ⊥) )
{ the first 3 steps of this calculation, in reverse order }
( ↓ a, b, c • a = f (a, g(a, ⊥, h(a, ⊥, ⊥)), c) ∧
b = g(a, b, c) ∧ c = h(a, b, c) )
{ Lemma 0 with G := g }
( ↓ a, b, c • a = f (a, g(a, ⊥, h(a, ⊥, ⊥)), c) ∧
b = g(a, ⊥, c) ∧ c = h(a, b, c) )
{ substitute equals for equals }
( ↓ a, b, c • a = f (a, g(a, ⊥, h(a, ⊥, ⊥)), c) ∧
b = g(a, ⊥, c) ∧ c = h(a, g(a, ⊥, c), c) )
{ Lemma 0 with G := ( λ a, b, c • h(a, g(a, ⊥, c), c) ) }
KRML 118 / VK0099 / MSR-TR-2003-08, p.7
=
=
( ↓ a, b, c • a = f (a, g(a, ⊥, h(a, ⊥, ⊥)), c) ∧
b = g(a, ⊥, c) ∧ c = h(a, g(a, ⊥, ⊥), ⊥) )
{ substitute equals for equals }
( ↓ a, b, c • a = f (a, g(a, ⊥, h(a, ⊥, ⊥)), h(a, g(a, ⊥, ⊥), ⊥)) ∧
b = g(a, ⊥, c) ∧ c = h(a, g(a, ⊥, ⊥), ⊥) )
{ Lemma 0 with G :=
( λ a, b, c • f (a, g(a, ⊥, h(a, ⊥, ⊥)), h(a, g(a, ⊥, ⊥), ⊥)) ) }
( ↓ a, b, c • a = f (⊥, g(⊥, ⊥, h(⊥, ⊥, ⊥)), h(⊥, g(⊥, ⊥, ⊥), ⊥)) ∧
b = g(a, ⊥, c) ∧ c = h(a, g(a, ⊥, ⊥), ⊥) )
This calculation shows that an expression for the least solution of a in (2) is
f (⊥, g(⊥, ⊥, h(⊥, ⊥, ⊥)), h(⊥, g(⊥, ⊥, ⊥), ⊥))
and similarly for b and c .
Our main result is that the Pruned Closed Form is the least fixpoint in Bn for
any n . In the next section, we prove this result directly, not using Lemma 0.
2 The theorem
We are given n ≥ 1 monotonic functions f1 , . . . , fn : Bn → B , where B is the
boolean domain {0, 1} ordered by ≤ (with 0 ≤ 1 ). To represent an indexed
→
n -tuple of things, like a list of booleans x1 , . . . , xn , we write −
x . The fact that
the given functions are monotonic is written as follows, for any index i and any
→
→
tuples of booleans −
x and −
y :
→→
−
→
−
→
→
x ≤−
y ⇒ f i .−
x ≤ fi .−
y
where an infix dot (with the highest operator precedence) denotes function appli→
−
cation, and the order ≤ is the component-wise ordering of tuples:
→→
−
→
−
x ≤−
y ≡ ( ∀ i • xi ≤ yi )
We are interested in viewing the functions as specifying a system of equations,
namely:
x1 , . . . , xn : x1 = f1 .(x1 , . . . , xn )
...
xn = fn .(x1 , . . . , xn )
(3)
KRML 118 / VK0099 / MSR-TR-2003-08, p.8
where the variables to the left of the colon show the unknowns. We take a tuple of
→
−
functions (f1 , . . . , fn ) , which we can also write as f , to itself be a function, one
which produces a tuple from the results of applying the given argument to each of
→
the functions. For example, for the functions given above and an argument −
x ,
we have:
→
→
→
(f1 , . . . , fn ).−
x = (f1 .−
x , . . . , fn .−
x)
Thus, we can write the system (3) of equations as:
→→
−
→
−
→
x : −
x = f .−
x
→
−
→
We are interested in the least (in the sense of the ordering ≤ ) solution −
x that
satisfies this equation. That is, we are interested in the least fixpoint of the function
→
−
→
−
f . Because the lattice of boolean n -tuples has height n , the least fixpoint of f
−
→
can be reached by applying f n times starting from the bottom element of the
→
−
lattice. That is, the least fixpoint of f is given by:
−n −
→
→
f .0
→
−
where exponentiation denotes successive function applications and 0 is the tuple
of n 0’s.
To precisely specify the Pruned Closed Form, we introduce a notation that
keeps track of which functions have been applied in the enclosing context. In
particular, we use a set that contains the indices of the functions already applied.
Formally, we define the following family of functions, for any index i and set S
of indices:
fi ◦ (gS ∪{i},1 , . . . , gS ∪{i},n ) if i 6∈ S
gS ,i =
−
(λ→
x • 0)
if i ∈ S
Taking advantage of our previous notation and using 0̇ to denote the function that
always returns 0 (that is, the boolean 0 extended pointwise to a boolean function),
we can write the definition of g as follows:
−→
fi ◦ −
g−
S ∪{i} if i 6∈ S
gS ,i =
0̇
if i ∈ S
Our goal is now to prove the following:
Theorem 1
−
→n −
→
−
→
−
→
g∅ . 0 = f . 0
KRML 118 / VK0099 / MSR-TR-2003-08, p.9
3 Proof
We start by proving some lemmas that we use in the proof of this theorem.
Lemma 2 For any index i and for any S $ {1, . . . , n} ,
→n−|S |−1 −
−
→
−
→
gS ,i . 0 ≤ (fi ◦ f
). 0
Proof. By induction on n − |S | . Let T denote S ∪ {i } . We consider three cases.
C ASE i ∈ S :
→
−
gS ,i . 0
=
{ definition of g , since i ∈ S }
→
−
0̇. 0
=
{ definition of 0̇ }
0
≤
{ 0 is bottom element of ≤ }
→n−|S |−1 −
−
→
(fi ◦ f
). 0
C ASE i 6∈ S ∧ |T | < n :
→
−
gS ,i . 0
=
{ definition of g , since i 6∈ S }
→
−
−
→
(fi ◦ gT ). 0
=
→
−
(fi ◦ (gT ,1, . . . , gT ,n )). 0
→
−
=
{ distribute . 0 }
→
−
→
−
fi .(gT ,1. 0 , . . . , gT ,n . 0 )
≤
{ for each index j , induction hypothesis with i , S := j , T , since
|S | + 1 = |T | < n ; and monotonicity of fi }
→n−|S |−2 −
−
→n−|S |−2 −
−
→
→
fi .((f1 ◦ f
). 0 , . . . , (fn ◦ f
). 0 )
→
−
=
{ distribute . 0 }
→n−|S |−2
−
→n−|S |−2 −
−
→
(fi ◦ (f1 ◦ f
, . . . , fn ◦ f
)). 0
→n−|S |−2
−
=
{ distribute ◦ f
}
→n−|S |−2 −
−
→
(fi ◦ (f1 , . . . , fn ) ◦ f
). 0
=
{ exponentiation }
→n−|S |−1 −
−
→
(fi ◦ f
). 0
KRML 118 / VK0099 / MSR-TR-2003-08, p.10
C ASE i 6∈ S ∧ |T | = n :
→
−
gS ,i . 0
=
{ see first 3 steps of previous case }
→
−
→
−
fi .(gT ,1. 0 , . . . , gT ,n . 0 )
=
{ for each index j , j ∈ T , so gT ,j = 0̇ }
→
−
fi . 0
=
{ |S | = n − 1 , so f n−|S |−1 is the identity function }
→
−
(fi ◦ f n−|S |−1 ). 0
The following corollary of Lemma 2 proves one direction of Theorem 1.
Corollary 3
→n −
→−
−
→
→−
−
→
g∅ . 0 ≤ f . 0
Proof.
→
−
−
→
g∅ . 0
=
−
→
≤
=
=
→
−
→
−
(g∅,1 . 0 , . . . , g∅,n . 0 )
{ for each index j , Lemma 2 with i , S := j , ∅ }
→n−1 −
−
→n−1 −
−
→
→
((f1 ◦ f
). 0 , . . . , (fn ◦ f
). 0 )
→n−1
−
→
−
{ distribute . 0 and ◦ f
}
→n−1 −
−
→
((f1 , . . . , fn ) ◦ f
). 0
{ exponentiation }
→n −
−
→
f .0
To support the remaining lemmas, we define one more family of functions.
For any index i and set S of indices,
fi if i 6∈ S
hS ,i =
0̇ if i ∈ S
Lemma 4 For any index i , monotonic function H : Bn → Bn , and m ≥ 0 ,
→
−
→
0 ≤ p ≤ m • (fi ◦ H p ).−
(fi ◦ H m ). 0 = 0 ⇒ ( ∀ p
0 = 0)
Proof. We prove the term of the quantification as follows:
KRML 118 / VK0099 / MSR-TR-2003-08, p.11
≤
=
=
→
−
(fi ◦ H p ). 0
− −
→
→
→
−
{ monotonicity of fi and H , since 0 ≤ H m−p . 0 }
→
−
(fi ◦ H p ).(H m−p . 0 )
→
−
(fi ◦ H m ). 0
{ antecedent }
0
Lemma 5 For any index i , set S of indices, m ≥ 0 , and T = S ∪ {i } ,
−
→m −
−
→p →
−
→p −
→
→
0 ≤ p ≤ m • hS .−
(fi ◦ hS ). 0 = 0 ⇒ ( ∀ p
0 = hT . 0 )
Proof. If i ∈ S , then S = T and the consequent follows trivially. For i 6∈ S ,
we prove the term of the quantification by induction on p .
C ASE p = 0 : Trivial—exponentiation with 0 gives identity function.
C ASE p > 0 :
−
→p −
→
hS . 0
=
{ exponentiation, since p > 0 }
−
→ −
→p−1 −
→
hS .(hS . 0 )
−
→p−1 −
→
=
{ distribute .(hS . 0 ) }
−
→p−1 −
−
→p−1 −
→
→
(hS ,1.(hS . 0 ), . . . , hS ,n .(hS . 0 ))
−
→p−1 −
−
→p−1 −
→
→
=
{ for any index j , hS ,j .(hS . 0 ) = hT ,j .(hS . 0 ) , see below }
−
→p−1 −
−
→p−1 −
→
→
(hT ,1.(hS . 0 ), . . . , hT ,n .(hS . 0 ))
−
→p−1 −
→
=
{ distribute .(hS . 0 ) }
−
→ −
→p−1 −
→
hT .(hS . 0 )
=
{ induction hypothesis with p := p − 1 }
−
→ −
→p−1 −
→
hT .(hT . 0 )
=
{ exponentiation }
−
→p −
→
hT . 0
Now for the proof of the third step in the calculation above. If j 6= i , then
j ∈ S ≡ j ∈ T , so hS ,j = hT ,j . If j = i , then:
=
−
→p−1 −
→
hS ,i .(hS . 0 )
{ definition of h , since i 6∈ S }
KRML 118 / VK0099 / MSR-TR-2003-08, p.12
=
=
=
−
→p−1 −
→
fi .(hS . 0 )
−
→
{ Lemma 4 with H , p := hS , p − 1 , using the antecedent of
Lemma 5 to fulfill the antecedent of Lemma 4 }
0
{ definition of 0̇ }
−
→p−1 −
→
0̇.(hS . 0 )
{ definition of h , since i ∈ T }
−
→p−1 −
→
hT ,i .(hS . 0 )
We need one more lemma.
Lemma 6 For any index i , set S of indices, and m satisfying 0 ≤ m ≤ n −|S | ,
−
→m −
→
→
−
(hS ,i ◦ hS ). 0 ≤ gS ,i . 0
Proof. By induction on m . We consider three cases.
C ASE i ∈ S :
−
→m
hS ,i ◦ hS
=
{ definition of h , since i ∈ S }
−
→m
0̇ ◦ hS
=
{ 0̇ is left zero element of ◦ }
0̇
=
{ definition of g , since i ∈ S }
gS ,i
C ASE i 6∈ S ∧ m = 0 :
−
→m −
→
(hS ,i ◦ hS ). 0
=
{ exponentiation, since m = 0 }
→
−
hS ,i . 0
=
{ definition of h , since i 6∈ S }
→
−
fi . 0
→−
−
→
→ →−
≤
{ monotonicity of fi , since 0 ≤ −
gT . 0 }
→
−
fi .(−
g→
T. 0 )
=
{ definition of g , since i 6∈ S }
→
−
gS ,i . 0
(4)
KRML 118 / VK0099 / MSR-TR-2003-08, p.13
C ASE i 6∈ S ∧ m > 0 : It suffices to prove that the left-hand side of (4) is 0
whenever the right-hand side is 0. Therefore, we assume the latter to be 0:
→
−
gS ,i . 0 = 0
(5)
and prove the former to be 0:
−
→m −
→
(hS ,i ◦ hS ). 0
=
{ definition of h , since i 6∈ S }
−
→m −
→
(fi ◦ hS ). 0
=
{ exponentiation, since m > 0 }
−
→m−1 −
→
(fi ◦ (hS ,1 , . . . , hS ,n ) ◦ hS
). 0
−
→m−1
→
−
=
{ distribute ◦ hS
and . 0 }
−
→m−1 −
−
→m−1 −
→
→
fi .((hS ,1 ◦ hS
). 0 , . . . , (hS ,n ◦ hS
). 0 )
≤
{ (6), see below; and monotonicity of fi }
→
−
→
−
fi .(gT ,1. 0 , . . . , gT ,n . 0 )
=
→
−
(fi ◦ −
g→
T ). 0
=
{ definition of g , since i 6∈ S }
→
−
gS ,i . 0
=
{ assumption (5) }
0
In this calculation, we used the following fact: for every index j ,
−
→m−1 −
→
→
−
(hS ,j ◦ hS
). 0 ≤ gT ,j . 0
(6)
which we now prove. We divide the proof of (6) up into two sub-cases.
−
→m−1 −
→
S UB - CASE (hS ,j ◦ hS
). 0 = 0 : Formula (6) follows immediately.
−
→m−1 −
→
S UB - CASE (hS ,j ◦ hS
). 0 6= 0 : First, we derive some consequences of
assumption (5):
→
−
gS ,i . 0 = 0
⇒
{ induction hypothesis with S , i , m := S , i , m − 1 }
−
→m−1 −
→
(hS ,i ◦ hS
(7)
). 0 = 0
=
{ definition of h , since i 6∈ S }
−
→m−1 −
→
(fi ◦ hS
). 0 = 0
⇒
{ Lemma 5 with m, p := m − 1, m − 1 }
−
→m−1 −
−
→m−1 −
→
→
hS
(8)
. 0 = hT
.0
KRML 118 / VK0099 / MSR-TR-2003-08, p.14
Now, calculating from the assumption we made in this sub-case:
−
→m−1 −
→
(hS ,j ◦ hS
). 0 6= 0
=
{ (7) }
−
→m−1 −
→
(hS ,j ◦ hS
). 0 6= 0 ∧ i 6= j
⇒
{ i 6= j , so j ∈ S ≡ j ∈ T , so hS ,j = hT ,j }
−
→m−1 −
→
(hT ,j ◦ hS
). 0 6= 0
=
{ (8) }
−
→m−1 −
→
(hT ,j ◦ hT
). 0 6= 0
⇒
{ induction hypothesis with S , i , m := T , j , m − 1 }
→
−
gT ,j . 0 6= 0
⇒
(6)
This concludes the proof of Lemma 6.
And finally, the proof of the theorem:
Proof of Theorem 1. The proof is a ping-pong argument.
→
−
→
−
g∅ . 0
→
−
≤
{ Corollary 3 }
→n −
−
→
f .0
—ping!
=
{ exponentiation, since n ≥ 1 }
→n−1 −
−
→
((f1 , . . . , fn ) ◦ f
). 0
→n−1
−
→
−
=
{ distribute ◦ f
and . 0 }
→n−1 −
−
→n−1 −
−
→
→
((f1 ◦ f
). 0 , . . . , (fn ◦ f
). 0 )
=
{ by definition of h , h∅,i = fi for each index i ; and
→ −
−
→
thus also h∅ = f }
→n−1 −
−
→n−1 −
−
→
→
((h∅,1 ◦ h∅ ). 0 , . . . , (h∅,n ◦ h∅
). 0 )
→
−
≤
{ Lemma 6 with S , i , m := ∅, i , n − 1 for each i }
→
−
→
−
(g∅,1 . 0 , . . . , g∅,n . 0 )
→
−
=
{ distribute . 0 }
→
−
→
−
g∅ . 0
—pong!
4 Related Work and Acknowledgments
Our theorem has already found a use, namely in the translation of boolean programs into satisfiability formulas [3].
KRML 118 / VK0099 / MSR-TR-2003-08, p.15
Before we knew of the Bekić-Leszczyłowski Theorem, one of us (Kuncak)
proved the theorem as detailed in Section 3. Tony Hoare then proposed a way to
prove the theorem in a way that would eliminate recursive uses of variables, one
by one. In doing this, Hoare also proved what essentially amounts to the BekićLeszczyłowski Theorem, appealing only to the Tarski Fixpoint Theorem [5]. We
elaborated this format in Section 1, to whose formulation Carroll Morgan also
contributed. We learnt about the Bekić-Leszczyłowski Theorem from Patrick
Cousot. The theorem is often called simply the Bekić Theorem, but de Bakker [0]
traces an independent proof thereof to Leszczyłowski. Finally, we are grateful for
feedback from the Eindhoven Tuesday Afternoon Club and the participants of the
IFIP WG 2.3 meeting in Biarritz, France (March 2003).
References
[0] J. de Bakker. Mathematical Theory of Program Correctness. Prentice-Hall,
1980.
[1] Hans Bekić. Definable operation in general algebras, and the theory of automata and flowcharts. In Cliff B. Jones, editor, Programming Languages and
Their Definition—Hans Bekić (1936–1982), volume 177 of Lecture Notes in
Computer Science, pages 30–55. Springer, 1984.
[2] Stephen Cole Kleene. Introduction to Metamathematics. D. Van Nostrand,
New York, 1952.
[3] K. Rustan M. Leino. A SAT characterization of boolean-program correctness.
In Thomas Ball and Sriram K. Rajamani, editors, Model Checking Software,
volume 2648 of Lecture Notes in Computer Science, pages 104–120. Springer,
May 2003.
[4] Jacek Leszczyłowski. A theorem on resolving equations in the space of languages. Bull. Acad. Polon. Sci., 19:967–970, 1971.
[5] Alfred Tarski. A lattice-theoretical fixpoint theorem and its applications. Pacific Journal of Mathematics, 5:285–309, 1955.
| 6cs.PL
|
A Modular Structural Operational Semantics
for Delimited Continuations
Neil Sculthorpe
Paolo Torrini
Peter D. Mosses
PL AN C OMP S Project
Department of Computer Science
Swansea University, UK
[email protected]
GRAC E FUL Project
Department of Computer Science
KU Leuven, Belgium
[email protected]
PL AN C OMP S Project
Department of Computer Science
Swansea University, UK
[email protected]
It has been an open question as to whether the Modular Structural Operational Semantics framework can express the dynamic semantics of call/cc. This paper shows that it can, and furthermore,
demonstrates that it can express the more general delimited control operators control and shift.
1
Introduction
Modular Structural Operational Semantics (MSOS) [25, 26, 27] is a variant of the well-known Structural
Operational Semantics (SOS) framework [29]. The principal innovation of MSOS relative to SOS is that
it allows the semantics of a programming construct to be specified independently of any semantic entities
with which it does not directly interact. For example, function application can be specified by MSOS
rules without mentioning stores or exception propagation.
While it is known that MSOS can specify the semantics of programming constructs for exception
handling [7, 8, 25], it has been unclear whether MSOS can specify more complex control-flow operators,
such as call/cc [1, 9]. Indeed, the perceived difficulty of handling control operators has been regarded
as one of the main limitations of MSOS relative to other modular semantic frameworks (e.g. [30, Section 2]). This paper demonstrates that the dynamic semantics of call/cc can be specified in MSOS, with
no extensions to the MSOS framework required. We approach this by first specifying the more general
delimited control operators control [18, 19, 34] and shift [12, 13, 14], and then specifying call/cc in terms
of control. In contrast to most other operational specifications of control operators given in direct style
(e.g. [18, 22, 24, 33]), ours is based on labelled transitions, rather than on evaluation contexts.
We will begin by giving a brief overview of delimited continuations (Section 2) and MSOS (Section 3). The material in these two sections is not novel, and can be skipped by a familiar reader. We will
then present our MSOS specification of the dynamic semantics of delimited control operators (Section 4).
To ensure that our MSOS specification does indeed define the same control operators as described in the
literature, we provide a proof of equivalence between our specification and one based on evaluation
contexts (Section 5).
2
Delimited Continuations
At any point in the execution of a program, the current continuation represents the rest of the computation. In a meta-language sense, a continuation can be understood as a context in which a program term
can be evaluated. Control operators allow the current continuation to be treated as an object in the language, by reifying it as a first-class abstraction that can be applied and manipulated. The classic example
of a control operator is call/cc [1, 9].
Olivier Danvy & Ugo de’Liguoro (Eds.): WoC’15 Postproceedings
EPTCS 212, 2016, pp. 63–80, doi:10.4204/EPTCS.212.5
c Neil Sculthorpe, Paolo Torrini & Peter D. Mosses
This work is licensed under the
Creative Commons Attribution License.
A Modular Structural Operational Semantics for Delimited Continuations
64
Delimited continuations generalise the notion of a continuation to allow representations of partial
contexts, relying on a distinction between inner and outer context. Control operators that manipulate
delimited continuations are always associated with control delimiters. The most well-known delimited
control operators are control (associated with the prompt delimiter) [18, 19, 34] and shift (associated with
the reset delimiter) [12, 13, 14], both of which can be used to simulate call/cc. The general idea of control
and shift is to capture the current continuation up to the innermost enclosing delimiter, representing
the inner context. We will give an informal description of control in this section. The formal MSOS
specification of control is given in Section 4, where we also specify shift and call/cc in terms of control.
control is a (call-by-value) unary operator that takes a higher-order function f as its argument, where
f expects a reified continuation as its argument. When executed, control reifies the current continuation,
up to the innermost enclosing prompt, as a function k. That inner context is then discarded and replaced
with the application f k. Other than its interaction with control, prompt is simply a unary operator that
evaluates its argument and returns the resulting value.
Let us consider some examples. In the following expression, the continuation k is bound to the
function (λ x. 2 ∗ x), the result of the prompt application is 14, and the expression evaluates to 15:
1 + prompt(2 ∗ control(λ k. k 7))
❀
15
A reified continuation can be applied multiple times, for example:
1 + prompt(2 ∗ control(λ k. k(k 7)))
29
❀
Furthermore, a continuation need not be applied at all. For example, in the following expression, the
multiplication by two is discarded:
1 + prompt(2 ∗ control(λ k. 7))
❀
8
In the preceding examples, the continuation k could have been computed statically. However, in
general, the current continuation is the context at the point in a program’s execution when control is executed, by which time some of the computation in the source program may already have been performed.
For example, the following program will print ABB:
prompt( print ′ A′ ; control(λ k. (k () ; k ())) ; print ′ B′ )
ABB
The command (print ′ A′ ) is executed before the control operator, so does not form part of the continuation
reified by control. In this case, k is bound to (λ x. (x ; print ′ B′ )), and so B is printed once for every
application of k.
Further examples of control can be found in the online test suite accompanying this paper [32], and
in the literature [18, 19].
3
Modular SOS
The rules in this paper will be presented using Implicitly Modular SOS (I-MSOS) [27], a variant of
MSOS that has a notational style similar to conventional SOS. I-MSOS can be viewed as syntactic sugar
for MSOS. We assume the reader is familiar with SOS (e.g. [3, 29]) and the basics of MSOS [25, 26, 27].
The key notational convenience of I-MSOS is that any semantic entities (e.g. stores or environments)
that are not mentioned in a rule are implicitly propagated between the premise(s) and conclusion, allowing entities that do not interact with the programming construct being specified to be omitted from
Neil Sculthorpe, Paolo Torrini & Peter D. Mosses
65
the rule. Two types of semantic entities are relevant to this paper: inherited entities (e.g. environments),
which, if unmentioned, are implicitly propagated from the conclusion to the premises, and observable
entities (emitted signals, e.g. exceptions), which, if unmentioned, are implicitly propagated from a sole
premise to the conclusion. Observable entities are required to have a default value, which is implicitly
used in the conclusion of rules that lack a premise and do not mention the entity. Note that by premise,
we refer specifically to a transition of the → relation, not any side conditions on the rule (which, for
notational convenience, we also write above the line).
To demonstrate the specification of control operators using I-MSOS rules, this paper will use the
funcon framework [8]. This framework contains an open collection of modular fundamental constructs
(funcons), each of which has its semantics specified independently by I-MSOS rules. Funcons facilitate formal specification of programming languages by serving as a target language for a specification
given by an inductive translation, in the style of denotational semantics. However, this paper is not concerned with the translation of control operators from any specific language: our aim is to give MSOS
specifications of control operators, and the funcon framework is a convenient environment for specifying
prototypical control operators. Examples of translations into funcons can be found in [8, 28].
We will now present some examples of funcons, and their specifications as small-step I-MSOS rules.
No familiarity with the funcon framework is required: for the purposes of understanding this paper the
funcons may simply be regarded as abstract syntax. We typeset funcon names in bold, meta-variables in
Capitalised Italic, and the names of semantic entities in sans-serif. When we come to funcons for control
operators, we will continue to use italic when referring to the control operator in general, and bold when
referring to the funcon specifically.
Figure 1 presents I-MSOS rules for the exception-handling funcons throw and catch [8]. The idea
is that throw emits an exception signal, and catch detects and handles that signal. The first argument of
catch is the expression to be evaluated, and the second argument (a function) is the exception handler.
Exception signals use an observable entity named exc, which is written as a label on the transition arrow.
The exc entity has either the value none, denoting the absence of an exception, or some(V ), denoting
the occurrence of an exception with value V . The side condition val(V ) requires the term V to be a value,
thereby controlling the order in which the rules can be applied. In the case of throw, first the argument is
evaluated to a value (Rule 1), and then an exception carrying that value is emitted (Rule 2). In the case of
catch, the first argument E is evaluated while no exception occurs (Rule 3). If an exception does occur,
then the handler H is applied to the exception value and the computation E is abandoned (Rule 4). If E
evaluates to a value V , then H is discarded and V is returned (Rule 5).
Observe that rules 1 and 5 do not mention the exc entity. In Rule 1 it is implicitly propagated from
premise to conclusion, and in Rule 5 it implicitly has the default value none. Also observe that none of
the rules in Figure 1 mention any other entities such as environments or stores; any such entities are also
implicitly propagated.
exc none
E −−−−−→ E ′
E → E′
throw(E) → throw(E ′ )
val(V )
exc some(V )
throw(V ) −−−−−−−→ stuck
exc none
(1)
catch(E, H) −−−−−→ catch(E ′ , H)
(3)
exc some(V )
E −−−−−−−→ E ′
(2)
exc none
catch(E, H) −−−−−→ apply(H,V )
val(V )
catch(V, H) → V
Figure 1: I-MSOS rules for exception handling.
(4)
(5)
A Modular Structural Operational Semantics for Delimited Continuations
66
ρ (I) = V
env ρ ⊢ bv(I) → V
env ρ ⊢ lambda(I, E) → closure(ρ , I, E)
val(closure(ρ , I, E))
E1 → E1′
apply(E1 , E2 ) → apply(E1′ , E2 )
(9)
val(V )
E → E′
apply(V, E) → apply(V, E ′ )
(10)
val(V1 )
val(V2 )
apply(closure(ρ , I,V1 ),V2 ) → V1
(11)
(6)
(7)
(8)
val(V )
env ({I 7→ V }/ρ ) ⊢ E → E ′
env ⊢ apply(closure(ρ , I, E),V ) → apply(closure(ρ , I, E ′ ),V )
(12)
Figure 2: I-MSOS rules for call-by-value lambda calculus.
Figure 2 presents I-MSOS rules for identifier lookup (bv, “bound-value”), abstraction (lambda), and
application (apply). Note that the closure funcon is a value constructor [7] (specified by Rule 8), and
thus has no transition rules of its own. We present these rules here for completeness, as these funcons
will be used when defining the semantics of control operators in Section 4.
Again, observe that rules 9–11 do not mention the environment env; it is propagated implicitly.
Furthermore, consider that none of the rules in Figure 1 mention the environment env, and none of
the rules in Figure 2 mention the exc signal. However, the modular nature of I-MSOS specifications
allows the two sets of rules to be combined without modification, with implicit propagation handling the
unmentioned entities. For comparison, in Figure 3 we present a conventional SOS specification of this
call-by-value lambda calculus combined with exception handling, in which both semantic entities are
mentioned explicitly in every rule.
ρ (I) = V
exc none
env ρ ⊢ bv(I) −−−−−→ V
exc X
env ρ ⊢ E −−−→ E ′
exc X
env ρ ⊢ throw(E) −−−→ throw(E ′ )
exc none
env ρ ⊢ lambda(I, E) −−−−−→ closure(ρ , I, E)
val(V )
exc X
env ρ ⊢ E1 −−−→ E1′
exc some(V )
env ρ ⊢ throw(V ) −−−−−−−→ stuck
exc X
env ρ ⊢ apply(E1 , E2 ) −−−→ apply(E1′ , E2 )
exc none
env ρ ⊢ E −−−−−→ E ′
exc none
env ρ ⊢ catch(E, H) −−−−−→ catch(E ′ , H)
exc some(V )
env ρ ⊢ E −−−−−−−→ E ′
exc none
env ρ ⊢ catch(E, H) −−−−−→ apply(H,V )
val(V )
exc none
env ρ ⊢ catch(V, H) −−−−−→ V
exc X
env ρ ⊢ E −−−→ E ′
val(V )
exc X
env ρ ⊢ apply(V, E) −−−→ apply(V, E ′ )
val(V )
exc X
env ({I 7→ V }/ρ ) ⊢ E −−−→ E ′
exc X
env ⊢ apply(closure(ρ , I, E),V ) −−−→
apply(closure(ρ , I, E ′ ),V )
val(V1 )
val(V2 )
exc none
env ρ ⊢ apply(closure(ρ , I,V1 ),V2 ) −−−−−→ V1
Figure 3: SOS rules for lambda calculus with exception handling.
Neil Sculthorpe, Paolo Torrini & Peter D. Mosses
4
67
I-MSOS Specifications of Control Operators
We now present a dynamic semantics for control operators in the MSOS framework. We specify control
and prompt directly, and then specify shift, reset and call/cc in terms of control and prompt. Our approach
is signal-based in a similar manner to the I-MSOS specifications of exceptions (Figure 1): a control
operator emits a signal when executed, and a delimiter catches that signal and handles it. Note that there
is no implicit top-level delimiter around a funcon program—a translation to funcons from a language
that does have an implicit top-level delimiter should insert an explicit top-level delimiter.
4.1 Overview of our Approach
Whether the semantics of control operators can be specified using MSOS has been considered an open
problem ([30, Section 2]). We suspect that this is because there is no explicit representation of a term’s
context in the MSOS framework—any given rule only has access to the current subterm and the contents
of any semantic entities—so it is not immediately obvious how to capture the context as an abstraction.
Our approach is to construct the current continuation of a control operator in the rule for its enclosing
delimiter. We achieve this by exploiting the way that a small-step semantics, for each step of computation, builds a derivation tree from the root of the program term to the current operation. Thus, for any
step at which a control operator is executed, not only will a rule for the control operator be part of the
derivation, but so too will a rule for the enclosing delimiter. At each such step, the current continuation
corresponds to an abstraction of the control operator (and its argument) from the subterm of the enclosing
delimiter, and thus can be constructed from that subterm.
We represent reified continuations as first-class abstractions, using the lambda funcon from Section 3. Constructing the abstraction is achieved in two stages: the rule for control replaces the occurrence of control (and its argument) with a fresh identifier, and the rule for prompt constructs the
abstraction from the updated subterm. At a first approximation, this suggests the following rules:
fresh-id(I)
ctrl some(F,I)
(13)
control(F) −−−−−−−−→ bv(I)
ctrl some(F,I)
E −−−−−−−−→ E ′
ctrl none
K = lambda(I, E ′ )
(14)
prompt(E) −−−−−→ prompt(apply(F, K))
The side condition fresh-id(I) requires that the identifier I introduced by this rule does not already
occur in the program. Rule 13 replaces the term control(F) with bv(I), and emits a signal (ctrl) containing the function F and the identifier I. The signal is then caught and handled by prompt in Rule 14. The
abstraction K representing the continuation of the executed control operator is constructed by combining
I with the updated subterm E ′ (which will now contain bv(I) in place of control(F)).
4.2 The Auxiliary Environment
There is one problem with the approach we have just outlined, which is that the identifier I is introduced
dynamically when the control operator executes, by which time closures may have already formed. In
particular, if control occurs inside the body of a lambda, and the enclosing prompt is outside that
lambda, then the bv(I) funcon would be introduced inside a closure that has already formed, and hence
A Modular Structural Operational Semantics for Delimited Continuations
68
does not contain a binding for I. For example, consider the evaluation of the following term:
prompt(lambda(x, control(lambda(k, bv(k)))))
→
{ by (7) }
prompt(closure(0,
/ x, control(lambda(k, bv(k)))))
→
{ by (14) }
prompt(apply(lambda(k, bv(k)), lambda(i, closure(0,
/ x, bv(i)))))
The occurrence of bv(i) is now inside a closure containing an empty environment. Were that closure to
be applied (say if this was a subterm of a larger program), then the body of the closure would get stuck,
as Rule 12 would provide an environment containing only x, which Rule 6 could not match.
This problem arises as a consequence of our choice to specify the semantics of lambda calculus using
environments and closures. If we had instead given a semantics using substitution, then this problem
would not have arisen. However, we prefer to use environments because they enable a more modular
specification: a substitution-based semantics requires substitution to be defined over every construct in
the language. Moreover, environments allow straightforward semantics for dynamic scope.
Our solution is to introduce an auxiliary environment that is not captured in closures. Figure 4
specifies aux-bv(I), which looks up the identifier I in this auxiliary environment, and aux-let-in(I,V, E),
which binds the identifier I to the value V in the auxiliary environment and scopes that binding over
the expression E. We make use of these funcons in the next subsection, where we give our complete
specification of control and prompt.
µ (I) = V
aux-env µ ⊢ aux-bv(I) → V
(15)
E1 → E1′
aux-let-in(I, E1 , E2 ) → aux-let-in(I, E1′ , E2 )
(16)
val(V )
aux-env ({I 7→ V }/µ ) ⊢ E → E ′
aux-env µ ⊢ aux-let-in(I,V, E) → aux-let-in(I,V, E ′ )
(17)
val(V1 )
val(V2 )
aux-let-in(I,V1 ,V2 ) → V2
(18)
Figure 4: I-MSOS rules for bindings in the auxiliary environment.
4.3 Dynamic Semantics of control and prompt
We specify control as follows:
E → E′
control(E) → control(E ′ )
val(F)
fresh-id(I)
ctrl some(F,I)
control(F) −−−−−−−−→ aux-bv(I)
(19)
(20)
Neil Sculthorpe, Paolo Torrini & Peter D. Mosses
69
Rule 19, in combination with the val(F) premise on Rule 20, ensures that the argument function is
evaluated to a closure before Rule 20 can be applied. Notice that Rule 20 uses aux-bv, in contrast to the
preliminary Rule 13 which used bv.
We then specify prompt as follows:
val(V )
prompt(V ) → V
(21)
ctrl none
E −−−−−→ E ′
ctrl none
(22)
prompt(E) −−−−−→ prompt(E ′ )
ctrl some(F,I)
E −−−−−−−−→ E ′
K = lambda(I, aux-let-in(I, bv(I), E ′ ))
ctrl none
(23)
prompt(E) −−−−−→ prompt(apply(F, K))
Rule 21 is the case when the argument is a value; the prompt is then discarded. Rule 22 evaluates the
argument expression while no ctrl signal is being emitted by that evaluation. Rule 23 handles the case
when a ctrl signal is detected, reifying the current continuation and passing it as an argument to the
function F. Notice that, unlike in the preliminary Rule 14, I is rebound using aux-let-in.
Rules 19–23 are our complete I-MSOS specification of the dynamic semantics of control and
prompt, relying only on the existence of the lambda-calculus and auxiliary-environment funcons from
figures 2 and 4. These rules are modular: they are valid independently of whether the control operators
coexist with a mutable store, exceptions, input/output signals, or other semantic entities. Except for the
use of an auxiliary environment, our rules correspond closely to those in specifications of control and
prompt based on evaluation contexts [18, 24]. However, our rules communicate between control and
prompt by emitting signals, and thus do not require evaluation contexts. In Section 5, we present a proof
of equivalence between our specification and a conventional one based on evaluation contexts.
4.4 Dynamic Semantics of shift and reset
The shift operator differs from control in that every application of a reified continuation is implicitly
wrapped in a delimiter, which has the effect of separating the context of that application from its inner
context [5]. This difference between control and shift is analogous to that between dynamic and static
scoping, insofar as with shift, the application of a reified continuation cannot access its context, in the
same way that a statically scoped function cannot access the environment in which it is applied.
A shift funcon can be specified in terms of control as follows:
E → E′
shift(E) → shift(E ′ )
val(F)
fresh-id(K)
fresh-id(X )
shift(F) → control(lambda(K, apply(F, lambda(X , reset(apply(bv(K), bv(X )))))))
(24)
(25)
The key point is the insertion of the reset delimiter; the rest of the lambda-term is merely an η -expansion
that exposes the application of the continuation K so that the delimiter can be inserted (following [5]).
Given this definition of shift, the reset delimiter coincides exactly with prompt:
reset(E) → prompt(E)
(26)
A Modular Structural Operational Semantics for Delimited Continuations
70
Alternatively, the insertion of the extra delimiter could be handled by the semantics of reset rather
than that of shift:
val(V )
(27)
reset(V ) → V
ctrl none
E −−−−−→ E ′
(28)
ctrl none
reset(E) −−−−−→ reset(E ′ )
ctrl some(F,I)
E −−−−−−−−→ E ′
K = lambda(I, reset(aux-let-in(I, bv(I), E ′ )))
ctrl none
(29)
reset(E) −−−−−→ reset(apply(F, K))
The only difference between rules 21–23 and rules 27–29 (other than the funcon names) is the definition
of K in Rule 29, which here has a delimiter wrapped around the body of the continuation. Given this
definition of reset, the shift operator now coincides exactly with control:
shift(E) → control(E)
(30)
This I-MSOS specification in Rules 27–30 is similar to the evaluation-context based specification of shift
and reset in [24, Section 2].
4.5 Dynamic Semantics of abort and call/cc
The call/cc operator is traditionally undelimited: it considers the current continuation to be the entirety
of the rest of the program. In a setting with delimited continuations, this can be simulated by requiring
there to be a single delimiter, and for it to appear at the top-level of the program. Otherwise, the two
distinguishing features of call/cc relative to control and shift are first that an applied continuation never
returns, and second that if the body of call/cc does not invoke a continuation, then the current continuation
is applied to the result of the call/cc application when it returns.
To specify call/cc, we follow Sitaram and Felleisen [34, Section 3] and first introduce an auxiliary
operator abort, and then specify call/cc in terms of control, prompt and abort. The purpose of abort is to
terminate a computation (up to the innermost enclosing prompt) with a given value:
E → E′
abort(E) → abort(E ′ )
(31)
val(V )
fresh-id(I)
abort(V ) → control(lambda(I,V ))
(32)
We achieve the first distinguishing feature of call/cc by placing an abort around any application of a continuation (preventing it from returning a value), and we achieve the second by applying the continuation
to the result of the F application (which resumes the current continuation if F returns a value):
E → E′
callcc(E) → callcc(E ′ )
val(F)
fresh-id(K)
fresh-id(X )
callcc(F) →
control(lambda(K, apply(bv(K), apply(F, lambda(X , abort(apply(bv(K), bv(X ))))))))
(33)
(34)
Neil Sculthorpe, Paolo Torrini & Peter D. Mosses
71
4.6 Other Control Effects
In Section 3 we presented a direct specification of exception handling using a dedicated semantic entity.
If throw and catch (Figure 1) were used in a program together with the control operators from this
section, this would give rise to two sets of independent control effects, each with independent delimiters.
An alternative would be to specify exception handling indirectly in terms of the control operators (e.g.
following Sitaram and Felleisen [34]), in which case the delimiters and semantic entity would be shared.
MSOS can specify either approach, as required by the language being specified.
Beyond the control operators discussed in this section, further and more general operators for manipulating delimited continuations exist, such as those of the CPS hierarchy [13]. These are beyond the
scope of this paper, and remain an avenue for future work.
5
Adequacy
Our SOS model of call-by-value lambda calculus extended with delimited control, which we have presented using I-MSOS rules, is provably equivalent to one based on the reduction semantics (RS) of
lambda terms where the evaluation strategy is specified using evaluation contexts. Reduction models for
delimited control based on evaluation contexts were originally introduced in [18] and refined in [24].
The adequacy proof in this section (Prop. 8) is carried out with respect to our version of those models in
a formalism that we call RC.
Our SOS model differs from reduction models in the framework it relies on. In particular, our SOS
model uses environments and signals, whereas RC uses substitution and evaluation contexts. Moreover,
there is a difference in the notion of value: in our SOS model function application is computed using
closures, whereas RS uses β -reduction and substitution. In order to focus on the operational content of
the models, it is convenient to get above these differences. We achieve this by embedding SOS in RS
with explicit congruence rules (an embedding that we call LS), and by lifting RC to an environment-based
formalism (called LR). We define a notion of adequacy between two systems, as an input-output relation,
parametric in a translation. We show adequacy of two systems by proving that they are derivationally
equivalent (in the sense of a step-wise relation), reasoning by induction on the structure of derivations.
Our adequacy proof for SOS and RC is split into three main parts: the equivalence of SOS and LS, of
LS and LR, and of LR and RC. A more challenging approach would involve giving a formal derivation
of an RC model from SOS along the lines of [11], but that goes beyond the scope of this paper.
Here we intend to focus on equivalence with respect to delimited control. Given the equivalence
between SOS and RC with respect to call-by-value lambda calculus (λV), we show that SOS and RC
are equivalent with respect to the extension of λV with delimited control (λDC). More specifically,
we define a syntactic representation of environments (standard and auxiliary ones) using contexts and
lambda terms. We use this representation to define LS as a lambda-term encoding of SOS. Adequacy
between SOS and LS is provable with respect to a simple translation relation.
We define LR as an environment-based version of RC obtained by lambda-lifting. We consider two
distinct extensions of the LR model of λV with delimited control. The first one, which we call LRDC, uses the lifted control rules of the original RC model, and thus equivalence with the RC model is
straightforward. The second one, which we call LX-DC, uses the LS version of the SOS control rules.
The difference between the LS model of λDC and LX-DC, which are provably equivalent, boils down
to that between SOS transitions, based on congruence rules and also using closures, and RC transitions,
based on evaluation contexts and using only lambda expressions. The adequacy of LX-DC and LR-DC,
proved with respect to the identity translation (Prop. 7), gives us the result of primary interest.
72
A Modular Structural Operational Semantics for Delimited Continuations
5.1 Reduction Semantics
Our presentation of reduction semantics with evaluation contexts (RC) follows the main lines of the λDC
model in [24]. Under the assumption that we only evaluate closed expressions, values and terms can be
defined as follows:
V = lambda(I, E)
(35)
E = V | bv(I) | apply(E, E) | control(E) | prompt(E)
(36)
A general notion of a context as a term with a hole can be defined by the following grammar:
C = [] | lambda(I,C) | apply(C, E) | apply(E,C) | prompt(C) | control(C)
(37)
The call-by-value (CBV) evaluation strategy can be specified using the more restrictive notion of a CBV
context (Q):
Q = [] | apply(Q, E) | apply(V, Q) | prompt(Q) | control(Q)
(38)
In order to represent delimited continuations, an even more restrictive notion is needed: a pure context
(P-context), which is a CBV-context that does not include control delimiters [24]:
P = [] | apply(P, E) | apply(V, P) | control(P)
(39)
The meta-linguistic notation C[E] (Q[E], P[E]) is used to represent a term factored into a context and
the subterm that fills the hole—we can think of this as a form of term annotation. This factorisation is
unique for the cases that we are considering.
The only reduction rules needed to specify λV are β -reduction and context propagation. These can
be presented as follows (giving us the RC-V model), using { 7→ } as meta-level notation for captureavoiding uniform substitution:
apply(lambda(I, E),V ) −→ E{bv(I) 7→ V }
(40)
E −→ E ′
Q[E] −→ Q[E ′ ]
(41)
The reduction rules for prompt and control can be formulated as follows (giving us the RC-DC model),
making use of P-contexts:
prompt(V ) −→ V
(42)
val(F)
K = lambda(I, P[bv(I)])
fresh-id(I)
prompt(P[control(F)]) −→ prompt(apply(F, K))
(43)
In a system based on reduction semantics, observational equivalence can be defined as the smallest congruence relation ≡ on terms that extends reduction equivalence with functional extensionality,
i.e. such that
∀ V. apply(F,V ) ≡ apply(F ′ ,V )
F ≡ F′
In presenting models based on RS, we typeset all construct names in sans-serif. We refer to SOS
values as ValSOS and to RC ones as ValRC . When needed, we subscript −→ and ≡ accordingly. We define
derivational equivalence and adequacy with respect to a translation relation (not mentioned in the case
that it is an identity), as follows.
Neil Sculthorpe, Paolo Torrini & Peter D. Mosses
73
Def. 1 Given two systems A and B, respectively defined on languages LA and LB with values ValA and
ValB , and a one-to-one relation R ⊆ (LA , LB ), we say that
1. A and B are adequate with respect to R (A ∼R B) whenever the following hold:
A) If E −→∗A V , with V ∈ ValA , then there exist Ea′ , Ea′′ ∈ LA , Eb′ , Eb′′ ∈ LB , V ′ ∈ ValB s.t.
E ≡A Ea′ , V ≡A Ea′′ , R(Ea′ , Eb′ ), R(Ea′′ , Eb′′ ), Eb′′ ≡B V ′ and Eb′ −→∗B V ′
B) If E −→∗B V , with V ∈ ValB , then there exist Ea′ , Ea′′ ∈ LA , Eb′ , Eb′′ ∈ LB , V ′ ∈ ValA s.t.
E ≡B Eb′ , V ≡B Eb′′ , R(Ea′ , Eb′ ), R(Ea′′ , Eb′′ ), Ea′′ ≡A V ′ and Ea′ −→∗B V ′
2. A and B are derivationally equivalent with respect to R whenever the following hold:
A) E1 −→A E2 whenever there exist E3 , E4 , E1′ , E2′ , E3′ , E4′ s.t. E1 ≡A E3 , E2 ≡A E4 , R(E3 , E3′ ),
R(E4 , E4′ ), E3′ ≡B E1′ , E4′ ≡B E2′ and E1′ −→∗B E2′
B) E1 −→B E2 whenever there exist E3 , E4 , E1′ , E2′ , E3′ , E4′ s.t. E1 ≡B E3 , E2 ≡B E4 , R(E3′ , E3 ),
R(E4′ , E4 ), E3′ ≡A E1′ , E4′ ≡A E2′ and E1′ −→∗A E2′
We define relational composition as R1 ◦ R2 = λ xy. ∃z. R1 (x, z) ∧ R2 (z, y).
5.2 Representing SOS as LS
In this section we define LS, as an encoding of SOS in lambda terms. Unlike reduction semantics, our
SOS models rely internally on a linguistic extension to account for closures and the auxiliary environment
notation. For this reason, we need an extended internal language, including the following additional
constructs: closure(ρ , I, E) for closures, aux-bv(I) for auxiliary identifier lookup, and aux-let-in(I, E, E)
for auxiliary let bindings; these constructs are not meant to be included in the source language definition.
In each expression aux-let-in(I, E, E ′ ), we require I to be used at most once in E ′ .
In order to represent environments, we introduce a notion of R-context:
R = [] | apply(lambda(I, R),V )
(44)
We tacitly assume that all bound variables in R are distinct. Each SOS transition specified by
env ρ ⊢ E −→ E ′
can be embedded as
Rρ [E] −→ Rρ [E ′ ]
where, for ρ = {I1 7→ V1 , . . . , In 7→ Vn }, Rρ = apply(lambda(In , (. . . , apply(lambda(I1 , []),V1 ), . . .)),Vn ).
We silently assume permutation in R-contexts. We introduce M-contexts to represent the auxiliary environment, in a similar manner to R-contexts, though using aux-let-in. In order to represent signals, we
extend this notion to one of S-context, introducing a new ternary value constructor ctrl, which is not part
of the expression definition but only of the RS representation of SOS.
M = [] | aux-let-in(I,V, M)
(45)
S = M | ctrl(V, I, M)
(46)
SOS transitions specified by
ctrl none
A) aux-env µ , env ρ ⊢ E −−−−−→ E ′
ctrl some(F,I)
B) aux-env µ , env ρ ⊢ E −−−−−−−−→ E ′
can be represented, respectively, as
A) Mµ [Rρ [E]] −→ Mµ [Rρ [E ′ ]]
B) Mµ [Rρ [E]] −→ ctrl(F, I, Mµ [Rρ [E ′ ]])
A Modular Structural Operational Semantics for Delimited Continuations
74
where, for µ = {I1 7→ V1 , . . . , In 7→ Vn }, Mµ = aux-let-in(In ,Vn , . . . aux-let-in(I1 ,V1 , [])). We assume that
S-bound variables are distinct from each other and from all the R-bound ones. As with R-contexts, we
silently assume permutation for M-contexts.
In this way, we define a one-to-one translation relation T( , ) between SOS configurations and LS
expressions. Applying the translation to the SOS rules gives us the LS models for λV and λDC (resp.
LS-V and LS-DC). In particular, the rule for bv, which uses the environment, can be expressed as follows:
S[apply(lambda(I, R[bv(I)]),V )] −→ S[apply(lambda(I, R[V ]),V )]
(47)
The aux-bv rule, which uses the auxiliary environment, takes the following form:
aux-let-in(I,V, M[R[aux-bv(I)]]) −→ aux-let-in(I,V, M[R[V ]])
(48)
The LS model for λDC extends LS-V with three more rules: the SOS rule for control,
val(F)
fresh-id(I)
M[R[control(F)]] −→ ctrl(F, I, M[R[aux-bv(I)]])
(49)
an additional congruence rule (which can only match Rule 49),
M[R[E]] −→ ctrl(F, I, M[R[E ′ ]])
M[R[P[E]]] −→ ctrl(F, I, M[R[P[E ′ ]]])
(50)
and an encoding of the SOS rule for control-in-prompt (Rule 23).
M[R[E]] −→ ctrl(F, I ′ , M[R[E ′ ]])
K = lambda(I, aux-let-in(I ′ , bv(I), E ′ ))
M[R[prompt(E)]] −→ M[R[prompt(apply(F, K)]]
fresh-id(I)
(51)
Since there is a one-to-one correspondence between LS and SOS transitions (treating R- and Mpermutations as silent transitions), and taking for simplicity the identity relation modulo reordering of
the environments as observational equivalence in SOS, the following is straightforward.
Prop. 1 The LS model of λV and the corresponding SOS one are derivationally equivalent and adequate
with respect to the translation T , and similarly for the LS and SOS models of λDC.
Proof: First we prove derivational equivalence, which is straightforward for λV. The LS control rules
correspond to the SOS ones, given the representation of the auxiliary environments and signals. Adequacy follows as the definition of value is the same in all these systems.
5.3 Lifting RC to LR
To facilitate comparison with LS, we define LR as an environment-based version of RC, using R- and
M-contexts to represent environments as in LS, and also extend the language with aux-let-in and closure.
In the LR models, the reduction rules can be specified as in RC, relying on evaluation contexts. For the
way aux-let-in and closure are used, no change is needed in the definition of context. However, since
reduction steps now have to be lifted by R- and M-contexts, we replace the single context propagation
rule that sufficed in RC with the four following rules: lifting, lifted congruence and R- and M-lowering.
Neil Sculthorpe, Paolo Torrini & Peter D. Mosses
75
E −→ E ′
M[R[E]] −→ M[R[E ′ ]]
(52)
M[R[E]] −→ M[R[E ′ ]]
M[R[Q[E]]] −→ M[R[Q[E ′ ]]]
(53)
M[apply(lambda(I, R[E]),V )] −→ M[apply(lambda(I, R[E ′ ]),V )]
M[R[apply(lambda(I, E),V )]] −→ M[R[apply(lambda(I, E ′ ),V )]]
(54)
aux-let-in(I,V, M[R[E]]) −→ aux-let-in(I,V, M[R[E ′ ]])
M[R[aux-let-in(I,V, E)]] −→ M[R[aux-let-in(I,V, E ′ )]]
(55)
We assume that LR models include R- and M-permutations, as well as the bv and aux-bv rules, and
Rule 50. Notice that evaluation can also apply to open terms, however we do not need to change our
definition of value, as substitution of free variables is dealt with as in LS, by the bv rule. We also need
the following rule to deal with closures:
ρ = {I1 7→ V1 , . . . In 7→ Vn }
free-vars(E) ⊆ {I1 , . . . , In }
closure(ρ , I, E) −→ apply(lambda(In , (. . . , apply(lambda(I1 , lambda(I, E)),V1 ), . . .)),Vn )
(56)
This gives us the LR model of λV (LR-V). We can extend this model with rules for delimited control
in two ways: to simulate RC (LR-DC), or to simulate SOS (LX-DC). The LR-DC rules for prompt and
control are those based on the RC one (i.e. they are the lifted version of Rules 42 and 43), and they do
not involve any use of the auxiliary notation. LR-DC is the extension of LR-V with these rules. On the
other hand, the LX-DC model is obtained by extending LR-V with the LS control rules (Rules 49 and
51, which rely on the auxiliary notation).
The following can be proved for all the systems we are considering, i.e. with respect to ≡X where
X ∈ {LS-V, LR-V, LS-DC, LX-DC, LR-DC}.
Prop. 2 A) V ∈ ValSOS whenever there exists V ′ ∈ ValRC such that V ≡X V ′ .
B) S[Rρ [apply(lambda(I, E),V1 )]] −→LS V2 with V1 ,V2 ∈ ValSOS whenever there exist V3 ,V4 ∈
ValRC such that S[[apply(closure(ρ , I, E),V3 )]] −→LR V4 , with V1 ≡X V3 and V2 ≡X V4 .
The following provable equivalences correspond respectively to the bv rule, to β -reduction, and to
aux-let-in elimination, for ≡X as before:
apply(lambda(I, bv(I)),V ) ≡X apply(lambda(I,V ),V )
(57)
apply(lambda(I, E),V ) ≡X E{bv(I) 7→ V }
(58)
aux-let-in(I,V, E) ≡X E{aux-bv(I) 7→ V }
(59)
The following can now be proved:
Prop. 3 apply(lambda(I, aux-let-in(I ′ , bv(I), P[aux-bv(I ′ )])),V ) ≡X apply(lambda(I, P[bv(I)]),V )
Proof: apply(lambda(I, aux-let-in(I ′ , bv(I), P[aux-bv(I ′ )])),V ) ≡X aux-let-in(I ′ ,V, P[aux-bv(I ′ )]), by
Equiv. 58.
aux-let-in(I ′ ,V, P[aux-bv(I ′ )]) ≡X P[V ], by Equiv. 59, observing that aux-bv(I ′ ) cannot occur free
in P, as it must be used at most once in P[aux-bv(I ′ )].
P[V ] ≡X apply(lambda(I, P[bv(I)]),V ), by Equiv. 58.
The following is an immediate consequence of Prop. 3, applying functional extensionality.
lambda(I, aux-let-in(I ′ , bv(I), P[aux-bv(I ′ )])) ≡X lambda(I, P[bv(I)])
(60)
76
A Modular Structural Operational Semantics for Delimited Continuations
5.4 Adequacy of SOS and RC
We first show that the LR models and the RC ones are equivalent.
Prop. 4 LR-V and RC-V are adequate, and so too are LR-DC and RC-DC.
Proof: The language of RC is included in that of LR, hence we can take the identity on RC (denoted
by IdRC ) as the translation. The LR models can be obtained by a lambda-lifting refactoring of the RC
models, and this gives equivalent systems, under a change of the evaluation strategy that affects only
the top level. We prove derivational equivalence by induction on the structure of derivations, relying
on Equiv. 59 and Rule 56 to eliminate the additional LR syntax, observing that aux-let-in and closure
are inessential in LR-DC (they can only be eliminated without leading to any new values). Adequacy
follows immediately as values are defined in the same way in the two systems.
We consider the relationship between the different representations of λV.
Prop. 5 The LS model of λV and the corresponding LR one are adequate.
Proof: The two models use the same language, hence we can take the identity translation. They differ
on congruence rules and reduction of function application. Congruence rules in LR are expressed using
CBV-contexts, unlike in LS, but both are equivalent specifications of CBV. For equivalence with respect
to values and function application, we rely on Prop. 2.
We extend this result to SOS-style delimited control.
Prop. 6 LS-DC and LX-DC are adequate.
Proof: We first prove derivational equivalence with respect to identity using Prop. 5 and the fact that the
two extensions are obtained by adding the same rules.
We finally compare SOS-style and RC-style delimited control.
Prop. 7 LX-DC and LR-DC are derivationally equivalent and adequate.
Proof: We prove derivational equivalence by induction on the structure of derivations, with respect to
the identity translation. The two systems are equivalent up to λV by Prop. 4, so the only interesting case
is delimited control, in which respect LX-DC stepwise behaves as LS-DC. The lifted version of the RC
prompt rule (Rule 42) is in both systems. Rules 49 and 50 can be added to LR-DC without expanding the
set of derivable values. Thus the only possible difference between the two systems is between the natural
LR control rule (the lifted version of Rule 43) of LR-DC, and the LS control-in-prompt rule (Rule 51)
of LX-DC. We now show that the two rules are interderivable (i.e. given the system with one rule, the
other one is derivable). First we observe that, by Eq. 60, the specification of the continuation K in either
rule is equivalent to that in the other, and therefore interchangeable.
From S to R: in order to derive the LR-DC rule from the LX-DC one, we observe that a lifted
expression M[R[P[control(F)]]], where F is a value, can be reduced to ctrl(F, I, M[R[P[aux-bv(I)]]]) in
LX-DC, using the control rule (Rule 49), and the applicable congruence rule (Rule 50). This gives us
the premise for the application of the LX-DC control-in-prompt rule to M[R[prompt(P[control(F)])]] in
a way that simulates the LR-DC control rule.
From R to S: in LX-DC (as in LS-DC) a one-step transition from M[R[E]] to ctrl(F, I, M[R[E ′ ]]) is
only possible provided E ≡ P[control(F)] and E ′ ≡ P[aux-bv(I)] for some P (possibly relying on the
conversion of closures to function applications). Therefore, the LR-DC control rule can be applied to
M[R[prompt(E)]] to simulate Rule 51.
Neil Sculthorpe, Paolo Torrini & Peter D. Mosses
77
Diagramatically, the overall proof can be presented as follows (where vertical arrows denote model
inclusion).
SOS-DC ∼T LS-DC ∼Id LX-DC ∼Id LR-DC ∼IdRC RC-DC
↑
↑
↑
↑
↑
SOS-V
∼T LS-V
∼Id LR-V
=
LR-V
∼IdRC RC-V
Prop. 8 SOS-DC and RC-DC are derivationally equivalent and adequate with respect to T ◦ IdRC
Proof: based on Props. 1, 4, 6, 7, using the fact that adequacy is transitive, by composition of the
translation relations, and by transitivity of observational equivalence.
6
Related Work
A direct way to specify control operators is by giving an operational semantics based on transition rules
and first-class continuations. We have taken this direct approach, though in contrast to most direct specifications of control operators (e.g. [18, 22, 23, 24, 30, 33]) our approach is based on emitting signals
via labelled transitions rather than on evaluation contexts. Control operators can also be given a denotational semantics by transformation to continuation-passing style (CPS) [12, 16, 31, 33], or a lower-level
operational specification by translation to abstract-machine code [6, 19]. At a higher level, algebraic
characterisations of control operators have been given in terms of equational theories [18, 23].
Denotationally, any function can be rewritten to CPS by taking the continuation (itself represented as
a function) as an additional argument, and applying that continuation to the value the function would have
returned. A straightforward extension of this transformation [13] suffices to express call/cc, shift and
reset; however, more sophisticated CPS transformations are needed to express control and prompt [33].
Felleisen’s [19] initial specification of control and prompt used a small-step operational semantics
without evaluation contexts. However, this specification otherwise differs quite significantly from ours,
being based on exchange rules that push control outwards through the term until it encounters a prompt.
As an exchange rule has to be defined for every other construct in the language, this approach is inherently not modular. Later specifications of control and prompt used evaluation contexts and algebraic
characterisations based on the notion of abstract continuations [18], where continuations are represented
as evaluation contexts and exchange rules are not needed. Felleisen [19] also gave a lower-level operational specification based on the CEK abstract machine, where continuations are treated as frame stacks.
The shift and reset operators were originally specified denotationally, in terms of CPS semantics [12,
13]. Continuations were treated as functions, relying on the meta-continuation approach [12] which
distinguishes between outer and inner continuations. Correspondingly, the meta-continuation transformation produces abstractions that take two continuation parameters, which can be further translated to
standard CPS. A big-step style operational semantics for shift has been given by Danvy and Yang [15],
and a specification based on evaluation contexts has been given by Kameyama and Hasegawa [23], together with an algebraic characterisation.
Giving a CPS semantics to control is significantly more complex than for shift [33]. This is because
the continuations reified by shift are always delimited when applied, and so can be treated as functions,
which is not the case for control. Different approaches to this problem have been developed, including
abstract continuations [18], the monadic framework in [16], and the operational framework in [6]. Relying on the introduction of recursive continuations, Shan [33] provides an alternative approach based
on a refined CPS transform. Conversely, the difference between control and shift can manifest itself
A Modular Structural Operational Semantics for Delimited Continuations
78
quite intuitively in the direct specification of these operators—whether in our I-MSOS specifications
(Section 4.4), or in specifications using evaluation contexts [18, 23, 24, 33].
As shown by Filinski [20], shift can be implemented in terms of call/cc and mutable state, and from
the point of view of expressiveness, any monad that is functionally expressible can be represented in
lambda calculus with shift and reset. Moreover, control and shift are equally expressive in the untyped
lambda calculus [33]. A direct implementation of control and shift has been given by Gasbichler and
Sperber [21]. A CPS-based implementation of control operators in a monadic framework has been given
by Dyvbig et al [16]. A semantics of call/cc based on an efficient implementation of evaluation contexts
is provided in the K Framework [30].
7
Conclusion
We have presented a dynamic semantics for control operators in the MSOS framework, settling the
question of whether MSOS is expressive enough for control operators. Our definitions are concise and
modular, and do not require the use of evaluation contexts. Definitions based on evaluation contexts are
often even more concise than the corresponding MSOS definitions, since a single alternative in a contextfree grammar for evaluation contexts subsumes an entire MSOS rule allowing evaluation of a particular
subexpression. However, such grammars are significantly less modular than MSOS rules: adding a new
control operator to a specified language may require duplication of a (potentially large) grammar [17,
e.g. pages 141–142]. (This inherent lack of modularity of evaluation context grammars is addressed in
the PLT Redex tools by the use of ellipsis.)
We initially validated our specifications through a suite of 70 test programs, which we accumulated
from examples in the literature on control operators ([1, 2, 4, 6, 9, 10, 12, 18, 19, 22, 33]). The language we used for testing was Caml Light, a pedagogical sublanguage of a precursor to OCaml, for
which we have an existing translation to funcons from a previous case study [8]. We extended Caml
Light with control operators, and specified the semantics of those operators as direct translations into
the corresponding funcons presented in this paper. The generated funcon programs were then tested by
our prototype funcon interpreter, which directly interprets their I-MSOS specifications. The suite of test
programs, and our accompanying translator and interpreter, are available online [32].
While the test programs demonstrated that we had successfully specified a control operator that
behaves very similarly to the operator control described in the literature, they did not prove that we had
specified exactly the same operator. We addressed this in Section 5, where we proved that our MSOS
specification is equivalent to a conventional specification using a reduction semantics based on evaluation
contexts (e.g. [18, 24]).
Acknowledgments: We thank Casper Bach Poulsen, Ferdinand Vesely and the anonymous reviewers
for feedback on earlier versions of this paper. We also thank Martin Churchill for his exploratory notes
on adding evaluation contexts to MSOS, and Olivier Danvy for suggesting additional test programs. The
reported work was supported by EPSRC grant (EP/I032495/1) to Swansea University for the PL AN C OMP S project and by EU funding (Horizon 2020, grant 640954) to KU Leuven for the GRAC E FUL
project.
Neil Sculthorpe, Paolo Torrini & Peter D. Mosses
79
References
[1] H. Abelson, R. K. Dybvig, C. T. Haynes, G. J. Rozas, N. I. Adams IV, D. P. Friedman, E. Kohlbecker,
G. L. Steele Jr., D. H. Bartley, R. Halstead, D. Oxley, G. J. Sussman, G. Brooks, C. Hanson, K. M. Pitman
& M. Wand (1998): Revised5 Report on the Algorithmic Language Scheme. Higher-Order and Symbolic
Computation 11(1), pp. 7–105, doi:10.1023/A:1010051815785.
[2] Kenichi Asai & Yukiyoshi Kameyama (2007): Polymorphic Delimited Continuations. In: Asian Symposium
on Programming Languages and Systems, Lecture Notes in Computer Science 4807, Springer, pp. 239–254,
doi:10.1007/978-3-540-76637-7 16.
[3] Egidio Astesiano (1991): Inductive and Operational Semantics. In: Formal Description of Programming
Concepts, IFIP State-of-the-Art Reports, Springer, pp. 51–136. ISBN 978-3-540-53961-2.
[4] Malgorzata Biernacka, Dariusz Biernacki & Olivier Danvy (2005): An Operational Foundation for Delimited Continuations in the CPS Hierarchy. Logical Methods in Computer Science 1(2), pp. 1–39,
doi:10.2168/LMCS-1(2:5)2005.
[5] Dariusz Biernacki & Olivier Danvy (2006): A Simple Proof of a Folklore Theorem about Delimited Control.
Journal of Functional Programming 16(3), pp. 269–280, doi:10.1017/S0956796805005782.
[6] Dariusz Biernacki, Olivier Danvy & Chung-chieh Shan (2006): On the Static and Dynamic Extents of Delimited Continuations. Science of Computer Programming 60(3), pp. 274–297, doi:10.1016/j.scico.2006.01.002.
[7] Martin Churchill & Peter D. Mosses (2013): Modular Bisimulation Theory for Computations and Values. In:
International Conference on Foundations of Software Science and Computation Structures, Lecture Notes in
Computer Science 7794, Springer, pp. 97–112, doi:10.1007/978-3-642-37075-5 7.
[8] Martin Churchill, Peter D. Mosses, Neil Sculthorpe & Paolo Torrini (2015): Reusable Components of Semantic Specifications. In: Transactions on Aspect-Oriented Software Development XII, Lecture Notes in
Computer Science 8989, Springer, pp. 132–179, doi:10.1007/978-3-662-46734-3 4.
[9] William Clinger (1987): The Scheme Environment: Continuations. SIGPLAN Lisp Pointers 1(2), pp. 22–28,
doi:10.1145/1317193.1317197.
[10] Olivier Danvy (2006): An Analytical Approach to Programs as Data Objects. DSc thesis, Department of
Computer Science, Aarhus University. Available at http://cs.au.dk/~danvy/DSc/.
[11] Olivier Danvy (2008): Defunctionalized Interpreters for Programming Languages. In: International Conference on Functional Programming, ACM, pp. 131–142, doi:10.1145/1411204.1411206.
[12] Olivier Danvy & Andrzej Filinski (1989):
A Functional Abstraction
texts.
Technical Report 89/12, DIKU, University of Copenhagen.
http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.43.4822.
of
Typed ConAvailable at
[13] Olivier Danvy & Andrzej Filinski (1990): Abstracting Control. In: Conference on LISP and Functional
Programming, ACM, pp. 151–160, doi:10.1145/91556.91622.
[14] Olivier Danvy & Andrzej Filinski (1992): Representing Control: A Study of the CPS Transformation. Mathematical Structures in Computer Science 2(4), pp. 361–391, doi:10.1017/S0960129500001535.
[15] Olivier Danvy & Zhe Yang (1999): An Operational Investigation of the CPS Hierarchy. In: European
Symposium on Programming Languages and Systems, Lecture Notes in Computer Science 1576, Springer,
pp. 224–242, doi:10.1007/3-540-49099-X 15.
[16] R. Kent Dyvbig, Simon Peyton Jones & Amr Sabry (2007): A Monadic Framework for Delimited Continuations. Journal of Functional Programming 17(6), pp. 687–730, doi:10.1017/S0956796807006259.
[17] Matthias Felleisen, Robert Bruce Findler & Matthew Flatt (2009): Semantics Engineering with PLT Redex.
MIT Press. ISBN 9780262062756.
[18] Matthias Felleisen, Mitch Wand, Daniel Friedman & Bruce Duba (1988): Abstract Continuations: A Mathematical Semantics for Handling Full Jumps. In: Conference on LISP and Functional Programming, ACM,
pp. 52–62, doi:10.1145/62678.62684.
80
A Modular Structural Operational Semantics for Delimited Continuations
[19] Mattias Felleisen (1988): The Theory and Practice of First-Class Prompts. In: Symposium on Principles of
Programming Languages, ACM, pp. 180–190, doi:10.1145/73560.73576.
[20] Andrzej Filinski (1994): Representing Monads. In: Symposium on Principles of Programming Languages,
ACM, pp. 446–457, doi:10.1145/174675.178047.
[21] Martin Gasbichler & Michael Sperber (2002): Final Shift for Call/cc: Direct Implementation of
Shift and Reset.
In: International Conference on Functional Programming, ACM, pp. 271–282,
doi:10.1145/581478.581504.
[22] Carl A. Gunter, Didier Rémy & Jon G. Riecke (1995): A Generalization of Exceptions and Control in ML-like
Languages. In: International Conference on Functional Programming Languages and Computer Architecture,
ACM, pp. 12–23, doi:10.1145/224164.224173.
[23] Yukiyoshi Kameyama & Masahito Hasegawa (2003): A Sound and Complete Axiomatization of Delimited Continuations. In: International Conference on Functional Programming, ACM, pp. 177–188,
doi:10.1145/944705.944722.
[24] Yukiyoshi Kameyama & Takuo Yonezawa (2008): Typed Dynamic Control Operators for Delimited Continuations. In: International Symposium on Functional and Logic Programming, Lecture Notes in Computer
Science 4989, Springer, pp. 239–254, doi:10.1007/978-3-540-78969-7 18.
[25] Peter D. Mosses (2002): Pragmatics of Modular SOS. In: International Conference on Algebraic
Methodology and Software Technology, Lecture Notes in Computer Science 2422, Springer, pp. 21–40,
doi:10.1007/3-540-45719-4 3.
[26] Peter D. Mosses (2004): Modular Structural Operational Semantics. Journal of Logic and Algebraic Programming 60–61, pp. 195–228, doi:10.1016/j.jlap.2004.03.008.
[27] Peter D. Mosses & Mark J. New (2009): Implicit Propagation in Structural Operational Semantics. In:
Workshop on Structural Operational Semantics, Electronic Notes in Theoretical Computer Science 229(4),
Elsevier, pp. 49–66, doi:10.1016/j.entcs.2009.07.073.
[28] Peter D. Mosses & Ferdinand Vesely (2014): FunKons: Component-Based Semantics in K. In: International
Workshop on Rewriting Logic and Its Applications, Lecture Notes in Computer Science 8663, Springer, pp.
213–229, doi:10.1007/978-3-319-12904-4 12.
[29] Gordon D. Plotkin (2004): A Structural Approach to Operational Semantics. Journal of Logic and Algebraic Programming 60–61, pp. 17–139, doi:10.1016/j.jlap.2004.05.001. Reprint of Technical Report FN-19,
DAIMI, Aarhus University, 1981.
[30] Grigore Roşu & Traian Florin Şerbănuţă (2010): An Overview of the K Semantic Framework. Journal of
Logic and Algebraic Programming 79(6), pp. 397–434, doi:10.1016/j.jlap.2010.03.012.
[31] Amr Sabry & Matthias Felleisen (1993): Reasoning About Programs in Continuation-passing Style. LISP
and Symbolic Computation 6(3–4), pp. 289–360, doi:10.1007/BF01019462.
[32] Neil Sculthorpe, Paolo Torrini & Peter D. Mosses (2016): A Modular Structural Operational Semantics for
Delimited Continuations: Additional Material. Available at http://www.plancomps.org/woc2016.
[33] Chung-chieh Shan (2007): A Static Simulation of Dynamic Delimited Control. Higher-Order and Symbolic
Computation 20(4), pp. 371–401, doi:10.1007/s10990-007-9010-4.
[34] Dorai Sitaram & Matthias Felleisen (1990): Control Delimiters and their Hierarchies. LISP and Symbolic
Computation 3(1), pp. 67–99, doi:10.1007/BF01806126.
| 6cs.PL
|
arXiv:1605.00781v2 [math.GR] 27 May 2017
CONFIGURATION EQUIVALENCE IS NOT
EQUIVALENT TO ISOMORPHISM
ALI REJALI AND MEISAM SOLEIMANI MALEKAN
Abstract. Giving a condition for the the amenability of groups,
Rosenblatt and Willis, first introduced the concept of configuration. From the beginning of the theory, the question whether the
concept of configuration equivalence coincides with the concept of
group isomorphism was posed. We negatively answer this question
by introducing two non-isomorphic, solvable and hence amenable
groups which are configuration equivalent. Also, we will prove this
conjecture, due to Rosenblatt and Willis, that configuration equivalent groups, both include the free non-Abelian group of same rank
or not. We show that two-sided equivalent groups have same class
numbers.
1. Introduction and Definitions
In this paper, all groups were assumed to be finitely generated and
discrete. Let G be a group, we denote the identity of the group G
by eG . Given any finite (ordered) subset g = (g1 , . . . , gn ) of G, there
is a Cayley graph denoted by Γ := Γ(G, g), with vertices being the
elements of G and the directed edges being from g to gi g for g ∈ G and
i ∈ {1, . . . , n}.
let G be a group, g = (g1 , . . . , gn ) be an ordered generating set and
Γ = Γ(G, g) be its corresponded Cayley graph. Assume that E =
{E1 , . . . , Em } is a finite partition of G, which can be considered as a
coloring of Γ by m colors.
A configuration C is an (n + 1)-tuple of colors (c0 , . . . , cn ) with each
ck being one of the m colors, and there are x0 , x1 , . . . , xn ∈ G such
that xi is the color ci , i = 0, 1, . . . , n, and for each i, xi = gi x0 . In
this case, we may say that (x0 , x1 , . . . , xn ) has the configuration C.
Date: February 26, 2018.
2010 Mathematics Subject Classification. 20F05, 20F16, 20F24, 43A07.
Key words and phrases. Configuration, Two-sided Configuration, Group Isomorphism, Conjugacy Class, Class Number, Paradoxical Decomposition .
The first authors would like to express their gratitude toward Banach Algebra
Center of Excellence for Mathematics, University of Isfahan. We also thank Yves
de Cornulier for the ideas in the proof of Theorem 4.3.
1
CONFIGURATION EQUIVALENCE IS NOT EQUIVALENT TO ISOMORPHISM 2
Simultaneously considering the right multiplication, the concept of twosided configuration is reached; A two-sided configuration is a (2n + 1)tuple C = (c0 , c1 , . . . , c2n ) satisfying ci ∈ {1, . . . , m}, i = 0, 1, . . . , 2n,
and there exists x ∈ Ec0 such that gi x ∈ Eci and xgi ∈ Eci+n for each
i ∈ {1, . . . , n}.
First, the concept of configuration was introduced by Rosenblatt and
Willis to give a characterization for the amenability of groups and then,
to characterize normal sets, the concept of two-sided configuration was
suggested.
For g and E as above, we called (g, E) a configuration pair. The
set of configurations (two-sided configurations, resp.) corresponding to
the configuration pair (g, E) will be denoted by Con(g, E) (Cont (g, E),
resp.). The set of all configuration and two-sided configuration sets of
G, denoted by Con(G) and Cont (G), respectively.
The origin of the theory of configuration, is the conjecture raised in
[8], that the combinatorial properties of configurations can be used to
characterize various kinds of behavior of groups (like the group being
Abelian or the group containing a non-Abelian free subgroup). This
conjecture leads to the notion of (two-sided) configuration equivalence;
A group G is configuration contained in a group H, written G - H, if
Con(G) ⊆ Con(H), and two groups G and H are configuration equivalent, written G ≈ H, if Con(G) = Con(H). The concepts of being
two-sided configuration contained, and two-sided configuration equivalent are similarly defined, the notations used denoting these concepts
are “-t ” and “≈t ”.
In [1], the first steps of the theory were taken. It was shown there
that finiteness and periodicity are the properties which can be characterized by configuration. In that paper, the authors proved that for two
configuration equivalent groups, the isomorphism classes of their finite
quotients were the same. The finite index property can be extended to
Abelian quotient property (see [2]). Also, it was shown in [1] that two
configuration equivalent groups, should satisfy in the same semi-group
laws, and in [4], this result was generalized by proving that same group
laws should be established in configuration equivalent groups. Hence,
in particular, being Abelian and the group property of being nilpotent
of class c are, in particular, another properties which can be characterized by configuration (see [1] and [2]). In [2], it was shown that if
G ≈ H, and G is a torsion free nilpotent group of Hirsch length h, then
so is H. It was interesting to know the answer to the question whether
being FC-group is conserved by equivalence of configuration. In [2], this
question was answered under the assumption of being-nilpotent. In [4],
this question was affirmatively answered without any extra hypothesis.
CONFIGURATION EQUIVALENCE IS NOT EQUIVALENT TO ISOMORPHISM 3
In addition, it was shown in that paper that the solubility of a group G
can be recovered from Con G. In our recent paper, [5], we showed that
the notion of normality can be obtained from two-sided configuration
equivalence. Also, in the presence of normality, we showed that if G
and H are two-sided equivalent, and if G has a normal subgroup N,
which the quotient G/N is finitely presented and G has a “recognizable” configuration pair w.r.t. N, then H contains a normal subgroup
N, such that G/N ∼
= H/N. It was shown, that the class of “polynomial
type” groups – involving finite, Abelian, free and polycyclic groups –
satisfied the “recognizability” condition.
We also interested in investigating the question: For which subclasses
of the class G of all groups, does configuration equivalence coincide
with isomorphism? In [1], this question was answered positively for the
class of finite, free and Abelian groups. In [2], it was shown that those
groups with the form of Zn ×F , where Z is the group of integers, n ∈ N
and F is an arbitrary finite group, are determined up to isomorphism
by their configurations. In [2], it was proved that if G ≈ D∞ , where
D∞ is the infinite dihedral group, then G ∼
= D∞ .
In [4], we pointed out that it was the existence of “golden” configuration pairs which implied isomorphism. Indeed, we showed that, in
the class of finitely presented Hopfian groups with golden configuration
pair, configuration equivalence coincided with isomorphism.
In the light of two-sided configuration, it was proved in [5] that,
for “polynomial type” groups, and for groups with finite commutator
subgroup, the “≈t ” and “∼
=” are equivalent. Specially, for polycyclic
or FC–groups, these two notation are equivalent.
We will prove that if G and H are two configuration equivalent
groups, and if G contains Fn , the non-Abelian free group of rank n,
then H, also, contains Fn .
In the present paper, we define configuration sets for a finite sigma
algebras of a group, and with the help of them, we will show that twosided equivalence groups, have the same number of normal subgroups
of finite index n, n ∈ N. Also, if G ≈t H, then the cardinality of normal
subgroups which their quotient are polycyclic is the same in both G
and H. The class number- the number of different conjugacy classeswill be shown to be equal for two-sided equivalent groups. In the class
of two-sided equivalent groups which have a finite class number, we will
study some type of subgroups, and will show that in this class the set
of finite, and polycyclic subgroups should be the same.
CONFIGURATION EQUIVALENCE IS NOT EQUIVALENT TO ISOMORPHISM 4
The question whether (two-sided) configuration equivalence implies
isomorphism has been seen to be open since the beginning of the theory of configuration. We negatively answer this open question by presenting two non-isomorphic solvable and hence amenable groups, both
having the same (two-sided) configuration sets. Like our two recent
papers ([4] and [5]), we will use the following notation:
Notation. Let G be a group with g = (g1 , . . . , gn ) as its ordered subset.
Let p ∈ N, J ∈ {1, 2, . . . , n}p and ρ ∈ {±1}p . We denote the product
Qp
ρ(i)
i=1 gJ(i) by W (J, ρ; g). We call the pair (J, ρ) a representative pair in
g and W (J, ρ; g) a word corresponding to (J, ρ) in g.
When we speak of a representative pair, (J, ρ), we assume the same
number of components for J and ρ. The number of components of J,
denotes by n(J).
2. Configuration and Finite Sigma Algebras
What is really important in configuration is the image of subsets of
a group G, under left translations by finite subsets of G. So, it seems
that for sets Con(g, E), or Cont (g, E) of a group G, we can replace E
by a sigma algebra A, and indeed we can. we involve sigma algebras
in the theory of configuration as follows:
Let G be a group. There is a correspondence between the finite sigma
algebras of G, and finite partitions of G. Indeed, for a finite sigma
algebra A, the set of atoms of A, E := {E1 , . . . , Em } is a partition
of G. We denote the atomic sets of a sigma algebra A by atom(A).
Also, if C is a finite collection of subsets of G, we use σ(C) to denote
the sigma algebra generated by C. In the following, we always consider
sigma algebras to be finite.
Now, we try to rewrite our symbols in configuration for sigma algebras; For a sigma algebra A, we define Con(g, A) to be Con(g, atom(A)).
Similarly, we can define Cont (g, A). Remember following efficient symbol from [5]:
Notation. Let G and H be two groups with g and h as their generating
set, respectively. Coloring Cayley graph Γ1 := Γ(G, g) and Γ2 :=
Γ(H, h) with same colors, we get partitions E and F of G and H,
respectively. For two sets E ∈ E and F ∈ F , we write EsF to show
that we have two sets of the same color. In particular, If we have
Con(g, E) = Con(h, F) or Cont (g, E) = Cont (h, F), for configuration pairs
(g, E) and (h, F ) for groups G and H, respectively; This implies that,
their corresponded Cayley graphs Γ1 := Γ(G, g) and Γ2 := Γ(H, h),
colored with same colors.
CONFIGURATION EQUIVALENCE IS NOT EQUIVALENT TO ISOMORPHISM 5
We can also use s for sigma algebras; Let E := {E1 , . . . , Em } and
F := {F1 , . . . , Fm } be partitions of G and H resp. such that Ei sFi ,
i = 1, . . . , m. For A ∈ σ(E) and B ∈ σ(F ), say AsB, when
{k : Ek ∩ A 6= ∅} = {k : Fk ∩ B 6= ∅}
In other words, if AsB, and A = Ei1 ∪· · ·∪Eij , then B = Fi1 ∪· · ·∪Fij .
The following technical lemma will be used in the following:
Lemma 2.1. Let G and H be two groups with sigma algebras A and B,
resp. and Con(g, A) = Con(h, B). Suppose A1 , A2 ∈ A and B1 , B2 ∈ B,
are such that Ai sBi , i = 1, 2, and g ∈ g corresponded to h ∈ h. we
have
(1) If gA1 ⊆ A2 , then hB1 ⊆ B2 ,
(2) If gA1 = A2 , then hB1 = B2
Proof. Set
g = (g1 , . . . , gn ),
atom(A) = {E1 , . . . , Em }
h = (h1 , . . . , hn ),
atom(B) = {F1 , . . . , Fm }
Without loss of generality, let g and h match g1 and h1 . Also, set
Ik := {i : Ei ∩ Ak 6= ∅},
k = 1, 2
So, by assumptions,
Ak =
[
Ei ,
Bk =
i∈Ik
[
Fi
(k = 1, 2)
i∈Ik
Now, for C = (c0 , c1 , . . . , cn ) in Con(g, A), c1 ∈ I2 if c0 ∈ I1 , this proves
(1).
For proving (2), note that if C = (c0 , c1 , . . . , cn ) in Con(g, A), then
c0 ∈ I1 , if and only if c1 ∈ I2 .
If Cont (g, E) = Cont (h, F ), one can easily get an analog of the above
lemma for left multiplication replaced with right multiplication.
Let G and H be two groups. Consider partitions R and S of G and
H respectively, and their refinements, R′ and S ′ . Assume that
R = {R1 , . . . , Rl },
R′ = {R1′ , . . . , Rk′ }
S = {S1 , . . . , Sl },
S ′ = {S1′ , . . . , Sk′ }
We may say that these two pairs (R′ , R) and (S ′ , S) are similar and
may write (R′ , R) ∼ (S ′ , S), if
{i : Rj′ ∩ Ri 6= ∅} = {i : Sj′ ∩ Si 6= ∅}
CONFIGURATION EQUIVALENCE IS NOT EQUIVALENT TO ISOMORPHISM 6
for each j = 1, . . . , k. Also, for sigma algebras A and B, resp., and
their sigma subalgebras, A′ and B′ , resp. we say (A, A′) and (B, B′ )
are similar, written (A, A′ ) ∼ (B, B′ ), if
(atom(A), atom(A′ )) ∼ (atom(B), atom(B)′ )
Now, we can rewrite [4, Lemma 3.2] in sigma algebras:
Lemma 2.2. Let A and B be sigma algebras of groups G and H, resp.
Suppose g and h are generating sets of G and H such that Con(g, A) =
Con(h, B). If A′ and B′ are sigma sub-algebras of A and B, such that
(A, A′) ∼ (B, B′ ), then Con(g, A′ ) = Con(h, B′ ).
The sigma algebraic version of [5, Lemma 3.2] is similarly obtained.
By use of sigma algebras, equality of Tarski numbers of configuration
equivalent groups, is easily obtained. Recall the following definition of
paradoxical decomposition and Tarski number:
Definition 2.3. A group G admits a paradoxical decomposition if there
exist positive integers m and n, disjoint subsets P1 , . . . , Pm , Q1 , . .
. , Qn of G and elements x1 , . . . , xm , y1 , . . . , yn of G such that
m
n
[
[
G=
xi Pi =
yj Qi
i=1
j=1
The minimal possible value of m + n in a paradoxical decomposition of
G is the Tarski number of G and denoted by τ (G). If a group G doesn’t
have a paradoxical decomposition, it means that G is amenable; In this
case we will define τ (G) to be ∞.
Theorem 2.4. Let groups G and H be configuration equivalent. Then
τ (G) = τ (H).
Proof. If G is amenable, so is H, and we have done. Hence without
loss of generality, suppose that τ (G) is finite. Fix a generating set
of G, say g0 . Let m and n, (P1 , . . . , Pm ), (Q1 , . . . , Qn ) and elements
x := (x1 , . . . , xm ), y := (y1 , . . . , yn ) be as in Definition 2.3. Consider
the sigma algebra A, generated by sets
Pi , xr Pi , Qj , ys Qj ,
i, r = 1, . . . , m, j, s = 1, . . . , n
Set g := g0 ⊕ x ⊕ y. Since G ≈ H, there exists a generating set
h = h0 ⊕ u ⊕ v and sigma algebra B of G, such that
Con(g, A) = Con(h, B)
where h0 , u := (u1 , . . . , um ), v := (v1 , . . . , vn ) are corresponded to g0 ,
x and y, resp. Suppose that Ci , Dj in B are such that Ci sPi and
Dj sQj , i = 1, . . . , m, j = 1, . . . , m. Therefore, Lemma 2.1 leads
CONFIGURATION EQUIVALENCE IS NOT EQUIVALENT TO ISOMORPHISM 7
to ur Ci sxr Pi and vs Dj sys Qj . So, u, v, Ci ’s and Dj ’s satisfy the
conditions of Definition 2.3, and hence, τ (H) ≤ τ (G). By the symetry,
we have also τ (G) ≤ τ (H), which completes the proof.
We say that a group G admits a G = nG decomposition, n ∈ N, if
there exists disjoint sets
{Pi,j : i = 1, . . . , n, j = 1, . . . , k(i)}
along with elements x(i) = (xi,1 , . . . , xi,k(i) ), i = 1, . . . , n, of G, such
that
k(i)
[
xi,j Pi,j (i = 1, . . . , n)
G=
j=1
So, paradoxical decomposition of a group G, is a G = 2G decomposition. If a group G, P
admits a G = nG decomposition, we denote the
minimal amount of ni=1 k(i) by τn (G). Using the method of proof of
the above theorem, we can easily show that
Theorem 2.5. Let G and H be two configuration equivalent group.
If G admits a G = nG decomposition, n ∈ N, then H also admits a
H = nH decomposition, and τn (G) = τn (H).
Subgroups of (two-sided) configuration equivalent groups have not
been studied yet. However, By the use of Tarski number, some pieces
of information about non-Abelian free subgroups can be obtained. By
a theorem of Jónsson and Dekker (see, for example, [16, Theorem
5.8.38]), τ (G) = 4 if and only if G contains a non-Abelian free subgroup. This and [7, Theorem 1], implies that if groups G and H are
configuration equivalent, then they both have non-Abelian free subgroups or not. In fact, we have more, and before stating the result,
some reminders are needed; The power function for a polynomial type
group G is a function ς on G with following properties:
(1) ς(G) is a finite set, and
(2) ς −1 (ς(eG )) = {eG }.
We associate a (finite) partition with a power function, called ς–partition
of G consisting of disjoint subsets, E(ς(g)) := ς −1 (ς(g)), g ∈ G.
(3) There is a generating set g of G which, for each g ∈ G, we can
find a representative pair (Jg , σg ) in g such that if Con(g, E) =
Con(h, F ), then
W (Jg , σg ; h)F (ς(eG )) ⊆ F (ς(W (Jg , σg ; g))
where F (ς(g)) considered to be the same as E(ς(g)) in color.
CONFIGURATION EQUIVALENCE IS NOT EQUIVALENT TO ISOMORPHISM 8
By use of the power function, we can speak about some subgroups of
configuration equivalent groups;
Theorem 2.6. Let two groups G and H be configuration equivalent.
Suppose that G contains a non-Abelian free subgroup of rank n, n ∈ N.
Then H contains a non-Abelian free subgroup of rank n.
Proof. Fix a generating set of G, say g. Let g0 be a generating set of
a subgroup G0 of G, which is free of rank n. Assume that ς is the
power function, and E0 is a ς-partition of G0 . Assume A is the sigma
algebra generated by set E(ς(g)), g ∈ G0 . Now, if Con(g0 ⊕ g, A) =
Con(h0 ⊕ h, B), and if we denote the set F in F , where F sE(ς(g)), by
F (ς(g)), then one can easily check that, for an arbitrary representative
pair in g0 , say (J, ρ), we have
W (J, ρ; h0 )F (ς(eG )) ⊆ F (ς(W (J, ρ; g0 ))
Hence, W (J, ρ; h0 ) = eH if and only if W (J, ρ; g0 ) = eG , and this means
H0 := hh0 i, is a subgroup of H, which is free of rank n.
In the following conjugacy classes are involved; For an element g in a
group G, we denote the conjugacy class of g by Cl(g). It is obvious that
a normal subset (see [5, Definition 2.1.]) of G is nothing but a disjoint
union of conjugacy classes of G. We will show in the following theorem
that, if two groups G and H are two-sided configuration equivalent,
then their class numbers will be the same:
Theorem 2.7. Let finitely generated groups G and H be two-sided
configuration equivalent. Then their class numbers are the same.
Proof. Let g = (g1 , . . . , gn ) be a generating set of G. Suppose the
class number of G is at least N, N ∈ N. So, there are elements xi ,
i = 1, . . . , N in G, such that Cl(xi ), i = 1, . . . , N, are pairwise disjoint.
Consider a finite sigma algebra A of G, containing the following sets,
Cl(xi ), Cl(xi )gj−1 ,
i = 1, . . . , N, j = 1, . . . , n
Assume B is a sigma algebra of G along with a generating set h, such
that Cont (g, A) = Cont (h, B). If elements Bi in B, i = 1, . . . , N are
such that Bi sCl(xi ), then we get by Lemma 2.1 that Bi , i = 1, . . . , N,
are normal sets, so the class number of H is at least N. This completes
the proof.
It is obvious that Cl(x) = {x} if and only if x is a central element.
Suppose that the class number of G is finite, this implies that Z(G) is
a finite subgroup of G, where Z(G) stands for the center of G. In the
below lemma, by using of the finiteness of the class number, we present
CONFIGURATION EQUIVALENCE IS NOT EQUIVALENT TO ISOMORPHISM 9
a certain type of configuration pair, which make it possible to study
subgroups more efficiently:
Lemma 2.8. Let G be a finitely generated group with finite class number. Fix a generating set g of G. Then there is a partition E of G,
containing {eG }, such that if Cont (g, E) = Cont (h, F ) for a configuration pair (h, F ) of a group H, and if {eG }sF ∈ F , then F will be
singleton, and without loss of generality, we can assume that F = {eH }.
Proof. Suppose that g = (g1 , . . . , gn ), and Z(G) = {z0 = eG , z1 , . . . , zr },
Let elements x1 , . . . , xs ∈ G be such that
Cl(zi ), Cl(xj ),
i = 0, 1, . . . , r, j = 1, . . . , s
are all of the conjugacy classes. Now, consider the sigma algebra A
generated by the following subsets of G,
{zi }, {zi gk−1 }, {xj }, Cl(xj ) \ {xj }, Cl(xj )gk−1
for i = 0, 1, . . . , r, j = 1, . . . , s, and k = 1, . . . , n. If Cont (g, A) =
Cont (h, B) for a sigma algebra B and a generating set h of a group H,
and if Ki and Lj in B are such that
Ki s{zi }, Lj sCl(xj ) i = 0, 1, . . . , r, j = 1, . . . , s
These sets are all normal (see Lemma 2.1), so by Theorem 2.7, there
are elements wi , and yj , in H such that
Ki = Cl(wi ), Lj = Cl(yj ),
i = 0, 1, . . . , r, j = 1, . . . , s
Since sets Cl(yj ) can be written as a union of at least two atoms, so wi
are all central, and without loss of generality we can assume w0 = eH .
Therefore F = atom(B) works well.
By the power function of polynomial type groups, we can say that
in two-sided configuration equivalent groups with finite class number,
the polynomial type subgroups are isomorphic:
Theorem 2.9. Let two groups G and H be two-sided configuration
equivalent. Assume that these groups have finite class number. Then
for each polynomial type subgroup, G0 of G, H contains a subgroup H0 ,
which is isomorphic to G0 . Furthermore G and H have isomorphic
centers.
Proof. Fix a generating set g for G. Let ς, E0 and g0 be such that the
properties (1) to (3) on page in page 7, are established for G0 . Applying
Lemma 2.8, we get a partition E of G. Now. if
Cont (g0 ⊕ g, σ(E1 ∪ E)) = Cont (h0 ⊕ h, B)
CONFIGURATION EQUIVALENCE IS NOT EQUIVALENT TO ISOMORPHISM10
by Lemma 2.8, property (3), one can easily see that G0 ∼
= H0 , where
H0 := hh0 i.
Since Z(G) becomes a finite subgroup, for G has a finite class number, by adding it to g, repeating the proof of 2.8, one can easily obtain
that Z(G) ∼
= Z(H).
It is of interest to study finitely generated groups with finite class
numbers. For each large prime number p, there exists a 2-generated infinite group of exponent p which has exactly p conjugacy classes (that’s
[12, Theorem 41.2]). Osin [3] has recently constructed a finitely generated example (that was a major breakthrough since the problem was
open for more than 60 years). All of these groups are infinitely presented and it is still an open problem, i.e., the question whether there
is an infinite finitely presented group with a finite class number. If such
a group exists, the following theorem will make sense:
Theorem 2.10. Let finitely presented group G has finite class number.
Suppose that G ≈ H, which H is a finitely generated group. Then H
is a quotient of G.
Proof. Fix a generating set of G, say g. Apply Lemma 2.8 to get a
partition E of G. Consider a set of defining relators
{W (Ji , ρi ; g) : i = 1, . . . , m}
Let A be a sigma algebra of G, generated by sets in E and
{W (J, ρ; g)}
n(J) ≤ max{n(Ji ) : i = 1, . . . , m}
But G ≈ H, so there is a sigma algebra B in H, such that Cont (g, A) =
Cont (h, B). Lemma 2.2, Lemma 2.8, imply that we can assume {eH } ∈
B, {eG }s{eH }, and
{W (J, ρ; g)}s{W (J, ρ; h)}, for (J, ρ) with n(J) ≤ max{n(Ji ) : i = 1, . . . , m}
so, in particular, W (Ji , ρi ; h) = eH , i = 1, . . . , m, and hence the map
G → H,
W (I, δ; g) 7→ W (I, δ; h)
where (I, δ) ranges over arbitrary representative pairs, introduced an
epimorphism.
3. Two-sided Configuration and Quotients
In [5], some results about quotients of two-sided configuration equivalent groups were obtained. In this section, we study the number of
quotients.
CONFIGURATION EQUIVALENCE IS NOT EQUIVALENT TO ISOMORPHISM11
Let N be a normal subgroup of G; We denote the quotient map,
G → G/N by qN , and in the cases where there is no ambiguity, we
may drop N. Recall that if E is a partition of G/N, then
q −1 (E) := {q −1 (E) : E ∈ E}
becomes a partition of G. We refer to such a partition, when we say
an N–partition of G. One can easily check that the intersection of
two N–partitions is itself an N–partition. An N–refinement of an
N–partition Ê, is a refinement, Eˆ′ , which is itself an N–partition.
Let g = (g1 , . . . , gk ) be an ordered subset of G, such that q(g) :=
(q(g1), . . . , q(gk )) is a generating set of G/N, then by [1, Lemma 6.2.],
there is an ordered subset n of N, such that ĝ = g ⊕ n becomes a
generating set of G, called N-extension of g. By an N–configuration
pair we mean a configuration pair (ĝ, Ê) such that g is an N–extension
generating set, and Ê is an N–partition of G.
The notions of “preserving presentation” and “recognizability” of a
configuration pair are defined in [5]. Lets first consider the concept of
recognizable configuration pair:
Definition 3.1. Let G be a group with a normal subgroup N and
an N–configuration pair (ĝ, Ê). We may say that (ĝ, Ê) is recognizable
w.r.t. N, if whenever Cont (ĝ, E) = Cont (ĥ, F̂), for a configuration pair
(ĥ, F̂ ) of a groups H, then for every g ∈ G \ N, there is a representative
pair, (Jg , ρg ) such that
g = W (Jg , ρg ; ĝ) and W (Jg , ρg ; ĥ)F ∩ F = ∅
where F ∈ F is in the same color as an element of E which contains
N.
As a consequence of Lemma 2.2, one can easily show that (see [4,
Lemma 3.3.]):
Lemma 3.2. Let G be a group with a normal subgroup N. Assume
(ĝ, Ê) is a recognizable configuration pair w.r.t. N. Then for each N–
refinement Eˆ′ of Ê, the configuration pair (ĝ, Eˆ′ ) is recognizable w.r.t.
N.
What really makes working with recognizable configuration pair useful, is accompaniment of this notion to the following one:
Definition 3.3. Let G, N, g, ĝ be regarded as above. We say that a
configuration pair (ĝ, E) preserves presentation w.r.t. N, if the following are held:
(I) N ∈ E,
CONFIGURATION EQUIVALENCE IS NOT EQUIVALENT TO ISOMORPHISM12
(II) if Cont (ĝ, Ê) = Cont (ĥ, F̂), for a configuration pair (ĥ, F̂) of a
groups H, and NsF ∈ F̂ , then F is normal and the function bellow,
defined on cosets of N,
W (J, ρ; ĝ)N 7→ W (J, ρ; ĥ)F
is considered to be well-defined, where (J, ρ) is an arbitrary representative pair.
Remark 1. By [5, Theorem 2.5], we know that if G/N is finitely
presented, then for each N–configuration pair (ĝ, Ê), there exists an
N–refinement Eˆ′ of Ê, such that (ĝ, Eˆ′ ) preserves presentation w.r.t.
N.
Let R and S be two partitions of a group G. Define the partition
R ∩ S as follows:
atom(σ(R ∪ S))
regarding this notation and Lemma 2.2 as well, we can see that it does
not matter, if we work with one or more than one partition of G, more
precisely:
Lemma 3.4. Let G be a group with a generating set g. Assume that
{R1 , . . . , Rr }, be a collection of partitions of G and G ≈t H for a group
H. Then, there are, a generating set h, and a collection {S1 , . . . , Sr }
of partitions of H, such that
Cont (g, Ri ) = Cont (h, Si ) i = 1, . . . , r
T
Proof. Set R = ri=1 Ri . Let (h, S) be a configuration pair of H
such that Cont (g, R) = Cont (h, S). If Si is a partition of H, such
that (R, Ri ) ∼ (S, Si ), i = 1, . . . , r. Then, Lemma 2.2, implies that
Cont (g, Ri ) = Cont (h, Si ), i = 1, . . . , r.
With the concept of two-sided configuration, we can study the number of finite index subgroups; To this end, we provide the following
definition:
Definition 3.5. Let G be a group, and C be a finite collection of
normal subgroups of G. We may say G is recognizable w.r.t C, if the
following are held:
(1) C is closed under intersection,
(2) G/N is finitely presented, for all N ∈ C, and
(3) There exists a generating set g of G, and a collection {E(N) :
N ∈ C} of partition of G such that (g, E(N)) is a recognizable
configuration pair w.r.t. N.
CONFIGURATION EQUIVALENCE IS NOT EQUIVALENT TO ISOMORPHISM13
In the case which there is a collection C, with properties in the above
definition, the next theorem is worthy of attention:
Theorem 3.6. Suppose a group G is recognizable w.r.t. a collection
C of its normal subgroups. Let a group H be two-sided configuration
equivalent with G. Then there is a collection {NN : N ∈ C} of normal
subgroups of H such that
(1) G ∼
= H ,
N
NN
(2) NN ∩M = NN ∩ NM , for N and M in C.
Proof. Let g and {E(N) : N ∈ C} be regarded as in Definition 3.5. By
Lemma 3.4, there are a generating set h and partitions {F (N) : N ∈
C}, such that
Cont (g, E(N)) = Cont (h, F (N)) (N ∈ C)
Hence, by [5, Lemma 3.4.], we obtain normal subgroups {NN : N ∈ C},
along with following isomorphisms
H
G
→
,
N
NN
W (J, ρ; g)N 7→ W (J, ρ; h)NN
With the above isomorphisms, the equation in (2) are easily proved.
We know by [10, Theorem 21.4] that, the number of subgroups of
finite index n in a finitely generated group G is finite. Since the intersection of finite index subgroups is again a finite index subgroup, by
the above theorem, we obtain:
Corollary 3.7. Let G and H be finitely generated groups with the same
two-sided configuration sets, and let n ∈ N. Then G and H contain
exactly the same number of normal subgroups of index n. Moreover,
we have
Y
Y
{H/N : |H : N| < ∞}
{G/N : |G : N| < ∞} ∼
=
Proof. Suppose {N1 , . . . , Nr } is the collection of all normal subgroups
of index n. Let C be a collection of normal subgroups of G obtained
from intersection of Nk ’s. For N ∈ C, assume that the power function of
G/N yields a configuration pair (qN (gN ), qN (E(N))), for an ordered set
Tr
r
gN and partition E(N) of G. Let g = ⊕\
k=1 gNk be a
k=1 Nk –extension
generating set of G. Hence, one can see, by g and C as above, the
conditions of Definition 3.5 are satisfied, and therefore, Theorem 3.6,
shows the existence of at least r normal subgroups of index n in H.
The symmetry in the concept of configuration completes the proof.
CONFIGURATION EQUIVALENCE IS NOT EQUIVALENT TO ISOMORPHISM14
We denote by P(G), the collection of normal subgroups of G, which
their quotient in G are polycyclic. Since every polycyclic group has
a recognizable configuration pair, and since P(G) is closed under intersection, an argument like the one in proof of the previous corollary
leads to:
Corollary 3.8. Let G and H be finitely generated groups with G ≈t H.
Then P(G) and P(H) have the same cardinality. Furthermore if P(G)
is finite, then there is a bijection Ψ : P(G) → P(H) such that
(1) For every N and M in P(G), Ψ(N ∩ M) = Ψ(N) ∩ Ψ(M),
(2) For each N ∈ P(G), we have G/N ∼
= H/Ψ(N).
4. The Concept of Configuration Equivalence is not
Equivalent to Isomorphism
The question whether the concepts of the configuration equivalence
and isomorphism of groups are the same has been open since the beginning of configuration theory. Indeed, the answer is “No”, there
are non-isomorphic groups with the same two-sided configuration sets.
It is worth noting that, the groups in our example are solvable, thus
this natural conjecture that these two concepts may be equivalent for
amenable groups will be rejected. To provide our example, we may
need to provide the following two technical lemmas:
Lemma 4.1. Let φ : G → H be an epimorphism of groups. Suppose
that g is a generating set of G, and F is a partition of H, then
Cont (g, φ−1 (F )) = Cont (φ(g), F )
Proof. Assume that C ∈ Cont (g, φ−1(F )). So, there exists (x0 , x1 , . . . , x2n )
having configuration C, that means, xk ∈ φ−1 (Fck ), k = 0, 1, . . . , 2n
and,
xk = gk x0 , xk+n = x0 gk (k = 1, . . . , n)
Therefore φ(xk ) ∈ Fck , and
φ(xk ) = φ(gk )φ(x0 ),
φ(xk+n ) = φ(x0 )φ(gk )
Hence, (φ(x0 ), φ(x1 ), . . . , φ(x2n )) has configuration C in Cont (φ(g), F ).
Conversely, let a tuple (y0 , y1 , . . . , y2n ) have configuration C ∈ Cont (φ(g), F).
That means, yk ∈ Fck , k = 0, 1, . . . , 2n and
yk = φ(gk )y0 ,
yk+n = y0 φ(gk ) (k = 1, . . . , n)
Choose x0 ∈ φ−1 (y0). Then
(x0 , g1 x0 , . . . , gn x0 , x0 g1 , . . . , x0 gn )
in G has configuration C, belonging to Cont (g, φ−1 (F )).
CONFIGURATION EQUIVALENCE IS NOT EQUIVALENT TO ISOMORPHISM15
We may call an homomorphism φ : G → H, generating-surjection,
if every ordered generating set of H is an image of an ordered generating set of G. It is clear that each generating-surjection, becomes
an epimorphism. We call two groups G and H generating-bijective, if
there are generating-surjections φ : G → H, and ψ : H → G. From the
following lemma, it will be understood that generating-bijective groups
have same two-sided configuration sets:
Lemma 4.2. Let G and H be two finitely generated groups. Suppose
there is a generating-surjection φ : G → H. Then Cont (H) ⊆ Cont (G).
Proof. Consider a configuration pair (h, F ) of H. There exists a generating set g of G such that h = φ(g). Thus, by previous lemma,
Cont (h, F ) = Cont (φ(g), F ) = Cont (g, φ−1 (F ))
This completes the proof.
It is obvious that isomorphic groups are generating-bijective, but the
converse is not always true, as the following theorem will show this:
Theorem 4.3. There exist non-isomorphic finitely generating groups
with the same two-sided configuration sets.
Proof. Put R := Z[t, t−1 ], the ring of Laurent polynomials. Let K be
a group of matrices
1 B D
0 A C
0 0 1
where B, C and D belongs to R, and A ∈ hti. The group B is easily
checked to be finitely generated; Indeed, if we denote the above matrix
by (A, B, C, D), then (k1 , k2 , k3 ), in which
k1 = (t, 0, 0, 0) k2 = (1, 1, 0, 0) k3 = (1, 0, 1, 0)
becomes a generating set of K, because of the following equations:
k
km
1 = (t , 0, 0, 0),
−m
km
= (1, 0, tm , 0),
1 k3 k1
k
m
k−m
1 k2 k1 = (1, t , 0, 0)
m
m
[k3 , k−m
1 k2 k1 ] = (1, 0, 0, t )
where m is an integer. The center of this group, Z(K), consists of
unipotent matrices with a single possibly non-trivial element in the
upper right corner
Z(K) = {(1, 0, 0, D) : D ∈ R}
It is clearly isomorphic to R. We can rewrite the product of the group
as bellow:
(A, B, C, D)(X, Y, Z, W ) = (AX, BX + Y, C + AW, D + BW + Z)
CONFIGURATION EQUIVALENCE IS NOT EQUIVALENT TO ISOMORPHISM16
By the above equality, one can see that the map
Φ : K → K,
(A, B, C, D) 7→ (A, B, tC, tD)
introduced an automorphism of K.
Set Nm := tm Z[t], m ∈ Z. The automorphism Φ, implies that
K/Nm ∼
= K/Φ(Nm ) = K/Nm+1 , m ∈ Z. Now, let G := K/N0 and
H := K/(2Z ⊕ N1 ). These groups are not isomorphic, for G is torsion
free and H is not.
Note that if k is an ordered set of K, such that its image under the
natural quotient map forms a generating set of K/Nm, then, also, the
image of Φ(k) will form a generating set of K/Nm . Thus, G and H are
generating-bijective. This and Lemma 4.2, complete the proof.
Remark 2. The groups G and H in the above theorem are solvable.
this shows that the two concepts, configuration equivalence and isomorphism, are not equivalent for solvable and hence amenable groups.
these groups are not finitely presented and they are not residually finite,
for they are not Hopfian.
References
[1] A. Abdollahi, A. Rejali and G. A. Willis, Group Properties Characterized by
Configurations. Illinois J. Math. 2004. 48(3): 861-873.
[2] A. Abdollahi, A. Rejali and A. Yousofzadeh, Configuration of Nilpotent Groups
and Isomorphism. J. Algebra Appl. 2009. 8(3): 339-350.
[3] Denis Osin, Small cancellations over relatively hyperbolic groups and embedding
theorems. Ann. of Math. (2) 172 (2010), no. 1, 139.
[4] A. Rejali and M. Soleimani Malekan, Strong Configuration Equivalence and
Isomorphism, http://arxiv.org/abs/1510.07209.
[5] A. Rejali and M. Soleimani Malekan, Two-sided Configuration of Polycyclic
Groups and Isomorphism,http://arxiv.org/abs/1512.03021.
[6] A. Rejali and A. Yousofzadeh, Group Properties Characterized by Two-sided
Configurations, Algebra Colloq. 17:4 (2010) 583-594.
[7] A. Tavakoli, A. Rejali, A. Yousofzadeh, A. Abdollahi, A Note About Configuration of A Group, Matematika, 2014, Volume 30, Number 2, 117-121.
[8] J. M. Rosenblatt and G. A. Willis, Weak Convergence Is Not Strong Convergence for Amenable Groups, Canad. Math. Bull. 44 (2001), 231-241. MR
2002d:43001.
[9] E. A. Walker, Cancellation in Direct Sums of Groups, Proc. Amer. Math. Soc.,
Vol. 7, No. 5 (Oct., 1956), pp. 898-902.
[10] O. Bogopolski, Introduction to Group Theory, European Mathematical Society,
2008.
[11] W. Magnus, A. Karrass, D. Solitar, Combinatorial Group Theory: Presentations of Groups in Terms of Generators and Relations, Dover Publications,
INC, New York, 2004.
[12] A. Yu. Oĺshanskii, Geometry of Defining Relations in Groups, Kluwer, 1991.
CONFIGURATION EQUIVALENCE IS NOT EQUIVALENT TO ISOMORPHISM17
[13] D. J. S. Robinson, A Course in the Theory of Groups, 2nd edition, Springer,
New York, 1983.
[14] D. Segal, Polycyclic groups, Cambridge University Press, London, 1983.
[15] C. C. Sims, Computation with Finitely Presented Groups. Cambridge: Cambridge University Press, 1994.
[16] M. Sapir, Combinatorial algebra: syntax and semantics,
http://www.math.vanderbilt.edu/msapir/book/b2.pdf.
[17] M. J. Tomkinsin, FC-groups, 2nd edition, Pitman publishing limited, London,
1984.
University of Isfahan
Current address: Department of Mathematics, Faculty of Sciences , University
of Isfahan, Isfahan, 81746-73441, Iran
E-mail address: [email protected]
Department of Mathematics, Ph. D student, University of Isfahan,
Isfahan, 81746-73441, Iran
E-mail address: [email protected]
| 4math.GR
|
arXiv:1705.10957v2 [math.AC] 14 Dec 2017
NEARLY COMMUTING MATRICES
ZHIBEK KADYRSIZOVA
Abstract. We prove that the algebraic set of pairs of matrices with a diagonal commutator over a field of positive prime characteristic, its irreducible
components, and their intersection are F -pure when the size of matrices is
equal to 3. Furthermore, we show that this algebraic set is reduced and the
intersection of its irreducible components is irreducible in any characteristic
for pairs of matrices of any size. In addition, we discuss various conjectures
on the singularities of these algebraic sets and the system of parameters on
the corresponding coordinate rings.
Keywords: Frobenius, singularities, F -purity, commuting matrices
1. Introduction and preliminaries
In this paper we study algebraic sets of pairs of matrices such that their
commutator is either nonzero diagonal or zero. We also consider some other
related algebraic sets. First let us define relevant notions.
Let X = (xij )1≤i,j≤n and Y = (yij )1≤i,j≤n be n×n matrices of indeterminates
over a field K. Let R = K[X, Y ] be the polynomial ring in {xij , yij }1≤i,j≤n and
let I denote the ideal generated by the off-diagonal entries of the commutator
matrix XY − Y X and J denote the ideal generated by the entries of XY −
Y X. The ideal I defines the algebraic set of pairs of matrices with a diagonal
commutator and is called the algebraic set of nearly commuting matrices. The
ideal J defines the algebraic set of pairs of commuting matrices.
1
2
ZHIBEK KADYRSIZOVA
Let uij denote the (i, j)th entry of the matrix XY −Y X. Then I = (uij | 1 ≤
i 6= j ≤ n) and J = (uij | 1 ≤ i, j ≤ n).
Theorem 1 ([Ger61]). The algebraic set of commuting matrices is irreducible,
i.e., it is a variety. Equivalently, Rad(J) is prime.
The following results are due to A. Knutson [Knu05], when the characteristic
of the field is 0, and to H.Young [You11] in all characteristics.
Theorem 2 ([Knu05], [You11]). The algebraic set of nearly commuting matrices is a complete intersection, with the variety of commuting matrices as
one of its irreducible components. In particular, the set {uij |1 ≤ i 6= j ≤ n} is
a regular sequence and the dimension of R/I is n2 + n.
Theorem 3 ([Knu05], [You11]). When K has characteristic zero, I is a radical
ideal.
A. Knutson in his paper [Knu05] conjectured that V(I) has only two irreducible components and it was proved in all characteristics by H.Young in his
thesis, [You11].
Theorem 4 ([You11]). If n ≥ 2, the algebraic set of nearly commuting matrices has two irreducible components, one of which is the variety of commuting
matrices and the other is the so-called skew component. That is, I has two
minimal primes, one of which is Rad(J).
Let P = Rad(J) and let Q denote the other minimal prime of I, i.e.,
T
Rad(I) = P Q.
The following conjecture was made in 1982 by M. Artin and M. Hochster.
NEARLY COMMUTING MATRICES
3
Conjecture 1. J is reduced, i.e., J = P .
It was answered positively by Mary Thompson in her thesis in the case of
3 × 3 matrices.
Theorem 5 ([Tho85]). R/J is a Cohen-Macaulay domain when n = 3.
Now let us go back to algebraic sets of nearly commuting matrices and their
irreducible components. First, we take a look at what we have when n = 1, 2.
When n = 1, everything is trivial. More precisely, I = P = Q = K[x11 , y11 ].
When n = 2, without loss of generality we may replace X and Y by X −x22 In
and Y − y22 In respectively. Here In is the identity matrix of size n. Denote
′
x′11 = x11 − x22 , y11
= y11 − y22 . Then the generators of I are 2 by 2 minors
u12 =
x′11 x12
,
′
y11
y12
u21 = −
x′11 x21
.
′
y11
y21
The diagonal entries of XY − Y X are
u11 =
x12 x21
y12 y21
,
u22 = −
x12 x21
.
y12 y21
Then J is the ideal generated by size 2 minors of
x′11
x12 x21
′
y11
y12
′
′
(x11 , y11
).
′ p−1 ′
′ p−1
(u12 u21 )p−1 = (x′11 y12 − x12 y11
) (x11 y21 − x21 y11
)
=
and
y21
Moreover,
therefore, J = P is prime. We also have that Q =
T
′
I = P Q is radical and P + Q = (x12 y21 − x21 y12 , x′11 , y11
) is prime.
We have that
4
ZHIBEK KADYRSIZOVA
p−1 p−1
X
X
α=0 β=0
(−1)
α+β
p−1 p−1
′ 2(p−1)−α−β p−1−α α p−1−β β
(x′11 )α+β (y11
)
x12
y12 x21
y21 .
β
α
′
Therefore, (u12 u21 )p−1 has a monomial term (x′11 y11
x12 y21 )p−1 with coefficient
(−1)p−1 . Since I [p] : I = (u12 u21 )p−1 +I [p] , R/I is F -pure, see Fedder’s criterion
Lemma 2. Furthermore, determinantal rings R/P , R/Q, R/(P + Q) are F regular, see [HH94].
Therefore, for the rest of the paper we shall use the following notations.
Notation 1. Let n ≥ 3 be an integer. Let X = (xij )1≤i,j≤n and Y = (yij )1≤i,j≤n
be n × n matrices of indeterminates over a field K. Let R = K[X, Y ] be the
polynomial ring in {xij , yij }1≤i,j≤n and let I denote the ideal generated by the
off-diagonal entries of the commutator matrix XY − Y X and J denote the
ideal generated by the entries of XY − Y X. Let P denote the radical of J
and Q be the other minimal prime of Rad(I).
We prove the following results in this paper.
Theorem 6. Let R be a ring as in Notation 1. Assume also that the field K
has positive prime characteristic. Then R/I, R/P , R/Q and R/(P +Q) are F pure rings when n = 3. In other words, the algebraic set of nearly commuting
matrices of size 3, its irreducible components and their intersection are F -pure.
In particular, the skew component is reduced in this case.
Theorem 7. Let R be a ring as in Notation 1. Then R/I is reduced. In other
words, the algebraic set of nearly commuting matrices is reduced for matrices
of all sizes and in all characteristics.
Theorem 8. The intersection of the variety of commuting matrices and the
skew component is irreducible, that is, Rad(P + Q) is prime.
NEARLY COMMUTING MATRICES
5
2. F -purity
In this section we show that the coordinate ring of the algebraic set of pairs of
matrices with a diagonal commutator is F -pure in the case of 3 by 3 matrices.
Moreover, we also show that it implies the corresponding fact for its irreducible
components, the variety of commuting matrices and the skew-component, and
their intersection.
First we state two lemmas due to R. Fedder and they include a criterion
for F -purity for finitely generated K-algebras and which has a particularly
convenient form for complete intersections.
Lemma 1 (Fedder [Fed87]). Let S be a regular local ring or a polynomial
ring over a field. If S has characteristic p > 0 and I is an unmixed proper
ideal (homogeneous in the polynomial case) with the primary decomposition
T
T
I = ni=1 Ai , then I [p] : I = ni=1 (A[p] : A).
Lemma 2 (Fedder’s criterion [Fed87]). Let (S, m) be a regular local ring or a
polynomial ring over a field with its (homogeneous) maximal ideal. If S has
characteristic p > 0 and I is a proper ideal (homogeneous in the polynomial
case), then S/I is F -pure if and only if I [p] : I 6⊂ m[p] .
The next result is a straightforward consequence of the above two lemmas.
It will prove to be quite useful for us.
Lemma 3. Let S be a regular local ring or a polynomial ring over a field.
Suppose that S has characteristic p > 0 and I is an ideal of S (homogeneous
T
in the polynomial case). Suppose also that S/I is F -pure and I = ni=1 Ai is
the primary decomposition. Then S/(Ai1 + . . . Aim ) is F -pure for all 1 ≤ i1 <
... < im ≤ n and for all 1 ≤ m ≤ n.
6
ZHIBEK KADYRSIZOVA
T
[p]
Proof. Observe first that (Ai1 + . . . Aim )[p] :S (Ai1 + . . . Aim ) ⊇ m
j=1 (Aij :
T
[p]
Aij ) ⊇ ni=1 (Ai : Ai ) = (I [p] : I). The rest is immediate from Lemma 1
and Lemma 2.
The above lemma is closely related to results on compatibly split ideals, cf.
[ST12].
Immediately we get the corresponding result for our algebraic set.
Corollary 1. Suppose that the coordinate ring of the algebraic set of nearly
commuting matrices R/I is F -pure. Then R/P , R/Q and R/(P + Q) are
F -pure.
Next, we use Fedder’s criterion to show F -purity of R/I in case when n = 3.
Theorem 9. Let K be a field of characteristic p > 0 and let n = 3. Let R be
a ring as in Notation 1. Then R/I is F -pure.
Proof. Recall that I is generated by a regular sequence {uij |1 ≤ i 6= j ≤ n}.
Q
p−1
[p]
Therefore, I : I =
R + I [p] . Thus by Fedder’s criterion it
1≤i6=j≤n uij
Q
is sufficient to prove that 1≤i6=j≤n up−1
∈
/ m[p] , where m is the homogeneous
ij
maximal ideal in R. We show this by proving the following claim.
Claim. If µ = x12 x13 x21 x23 x31 x33 y11 y12 y23 y31 y32 y33 , then µp−1 is a monomial
Q
term of 1≤i6=j≤3 up−1
with a nonzero coefficient modulo p.
ij
Proof. We compute the coefficient of µp−1. It can be obtained by choosing a
monomial from every uij in the following way:
u12 :
(−x12 y11 )α1 (x13 y32 )β1
u13 :
(−x23 y12 )α2 (x12 y23 )β2 (−x13 y11 )γ2 (x13 y33 )δ2
NEARLY COMMUTING MATRICES
u21 :
(−x31 y23 )α3 (x21 y11 )β3 (x23 y31 )γ3
u23 :
(x23 y33 )α4 (−x33 y23 )β4
u31 :
(−x21 y32 )α5 (x33 y31 )β5 (−x31 y33 )γ5 (x31 y11 )δ5
u32 :
(x31 y12 )α6 (−x12 y31 )β6 (x33 y32 )γ6
7
Then the exponents Ast and Bst of each xst and yst respectively are
A12 = α1 + β2 + β6
A23 = α2 + γ3 + α4
A13 = β1 + γ2 + δ2
A31 = α3 + γ5 + δ5 + α6
A21 = β3 + α5
A33 = β4 + β5 + γ6
B11 = α1 + γ2 + β3 + δ5
B31 = γ3 + β5 + β6
B12 = α2 + α6
B32 = β1 + α5 + γ6
B23 = β2 + α3 + β4
B33 = δ2 + α4 + γ5
In addition, denote
C12 = α1 + β1 ,
C23 = α4 + β4 ,
C13 = α2 + β2 + γ2 + δ2 ,
C31 = α5 + β5 + γ5 + δ5 ,
C21 = α3 + β3 + γ3 ,
C32 = α6 + β6 + γ6 .
Our goal is to find all nonnegative integer tuples α = (α1 , . . . , α6 ), β =
(β1 , . . . , β6 ), γ = (γ2 , γ3 , γ5 , γ6 ), δ = (δ5 , δ6 ) such that Ast = p − 1, Bst = p − 1
for all 1 ≤ s, t ≤ 3 and Cij = p − 1 for all 1 ≤ i 6= j ≤ 3.
Notice that the linear system does not have a nonzero determinant: the sum
of the first 12 equations is twice the sum of the rest 6 equations. Therefore,
there is not a unique solution.
The above linear system can be solved using standard methods from linear
algebra and has the following solution:
8
ZHIBEK KADYRSIZOVA
α
β
=
γ
δ
a
b
d
p−1−a p−1−a−b p−1−a
−
a−c
a−d
−
c
−
p−1−b−a+d
a+b−d
−
−
p−1−b
p−1−a−b+d
b
b+a−d−c
0
c−a
−
a
where the column vector [α, β, γ, δ] represents the matrix of solutions and
a, b, c, d are arbitrary non-negative integers.
Since we look for non-negative integer solutions we must have that a = c
and a, b ≥ d and a + b ≤ p − 1. Hence we have that
α
β
=
γ
δ
a
b
d
p−1−b−a+d
a
p−1−b
p−1−a p−1−a−b p−1−a
a+b−d
p−1−a−b+d
b
−
0
a−d
−
b−d
0
−
a
−
−
0
−
Therefore, the coefficient of µp−1 is the sum of expressions of the form
(−1)α1 +α2 +γ2 +α3 +β4 +α5 +γ5 +β6 ((p − 1)!)6 /(α1 ! . . . α6 !β1 ! . . . β6 !γ2 !γ3 !γ5 !γ6 !δ5 !δ6 !)
NEARLY COMMUTING MATRICES
9
where α = (α1 , . . . , α6 ), β = (β1 , . . . , β6 ), γ = (γ2 , γ3 , γ5, γ6 ), δ = (δ5 , δ6 ) run
over all solutions of the linear system above. That is,
(p−1)/2
X
d=0
X
(−1)
a−d
a,b≥d, a+b≤p−1
2
2
2
b
p−1−b a+b−d
p−1
p−1
p−1
d
b
a
a+b−d
b
a
which modulo p is equivalent to
(p−1)/2
X
d=0
X
(−1)
a−d
a,b≥d, a+b≤p−1
p−1−b
a
b
a+b−d
d
b
It also can be written as
(p−1)/2
X
d=0
or
X
(−1)
a−d
a,b≥d, a+b≤p−1
a+b−d
p−1−b
a−d b−d d
a
p−1 b p−1−b
X
X X
b
a+b−d
a−d p − 1 − b
(−1)
d
b
a
b=0 d=0 a=d
The following lemma shows that the above expression is equal to 1 for all
values of p. In fact, for this purpose p does not have to be prime.
Lemma 4. Let Cm =
1 for all m ≥ 1.
Pm Pb
b=0
d=0
Pm−b
a=d
(−1)a−d
m−b
a
b
a+b−d
b
d
. Then Cm =
Proof. We prove a stronger statement.
Claim. Let Bm,b =
Pb
d=0
Pm−b
Bm,b
a=d
(−1)a−d
m−b
a
a+b−d
b
b
. Then for all m ≥ 1
d
0 if 0 ≤ b ≤ m − 1;
=
1
if b = m.
P P0
a−d m−b a+b−d b
(−1)
= 1 and
Proof. First observe that Bm,m = m
a=d
d=0
a
b
d
Pm
Bm,0 = a=0 (−1)a ma = 0. Hence we may assume that 0 < b < m.
10
ZHIBEK KADYRSIZOVA
Let Am,b,d =
Pm−b
a=d
m−b
a
(−1)a
Consider the difference
a+b−d
b
, then Bm,b =
Pb
d=0 (−1)
Am,b,d .
d b
d
Am,b,d − Am,b,d+1 =
m−b
m−b
X
X
a+b−d−1
a+b−d
a m−b
a m−b
=
−
(−1)
(−1)
b
a
b
a
a=d+1
a=d
m−b
X
a+b−d−1
a+b−d
m−b
a m−b
)=
−
(
+
(−1)
(−1)
b
b
a
d
a=d+1
d
Using Pascal’s identity, we get
m−b
X
a+b−d−1
m−b
a m−b
=
(−1)
+
(−1)
b−1
a
d
d
a=d+1
m−b a+b−d−1
=
(−1)
b
−
1
a
a=d
m−b
X
a
m−1−(b−1)
X
(−1)
a=d
a
m − 1 − (b − 1) a + (b − 1) − d
.
b−1
a
Thus we have that
Am,b,d − Am,b,d+1 = Am−1,b−1,d for all m − 1 ≥ b ≥ d + 1 and d ≥ 0.
Therefore,
Bm−1,b−1
b−1
X
b−1
d b−1
(Am,b,d −Am,b,d+1 ) =
(−1)
Am−1,b−1,d =
(−1)
=
d
d
d=0
d=0
b−1
X
b−1
X
d
(−1)
d=0
d
b−1
X
b−1
d b−1
Am,b,d+1 =
(−1)
Am,b,d −
d
d
d=0
b
X
b b−d
d−1 b − 1
(−1)
Am,b,d =
(−1)
Am,b,d −
d
d
−
1
b
d=0
d=1
b−1
X
d
NEARLY COMMUTING MATRICES
11
b
X
b b−d
d b d
(−1)
Am,b,d +
(−1)
Am,b,d =
d
d b
b
d=0
d=1
b−1
X
d
b
Am,b,d + Am,b,0 + (−1)b Am,b,b =
(−1)
d
d=1
b−1
X
d
b
Am,b,d = Bm,b .
(−1)
d
d=0
b
X
d
Thus we have that Bm−1,b−1 = Bm,b for all m ≥ 1 and m − 1 ≥ b ≥ 1.
In case m = 1, we only have B1,0 = 0. Finally, use induction on m to
conclude that Bm,b = 0 for all m ≥ 1 and m − 1 ≥ b.
P
Thus, Cm = m
b=0 Bm,b = 1.
Finally, we complete the proof of Theorem 9. We have that
m[p] and R/I is F -pure when n = 3.
Q
1≤i6=j≤n
up−1
∈
/
ij
Corollary 2. Let R be a ring as in Notation 1. When n = 3, R/P, R/Q and
R/(P + Q) are F -pure Cohen-Macaulay rings and R/(P + Q) is Gorenstein.
Proof. By [Tho85], R/P is a Cohen-Macaulay ring when n = 3. Since the
ideals P and Q are linked via I, that is I : P = Q and I : Q = P , we have
that R/Q is also Cohen-Macaulay, see [PS74]. Moreover, the theory of linkage
also implies that (P + Q)/P and (P + Q)/Q are isomorphic to the canonical
modules of R/P and R/Q, respectively. Hence R/(P + Q) is Gorenstein of
dimension n2 + n − 1.
Corollary 3. Let R be a ring as in Notation 1. Then P + Q is radical when
n = 3.
12
ZHIBEK KADYRSIZOVA
Remark. We prove in the next section that for all n, the radical of P + Q is
prime, which implies that P + Q is prime when n = 3. In particular, we have
that R/(P + Q) is a domain when n = 3.
3. Irreducibility of P + Q
In this section we prove that the intersection of the variety of commuting
matrices and the skew-component is irreducible. But first we define some
notions.
Definition. Let X be an n by n matrix of indeterminates. Then D(X) is an
n by n matrix whose ith column is defined by the diagonal entries of X i−1
numbered from upper left corner to lower right corner. Let P(X) denote the
determinant of D(X).
Theorem 10 ([You11]). P(X) is an irreducible polynomial.
Remark. P(X) = P(X − aI), where a ∈ K and I ∈ Mn (K) is the identity
matrix.
The next two lemmas are due to H. Young. They give us the connection
between the variety defined by P(X) and the algebraic set of nearly commuting
matrices.
Lemma 5 ([You11]). Given an n × n matrix A, if there exists a matrix B
such that [A, B] is a non-zero diagonal matrix, then P(A) = 0.
Lemma 6 ([You11]). There is a dense open set U in the variety defined by
P(X) where for every point A in U, there exists a matrix B such that [A, B]
is a nonzero diagonal matrix.
NEARLY COMMUTING MATRICES
13
The following notion of a discriminant is of significant importance in matrix
theory. We use it in this section in order to reduce our study to the case when
commuting matrices have a particularly simple characterization.
Definition. Let A ∈ Mn (K). Then the discriminant ∆(A) of A is the discriminant of its characteristic polynomial. That is, if K contains all the eigenvalues
Q
λ1 , . . . , λn of A, then ∆(A) = 1≤i<j≤n (λi − λj )2 .
Fact. Let A ∈ Mn (K) be a matrix such that ∆(A) 6= 0, or equivalently, A
has distinct eigenvalues. Then a matrix B commutes with A if and only if B
is a polynomial in A of degree at most n − 1, see Theorem 3.2.4.2 [HJ85].
Remark. P(X) is an irreducible polynomial of degree n(n − 1)/2 and ∆(X) is
a polynomial of degree n(n − 1). Moreover, when n ≥ 3, P(X) does not divide
∆(X). This can be proved by showing that there exists a matrix A with the
property that P(A) = 0 while ∆(A) 6= 0. For example, for this purpose one
can use the following matrices.
0 1 0 ... 0 0
0 0 1 ... 0 0
0
0
en =
if p ∤ n, and E
, otherwise.
En =
...
...
0 En−1
0 0 0 ... 0 1
1 0 0 ... 0 0
en .
The characteristic polynomials are xn − 1 for En and x(xn−1 − 1) for E
Here is the outline for how we prove the main result of this section, that is,
Rad(P + Q) is prime.
(1) dim V(P + Q) = n2 + n − 1 and V(P + Q) is equidimensional.
14
ZHIBEK KADYRSIZOVA
(2) dim V(P + Q, ∆(X)) ≤ n2 + n − 2, that is, ∆(X) is not in any of the
minimal primes of P + Q.
(3) Rad(P + Q)R∆(X) is a prime ideal.
We observe that P + Q has no minimal primes of height larger than one
over P and Q. First we need the following theorem due to R. Hartshorne.
Theorem 11 ([Har62] Proposition 2.1). Let A be a Noetherian local ring with
the maximal ideal m. If Spec(A) − {m} is disconnected, then the depth of A is
at most 1.
Lemma 7. Let P and Q be ideals as in Notation 1. Then every minimal
prime of P + Q has height n2 − n + 1.
Proof. Suppose that there exists a minimal prime ideal T of P + Q of height at
least ht(I)+2. Localize at T . Then (P +Q)(R/I)T is T (R/I)T -primary. Moreover, V(P ) and V(Q) are disjoint on the punctured spectrum Spec((R/I)T ) −
{T (R/I)T }. However, the above theorem shows that this is not possible.
Now let us define the set-up which we need to state and prove our next
result.
Let m be a positive integer such that m ≤ n. Fix a partition (h1 , . . . , hm )
of n, that is, choose positive integers h1 , . . . , hm such that h1 + . . . + hm = n.
Let Ji be an upper triangular Jordan form of a nilpotent matrix of size hi . For
each hi there are finitely many choices of Ji . Let J = (J1 , . . . , Jm ) and let Ii
denote the identity matrix of size hi .
For any m-tuple λ = λ1 , . . . , λm of distinct elements of K, let J(λ) =
J(λ1 , . . . , λm ) be a matrix such that for all 1 ≤ i ≤ m, the blocks λi Ii + Ji are
on the main diagonal. That is, J(λ) is the direct sum of matrices λi Ii + Ji .
NEARLY COMMUTING MATRICES
15
Let Λ = {(λ1 , . . . , λm ) ∈ Am | λi 6= λj for all 1 ≤ i 6= j ≤ m}. It is an open
subset of Am and therefore is irreducible and has dimension m. Let
WJ = {A ∈ Mn (K) | A is similar to some J(λ1 , . . . , λm ) with
λ1 , . . . , λm ∈ K distinct}.
Let cJ denote the dimension of the set of matrices that commute with J(λ),
for some λ. This number is independent of the choice of λ, since J(λ) commutes
with a matrix A if and only if A is a direct sum of matrices Ai such that each
Ai has size hi and Ai commutes with Ji . Moreover, cJ is the dimension of the
set of invertible matrices that commute with J(λ), for some λ.
Lemma 8. The dimension of WJ is n2 − cJ + m.
Proof. Define a surjective map of algebraic sets
θ : GLn (K) × Λ → WJ
such that
(U, λ1 , . . . , λm ) → U −1 J(λ1 , . . . , λm )U.
Fix λ = (λ1 , . . . , λm ). Then
θ−1 (J(λ)) = {(U, µ) ∈ GLn (K) × Λ | U −1 J(µ)U = J(λ)}
and it has the dimension of the set
{U ∈ GLn (K) | J(λ)U = UJ(λ)},
that is, it is the set of all invertible matrices commuting with J(λ).
16
ZHIBEK KADYRSIZOVA
Let M = U −1 J(λ)U ∈ WJ and let (V, µ) ∈ θ−1 (M). Then U −1 J(λ)U =
V −1 J(µ)V for some µ and J(µ) = (UV −1 )−1 J(λ)(UV −1 ). Hence, (V, µ) ∈
θ−1 (J(λ))U. Therefore, θ−1 (J(λ)) and θ−1 (M) have the same dimension.
Since the dimension of WJ is the dimension of GLn (K) × Λ minus the dimension of a generic fiber θ−1 (J(λ)), we have that the dimension of WJ is
n2 − cJ + m.
Moreover, the set of pairs of matrices (A, B) ∈ Mn (K) × Mn (K) such that
A and B commute has dimension (n2 − cJ + m) + cJ = n2 + m ≤ n2 + n.
Claim. Let
W = {(A, B) ∈ Mn (K)×Mn (K)| [A, B] = 0, ∆(A) = 0, P(A) = 0, P(B) = 0},
then there is an injective map
Ψ : V(P + Q, ∆(X)) → W
so that
(A, B) → (A, B).
Proof. Let (A, B) ∈ V(Q, ∆(X)) − V(P + Q, ∆(X)). Then by Lemma 5,
P(A) = P(B) = 0. Therefore, (A, B) ∈ W . Since V(Q) is the closure of
V(Q)−V(P +Q) we have that P(A) = P(B) = 0 for all (A, B) ∈ V(Q, ∆(X)).
Hence, V(P + Q, ∆(X)) ⊆ W .
Claim. The dimension of the set W = {(A, B) ∈ Mn (K) × Mn (K)| [A, B] =
0, ∆(A) = 0, P(A) = 0, P(B) = 0} is at most n2 + n − 2.
Proof. Let V = {(A, B) ∈ Mn (K) × Mn (K)| [A, B] = 0, ∆(A) = 0} and Vm =
{(A, B) ∈ V | A has m distinct eigenvalues }. Then we have that dim Vm =
NEARLY COMMUTING MATRICES
n2 + m and V =
Sn−1
m=1
17
Vm . Therefore, dim V ≤ n2 + n − 1. Notice that since
∆(A) = 0, m ≤ n − 1.
Similarly, let Wm = {(A, B) ∈ W | A has m distinct eigenvalues }. Then
Sn−1
Wm . For each value of m, Wm ⊆ Vm . Therefore, the dimension
W = m=1
of W is at most n2 + n − 1. Moreover, W is a closed subset of V defined by
the vanishing of P(X) and P(Y ). To prove the claim we need to show that
dim W cannot be n2 + n − 1. We do this by showing that W does not contain
any component of V of dimension n2 + n − 1. In other words, we show that
there are pairs of matrices (A, B) ∈ V but not in W , i.e., either P(A) 6= 0 or
P(B) 6= 0.
Let A ∈ Mn (K) be a matrix with distinct eigenvalues λ = λ1 = λ2 , λ3 , . . . , λn .
Then A is similar to a Jordan matrix
λ
0
Case 1. A is similar to J =
0
...
0
in two possible forms.
0 0 0 ... 0
λ 0 0 ... 0
0 λ3 0 . . . 0
...
...
0 0 0 . . . λn
Take B = diag(a1 , . . . , an ) be a diagonal matrix with distinct entries on the
Q
diagonal. Then [A, B] = 0 and P(B) = 1≤i<j≤n (aj − ai ) 6= 0.
Case 2. A is similar to J =
λ
1
0
0 ...
0
λ
0
0 ...
0
0 λ3 0 . . .
...
0
...
0
0
...
0 ...
0
0
0
λn
18
ZHIBEK KADYRSIZOVA
Write J =
J0
0
0
, where J0 =
J1
λ4 0
0 λ5
0 λ 0 and J1 =
0 0 λ3
0 0
λ 1
0
Take an n by n block-diagonal matrix U =
U0
0
0
U1
...
0
...
0
...
. . . λn
such that U0 =
0 0 1
1 1 0 and U1 ∈ Mn−3 (K) is the identity matrix.
0 1 1
−1
−1
U
0
0
U
J
U
0
and U −1 JU = 0 0 0
.
Then U −1 =
−1
0 U1
0
J1
Our goal is to show that P(U −1 JU) 6= 0. First, we prove it for the case of
3 by 3 matrices, i.e., for U0−1 J0 U0 .
Observe that P(U0−1 J0 U0 ) = P(U0−1 J0 U0 − λI) = P(U0−1 (J0 − λI)U0 ).
Denote M = U −1 (J − λI)U and M0 = U0−1 (J0 − λI)U0 .
We have that
U0−1
1 1 −1
= −1 0 1
1 0 0
and
1
1 − (λ3 − λ) −(λ3 − λ)
M0 = U0−1 (J0 − λI)U0 = −1 −1 + (λ3 − λ) (λ3 − λ)
1
1
0
.
.
NEARLY COMMUTING MATRICES
Moreover,
0 −(λ3 − λ)2 −(λ3 − λ)2
M02 = 0
0
(λ3 − λ)
2
(λ3 − λ)
0
2
0
19
.
In particular, the diagonal diag(M0i ) = (0, (λ3 − λ)i , 0) for all i ≥ 2. Then
1
1
0
P(M0 ) = det 1 −1 + (λ3 − λ) (λ3 − λ)2
1
0
0
= (λ3 − λ)2 6= 0.
Finally,
P(M) = det
1
1
1
0
−1 + λ3 − λ (λ3 − λ)
1
0
1
λ4 − λ
2
0
(λ3 − λ)
...
3
0
...
λn − λ
. . . (λ3 − λ)
...
n−1
0
...
(λn − λ)2 (λn − λ)3 . . . (λn − λ)n−1
n
Y
=
(λi − λ)2
i=3
0
(λ4 − λ)2 (λ4 − λ)3 . . . (λ4 − λ)n−1
...
1
0
Y
=
(λj − λi ) 6= 0.
3≤i<j≤n
The final expression for the determinant is nonzero. Hence dim W ≤ n2 +n−2.
Thus we have that ∆(X) is not in any of the minimal primes of P + Q.
Now we prove that P + Q has only one minimal prime.
20
ZHIBEK KADYRSIZOVA
Theorem 12. Let P and Q be as in Notation 1. Then V(P +Q) is irreducible,
i.e., Rad(P + Q) is prime.
Proof. Let U be a dense open subset in the algebraic set defined by P(X) as in
Lemma 6. Let A ∈ Mn (K) be such that P(A) = 0. Suppose that A ∈ U. Then
by Lemma 6 there exists a matrix B such that (A, B) is in the skew-component
of the algebraic set of nearly commuting matrices, that is (A, B) ∈ V(Q). Let
K[t] be a polynomial ring in one independent variable t. Fix any f ∈ K[t].
Then (A, cB + f (A)) ∈ V(Q) for all c ∈ K − {0}. Since Q defines a closed set,
we must have that (A, f (A)) ∈ V(Q), i.e., when c = 0 as well. Since U is a
dense subset in V(P(X)), (A, f (A)) ∈ V(Q) for all A ∈ V(P(X)). Recall that
f was an arbitrary element of K[t].
Now assume also that ∆(A) 6= 0. Then every matrix B that commutes with
A is a polynomial in A of degree at most n − 1. Thus
V(P )∆(X) = {(A, f (A)) | ∆(A) 6= 0 and f is a polynomial of degree at most n−1}.
Moreover, since V(P + Q) ⊂ V(P ), every element of V(P + Q)∆(X) is of the
form (A, f (A)), where P(A) = 0 and f is a polynomial of degree at most n−1.
Identify polynomials f ∈ K[t] of degree at most n − 1 with An . Then we
can consider a map
V(P(X))∆(X) × An → V(P + Q)∆(X)
such that
(A, f ) → (A, f (A)).
Moreover, this map is a bijective morphism. Therefore, V(P + Q)∆(X) is
irreducible. If V(P + Q) is not irreducible, then its nontrivial irreducible
NEARLY COMMUTING MATRICES
21
decomposition will give us a nontrivial irreducible decomposition of V(P +
Q)∆(X) , since ∆(X) is not in any of the minimal primes of P + Q. Thus the
result.
Corollary 4. Let P and Q be as in Notation 1. Then, when n = 3, P + Q is
prime.
4. The ideal of nearly commuting matrices is a radical ideal
In this section we prove that I is a radical ideal in all characteristics. We
T
know that Rad(I) = P Q and I is unmixed as the heights of P and Q are
equal to n2 − n. To prove the result it is sufficient to show that I becomes
prime or radical once we localize at P or Q.
Theorem 13. The defining ideal of the algebraic set of nearly commuting
matrices is radical.
Proof. For simplicity of notation, let P denote P(X).
T
T
We have that K[X] P = (0), since every f ∈ K[X] P must vanish
when we set X = Y . Therefore, W = K[X] − {0} is disjoint from P and
hence from I. Localize at P . Then we have an injective homomorphism of
K[X, Y ]/I-modules
(K[X, Y ]/I)P ֒→ (K(X)[Y ]/I)P ∼
= (L[Y ]/I)P ,
where L = K(X) and now I is an ideal generated by n2 − n linear equations
in the entries of Y with coefficients in L. We can always choose at least n
variables yij , (i, j) ∈ Λ, and write the rest of them as L-linear combinations of
the chosen ones. Thus (K[X, Y ]/I)P ֒→ L[yij ](i,j)∈Λ and IK[X, Y ]P is prime.
22
ZHIBEK KADYRSIZOVA
T
Next observe that K[X] Q = (P). Clearly, (P) ⊆ Q. To prove the other
T
direction, let f ∈ K[X] Q be nonzero. Then by Lemma 5, f ∈ (P). In other
words, for all A ∈ Mn (K) such that A ∈ V(Q) and such that there exists a
matrix B with the property that [A, B] is nonzero diagonal, then P(A) = 0.
Therefore, we have an injective homomorphism of K[X, Y ]/I-modules
(K[X, Y ]/I)Q ֒→ (V [Y ]/I)Q ,
where V = K[X](P) is a discrete valuation domain. Then generators of I
become linear polynomials in the entries of Y with coefficients in V . Let
B be the matrix of coefficients of this linear system such that its rows are
indexed by (i, j) for 1 ≤ i 6= j ≤ n and columns are indexed by (h, k) for all
1 ≤ h, k ≤ n. Then for i 6= h and k 6= j B has an entry xih in the (i, k), (h, k)
spot, has an entry −xkj in the (i, j), (i, k) spot, and xii − xjj in the (i, j)(i, j)
spot and zero everywhere else. Let y1 , . . . , yn2 denote the entries of Y such
that y(i−1)n+j = yij . In V [Y ], I is generated by the entries of the matrix
y1
y2
B
.
...
y n2
By doing elementary row operations over V , we can transform B into a
diagonal matrix C. This gives new generators of I. To prove that IV [Y ] is
radical, it is sufficient to show that the diagonal entries in C have order at
most one in V . To this end it reduces to show that C has rank n2 − n and
the ideal generated by the minors of C of size n2 − n cannot be contained in
NEARLY COMMUTING MATRICES
23
P 2 V . But then it is sufficient to prove this for the original matrix B. Hence
it suffices to show:
Claim.
(1) The submatrix B0 of B obtained from the first n2 − n columns
has nonzero determinant in V .
(2) The determinant of B0 is in (P) − (P 2 ).
Proof.
(1) It is sufficient to prove the first part of the claim over K(X) = frac(V ), i.e.,
after we invert P. In this case, since X and Y nearly commute, they must
commute, see Lemma 5. Moreover, X is a generic matrix, hence its discriminant is nonzero and is not divisible by P. Thus X has distinct eigenvalues
and Y is a polynomial in X of degree at most n − 1. Write B = [B0 |B1 ], then
our equations become
y1
y2
B0
...
yn2 −n
yn2 −n+1
yn2 −n+2
+ B1
...
y n2
= 0.
Notice that B0 is invertible if and only if for every choice of the values for
[yn2 −n+1 , . . . , yn2 ] there is a unique solution for the above equation.
Furthermore, the bottom rows of X 0 , X, . . . , X n−1 are linearly independent
for a generic matrix X. This is true because it even holds for the permutation
24
ZHIBEK KADYRSIZOVA
matrix
0 1
0 0
E=
...
0 0
1 0
0 ... 0 0
1 ... 0 0
...
0 ... 0 1
0 ... 0 0
for which the bottom rows of E 0 , E, . . . , E n−1 are the standard basis vectors
ei for 1 ≤ i ≤ n.
Thus, given any bottom row ρ of Y , there exist α0 , . . . , αn−1 ∈ K(X) such
that ρ equals the bottom row of α0 + α1 X + . . . + αn−1 X n−1 . That is, such
a Y is uniquely determined by the entries of its bottom row. Therefore, B0 is
invertible in K(X).
(2) First, let us show that det B0 ∈ (P). For any matrix A in an open dense
subset of the closed set defined by P, there exists a matrix A′ such that the
commutator [A, A′ ] is a nonzero diagonal matrix, see Lemma 6 . Hence, for
all c ∈ K − {0} and for all f ∈ K[X] polynomials of degree at most n − 1,
(A, cA′ + f (A)) ∈ V(I). Therefore, the space of solutions of
y1
y2
B·
=0
...
y n2
has dimension n + 1, but we showed in (1) that it must be n. Therefore, the
minors of B must vanish whenever P vanishes.
NEARLY COMMUTING MATRICES
25
Notice that the degree of the polynomial P is n(n − 1)/2, while the degree
of det B0 is n(n − 1). Therefore, to prove part (2) it is sufficient to show that
det B0 is not a K-scalar multiple of P 2 .
Now let us put grading on the entries of X and Y . Let deg xij = deg yij =
i − j. Then their products XY and Y X and sums have this property as
well: deg (XY )ij = i − j and deg (X + Y )ij = i − j. Therefore, so does the
commutator matrix XY − Y X. In fact, any polynomial in X and Y has this
property. Notice that the diagonal entries have degree 0, thus P has degree
0. However, this is not the case for the determinant of the matrix B0 . The
nonzero entry corresponding to (i, j), (k, h) has degree i − j + h − k. Therefore,
if a product of the entries is a nonzero term of the determinant of B0 , then its
P
P
P
degree is 1≤i6=j≤n 1≤h≤n 1≤k<n (i − j + h − k) = n2 (n − 1)2 /2 6= 0 for all
n ≥ 2. Hence det B0 cannot be a K-scalar multiple of P 2 . That is, when we
factor out P from the minors of B, the remaining expression is not divisible
by P.
Now we are ready to finish our discussion. Let C = [C0 |C1 ] be a matrix
that is obtained from B by elementary row transformations so that C0 =
diag(cii )1≤i≤n2 −n is a diagonal matrix. We proved that there exists 1 ≤ k ≤
n2 −n with the property that cii is a unit in V for all i 6= k and ckk ∈ (P)−(P 2 ).
Denote ckk = αP, where α is a unit in V .
26
ZHIBEK KADYRSIZOVA
The ideal I is generated by the following equations
c11 0 0 . . .
0
y1
0 c22 0 . . .
0
y2
...
...
...
2
2
0
0 0 . . . cn −n−1,n −n−1
0
yn2 −n
0
0 0 ...
0
cn2 −n,n2 −n
yn2 −n+1
yn2 −n+2
+C1
...
y n2
.
Pn
Then V [Y ]/I ∼
= V [yn2 −n+1 , yn2 −n+2 , . . . , yn2 ][ykk ]/(αP − j=1 (C1 )kj yn2 −n+j )
is reduced. To show this we consider two cases. If all (C1 )kj ∈ P, then
the last factor ring is isomorphic to V [yn2 −n+1 , yn2 −n+2 , . . . , yn2 ][z]/(zP). If
there is j so that (C1 )kj is a unit, then the factor ring is isomorphic to
V [yi ]n2 −n+1≤i6=j≤n2 [ykk ]. In either case, it is reduced. Therefore, since we have
an injective map (R/I)Q ֒→ (V [Y ]/I)Q , IRQ is radical.
5. Conjectures
In this section we state conjectures that we have made while doing the
research. Many of them appeared as a result of computations performed on a
computer algebra program Macaulay2, [GS].
Conjecture 2. Let R be as in Notation 1. Then R/P , R/Q and R/(P + Q)
are F -regular.
Remark. In the case when n = 2 the conjecture is true.
The following lemma allows us to reduce the above conjecture to the F regularity of R/(P + Q).
NEARLY COMMUTING MATRICES
27
Lemma 9. Let R be a Noetherian local or N-graded ring of prime characteristic p > 0 and let I be an ideal (homogeneous in the graded case) generated by
a regular sequence. Let P and Q be ideals of R of the same height such that P
T
and Q are linked via I = P Q. Let R/P be Cohen-Macaulay. Suppose that
R/(P + Q) is F -regular (or equivalently, F -rational). Then R/P and R/Q are
F -regular.
Proof. By [PS74], R/Q is Cohen-Macaulay and has the canonical module isomorphic to (P + Q)/Q. Similarly, the canonical module of R/P is (P + Q)/P .
Then R/(P + Q) is Gorenstein, hence it is F -rational if and only if it is F regular.
Recall that a graded ring R is F -regular if and only if Rm is F -regular,
[LS99]. Then R/(P + Q) is F -rational if and only if its localization at the
homogeneous maximal ideal is F -rational. Then by applying Corollary 2.9 in
[Ene03] we have that F -rationality of R/(P + Q) implies F -regularity of R/P
and R/Q.
Thus if we want to prove that the variety of commuting matrices and the
skew component are F -regular, it is sufficient to prove the statement for their
intersection. Of course we need to know whether R/P is Cohen-Macaulay.
Conjecture 3. R/I is F -pure for all n.
The above conjecture can be solved by proving the following one.
Conjecture 4. Let µ =
Qn−1
i=1
Qn
i=1,j=1
Then µp−1 is a monomial term of
modulo p.
xij yij
xii yi,n−i+1 ·xn,n−1 ·yn−1,1
Q
1≤i6=j≤n
.
upij with coefficient equal to 1
28
ZHIBEK KADYRSIZOVA
Remark. The above monomial can be obtained taking the product of all the
variables and dividing by the variables according to the following pattern:
denote by ⋆ the variable to be divided out.
⋆
x12
...
x1,n−2
x1,n−1
x1n
x21
⋆
...
x2,n−2
x2,n−1
x2n
...
X=
...
xn−2,1 xn−2,2 . . .
⋆
xn−1,1 xn−1,2 . . . xn−1,n−2
xn,1
xn,2
...
xn,n−2
,
xn−2,n−1 xn−1,n
⋆
xn−1,n
⋆
xn,n
y11
y12
y1,3
...
y1,n−1
⋆
y21
y22
y2,3
...
⋆
y2n
...
Y =
yn−2,1 yn−2,2
⋆
⋆
yn,1
yn,2
...
⋆
. . . yn−2,n−1 yn−2,n
yn−1,3 . . . yn−1,n−1 yn−1,n
yn,3
...
yn,n−1
yn,n
Conjecture 5. Let X be a matrix of indeterminates of size n over a field K.
Let P(X) be the irreducible polynomial as in Definition 3. Then K[X]/P(X)
is F -regular.
Conjecture 6. The following is a regular sequence on R/I and hence a part
of a system of parameters on R/J and R/Q.
xst − yt,θ(s,t) , x1n , xnn , x11 − y2n
NEARLY COMMUTING MATRICES
(s + t)mod n,
for all 1 ≤ s, t, ≤ n and where θ(s, t) =
n,
29
if s + t 6= n;
if s + t = n.
Remark. The conjecture was verified by using Macaulay2 software when n =
3, 4 over K = Q and in some small prime characteristics.
In the case when n = 3, this is equivalent to the following identifications of
variables in matrices X and Y
x11 x12
X=
x21 x22 x22 ,
x31 x32
x31 x11 x21
0
0
Y =
x22 x32 x12 .
0
x22
0
Conjecture 7. Let Z ⊆ {uij | 1 ≤ i 6= j ≤ n} be any subset of cardinality at
most n2 − n − 1. Let IZ be the ideal of R generated by the elements of Z.
Then R/IZ is F -regular. In particular, IZ is a prime ideal.
6. Acknowledgement
This paper is the part of the thesis written by the author during her doctorate study at the University of Michigan. The author would like to express
an enormous gratitude to Mel Hochster for the support and guidance received
while working on the paper. This work was supported in part by NSF grant
DMS-1401384 and the Barbour Scholarship at the University of Michigan.
30
ZHIBEK KADYRSIZOVA
References
[Ene03] F. Enescu, Applications of pseudocanonical covers to tight closure problems, J. Pure
Appl. Algebra 178(2) (2003), 159–167.
[Fed87] R. Fedder, F-purity and rational singularity in graded complete intersection rings,
Transactions of the American Mathematical Society 301 (1987), no. 1, 47–62.
[Ger61] M. Gerstenhaber, On dominance and varieties of commuting matrices, Ann. Math.
73 (1961), 324–348.
[GS]
D. Grayson and M. Stillman,
tem
for
algebraic
geometry
Macaulay 2:
and
commutative
a computer algebra sysalgebra,
available
at
http://www.math.uiuc.edu/Macaulay2.
[Har62] R. Hartshorne, Complete intersections and connectedness, American Journal of
Mathematics 84 (1962), no. 3, 497–508.
[HH94] M. Hochster and C. Huneke, Tight closure of parameter ideals and splitting in
module-finite extensions, J. Algebraic Geom. 3 (1994), no. 4, 599–670.
[HJ85]
R. Horn and C. Johnson, Matrix analysis, Cambridge Univ. Press, Cambridge, UK,
1985.
[Knu05] A. Knutson, Some schemes related to the commuting variety, Journal of Algebraic
Geometry 14 (2005), 283–294.
[LS99]
G. Lyubeznik and K. Smith, Strong and weak F-regularity are equivalent for graded
rings, American Journal of Mathematics 121 (1999), no. 6, 1279–1290.
[PS74]
C. Peskine and L. Szpiro, Liaison des variétés algébriques, i, Invent. Math. 26
(1974), 271–302.
[ST12]
K. Schwede and K. Tucker, A survey of test ideals, Progress in Commutative Algebra 2, Closures, Finiteness and Factorization, Walter de Gruyter GmbH and Co.
KG Berlin (2012), 39–99.
[Tho85] M. Thompson, Topics in the ideal theory of commutative Noetherian rings, Ph.D.
thesis, University of Michigan, 1985.
[You11] H. Young, Components of algebraic sets of commuting and nearly commuting matrices, Ph.D. thesis, University of Michigan, 2011.
NEARLY COMMUTING MATRICES
Email address: [email protected]
Department of Mathematics,
School of Science and Technology,
Nazarbayev University
53 Kabanbay Batyr Ave, Astana 010000, Kazakhstan
31
| 0math.AC
|
Learning and Real-time Classification of
Hand-written Digits With Spiking Neural Networks
Shruti R. Kulkarni, John M. Alexiades, Bipin Rajendran
arXiv:1711.03637v1 [stat.ML] 9 Nov 2017
Department of Electrical and Computer Engineering,
New Jersey Institute of Technology, Newark, NJ, 07102, USA
Email: {srk68, jma59, bipin}@njit.edu
Abstract—We describe a novel spiking neural network (SNN)
for automated, real-time handwritten digit classification and its
implementation on a GP-GPU platform. Information processing
within the network, from feature extraction to classification is
implemented by mimicking the basic aspects of neuronal spike
initiation and propagation in the brain. The feature extraction
layer of the SNN uses fixed synaptic weight maps to extract
the key features of the image and the classifier layer uses the
recently developed NormAD approximate gradient descent based
supervised learning algorithm for spiking neural networks to
adjust the synaptic weights. On the standard MNIST database
images of handwritten digits, our network achieves an accuracy
of 99.80% on the training set and 98.06% on the test set, with
nearly 7× fewer parameters compared to the state-of-the-art
spiking networks. We further use this network in a GPU based
user-interface system demonstrating real-time SNN simulation to
infer digits written by different users. On a test set of 500 such
images, this real-time platform achieves an accuracy exceeding
97% while making a prediction within an SNN emulation time
of less than 100 ms.
Index Terms—Spiking neural networks, classification, supervised learning, GPU based acceleration, real-time processing
I. I NTRODUCTION
The human brain is a computational marvel compared to
man-made systems, both in its ability to learn to execute highly
complex cognitive tasks, as well as in its energy efficiency. The
computational efficiency of the brain stems from its use of
sparsely issued binary signals or spikes to encode and process
information. Inspired by this, spiking neural networks (SNNs)
have been proposed as a computational framework for learning
and inference [1]. General purpose graphical processing units
(GP-GPUs) have become an ideal platform for accelerated
implementation of large scale machine learning algorithms [2].
There have been multiple GPU based implementations for
simulating large SNNs [3]–[8], with most of these targeting
the forward communication of spikes through large networks
of spiking neurons and/or local weight update based on
spike timing difference. In contrast, we demonstrate a highly
optimized real time implementation scheme for spike based
supervised learning on GPU platforms and use the framework
for real time inference on digits captured from different users
through a touch-screen interface.
Previous efforts to develop deep convolutional spiking networks started by using second generation artificial neural
This research was supported in part by grants from National Science
Foundation Award 1710009, CISCO, and the McNair Fellowship Program.
networks (ANNs) with back-propagation of errors to train the
network and thereafter converting it into spiking versions [9]–
[12]. There have been several supervised learning algorithms
proposed to train the SNNs, by explicitly using the time
of spikes of neurons to encode information, and to derive
the appropriate weight update rules to minimize the distance
between desired spike times and observed spike times in
a network [13]–[17]. We use the Normalized Approximate
Descent (NormAD) algorithm to design a system to identify
handwritten digits. The NormAD algorithm has shown superior convergence speed compared to other methods such as the
Remote Supervised Method (ReSuMe) [13].
Our SNN is trained on the MNIST database consisting of
60, 000 training images and 10, 000 test images [18]. The
highest accuracy SNN for the MNIST was reported in [16],
where a two-stage convolution neural network achieved an
accuracy of 99.31% on the test set. Our network, in contrast,
has just three layers, with about 82, 000 learning synapses (7×
fewer parameters compared to [16]) and achieves an accuracy
of 98.06% on the MNIST test dataset.
The paper is organized as follows. The computational units
of the SNN and the network architecture are described in
section II. Section III details how the network simulation is
divided among different CUDA kernels. The user-interface
system and the image pre-processing steps are explained in
Section IV. We present the results of our network simulation
and speed related optimizations in Section V. Section VI
concludes our GPU based system implementation study.
II. S PIKING N EURAL N ETWORK
The basic units of an SNN are spiking neurons and synapses
interconnecting them. For computational tractability, we use
the leaky integrate and fire (LIF) model of neurons, where the
evolution of the membrane potential, Vm (t) is described by:
dVm (t)
= −gL (Vm (t) − EL ) + I(t)
(1)
C
dt
Here I(t) is the total input current, EL is the resting potential, and C (300 pF) and gL (30 nS) model the membrane
capacitance and leak conductance, respectively [13]. Once the
membrane potential crosses a threshold (Vm (t) ≥ VT ), it is
reset to its resting value EL and remains at that value till
the neuron comes out of its refractory period (tref = 3 ms).
The synapse, with weight wk,l connecting input neuron k to
output neuron l, transforms the incoming spikes (arriving at
©2017 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including
reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or
reuse of any copyrighted component of this work in other works.
times t1k , t2k , . . .) into a post-synaptic current (Ik,l ), based on
the following transformation,
X
(2)
ck (t) =
δ(t − tik ) ∗ e−t/τ1 − e−t/τ2
i
Ik,l (t) = wk,l × ck (t)
(3)
Here, the summed δ function represents the incoming spike
train and the double decaying exponentials with τ1 = 5 ms
and τ2 = 1.25 ms represent the synaptic kernel. These values
closely match the biological time constants [19].
in the output layer generates a spike train with a frequency
close to 285 Hz and the other output neurons issue no spikes
during the presentation duration, T (set to 100 ms in baseline
experiments). T is also a hyper-parameter of our network,
and its effect on the network’s classification ability will be
discussed in section V. This layer also has lateral inhibitory
connections that helps to prevent the non-label neurons from
spiking for a given input. The output neuron with the highest
number of spikes is declared the winner of the classification.
B. Learning layer
A. Network architecture & spike encoding
We use a three-layered network where hidden layer performs feature extraction and the output layer performs classification (see Fig. 1). The network is designed to take input
from 28×28 pixel MNIST digit image. We translate this pixel
value into a set of spike streams, by passing the pixels as
currents to a layer of 28 × 28 neurons (first layer). The current
i(k) applied to a neuron corresponding to pixel value k, in the
range [0, 255] is obtained by the following linear relation:
i(k) = I0 + k × Ip
(4)
where Ip = 101.2 pA is a scaling factor, and I0 = 2700 pA is
the minimum current above which an LIF neuron can generate
a spike (for the parameters chosen in equation 1). These spike
streams are then weighted with twelve 3 × 3 synaptic weight
maps (or filters), with a priori chosen values to generate
equivalent current streams using equations 2 and 3. These
12 spatial filter maps are chosen to detect various edges and
corners in the image.
Convolution layer
Input
currents
28 x 28 neurons
28x28 pixel
Input Image
Input Layer
Spike
Trains
Twelve spatial
filters of size 3x3
Learning Synapses
(Fully Connected Layer)
10 neurons
26x26x12 neurons
Output Layer
All-to-all lateral inhibitory
connections
The 8112×10 synapses connecting the hidden layer neurons
to the 10 output layer neurons are modified during the course
of training using the NormAD rule [13]. The strength of the
weights are adjusted based on the error between the observed
and desired spike streams (e(t) = S d (t) − S o (t)) and the
ˆ
term d(t),
denoting the effect of incoming spike kernels on
the neuron’s membrane potential, according to the relation:
Z T
d̂(t)
dt,
d̂(t) = c(t) ∗ ĥ(t)
(5)
∆w = r
e(t)
k
d̂(t)k
0
where ĥ(t) = (1/C) exp(−t/τL ) represents the neuron’s
impulse response with τL = 1 ms, and r is the learning rate.
III. CUDA IMPLEMENTATION
The SNN is implemented on a GPU platform using the
CUDA-C programming framework. A GPU is divided into
streaming multiprocessors (SM), each of which consists of
stream processors (SP) that are optimized to execute math
operations. The CUDA-C programming framework exploits
the hardware parallelism of GPUs and launches jobs on the
GPU in a grid of blocks each mapped to an SM. The blocks
are further divided into multiple threads, each of which is
scheduled to run on an SP, also called a CUDA core. Since
memory transfer between CPU and GPU local memory is one
of the main bottlenecks, all network variables (i.e., neuron
membrane potentials and synaptic currents) are declared in the
global GPU memory in our implementation. The simulation
equations (1), (2) and (3) are evaluated numerically in an
iterative manner at each time step.
Next epoch
Next time step
The output layer consists of 10 neurons, one for each of
the ten digits. We train the network so that the correct neuron
Read in
all images
from MNIST
Generate
spike trains
GPU device
memory
LIF Layer -1
784 neurons
Image
(28x28)
Fig. 1: The 28×28 pixel images from the MNIST database are
converted to spike trains, which are presented for a duration
T , weighted with twelve 3 × 3 synaptic weight maps (below)
resulting in twelve 26 × 26 current streams and then feed to
the corresponding feature map neurons. There are 10 output
neurons corresponding to each digit. The weights of the fullyconnected feed-forward synapses to the output layer neurons
(8112 × 10) are adjusted using the NormAD learning rule
[13]. Additionally, the output layer neurons also have lateral
inhibitory connections.
LIF Layer-2
(8112 neurons)
Synaptic
current
Current summation
(10 currents)
Spike
Trains
Synaptic currents LIF Layer-3
12 Convolution kernels
(10 neurons)
(26x26 synapses per map) (8112x10 synapses)
Spike
errors
Weight Update
(8112x10 synapses)
Netowork
Output
Compute del_w
Compute d_hat
(8112x10 synapses)
(8112 Spike Trains)
Fig. 2: Diagram showing the different variables of the network
being computed each time step and how the signals flow across
different layers. The dimensions within the brackets are the
sizes of those variables and their respective CUDA kernels.
Fig. 2 shows the forward pass and backward pass for weight
update during the training phase. Image pixels read into the
GPU memory are passed as currents to layer one neurons (grid
size of 28 × 28) for the presentation duration, T . The filtering
process involves 2D convolution of the incoming spike kernels
and the weight matrix (3 × 3). The computation is parallelized
across 12 CUDA kernels, each with a grid size of 26 × 26
threads. Each thread computes the current to the hidden layer
neurons, indexed as a 2D-array i, j, {0 ≤ i, j, ≤ 25} at a timestep n, based on the following spatial convolution relation:
2 X
2
X
Iin (i, j, n) =
wconv (a, b) × c(i + a, j + b, n) (6)
a=0 b=0
where c represents the synaptic kernel (equation 2) calculated
from the spike trains of the 28 × 28 pixels and wconv (a, b)
represents each of the weights from the 3 × 3 filter matrix.
The membrane potential of an array of k LIF neurons, for
applied current I(n) (as described in equation 1) is evaluated
using the second order Runge-Kutta method as:
k1 = [−gL (Vm (n) − EL ) + I(n)]/C
(7)
k2 = [−gL (Vm (n) + k1 ∆t − EL ) + I(n)]/C
(8)
Vm (n + 1) = Vm (n) + [(k1 + k2 )∆t/2]
(9)
Each thread k independently checks if the membrane potential
has exceeded the threshold to artificially reset it.
If Vmk (n + 1) ≥ VT ⇒ Vmk (n + 1) = EL
IV. R EAL - TIME INFERENCE ON USER DATA
We used the CUDA based SNN described in the previous
section, to design a user interface that can capture and identify
the images of digits written by users in real-time from a touchscreen interface. The drawing application to capture the digit
drawn by the user is built using OpenCV, an image processing
library [21]. The captured image from the touch screen is
pre-processed using standard methods similar to that used to
generate the MNIST dataset images [18]. We convert the user
drawn images to the required format which is a grayscale
image of size 28 × 28 pixels. The network is implemented on
the NVIDIA GTX 860M GPU which has 640 CUDA cores.
The preprocessing phase takes about 15 ms and this image
is then passed to the trained SNN for inference. The CUDA
process takes about 300 ms to initialize the network in the GPU
memory, after which the network simulation time depends on
the presentation time T and the time step interval ∆t.
(a).
(b).
(10)
Refractory period is implemented by storing the latest spike
issue time, nlast
of each neuron in a vector R; the membrane
k
potential of a neuron is updated only when the current time
step n > nlast
+ (tref /∆t).
k
The synaptic current from neuron k in hidden layer to
neuron l in output layer as given in equation 3 can be re-written
to be evaluated in an iterative manner, thereby avoiding the
evaluation of expensive exponential of the difference between
the current time n and previous spike times nik . The synaptic
current computation, at time step n, for each of the (k, l)
synapse is spawned in CUDA across 8112 × 10 kernels as:
ak (n) = ak (n − 1) × exp(−∆t/τ1 ) + δ(n − nik )
(11)
nik )
(12)
bk (n) = bk (n − 1) × exp(−∆t/τ2 ) + δ(n −
update the synaptic weights in parallel. The evaluation of the
total synaptic current and the norm is performed using parallel
reduction in CUDA [20]. During the inference or testing phase,
we calculate the synaptic currents and membrane potentials of
neurons in both layers to determine spike times, but do not
evaluate the d̂ term and the weight update.
ck (n) = ak (n) − bk (n)
(13)
Ik,l = wk,l × ck (n)
(14)
where ak (n) and bk (n) represent the rising and falling regions
of the double exponential synaptic kernel. The strength of the
synapses between the hidden and output layers is initialized
to zero during training. At every time step, the error function
for each output neuron is calculated, based on the difference
between the observed and desired spikes. Next, dˆk (equation
5) for the spikes originating from neuron k is computed as:
dˆk (n) = dˆk (n − 1)e−∆t/τL + (ck (n)∆t)/C
(15)
Once dˆk (n) is evaluated, we compute its norm across all
k neurons and determine the instantaneous ∆wk,l (n) for all
the 81, 120 synapses in parallel, if there is a spike error. At
the end of presentation, the accumulated ∆wk,l is used to
Drawn Image
Inverted Image
Bounding
boxed Image
Normalized and
center-massed
28x28 Image
Fig. 3: (a) Outline of the preprocessing steps used to convert
the user input to a 28×28 image that is fed to the network, (b)
Examples of user input (left) and the pre-processed 28 × 28
pixel images fed to the SNN (right).
A. Image Preprocessing
Fig. 3(a) shows the preprocessing steps used to create the
input signal to the SNN from the captured image and Fig. 3(b)
shows some sample pre-processed images. The image captured
from the user is first binarized by thresholding and cropped
to remove excess background. The image is resized to 20
pixels along its longer dimension, while maintaining its aspect
ratio. Thereafter, the resized image is placed in a 28 × 28
bounding box such that the image’s center of mass coincides
with the center of the bounding box. Finally, the image is
passed through a blurring filter to create gray-scale images
similar to the ones in the MNIST dataset.
V. R ESULTS
We trained the network on the MNIST training data-set
consisting of 60, 000 images, for 20 epochs. Our network
achieves an error of 0.2% on the training set and 1.94% on
the test set with a time step of ∆t = 0.1 ms when the network
is simulated for T = 100 ms. Table I lists the state-of-theart networks (ANN and SNN) for the MNIST classification
problem. It can be seen that though these networks have
classification accuracies exceeding 99%, they use more than
7× the number of parameters compared to our network, which
is designed to simplify the computational load in developing
real-time system.
TABLE I: Comparison of our SNN with state-of-the-art
Network and learning algorithm
Deep Learning [22]
ANN converted to SNN [9]
4-layer convolution SNN [16]
SNN, with NormAD (this work)
Learning synapses
2, 508, 470
1, 422, 848
581, 520
81, 120
Accuracy
99.79%
99.12%
99.31%
98.06%
Test accuracy (%)
If the integration time step interval used during inference
is 1 ms (i.e., approximating the neuronal integration) instead
of 0.1 ms, the MNIST test error increases only by about 0.4%
(see Fig. 4(a)), but there is a 10× reduction in the processing
time. Hence, for our touch screen based interface system we
simulate the SNN with ∆t of 1 ms to infer the users’ digits.
When each digit is presented for T = 75 ms, the network
can be simulated in an average wall clock time of 65 ms,
making real-time processing possible (Fig. 4(b)). We tested
(a).
0
28 x 28
Image
98.0
4
9
97.8
dt=0.1 ms
dt=1.0 ms
97.6
97.4
50
100
150
Presentation time [ms]
Image
Capture
t=0
Preprocessing
complete
t=15
(b).
Output
Spikes
SNN
prediction
and result
display
t=80
Time [ms]
Fig. 4: (a). MNIST test-set accuracy as a function of presentation time and the integration time step ∆t. (b) Various stages
of classifying a user’s input: the image pre-processing takes
15 ms and the 75 ms SNN emulation is completed in real-time.
the network’s accuracy with ∆t = 1 ms on a set of 500
handwritten digits collected from various users through our
user-interface system. At T = 75 ms, we measure an accuracy
of 97.4% on our set of 500 captured images, while on the
MNIST test-set it was 97.68%. The slight loss in performance
compared to the MNIST dataset is attributed to the deviations
from the statistical characteristics of the captured images
compared to the MNIST dataset.
VI. C ONCLUSION
We developed a simple three-layer spiking neural network
that performs spike encoding, feature extraction, and classification. All information processing and learning within the
network is performed entirely in the spike domain. With
approximately 7 times lesser number of synaptic weight parameters compared to the state of the art spiking networks,
we show that our approach achieves classification accuracy
exceeding 99% on the training set of the MNIST database and
98.06% on its test set. The trained network implemented on the
CUDA parallel computing platform is also able to successfully
identify digits written by users in real-time, demonstrating its
true generalization capability.
We have also demonstrated a general framework for implementing spike based neural networks and supervised learning
with non-local weight update rules on a GPU platform. At each
time step, the neuronal spike transmission, synaptic current
computation and weight update calculation for the network
are all executed in parallel in this framework. Using this
GPU implementation, we demonstrated a touch-screen based
platform for real-time classification of user-generated images.
R EFERENCES
[1] W. Maass, “Networks of spiking neurons: The third generation of neural
network models,” Neural networks, vol. 10, no. 9, pp. 1659–1671, 1997.
[2] A. Coates et al., “Deep learning with COTS HPC systems,” in 30th
International Conference on Machine Learning, 2013.
[3] A. K. Fidjeland and M. P. Shanahan, “Accelerated simulation of spiking
neural networks using GPUs,” in The 2010 International Joint Conference on Neural Networks (IJCNN). IEEE, 2010, pp. 1–8.
[4] J. M. Nageswaran et al., “A configurable simulation environment for the
efficient simulation of large-scale spiking neural networks on graphics
processors,” Neural networks, vol. 22, no. 5, pp. 791–800, 2009.
[5] E. Yavuz, J. Turner, and T. Nowotny, “GeNN: a code generation
framework for accelerated brain simulations,” Nature Scientific Reports,
vol. 6, no. 18854, 2016.
[6] D. Yudanov et al., “GPU-based simulation of spiking neural networks
with real-time performance and high accuracy,” in International Joint
Conference on Neural Networks, July 2010.
[7] F. Naveros et al., “Event-and time-driven techniques using parallel CPUGPU co-processing for Spiking Neural Networks,” Front. in Neuroinformatics, vol. 11, 2017.
[8] J. L. Krichmar, P. Coussy, and N. Dutt, “Large-scale spiking neural networks using neuromorphic hardware compatible models,” ACM Journal
on Emerging Technologies in Computing Systems (JETC), vol. 11, no. 4,
p. 36, 2015.
[9] P. U. Diehl et al., “Fast-classifying, high-accuracy spiking deep networks
through weight and threshold balancing,” in 2015 International Joint
Conference on Neural Networks (IJCNN), July 2015, pp. 1–8.
[10] Y. Cao, Y. Chen, and D. Khosla, “Spiking deep convolutional neural
networks for energy-efficient object recognition,” International Journal
of Computer Vision, vol. 113, no. 1, pp. 54–66, 2015.
[11] B. Rueckauer et al., “Theory and tools for the conversion of
analog to spiking convolutional neural networks,” arXiv preprint
arXiv:1612.04052, 2016.
[12] E. Hunsberger and C. Eliasmith, “Training Spiking Deep Networks for
Neuromorphic Hardware,” arXiv preprint arXiv:1611.05141, 2016.
[13] N. Anwani and B. Rajendran, “NormAD -Normalized Approximate
Descent based supervised learning rule for spiking neurons,” in International Joint Conference on Neural Networks, July 2015, pp. 1–8.
[14] F. Ponulak and A. Kasinski, “Supervised learning in spiking neural
networks with ReSuMe: sequence learning, classification, and spike
shifting,” Neural Computation, vol. 22, no. 2, pp. 467–510, 2010.
[15] A. Mohemmed et al., “SPAN: Spike Pattern Association Neuron for
Learning Spatio-Temporal Spike Patterns,” International Journal of
Neural Systems, vol. 22, no. 04, 2012.
[16] J. H. Lee, T. Delbruck, and M. Pfeiffer, “Training Deep Spiking Neural
Networks using Backpropagation,” Frontiers in Neuroscience, vol. 10,
p. 508, 2016.
[17] W. W. Lee, S. L. Kukreja, and N. V. Thakor, “CONE: ConvexOptimized-Synaptic Efficacies for Temporally Precise Spike Mapping,”
IEEE Transactions on Neural Networks and Learning Systems, vol. 28,
no. 4, pp. 849–861, April 2017.
[18] “The MNIST database of handwritten digits,” available at
http://yann.lecun.com/exdb/mnist/.
[19] P. Dayan, L. Abbott et al., “Theoretical neuroscience: computational
and mathematical modeling of neural systems,” Journal of Cognitive
Neuroscience, vol. 15, no. 1, pp. 154–155, 2003.
[20] M. Harris, “Optimizing parallel reduction in CUDA,” available at
http://bit.ly/2gd7fSb.
[21] I. Culjak, D. Abram, T. Pribanic, H. Dzapo, and M. Cifrek, “A brief
introduction to OpenCV,” in 2012 Proceedings of the 35th International
Convention MIPRO, May 2012, pp. 1725–1730.
[22] L. Wan et al., “Regularization of Neural Networks using DropConnect,”
in Proceedings of the 30th International Conference on Machine Learning, 2013, pp. 1058–1066.
| 9cs.NE
|
arXiv:1608.01654v2 [cs.PL] 7 Nov 2016
Hypercollecting Semantics
and its Application to Static Analysis of Information Flow
Mounir Assaf
David A. Naumann
Julien Signoles
Stevens Institute of Technology,
Hoboken, US
[email protected]
Stevens Institute of Technology,
Hoboken, US
[email protected]
Software Reliability and Security Lab,
CEA LIST, Saclay, FR
[email protected]
Éric Totel
Frédéric Tronel
CIDRE, CentraleSupélec,
Rennes, FR
[email protected]
CIDRE, CentraleSupélec,
Rennes, FR
[email protected]
Abstract
We show how static analysis for secure information flow can be expressed and proved correct entirely within the framework of abstract
interpretation. The key idea is to define a Galois connection that
directly approximates the hyperproperty of interest. To enable use
of such Galois connections, we introduce a fixpoint characterisation
of hypercollecting semantics, i.e. a “set of sets” transformer. This
makes it possible to systematically derive static analyses for hyperproperties entirely within the calculational framework of abstract
interpretation. We evaluate this technique by deriving example static
analyses. For qualitative information flow, we derive a dependence
analysis similar to the logic of Amtoft and Banerjee (SAS’04) and
the type system of Hunt and Sands (POPL’06). For quantitative information flow, we derive a novel cardinality analysis that bounds the
leakage conveyed by a program instead of simply deciding whether
it exists. This encompasses problems that are hypersafety but not
k-safety. We put the framework to use and introduce variations
that achieve precision rivalling the most recent and precise static
analyses for information flow.
Categories and Subject Descriptors D.2.4 [Software Engineering]: Software/Program Verification–Assertion checkers; D.3 [Programming Languages]; F.3.1 [Logics and meanings of programs]:
Semantics of Programming Language
Keywords static analysis, abstract interpretation, information flow,
hyperproperties
1.
Introduction
Most static analyses tell something about all executions of a program.
This is needed, for example, to validate compiler optimizations.
Functional correctness is also formulated in terms of a predicate on
observable behaviours, i.e. more or less abstract execution traces: A
Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee
provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and
the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored.
Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, or to redistribute to lists, contact
the Owner/Author(s). Request permissions from [email protected] or Publications Dept., ACM, Inc., fax +1 (212)
869-0481.
POPL ’17, January 18 - 20, 2017, Paris, France
Copyright c 2017 held by owner/author(s). Publication rights licensed to ACM.
ACM 978-1-4503-4660-3/17/01. . .$15.00
DOI: http://dx.doi.org/10.1145/http://dx.doi.org/10.1145/3009837.3009889
program is correct if all its traces satisfy the predicate. By contrast
with such trace properties, extensional definitions of dependences
involve more than one trace. To express that the final value of a
variable x may depend only on the initial value of a variable y, the
requirement—known as noninterference in the security literature
(Sabelfeld and Myers 2003)—is that any two traces with the same
initial value for y result in the same final value for x. Sophisticated
information flow policies allow dependences subject to quantitative
bounds—and their formalisations involve more than two traces,
sometimes unboundedly many.
For secure information flow formulated as decision problems, the
theory of hyperproperties classifies the simplest form of noninterference as 2-safety and some quantitative flow properties as hypersafety
properties (Clarkson and Schneider 2010). A number of approaches
have been explored for analysis of dependences, including type systems, program logics, and dependence graphs. Several works have
used abstract interpretation in some way. One approach to 2-safety is
by forming a product program that encodes execution pairs (Barthe
et al. 2004; Terauchi and Aiken 2005; Darvas et al. 2005), thereby
reducing the problem to ordinary safety which can be checked by
abstract interpretation (Kovács et al. 2013) or other means. Alternatively, a 2-safety property can be checked by dedicated analyses
which may rely in part on ordinary abstract interpretations for trace
properties (Amtoft et al. 2006).
The theory of abstract interpretation serves to specify and
guide the design of static analyses. It is well known that effective
application of the theory requires choosing an appropriate notion
of observable behaviour for the property of interest (Cousot 2002;
Bertrane et al. 2012, 2015). Once a notion of “trace” is chosen, one
has a program semantics and “all executions” can be formalized in
terms of collecting semantics, which can be used to define a trace
property of interest, and thus to specify an abstract interpretation
(Cousot and Cousot 1977, 1979; Cousot 1999).
The foundation of abstract interpretation is quite general, based
on Galois connections between semantic domains on which collecting semantics is defined. Clarkson and Schneider (2010) formalize
the notion of hyperproperty in a very general way, as a set of sets
of traces. Remarkably, prior works using abstract interpretation for
secure information flow do not directly address the set-of-sets dimension and instead involve various ad hoc formulations. This paper
presents a new approach of deriving information flow static analyses
within the calculational framework of abstract interpretation.
First contribution. We lift collecting semantics to sets of trace
sets, dubbed hypercollecting semantics, in a fixpoint formulation
which is not simply the lifted direct image. This can be composed
with Galois connections that specify hyperproperties beyond 2safety, without recourse to ad hoc additional notions. On the basis
of this foundational advance, it becomes possible to derive static
analyses entirely within the calculational framework of abstract
interpretation (Cousot and Cousot 1977, 1979; Cousot 1999).
Second contribution. We use hypercollecting semantics to
derive an analysis for ordinary dependences. This can be seen as a
rational reconstruction of both the type system of Hunt and Sands
(2006, 2011) and the logic of Amtoft and Banerjee (2004). They
determine, for each variable x, a conservative approximation of the
variables y whose initial values influence the final value of x.
Third contribution. We derive a novel analysis for quantitative
information flow. This shows the benefit of taking hyperproperties
seriously by means of abstract interpretation. For noninterference,
once the variables y on which x depends have fixed values, there
can be only one final value for x. For quantitative information flow,
one is interested in measuring the extent to which other variables
influence x: for a given range of variation for the “high inputs”,
what is the range of variation for the final values of x? We directly
address this question as a hyperproperty: given a set of traces that
agree only on the low inputs, what is the cardinality of the possible
final values for x? Using the hypercollecting semantics, we derive
a novel cardinality abstraction. We show how it can be used for
analysis of quantitative information problems including a bounding
problem which is not k-safety for any k.
The calculational approach disentangles key design decisions
and it enabled us to identify opportunities for improving precision.
We assess the precision of our analyses and provide a formal
characterisation of precision for a quantitative information flow
analysis vis a vis qualitative. Versions of our analyses rival state of
the art analyses for qualitative and quantitative information flow.
Our technical development uses the simplest programming
language and semantic model in which the ideas can be exposed.
One benefit of working entirely within the framework of abstract
interpretation is that a wide range of semantics and analyses are
already available for rich programming languages.
Outline. Following the background (Section 2), we introduce
domains and Galois connections for hyperproperties (Section 3)
and hypercollecting semantics (Section 4). Hyperproperties for
information flow are defined in Section 5. We use the framework to
derive the static analyses in Section 6 and Section 7. Section 8 uses
examples to evaluate the precision of the analyses, and shows how
existing analyses can be leveraged to improve precision. We discuss
related work (Section 9) and conclude. Appendices provide detailed
proofs for all results, as well as a table of symbols.
2.
Background: Collecting Semantics, Galois
Connections
The formal development uses deterministic imperative programs
over integer variables. Let n range over literal integers Z, x over variables, and ⊕ (resp. cmp) over some arithmetic (resp. comparison)
operators.
c ::= skip | x := e | c1 ; c2 | if b then c1 else c2 | while b do c
e ::= n | x | e1 ⊕ e2 | b
b ::= e1 cmp e2
Different program analyses may consider different semantic
domains as needed to express a given class of program properties.
For imperative programs, the usual domains are based on states
σ ∈ States that map each variable to a value (Winskel 1993). Some
P(States∗ )
P(Trc)
P(States)
Figure 1. Fragment of the hierarchy of semantic domains
abstraction
(−−−−−→)
program properties require the use of traces that include intermediate
states; others can use more abstract domains. For information flow
properties involving intermediate outputs, or restricted to explicit
data flow (Schoepe et al. 2016), details about intermediate steps
are needed. By contrast, bounding the range of variables can be
expressed in terms of final states. As another example, consider
determining which variables are left unchanged: To express this, we
need both initial and final states.
In this paper we use the succinct term trace for elements of
Trc defined by Trc , States × States, interpreting t ∈ Trc as an
initial and final state. In the literature, these are known as relational
traces, by contrast with maximal trace semantics using the set
States∗ of finite sequences. A uniform framework describes the
relationships and correspondences between these and many other
semantic domains using Galois connections (Cousot 2002). Three
of these domains are depicted in Figure 1.
Given partially ordered sets C, A, the monotone functions
α ∈ C → A and γ ∈ A → C comprise a Galois connection,
γ
−
− (A, v), provided they satisfy
a proposition we write (C, ≤) ←
−−
α→
α(c) v a iff c ≤ γ(a) for all c ∈ C, a ∈ A.
For example, to specify an analysis that determines which
variables are never changed, let A be sets of variables. Define
α ∈ P(Trc) → P(Vars) by α(T ) = {x | ∀(σ, σ 0 ) ∈ T, σ(x) =
σ 0 (x)} and γ(X) = {(σ, σ 0 ) | ∀x ∈ X, σ(x) = σ 0 (x)}. Then
γ
−
− (P(V ar), ⊇).
(P(Trc), ⊆) ←
−−
α→
For the hierarchy of usual domains, depicted in Figure 1, the
connections are defined by an “element-wise abstraction”. Define
elt ∈ States∗ → Trc by elt(σ0 σ1 . . . σn ) , (σ0 , σn ). This lifts to
an abstraction P(States∗ ) → P(Trc).
Lemma 1 Element-wise abstraction. Let elt ∈ C → A be a
function between sets. Let αelt (C) , {elt(c) | c ∈ C} and
γelt
−−
−−
− (P(A), ⊆).
γelt (A) , {c | elt(c) ∈ A}. Then (P(C), ⊆) ←
−−
→
α−
elt
The domain P(States), which suffices to describe the final
reachable states of a program, is an abstraction of the relational
domain P(Trc), by elt(σ, τ ) , τ . In this paper we focus on the
domain Trc because it is the simplest that can express dependences.
Program semantics. We define both the denotational semantics
JcK ∈ Trc⊥ → Trc⊥ of commands and the denotational semantics
JeK ∈ Trc → Val of expressions. Here Val , Z and Trc⊥ adds
bottom element ⊥ using the flat ordering.
JcK ∈ Trc⊥ → Trc⊥
Standard semantics of commands
JcK⊥ , ⊥
Jx := eK(σ, τ ) , (σ, τ [x 7→ JeK(σ, τ )])
Jc1 ; c2 Kt , Jc2 K ◦ Jc1 Kt
JskipKt , t
(
Jc1 Kt
Jif b then c1 else c2 Kt ,
Jc2 Kt
if JbKt = 1
if JbKt = 0
Jwhile b do cKt , (lfp4̇
(λt.⊥) F)(t)
(
t
where F(w)(t) ,
w
◦
JcKt
if JbKt = 0
otherwise
Let t be a trace (σ, τ ). The denotation JeKt evaluates e in the
“current state”, τ . (In Sect. 5 we also use JeKpre t which evaluates e
in the initial state, σ.) The denotation JcKt is (σ, τ 0 ) where execution
of c in τ leads to τ 0 . The denotation is ⊥ in case c diverges
from τ . Boolean expressions evaluate to either 0 or 1. We assume
programs do not go wrong. We denote by 4̇ the point-wise lifting
to Trc⊥ → Trc⊥ of the approximation order 4 on Trc⊥ .
The terminating computations of c can be written as its image
on the initial traces: {JcKt | t ∈ IniTrc and JcKt 6= ⊥} where
IniTrc , {(σ, σ) | σ ∈ States}
To specify properties that hold for all executions we use collecting semantics which lifts the denotational semantics to arbitrary
sets T ∈ P(Trc) of traces. The idea is that ⦃c⦄T is the direct image
of JcK on T . To be precise, in this paper we focus on terminationinsensitive properties, and thus ⦃c⦄T is the set of non-⊥ traces t0
such that JcKt = t0 for some t ∈ T . Later we also use the collecting
semantics of expressions: ⦃e⦄T , {JeKt | t ∈ T }.
Importantly, the collecting semantics ⦃c⦄ ∈ P(Trc) → P(Trc)
can be defined compositionally using fixpoints (Cousot 2002, Sec.
7). For conditional guard b, write ⦃ grdb ⦄ for the filter defined by
⦃ grdb ⦄T , {t ∈ T | JbKt = 1}.
⦃c⦄ ∈ P(Trc) → P(Trc)
Collecting semantics
⦃x := e⦄T , {Jx := eKt | t ∈ T }
⦃c1 ; c2 ⦄T , ⦃c2 ⦄ ◦ ⦃c1 ⦄T
⦃skip⦄T , T
⦃if b then c1 else c2 ⦄T , ⦃c1 ⦄ ◦ ⦃ grdb ⦄T ∪ ⦃c2 ⦄ ◦ ⦃ grd¬b ⦄T
⦃while b do c⦄T , ⦃ grd¬b ⦄ lfp⊆
T ⦃if b then c else skip⦄
The clause for while loops uses the denotation of a constructed
conditional command as a definitional shorthand—its denotation is
compositional.
γ
−
− (A, v), such as
Given a Galois connection (P(Trc), ⊆) ←
−−
α→
the one for unmodified variables, the desired analysis is specified as
α ◦ ⦃c⦄ ◦ γ. Since it is not computable in general, we only require
an approximation f ] ∈ A → A that is sound in this sense:
α ◦ ⦃c⦄ ◦ γ v̇ f ]
(1)
where v̇ denotes the point-wise lifting of the partial order v.
To explain the significance of this specification, suppose one
wishes to prove program c satisfies a trace property T ∈ P(Trc),
i.e. to prove that ⦃c⦄(IniTrc) ⊆ T . Given eq. (1) it suffices to find
an abstract value a that approximates IniTrc, i.e. IniTrc ⊆ γ(a),
and show that
γ(f ] (a)) ⊆ T
(2)
]
˙
eq. (1) is equivalent to ⦃c⦄ ◦ γ ⊆ γ ◦ f by a property of
Galois connections. So eq. (2) implies ⦃c⦄(γ(a)) ⊆ T which (by
monotonicity of ⦃c⦄) implies ⦃c⦄(IniTrc) ⊆ ⦃c⦄(γ(a)) ⊆ T .
The beauty of specification eq. (1) is that f ] can be obtained as
an abstract interpretation ⦃c⦄] , derived systematically for all c by
calculating from the left side of eq. (1) as shown by Cousot (1999).
3.
Domains and Galois Connections for
Hyperproperties
To express hyperproperties, we need Galois connections for domains
that involve sets of sets of observable behaviours. This section spells
out how such powerset domains form a hierarchy as illustrated
along the top of Figure 2. We describe how dependences and
P(P(States∗ ))
P(P(Trc))
P(P(States))
P(States∗ )
P(Trc)
P(States)
abstraction
Figure 2. Extended hierarchy of semantic domains (−−−−−→)
cardinalities for quantitative information flow can be formulated
as Galois connections. We spell out a methodology whereby the
standard notions and techniques of abstract interpretation can be
applied to specify and derive—in the same form as Equation (1)—
static analyses for hyperproperties.
As a first example, consider the condition: the final value of x
depends only on the initial value of y. Its expression needs, at least,
two traces: If two traces, denoted by (σ, σ 0 ) and (τ, τ 0 ), agree on
the initial value of y then they agree on the final value of x. That is,
σ(y) = τ (y) implies σ 0 (x) = τ 0 (x). This must hold for any two
traces of the program. This is equivalent to the following: For all
sets T of traces, if traces in T all agree on the initial value of y then
they all agree on the final value of x. Later we extend this example
to an analysis that infers which dependences hold.
Consider the problem of quantifying information flow with mincapacity (Smith 2009). For a program on two integer variables h, l,
the problem is to infer how much information is conveyed via l
about h: considering some traces that agree on the initial value of l,
how many final values are possible for l. For example, the program
l := (h mod 2) + l has two final values for l, for each initial l,
though there are many possible initial values for h. This cardinality
problem generalizes prior work on quantitative flow analysis, where
typically low inputs are not considered.
Whereas the simple dependence problem can be formulated in
terms of 2 traces, the cardinality problem involves trace sets of
unbounded size. In the terminology of hyperproperties, it is not a
k-safety hyperproperty for any k (Yasuoka and Terauchi 2011, Sec.
3), although it is hypersafety (Clarkson and Schneider 2010). For
a fixed k, the problem “variable l has at most k − 1 final values” is
k-safety, which means it can be formulated in terms of sets with at
most k traces.
It turns out that by using Galois connections on sets of sets, we
can develop a general theory that encompasses many hyperproperties
and which enables derivation of interesting abstract interpreters. For
our applications, we use relational traces as the notion of observable
behavior, and thus P(P(Trc)). The approach works as well for other
notions, so there is a hierarchy of domains as shown at the top of Figure 2, in parallel with the ordinary hierarchy shown along the bottom.
The abstractions of this hierarchy are obtained by lifting each
abstraction between two standard collecting semantics (Cousot
2002) to their hypercollecting versions, by element-wise abstraction (Lemma 1). For instance, Lemma 1 justifies the abstraction
between P(P(Trc)) and P(P(States)), by lifting the abstraction
between P(Trc) and P(States) (Cousot 2002, Sec. 8). Additionally, the diagonal lines in Figure 2 represent abstractions between
hypercollecting semantics defined over some form of observations
and the corresponding collecting semantics defined over the same
observations.
Lemma 2 . Let C be a set. Define αhpp (C) , ∪C∈C C and
γhpp (C) , P(C). These form a Galois connection:
γhpp
−−
−−−
− (P(C), ⊆)
(P(P(C)), ⊆) ←
−−
→
α −−
hpp
It is noted by Clarkson and Schneider (2010) that any trace
property can be lifted to a unique hyperproperty; this lifting is
exactly the concretisation γhpp of Lemma 2. Although the model
of Clarkson and Schneider (2010) is quite general, it does focus on
infinite traces. But hyperproperties can be formulated in terms of
other notions of observation, as illustrated in Figure 2.
Cardinality abstraction. To lay the groundwork for our quantitative information flow analysis, we consider abstracting a set of
values by its cardinality. Cardinality is one ingredient in many quantitative information flow analyses estimating the amount of sensitive
information a program may leak (Smith 2009; Backes et al. 2009;
Braun et al. 2009; Köpf and Rybalchenko 2013; Mardziel et al.
2013; Doychev et al. 2013). The lattice of abstract representations
we consider is the set
[0, ∞] , N ∪ {∞}
where ∞ denotes an infinite cardinal number. We use the natural order ≤, and max as a join. Consider the abstraction operator crdval ∈ P(Val) → [0, ∞] computing cardinality and
given by crdval(V ) , |V |. This operator crdval is not additive, i.e. it does not preserve joins; e.g. crdval({1, 2} ∪ {2, 3}) 6=
max(crdval({1, 2}), crdval({2, 3})). Thus, there exists no associated concretisation f for which crdval is the lower adjoint in a
Galois connection. Yet, we can lift the abstraction operator crdval
to a Galois connection over P(P(Val)) through what is called a
supremus abstraction (Cousot 2002, p.52).
Lemma 3 Supremus abstraction. Let elt ∈ C → A be a function
from a set C, with codomain forming a complete lattice (A, v). Let
αelt (C) , tc∈C elt(c) and γelt (a) , {c ∈ C | elt(c) v a}. Then
γelt
−−
−−
− (A, v)
(P(C), ⊆) ←
−−
→
α−
elt
For example, define αcrdval (V) , maxV ∈V crdval(V ) and
γcrdval (n) , {V | crdval(V ) ≤ n}. Thus we obtain a Galois
γcrdval
←
−−
−−
−−
−−
−
− ([0, ∞] , ≤).
connection (P(P(Val)), ⊆) −
−
→
αcrdval
As another example let us consider, in simplified form, an
ingredient in dependency or noninterference analysis. For program
variable x, agreex ∈ P(States) → {tt, ff} determines whether a
set of states contains only states that all agree on x’s value:
agreex (Σ) , (∀σ, σ 0 ∈ Σ, JxKσ = JxKσ 0 )
Function agreex is not additive, so it is not part of a Galois
connection from P(States) to {tt, ff}. The same problem arises
with agreements on multiple variables, and with more concrete
domains like the finite maximal trace semantics P(States∗ ).
We lift the operator agreex to a Galois connection over
P(P(States)). A supremus abstraction yields
αagreex (S) , (∀Σ ∈ S, agreex (Σ))
γagreex (bv) , {Σ | agreex (Σ) ⇐= bv}
abstract concretisation/denotation of a security requirement yields a
set of sets of traces, namely a hyperproperty. Hints of this intuition
appear in the literature (McLean 1994; Volpano 1999; Rushby 2001;
Zakinthinos and Lerner 1997); e.g. security policies “are predicates
on sets of traces (i.e. they are higher order)” (Rushby 2001, p.2).
However, only recently has a comprehensive framework proposed a
sharp characterisation of security policies as hyperproperties (Clarkson and Schneider 2008, 2010).
Abstract interpretation of hyperproperties. The basic methodology for the verification of a hyperproperty HP, may be described
as follows:
Step 1. Design approximate representations forming a complete
lattice A, choose a collecting semantics C among the extended
hierarchy (set of sets domains, e.g. P(P(Trc))), and define α, γ
γ
−
− (A, v).
for a Galois connection (C, ≤) ←
−−
α→
Step 2. Compute an approximation a ∈ A of the semantics C ∈ C
of the program P of interest.
Step 3. Prove that the inferred approximation a implies that P
satisfies HP. The concretisation γ(a) is a set of trace sets,
of which the program’s trace set is a member—by contrast to
approximations of trace properties, which infer a single trace set
of which the program trace set is a subset. Then, it suffices to
prove γ(a) ⊆ HP.
Step 1 is guided by the need to have γ(a) ⊆ HP, i.e. a describes
a hyperproperty that implies HP. The calculational design (Cousot
1999) of abstract domains greatly systematises Step 2, by relying on
the Galois connection defined in Step 1. Collecting semantics can be
adapted to the additional structure of sets, as we show in Section 4.
4.
Hypercollecting Semantics
In the following, we introduce a hypercollecting semantics defined
over sets T ∈ P(P(Trc)) of sets of traces. This is used in
subsequent sections to derive static analyses.
Here is Step 2 of the methodology, spelled out in detail. Given
γ
−
− (A, v] ) built by the
a Galois connection (P(P(Trc)), ⊆) ←
−−
α→
supremus abstraction, and an approximation a of the initial traces
(i.e. IniTrc is in γ(a)), find an approximation a0 ∈ A of the
analysed program c, i.e. ⦃c⦄ IniTrc is in γ(a0 ). Then prove that the
program satisfies the hyperproperty HP of interest, i.e. γ(a0 ) ⊆ HP.
In order to compute a0 , we define a hypercollecting semantics
LcM ∈ P(P(Trc)) → P(P(Trc)). That will serve to derive—in
the manner of Equation (1)—a static analysis that is correct by
construction.
Hypercollecting semantics
γagreex
−−
−−
−−
−−
−
− ({tt, ff}, ⇐=).
so that (P(P(States)), ⊆) ←
−−
→
αagree
x
These examples are consistent with the many formulations of
noninterference (e.g. (Goguen and Meseguer 1982; Volpano and
Smith 1997; Giacobazzi and Mastroeni 2004; Amtoft and Banerjee
2004; Hunt and Sands 2006)) that motivated the characterisation of
information-flow security requirements as hyperproperties (Clarkson and Schneider 2010). Concretising an abstract value a can
be seen as defining the denotation of a type expression (as in, for
instance, Benton (2004, Sec. 3.3.1) and Hunt and Sands (1991)),
i.e. defining the set of objects that satisfy the description a. Thus,
concretising tt, when tt is interpreted as “satisfies a property requirement”, naturally yields a set of traces. Concretising tt, where
tt is interpreted as “satisfies a security requirement”, yields a set of
sets of traces.
Intuitively, the most abstract denotation/concretisation of a property requirement is defined in terms of a set of traces. The most
L c M ∈ P(P(Trc)) → P(P(Trc))
Lx := eMT , {⦃x := e⦄T | T ∈ T}
Lc1 ; c2 MT , Lc2 M ◦ Lc1 MT
LskipMT , T
Lif b then c1 else c2 MT ,
{⦃c1 ⦄ ◦ ⦃ grdb ⦄T ∪ ⦃c2 ⦄ ◦ ⦃ grd¬b ⦄T | T ∈ T}
Lwhile b do c MT , Lgrd¬b M lfp⊆
T Lif b then c else skipM
Lgrdb MT , {⦃ grdb ⦄T | T ∈ T}
Recall from Section 2 that standard collecting semantics is a
fixpoint-based formulation that captures the direct image on sets
of the underlying program semantics – this is proved, for example,
by Cachera and Pichardie (2010); Assaf and Naumann (2016). The
fixpoint formulation at the level of sets-of-sets we use is not simply
the direct image of the standard collecting semantics. The direct
image of the standard collecting semantics would yield a set of
(inner) fixpoints over sets of traces, whereas an outer fixpoint over
sets of sets of traces enables straightforward application of the
fixpoint transfer theorem.
Theorem 1 . For all c and all T ∈ P(Trc), ⦃c⦄T is in L c M{T }.
For a singleton {T }, the set LcM{T } ∈ P(P(Trc)) is not
necessarily a singleton set containing only the element ⦃c⦄T . If
c is a loop, LcM{T } yields a set of sets R of traces, where each set
R of traces contains only traces that exit the loop after less than
k iterations, for k ∈ N. We prove this theorem as corollary of the
following:
∀T ∈ P(P(Trc)), {⦃c⦄T | T ∈ T} ⊆ LcMT
This is proved by structural induction on commands. For loops, there
is a secondary induction on iterations of the loop body.
In summary, suppose one wishes to prove program c satisfies
hyperproperty HP ∈ P(P(Trc)), i.e. one wishes to prove that
⦃c⦄(IniTrc) ∈ HP. Suppose we have an approximation f ] of the
hypercollecting semantics, similarly to eq. (1), i.e.
]
α ◦ LcM ◦ γ v̇ f
]
(3)
Given eq. (3) it suffices to find an abstract value a that approximates
IniTrc, i.e. IniTrc ∈ γ(a), and show that:
γ(f ] (a)) ⊆ HP
(4)
˙ γ ◦ f ] by a property
Why? Equation (3) is equivalent to LcM γ ⊆
of Galois connections. So we have ⦃c⦄(IniTrc) ∈ LcM(γ(a)) ⊆
γ(f ] (a)) ⊆ HP using IniTrc ∈ γ(a), the Theorem, and eq. (4).
◦
5.
Information Flow
This section gives a number of technical definitions which build
up to the definition of Galois connections with which we specify
information flow policies explicitly as hyperproperties.
When a fixed main program is considered, we refer to it as P
and its variables as VarP . Our analyses are parametrised by the
program P to analyse, and an initial typing context Γ ∈ VarP → L
mapping each variable to a security level l ∈ L for its initial value.
We assume (L, v, t, u) is a finite lattice. In the most concrete case,
L may be defined as the universal flow lattice, i.e. the powerset of
variables P(VarP ), from which all other information flow types can
be inferred through a suitable abstraction (Hunt and Sands 2006,
Sec. 6.2); the initial typing context is then defined as λx.{x}.
Initial l-equivalence and variety. A key notion in information
flow is l-equivalence. Two states are l-equivalent iff they agree on
the values of variables having security level at most l. We introduce
the same notion over a set of traces, requiring that the initial states
are l-equivalent. Let us first denote by JeKpre ∈ Trc → Val
the evaluation of expression e in the initial state σ of a trace
(σ, τ ) ∈ Trc—unlike JeK ∈ Trc → Val which evaluates expression
e in the final state τ . Then, we denote by T |=Γ l the judgement
that all traces in a set T ⊆ Trc are initially l-equivalent, i.e. they all
initially agree on the value of variables up to a security level l ∈ L.
For example, in the case that L is the universal flow lattice, T |=Γ
{x, y} means ∀t1 , t2 ∈ T, JxKpre t1 = JxKpre t2 ∧ JyKpre t1 =
JyKpre t2 .
Initial l-equivalence
T |=Γ l iff. ∀t1 , t2 ∈ T, ∀x ∈ VarP ,
Γ(x) v l =⇒ JxKpre t1 = JxKpre t2
T |=Γ l
The notion of variety (Cohen 1977) underlies most definitions of
qualitative and quantitative information flow security. Information is
transmitted from a to b over execution of program P if by “varying
the initial value of a (exploring the variety in a), the resulting
value in b after P’s execution will also vary (showing that variety is
conveyed to b)” (Cohen 1977). We define the l-variety of expression
e, as the set of sets of values e may take, when considering only
initially l-equivalent traces. The variety is defined first as a function
Ol ⦃e⦄ ∈ P(Trc) → P(P(Val)) on trace sets, from which we
obtain a function Ol LeM ∈ P(P(Trc)) → P(P(Val)), on sets of
trace sets. Intuitively, l-variety of expression e is the variety that is
conveyed to e by varying only the input values of variables having a
security level l0 such that ¬(l0 v l).
Ol ⦃e⦄
l-variety
Ol LeM
Ol ⦃e⦄ ∈ P(Trc) → P(P(Val))
Ol ⦃e⦄T , {⦃e⦄R | R ⊆ T and R |=Γ l}
Ol LeM ∈ P(P(Trc)) → P(P(Val))
Ol LeMT , ∪T ∈T Ol ⦃e⦄T
Each set V ∈ Ol ⦃e⦄T of values results from initially lequivalent traces (R |=Γ l for R ⊆ T ). Thus, expression e does not
leak sensitive information to attackers having a security clearance
l ∈ L if Ol ⦃e⦄T is a set of singleton sets. Indeed, sensitive data
for attackers with security clearance l ∈ L is all data having
a security level l0 for which attackers do not have access (i.e.
¬(l0 v l) (Denning and Denning 1977)). Thus, if Ol ⦃e⦄T is a set of
singleton sets, this means that no matter how sensitive information
varies, this variety is not conveyed to expression e.
Besides a pedagogical purpose, we define l-variety Ol ⦃e⦄ (resp.
Ol LeM) instead of simply lifting the denotational semantics JeK of
expressions to sets of traces (resp. sets of sets of traces) since we
want to build modular abstractions of traces by relying on underlying
abstractions of values. Thus, l-variety enables us to pass information
about initially l-equivalent traces to the underlying domain of values
by keeping disjoint values that originate from traces that are not
initially l-equivalent.
Specifying information flow. We now have the ingredients needed
to describe information flow for command c, with respect to typing context Γ ∈ VarP → L. A quantitative security metric, introduced by Smith (2009, 2011), relies on min-entropy and mincapacity (Rényi 1961) in order to estimate the leakage of a program. Let us assume a program P that is characterized by a set
TP ∈ P(Trc) of traces, i.e. TP , ⦃ P ⦄ IniTrc. For simplicity, assume attackers only observe the value of a single variable x ∈ VarP .
(The generalization to multiple variables is straightforward.) The
leakage of P, as measured by min-capacity, to attackers having
security clearance l ∈ L is defined by
MLl , log2
◦
αcrdval
◦
Ol ⦃x⦄TP
(The definition of αcrdval follows Lemma 3.) For our purposes,
it suffices to know that this quantity aims to measure, in bits, the
remaining uncertainty about sensitive data for attackers with security
clearance l. Refer to the original work (Smith 2009) for more details.
Leaving aside the logarithm in the definition of MLl , a quantitative security requirement may enforce a limit on the amount of
information leaked to attackers with security clearance l ∈ L, by
requiring that the l-cardinality of variable x is less than or equal to
some non-negative integer k. We denote by SR(l, k, x) the hyperproperty that characterises this security requirement, i.e. the set of
program denotations satisfying it:
SR(l, k, x) , {T ∈ P(Trc) | αcrdval
◦
Ol ⦃x⦄T ≤ k}
Note that SR implicitly depends on the choice of initial typing Γ, as
does Ol ⦃x⦄T .
The termination-insensitive noninterference policy “the final
value of x depends only on the initial values of variables labelled
at most l” corresponds to the hyperproperty SR(l, 1, x). Therefore,
the program P satisfies SR(l, 1, x) if αcrdval ◦ Ol ⦃x⦄TP ≤ 1. Let
T = LPM{IniTrc}. Since TP is in T (Theorem 1), then P satisfies
SR(l, 1, x) if αcrdval ◦ Ol LxMT ≤ 1, by monotony of αcrdval and
by Ol ⦃x⦄TP ⊆ Ol LxMT from the definition of Ol L−M.
6.
Dependences
We rely on abstract interpretation to derive a static analysis similar
to existing ones inferring dependences (Amtoft and Banerjee 2004;
Hunt and Sands 2006; Amtoft et al. 2006; Hunt and Sands 2011).
Recall that our analyses are parametrised on a security lattice
L and program P. We denote by l ; x an atomic dependence
constraint, with l ∈ L and x ∈ VarP , read as “agreement up to
security level l leads to agreement on x”. It is an atomic pre-post
contract expressing that the final value of x must only depend on
initial values having at most security level l. Said otherwise, l ; x
states the noninterference of variable x from data that is sensitive
for attackers with security clearance l, i.e. all inputs having security
level l0 such that ¬(l0 v l).
Dependences are similar to information flow types (Hunt and
Sands 2006) and are the dual of independences assertions (Amtoft
and Banerjee 2004). Both interpretations are equivalent (Hunt and
Sands 2006, Sec. 5).
D ∈ Dep
Dep
Lattice of dependence constraints
Given a lattice L and program P, define
Dep , P({l ; x | l ∈ L, x ∈ VarP })
D 1 v\ D 2 , D 1 ⊇ D 2
agree
agree(V )
∈
,
αagree
αagree (V)
∈
,
γagree
γagree (bv)
∈
,
αagree
deptr
deptr(T )
∈
,
αdeptr
αdeptr (T)
∈
,
γdeptr
γdeptr (D)
∈
,
γagree
P(Val) → {tt, ff}
(∀v1 , v2 ∈ V, v1 = v2 )
P(P(Val)) → {tt, ff}
∧V ∈V agree(V )
{tt, ff} → P(P(Val))
{V ∈ P(Val) | agree(V ) ⇐= bv}
αdeptr
γdeptr
P(Trc) → Dep
{l ; x | l ∈ L, x ∈ VarP , αagree (Ol ⦃x⦄T )}
P(P(Trc)) → Dep
t\ T ∈T deptr(T )
Dep → P(P(Trc))
{T | deptr(T ) v\ D}
γdeptr
−−
−−−−
− (Dep, v\ )
(P(P(Trc)), ⊆) ←
−
−
→
α −−−
deptr
Note that deptr(T ) is the set of dependences l ; x for which
αagree (Ol ⦃x⦄T ) holds. For instance, the initial typing context
Γ ∈ VarP → L determines the initial dependences of a program:
αdeptr ({IniTrc})
= {l ; x | l ∈ L, x ∈ VarP and αagree (Ol ⦃x⦄ IniTrc)}
= {l ; x | l ∈ L, x ∈ VarP and Γ(x) v l}
l
We derive an approximation OD
LeM\ of l-variety Ol LeM. This
l
\
approximation OD LeM ∈ Dep → {tt, ff}, called l-agreement of
expression e, determines whether a set D of dependence constraints
guarantees that no variety is conveyed to expression e when the
inputs up to security level l are fixed. Notice that we use symbol \
and subscript D here, for contrast with similar notation using ] and
subscript C in later sections.
l
OD
LeM\ ∈ Dep → {tt, ff}
l
OD
LnM\ D , tt
In the rest of this section, L and P are fixed, together with a
typing context Γ ∈ VarP → L.
The semantic characterisation of dependences is tightly linked
to variety. An atomic constraint l ; x holds if no variety is
conveyed to x when the inputs up to security level l are fixed.
We use this intuition to define the Galois connections linking the
hypercollecting semantics and the lattice Dep, by instantiating the
supremus abstraction in Lemma 3.
The agreement abstraction approximates a set V ∈ P(P(Val))
by determining whether it contains variety.
agree
deptr
Dependence abstraction
l-agreement of expressions
D1 t\ D2 , D1 ∩ D2
Agreements abstraction
of level at most l. So αagree (Ol ⦃x⦄T ) holds just if there is at most
one final value.
l
OD
LxM\ D , (l ; x ∈ D)
l
l
l
OD
Le1 ⊕ e2 M\ D , OD
Le1 M\ D ∧ OD
Le2 M\ D
l
l
l
OD
Le1 cmp e2 M\ D , OD
Le1 M\ D ∧ OD
Le2 M\ D
l
Deriving the clauses defining OD
L−M\ amounts to a constructive
proof of the following.
l
Lemma 4 . OD
LeM\ is sound:
∀e, ∀l, ∀D,
αagree
◦
l
Ol LeM ◦ γdeptr (D) ⇐= OD
LeM\ D .
Dependence abstract semantics. We derive a dependence abstract
semantics LcM\ by approximating the hypercollecting semantics LcM.
This abstract semantics LcM\ ∈ Dep → Dep over-approximates the
dependence constraints that hold after execution of a command c,
on inputs satisfying initial dependence constraints.
We assume a static analysis approximating the variables that a
command modifies.
Mod ∈ Com → P(V ar)
Modifiable variables
−−−−− ({tt, ff}, ⇐=)
(P(P(Val)), ⊆) ←
−−
α−−−→
For all c, x, if there exists t, t ∈ Trc such that JcKt = t0 and
JxKpre t0 6= JxKt0 , then x ∈ Mod(c).
Note that γagree (tt) is {V ∈ P(Val) | agree(V )} and γagree (ff) is
P(Val). Also, agree(V ) iff |V | ≤ 1.
The dependence abstraction approximates a set T ∈ P(P(Trc))
by a dependence constraint D ∈ Dep. Recall that Ol ⦃x⦄T is the
set of final values for variable x in traces t ∈ T that agree on inputs
The abstract semantics of assignments x := e discards all atomic
constraints related to variable x in the input set D of constraints,
and adds atomic constraints l ; x if D guarantees l-agreement
for expression e. For conditionals, for each security level l, if
the input set D guarantees l-agreement of the conditional guard,
γagree
agree
0
the abstract semantics computes the join over the dependences of
both conditional branches, after projecting to only those atomic
constraints related to l (notation π l (−)). If D does not guarantee
l-agreement of the conditional guard, atomic constraints related to
both l and variables possibly modified are discarded. Intuitively, if D
guarantees l-agreement of the conditional guard, then l-agreement
over some variable x in both branches guarantees l-agreement over
x after the conditional command. Otherwise, the only l-agreements
that are guaranteed after the conditional are those that hold before
the conditional for variables that are not modified.
LcM\ ∈ Dep → Dep
Dependence abstract semantics
LskipM\ D , D
Comparison with previous analyses. Our dependence analysis is
similar to the logic of Amtoft and Banerjee (2004) as well as the flowsensitive type system of Hunt and Sands (2006). The relationship
between our sets D ∈ Dep of dependence constraints and the type
environments ∆ ∈ VarP → L of Hunt and Sands can be formalised
by the abstraction:
Lc1 ; c2 M\ D , Lc2 M\
◦
Lc1 M\ D
Lx := eM\ D ,
l
{l ; y ∈ D | y 6= x} ∪ {l ; x | l ∈ L, OD
LeM\ D}
Lif b then c1 else c2 M\ D ,
let D1 = Lc1 M\ D in
let D2 = Lc2 M\ D in
let W (
= Mod(if b then c1 else c2 ) in
l
S π l (D1 ) t\ π l (D2 )
LbM\ D
if OD
l
/ W } otherwise
l∈L {l ; x ∈ π (D) | x ∈
αhs
αhs (D)
∈
,
γhs
γhs (∆)
∈
,
Dep → VarP → L
λx. u {l | l ; x ∈ D}
(VarP → L) → Dep
{l ; x | x ∈ VarP , l ∈ L, ∆(x) v l}
This is in fact an isomorphism because of the way we interpret
dependences. Indeed, if l ; x holds, then also l0 ; x for all l0 ∈ L
such that l v l0 (cf. Corollary 4 in Appendix G.2). This observation
suggests reformulating the sets D ∈ Dep of dependence constraints
to contain only elements with minimal level, but we refrain from
doing so for simplicity of presentation.
Our dependence analysis is at least as precise as the type system
of Hunt and Sands. To state this result, we denote by ⊥L the bottom
element of the lattice L. We also assume that the modified variables
is precise enough to simulate the same effect as the program counter
used in the type system: Mod(c) is a subset of the variables that are
targets of assignments in c.
Theorem 3 . For all c, D0 , D ∈ Dep, ∆0 , ∆ ∈ VarP → L, where
⊥L ` ∆0 {c}∆, and D = LcM\ D0 , it holds that:
]
\
Lwhile b do cM\ D , lfpv
D Lif b then c1 else c2 M
π l (D) , {l ; x ∈ D | x ∈ VarP }
αhs (D0 ) v̇ ∆0 =⇒ αhs (D) v̇ ∆ .
Theorem 2 . The dependence semantics is sound:
αdeptr
\
◦
\
LcM ◦ γdeptr v̇ LcM\ .
7.
We denote by v̇ the point-wise lifting of the partial order v\ .
We can derive this abstract semantics by directly approximating the
relational hypercollecting semantics LcM through the dependence
Galois connection (αdeptr , γdeptr ). The derivation is by structural
induction on commands. It leverages mathematical properties of
Galois connections. We start with the specification of the best
abstract transformer αdeptr ◦ LcM ◦ γdeptr ∈ Dep → Dep, and
successively approximate it to finally obtain the definition of the
dependence abstract semantics for each form of command. The
derivation is the proof, and the obtained definition of the abstract
semantics is correct by construction.
Let us showcase the simplest derivation for a sequence of
commands in order to illustrate this process:
αdeptr ◦ Lc1 ; c2 M ◦ γdeptr
= HBy definition of the hypercollecting semanticsI
αdeptr
◦
\
Lc2 M ◦ Lc1 M ◦ γdeptr
v̇ HBy γdeptr
αdeptr
◦
\
◦
αdeptr is extensive I
Lc2 M ◦ γdeptr
◦
αdeptr
◦
Lc1 M ◦ γdeptr
v̇ HBy induction hypothesis αdeptr
Lc2 M\
◦
Lc1 M\
◦
Lattice of cardinality constraints
Card
C ∈ Card
For a program P and lattice L, we say C is a valid set of constraints
iff ∀x ∈ VarP , ∀l ∈ L, ∃!n ∈ [0, ∞] , l ; x#n ∈ C .
Let Card be the set of valid sets of constraints.
It is a complete lattice:
C1 v] C2 iff ∀l ; x#n1 ∈ C1 , ∃n2 ,
l ; x#n2 ∈ C2 ∧ n1 ≤ n2
C1 t] C2 , {l ; x# max(n1 , n2 ) |
l ; x#n1 ∈ C1 , l ; x#n2 ∈ C2 }
\
LcM ◦ γdeptr v̇ LcM\ I
, HTake this last approximation as the definition.I
Lc1 ; c2 M\
Cardinality Abstraction
Dependence analysis is only concerned with whether variety is conveyed. We refine this analysis by deriving a cardinality abstraction
that enumerates variety.
We denote by l ; x#n an atomic cardinality constraint where
l ∈ L, x ∈ VarP and n ∈ [0, ∞], read as “agreement up to security
level l leads to a variety of at most n values in variable x”.
Alternatively, we can leverage Galois connections to give the
analysis as an approximation of the cardinality analysis. We work
this out by Lemmas 6 and 7, introduced in Section 7.
In the rest of this section, L and P are fixed, together with a
typing context Γ ∈ VarP → L.
A valid constraint set is essentially a function from l and x to n.
So v] is essentially a pointwise order on functions, and we ensure
that v] is antisymmetric.
The cardinality abstraction relies on the abstraction αcrdval ,
introduced in Section 3, in order to approximate l-variety of a
variable into a cardinality n ∈ [0, ∞].
crdtr
Cardinality abstraction
crdtr
crdtr(T )
∈
,
αcrdtr
αcrdtr (T)
∈
,
γcrdtr
γcrdtr (C )
∈
,
αcrdtr
γcrdtr
P(Trc) → Card
{l ; x#n | l ∈ L, x ∈ VarP ,
n = αcrdval (Ol ⦃x⦄T ) }
P(P(Trc)) → Card
t] T ∈T crdtr(T )
Card → P(P(Trc))
{T | crdtr(T ) v] C }
Theorem 4 . The cardinality abstract semantics is sound:
γcrdtr
−−
−−−− (Card, v )
(P(P(Trc)), ⊆) ←
−−
α −−→
]
crdtr
The cardinality abstraction enables us to derive an approximation
l
l
OC
LeM] of l-variety Ol LeM. This approximation OC
LeM] ∈ Card →
[0, ∞], called l-cardinality of expression e, enumerates the l-variety
conveyed to expression e assuming a set C ∈ Card of cardinality
constraints holds. Note that the infinite cardinal ∞ is absorbing, i.e.
∀n, ∞ × n , ∞.
l
OC
LeM] ∈ Card → [0, ∞]
l-cardinality of expressions
l
OC
LnM] C , 1
l
OC
LxM] C , n where l ; x#n ∈ C
l
l
l
OC
Le1 ⊕ e2 M] C , OC
Le1 M] C × OC
Le2 M] C
l
l
l
OC
Le1 cmp e2 M] C , min 2, OC
Le1 M] C × OC
Le2 M] C
l
LeM] is sound:
Lemma 5 . OC
∀e, ∀l,
αcrdval
◦
l
˙ OC
Ol LeM ◦ γcrdtr ≤
LeM] .
We now derive a cardinality abstract semantics by approximating the relational hypercollecting semantics of Section 4. It uses
definitions to follow.
Cardinality abstract semantics
LskipM] C , C
abstract semantics of conditionals is also similar to dependences:
if the conditional guard does not convey l-variety, then all initially
l-equivalent traces follow the same execution path and the join
operator (defined as max over cardinality) over both conditional
branches over-approximates the l-cardinality after the conditional.
Otherwise, the l-cardinality over both conditional branches have
to be summed—for the variables that may be modified in the
conditional branches—to soundly approximate the l-cardinality after
the conditional.
LcM] ∈ Card → Card
Lc1 ; c2 M] C , Lc2 M]
◦
Lc1 M] C
Lx := eM] C ,
{l ; y#n ∈ C | y 6= x}
l
∪{l ; x#n | l ∈ L, x ∈ VarP , n = OC
LeM] C }
Lif b then c1 else c2 M] C ,
let C1 = Lc1 M] C in
let C2 = Lc2 M] C in
let W (
= Mod(if b then c1 else c2 ) in
l
S π l (C1 ) t] π l (C2 )
if OC
LbM] C = 1
l
]
l
l∈L π (C1 ) t add(W,π l (C )) π (C2 ) otherwise
]
]
Lwhile b do cM] C , lfpv
C Lif b then c1 else c2 M
π l (C ) , {l ; x#n ∈ CS | x ∈ VarP , n ∈ [0, ∞]}
C1 t] add(W,C0 ) C2 ,
x∈Var
S P \W {l ; x#n ∈ C0 }
∪ x∈W {l ; x#(n1 +n2 ) |
l ; x#nj ∈ Cj , j = 1, 2}
The abstract semantics of assignments x := e is similar in spirit
to the one for dependences: discard atomic constraints related to x,
and add new ones by computing l-cardinality of expression e. The
αcrdtr
◦
]
LcM ◦ γcrdtr v̇ LcM] .
The lattice Card is complete, although not finite. We may
define a widening operator ∇ ∈ Card × Card → Card to ensure
convergence of the analysis (Cousot and Cousot 1992)(Nielson et al.
1999)(Cortesi and Zanioli 2011, Sec. 4).
C1 ∇C2 , {l ; x#n | l ; x#n1 ∈ C1 , l ; x#n2 ∈ C2 ,
n = n1 ∇n2 }
n1 ∇n2 , if (n2 ≤ n1 ) then n1 else ∞
The occurrence of widening depends on the iteration strategy
employed by the static analyser. Widening accelerates or forces
the convergence of fixpoint computations. In the simplest setting,
the analyser passes as arguments to the widening operator the old
set C1 of cardinality as well as the new set C2 that is computed.
For each atomic cardinality constraint, the widening operator then
compares the old cardinality n1 to the new cardinality n2 . If the
cardinality is still strictly increasing (n2 > n1 ), the widening forces
the convergence by setting it to ∞. If the cardinality is decreasing,
the widening operator sets it to the maximum cardinality n1 in
order to force convergence and ensure the sequence of computed
cardinalities is stationary.
Min-capacity leakage. So far, we showed how one can derive
static analyses of hyperproperties—the abstract representations
themselves are interpreted as hyperproperties—by approximating
hypercollecting semantics. Let us now recall the security requirement SR(l, k, x) introduced in Section 4 in order to illustrate how
these analyses may prove that a program satisfies a hyperproperty,
i.e. Step 3 of the methodology in Section 3 (see also Equation (4)).
Consider a program P characterised by a set TP ∈ P(Trc) of
traces, i.e. TP is ⦃ P ⦄ IniTrc. How do we prove that P satisfies the
hyperproperty SR(l, k, x)? We can use the cardinality analysis to
prove that variable x has a l-cardinality that is at most k. Indeed,
if C approximates TP (i.e. αcrdtr ({TP }) v] C ) then αcrdval ◦
l
Ol ⦃x⦄TP ≤ OC
LxM] C . Thus, if the inferred l-cardinality of C is
at most k then program P is guaranteed to satisfy the hyperproperty
SR(l, k, x). We have {TP } ⊆ γcrdtr (C ) since C approximates TP
(i.e. αcrdtr ({TP }) v] C ). And we have γcrdtr (C ) ⊆ SR(l, k, x)
l
by assumption OC
LxM] C ≤ k. Hence TP ∈ SR(l, k, x).
The hyperproperty SR(l, k, x) is a (k + 1)-safety hyperproperty (Clarkson and Schneider 2010), i.e. it requires exhibiting
at most k + 1 traces in order to prove that a program does not
satisfy SR(l, k, x). For example, termination-insensitive noninterference for security level l, which corresponds to the hyperproperty
SR(l, 1, x), is 2-safety. A k-safety hyperproperty of a program can
be reduced to a safety property of a k-fold product program (Barthe
et al. 2004; Terauchi and Aiken 2005; Darvas et al. 2005; Clarkson
and Schneider 2010).
Various quantitative information flow properties are not k-safety.
For example, the bounding problem that the cardinality analysis
targets, namely min-capacity leakage is not a k-safety hyperproperty
for any k (Yasuoka and Terauchi 2011, Sec. 3). Instead, this
bounding problem is hypersafety (Clarkson and Schneider 2010).
Cardinalities vs. dependences. Just as quantitative security metrics are the natural generalisations of qualitative metrics such as noninterference, the cardinality abstraction is a natural generalisation of
dependence analysis. Instead of deciding if variety is conveyed, the
cardinality analysis enumerates this variety. In other words, dependences are abstractions of cardinalities. We can factor the Galois connections, e.g. (αagree , γagree ) is (αlqone ◦ αcrdval , γcrdval ◦ γlqone )
for suitable (αlqone , γlqone ).
improve precision. For simplicity, we consider a two point lattice
{L, H} and an initial typing context where variables yi are the
only low variables (Γ(yi ) = L). As is usual, low may flow to high
(L v H).
Consider the following program.
1
2
3
Lemma 6 . (αagree , γagree ) is the composition of two Galois connections (αcrdval , γcrdval ) and (αlqone , γlqone ) :
4
γlqone
γ
Listing 1. Leaking 1 bit of secret
crdval
−−
−
−−−
− ([0, ∞] , ≤) ←
−−−−− ({tt, ff}, ⇐=)
(P(P(Val)), ⊆) ←
−−
→
−−
α −−−
α−−−→
crdval
lqone
The cardinality abstraction determines that x has at most 2 values after the execution of the program in Listing 1, for initially
L-equivalent traces. For fixed low inputs, x has one value in the then
branch and one value in the else branch, and these cardinalities get
summed after the conditional since the conditional guard may evaluate to 2 different values. Thus, the cardinality abstraction proves
that this example program satisfies the hyperproperty SR(L, 2, x).
with:
(
tt
ff
(
1
γlqone (bv) ,
∞
αlqone (n) ,
if n ≤ 1
, and
otherwise.
if bv = tt
otherwise.
Stronger trace properties. Another way of proving a hyperproperty is by proving a stronger trace property. If a program is proven
to satisfy a trace property T ∈ P(Trc), then proving that T is
stronger than hyperproperty H ∈ P(P(Trc))—in the sense that
γhpp (T ) ⊆ H—guarantees the program satisfies the hyperproperty H. For instance, by proving for some program that an output
variable x ranges over an interval of integer values whose size is k,
we can prove that program satisfies SR(L, k, x).
However, approximating a hyperproperty by a trace property
may be too coarse for some programs, as we can illustrate with an
interval analysis (Cousot and Cousot 1977) on the example program
in Listing 1. Such an interval analysis loses too much precision in
the initial state of this program, since it maps all low input variables
y1 , y2 and y3 to [−∞, +∞]. After the conditional, it determines
that x belongs to the interval [−∞, +∞], which is a coarse overapproximation. Also, a polyhedron (Cousot and Halbwachs 1978)
does not capture the disjunction that is needed for this example
program (x = y2 or x = y3 ). Both abstract domains and many more
existing ones are not suitable for the task of inferring cardinalities
or dependences because they are convex. Using them as a basis to
extract counting information delivers an over-approximation of the
leakage, but a coarse one, especially in the presence of low inputs.
A disjunction of two polyhedra —through powerset domains,
disjunctive postconditions, or partitioning (Bourdoncle 1992)— is
as precise as the cardinality analysis for this example. However,
disjunctions are not tractable in general. As soon as one fixes a
maximum number of disjunctive elements (as in the quantitative
information flow analysis of Mardziel et al. (2011, 2013)) or defines
a widening operator to guarantee convergence, one loses the relative
precision wrt. classical dependence analyses (Amtoft and Banerjee
2004; Hunt and Sands 2006) that the cardinality analysis guarantees
(Cf. Corollary 1). Future work will investigate relying on cardinality analysis as a strategy guiding trace partitioning (Rival and
Mauborgne 2007). Combining our analyses with existing domains
will also deliver better precision.
Consider the following program.
Lemma 7 . (αdeptr , γdeptr ) is the composition of two Galois connections (αcrdtr , γcrdtr ) and (αlqonecc , γlqonecc ) :
γlqonecc
γ
crdtr
−−
−−−− (Card, v] ) ←
−−
−−−−−
− (Dep, v\ )
(P(P(Trc)), ⊆) ←
−−
−−
→
α −−→
α −−−−
crdtr
lqonecc
with:
αlqonecc (C ) , {l ;Sx | l ; x#n ∈ C and αlqone (n)}
γlqonecc (D) ,
{l ; x#n | n = γlqone (l ; x ∈ D)}
l∈L,x∈VarP
We use Lemmas 6 and 7 to abstract further the cardinality
abstract semantics and derive the correct by construction dependence
analysis of Section 6. This derivation, which can be found in
Appendix G, proves Lemma 4 and Theorem 2 stated earlier.
As a corollary and by Theorem 3, this also proves the precision of the cardinality analysis relative to Amtoft and Banerjee’s
logic (Amtoft and Banerjee 2004) as well as Hunt and Sands’ type
system (Hunt and Sands 2006, 2011).
Corollary 1 No leakage for well-typed programs. For all c,
C0 , C ∈ Card, ∆0 , ∆ ∈ VarP → L, where ⊥L ` ∆0 {c}∆,
and C = LcM] C0 , it holds that:
αhs
αlqonecc (C0 ) v̇ ∆0 =⇒
l
∀x ∈ VarP , l ∈ L, ∆(x) v l =⇒ OC
LxM] ≤ 1
◦
The cardinality analysis determines that there is no leakage for
programs that are “well-typed” by the flow-sensitive type system
of Hunt and Sands. By “well-typed”, we mean that the final typing
environment that is computed by the type system allows attackers
with security clearance l ∈ L to observe a variable x ∈ VarP .
To the best of our knowledge, the cardinality abstraction is
the first approximation-based analysis for quantitative information
flow that provides a formal precision guarantee wrt. traditional
analyses for qualitative information flow. This advantage makes
the cardinality analysis appealing even when interested in proving
a qualitative security policy such as non-interference, since the
cardinality abstraction provides quantitative information that may
assist in making better informed decisions if declassification is
necessary. Nonetheless, we need further experimentation to compare
to other quantitative analyses —see Section 9.
8.
Towards More Precision
This section introduces examples to evaluate the precision of the
analyses, and shows how existing analyses can be leveraged to
if (y1 ≥ secret ) then
x := y2
else
x := y3
1
2
if (y1 ≥ secret ) then x := y2 else x := y3 ;
o := x * y4
Listing 2. Leaking x
The cardinal abstraction determines that variable o leaks the two
possible values of x: for fixed low inputs, x has two possible values whereas y4 has one possible value. Relational abstract domains
such as polyhedra (Cousot and Halbwachs 1978) or octogons (Miné
2006a) do not support non-linear expressions, and therefore are
unable to compute a precise bound of the leakage for variable o.
Consider an analysis with a disjunction {x = y2 ∨ x = y3 } of polyhedra and linearisation over intervals (Miné 2006b). Linearisation of
expressions y2 ∗y4 and y3 ∗y4 will compute the following constraints
for variable o: {(o = y2 ∗ [−∞, +∞]) ∨ (o = y3 ∗ [−∞, +∞])} if
linearisation happens for the right side of expressions, or constraint
{(o = [−∞, +∞] ∗ y4 ) ∨ (o = [−∞, +∞] ∗ y4 )} if linearisation happens for the left side expressions. Two more combinations
of constraints are possible, but none will deduce that variable o
has at most 2 values, because the underlying domain of intervals
lacks the required precision. Linearisation over both intervals and
cardinalities delivers better precision.
We can now also improve the dependence abstraction:
αlqonecc
αcrdval
◦
l
Ol Lx2 M ◦ Lgrdx1 ==x2 M ◦ γcrdtr (C ) ≤ OC
Lx1 M] C
Therefore, we can deduce that:
αcrdtr
◦
Lgrdx1 ==x2 M ◦ γcrdtr (C )
v {l ; x#n ∈ C | x 6= x1 , x 6= x2 }
]
∪ {l ; x1 # min(n1 , n2 ), l ; x2 # min(n1 , n2 ) |
l ; x1 #n1 ∈ C , l ; x2 #n2 ∈ C }
, Lgrdx1 ==x2 M] C
For other comparison operators, we use as before Lgrdb M] C , C .
l ; x1 #n1 ∈ γlqonecc (D), l ; x2 #n2 ∈ γlqonecc (D)})
For other comparison operators, we also use Lgrdb M\ D , D.
With these new definitions, we can update the abstract semantics
of conditionals and loops, for both dependences and cardinalities, to
leverage the transfer functions Lgrd− M\ and Lgrd− M] .
Improved dependences abstract semantics LcM\ ∈ Dep → Dep
Lif b then c1 else c2 M\ D ,
let D1 = Lgrdb M\ ◦ Lc1 M\ D in
let D2 = Lgrd¬b M\ ◦ Lc2 M\ D in
let W (
= Mod(if b then c1 else c2 ) in
l
S π l (D1 ) t\ π l (D2 )
LbM\ D
if OD
l
/ W } otherwise
l∈L {l ; x ∈ π (D) | x ∈
The cardinality abstraction determines that initially L-equivalent
memories lead to a variety of at most 2 in the pointer p after the
conditional, whereas both y2 and y3 have a variety of 1. Assuming
an aliasing analysis determines that p may point to y2 or y3 , the
cardinality analysis determines that variable o has a variety of at
most 2, for initially L-equivalent memories.
l
Ol Lx1 M ◦ Lgrdx1 ==x2 M ◦ γcrdtr (C ) ≤ OC
Lx2 M] C
γlqonecc (D)
, Lgrdx1 ==x2 M\ D
Listing 3. Leaking 1 bit of secret
◦
◦
v {l ; x ∈ D | x 6= x1 , x 6= x2 }
∪ {l ; x1 , l ; x2 | l ; x1 ∈ D or l ; x2 ∈ D}
\
if (y1 ≥ secret ) then
p := &y2
else
p := &y3
o := * p
αcrdval
Lgrdx1 ==x2 M]
v αlqonecc ({l ; x#n ∈ γlqonecc (D) | x 6= x1 , x 6= x2 })
∪ αlqonecc ({l ; x1 # min(n1 , n2 ), l ; x2 # min(n1 , n2 ) |
\
Scaling to richer languages. We can rely on existing abstract
domains to support richer language constructs, e.g. pointers and
aliasing. Consider the following variation of Listing 1.
Improving precision. To improve precision of the cardinality
abstraction, we can augment it with existing abstract domains. One
shortcoming of the cardinality analysis is the fact that it is not
relational. Assuming attackers with security clearance L observe
both variables x and o after execution of the program in Listing 2,
the cardinality abstraction leads us to compute a leakage of two bits:
four different possible values, instead of only 2 possible values for
initially L-equivalent memories. Relying on a relational domain with
linearisation (Miné 2006b) over cardinalities captures the required
constraints {L ; x#2, L ; o#1 ∗ x} to compute a leakage of
only one bit; these constraints are to be interpreted as “initially
L-equivalent memories result in o being equal to one fixed integer
times x, and x having at most 2 values”.
We leave these extensions of cardinality analysis —and its abstraction as dependence analysis— for future work. In the following,
we focus on one particular improvement to both previous analyses in
order to gain more precision. We uncovered this case while deriving
the analyses, by relying on the calculational framework of abstract
interpretation. Indeed, notice that the following holds:
◦
Lwhile b do cM\ D , Lgrd¬b M\
]
◦
\
lfpv
D Lif b then c1 else c2 M
Improved cardinality abs. semantics
LcM] ∈ Card → Card
Lif b then c1 else c2 M] C ,
let C1 = Lgrdb M] ◦ Lc1 M] C in
let C2 = Lgrd¬b M] ◦ Lc2 M] C in
let W (
= Mod(if b then c1 else c2 ) in
l
S π l (C1 ) t] π l (C2 )
LbM] C = 1
if OC
l
]
l
l∈L π (C1 ) t add(W,π l (C )) π (C2 ) otherwise
Lwhile b do cM] C , Lgrd¬b M\
]
◦
]
lfpv
C Lif b then c1 else c2 M
To illustrate the benefits of this improvement, consider the
following example.
1
2
3
4
5
while ( secret != y3 ) do {
x := x +1;
secret := secret - 1;
}
o := secret ;
Listing 4. Improved precision
The cardinality analysis determines that initially L-equivalent
memories result in x having an infinity of values: the L-cardinality
of x grows until it is widened to ∞. In contrast, cardinalities also
determine that variables o and secret have only 1 value, assuming Lequivalent memories. This is because of the reduction that concerns
variable secret after the while loop, specifically Lgrdsecret==y3 M\ .
Similarly, the improved dependence analysis also determines that
both variables secret and o are low. These are sound precision gains
for termination-insensitive noninterference; Askarov et al. (2008)
discusses the guarantees provided by this security requirement.
Remarkably, this has been overlooked by many previous analyses.
In fact, this simple improvement makes our dependence analysis
strictly more precise than Amtoft and Banerjee (2004)’s and Hunt
and Sands (2006, 2011)’s analyses and incomparable to the more
recent dependence analysis of Müller et al. (2015).
Combination with intervals. Consider now the following
example inspired from Müller et al. (2015).
0
1
2
3
4
1
2
3
4
5
6
7
if ( secret == 0) then {
x := 0;
y := y + 1;
}
else {
x := 0;
}
5
6
7
8
9
10
11
Listing 5. Example program from Müller et al. (2015)
The analysis of Müller et al. (2015) determines that x is low,
whereas the cardinality abstraction determines that L-equivalent
memories result in at most 2 values for variable x, because it does
not track the actual values of variables. We can combine cardinality
with an interval analysis to be more precise in such cases, through a
reduced product (Cousot and Cousot 1979; Granger 1992; Cortesi
et al. 2013).
Assume a set StInt of interval environments provided with
˙ ],Int . Assume also
the usual partial order that we denote by ≤
Int
Int
a Galois connection (α , γ ) enabling the derivation of an
interval analysis as an approximation of a standard collecting
semantics defined over P(Trc). We can lift this Galois connection
to P(P(Trc)) to obtain a Galois connection by compositing with
(αhpp , γhpp ), to obtain (α0 , γ 0 ) , (αInt ◦ αhpp , γ Int ◦ γhpp ) with:
γhpp
γ Int
hpp
αInt
˙ ],Int )
−−
−−−
− (P(Trc), ⊆) ←
−−
−−
−
− (StInt, ≤
(P(P(Trc)), ⊆) ←
−−
→
−−
→
α −−
A Granger’s reduced product Granger (1992) for the cardinality
abstraction and an interval analysis may be defined as a pair
of functions toint ∈ Card × StInt → StInt and tocard ∈
Card × StInt → Card verifying the following conditions:
1. soundness:
γ 0 (toint(C , ı)) ∩ γcrdtr (C )
γ 0 (ı) ∩ γcrdtr (tocard(C , ı))
2. reduction:
toint(C , ı)
tocard(C , ı)
= γ 0 (ı) ∩ γcrdtr (C )
= γ 0 (ı) ∩ γcrdtr (C )
˙ ],Int ı
≤
v] C
Let us denote by size the function that returns the size of an
interval. One such Granger’s reduced product can be defined as:
tocard
tocard(C , ı)
∈
,
toint
toint(C , ı)
∈
,
Card × StInt → Card
{l ; x#n0 | l ; x#n ∈ C and
n0 = min (n, size ı(x))}
Card × StInt → Card
ı
Once enhanced with this reduced product, the cardinality analysis determines for the program in Listing 5, that L-equivalent memories result in at most one possible value for variable x.
The dependence analysis can be improved similarly, with a
reduction function defined as follows:
todep ∈ Dep × StInt → Dep
todep(D, ı) , D ∪ {l ; x | l ∈ L and size ı(x) = 1}
Once extended with a reduced product with intervals, the dependence analysis is also able to determine that variable x is low for
the program in Listing 5.
12
13
//L ; h#∞, L ; y1 #1, L ; y2 #1, L ; y3 #1
y1 := 1; //L ; y1 #1
if ( h == y1 ) then {
skip ; //L ; h#1, L ; y1 #1, L ; y2 #1
}
else {
y2 := 5; //L ; y1 #1, L ; y2 #1
while (y2 != 1) do {
y2 := y2 -1; //L ; y2 #1
y1 := y2 ; //L ; y1 #1
} //L ; y1 #1, L ; y2 #1
}
//L ; h#∞, L ; y1 #2, L ; y2 #2, L ; y3 #1
o := y1 * y3 ; //L ; o#2
Listing 6. No leakage for variable o
More reduced products. As a final example, let us consider
Listing 6, inspired by Besson et al. (2016, program 7), that we
annotate with the result of the improved cardinality abstraction. To
the best of our knowledge, no existing automated static analysis
determines that variable o is low at the end of this program. Also,
no prior monitor but the one recently presented by Besson et al.
(2016) accepts all executions of this program, assuming attackers
with clearance L can observe variable o.
For initially L-equivalent memories, the cardinality abstraction
determines that variables y1 , y2 and o have at most two values. This
result is precise for y2 , but not precise for y1 and o. As a challenge,
let us see what is required to gain more precision to determine that
both variables y1 and o have at most 1 possible value – they are low.
To tackle this challenge, we need to consider cardinality combined with an interval analysis and a simple relational domain tracking equalities. With the equality y1 = y2 at the exit of the loop, both
y1 and y2 will be reduced to the singleton interval [1, 1]. After the
conditional, we still deduce that y2 has at most 2 different values
thanks to the cardinality abstraction. Using intervals, we deduce that
variable y1 has only one value (singleton interval [1, 1]). And finally,
at the last assignment the cardinalities abstraction determines that
variable o has only one possible value. Similarly, this same combination of analyses can be put to use to let the dependence analysis
reach the desired precision.
9.
Related Work
Although noninterference has important applications, for many security requirements it is too strong. That is one motivation for research
in quantitative information flow analysis. In addition, a number of
works investigate weakenings of noninterference and downgrading
policies that are conditioned on events or data values (Askarov and
Sabelfeld 2007; Banerjee et al. 2008; Sabelfeld and Sands 2009;
Mastroeni and Banerjee 2011). Assaf (2015, Chapter 4) proposes
to take the guarantees provided by termination-insensitive noninterference (Askarov et al. 2008) as an explicit definition for security;
this Relative Secrecy requirement is inspired by Volpano and Smith
(2000) who propose a type-system preventing batch-job programs
from leaking secrets in polynomial time. Giacobazzi and Mastroeni
(2004) introduce abstract noninterference, which generalizes noninterference by means of abstract interpretations that specify, for
example, limits on the attacker’s power and the extent of partial
releases (declassification). The survey by Mastroeni (2013) further
generalizes the notion and highlights, among other things, its applicability to a range of underlying semantics. The Galois connections
in this work are at the level of trace sets, not sets of sets. Abstract
noninterference retains the explicit 2-run formulation (Volpano et al.
1996; Sabelfeld and Myers 2003): from two related initial states, two
executions lead to related final states. The relations are defined in
terms of abstract interpretations of the individual states/executions.
Mastroeni and Banerjee (2011) show how to infer indistinguishability relations—modelling attackers’ observations—to find the best
abstract noninterference policy that holds. The inference algorithm
iteratively refines the relation by using counter-examples and abstract domain completion (Cousot and Cousot 1979).
Set-of-sets structures occur in work on abstraction for nondeterministic programs, but in those works one level of sets are powerdomains for nondeterminacy; the properties considered are trace
properties (Schmidt 2009, 2012). Hunt and Sands (1991) develop a
binding time analysis and a strictness analysis (Hunt 1990) based on
partial equivalence relations: Their concretisations are sets of equivalence classes. Cousot and Cousot (1994) point out that this analysis
could be achieved by a collecting semantics over sets-of-sets, defined simply as a direct image. To the best of our knowledge this
has not been explored further in the literature, except in unpublished
work on which this paper builds (Assaf 2015; Assaf et al. 2016b).
Clarkson et al. (2014); Finkbeiner et al. (2015) extend temporal
logic with means to quantify over multiple traces in order to express
hyperproperties, and provide model checking algorithms for finite
space systems. Agrawal and Bonakdarpour (2016) introduce a
technique for runtime verification of k-safety properties.
The dependences analysis we derive is similar to the information flow logic of Amtoft and Banerjee (2004) and the equivalent
flow-sensitive type system of Hunt and Sands (2006). Amtoft and
Banerjee use the domain P(Trc) and on the basis of a relational
logic they validate a forward analysis. In effect their interpretation
of “independences” is a Galois connection with sets of sets, but
the analysis is not formulated or proved correct as an abstract interpretation. To deal with dynamically allocated state, Amtoft et al.
(2006) augment the relational assertions of information flow logic
with region assertions, which can be computed by abstract interpretation. This is used both to express agreement relations between the
two executions and to approximate modifiable locations. This approach is generalized in Banerjee et al. (2016) to a relational Hoare
logic for object-based programs that encompasses information flow
properties with conditional downgrading (Banerjee et al. 2008).
Müller et al. (2015) give a backwards analysis that infers dependencies and is proved strictly more precise than (Hunt and Sands
2006; Amtoft and Banerjee 2004). This is achieved by product
construction that facilitates inferring relations between variables in
executions that follow different control paths. Correctness of the
analysis is proved by way of a relational Hoare logic. The variations
of our proposed analyses, in Section 8, rivals theirs in terms of
precision—they are incomparable.
Our dependence analysis relies on an approximation of the modifiable variables, to soundly track implicit flows due to control flow,
instead of labelling a program counter variable pc to account for implicit flows (Sabelfeld and Myers 2003). Zanioli and Cortesi (2011)
also derive a similar analysis through a syntactic Galois connection—
a syntactic assignment z := x ∗ y is abstracted into a propositional
formula x → z∧y → z denoting an information flow from variables
x and y to variable z. The soundness of this analysis wrt. a semantic
property such as noninterference requires more justification, though
it is remarkable that the concretisation of propositional formula
yields, roughly speaking, a set of program texts. Zanotti (2002) also
provides an abstract interpretation account of a flow-insensitive type
system (Volpano et al. 1996) enforcing noninterference by guaranteeing a stronger safety property, namely that sensitive locations
should not influence public locations (Boudol 2008).
Kovács et al. (2013) explicitly formulate termination-insensitive
noninterference as an abstract interpretation, namely the “merge
over all twin computations” that makes explicit both the 2-safety
aspect and the need for an analysis to relate some aligned intermediate states. Their analysis, like many others, is based on reducing the
problem to a safety property of product programs. Sousa and Dillig
(2016) implement an algorithm that automates reasoning in a Hoare
logic for k-safety, implicitly constructing product programs; the
performance compares favorably with explicit construction of product programs. Program dependency graphs are another approach to
dependency, shown to be correct for noninterference by Wasserrab
et al. (2009) using slicing and a simulation argument.
Denning (1982, Chap. 5) proposes the first quantitative measure
of a program’s leakage in terms of Shannon entropy (Shannon
1948). Other quantitative metrics emerge in the literature (Braun
et al. 2009; Clarkson et al. 2009; Smith 2009; Dwork 2011; Smith
2011; Alvim et al. 2012). These quantitative security metrics model
different scenarios suitable for different policies. Most existing static
analyses for quantitative information flow leverage existing model
checking tools and abstract domains for safety; they prove that a
program satisfies a quantitative security requirement by proving a
stronger safety property. In contrast, the cardinal abstraction proves a
hyperproperty by inferring a stronger hyperproperty satisfied by the
analysed program. This is key to target quantitative information flow
in mutlilevel security lattices, beyond the 2-point lattice {L, H}.
Backes et al. (2009) synthesize equivalence classes induced by
outputs over low equivalent memories by relying on software model
checkers, in order to bound various quantitative metrics. Heusser and
Malacaria (2009) also rely on a similar technique to quantify information flow for database queries. Köpf and Rybalchenko (2010) note
that the exact computation of information-theoretic characteristics
is prohibitively hard, and propose to rely on approximation-based
analyses, among which are randomisation techniques and abstract
interpretation ones. They also propose to rely on a self-composed
product program to model a scenario where attackers may refine
their knowledge by influencing the low inputs. Klebanov (2014)
relies on similar techniques to handle programs with low inputs, and
uses polyhedra to synthesize linear constraints (Cousot and Halbwachs 1978) over variables. Mardziel et al. (2013) decide whether
answering a query on sensitive data augments attackers’ knowledge
beyond a certain threshold, by using probabilistic polyhedra.
10.
Conclusion
Galois connection-based semantic characterisations of program
analyses provide new perspectives and insights that lead to improved
techniques. We have extended the framework to fully encompass
hyperproperties, through a remarkable form of hypercollecting
semantics that enables calculational derivation of analyses. This
new foundation raises questions too numerous to list here.
One promising direction is to combine dependence and cardinality analysis with existing abstract domains, e.g. through advanced
symbolic methods (Miné 2006b), and partitioning (Handjieva and
Tzolovski 1998; Rival and Mauborgne 2007).
Static analysis of secure information flow has yet to catch up with
recent advances in dynamic information flow monitoring (Besson
et al. 2013; Bello et al. 2015; Hedin et al. 2015; Assaf and Naumann
2016; Besson et al. 2016). We discussed, in Section 8, how existing
static analyses may be of use to statically secure information flow.
It seems likely that hypercollecting semantics will also be of use for
dynamic analyses.
Acknowledgments
Thanks to Anindya Banerjee and the anonymous reviewers for
thoughtful comments and helpful feedback. This work was partially
supported by NSF awards CNS-1228930 and CCF-1649884, ANR
project AnaStaSec ANR-14-CE28-0014 and a CFR CEA Phd
Fellowship.
References
S. Agrawal and B. Bonakdarpour. Runtime verification of k-safety hyperproperties in HyperLTL. In IEEE Computer Security Foundations
Symposium, pages 239–252, 2016.
M. S. Alvim, K. Chatzikokolakis, C. Palamidessi, and G. Smith. Measuring
information leakage using generalized gain functions. In IEEE Computer
Security Foundations Symposium, pages 265–279, 2012.
T. Amtoft and A. Banerjee. Information flow analysis in logical form. In
Static Analysis Symposium, pages 100–115, 2004.
T. Amtoft, S. Bandhakavi, and A. Banerjee. A logic for information
flow in object-oriented programs. In ACM Symposium on Principles
of Programming Languages, pages 91–102, 2006.
A. Askarov and A. Sabelfeld. Gradual release: Unifying declassification,
encryption and key release policies. In IEEE Symposium on Security and
Privacy, 2007.
A. Askarov, S. Hunt, A. Sabelfeld, and D. Sands. Termination-insensitive
noninterference leaks more than just a bit. In European Symposium on
Research in Computer Security, volume 5283 of LNCS, 2008.
M. Assaf. From Qualitative to Quantitative Program Analysis : Permissive
Enforcement of Secure Information Flow. PhD thesis, Université de
Rennes 1, May 2015. https://hal.inria.fr/tel-01184857.
M. Assaf and D. Naumann. Calculational design of information flow
monitors. In IEEE Computer Security Foundations Symposium, pages
210–224, 2016.
M. Assaf, D. Naumann, J. Signoles, É. Totel, and F. Tronel. Hypercollecting
semantics and its application to static analysis of information flow.
Technical report, Apr. 2016a. URL https://arxiv.org/abs/1608.
01654.
M. Assaf, J. Signoles, É. Totel, and F. Tronel. The cardinal abstraction
for quantitative information flow. In Workshop on Foundations of
Computer Security (FCS), June 2016b. https://hal.inria.fr/hal01334604.
M. Backes, B. Köpf, and A. Rybalchenko. Automatic discovery and
quantification of information leaks. In IEEE Symposium on Security
and Privacy, pages 141–153. IEEE, 2009.
A. Banerjee, D. A. Naumann, and S. Rosenberg. Expressive declassification
policies and modular static enforcement. In IEEE Symposium on Security
and Privacy, pages 339–353, 2008.
A. Banerjee, D. A. Naumann, and M. Nikouei. Relational logic with framing
and hypotheses. In 36th IARCS Annual Conference on Foundations of
Software Technology and Theoretical Computer Science, 2016. To appear.
G. Barthe, P. R. D’Argenio, and T. Rezk. Secure information flow by selfcomposition. In IEEE Computer Security Foundations Workshop, pages
100–114, 2004.
L. Bello, D. Hedin, and A. Sabelfeld. Value sensitivity and observable
abstract values for information flow control. In Logic for Programming,
Artificial Intelligence, and Reasoning (LPAR), pages 63–78, 2015.
N. Benton. Simple relational correctness proofs for static analyses and program transformations. In ACM Symposium on Principles of Programming
Languages, pages 14–25, 2004.
J. Bertrane, P. Cousot, R. Cousot, J. Feret, L. Mauborgne, A. Miné, and
X. Rival. Static analysis and verification of aerospace software by abstract
interpretation. In AIAA Infotech@Aerospace 2010, 2012.
J. Bertrane, P. Cousot, R. Cousot, J. Feret, L. Mauborgne, A. Miné, and
X. Rival. Static analysis and verification of aerospace software by abstract
interpretation. Foundations and Trends in Programming Languages, 2
(2-3):71–190, 2015.
F. Besson, N. Bielova, and T. Jensen. Hybrid information flow monitoring
against web tracking. In IEEE Computer Security Foundations Symposium, pages 240–254. IEEE, 2013.
F. Besson, N. Bielova, and T. Jensen. Hybrid monitoring of attacker
knowledge. In IEEE Computer Security Foundations Symposium, pages
225–238, 2016.
G. Boudol. Secure information flow as a safety property. In Formal Aspects
in Security and Trust, pages 20–34, 2008.
F. Bourdoncle. Abstract interpretation by dynamic partitioning. Journal of
Functional Programming, 2(04):407–435, 1992.
C. Braun, K. Chatzikokolakis, and C. Palamidessi. Quantitative notions of
leakage for one-try attacks. In Mathematical Foundations of Programming Semantics (MFPS), volume 249, pages 75–91, 2009.
D. Cachera and D. Pichardie. A certified denotational abstract interpreter. In
Interactive Theorem Proving (ITP), pages 9–24, 2010.
M. R. Clarkson and F. B. Schneider. Hyperproperties. In IEEE Computer
Security Foundations Symposium, pages 51–65, 2008.
M. R. Clarkson and F. B. Schneider. Hyperproperties. Journal of Computer
Security, 18(6):1157–1210, 2010.
M. R. Clarkson, A. C. Myers, and F. B. Schneider. Quantifying information
flow with beliefs. Journal of Computer Security, 17:655–701, 2009.
M. R. Clarkson, B. Finkbeiner, M. Koleini, K. K. Micinski, M. N. Rabe,
and C. Sánchez. Temporal logics for hyperproperties. In Principles of
Security and Trust, volume 8414 of LNCS, pages 265–284, 2014.
E. Cohen. Information transmission in computational systems. In Proceedings of the sixth ACM Symposium on Operating Systems Principles, pages
133–139, 1977.
A. Cortesi and M. Zanioli. Widening and narrowing operators for abstract
interpretation. Computer Languages, Systems & Structures, pages 24–42,
2011.
A. Cortesi, G. Costantini, and P. Ferrara. A survey on product operators
in abstract interpretation. In Semantics, Abstract Interpretation, and
Reasoning about Programs: Essays Dedicated to David A. Schmidt on the
Occasion of his Sixtieth Birthday, volume 129 of EPTCS, pages 325–336,
2013.
P. Cousot. The calculational design of a generic abstract interpreter. In
M. Broy and R. Steinbrüggen, editors, Calculational System Design,
volume 173, pages 421–506. NATO ASI Series F. IOS Press, Amsterdam,
1999.
P. Cousot. Constructive design of a hierarchy of semantics of a transition
system by abstract interpretation. Theoretical Computer Science, 277
(1-2):47–103, 2002.
P. Cousot and R. Cousot. Abstract interpretation: a unified lattice model for
static analysis of programs by construction or approximation of fixpoints.
In ACM Symposium on Principles of Programming Languages, pages
238–252, 1977.
P. Cousot and R. Cousot. Systematic design of program analysis frameworks.
In ACM Symposium on Principles of Programming Languages, pages
269–282, 1979.
P. Cousot and R. Cousot. Comparing the galois connection and widening/narrowing approaches to abstract interpretation. In Programming Language
Implementation and Logic Programming (PLILP), pages 269–295, 1992.
P. Cousot and R. Cousot. Higher-order abstract interpretation (and application to comportment analysis generalizing strictness, termination,
projection and per analysis of functional languages). In International
Conference on Computer Languages (ICCL), pages 95–112, 1994.
P. Cousot and N. Halbwachs. Automatic discovery of linear restraints
among variables of a program. In ACM Symposium on Principles of
Programming Languages, pages 84–96, 1978.
Á. Darvas, R. Hähnle, and D. Sands. A theorem proving approach to analysis
of secure information flow. In Security in Pervasive Computing, pages
193–209, 2005.
D. E. R. Denning. Cryptography and Data Security. Addison-Wesley
Longman Publishing Co., Inc., 1982.
D. E. R. Denning and P. J. Denning. Certification of programs for secure
information flow. Communications of ACM, 20(7):504–513, 1977.
G. Doychev, D. Feld, B. Köpf, L. Mauborgne, and J. Reineke. Cacheaudit:
A tool for the static analysis of cache side channels. In USENIX Security
Symposium, pages 431–446, 2013.
C. Dwork. A firm foundation for private data analysis. Communications of
ACM, pages 86–95, 2011.
B. Finkbeiner, M. N. Rabe, and C. Sánchez. Algorithms for model checking
HyperLTL and HyperCTL ˆ*. In Computer Aided Verification, volume
9206 of LNCS, pages 30–48, 2015.
R. Giacobazzi and I. Mastroeni. Abstract non-interference: parameterizing
non-interference by abstract interpretation. In ACM Symposium on
Principles of Programming Languages, pages 186–197, 2004.
J. A. Goguen and J. Meseguer. Security policies and security models. In
IEEE Symposium on Security and Privacy, pages 11–20, 1982.
P. Granger. Improving the results of static analyses programs by local decreasing iteration. In Foundations of Software Technology and Theoretical
Computer Science, volume 652, pages 68–79, 1992.
M. Handjieva and S. Tzolovski. Refining dtatic analyses by trace-based partitioning using control flow. In International Static Analysis Symposium,
1998.
D. Hedin, L. Bello, and A. Sabelfeld. Value-sensitive hybrid information
flow control for a JavaScript-Like language. In IEEE Computer Security
Foundations Symposium, pages 351–365, 2015.
J. Heusser and P. Malacaria. Applied quantitative information flow and
statistical databases. In Formal Aspects in Security and Trust, pages
96–110, 2009.
S. Hunt. PERs generalize projections for strictness analysis (extended
abstract). In Proceedings of the Third Annual Glasgow Workshop on
Functional Programming, 1990.
S. Hunt and D. Sands. Binding time analysis: A new PERspective. In
Proceedings of the Symposium on Partial Evaluation and SemanticsBased Program Manipulation, PEPM’91, Yale University, New Haven,
Connecticut, USA, June 17-19, 1991, pages 154–165, 1991.
S. Hunt and D. Sands. On flow-sensitive security types. In ACM Symposium
on Principles of Programming Languages, pages 79–90, 2006.
S. Hunt and D. Sands. From exponential to polynomial-time security typing
via principal types. In ACM Workshop on Programming Languages and
Analysis for Security, pages 297–316, 2011.
V. Klebanov. Precise quantitative information flow analysis - a symbolic
approach. Theoretical Computer Science, 538:124–139, 2014.
B. Köpf and A. Rybalchenko. Approximation and randomization for
quantitative information-flow analysis. In IEEE Computer Security
Foundations Symposium, pages 3–14, 2010.
B. Köpf and A. Rybalchenko. Automation of quantitative information-flow
analysis. In Formal Methods for Dynamical Systems - 13th International
School on Formal Methods for the Design of Computer, Communication,
and Software Systems, volume 7938 of LNCS, pages 1–28, 2013.
M. Kovács, H. Seidl, and B. Finkbeiner. Relational abstract interpretation for
the verification of 2-hypersafety properties. In ACM SIGSAC conference
on Computer and Communications Security, pages 211–222, 2013.
P. Mardziel, S. Magill, M. Hicks, and M. Srivatsa. Dynamic enforcement
of knowledge-based security policies. In IEEE Computer Security
Foundations Symposium, pages 114–128. IEEE, 2011.
P. Mardziel, S. Magill, M. Hicks, and M. Srivatsa. Dynamic enforcement of
knowledge-based security policies using probabilistic abstract interpretation. Journal of Computer Security, 21(4):463–532, 2013.
I. Mastroeni. Abstract interpretation-based approaches to security - A
survey on abstract non-interference and its challenging applications.
In Semantics, Abstract Interpretation, and Reasoning about Programs:
Essays Dedicated to David A. Schmidt on the Occasion of his Sixtieth
Birthday, volume 129 of EPTCS, pages 41–65, 2013.
I. Mastroeni and A. Banerjee. Modelling declassification policies using
abstract domain completeness. Mathematical Structures in Computer
Science, 21(06):1253–1299, 2011.
J. McLean. A general theory of composition for trace sets closed under
selective interleaving functions. In IEEE Symposium on Security and
Privacy, pages 79–93, 1994.
A. Miné. The octagon abstract domain. Higher-order and symbolic
computation, 19(1):31–100, 2006a.
A. Miné. Symbolic methods to enhance the precision of numerical abstract
domains. In Verification, Model Checking, and Abstract Interpretation,
pages 348–363, 2006b.
C. Müller, M. Kovács, and H. Seidl. An analysis of universal information
flow based on self-composition. In IEEE Computer Security Foundations
Symposium, pages 380–393, 2015.
F. Nielson, H. R. Nielson, and C. Hankin. Principles of Program Analysis.
Springer, 1999.
A. Rényi. On measures of entropy and information. In the Fourth Berkeley
Symposium on Mathematical Statistics and Probability, 1961.
X. Rival and L. Mauborgne. The trace partitioning abstract domain. ACM
Transactions on Programming Languages and Systems, 29(5):26, 2007.
J. Rushby. Security requirements specifications: How and what. In Symposium on Requirements Engineering for Information Security (SREIS),
2001.
A. Sabelfeld and A. C. Myers. Language-based information-flow security.
IEEE Journal on Selected Areas in Communications, 21(1):5–19, 2003.
A. Sabelfeld and D. Sands. Declassification: Dimensions and principles.
Journal of Computer Security, 17(5), 2009.
D. A. Schmidt. Abstract interpretation from a topological perspective. In
Static Analysis, 16th International Symposium, volume 5673 of LNCS,
pages 293–308, 2009.
D. A. Schmidt. Inverse-limit and topological aspects of abstract interpretation. Theoretical Computer Science, 430:23–42, 2012.
D. Schoepe, M. Balliu, B. C. Pierce, and A. Sabelfeld. Explicit secrecy: A
policy for taint tracking. In IEEE European Symposium on Security and
Privacy, pages 15–30, 2016.
C. E. Shannon. A mathematical theory of communication. The Bell System
Technical Journal, 27:379–423, 1948.
G. Smith. On the foundations of quantitative information flow. In International Conference on Foundations of Software Science and Computational
Structures, pages 288–302, 2009.
G. Smith. Quantifying information flow using min-entropy. In Quantitative
Evaluation of Systems (QEST), 2011 Eighth International Conference on,
pages 159–167. IEEE, 2011.
M. Sousa and I. Dillig. Cartesian Hoare logic for verifying k-safety
properties. In ACM Conference on Programming Language Design
and Implementation, pages 57–69, 2016.
T. Terauchi and A. Aiken. Secure information flow as a safety problem. In
Static Analysis Symposium, pages 352–367, 2005.
D. Volpano and G. Smith. Eliminating covert flows with minimum typings.
In IEEE Computer Security Foundations Workshop, pages 156–168, 1997.
D. Volpano and G. Smith. Verifying secrets and relative secrecy. In ACM
Symposium on Principles of Programming Languages, pages 268–276,
2000.
D. Volpano, C. Irvine, and G. Smith. A sound type system for secure flow
analysis. Journal of Computer Security, 4(2-3):167–187, 1996.
D. M. Volpano. Safety versus secrecy. In Static Analysis Symposium, pages
303–311, 1999.
D. Wasserrab, D. Lohner, and G. Snelting. On PDG-based noninterference
and its modular proof. In ACM Workshop on Programming Languages
and Analysis for Security, pages 31–44, 2009.
G. Winskel. The Formal Semantics of Programming Languages: an
Introduction. Cambridge, 1993.
H. Yasuoka and T. Terauchi. On bounding problems of quantitative
information flow. Journal of Computer Security, 19(6):1029–1082, 2011.
A. Zakinthinos and S. Lerner. A general theory of security properties. In
IEEE Symposium on Security and Privacy, pages 94–102, 1997.
M. Zanioli and A. Cortesi. Information leakage analysis by abstract
interpretation. In SOFSEM 2011: Theory and Practice of Computer
Science, pages 545–557, 2011.
M. Zanotti. Security typings by abstract interpretation. In Static Analysis
Symposium, volume 2477, pages 360–375, 2002.
Appendix A.
Symbols
Val
∞
v ∈ Val
V ∈ P(Val)
V ∈ P(P(Val))
Trc
t ∈ Trc
T ∈ P(Trc)
T ∈ P(P(Trc))
States
σ ∈ States
Σ ∈ P(States)
S ∈ P(P(States))
States∗ , ∪n∈N Statesn
a set of integers
an infinite cardinal number
an integer
a set of values
a set of sets of values
the set of (relational) traces
a trace
a set of traces
a set of sets of traces
the set of states
a state
a set of states
a set of sets of states
the set of finite sequence of states
VarP
L
l∈L
Γ ∈ VarP → L
the set of variables of a program P
a multilevel security lattice
a security level
an initial typing context
γ
−
−A
C←
−−
α→
a Galois connection
JcK ∈ Trc → Trc
JeK ∈ Trc → Val
JeKpre ∈ Trc → Val
⦃c⦄ ∈ P(Trc) → P(Trc)
LcM ∈ P(P(Trc)) → P(P(Trc))
denotational semantics of commands
value of e in final state
value of e in initial state
collecting semantics
hypercollecting semantics
l;x
atomic dependence: “agreement up to security
level l leads to agreement on x”
A set of atomic dependency constraints
atomic cardinality: “agreement up to security
level l leads to an l-cardinality of n values for x”
A valid set of atomic cardinality constraints
D ∈ Dep
l ; x#n
C ∈ Card
Appendix B.
Background: Collecting Semantics, Galois Connections
Lemma 1 Element-wise abstraction. Let elt ∈ C → A be a function between sets. Let αelt (C) , {elt(c) | c ∈ C} and
γelt
−−
−−
− (P(A), ⊆).
γelt (A) , {c | elt(c) ∈ A}. Then (P(C), ⊆) ←
−−
→
α−
elt
Proof.
Let C ∈ P(C) and A ∈ P(A).
αelt (C) ⊆ A ⇐⇒ {elt(c) | c ∈ C} ⊆ A
⇐⇒ ∀c ∈ C, elt(c) ∈ A
⇐⇒ C ⊆ {c | elt(c) ∈ A}
⇐⇒ C ⊆ γelt (A)
Appendix C.
Domains and Galois Connections for Hyperproperties
Lemma 2 . Let C be a set. Define αhpp (C) , ∪C∈C C and γhpp (C) , P(C). These form a Galois connection:
γhpp
−−
−−−
− (P(C), ⊆)
(P(P(C)), ⊆) ←
−−
→
α −−
hpp
Proof.
This is a special case of the supremus abstraction (Cousot 2002, p.52), that is defined in Lemma 3. Indeed, we can instantiate a supremus
γhpp
−−
−−−
− P(C), with αhpp (C) = ∪C∈C C
abstraction by taking hpp , id (∈ P(C) → P(C)). We thus obtain a Galois connection P(P(C)) ←
−−
→
α −−
hpp
and γhpp (C) = {C 0 ∈ P(C) | C 0 ⊆ C} (= P(C)). Notice here that the powerset of a set C, provided with set inclusion as a partial order, is
a complete lattice as required by the supremus abstraction.
Lemma 3 Supremus abstraction. Let elt ∈ C → A be a function from a set C, with codomain forming a complete lattice (A, v). Let
αelt (C) , tc∈C elt(c) and γelt (a) , {c ∈ C | elt(c) v a}. Then
γ
elt
−−
−
−
− (A, v)
(P(C), ⊆) ←
−−
→
α−
elt
Proof.
Notice that the assumption that the lattice (A, v, t) is complete guarantees that αelt (C) is well-defined: the set {elt(c) | c ∈ C} does
have a supremum.
Let C ∈ P(C) and a ∈ A. The proof goes by definitions.
αelt (C) v a ⇐⇒ tc∈C elt(c) v a
⇐⇒ ∀c ∈ C, elt(c) v a
⇐⇒ C ⊆ {c ∈ C | elt(c) v a}
⇐⇒ C ⊆ γelt (a)
Appendix D.
Hypercollecting Semantics
Before proving the main result of this section in Theorem 1, we will first prove Lemma 8.
Both proofs of Lemma 8 and Theorem 1 are by structural induction. Most cases follow from definitions. The important cases are for while
loops and the proof technique is a classical one when using a denotational semantics. E.g., in order to prove equality of two denotations
characterised as a fixpoint, it suffices to introduce two sequences that converge towards the fixpoint characterisations and prove equality of
these sequences. This ensures that their limits – the denotations characterised as a fixpoint – are equal.
Let us now prove Lemma 8 – this lemma is used later in the proof case of while loops for Theorem 1.
Lemma 8 . For all commands c, for all sets of traces T ∈ P(Trc), the standard collecting semantics (Section 2) can be expressed as the direct
image of the denotational semantics :
⦃c⦄T = {JcKt ∈ Trc | t ∈ T }
Proof.
The proof proceeds by structural induction on commands. The most important case is the case of while loops.
1 – Case skip:
⦃skip⦄T = T = {JskipKt | t ∈ T }
2 – Case x := e:
⦃x := e⦄T = {Jx := eKt | t ∈ T }
3 – Case c1 ; c2 :
⦃c1 ; c2 ⦄T = ⦃c2 ⦄ ◦ ⦃c1 ⦄T
= HBy induction on c1 I
⦃c2 ⦄({Jc1 Kt ∈ Trc | t ∈ T })
= HBy induction on c2 I
{Jc2 K ◦ Jc1 Kt ∈ Trc | t ∈ T }
= {Jc1 ; c2 Kt ∈ Trc | t ∈ T }
4 – Case if (b) then c1 else c2 :
⦃if (b) then c1 else c2 ⦄T = ⦃c1 ⦄ ◦ ⦃ grdb ⦄T ∪ ⦃c2 ⦄ ◦ ⦃ grd¬b ⦄T
= HBy induction hypothesis on both c1 and c2 I
{Jc1 Kt ∈ Trc | t ∈ ⦃ grdb ⦄T } ∪ {Jc2 Kt ∈ Trc | t ∈ ⦃ grd¬b ⦄T }
= {Jif (b) then c1 else c2 Kt ∈ Trc | t ∈ T }
5 – Case while (b) do c:
5.1 – Let us first prove the following intermediate result:
∀T ∈ P(Trc),
b
◦ ⦃ grd ⦄X .
λX.
T
∪
⦃c⦄
{Jwhile (b) do cKt ∈ Trc | t ∈ T } = ⦃ grd¬b ⦄ lfp⊆
∅
Indeed, let the sequence (xTn )n≥0 be defined as:
with F defined as
xTn , {F (n) (⊥)(t) ∈ Trc | t ∈ T }
(
t
F(w)(t) ,
w
◦
JcKt
if JbKt = 0
otherwise
Notice that for all t ∈ T , the sequence (F (n) (⊥)(t))n≥0 converges and is equal to the evaluation of the while loop in the state
t (i.e. Jwhile b do cKt = F (∞) (⊥)(t)), by definition of the denotational semantics of loops; thus, the sequence xTn converges to
{Jwhile b do cKt ∈ Trc | t ∈ T }.
Let also the sequences (ynT )n≥0 and (gnT )n≥0 be defined as:
ynT , ⦃ grd¬b ⦄gnT
T
gn+1
, T ∪ ⦃c⦄ ◦ ⦃ grdb ⦄gnT
g0T , ∅
⊆
b
b
◦
◦
Notice that for all T ∈ P(Trc), the sequence gnT converges to lfp⊆
∅ λX. T ∪ ⦃c⦄ ⦃ grd ⦄X (or written otherwise: lfpT λX. ⦃c⦄ ⦃ grd ⦄X).
⊆
T
¬b
b
◦
This also means that the sequence yn converges to ⦃ grd ⦄( lfp∅ λX. T ∪ ⦃c⦄ ⦃ grd ⦄X ).
Thus, it suffices to prove that:
∀T ∈ P(Trc), ∀n ∈ N, xTn = ynT .
The proof proceeds by induction on n.
- xT0 = ∅ = y0T
- Let n ∈ N such that: ∀T ∈ P(Trc), xTn = ynT . Then:
xTn+1 = {F (n+1) (⊥)(t) ∈ Trc | t ∈ T }
= ⦃ grd¬b ⦄T ∪ {F (n) (⊥)(JcKt) ∈ Trc | σ ∈ ⦃ grdb ⦄T }
= ⦃ grd¬b ⦄T ∪ {F (n) (⊥)(t) ∈ Trc | t ∈ ⦃c⦄ ◦ ⦃ grdb ⦄T }
◦⦃ grdb
= ⦃ grd¬b ⦄T ∪ x⦃c⦄
n
⦄T
= HBy induction hypothesisI
◦⦃ grdb
⦄T
b
⦄T
⦃ grd¬b ⦄T ∪ yn⦃c⦄
= HBy definition of yn⦃c⦄
¬b
◦⦃ grd
¬b
⦄T ∪ ⦃ grd
b
◦
= ⦃ grd ⦄ T ∪ gn⦃c⦄ ⦃ grd ⦄T
[
= HBecause for all T , gnT =
⦃ grd
I
b
◦
⦄gn⦃c⦄ ⦃ grd ⦄T
¬b
0≤k≤n−1
(⦃c⦄ ◦ ⦃ grdb ⦄)(k) (T ) I
T
⦃ grd¬b ⦄gn+1
T
= yn+1
5.2 – Let us now prove that :
b
⊆
◦
lfp⊆
∅ λX. T ∪ ⦃c⦄ ⦃ grd ⦄X = lfp∅ λX.T ∪ ⦃if b then c else skip⦄X
Indeed, let the sequence (fnT )n≥0 be defined as:
f0T , ∅
T
fn+1
, T ∪ ⦃if b then c else skip⦄fnT
Therefore, by induction on n ∈ N, it holds that fn = gn :
- f0T = g0T = ∅.
- let n ∈ N, such that fnT = gnT . Then:
T
gn+1
= Σ ∪ ⦃c⦄ ◦ ⦃ grdb ⦄gnT
T
= HSince ⦃ grd¬b ⦄gnT ⊆ gnT ⊆ gn+1
I
T ∪ ⦃c⦄ ◦ ⦃ grdb ⦄gnT ∪ ⦃ grd¬b ⦄gnT
= T ∪ ⦃if b then c else skip⦄gnT
= HBy induction hypothesisI
T ∪ ⦃if b then c else skip⦄fnT
T
= fn+1
This concludes our induction on n.
Thus, by passing to the limit of both sequences, we obtain the desired result.
5.3 – Finally, we can conclude:
⦃while b do c⦄T =⦃ grd¬b ⦄ lfp⊆
T ⦃if b then c else skip⦄
= H by intermediate result 5.2 I
b
⦃ grd¬b ⦄ lfp⊆
T ⦃c⦄⦃ grd ⦄
= H by intermediate result 5.1I
{⦃while (b) do c⦄t ∈ Trc | t ∈ T }
We conclude this proof by structural induction, and Cases 1 to 5.
Theorem 1 . For all c and all T ∈ P(Trc), ⦃c⦄T is in L c M{T }.
Proof.
We prove the theorem as a corollary of this more general result:
∀T ∈ P(P(Trc)), {⦃c⦄T | T ∈ T} ⊆ LcMT
This proof proceeds by structural induction on commands. The most important case is the one for while loops; the other ones follow from
definition.
1 – Case skip :
2 – Case x := e:
LskipMT = {⦃skip⦄T | T ∈ T} ⊇ {⦃skip⦄T | T ∈ T}
3 – Case c1 ; c2 :
Lx := eMT = {⦃x := e⦄T | T ∈ T} ⊇ {⦃x := e⦄T | T ∈ T}
Lc1 ; c2 MT = Lc2 M ◦ Lc1 MT
⊇ Hby structural induction on c1 , and monotonicity of the hypercollecting semantics LcMI
Lc2 M({⦃c1 ⦄T | T ∈ T})
⊇ Hby structural induction on c2 I
{⦃c2 ⦄T 0 | T 0 ∈ {⦃c1 ⦄T | T ∈ T}}
= {⦃c2 ⦄ ◦ ⦃c1 ⦄T | T ∈ T}
= {⦃c1 ; c2 ⦄T | T ∈ T}
4 – Case if (b) then c1 else c2 :
Lif (b) then c1 else c2 MT = {⦃if (b) then c1 else c2 ⦄T | T ∈ T}
⊇ {⦃if (b) then c1 else c2 ⦄T | T ∈ T}
5 – Case while (b) do c:
Let (XTn )n∈N be the sequence defined as
n
o
XTn , {F (n) (⊥)(t) ∈ Trc | t ∈ T } | T ∈ T for n ≥ 1,
XT0 = ∅
where:
F(w)(t) ,
(
t
w
◦
JcKt
if JbKt = 0
otherwise
Notice that the limit of the sequence xTn , {F (n) (⊥)(t) ∈ Trc | t ∈ T } is the ordinary collecting semantics ⦃while (b) do c⦄T of the while
loop, as proved in Lemma 8. Thus, the sequence XTn converges to {⦃while (b) do c⦄T | T ∈ T}.
Let also (YTn )n∈N and (GTn )n∈N be the sequences defined as
GTn+1 , T ∪ Lif (b) then c else skipMGTn for n ≥ 0,
GT0 , ∅
YTn , Lgrd¬b MGTn
is the hypercollecting semantics of the while loop (Lwhile (b) do cMT).
Notice that the limit of
Thus, it suffices to prove that the sequences XTn and YTn verify the following result ∀T ∈ P(P(Trc)), ∀n ∈ N, XTn+1 ⊆ YTn+1 ; passing to
the limit in this inequality leads to the required result ∀T ∈ P(P(Trc)), {⦃while (b) do c⦄T | T ∈ T} ⊆ Lwhile (b) do cMT.
We prove the following more precise characterisation of the sequences XTn and YTn (this implies XTn+1 ⊆ YTn+1 ):
YTn
∀n ∈ N, ∀T ∈ P(P(Trc)), YTn+1 = YTn ∪ XTn+1
The remaining of this proof proceeds by induction on n ∈ N.
- case n = 0:
YT1 = Lgrd¬b MGT1
= Lgrd¬b MT
= {⦃ grd¬b ⦄T | T ∈ T}
= {{F (1) (⊥)(t) ∈ Trc | t ∈ T } | T ∈ T}
= Hsince YT0 = ∅ and by definition of XT1 I
YT0 ∪ XT1
- Let n ∈ N such that YTn+1 = YTn ∪ XTn+1 . Then:
YTn+2 = Lgrd¬b MGTn+2
= Lgrd¬b M T ∪ Lif (b) then c1 else skipMGTn+1
= Lgrd¬b MT ∪ Lgrd¬b M ◦ Lif (b) then c1 else skipMGTn+1
= H ∀n ∈ N, GTn+1 = ∪0≤k≤n Lif (b) then c1 else c2 M(k) TI
Lgrd¬b MT ∪ Lgrd¬b M ∪1≤k≤n+1 Lif (b) then c else skipM(k) T
= ∪0≤k≤n+1 Lgrd¬b M ◦ Lif (b) then c else skipM(k) T
= ∪0≤k≤n Lgrd¬b M ◦ Lif (b) then c else skipM(k) T ∪ Lgrd¬b M ◦ Lif (b) then c else skipM(n+1) T
= Lgrd¬b M ◦ ∪0≤k≤n Lif (b) then c else skipM(k) T ∪ Lgrd¬b M ◦ Lif (b) then c else skipM(n+1) T
= Lgrd¬b MGTn+1 ∪ Lgrd¬b M ◦ Lif (b) then c else skipM(n+1) T
= YTn+1 ∪ Lgrd¬b M ◦ Lif (b) then c else skipM(n+1) T
= YTn+1 ∪ {⦃ grd¬b ⦄ ◦ ⦃if (b) then c1 else skip⦄(n+1) T | T ∈ T}
= Hthe set ⦃ grd¬b ⦄ ◦ ⦃if (b) then c1 else c2 ⦄(n+1) T is the set of traces exiting the loop body after n+1 or less iterations:I
H it is equal to {F (n+2) (⊥)(t) | t ∈ T } by definition of FI
n
o
YTn+1 ∪ {F (n+2) (⊥)(t) | t ∈ T } | T ∈ T
= YTn+1 ∪ XTn+2
This concludes our induction on n.
We conclude this proof by structural induction, and Cases 1 to 5.
Appendix E.
Dependences
γdeptr
−−
−−−−
− Dep.
Lemma 9 . (αdeptr , γdeptr ) yields a Galois connection: P(P(Trc)) ←
−
−
→
α −−−
deptr
Proof.
The lattice Dep is finite, therefore complete. Thus, this is a Galois connection since it is an instance of the supremus abstraction presented
in Lemma 3.
γagree
−−−−− {tt, ff}.
The same reasoning applies for P(P(Val)) ←
−−
α−−−→
agree
The proofs of of both Lemma 4 and Theorem 2 are deferred to Appendix G: as we explain after Lemmas 6 and 7, we derive the dependence
abstract semantics as an approximation of the cardinality semantics.
Appendix F.
Cardinality Abstraction
γ
crdtr
−−
−−−− Card.
Lemma 10 . (αcrdtr , γcrdtr ) yields a Galois connection: P(P(Trc)) ←
−−
α −−→
crdtr
Proof. The lattice Card is complete, since all subsets of Card have an infimum and a supremum wrt. partial order v] , notably because the
closed interval [0, ∞] is complete wrt. partial order ≤. Thus, this is an instance of the supremus abstraction Lemma 3.
l
Lemma 5 . OC
LeM] is sound:
∀e, ∀l,
αcrdval
◦
l
˙ OC
Ol LeM ◦ γcrdtr ≤
LeM] .
Proof.
The derivation proof is by structural induction on expressions. In each case we start from the left side and derive the definition on the right
side. The interesting case is for binary arithmetic operations.
1 – Case : integer literal n
Let l ∈ L, and C ∈ Card.
Ol LnM ◦ γcrdtr (C )
= αcrdval ◦ Ol LnM {T | crdtr(T ) v] C }
αcrdval
◦
= αcrdval ∪T ∈γcrdtr (C ) {⦃n⦄R | R ⊆ T and R |=Γ l}
= αcrdval ({⦃n⦄R | R ⊆ T, R |=Γ l and T ∈ γcrdtr (C )})
≤ HNB: precision loss for simplicity of presentation, when C is bottomI
αcrdval ({{n}})
=
max crdval(V )
V ∈{{n}}
=1
l
, OC
LnM] C
l
LnM] C is being defined.
Here we use , to indicate that OC
2 – Case : variable id
Let l ∈ L, and C ∈ Card.
Ol LidM ◦ γcrdtr (C )
= αcrdval ∪T ∈γcrdtr (C ) Ol ⦃id⦄T
αcrdval
◦
= Hαcrdval preserves joinsI
max
αcrdval Ol ⦃id⦄T
T ∈γcrdtr (C )
= n H where id ; l#n ∈ αcrdtr
≤ Hαcrdtr
◦
γcrdtr (C )I
γcrdtr is reductive : αcrdtr
n H where id ; l#n ∈ C I
3 – Case : e1 ⊕ e2
Let l ∈ L, and C ∈ Card.
◦
◦
γcrdtr (C ) v] C I
l
, OC
LidM] C
αcrdval
◦
Ol Le1 ⊕ e2 M ◦ γcrdtr (C )
= αcrdval ({⦃e1 ⊕ e2 ⦄R | R ⊆ T, R |=Γ l and T ∈ γcrdtr (C )})
≤ αcrdval ({⦃e1 ⦄R | R ⊆ T, R |=Γ l and T ∈ γcrdtr (C )}) ×
αcrdval ({⦃e2 ⦄R | R ⊆ T, R |=Γ l and T ∈ γcrdtr (C )})
= αcrdval
◦
Ol Le1 M ◦ γcrdtr (C ) × αcrdval
◦
≤ HBy induction hypothesisI
Ol Le2 M ◦ γcrdtr (C )
l
l
OC
Le1 M] C × OC
Le2 M] C
l
, OC
Le1 ⊕ e2 M] C
4 – Case : e1 cmp e2
This derivation is similar to case e1 ⊕ e2 , with the difference that booleans evaluate to at most 2 different values, 1 or 0.
Ol Le1 cmp e2 M ◦ γcrdtr (C )
l
l
≤ min 2, OC
Le1 M] C × OC
Le2 M] C
αcrdval
◦
l
, OC
Le1 cmp e2 M] C
5 – Case : conclusion
We conclude by structural induction on expressions, and cases 1 to 4.
Theorem 4 . The cardinality abstract semantics is sound:
αcrdtr
◦
]
LcM ◦ γcrdtr v̇ LcM] .
Proof.
The derivation proof is by structural induction on commands. The interesting case is for conditionals.
1 – Case : skip
Let C ∈ Card.
αcrdtr
LskipM ◦ γcrdtr (C )
◦
= αcrdtr
◦
]
v Hαcrdtr
C
γcrdtr (C )
◦
γcrdtr is reductive: αcrdtr
◦
γcrdtr (C ) v] C I
, LskipM] C
2 – Case : c1 ; c2
αcrdtr
◦
Lc1 ; c2 M ◦ γcrdtr (C )
= αcrdtr
]
v
αcrdtr
◦
Lc2 M ◦ Lc1 M ◦ γcrdtr (C )
αcrdtr is extensive, Lc2 M and αcrdtr are monotoneI
Lc2 M ◦ γcrdtr ◦ αcrdtr ◦ Lc1 M ◦ γcrdtr (C )
Hγcrdtr ◦
◦
v] HBy induction hypothesisI
Lc2 M]
◦
Lc1 M] C
, Lc1 ; c2 M] C
3 – Case : id := e
3.1 – We first proceed towards an intermediate derivation:
αcrdtr
=
◦
Lid := eM ◦ γcrdtr (C )
G]
crdtr(T )
T ∈Lid:=eM◦γcrdtr (C )
G]
=
[
T ∈Lid:=eM◦γcrdtr (C )
{l ; x#n | n = αcrdval (O ⦃x⦄T ) }
l
l∈L,x∈VarP
=
[
l∈L,x∈VarP
T ∈Lid:=eM◦γcrdtr (C )
l ; x#n | n =
{l ; x#n | n = αcrdval (O ⦃x⦄T ) }
l
l∈L,x∈VarP
=
G]
[
max
T ∈Lid:=eM◦γcrdtr (C )
αcrdval (Ol ⦃x⦄T )
We now consider two cases: variables that are not modified by the assignment, and variable that are.
3.2 – Case x 6= id:
Notice that ∀l ∈ L, ∀x ∈ VarP , such that x 6= id, ∀T ∈ γcrdtr (C ):
Ol ⦃x⦄T = Ol ⦃x⦄ ⦃id := e⦄T
Thus:
max
αcrdval (Ol ⦃x⦄T )
T ∈Lid:=eM◦γcrdtr (C )
=
max
T ∈γcrdtr (C )
αcrdval (Ol ⦃x⦄T )
= Hαcrdval preserves joinsI
[
l
αcrdval
O ⦃x⦄T
T ∈γcrdtr (C )
= HBy definition of Ol LxMI
αcrdval Ol LxM (γcrdtr (C ))
= αcrdval
◦
Ol LxM ◦ γcrdtr (C )
l
≤ HBy soundness of OC
LxM] C , Lemma 5I
l
OC
LxM] C
= n where l ; x#n ∈ C
3.3 – Case x is id :
∀l ∈ L, we have :
max
αcrdval (Ol ⦃id⦄T )
T ∈Lid:=eM◦γcrdtr (C )
=
max
T ∈γcrdtr (C )
= αcrdval
◦
αcrdval (Ol ⦃e⦄T )
Ol LeM ◦ γcrdtr (C )
l
≤ HBy soundness of OC
LxM] C , Lemma 5I
l
OC
LeM] C
3.4 – Final derivation:
αcrdtr
◦
Lid := eM ◦ γcrdtr (C )
= HRecall the intermediate derivation in Case 3.1I
[
l ; x#n | n =
max
T ∈Lid:=eM◦γcrdtr (C )
l∈L,x∈VarP
αcrdval (Ol ⦃x⦄T )
v] HBy Cases 3.2 and 3.3I
n
o
[
[
l
]
l ; x#n ∈ C } ∪ {l ; id#OC LeM C
x∈VarP \{id}
l∈L
HNB: this set of constraints remains valid, owing to exclusion of id on the leftI
n
o
l
= l ; x#n ∈ C | x 6= id} ∪ {l ; id#OC
LeM] C | l ∈ L
, Lid := eM] C
4 – Case if b then c1 else c0 :
4.1 – Intermediate derivation:
αcrdtr
=
◦
Lif b then c1 else c0 M ◦ γcrdtr (C )
G]
crdtr(T )
T ∈Lif b c1 else c0 M◦γcrdtr (C )
=
G]
[
T ∈Lif b c1 else c0 M◦γcrdtr (C )
l∈L,x∈VarP
n
o
l ; x#αcrdval (O ⦃x⦄T )
l
=
G]
[
l∈L,x∈VarP
=
n
T ∈Lif b c1 else c0 M◦γcrdtr (C )
l ; x#
[
o
l ; x#αcrdval (O ⦃x⦄T )
l
l∈L,x∈VarP
max
T ∈Lif b c1 else c0 M◦γcrdtr (C )
αcrdval (Ol ⦃x⦄T )
l
4.2 – Case OC
LbM] C = 1 :
l
Let l ∈ L, and assume OC
LbM] C = 1. Let x ∈ VarP .
0
◦
∀T ∈ Lif b c1 else c0 M γcrdtr (C ), exists T ∈ γcrdtr (C ) such that T 0 = ⦃if b c1 else c0 ⦄T .
(Since LcM is not just the lifting of ⦃c⦄ to a set of sets (semantics of loops is not), in general if T 0 ∈ LcMT, we only have the existence of
T ∈ T such that T 0 ⊆ ⦃c⦄T . Here, we also rely on the fact that γcrdtr (C ) is a subset-closed. This is merely a convenient shortcut to avoid
lengthy details; it should be possible to use only the fact that T 0 ⊆ ⦃c⦄T to perform the same derivation. )
Let T 0 ∈ Lif b c1 else c0 M ◦ γcrdtr (C ), and T ∈ γcrdtr (C ) such that T 0 = ⦃if b c1 else c0 ⦄T .
l
˙ OC
Since αcrdval ◦ Ol LbM ◦ γcrdtr (C ) ≤
LbM] C (= 1), ∀R ⊆ T such that R |=Γ l, the traces r ∈ R all evaluate b to 1 or (exclusively) to
0; i.e. the sets R ⊆ T such that R |=Γ l are partitioned into the sets evaluating b to 1, and those evaluating b to 0.
Therefore, ∀R0 ⊆ T 0 such that R0 |=Γ l, exists R ∈ T and j ∈ {0, 1} such that ⦃b⦄R = {j} and ⦃cj ⦄R = R0 .
Thus,
αcrdval (Ol ⦃x⦄T 0 )
= αcrdval ({⦃x⦄R0 | R0 ⊆ T 0 and R0 |=Γ l})
[
= αcrdval
{⦃x⦄(⦃if b c1 else c0 ⦄R)}
R⊆T and R|=Γ l
= αcrdval
[
[
j∈{0,1} R⊆T and R|=Γ l and ⦃b⦄R={j}
= max αcrdval
j∈{0,1}
{⦃x⦄(⦃cj ⦄R)}
[
R⊆T and R|=Γ l and ⦃b⦄R={j}
{⦃x⦄(⦃cj ⦄R)}
≤ Hαcrdval is monotoneI
max αcrdval ◦ Ol LxM ◦ Lcj M ◦ γcrdtr (C )
j∈{0,1}
≤ Hαcrdval ◦ Ol LxM is monotone, γcrdtr ◦ αcrdtr extensive I
max αcrdval ◦ Ol LxM ◦ γcrdtr ◦ αcrdtr Lcj M ◦ γcrdtr (C )
j∈{0,1}
≤ HBy induction hypothesisI
max αcrdval ◦ Ol LxM ◦ γcrdtr
j∈{0,1}
◦
Lcj M] C
≤ HBy soundness of abstract variety, Lemma 5I
l
max OC
LxM] ◦ Lcj M] C
j∈{0,1}
= max nj where l ; x#nj ∈ Lcj M] C
j∈{0,1}
l
4.3 – Case OC
LbM] C > 1, x ∈
/ Mod(if b c1 else c0 ) :
l
Let l ∈ L, and assume OC
LbM] C > 1. Let x ∈ VarP .
Let T 0 ∈ Lif b c1 else c0 M ◦ γcrdtr (C ), and T ∈ γcrdtr (C ) such that T 0 = ⦃if b c1 else c0 ⦄T .
Notice first that if x ∈
/ Mod(if b c1 else c0 ), then:
αcrdval (Ol ⦃x⦄T 0 ) = αcrdval (Ol ⦃x⦄T )
≤ Hαcrdval is monotoneI
αcrdval
l
4.4 – Case OC
LbM] C > 1, x ∈ Mod(if b c1 else c0 ) :
l
Let l ∈ L, and assume OC
LbM] C > 1. Let x ∈ VarP .
◦
Ol LxM ◦ γcrdtr (C )
l
≤ OC
LxM] C
= n s.t l ; x#n ∈ C
Let T 0 ∈ Lif b c1 else c0 M ◦ γcrdtr (C ), and T ∈ γcrdtr (C ) such that T 0 = ⦃if b c1 else c0 ⦄T .
αcrdval (Ol ⦃x⦄T 0 )
= αcrdval (Ol ⦃x⦄ ◦ ⦃if b c1 else c0 ⦄T )
≤ αcrdval (Ol ⦃x⦄ ⦃c1 ⦄ ◦ ⦃ grdb ⦄T ∪ ⦃c0 ⦄ ◦ ⦃ grd¬b ⦄T
≤ αcrdval Ol ⦃x⦄ ◦ ⦃c1 ⦄ ◦ ⦃ grdb ⦄T + αcrdval Ol ⦃x⦄ ◦ ⦃c0 ⦄ ◦ ⦃ grd¬b ⦄T
≤ HBy monotonicity, T ∈ γcrdtr (C ) and Theorem 1I
αcrdval
◦
l
≤ OC
LxM]
Ol LxM ◦ Lc1 M ◦ Lgrd¬b M ◦ γcrdtr (C ) + αcrdval
Lc1 M]
◦
◦
l
Lgrd¬b M] C + OC
LxM]
◦
Lc0 M]
◦
◦
Ol LxM ◦ Lc0 M ◦ Lgrd¬b M ◦ γcrdtr (C )
Lgrd¬b M] C
≤ HAs a first approximation, we simply use Lgrdb M] C v] C . We refine this in Section 8I
l
OC
LxM]
◦
l
Lc1 M] C + OC
LxM]
◦
Lc2 M] C
= n1 + n2 s.t l ; x#n1 ∈ Lc1 M] C and l ; x#n2 ∈ Lc2 M] C
4.4 – Final derivation:
αcrdtr
◦
Lif b then c1 else c0 M ◦ γcrdtr (C )
= HBy the intermediate derivation in case 4.1I
[
l ; x#
max
αcrdval (Ol ⦃x⦄T )
T ∈Lif b c1 else c0 M◦γcrdtr (C )
l∈L,x∈VarP
(
≤
π l (Lc1 M] C ) t] π l (Lc2 M] C )
π l (Lc1 M] C ) t] add(if b c1 else c0 ,πl (C )) π l (Lc1 M] C )
[
l∈L
l
LbM] = 1
if OC
otherwise
with
π l (C ) , {l ; x#n ∈ C | x ∈ VarP , n ∈ [0, ∞]}
and
C1 t] add(com,C0 ) C2 ,
[
{l ; x#n | n , n1 + n2 s.t l ; x#nj ∈ Cj , j = 1, 2}
x∈Mod(com)
[
{l ; x#n ∈ C0 }
x∈VarP \ Mod(com)
5 – Case while b do c:
αcrdtr
◦
Lwhile b do cM]
= αcrdtr
◦ γcrdtr (C )
¬b
⊆
◦ Lgrd M lfp
γcrdtr (C ) Lif b then c else skipM
v] Hαcrdtr ,Lgrd¬b M are monotone, γcrdtr ◦ αcrdtr is extensiveI
αcrdtr ◦ Lgrd¬b M ◦ γcrdtr ◦ αcrdtr lfp⊆
γcrdtr (C ) Lif b then c else skipM
v] HBy assuming Lgrdb M] is soundI
Lgrd¬b M] ◦ αcrdtr lfp⊆
γcrdtr (C ) Lif b then c else skipM
v] HBy the fixpoint transfer theoremI
Lgrd¬b M]
]
◦
]
lfpv
C Lif b then c1 else c2 M
]
v] Hprecision loss for simplicity as a first approximation, Lgrd¬b M] v̇ idI
]
]
lfpv
C Lif b then c1 else c2 M
, Lwhile b do cM] C
6 – Case : conclusion
We conclude by structural induction on commands and cases 1 to 5.
Appendix G.
Appendix G.1
Dependencies reloaded
Soundness proof for dependences semantics
As noted in the text, we can derive the dependency analysis by calculation from its specification. The derivation looks similar to the one in
Appendix F for the cardinality abstraction. So here we choose a different way of proving soundness for dependency analysis. We formulate it
as an abstraction of the cardinality abstraction. This is another illustration of the benefit gained from working with hyperproperties entirely
within the framework of abstract interpretation.
This proof of soundness also implies that the cardinality abstraction is at least as precise as the type system of Hunt and Sands (Hunt and
Sands 2006) and the logic of Amtoft and Banerjee (Amtoft and Banerjee 2004), as a corollary of Theorem 3.
Lemma 6 . (αagree , γagree ) is the composition of two Galois connections (αcrdval , γcrdval ) and (αlqone , γlqone ) :
γcrdval
γlqone
crdval
lqone
−−
−−−−
− ([0, ∞] , ≤) ←
−−−−− ({tt, ff}, ⇐=)
(P(P(Val)), ⊆) ←
−−
→
−−
α −−−
α−−−→
with:
(
if n ≤ 1
, and
otherwise.
tt
ff
(
1
γlqone (bv) ,
∞
αlqone (n) ,
if bv = tt
otherwise.
Proof.
Notice that:
agree(V ) , (∀v1 , v2 ∈ V, v1 = v2 ) = (crdval(V ) ≤ 1)
Also,
αagree (V) , ∧V ∈V agree(V )
= ∧V ∈V (crdval(V ) ≤ 1)
= max crdval(V ) ≤ 1
V ∈V
= αcrdval (V) ≤ 1
= αlqone
◦
αcrdval (V)
(
where αlqone (n) ,
tt if n ≤ 1
ff otherwise
And,
γagree (bv) , {V ∈ P(Val) | agree(V ) ⇐= bv}
= {V ∈ P(Val) | crdval(V ) ≤ γlqone (bv)} where
(
1
if bv = tt
γlqone (bv) ,
∞ otherwise
= γcrdval
◦
γlqone (bv)
γlqone
−−−−− {tt, ff}:
Notice that [0, ∞] ←
−−
α−−−→
lqone
∀n ∈ [0, ∞] , ∀ bv ∈ {tt, ff},
Thus, we obtain αagree = αlqone
◦
αlqone (n) ⇐= bv iff. n ≤ γlqone (bv)
αcrdval , as well as γagree = γcrdval
◦
γlqone :
γcrdval
γlqone
crdval
lqone
−−−−− ({tt, ff}, ⇐=)
−−
−−−−
− ([0, ∞] , ≤) ←
(P(P(Val)), ⊆) ←
−−
→
−−
α −−−
α−−−→
Lemma 7 . (αdeptr , γdeptr ) is the composition of two Galois connections (αcrdtr , γcrdtr ) and (αlqonecc , γlqonecc ) :
γcrdtr
γlqonecc
crdtr
lqonecc
−−
−−−− (Card, v] ) ←
−−
−−−−−
− (Dep, v\ )
(P(P(Trc)), ⊆) ←
−−
−−
→
α −−→
α −−−−
with:
αlqonecc (C ) , {l ;Sx | l ; x#n ∈ C and αlqone (n)}
γlqonecc (D) ,
{l ; x#n | n = γlqone (l ; x ∈ D)}
l∈L,x∈VarP
Proof.
First,
αdeptr (T) =
G\
deptr(T )
T ∈T
=
G\
[
T ∈T l∈L,x∈VarP
{l ; x | αagree (Ol ⦃x⦄T )}
= HBy the decomposition in Case 1I
G\
[
{l ; x | αlqone
◦
αcrdval (Ol ⦃x⦄T )}
T ∈T l∈L,x∈VarP
=
G\
αlqonecc
T ∈T
=
G\
[
{l ; x#n | n = αcrdval (O ⦃x⦄T )}
l
l∈L,x∈VarP
Hwith αlqonecc (C ) , {l ; x | l ; x#n ∈ C and αlqone (n)}I
αlqonecc
◦
crdtr(T )
T ∈T
= Hαlqonecc preserves unionsI
!
G]
αlqonecc
crdtr(T )
T ∈T
= αlqonecc
◦
αcrdtr (T)
Also,
γdeptr (D) = {T | deptr(T ) v\ D}
= {T | αlqonecc
◦
= {T | αlqonecc
◦
crdtr(T ) v\ D}
crdtr(T ) ⊇ D}
= {T | ∀l ; x ∈ D, l ; x ∈ αlqonecc
= T | ∀l ; x ∈ D,
◦
crdtr(T )}
l ; x ∈ αlqonecc ∪l∈L,x∈VarP {l ; x#n | n = αcrdval (Ol ⦃x⦄T ) }
= {T | ∀l ; x ∈ D, αlqone (αcrdval (Ol ⦃x⦄T ))}
= {T | ∀l ; x ∈ D, αcrdval (Ol ⦃x⦄T ) ≤ 1}
= γcrdtr
◦
γlqonecc (D)
Hγlqonecc (D) ,
Therefore, we have αdeptr = αlqonecc
◦
[
l∈L,x∈VarP
{l ; x#n | n = γlqone (l ; x ∈ D)}I
αcrdtr , and γdeptr = γcrdtr
◦
γlqonecc , with:
γcrdtr
γlqonecc
crdtr
lqonecc
−−
−−−− (Card, v] ) ←
−−
−−−−−
− (Dep, v\ )
(P(P(Trc)), ⊆) ←
−−
−−
→
α −−→
α −−−−
l
Lemma 4 . OD
LeM\ is sound:
∀e, ∀l, ∀D,
αagree
◦
l
Ol LeM ◦ γdeptr (D) ⇐= OD
LeM\ D .
Proof.
l
1 – Derivation of Agreements OD
LeM\ up to l as an abstraction of cardinalities up to security level l.
αagree
◦
Ol LnM ◦ γdeptr (D)
= αlqone
Henceforth, we will derive
1.1 – Case : n
Let l ∈ L, D ∈ Dep.
l
OD
LeM\
⇐=
◦
αcrdval
◦
Ol LnM ◦ γdeptr
l
αlqone ◦ OC
LeM] ◦
◦
γlqonecc (D)
γlqonecc (D)
l
as an abstraction of cardinalities OC
LeM] . This derivation goes by structural induction on expressions.
αlqone
◦
l
OC
LnM]
◦
γlqonecc (D) = αlqone (1)
= tt
l
H, OD
LnM\ I
1.2 – Case : id Let l ∈ L, D ∈ Dep.
αlqone
◦
l
OC
LidM]
◦
γlqonecc (D) = αlqone (n) where l ; id#n ∈ γlqonecc (D)
= (l ; id ∈ D)
1.3 – Case : e1 ⊕ e2 Let l ∈ L, D ∈ Dep.
l
H, OD
LidM\ I
l
OC
Le1 ⊕ e2 M] ◦ γlqonecc (D)
l
l
= αlqone (OC
Le1 M] ◦ γlqonecc (D)) × (OC
Le2 M]
αlqone
◦
⇐= αlqone
=
◦
l
OC
Le1 M]
l
OD
Le1 M\ D
∧
γlqonecc (D) ∧ αlqone
◦
l
OD
Le2 M\ D
H,
1.4 – Case : e1 cmp e2
This case is similar to case 1.3.
1.5 – Case: conclusion
We conclude by structural induction on expressions.
l
OD
Le1
◦
◦
γlqonecc (D))
l
OC
Le2 M]
◦
γlqonecc (D)
⊕ e2 M DI
\
Theorem 2 . The dependence semantics is sound:
αdeptr
Proof.
Recall that we have αdeptr = αlqonecc
◦
◦
\
LcM ◦ γdeptr v̇ LcM\ .
αcrdtr , and γdeptr = γcrdtr
◦
γlqonecc , with:
γcrdtr
γlqonecc
crdtr
lqonecc
−−
−−−−−
− Dep
−−
−−−− Card ←
P(P(Trc)) ←
−−
→
−−
α −−−−
α −−→
Since,
αdeptr
◦
LcM ◦ γdeptr (D)
= αlqonecc
v\ αlqonecc
◦
αcrdtr
◦
LcM]
◦
◦
LcM ◦ γcrdtr
◦
γlqonecc (D)
γlqonecc (D)
We will continue the derivation of dependences abstract semantics LcM\ as an abstraction of LcM] .
We make explicit 2 derivations, for assignments and conditionals. The other cases are similar to the derivation of the cardinalities abstract
semantics.
1 – Case : id := e
Lid := eM] ◦ γlqonecc (D)
l
= αlqonecc {l ; x#n ∈ γlqonecc (D) | x 6= id} ∪ {l ; id#n | n , OC
LeM]
αlqonecc
◦
◦
γlqonecc (D), l ∈ L}
l
= {l ; x ∈ D | x 6= id} ∪ {l ; id | OD
LeM\ }
2 – Case if b then c1 else c2 :
αlqonecc
=
⇐=
◦
Lif b then c1 else c2 M]
γlqonecc (D)
◦
l
if OC
LbM]
◦
γlqonecc (D) = 1
otherwise
let D1 = Lc1 M D in
let D2 = Lc2 M\ D in
let W
( = Mod(if b c1 else c0 ) in
l
S π l (D1 ) t\ π l (D2 )
if OD
LbM\ D
l
l∈L π (D) \ {l ; x | x ∈ W } otherwise
We conclude by structural induction on commands
Appendix G.2
◦
let C1 = Lc1 M γlqonecc (D) in
let C2 = Lc2 M] ◦ γlqonecc (D) in
let W =
Mod(if
b c1 else c0 ) in
l
] l
π (C1 ) t π (C2 )
π l (C )
S
1
αlqonecc
t] add(W,πl (C ))
l∈L
π l (C2 )
]
Precision proof
Lemma 11 . For all l, l0 ∈ L, for all T ∈ P(Trc):
\
.
0
l v l0 =⇒ Ol ⦃e⦄T ⊆ Ol ⦃e⦄T
Proof.
Assume l v l0 . Then, for all R ⊆ T ,
R |=Γ l0 =⇒ R |=Γ l
Thus,
Therefore, it holds that:
{R | R ⊆ T and R |=Γ l0 } ⊆ {R | R ⊆ T and R |=Γ l}
0
Ol ⦃e⦄T ⊆ Ol ⦃e⦄T
Corollary 2 . For all l, l0 ∈ L, for all T ∈ P(P(Trc)):
0
l v l0 =⇒ Ol LeMT ⊆ Ol LeMT
Proof. This is a direct result from Lemma 11 and definition of Ol LeM.
Corollary 3 . For all l, l0 ∈ L, for all id, for all T ∈ P(P(Trc)):
l v l0 =⇒
l ; id ∈ αdeptr (T) =⇒ l0 ; id ∈ αdeptr (T)
Proof.
0
Let us assume l v l0 . By Corollary 2, we have Ol LidMT ⊆ Ol LidMT.
0
And by monotonicity of αagree , we have αagree (Ol LidMT) ⇐= αagree (Ol LidMT).
l
Thus, if l ; id ∈ αdeptr (T), then αagree (O LidMT) = tt and also
0
αagree (Ol LidMT) = tt, thus l0 ; D ∈ αdeptr (T).
Corollary 4 . For all l, l0 ∈ L, for all id, for all D ∈ Dep:
l v l0 =⇒ γdeptr (D ∪ {l ; id}) = γdeptr (D ∪ {l ; id, l0 ; id})
Proof.
1 – Note that D ∪ {l ; id, l0 ; id} ⊇ D ∪ {l ; id}, thus
D ∪ {l ; id, l0 ; id} v\ D ∪ {l ; id}
Therefore, by monotony of γdeptr :
γdeptr (D ∪ {l ; id, l0 ; id}) ⊆ γdeptr (D ∪ {l ; id})
2 – Also, let T ∈ γdeptr (D ∪ {l ; id}).
We have deptr(T ) v\ D ∪ {l ; id} by definition of γdeptr .
Thus, l ; id ∈ αdeptr (T ) and also l0 ; id ∈ αdeptr (T ) by Corollary 3. This also means that deptr(T ) v\ D ∪ {l ; id, l0 ; id}.
Finally, T ∈ γdeptr (D ∪ {l ; id, l0 ; id}) and
γdeptr (D ∪ {l ; id, l0 ; id}) ⊇ γdeptr (D ∪ {l ; id})
This concludes our proof by Case 1 and 2.
Henceforth, we will assume that all D ∈ Dep are well formed, meaning that ∀l, l0 ∈ Dep, l ; id ∈ D =⇒ l0 ; id ∈ D.
We conjecture that this can be proven for the dependence analysis we have derived: given a well formed initial set of dependence constraints,
the analysis always yields a well formed set of dependence constraints. For simplicity, we will use Corollary 4 to argue that we can still
augment any set of dependence constraints to ensure it is well formed by adding the appropriate atomic constraints. An alternative approach
would reduce the set of dependence constraints, and change slightly the abstract semantics in order to leverage Corollary 4 and guarantee the
same precision, but we refrain from doing so for simplicity.
We consider the constructive version of Hunt and Sands’ flow sensitive type system, proposed in (Hunt and Sands 2011).
Lemma 12 . For all e, D ∈ Dep, ∆ ∈ VarP → L, l ∈ L such that ∆ ` e : l, it holds that:
l
αhs (D) v̇ ∆ =⇒ OD
LeM\ D = tt
Proof.
The proof proceeds by structural induction on expressions.
1 – Case n:
l
By definition of OD
LnM\ , we have:
l
∀D, ∀l ∈ L, OD
LnM\ D = tt
2 – Case id:
By definition of the type system, we have ∆(id) = l. Thus:
αhs (D) v̇ ∆ =⇒ u{l0 | l0 ; id ∈ D} v l
=⇒ Hsince D is assumed well-formedI
l ; id ∈ D
l
=⇒ OD
LidM\ D = tt
3 – Case e1 ⊕ e2 :
By definition of the type system, there is l1 , l2 such that ∆ ` e1 : l1 and ∆ ` e2 : l2 , with l1 t l2 = l.
Thus, by induction on e1 and e2 , and assuming αhs (D) v̇ ∆, we have:
l1
l2
OD
Le1 M\ D = tt ∧ OD
Le2 M\ D = tt
Thus, since D is well formed and l1 v l and l2 v l, it holds that:
Therefore, also:
l
l
OD
Le1 M\ D = tt ∧ OD
Le2 M\ D = tt
l
OD
Le1 ⊕ e2 M\ D = tt
4 – Case e1 cmp e2 :
This case is similar to Case 3.
5 – We conclude by structural induction and Cases 1 to 4.
Let us denote by ⊥L ∈ L the bottom element of the lattice L.
Theorem 3 . For all c, D0 , D ∈ Dep, ∆0 , ∆ ∈ VarP → L, where ⊥L ` ∆0 {c}∆, and D = LcM\ D0 , it holds that:
αhs (D0 ) v̇ ∆0 =⇒ αhs (D) v̇ ∆ .
Proof.
The proof goes by structural induction on commands. The conditional case explicitly assumes that the modified variables analysis is precise
enough, to enable the simulation of the program counter. This can be achieved by collecting variable names in a while language.
1 – Case skip : this case stems from the premice.
2 – Case id := e :
Assume αhs (D0 ) v̇ ∆0 .
2.1 – Case : x 6= id
Then, for all x 6= id, ∆(x) = ∆0 (x).
Also, αhs (D)(x) = αhs (D0 )(x) v ∆0 (x) = ∆(x).
2.2 – Case : x = id
Otherwise, ∆ = ∆0 [id 7→ l], where ∆0 ` e : l.
l
By Lemma 12, since αhs (D0 ) v̇ ∆0 , we have OD
LeM\ D0 = tt.
Thus, l ; id ∈ D and :
αhs (D)(id) v ∆(id)
2.3 – Finally, by Cases 2.2 and 2.3, we have :
αhs (D) v̇ ∆
3 – Case c1 ; c2 :
This case proceeds by induction on both c1 and c2 by remarking that the type system types both command c1 and c2 in ⊥L .
4 – Case if (b) then c1 else c2 :
Assume αhs (D0 ) v̇ ∆0 . Let lb , ∆1 , ∆2 such that: lb ` ∆0 {c1 }∆1 and lb ` ∆0 {c2 }∆2 , with ∆ = ∆1 ṫ ∆2 .
Also, let ∆01 and ∆02 such that: ⊥L ` ∆0 {c1 }∆01 and ⊥L ` ∆0 {c2 }∆02 , with:
∆1 = ∆01 [id 7→ ∆01 (id) t lb , ∀id ∈ Mod(c1 )], ∆2 = ∆02 [id 7→ ∆02 (id) t lb , ∀id ∈ Mod(c2 )].
Intuitively, the program counter pc can be simulated by a modified variables analysis that is precise enough. For a while language, this can
be achieved simply by collecting variable names.
Let D1 , D2 ∈ Dep such that: Lc1 M\ D0 = D1 and Lc2 M\ D0 = D2 .
Then, assuming
W = M od(if (b) then c1 else c2 ), we have:
l
\ l
l
\
π
(D
1 ) t π (D2 ) if OD LbM D
S
l
D=
{l ; x ∈ π (D) |
l∈L
x∈
/ W}
otherwise
4.1 – By induction on c1 , we have αhs (D1 ) v̇ ∆01 .
4.2 – By induction on c2 , we have αhs (D2 ) v̇ ∆02 .
4.3 – Assume x 6∈ W , and prove αhs (D)(x) v ∆(x).
Since x 6∈ W , we have ∆(x) = ∆0 (x).
Therefore, αhs (D0 ) v̇ ∆0 implies ∆(x) ; x ∈ D0 .
Thus, since x 6∈ W , we have ∆(x) ; x ∈ D1 , and ∆(x) ; x ∈ D2 (atomic constraints related to variables not explicitly written in c1 are
not discarded from D0 , and likewise for those that are not explicitly written in c2 ). Thus, ∆(x) ; x ∈ D, meaning that αhs (D)(x) v ∆(x).
4.4 – Assume x ∈ W amd prove αhs (D)(x) v ∆(x):
We have lb v ∆(x) since x is explicitly written in one of the branches at least.
Also, by 4.1 and 4.2, we have αhs (D1 )(x) v ∆01 (x), and αhs (D2 )(x) v ∆02 (x). Meaning that
αhs (D1 t\ D2 )(x) = αhs (D1 )(x) t αhs (D2 )(x) v ∆01 (x) t ∆02 (x) v ∆(x)
Notice that ∆01 and ∆02 are well formed. Thus, exists lx such that lb v lx , such that lx ; x ∈ D1 t\ D2 , and lx v ∆(x).
l
And since ∀l ∈ L, such that lb v l, we also have OD
LbM\ D0 = tt by using Lemma 12 and D0 is well formed.
l
l
\ l
Thus, ∀l ∈ L, such that lb v l, π (D) = π (D1 ) t π (D2 ) = π l (D1 t\ D2 ), i.e. lx ; x ∈ D
Thus, αhs (D)(x) v ∆(x).
5 – Case : while (b) then c
Assume αhs (D0 ) v̇ ∆0 .
The output type environment ∆ is defined by:
∆ = lfp λ∆v .let ∆0 s.t. ⊥L t ∆v (a) ` ∆v {c}∆0 in ∆0 ṫ ∆0
Or written differently, ∆ is given by:
∆ = lfp λ∆v .let ∆0 s.t. ⊥L ` ∆v {if (b) then c else skip}∆0 in ∆0 ṫ ∆0
Let (∆n ) be the sequence defined as
∆n+1 = let ∆0 s.t. ⊥L ` ∆n {if (b) then c else skip}∆0 in ∆0 ṫ ∆0
Also, let (Dn ) be the sequence defined as Dn+1 = D0 t\ Lif (b) then c else skipM\ Dn
Then, we prove by induction on n that αhs (Dn ) v ∆n .
5.1 – Case n = 0.
This case holds by assumption αhs (D0 ) v̇ ∆0 .
5.2 – Case: Assume αhs (Dn ) v ∆n , and prove αhs (Dn+1 ) v ∆n+1 .
Let ∆0 such that ⊥L ` ∆n {if (b) then c else skip}∆0
By assumption, we have αhs (Dn ) v ∆n . Thus, by using the same proof in Case 4, we have
αhs (Lif (b) then c else skipM\ Dn ) v̇ ∆0
Therefore,
αhs (Lif (b) then c else skipM\ Dn ) ṫ αhs (D0 ) v̇ (∆0 ṫ ∆0 )
Therefore, αhs (Dn+1 ) v ∆n+1 . which proves that both least fixpoints are equal.
6 – Finally, we conclude by Cases 1–5, and structural induction on commands.
| 6cs.PL
|
1
Dynamic Topology Adaptation Based on Adaptive Link
Selection Algorithms for Distributed Estimation
arXiv:1510.05324v1 [cs.SY] 18 Oct 2015
Songcen Xu*, Rodrigo C. de Lamare, Senior Member, IEEE, and H. Vincent Poor, Fellow, IEEE
Abstract—This paper presents adaptive link selection algorithms for distributed estimation and considers their application
to wireless sensor networks and smart grids. In particular, exhaustive search–based least–mean–squares(LMS)/recursive least
squares(RLS) link selection algorithms and sparsity–inspired
LMS/RLS link selection algorithms that can exploit the topology
of networks with poor–quality links are considered. The proposed
link selection algorithms are then analyzed in terms of their stability, steady–state and tracking performance, and computational
complexity. In comparison with existing centralized or distributed
estimation strategies, key features of the proposed algorithms
are: 1) more accurate estimates and faster convergence speed
can be obtained; and 2) the network is equipped with the ability
of link selection that can circumvent link failures and improve
the estimation performance. The performance of the proposed
algorithms for distributed estimation is illustrated via simulations
in applications of wireless sensor networks and smart grids.
Index Terms—Adaptive link selection, distributed estimation,
wireless sensor networks, smart grids.
I. I NTRODUCTION
ISTRIBUTED signal processing algorithms have become
a key approach for statistical inference in wireless networks and applications such as wireless sensor networks and
smart grids [1], [2], [3], [4]. It is well known that distributed
processing techniques deal with the extraction of information
from data collected at nodes that are distributed over a
geographic area [1]. In this context, for each specific node,
a set of neighbor nodes collect their local information and
transmit the estimates to a specific node. Then, each specific
node combines the collected information together with its local
estimate to generate an improved estimate.
D
A. Prior and Related Work
Several works in the literature have proposed strategies for
distributed processing which include incremental [1], [5], [6],
[7], diffusion [2], [8], sparsity–aware [3], [9], [10], [11] and
consensus–based strategies [4]. With the incremental strategy,
the processing follows a Hamiltonian cycle, i.e., the information flows through these nodes in one direction, which means
each node passes the information to its adjacent node in a
uniform direction. However, in order to determine a cyclic
path that covers all nodes, this method needs to solve an
NP–hard problem. In addition, when any of the nodes fails,
S. Xu* is with the Communications Research Group, Department of
Electronics, University of York, YO10 5DD York, U.K. (e-mail: [email protected]).
R. C. de Lamare is with CETUC / PUC-Rio, Brazil and Department of
Electronics, University of York, U.K. (e-mail: [email protected]).
H. V. Poor is with the Department of Electrical Engineering, Princeton
University, Princeton NJ 08544 USA (e-mail: [email protected]).
Part of this work has been presented at the 2013 IEEE International
Conference on Acoustics, Speech, and Signal Processing, Vancouver, Canada
and 2013 IEEE International Workshop on Computational Advances in MultiSensor Adaptive Processing, Saint Martin. EDICS: NET-DISP, NET-ADEG,
NET-GRPH, NET-SPRS
data communication through the cycle is interrupted and the
distributed processing breaks down [1].
In distributed diffusion strategies [2], [9], the neighbors
for each node are fixed and the combining coefficients are
calculated after the network topology is deployed and starts
its operation. One disadvantage of this approach is that the
estimation procedure may be affected by poorly performing
links. More specifically, the fixed neighbors and the pre–
calculated combining coefficients may not provide an optimized estimation performance for each specified node because
there are links that are more severely affected by noise or
fading. Moreover, when the number of neighbor nodes is large,
each node requires a large bandwidth and transmit power. Prior
work on topology design and adjustment techniques includes
the studies in [12], [13] and [14], which are not dynamic in
the sense that they cannot track changes in the network and
mitigate the effects of poor links.
B. Contributions
This paper proposes and studies adaptive link selection
algorithms for distributed estimation problems. Specifically,
we develop adaptive link selection algorithms that can exploit
the knowledge of poor links by selecting a subset of data
from neighbor nodes. The first approach consists of exhaustive search–based least–mean–squares(LMS)/recursive least
squares(RLS) link selection (ES–LMS/ES–RLS) algorithms,
whereas the second technique is based on sparsity–inspired
LMS/RLS link selection (SI–LMS/SI–RLS) algorithms. With
both approaches, distributed processing can be divided into
two steps. The first step is called the adaptation step, in
which each node employs LMS or RLS to perform the adaptation through its local information. Following the adaptation
step, each node will combine its collected estimates from its
neighbors and local estimate, through the proposed adaptive
link selection algorithms. The proposed algorithms result in
improved estimation performance in terms of the mean–square
error (MSE) associated with the estimates. In contrast to
previously reported techniques, a key feature of the proposed
algorithms is that the combination step involves only a subset
of the data associated with the best performing links.
In the ES–LMS and ES–RLS algorithms, we consider all
possible combinations for each node with its neighbors and
choose the combination associated with the smallest MSE
value. In the SI–LMS and SI–RLS algorithms, we incorporate a reweighted zero attraction (RZA) strategy into the
adaptive link selection algorithms. The RZA approach is
often employed in applications dealing with sparse systems
in such a way that it shrinks the small values in the parameter vector to zero, which results in better convergence and
steady–state performance. Unlike prior work with sparsity–
aware algorithms [3], [15], [16], [17], the proposed SI–LMS
2
and SI–RLS algorithms exploit the possible sparsity of the
MSE values associated with each of the links in a different
way. In contrast to existing methods that shrink the signal
samples to zero, SI–LMS and SI–RLS shrink to zero the links
that have poor performance or high MSE values. By using
the SI–LMS and SI–RLS algorithms, data associated with
unsatisfactory performance will be discarded, which means the
effective network topology used in the estimation procedure
will change as well. Although the physical topology is not
changed by the proposed algorithms, the choice of the data
coming from the neighbor nodes for each node is dynamic,
leads to the change of combination weights and results in
improved performance. We also remark that the topology could
be altered with the aid of the proposed algorithms and a
feedback channel which could inform the nodes whether they
should be switched off or not. The proposed algorithms are
considered for wireless sensor networks and also as a tool for
distributed state estimation that could be used in smart grids.
In summary, the main contributions of this paper are:
• We present adaptive link selection algorithms for distributed estimation that are able to achieve significantly
better performance than existing algorithms.
• We devise distributed LMS and RLS algorithms with link
selection capabilities to perform distributed estimation.
• We analyze the MSE convergence and tracking performance of the proposed algorithms and their computational complexities and we derive analytical formulas to
predict their MSE performance.
• A simulation study of the proposed and existing distributed estimation algorithms is conducted along with
applications in wireless sensor networks and smart grids.
This paper is organized as follows. Section II describes the
system model and the problem statement. In Section III, the
proposed link selection algorithms are introduced. We analyze
the proposed algorithms in terms of their stability, steady–state
and tracking performance, and computational complexity in
Section IV. The numerical simulation results are provided in
Section V. Finally, we conclude the paper in Section VI.
Notation: We use boldface upper case letters to denote
matrices and boldface lower case letters to denote vectors. We
use (·)T and (·)−1 denote the transpose and inverse operators
respectively, (·)H for conjugate transposition and (·)∗ for
complex conjugate.
II. S YSTEM M ODEL AND P ROBLEM S TATEMENT
k
Nk
Fig. 1.
Network topology with N nodes
We consider a set of N nodes, which have limited processing capabilities, distributed over a given geographical
area as depicted in Fig. 1. The nodes are connected and
form a network, which is assumed to be partially connected
because nodes can exchange information only with neighbors
determined by the connectivity topology. We call a network
with this property a partially connected network whereas a
fully connected network means that data broadcast by a node
can be captured by all other nodes in the network in one hop
[18]. We can think of this network as a wireless network,
but our analysis also applies to wired networks such as power
grids. In our work, in order to perform link selection strategies,
we assume that each node has at least two neighbors.
The aim of the network is to estimate an unknown parameter
vector ω 0 , which has length M . At every time instant i, each
node k takes a scalar measurement dk (i) according to
dk (i) = ω H
0 xk (i) + nk (i),
i = 1, 2, . . . , I,
(1)
where xk (i) is the M ×1 random regression input signal vector
and nk (i) denotes the Gaussian noise at each node with zero
2
mean and variance σn,k
. This linear model is able to capture or
approximate well many input-output relations for estimation
purposes [19] and we assume I > M . To compute an estimate
of ω in a distributed fashion, we need each node to minimize
the MSE cost function [2]
2
Jωk (i) ω k (i) = E dk (i) − ω H
(2)
k (i)xk (i) ,
where E denotes expectation and ωk (i) is the estimated vector
generated by node k at time instant i. Equation (2) is also
the definition of the MSE. To solve this problem, diffusion
strategies have been proposed in [2], [8] and [20]. In these
strategies, the estimate for each node is generated through a
fixed combination strategy given by
X
ω k (i) =
ckl ψ l (i),
(3)
l∈Nk
where Nk denotes the set of neighbors of node k including
node k itself, ckl ≥ 0 is the combining coefficient and ψ l (i)
is the local estimate generated by node l through its local
information.
There are many ways to calculate the combining coefficient
ckl which include the Hastings [21], the Metropolis [22], the
Laplacian [23] and the nearest neighbor [24] rules. In this
work, due to its simplicity and good performance we adopt
the Metropolis rule [22] given by
max{|N1 |,|N |} , if k 6= l are linked
k
l
P
(4)
ckl =
ckl , for k = l.
1−
l∈Nk /k
where |Nk | denotes the cardinality of Nk .
The set of coefficients ckl should satisfy [2]
X
ckl = 1.
(5)
l∈Nk ∀k
For the combination strategy mentioned in (3), the choice
of neighbors for each node is fixed, which results in some
problems and limitations, namely:
• Some nodes may face high levels of noise or interference,
which may lead to inaccurate estimates.
• When the number of neighbors for each node is high,
large communication bandwidth and high transmit power
are required.
3
Some nodes may shut down or collapse due to network
problems. As a result, local estimates to their neighbors
may be affected.
Under such circumstances, a performance degradation is likely
to occur when the network cannot discard the contribution
of poorly performing links and their associated data in the
estimation procedure. In the next section, the proposed adaptive link selection algorithms are presented, which equip a
network with the ability to improve the estimation procedure.
In the proposed scheme, each node is able to dynamically
select the data coming from its neighbors in order to optimize
the performance of distributed estimation techniques.
TABLE I
T HE ES-LMS A LGORITHM
•
III. P ROPOSED A DAPTIVE L INK S ELECTION A LGORITHMS
In this section, we present the proposed adaptive link
selection algorithms. The goal of the proposed algorithms is
to optimize the distributed estimation and improve the performance of the network by dynamically changing the topology.
These algorithmic strategies give the nodes the ability to
choose their neighbors based on their MSE performance. We
develop two categories of adaptive link selection algorithms;
the first one is based on an exhaustive search, while the second
is based on a sparsity–inspired relaxation. The details will be
illustrated in the following subsections.
A. Exhaustive Search–Based LMS/RLS Link Selection
The proposed ES–LMS and ES–RLS algorithms employ
an exhaustive search to select the links that yield the best
performance in terms of MSE. First, we describe how we
define the adaptation step for these two strategies. In the ES–
LMS algorithm, we employ the adaptation strategy given by
∗
ψ k (i) = ω k (i − 1) + µk xk (i) dk (i) − ωH
k (i − 1)xk (i) , (6)
where µk is the step size for each node. In the ES–RLS
algorithm, we employ the following steps for the adaptation:
φ−1 (i) =
λ−1 φ−1 (i − 1)
λ−2 φ−1 (i − 1)x(i)xH (i)φ−1 (i − 1)
−
, (7)
1 + λ−1 xH (i)φ−1 (i − 1)x(i)
where λ is the forgetting factor. Then, we let
P (i) = φ−1 (i)
(8)
and
λ−1 P (i)x(i)
.
(9)
1 + λ−1 xH (i)P (i)x(i)
∗
ψ k (i) = ω k (i − 1) + k(i) dk (i) − ωH
k (i − 1)xk (i) , (10)
k(i) =
P (i + 1) = λ−1 P (i) − λ−1 k(i)xH (i)P (i).
(11)
Following the adaptation step, we introduce the combination
step for both ES–LMS and ES–RLS algorithms, based on an
exhaustive search strategy. At first, we introduce a tentative
set Ωk using a combinatorial approach described by
Ωk ∈ 2Nk \∅,
(12)
where the set Ωk is a nonempty set with 2Nk elements. After
the tentative set Ωk is defined, we write the cost function (2)
for each node as
2
Jψ(i) ψ(i) , E dk (i) − ψ H (i)xk (i) ,
(13)
Initialize: ω k (0)=0
For each time instant i=1,2, . . . , I
For each node k=1,2, . . . , N
∗
ψ k (i) = ω k (i − 1) + µk xk (i)[dk (i) − ω H
k (i − 1)xk (i)]
end
For each node k=1,2, . . . , N
find all possible sets of Ωk
P
ckl (i)ψ l (i)]H xk (i)
eΩk (i) = dk (i) − [
l∈Ωk
b k (i) = arg min |eΩ (i)|
Ω
k
PΩk
ω k (i) =
ckl (i)ψ l (i)
end
end
b (i)
l∈Ω
k
where
ψ(i) ,
X
ckl (i)ψ l (i)
(14)
l∈Ωk
is the local estimator and ψ l (i) is calculated through (6) or
(10), depending on the algorithm, i.e., ES–LMS or ES–RLS.
With different choices of the set Ωk , the combining coefficients
ckl will be re–calculated through (4), to ensure condition (5)
is satisfied.
Then, we introduce the error pattern for each node, which
is defined as
X
H
eΩk (i) , dk (i) −
ckl (i)ψ l (i) xk (i).
(15)
l∈Ωk
For each node k, the strategy that finds the best set Ωk (i) must
solve the following optimization problem:
Ω̂k (i) = arg
min
Ωk ∈2Nk \∅
|eΩk (i)|.
(16)
After all steps have been completed, the combination step in
(3) is performed as described by
X
ω k (i) =
ckl (i)ψ l (i).
(17)
b k (i)
l∈Ω
At this stage, the main steps of the ES–LMS and ES–RLS
algorithms have been completed. The proposed ES–LMS and
b k (i) that minimizes the error
ES–RLS algorithms find the set Ω
pattern in (15) and (16) and then use this set of nodes to
obtain ωk (i) through (17). The ES–LMS/ES–RLS algorithms
are briefly summarized as follows:
Step 1 Each node performs the adaptation through its local
information based on the LMS or RLS algorithm.
Step 2 Each node finds the best set Ωk (i), which satisfies
(16).
Step 3 Each node combines the information obtained from
its best set of neighbors through (17).
The details of the proposed ES–LMS and ES–RLS algorithms
are shown in Tables I and II. When the ES–LMS and ES–RLS
algorithms are implemented in networks with a large number
of small and low–power sensors, the computational complexity
cost may become high, as the algorithm in (16) requires an
exhaustive search and needs more computations to examine
all the possible sets Ωk (i) at each time instant.
4
TABLE II
T HE ES-RLS A LGORITHM
Initialize: ω k (0)=0
phi−1 (0) = δ −1 I, δ = small positive constant
For each time instant i=1,2, . . . , I
For each node k=1,2, . . . , N
φ−1 (i) = λ−1 φ−1 (i − 1)
−1
λ−2 φ−1 (i − 1)xk (i)xH
(i − 1)
k (i)φ
−
H
−1
−1
1 + λ xk (i)φ (i − 1)xk (i)
P (i) = φ−1 (i)
λ−1 P (i)xk (i)
k(i) =
1 + λ−1 xH
k (i)P (i)xk (i)
∗
ψ k (i) = ω k (i − 1) + k(i)[dk (i) − ω H
k (i − 1)xk (i)]
−1
P (i + 1) = λ P (i) − λ−1 k(i)xH
(i)P
(i)
k
end
For each node k=1,2, . . . , N
find all possible sets ofP
Ωk
ckl (i)ψ l (i)]H xk (i)
eΩk (i) = dk (i) − [
l∈Ωk
end
end
b k (i) = arg min |eΩ (i)|
Ω
k
PΩk
ω k (i) =
ckl (i)ψ l (i)
l∈Nk
b (i)
l∈Ω
k
B. Sparsity–Inspired LMS/RLS Link Selection
The ES–LMS/ES–RLS algorithms previously outlined need
to examine all possible sets to find a solution at each time
instant, which might result in high computational complexity
for large networks operating in time–varying scenarios. To
solve the combinatorial problem with reduced complexity,
we propose sparsity-inspired based SI–LMS and SI–RLS
algorithms, which are as simple as standard diffusion LMS or
RLS algorithms and are suitable for adaptive implementations
and scenarios where the parameters to be estimated are slowly
time–varying. The zero–attracting strategy (ZA), reweighted
zero–attracting strategy (RZA) and zero–forcing (ZF) are
reported in [25], [3], [26], [27], [28], [29], [30], [31], [32],
[33], [34], [35], [36], [37], [38], [39], [40] as for sparsity
aware techniques. These approaches are usually employed in
applications dealing with sparse systems in scenarios where
they shrink the small values in the parameter vector to zero,
which results in better convergence rate and steady–state
performance. Unlike existing methods that shrink the signal
samples to zero, the proposed SI–LMS and SI–RLS algorithms
shrink to zero the links that have poor performance or high
MSE values. To detail the novelty of the proposed sparsity–
inspired LMS/RLS link selection algorithms, we illustrate the
processing in Fig.2.
MSE Value
MSE Value
Sparsity Aware
Technique
Nodes
a ) Sparsity Aware Technique
MSE Value
Nodes
MSE Value
SILMS/SIRLS
Algorithms
Nodes
b ) SILLS and SIRLS Algorithms
Fig. 2.
in Fig. 2 (b). In the following, we will show how the proposed
SI–LMS/SI–RLS algorithms are employed to realize the link
selection strategy automatically.
In the adaptation step, we follow the same procedure in
(6)–(10) as that of the ES–LMS and ES–RLS algorithms for
the SI–LMS and SI–RLS algorithms, respectively. Then we
reformulate the combination step. First, we introduce the log–
sum penalty into the combination step in (3). Different penalty
terms have been considered for this task. We have adopted
a heuristic approach [3], [41] known as reweighted zero–
attracting strategy into the combination step in (3) because this
strategy has shown an excellent performance and is simple
to implement. The regularization function with the log–sum
penalty is defined as:
X
f1 (ek (i)) =
log 1 + ε|ekl (i)| ,
(18)
Nodes
Sparsity aware signal processing strategies
Fig. 2 (a) shows a standard type of sparsity–aware processing. We can see that, after being processed by a sparsity–
aware algorithm, the nodes with small MSE values will be
shrunk to zero. In contrast, the proposed SI–LMS and SI–RLS
algorithms will keep the nodes with lower MSE values and
shrink the nodes with large MSE values to zero as illustrated
where the error pattern ekl (i)(l ∈ Nk ), which stands for the
neighbor node l of node k including node k itself, is defined
as
ekl (i) , dk (i) − ψ H
(19)
l (i)xk (i)
and ε is the shrinkage magnitude. Then, we introduce the vector and matrix quantities required to describe the combination
step. We first define a vector ck that contains the combining
coefficients for each neighbor of node k including node k itself
as described by
ck , [ckl ],
l ∈ Nk .
(20)
Then, we define a matrix Ψk that includes all the estimated
vectors, which are generated after the adaptation step of SI–
LMS and of SI–RLS for each neighbor of node k including
node k itself as given by
Ψk , [ψ l (i)],
l ∈ Nk .
(21)
Note that the adaptation steps of SI–LMS and SI–RLS are
identical to (6) and (10), respectively. An error vector êk
that contains all error values calculated through (19) for each
neighbor of node k including node k itself is expressed by
êk , [ekl (i)],
l ∈ Nk .
(22)
Here, we use a hat to distinguish êk defined above from the
original error ek . To devise the sparsity–inspired approach, we
have modified the vector êk in the following way:
1) The element with largest absolute value |ekl (i)| in êk
will be employed as |ekl (i)|.
2) The element with smallest absolute value will be set
to −|ekl (i)|. This process will ensure the node with
smallest error pattern has a reward on its combining
coefficient.
3) The remaining entries will be set to zero.
At this point, the combination step can be defined as [41]
ωk (i) =
|Nk |
X
j=1
∂f1 (êk [j])
Ψk [j],
ck [j] − ρ
∂êk [j]
(23)
where ck [j], êk [j] and Ψk [j] stand for the jth element in
the ck , êk and Ψk . The parameter ρ is used to control the
5
algorithm’s shrinkage intensity. We then calculate the partial
derivative of êk [j]:
∂ log(1 + ε|ekl (i)|)
∂f1 (êk [j])
=
∂êk [j]
∂ ekl (i)
sign(ekl (i))
l ∈ Nk
=ε
1 + ε|ekl (i)|
sign(êk [j])
=ε
.
(24)
1 + ε|êk [j]|
|N
Pk |
∂f1 (êk [j])
ck [j] − ρ ∂êk [j]
To ensure that
= 1, we replace
j=1
êk [j] with ξmin in the denominator, where the parameter ξmin
stands for the minimum absolute value of ekl (i) in êk . Then,
(24) can be rewritten as
TABLE III
T HE SI-LMS AND SI-RLS A LGORITHMS
Initialize: ω k (−1)=0
P (0) = δ −1 I, δ = small positive constant
For each time instant i=1,2, . . . , I
For each node k=1,2, . . . , N
The adaptation step for computing ψ k (i)
is exactly the same as the ES-LMS and ES-RLS
for the SI-LMS and SI-RLS algorithms respectively
end
For each node k=1,2, . . . , N
ekl (i) = dk (i) − xH
l ∈ Nk
k (i)ψ l (i)
ck = [ckl ] l ∈ Nk
Ψk = [ψ l (i)] l ∈ Nk
ek = [ekl (i)] l ∈ Nk
Find the maximum and minimum absolute terms in ek
Modified ek as ek =[0· · ·0,|ekl (i)|,0· · ·0,−|ekl (i)|,0· · ·0]
| {z }
| {z }
max
min
ξmin = min |ekl (i)|
|Nk |
P
sign(e [j])
ck [j] − ρε 1+ε|ξk | Ψk [j]
ω k (i) =
j=1
∂f1 (êk [j])
sign(êk [j])
=ε
.
∂êk [j]
1 + ε|ξmin |
(25)
At this stage, the MSE cost function governs the adaptation
step, while the combination step employs the combining
coefficients with the derivative of the log-sum penalty which
performs shrinkage and selects the set of estimates from
the neighbor nodes with the best performance. The function
sign(a) is defined as
a/|a|
a 6= 0
sign(a) =
(26)
0
a = 0.
Then, by inserting (25) into (23), the proposed combination
step is given by
|Nk |
X
sign(êk [j])
ω k (i) =
ck [j] − ρε
Ψk [j].
(27)
1 + ε|ξmin |
j=1
k [j])
Note that the condition ck [j] − ρε sign(ê
1+ε|ξmin | ≥ 0 is enforced
k [j])
in (27). When ck [j] − ρε sign(ê
1+ε|ξmin | = 0, it means that the
corresponding node has been discarded from the combination
step. In the following time instant, if this node still has the
largest MSE, there will be no changes in the combining
coefficients for this set of nodes.
To guarantee the stability, the parameter ρ is assumed to
be sufficiently small and the penalty takes effect only on
the element in êk for which the magnitude is comparable
to 1/ε [3]. Moreover, there is little shrinkage exerted on the
element in êk whose |êk [j]| ≪ 1/ε. The SI–LMS and SI–
RLS algorithms perform link selection by the adjustment of
the combining coefficients through (27). At this point, it should
be emphasized that:
• The process in (27) satisfies condition (5), as the penalty
and reward amounts are the same for the nodes with
maximum and minimum error pattern, respectively.
• When computing (27), there are no matrix–vector multiplications. Therefore, no additional complexity is introduced. As described in (23), only the jth element in
ck , êk and Ψk are used for calculation.
For the neighbor node with the largest MSE value, after the
modifications of êk , its ekl (i) value in ek will be a positive
k [j])
number which will lead to the term ρε sign(ê
1+ε|ξmin | in (27) being
positive too. This means that the combining coefficient for this
min
end
end
node will be shrunk and the weight for this node to build ω k (i)
will be shrunk too. In other words, when a node encounters
high noise or interference levels, the corresponding MSE value
might be large. As a result, we need to reduce the contribution
of this group of nodes.
In contrast, for the neighbor node with the smallest MSE,
as its ekl (i) value in ek will be a negative number, the term
k [j])
ρε sign(ê
1+ε|ξmin | in (27) will be negative too. As a result, the
weight for this node associated with the smallest MSE to build
ωk (i) will be increased. For the remaining neighbor nodes, the
k [j])
entry ekl (i) in ek is zero, which means the term ρε sign(ê
1+ε|ξmin |
in (27) is zero and there is no change for the weights to build
ωk (i). The main steps for the proposed SI–LMS and SI–RLS
algorithms are listed as follows:
Step 1 Each node carries out the adaptation through its local
information based on the LMS or RLS algorithm.
Step 2 Each node calculates the error pattern through (19).
Step 3 Each node modifies the error vector ek .
Step 4 Each node combines the information obtained from
its selected neighbors through (27).
The SI–LMS and SI–RLS algorithms are detailed in Table
III. For the ES–LMS/ES–RLS and SI–LMS/SI–RLS algorithms, we design different combination steps and employ
the same adaptation procedure, which means the proposed
algorithms have the ability to equip any diffusion–type wireless networks operating with other than the LMS and RLS
algorithms. This includes, for example, the diffusion conjugate
gradient strategy [42].
IV. A NALYSIS OF THE PROPOSED ALGORITHMS
In this section, a statistical analysis of the proposed algorithms is developed, including a stability analysis and an
MSE analysis of the steady–state and tracking performance.
In addition, the computational complexity of the proposed
algorithms is also detailed. Before we start the analysis, we
make some assumptions that are common in the literature [19].
Assumption I: The weight-error vector εk (i) and the input
signal vector xk (i) are statistically independent, and the
weight–error vector for node k is defined as
εk (i) , ωk (i) − ω0 ,
(28)
6
where ω 0 denotes the optimum Wiener solution of the actual
parameter vector to be estimated, and ωk (i) is the estimate
produced by the proposed algorithms at time instant i.
Assumption II: The input signal vector xl (i) is drawn from
a stochastic process, which is ergodic in the autocorrelation
function [19].
Assumption III: The M × 1 vector q(i) represents a stationary sequence of independent zero–mean vectors and positive
definite autocorrelation matrix Q = E[q(i)q H (i)], which is
independent of xk (i), nk (i) and εl (i).
Assumption IV (Independence): All regressor input signals
xk (i) are spatially and temporally independent.
A. Stability Analysis
In general, to ensure that a partially-connected network
can converge to the global network performance, information
should be propagated across the network [43]. The work in
[12] shows that it is central to the performance that each node
should be able to reach the other nodes through one or multiple
hops [43]. In this section, we discuss the stability analysis of
the proposed ES–LMS and SI–LMS algorithms.
First, we substitute (6) into (17) and obtain
X
ω k (i + 1) =
ckl (i)ψ l (i + 1)
We also define an N × N matrix C where the combining
coefficients {ckl } correspond to the {l, k} entries of the matrix
C and the N M ×N M matrix C G with a Kronecker structure:
CG = C ⊗ IM
(35)
where ⊗ denotes the Kronecker product.
By inserting el (i + 1) = e0−l − εH
l (i)xl (i + 1) into (33),
the global version of (30) can then be written as
ε(i + 1) = C TG I − M D(i + 1) ε(i) + C TG M g(i + 1), (36)
where e0−l is the estimation error produced by the Wiener
filter for node l as described by
e0−l = dl (i) − ω H
0 xl (i).
(37)
D , E[D(i + 1)]
= diag{R1 , ..., RN }
(38)
If we define
and take the expectation of (36), we arrive at
E{ε(i + 1)} = C TG I − M D E{ε(i)}.
(39)
Before we proceed, let us define X = I − M D and introduce
Lemma 1:
Lemma 1: Let C G and X denote arbitrary N M × N M mab
l∈Ωk (i)
X
trices,
where C G has real, non-negative entries, with columns
=
[ωl (i) + µl xl (i + 1)e∗l (i + 1)]ckl (i)
adding up to one. Then, the matrix Y = C TG X is stable for
b k (i)
l∈Ω
any choice of C G if and only if X is stable.
X
∗
Proof : Assume that X is stable, it is true that for every
=
[ω0 + εl (i) + µl xl (i + 1)el (i + 1)]ckl (i)
square
matrix X and every α > 0, there exists a submultib k (i)
l∈Ω
plicative matrix norm || · ||τ that satisfies ||X||τ ≤ τ (X) + α,
X
X
=
ω 0 ckl +
[εl (i) + µl xl (i + 1)e∗l (i + 1)]c
where
kl (i)the submultiplicative matrix norm ||AB|| ≤ ||A||·||B||
b k (i)
b k (i)
holds and τ (X) is the spectral radius of X [44], [45]. Since
l∈Ω
l∈Ω
X
X is stable, τ (X) < 1, and we can choose α > 0 such that
subject to
ckl (i) = 1
τ (X) + α = v < 1 and ||X||τ ≤ v < 1. Then we obtain [8]
l
X
||Y i ||τ = ||(C TG X)i ||τ
= ω0 +
[εl (i) + µl xl (i + 1)e∗l (i + 1)]ckl (i).
b k (i)
l∈Ω
(29)
≤ ||(C TG )i ||τ · ||X i ||τ
≤v
i
(40)
||(C TG )i ||τ .
Then, we have
Since C TG has non–negative entries with columns that add up
to one, it is element–wise bounded by unity. This means its
εk (i + 1) = ω0 − ω 0 +
[εl (i) + µl xl (i + 1)e∗l (i + 1)]ckl (i)
Frobenius norm is bounded as well and by the equivalence of
b k (i)
l∈Ω
X
norms, so is any norm, in particular ||(C TG )i ||τ . As a result,
=
[εl (i) + µl xl (i + 1)e∗l (i + 1)]ckl (i).
we have
b k (i)
l∈Ω
lim ||Y i ||τ = 0,
(41)
i→∞
(30)
where Y i converges to the zero matrix for large i. This
By employing Assumption IV, we start with (30) for the ES–
establishes the stability of (C TG X)i .
LMS algorithm and define the global vectors and matrices:
A square matrix X is stable if it satisfies X i → 0 as i →
T
ε(i + 1) , [ε1 (i + 1), · · · , εN (i + 1)]
(31) ∞. In view of Lemma 1 and (82), we need the matrix I −M D
to be stable. As a result, it requires I − µk Rk to be stable for
M , diag{µ1 I M , ..., µN I M }
(32) all k, which holds if the following condition is satisfied:
X
H
D(i + 1) , diag{x1 (i + 1)xH
1 (i + 1), ..., xN (i + 1)xN (i + 1)}
(33)
and the N M × 1 vector
g(i + 1) = [x1 (i + 1)n1 (i + 1), · · · , xN (i + 1)nN (i + 1)]T .
(34)
0 < µk <
2
λmax Rk
(42)
where λmax Rk is the largest eigenvalue of the correlation
matrix Rk . The difference between the ES–LMS and SI–
LMS algorithms is the strategy to calculate the matrix C.
7
Lemma 1 indicates that for any choice of C, only X needs
to be stable. As a result, SI–LMS has the same convergence
condition as in (42). Given the convergence conditions, the
proposed ES–LMS/ES–RLS and SI–LMS/SI–RLS algorithms
will adapt according to the network connectivity by choosing
the group of nodes with the best available performance to
construct their estimates. Comparing the results in (42) with
the existing algorithms, it can be seen that the proposed link
selection techniques change the set of combining coefficients,
which are indicated in C G , as the matrix C employs the
b k (i).
chosen set Ω
=
ek (i) = dk (i) − ωH
k (i)xk (i)
= dk (i) − [ω 0 − εk (i)]H xk (i)
H
= dk (i) − ωH
0 xk (i) + εk (i)xk (i)
= e0−k + εH
k (i)xk (i).
With Assumption I, the MSE expression can be derived as
2
Jmse−k (i) = E[|ek (i)| ]
∗
H
= E e0−k + εH
(i)x
(i)
e
+
x
(i)ε
(i)
k
k
k
0
k
H
= Jmin−k + E[εH
k (i)xk (i)xk (i)εk (i)]
= Jmin−k +
H
tr{E[εk (i)εH
k (i)xk (i)xk (i)]}
H
tr{E[xk (i)xk (i)]E[εk (i)εH
k (i)]}
= Jmin−k +
= Jmin−k + tr{Rk (i)K k (i)},
(44)
where tr(·) denotes the trace of a matrix and Jmin−k is the
minimum mean–square error (MMSE) for node k [19]:
−1
2
Jmin−k = σd,k
− pH
k (i)Rk (i)pk (i),
(45)
Rk (i) = E[xk (i)xH
k (i)] is the correlation matrix of the inputs
for node k, pk (i) = E[xk (i)d∗k (i)] is the cross–correlation
vector between the inputs and the measurement dk (i), and
K k (i) = E[εk (i)εH
k (i)] is the weight–error correlation matrix.
From [19], the EMSE is defined as the difference between the
mean–square error at time instant i and the minimum mean–
square error. Then, we can write
Jex−k (i) = Jmse−k (i) − Jmin−k
= tr{Rk (i)K k (i)}.
(46)
For the proposed adaptive link selection algorithms, we will
derive the EMSE formulas separately based on (46) and
we assume that the input signal is modeled as a stationary
process.
1) ES–LMS: To update the estimate ω k (i), we employ
X
ω k (i + 1) =
ckl (i)ψ l (i + 1)
=
b k (i)
l∈Ω
X
b k (i)
l∈Ω
ckl (i)[ω l (i) + µl xl (i + 1)e∗l (i + 1)]
− xH
l (i + 1)ω l (i))].
(47)
Then, subtracting ω 0 from both sides of (47), we arrive at
X
εk (i + 1) =
ckl (i)[ωl (i) + µl xl (i + 1)(dl (i + 1)
b k (i)
l∈Ω
=
X
b k (i)
l∈Ω
=
X
b k (i)
l∈Ω
(43)
ckl (i)[ωl (i) + µl xl (i + 1)(dl (i + 1)
b k (i)
l∈Ω
B. MSE Steady–State Analysis
In this part of the analysis, we devise formulas to predict
the excess MSE (EMSE) of the proposed algorithms. The error
signal at node k can be expressed as
X
− xH
l (i + 1)ω l (i))] −
X
ckl (i)ω 0
b k (i)
l∈Ω
ckl (i) εl (i) + µl xl (i + 1) dl (i + 1)
−
xH
l (i
+ 1)(εl (i) + ω 0 )
ckl (i) εl (i) + µl xl (i + 1) dl (i + 1)
−
xH
l (i
xH
l (i
+ 1)εl (i) −
+ 1)ω0
X
=
ckl (i) εl (i) − µl xl (i + 1)xH
l (i + 1)εl (i)
b k (i)
l∈Ω
+ µl xl (i + 1)e∗0−l (i + 1)
X
=
ckl (i) I − µl xl (i + 1)xH
l (i + 1) εl (i)
b k (i)
l∈Ω
+ µl xl (i + 1)e∗0−l (i + 1) .
Let us introduce the random variables αkl (i):
b k (i)
1, if l ∈ Ω
αkl (i) =
0, otherwise.
(48)
(49)
At each time instant, each node will generate data associated
with network covariance matrices Ak with size N × N which
reflect the network topology, according to the exhaustive
search strategy. In the network covariance matrices Ak , a value
equal to 1 means nodes k and l are linked and a value 0 means
nodes k and l are not linked.
For example, suppose a network has 5 nodes. For node
3, there are two neighbor nodes, namely, nodes 2 and 5.
Through Eq. (12), the possible configurations of set Ω3 are
{3, 2}, {3, 5} and {3, 2, 5}. Evaluating all the possible sets
for Ω3 , the relevant covariance matrices A3 with size 5 × 5
at time instant i are described in Table IV.
Then, the coefficients αkl are obtained according to the
covariance matrices Ak . In this example, the three sets of
αkl are respectively shown in Table V.
The parameters ckl will then be calculated through Eq. (4)
for different choices of matrices Ak and coefficients αkl . After
αkl and ckl are calculated, the error pattern for each possible
Ωk will be calculated through (15) and the set with the smallest
error will be selected according to (16).
With the newly defined αkl , (48) can be rewritten as
X
εk (i + 1) =
αkl (i)ckl (i) I − µl xl (i + 1)xH
l (i + 1) εl (i)
l∈Nk
8
TABLE IV
C OVARIANCE MATRICES A3 FOR DIFFERENT SETS OF Ω3
(a) {3, 2}
2
3 4
0
1
0
1
1 0
0
0
1
1
2
3
4
5
(c) {3, 2, 5}
2
3 4
0
1
0
1
1 0
0
1
1
1
2
3
4
5
5
(b) {3, 5}
2
3
4
0
0
0 0
1
0
0
1
1
1
2
3
4
5
0
5
1
where Rl,q (i + 1) = E[xl (i + 1)xH
q (i + 1)] and K l,q (i) =
E[εl (i)εH
q (i)]. To further simplify the analysis, we assume that
the samples of the input signal xk (i) are uncorrelated, i.e.,
2
2
Rk = σx,k
I with σx,k
being the variance. Using the diagonal
2
matrices Rk = Λk = σx,k
I and Rl,q = Λl,q = σx,l σx,q I we
can write
X
K k (i + 1) =
E α2kl (i)c2kl (i)
I − µl Λl K l (i) I − µl Λl
l∈Nk
+ µ2l e0−l (i + 1)e∗0−l (i + 1)Λl
X
+
E αkl (i)αkq (i)ckl (i)ckq (i)
5
1
l,q∈Nk
l6=q
TABLE V
C OEFFICIENTS αkl
(a) {3, 2}
α31 = 0
α32 = 1
α33 = 1
α34 = 0
α35 = 0
FOR DIFFERENT SETS OF
(b) {3, 5}
α31 = 0
α32 = 0
α33 = 1
α34 = 0
α35 = 1
×
Ω3
l,q∈Nk
l6=q
× I − µl Λl + µl µq e0−q (i +
(50)
Starting from (46), we then focus on K k (i + 1).
K k (i + 1) = E[εk (i + 1)εH
k (i + 1)].
(51)
In (50), the term αkl (i) is determined through the network
topology for each subset, while the term ckl (i) is calculated
through the Metropolis rule. We assume that αkl (i) and ckl (i)
are statistically independent from the other terms in (50). Upon
convergence, the parameters αkl (i) and ckl (i) do not vary
b k (i) for
because at steady state the choice of the subset Ω
each node k will be fixed. Then, under these assumptions,
substituting (50) into (51) we arrive at:
X
K k (i + 1) =
E α2kl (i)c2kl (i)
I − µl Rl (i + 1) K l (i)
l∈Nk
l,q∈Nk
l6=q
× I − µl Rl (i + 1) K l,q (i) I − µq Rl (i + 1)
+ µl µq e0−l (i + 1)e∗0−q (i + 1)Rl,q (i + 1)
X
+
E αkl (i)αkq (i)ckl (i)ckq (i)
l,q∈Nk
l6=q
I − µq Rq (i + 1) K H
l,q (i) I − µl Rl (i + 1)
H
∗
+ µl µq e0−q (i + 1)e0−l (i + 1)Rl,q (i + 1) (52)
×
1)e∗0−l (i
+
1)ΛH
l,q
.
(53)
Due to the structure of the above equations, the approximations
and the quantities involved, we can decouple (53) into
X
1 − µl λnl Kln (i) 1 − µl λnl
Kkn (i + 1) =
E α2kl (i)c2kl (i)
l∈Nk
+ µ2l e0−l (i + 1)e∗0−l (i + 1)λnl
X
n
+
(i)
E αkl (i)αkq (i)ckl (i)ckq (i)
1 − µl λnl Kl,q
l,q∈Nk
l6=q
× 1−
× I − µl Rl (i + 1) + µ2l e0−l (i + 1)e∗0−l (i + 1)
X
× Rl (i + 1) +
E αkl (i)αkq (i)ckl (i)ckq (i)
I − µl Λl K l,q (i)
× I − µq Λq + µl µq e0−l (i + 1)e∗0−q (i + 1)Λl,q
X
+
E αkl (i)αkq (i)ckl (i)ckq (i)
I − µq Λq K H
l,q (i)
(c) {3, 2, 5}
α31 = 0
α32 = 1
α33 = 1
α34 = 0
α35 = 1
∗
+ µl xl (i + 1)e0−l (i + 1) .
1)e∗0−q (i
1)λnl,q
1)e∗0−l (i
1)λnl,q
µq λnq
+ µl µq e0−l (i +
+
X
n H
+
) (i)
E αkl (i)αkq (i)ckl (i)ckq (i)
1 − µq λnq (Kl,q
l,q∈Nk
l6=q
× 1−
µl λnl
+ µl µq e0−q (i +
+
,
(54)
where Kkn (i + 1) is the nth element of the main diagonal of
K k (i + 1). With the assumption that αkl (i) and ckl (i) are
statistically independent from the other terms in (50), we can
rewrite (54) as
X
2
E α2kl (i) E c2kl (i)
Kkn (i + 1) =
1 − µl λnl Kln (i)
l∈Nk
+ µ2l e0−l (i + 1)e∗0−l (i + 1)λnl
X
+2×
E αkl (i)αkq (i) E ckl (i)ckq (i)
1 − µl λnl
l,q∈Nk
l6=q
9
TABLE VI
C OVARIANCE MATRIX A3 UPON CONVERGENCE
From [41], we employ the relation E[sign(ekl )] ≈
sign(e0−k ). According to Eqs. (1) and (37), when the proposed
1
2
3
4
5
algorithm converges at node k or the time instant i goes to
1
0
infinity, we assume that the error e0−k will be equal to the
2
1
noise variance at node k. Then, the asymptotic value hkl can
3
0
1
1
0
1
4
0
be divided into three situations due to the rule of the SI–LMS
5
1
algorithm:
sign(|e0−k |)
for the node with the largest MSE
ρε 1+ε|e0−k |
n
∗
n
n
sign(−|e
|)
0−k
.
× 1 − µq λq Kl,q (i) + µl µq e0−l (i + 1)e0−q (i + 1)λh
ρε 1+ε|e0−k | for the node with the smallest MSE
l,qkl =
0
for all the remaining nodes.
(55)
(60)
By taking i → ∞, we can obtain (56). It should be noticed Under this situation, after the time instant i goes to infinity, the
that with the assumption that upon convergence the choice of parameters hkl for each neighbor node of node k can be obcovariance matrix Ak for node k is fixed, which means it is tained through (60) and the quantity hkl will be deterministic
deterministic and does not vary. In the above example, we and can be taken out from the expectation.
assume the choice of A3 is fixed as show in Table VI.
At last, removing the random variables αkl (i) and inserting
Then the coefficients αkl will also be fixed and given by
(58), (59) into (56), the asymptotic values Kkn for the SI-LMS
algorithm are obtained as in (61). At this point, the theoretical
α31 = 0
results are deterministic.
α32 = 1
Then, the MSE for SI–LMS algorithm is given by
α33 = 1
α
=
0
34
M
X
2
α35 = 1
Jmse−k = Jmin−k + M σx,k
(62)
Kkn (SI-LMS).
n=1
as well as the parameters ckl that are computed using the
Metropolis combining rule. As a result, the coefficients αkl
3) ES–RLS: For the proposed ES–RLS algorithm, we start
and the coefficients ckl are deterministic and can be taken out from (10), after inserting (10) into (17), we have
from the expectation.
X
ω k (i + 1) =
ckl (i)ψ l (i + 1)
The MSE is then given by
Jmse−k = Jmin−k +
2
M σx,k
M
X
Kkn (ES-LMS).
(57)
=
n=1
2) SI–LMS: For the SI–LMS algorithm, we do not need
to consider all possible combinations. This algorithm simply
adjusts the combining coefficients for each node with its
neighbors in order to select the neighbor nodes that yield
the smallest MSE values. Thus, we redefine the combining
coefficients through (27)
ckl−new = ckl − ρε
sign(|ekl |)
(l ∈ Nk ).
1 + ε|ξmin |
(58)
For each node k, at time instant i, after it received the
estimates from all its neighbors, it calculates the error pattern
ekl (i) for every estimate received through Eq. (19) and find the
nodes with the largest and smallest error. An error vector êk
is then defined through (22), which contains all error pattern
ekl (i) for node k.
Then a procedure which is detailed after Eq. (22) is
carried out and modifies the error vector êk . For example,
suppose node 5 has three neighbor nodes, which are nodes
3, 6 and 8. The error vector ê5 has the form described by
ê5 = [e53 , e55 , e56 , e58 ] = [0.023, 0.052, −0.0004, −0.012].
After the modification, the error vector ê5 will be edited as
ê5 = [0, 0.052, −0.0004, 0]. The quantity hkl is then defined
as
sign(|ekl |)
(l ∈ Nk ),
(59)
hkl = ρε
1 + ε|ξmin |
and the term ’error pattern’ ekl in (59) is from the modified
error vector êk .
=
b k (i)
l∈Ω
X
b k (i)
l∈Ω
X
ckl (i)[ω l (i) + kl (i + 1)e∗l (i + 1)]
ckl (i)[ω l (i) + kl (i + 1)(dl (i + 1)
b k (i)
l∈Ω
− xH
l (i + 1)ω l (i))].
(63)
Then, subtracting the ω 0 from both sides of (47), we arrive at
X
εk (i + 1) =
ckl (i)[ω l (i) + kl (i + 1)(dl (i + 1)
b k (i)
l∈Ω
=
X
b k (i)
l∈Ω
=
X
b k (i)
l∈Ω
− xH
l (i + 1)ω l (i))] −
X
ckl (i)ω 0
b k (i)
l∈Ω
ckl (i) εl (i) + kl (i + 1) dl (i + 1)
−
xH
l (i
+ 1)(εl (i) + ω0 )
ckl (i) I − kl (i + 1)xH
l (i + 1) εl (i)
+ kl (i +
1)e∗0−l (i
+ 1) .
(64)
Then, with the random variables αkl (i), (64) can be rewritten
as
X
εk (i + 1) =
αkl (i)ckl (i) I − kl (i + 1)xH
l (i + 1) εl (i)
l∈Nk
10
P
l∈Nk
Kkn (ES-LMS) =
1−
P
l∈Nk
P
l∈Nk
Kkn (SI-LMS) =
1−
P
2
α2kl c2kl (1 − µl λn
l ) −2
(ckl − hkl )2 µ2l Jmin−l λn
l +2
l∈Nk
+ kl (i +
α2kl c2kl µ2l Jmin−l λn
l +2
2
(ckl − hkl )2 (1 − µl λn
l ) −2
1)e∗0−l (i
+ 1) .
(65)
P
l,q∈Nk
l6=q
P
l,q∈Nk
l6=q
P
l,q∈Nk
l6=q
P
l,q∈Nk
l6=q
n
αkl αkq ckl ckq µl µq e0−l e∗
0−q λl,q
n
αkl αkq ckl ckq (1 − µl λn
l )(1 − µq λq )
.
(56)
n
(ckl − hkl )(ckq − hkq )µl µq e0−l e∗
0−q λl,q
n
(ckl − hkl )(ckq − hkq )(1 − µl λn
l )(1 − µq λq )
+
X
l,q∈Nk
l6=q
.
1−
E αkl (i)αkq (i)ckl (i)ckq (i)
(61)
1 2
i−M
Since kl (i + 1) = Φ−1
l (i + 1)xl (i + 1) [19], we can modify
λnl,q e0−q (i + 1)e∗0−l (i + 1)
n
H
the (65) as
× (Kl,q (i)) +
(i − M )2 λnq λnl
X
−1
H
(69)
εk (i + 1) =
αkl (i)ckl (i) I − Φl (i + 1)xl (i + 1)xl (i + 1) εl (i)
n
l∈Nk
where Kk (i + 1) is the nth elements of the main diagonals
of K k (i + 1). With the assumption that, upon convergence,
−1
∗
+ Φl (i + 1)xl (i + 1)e0−l (i + 1) .
αkl and ckl do not vary, because at steady state, the choice
b k (i) for each node k will be fixed, we can rewrite
(66) of subset Ω
(69)
as
(70).
Then, the MSE is given by
At this point, if we compare (66) with (50), we can find that
the difference between (66) and (50) is, the Φ−1
l (i+1) in (66)
has replaced the µl in (50). From [19], we also have
1
R−1 (i + 1) for i > M + 1. (67)
i−M l
As a result, we can arrive
X
Λ−1 Λl
K l (i)
K k (i + 1) =
E α2kl (i)c2kl (i)
I− l
i−M
2
Jmse−k (i + 1) = Jmin−k + M σx,k
M
X
n=1
Kkn (i + 1)(ES-RLS).
(71)
On the basis of (70), we have that when i tends to infinity,
the MSE approaches the MMSE in theory [19].
4) SI–RLS: For the proposed SI–RLS algorithm, we insert
(58) into (70), remove the random variables αkl (i), and
following the same procedure as for the SI–LMS algorithm,
l∈Nk
wecan obtain (72), where hkl and hkq satisfy the rule in (60).
Λ−1
Λl Λ−1
Λl Λ−1
∗
l
l
Then, the MSE is given by
× I−
+ l
e
(i
+
1)e
(i
+
1)
0−l
0−l
i−M
(i − M )2
M
X
X
2
Λ−1
Kkn (i + 1)(SI-RLS).
l Λl Jmse−k (i + 1) = Jmin−k + M σx,k
+
E αkl (i)αkq (i)ckl (i)ckq (i)
I−
i−M
n=1
l,q∈Nk
(73)
l6=q
In
conclusion,
according
to
(61)
and
(72),
with
the
help
of
−1
Λ−1
Λq Λ−1
q
l Λl,q Λq
modified
combining
coefficients,
for
the
proposed
SI–type
+
× K l,q (i) I −
e0−l (i + 1)
i−M
(i − M )2
algorithms, the neighbor node with lowest MSE contributes
X
the most to the combination, while the neighbor node with
∗
× e0−q (i + 1) +
E αkl (i)αkq (i)ckl (i)ckq (i)
the highest MSE contributes the least. Therefore, the proposed
l,q∈Nk
SI–type algorithms perform better than the standard diffusion
l6=q
algorithms with fixed combining coefficients.
−1
−1
Λq Λq H
Λ Λl
K l,q (i) I − l
× I−
i−M
i−M
C. Tracking Analysis
−1 H
−1
Λq Λl,q Λl
In this section, we assess the proposed ES–LMS/RLS and
e0−q (i + 1)e∗0−l (i + 1) .
(68)
+
SI–LMS/RLS algorithms in a non–stationary environment, in
(i − M )2
Due to the structure of the above equations, the approximations which the algorithms have to track the minimum point of
the error–performance surface [46], [47]. In the time–varying
and the quantities involved, we can decouple (68) into
scenarios of interest, the optimum estimate is assumed to vary
X
1
2
according to the model ω 0 (i + 1) = βω 0 (i) + q(i), where
Kln (i)
Kkn (i + 1) =
E α2kl (i)c2kl (i)
1−
i−M
q(i) denotes a random perturbation [44] and β = 1 in order to
l∈Nk
facilitate the analysis. This is typical in the context of tracking
e0−l (i + 1)e∗0−l (i + 1)
analysis of adaptive algorithms [19], [48], [49].
+
λnl (i − M )2
1) ES–LMS: For the tracking analysis of the ES–LMS
X
1 2algorithm, we employ Assumption III and start from (47).
+
E αkl (i)αkq (i)ckl (i)ckq (i)
1−
i − M After subtracting the ω0 (i + 1) from both sides of (47), we
l,q∈Nk
obtain
l6=q
X
n
∗
λ
e
(i
+
1)e
(i
+
1)
εk (i + 1) =
ckl (i)[ω l (i) + µl xl (i + 1)(dl (i + 1)
0−q
l,q 0−l
n
× Kl,q
(i) +
b k (i)
(i − M )2 λnl λnq
l∈Ω
E[Φ−1
l (i + 1)] =
11
+ 1)(ES-RLS) =
1−
1−
2
P
(ckl − hkl )2 1 −
X
1
i−M
− xH
l (i + 1)ω l (i))] −
b k (i)
l∈Ω
X
b k (i)
l∈Ω
l,q∈Nk
l6=q
P
−2
n
(i − M)2 λn
l λq
αkl αkq ckl ckq 1 −
l,q∈Nk
l6=q
P
∗
λn
l,q e0−l e0−q
(ckl − hkl )(ckq − hkq )
l,q∈Nk
l6=q
−2
P
1
i−M
(70)
∗
λn
l,q e0−l e0−q
n
(i − M)2 λn
l λq
(ckl − hkl )(ckq − hkq ) 1 −
l,q∈Nk
l6=q
2 .
1
i−M
2 .
(72)
INSTANT
Adaptation Method
LMS
RLS
Multiplications
8M + 2
4M 2 + 16M + 1
Additions
8M
4M 2 + 12M − 1
Divisions
1
TABLE VIII
C OMPUTATIONAL COMPLEXITY FOR COMBINATION STEP PER NODE PER TIME
ckl (i) ω 0 (i) + q(i)
INSTANT
ckl (i) εl (i) + µl xl (i + 1) dl (i + 1)
Algorithms
ES–LMS/RLS
− q(i)
− xH
(i
+
1)(ε
(i)
+
ω
)
l
0
l
X
=
ckl (i) I − µl xl (i + 1)xH
l (i + 1) εl (i)
b k (i)
l∈Ω
αkl αkq ckl ckq
TABLE VII
ckl (i)[ω l (i) + µl xl (i + 1)(dl (i + 1)
X
P
ckl (i)ω 0 (i + 1)COMPUTATIONAL COMPLEXITY FOR THE ADAPTATION STEP PER NODE PER TIME
b k (i)
l∈Ω
X
b k (i)
l∈Ω
=
+2
(ckl − hkl )2
− xH
l (i + 1)ω l (i))] −
1
i−M
Jmin−l
2
λn
l (i − M)
P
l∈Nk
=
2
α2kl c2kl 1 −
l∈Nk
+ 1)(SI-RLS) =
P
l∈Nk
Kkn (i
+2
α2kl c2kl
l∈Nk
Kkn (i
Jmin−l
2
λn
l (i − M)
P
+ µl xl (i + 1)e∗0−l (i + 1) − q(i).
SI–LMS/RLS
Multiplications
T!
t!(T − t)!
(2M + 3)|Nk |
M(t + 1)
Additions
T!
t!(T − t)!
(M + 2)|Nk |
Divisions
Mt
|Nk |
4) SI–RLS: We start from (73), and after a similar procedure to that of the SI–LMS algorithm, we have
2
Jmse−k (i + 1) = Jmin−k + M σx,k
M
X
Kkn (i + 1)(SI-RLS)
n=1
(74)
Using Assumption III, we can arrive at
2
+ M σx,k
tr{Q}.
(80)
In conclusion, for time-varying scenarios there is only one
2
additional term M σx,k
tr{Q} in the MSE expression for all
algorithms, and this additional term has the same value for
all algorithms. As a result, the proposed SI–type algorithms
still perform better than the standard diffusion algorithms
with fixed combining coefficients, according to the conclusion
obtained in the previous subsection.
Jex−k (i + 1) = tr{Rk (i + 1)K k (i + 1)} + tr{Rk (i + 1)Q}.
(75)
The first part on the right side of (75), has already been
obtained in the MSE steady–state analysis part in Section IV
B. The second part can be decomposed as
H
D. Computational Complexity
tr{Rk (i + 1)Q} = tr E xk (i + 1)xH
(i
+
1)
E
q(i)q
(i)
k
In the analysis of the computational cost of the algorithms
2
= M σx,k
tr{Q}.
(76) studied, we assume complex-valued data and first analyze the
adaptation step. For both ES–LMS/RLS and SI–LMS/RLS
The MSE is then obtained as
algorithms, the adaptation cost depends on the type of reM
cursions (LMS or RLS) that each strategy employs. The
X
2
2
Jmse−k = Jmin−k +M σx,k
Kkn (ES-LMS)+M σx,k
tr{Q}. details are shown in Table VII. For the combination step,
n=1
we analyze the computational complexity in Table VIII. The
(77)
overall complexity for each algorithm is summarized in Table
2) SI–LMS: For the SI–LMS recursions, we follow the
IX. In the above three tables, T is the total number of nodes
same procedure as for the ES-LMS algorithm, and obtain
linked to node k including node k itself and t is the number of
M
nodes
chosen from T . The proposed algorithms require extra
X
2
2
Jmse−k = Jmin−k +M σx,k
Kkn (SI-LMS)+M σx,k
tr{Q}. computations as compared to the existing distributed LMS and
n=1
RLS algorithms. This extra cost ranges from a small additional
(78) number of operations for the SI-LMS/RLS algorithms to a
3) ES–RLS: For the SI–RLS algorithm, we follow the same more significant extra cost that depends on T for the ESprocedure as for the ES–LMS algorithm and arrive at
LMS/RLS algorithms.
2
Jmse−k (i + 1) = Jmin−k + M σx,k
M
X
Kkn (i + 1)(ES-RLS)
n=1
2
+ M σx,k
tr{Q}.
(79)
V. S IMULATIONS
In this section, we investigate the performance of the
proposed link selection strategies for distributed estimation
12
TABLE IX
C OMPUTATIONAL
Algorithm
COMPLEXITY PER NODE PER TIME INSTANT
Multiplications
(t + 1)T !
+8 M +2
t!(T
−
t)!
(t + 1)T !
+ 16 M + 1
4M 2 +
t!(T − t)!
(8 + 2Nk )M + 3|Nk | + 2
2
4M + (16 + 2|Nk |)M + 3|Nk | + 1
ES–LMS
ES–RLS
SI–LMS
SI–RLS
Additions
T!
+8 M
(t
−
1)!(T
−
t)!
T!
4M 2 +
+ 12 M − 1
(t − 1)!(T − t)!
(8 + |Nk |)M + 2|Nk |
2
4M + (12 + |Nk |)M + 2|Nk | − 1
Divisions
20
1
|Nk |
|Nk | + 1
−5
14
Diffusion LMS [2]
Single−−Link Strategy [29]
ES−−LMS
SI−−LMS
Diffusion RLS [28]
ES−−RLS
SI−−RLS
MMSE Bound
12
11
18
−10
16
10
−15
4
14
18
20
12
MSE (dB)
−20
10
9
15
7
8
19
−25
8
6
−30
13
4
3
6
17
−35
2
16
0
0
2
4
6
2
1
8
10
12
5
14
16
18
20
−40
0
Fig. 3.
80
160
240
320
400
480
560
640
720
800
Time, i
Diffusion wireless sensor networks topology with 20 nodes
Fig. 4.
Network MSE curves in a static scenario
in two scenarios: wireless sensor networks and smart grids.
In these applications, we simulate the proposed link selection
strategies in both static and time–varying scenarios. We also
show the analytical results for the MSE steady–state and
tracking performances that we obtained in Section IV.
0
Diffusion LMS [2]
Single−−Link Strategy [29]
ES−−LMS
SI−−LMS
Diffusion RLS [28]
ES−−RLS
SI−−RLS
MMSE Bound
−5
MSE (dB)
−10
A. Diffusion Wireless Sensor Networks
−15
−20
In this subsection, we compare the proposed ES–LMS/ES–
RLS and SI–LMS/SI–RLS algorithms with the diffusion LMS
algorithm [2], the diffusion RLS algorithm [50] and the single–
link strategy [51] in terms of their MSE performance. The network topology is illustrated in Fig. 3 and we employ N = 20
nodes in the simulations. The length of the unknown parameter
vector ω 0 is M = 10 and it is generated randomly. The input
signal is generated as xk (i) = [xk (i) xk (i − 1) ... xk (i −
M + 1)] and xk (i) = uk (i) + αk xk (i − 1), where αk is a
correlation coefficient and uk (i) is a white noise process with
2
variance σu,k
= 1 − |αk |2 , to ensure the variance of xk (i) is
2
σx,k = 1. The noise samples are modeled as circular Gaussian
2
noise with zero mean and variance σn,k
= 0.001. The step
size for the diffusion LMS ES–LMS and SI–LMS algorithms
is µ = 0.045. For the diffusion RLS algorithm, both ES–RLS
and SI–RLS, the forgetting factor λ is set to 0.97 and δ is
equal to 0.81. In the static scenario, the sparsity parameters
of the SI–LMS/SI–RLS algorithms are set to ρ = 4 × 10−3
and ε = 10. The Metropolis rule is used to calculate the
combining coefficients ckl . The MSE and MMSE are defined
as in (2) and (45), respectively. The results are averaged over
100 independent runs.
In Fig. 4, we can see that ES–RLS has the best performance
for both steady-state MSE and convergence rate, and obtains
a gain of about 8 dB over the standard diffusion RLS algorithm. SI–RLS is a bit worse than the ES–RLS, but is still
significantly better than the standard diffusion RLS algorithm
by about 5 dB. Regarding the complexity and processing time,
SI–RLS is as simple as the standard diffusion RLS algorithm,
−25
−30
0
80
160
240
320
400
480
560
640
720
800
Time, i
Fig. 5.
Network MSE curves in a time-varying scenario
while ES–RLS is more complex. The proposed ES–LMS and
SI–LMS algorithms are superior to the standard diffusion
LMS algorithm. In the time–varying scenario, the sparsity
parameters of the SI–LMS and SI–RLS algorithms are set to
ρ = 6 × 10−3 and ε = 10. The unknown parameter vector ω 0
varies according to the first–order Markov vector process:
ω0 (i + 1) = βω 0 (i) + z(i),
(81)
where z(i) is an independent zero–mean Gaussian vector
process with variance σz2 = 0.01 and β = 0.98.
Fig. 5 shows that, similarly to the static scenario, ES–RLS
has the best performance, and obtains a 5 dB gain over the
standard diffusion RLS algorithm. SI–RLS is slightly worse
than the ES–RLS, but is still better than the standard diffusion
RLS algorithm by about 3 dB. The proposed ES–LMS and SI–
LMS algorithms have the same advantage over the standard
diffusion LMS algorithm in the time-varying scenario. Notice
that in the scenario with large |Nk |, the proposed SI-type
algorithms still have a better performance when compared with
the standard techniques. To illustrate the link selection for the
ES–type algorithms, we provide Figs. 6 and 7. From these
two figures, we can see that upon convergence the proposed
13
Link selection state for Node 16 with ES−−LMS algorithm
static scenarios
{11, 15, 18, 20}
−10
b) SI−LMS, N=20 nodes, µ=0.045,
ρ=4*10e−3, ε=10
0
Simulation
Theory
a) ES−LMS, N=20 nodes, µ=0.045
−5
Simulation
Theory
−5
−5
−10
−15
−10
−15
{12,15,17,18}
{11,12,14,15,18,20}
−30
{11,12,14,15,18,20}
−20
MSE (dB)
−25
MSE (dB)
MSE (dB)
−15
−20
−25
−20
−25
−35
−30
−40
0
Fig. 6.
80
160
240
320
400
Time, i
480
560
640
−30
720
−35
Set of selected links for node 16 with ES–LMS in a static scenario
−35
Link selection state for Node 16 with ES−LMS algorithm
time−−varying scenarios
0
−40
0
{12,14,17}
10
20
SNR (dB)
−40
0
30
10
20
SNR (dB)
30
−5
Fig. 8.
MSE (dB)
−10
MSE performance against SNR for ES–LMS and SI–LMS
a) ES−RLS, N=20 nodes, λ=0.97,
δ=0.811
0
Simulation
Theory
−15
{11,12,17,18}
b) SI−RLS, N=20 nodes, λ=0.97,
δ=0.811, ρ=4*10e−3, ε=10
−5
−5
Simulation
Theory
−10
−20
{11,12,14,17,18,20}
{11,12,14,17,18,20}
−10
−15
−25
160
240
320
400
Time, i
480
560
640
720
Fig. 7. Link selection state for node 16 with ES–LMS in a time-varying
scenario
bk.
algorithms converge to a fixed selected set of links Ω
B. MSE Analytical Results
The aim of this section is to validate the analytical results
obtained in Section IV. First, we verify the MSE steady–state
performance. Specifically, we compare the analytical results in
(57), (62), (71) and (73) to the results obtained by simulations
under different SNR values. The SNR indicates the signal
variance to noise variance ratio. We assess the MSE against
the SNR, as show in Figs. 8 and 9. For ES–RLS and SI–
RLS algorithms, we use (71) and (73) to compute the MSE
after convergence. The results show that the analytical curves
coincide with those obtained by simulations and are within
0.01 dB of each other, which indicates the validity of the
analysis. We have assessed the proposed algorithms with SNR
equal to 0dB, 10dB, 20dB and 30dB, respectively, with 20
nodes in the network. For the other parameters, we follow the
same definitions used to obtain the network MSE curves in
a static scenario. The details have been shown on the top of
each sub figure in Figs. 8 and 9. The theoretical curves for
ES–LMS/RLS and SI–LMS/RLS are all close to the simulation
curves and remain within 0.01 dB of one another.
The tracking analysis of the proposed algorithms in a time–
varying scenario is discussed as follows. Here, we verify that
the results (77), (78), (79) and (80) of the subsection on
the tracking analysis can provide a means of estimating the
MSE (dB)
80
MSE (dB)
−15
−30
0
−20
−20
−25
−25
−30
−30
−35
−35
−40
0
Fig. 9.
10
20
SNR (dB)
30
−40
0
10
20
SNR (dB)
30
MSE performance against SNR for ES–RLS and SI–RLS
MSE. We consider the same model as in (81). In the next
examples, we employ N = 20 nodes in the network and the
same parameters used to obtain the network MSE curves in a
time–varying scenario. A comparison of the curves obtained
by simulations and by the analytical formulas is shown in
Figs. 10 and 11. From these curves, we can verify that the gap
between the simulation and analysis results are within 0.02dB
under different SNR values. The details of the parameters are
shown on the top of each sub figure in Figs. 10 and 11.
C. Smart Grids
The proposed algorithms provide a cost–effective tool that
could be used for distributed state estimation in smart grid
applications. In order to test the proposed algorithms in a
possible smart grid scenario, we consider the IEEE 14–bus
system [52], where 14 is the number of substations. At every
14
a) ES−LMS, N=20 nodes, µ=0.045
0
Simulation
Theory
b) SI−LMS, N=20 nodes, µ=0.045,
ρ=6*10e−3, ε=10
0
Simulation
Theory
−5
−5
where xk (i) is the measurement Jacobian vector for bus k.
Then, the aim of the distributed estimation algorithm is to
compute an estimate of ω 0 , which can minimize the cost
function given by
2
Jω k (i) (ωk (i)) = E|dk (i) − xH
k (i)ω k (i)| .
We compare the proposed algorithms with the M–CSE algorithm [4], the single link strategy [51], the standard diffusion
RLS algorithm [50] and the standard diffusion LMS algorithm
[2] in terms of MSE performance and the Phase Angle Gap.
The MSE comparison is used to determine the accuracy of the
algorithms, and the Phase Angle Gap is used to compare the
rate of convergence. In our scenario, ’Phase Angle Gap’ refers
to the phase angle difference between the actual parameter
vector or target ω 0 and the estimate ωk (i) for all buses. We
define the IEEE–14 bus system as in Fig. 12.
−10
MSE (dB)
MSE (dB)
−10
−15
−15
−20
−20
−25
−25
(84)
13
−30
0
10
20
SNR (dB)
−30
0
30
14
12
10
20
SNR (dB)
30
10
Fig. 10. MSE performance against SNR for ES–LMS and SI–LMS in a
time-varying scenario
11
6
9
a) ES−RLS, N=20 nodes, λ=0.97,
δ=0.811
−5
Simulation
Theory
8
b) SI−RLS, N=20 nodes, λ=0.97,
δ=0.811, ρ=6*10e−3, ε=10
0
Simulation
Theory
7
1
5
4
3
2
−10
−5
Fig. 12.
−20
−15
−25
−20
−30
−25
−35
0
All buses are corrupted by additive white Gaussian noise
with equal variance σ 2 = 0.001. The step size for the
standard diffusion LMS [2], the proposed ES–LMS and SI–
LMS algorithms is 0.018. The parameter vector ω0 is set to
an all–one vector. For the diffusion RLS, ES–RLS and SI–
RLS algorithms the forgetting factor λ is set to 0.945 and δ is
equal to 0.001. The sparsity parameters of the SI–LMS/RLS
algorithms are set to ρ = 0.07 and ε = 10. The results are
averaged over 100 independent runs. We simulate the proposed
algorithms for smart grids under a static scenario.
−10
MSE (dB)
MSE (dB)
−15
10
20
SNR (dB)
30
−30
0
IEEE 14–bus system for simulation
0
10
20
SNR (dB)
Diffusion RLS[28]
SI−−RLS
Diffusion LMS[2]
SI−−LMS
M−CSE[4]
ES−−LMS
Single−−Link Strategy[29]
ES−−RLS
MMSE Bound
30
−5
Fig. 11. MSE performance against SNR for ES–RLS and SI–RLS in a
time-varying scenario
−10
time instant i, each bus k, k = 1, 2, . . . , 14, takes a scalar
measurement dk (i) according to
dk (i) = Xk ω 0 (i) + nk (i), k = 1, 2, . . . , 14,
(82)
where ω0 (i) is the state vector of the entire interconnected
system, Xk (ω 0 (i)) is a nonlinear measurement function of
bus k. The quantity nk (i) is the measurement error with mean
equal to zero and which corresponds to bus k.
Initially, we focus on the linearized DC state estimation
problem. The system is built with 1.0 per unit (p.u) voltage
magnitudes at all buses and j1.0 p.u. branch impedance. Then,
the state vector ω 0 (i) is taken as the voltage phase angle vector
ω 0 for all buses. Therefore, the nonlinear measurement model
for state estimation (82) is approximated by
dk (i) = xH
k (i)ω 0 + nk (i),
k = 1, 2, . . . , 14,
(83)
MSE(dB)
−15
−20
−25
−30
−35
−40
0
100
200
300
400
500
600
700
800
900
Time, i
Fig. 13.
MSE performance curves for smart grids.
From Fig. 13, it can be seen that ES–RLS has the best
performance, and significantly outperforms the standard diffusion LMS [2] and the M–CSE [4] algorithms. The ES–LMS is
slightly worse than ES–RLS, which outperforms the remaining
techniques. SI–RLS is worse than ES–LMS but is still better
15
than SI–LMS, while SI–LMS remains better than the diffusion
RLS, LMS, M–CSE algorithms and the single link strategy.
Phase Angle Gap Comparision For Bus 5
0
−0.1
Phase Angle Gap
−0.2
−0.3
Diffusion RLS[28]
SI−−RLS
Diffusion LMS[2]
SI−−LMS
M−−CSE[4]
ES−−LMS
Single Link Strategy[29]
ES−−RLS
−0.4
−0.5
0
10
20
30
40
50
60
70
80
90
Time,i
Fig. 14.
Phase Angle Gap comparison for Bus 5.
In order to compare the convergence rate, we employ the
’Phase Angle Gap’ to describe the results. We choose bus 5
and the first 90 iterations as an example to illustrate our results.
In Fig. 14, ES–RLS has the fastest convergence rate, while SI–
LMS is the second fastest, followed by the standard diffusion
RLS, ES–LMS, SI–LMS, and the standard diffusion LMS
algorithms. The M–CSE algorithm and the single link strategy
have the worst performance. The estimates ω k (i) obtained by
the proposed algorithms quickly reach the target ω 0 , which
means the Phase Angle Gap will converge to zero.
VI. C ONCLUSIONS
In this paper, we have proposed ES–LMS/RLS and SI–
LMS/RLS algorithms for distributed estimation in applications
such as wireless sensor networks and smart grids. We have
compared the proposed algorithms with existing methods. We
have also devised analytical expressions to predict their MSE
steady–state performance and tracking behavior. Simulation
experiments have been conducted to verify the analytical
results and illustrate that the proposed algorithms significantly
outperform the existing strategies, in both static and time–
varying scenarios, in examples of wireless sensor networks
and smart grids.
VII. ACKNOWLEDGEMENTS
The authors wish to thank the anonymous reviewers, whose
comments and suggestions have greatly improved the presentation of these results.
R EFERENCES
[1] C. G. Lopes and A. H. Sayed, “Incremental adaptive strategies over
distributed networks,” IEEE Trans. Signal Process., vol. 48, no. 8, pp.
223–229, Aug 2007.
[2] ——, “Diffusion least-mean squares over adaptive networks: Formulation and performance analysis,” IEEE Trans. Signal Process., vol. 56,
no. 7, pp. 3122–3136, July 2008.
[3] Y. Chen, Y. Gu, and A. Hero, “Sparse LMS for system identification,”
Proc. IEEE ICASSP, pp. 3125–3128, Taipei, Taiwan, May 2009.
[4] L. Xie, D.-H. Choi, S. Kar, and H. V. Poor, “Fully distributed state
estimation for wide-area monitoring systems,” IEEE Trans. Smart Grid,
vol. 3, no. 3, pp. 1154–1169, September 2012.
[5] D. Bertsekas, “A new class of incremental gradient methods for least
squares problems,” SIAM J. Optim, vol. 7, no. 4, p. 913926, 1997.
[6] A. Nedic and D. Bertsekas, “Incremental subgradient methods for
nondifferentiable optimization,” SIAM J. Optim, vol. 12, no. 1, p.
109138, 2001.
[7] M. G. Rabbat and R. D. Nowak, “Quantized incremental algorithms for
distributed optimization,” IEEE J. Sel. Areas Commun, vol. 23, no. 4,
p. 798808, Apr 2005.
[8] F. S. Cattivelli and A. H. Sayed, “Diffusion lms strategies for distributed
estimation,” IEEE Trans. Signal Process., vol. 58, p. 10351048, March
2010.
[9] P. D. Lorenzo, S. Barbarossa, and A. H. Sayed, “Sparse diffusion
LMS for distributed adaptive estimation,” Proc. IEEE International
Conference on Acoustics, Speech, and Signal Processing, pp. 3281–
3284, Kyoto, Japan, March 2012.
[10] S. Xu, R. C. de Lamare, and H. V. Poor, “Distributed compressed
estimation based on compressive sensing,” IEEE Signal Processing
Letters, vol. 22, no. 9, pp. 1311–1315, Sept 2015.
[11] ——, “Adaptive link selection algorithms for distributed estimation,”
EURASIP Journal on Advances in Signal Processing, vol. 2015, no. 1,
pp. 1–22, October 2015.
[12] C. Lopes and A. Sayed, “Diffusion adaptive networks with changing
topologies,” in Proc. IEEE International Conference on Acoustics,
Speech and Signal Processing, Las Vegas, 2008, pp. 3285–3288.
[13] B. Fadlallah and J. Principe, “Diffusion least-mean squares over adaptive
networks with dynamic topologies,” in Proc. IEEE International Joint
Conference on Neural Networks, Dallas, TX, USA 2013.
[14] T. Wimalajeewa and S. Jayaweera, “Distributed node selection for
sequential estimation over noisy communication channels,” IEEE Trans.
Wirel. Commun., vol. 9, no. 7, pp. 2290–2301, July 2010.
[15] R. C. de Lamare and R. Sampaio-Neto, “Adaptive reduced-rank processing based on joint and iterative interpolation, decimation and filtering,”
IEEE Trans. Signal Process., vol. 57, no. 7, pp. 2503–2514, July 2009.
[16] R. C. de Lamare and P. S. R. Diniz, “Set-membership adaptive algorithms based on time-varying error bounds for CDMA interference
suppression,” IEEE Trans. Vehi. Techn., vol. 58, no. 2, pp. 644–654,
February 2009.
[17] L. Guo and Y. F. Huang, “Frequency-domain set-membership filtering
and its applications,” IEEE Trans. Signal Process., vol. 55, no. 4, pp.
1326–1338, April 2007.
[18] A. Bertrand and M. Moonen, “Distributed adaptive node–specific signal
estimation in fully connected sensor networks–part II: Simultaneous and
asynchronous node updating,” IEEE Trans. Signal Process., vol. 58,
no. 10, pp. 5292–5306, 2010.
[19] S. Haykin, Adaptive Filter Theory, 4th ed. Upper Saddle River, NJ,
USA: Prentice Hall, 2002.
[20] L. Li and J. A. Chambers, “Distributed adaptive estimation based on
the APA algorithm over diffusion networks with changing topology,”
in Proc. IEEE Statist. Signal Process. Workshop, pp. 757–760, Cardiff,
Wales, September 2009.
[21] X. Zhao and A. H. Sayed, “Performance limits for distributed estimation
over lms adaptive networks,” IEEE Trans. Signal Process., vol. 60,
no. 10, pp. 5107–5124, October 2012.
[22] L. Xiao and S. Boyd, “Fast linear iterations for distributed averaging,”
Syst. Control Lett., vol. 53, no. 1, pp. 65–78, September 2004.
[23] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks of
agents with switching topology and time-delays,” IEEE Trans. Autom.
Control., vol. 49, pp. 1520–1533, September 2004.
[24] A. Jadbabaie, J. Lin, and A. S. Morse, “Coordination of groups of mobile
autonomous agents using nearest neighbor rules,” IEEE Trans. Autom.
Control, vol. 48, no. 6, pp. 988–1001, June 2003.
[25] R. C. de Lamare and R. C. de Lamare, “Adaptive reduced-rank mmse
filtering with interpolated fir filters and adaptive interpolators,” IEEE
Signal Processing Letters, vol. 12, no. 3, March 2005.
[26] R. Meng, R. C. de Lamare, and V. H. Nascimento, “Sparsity-aware
affine projection adaptive algorithms for system identification,” in Proc.
Sensor Signal Processing for Defence Conference, London, UK, 2011.
[27] Z. Yang, R. de Lamare, and X. Li, “Sparsity-aware space-time adaptive
processing algorithms with l1-norm regularisation for airborne radar,”
Signal Processing, IET, vol. 6, no. 5, pp. 413–423, July 2012.
[28] ——, “L1-regularized stap algorithms with a generalized sidelobe canceler architecture for airborne radar,” Signal Processing, IEEE Transactions on, vol. 60, no. 2, pp. 674–686, Feb 2012.
[29] R. C. de Lamare and R. C. de Lamare, “Sparsity-aware adaptive
algorithms based on alternating optimization and shrinkage,” IEEE
Signal Processing Letters, vol. 21, no. 2, pp. 225–229, January 2014.
16
[30] R. C. de Lamare and R. Sampaio-Neto, “Reduced–rank adaptive filtering
based on joint iterative optimization of adaptive filters,” IEEE Signal
Process. Lett., vol. 14, no. 12, pp. 980–983, December 2007.
[31] ——, “Reduced-rank space–time adaptive interference suppression with
joint iterative least squares algorithms for spread-spectrum systems,”
IEEE Transactions Vehicular Technology, vol. 59, no. 3, pp. 1217–1228,
March 2010.
[32] ——, “Adaptive reduced-rank equalization algorithms based on alternating optimization design techniques for MIMO systems,” IEEE
Transactions on Vehicular Technology, vol. 60, no. 6, pp. 2482–2494,
July 2011.
[33] ——, “Adaptive reduced-rank processing based on joint and iterative
interpolation, decimation, and filtering,” IEEE Transactions on Signal
Processing, vol. 57, no. 7, pp. 2503–2514, July 2009.
[34] R. Fa, R. C. de Lamare, and L. Wang, “Reduced-rank stap schemes
for airborne radar based on switched joint interpolation, decimation and
filtering algorithm,” IEEE Transactions on Signal Processing, vol. 58,
no. 8, pp. 4182–4194, August 2010.
[35] S. Li, R. C. de Lamare, and R. Fa, “Reduced-rank linear interference
suppression for ds-uwb systems based on switched approximations of
adaptive basis functions,” IEEE Transactions on Vehicular Technology,
vol. 60, no. 2, pp. 485–497, Feb 2011.
[36] R. C. de Lamare, R. Sampaio-Neto, and M. Haardt, “Blind adaptive constrained constant-modulus reduced-rank interference suppression
algorithms based on interpolation and switched decimation,” IEEE
Transactions on Signal Processing, vol. 59, no. 2, pp. 681–695, Feb
2011.
[37] M. L. Honig and J. S. Goldstein, “Adaptive reduced-rank interference
suppression based on the multistage wiener filter,” IEEE Transactions
on Communications, vol. 50, no. 6, June 2002.
[38] R. C. de Lamare, M. Haardt, and R. Sampaio-Neto, “Blind adaptive
constrained reduced-rank parameter estimation based on constant modulus design for cdma interference suppression,” IEEE Transactions on
Signal Processing, vol. 56, no. 6, June 2008.
[39] N. Song, R. C. de Lamare, M. Haardt, and M. Wolf, “Adaptive widely
linear reduced-rank interference suppression based on the multi-stage
wiener filter,” IEEE Transactions on Signal Processing, vol. 60, no. 8,
August 2012.
[40] H. Ruan and R. de Lamare, “Robust adaptive beamforming using a
low-complexity shrinkage-based mismatch estimation algorithm,” Signal
Processing Letters, IEEE, vol. 21, no. 1, pp. 60–64, Jan 2014.
[41] Y. Chen, Y. Gu, and A. Hero, “Regularized least-mean-square algorithms,” Technical Report for AFOSR, December 2010.
[42] S. Xu and R. C. de Lamare, “Distributed conjugate gradient strategies
for distributed estimation over sensor networks,” in Proc. Sensor Signal
Processing for Defence 2012, London, UK, 2012.
[43] F. Cattivelli and A. H. Sayed, “Diffusion strategies for distributed kalman
filtering and smoothing,” IEEE Trans. Autom. Control, vol. 55, no. 9,
pp. 2069–2084, September 2010.
[44] A. H. Sayed, Fundamentals of Adaptive Filtering. Hoboken, NJ, USA:
John Wiley&Sons, 2003.
[45] T. Kailath, A. H. Sayed, and B. Hassibi, Linear Estimation. Englewood
Cliffs, NJ, USA: Prentice-Hall, 2000.
[46] R. C. de Lamare and P. S. R. Diniz, “Blind adaptive interference
suppression based on set-membership constrained constant-modulus
algorithms with dynamic bounds,” IEEE Trans. Signal Process., vol. 61,
no. 5, pp. 1288–1301, March 2013.
[47] Y. Cai and R. C. de Lamare, “Low–complexity variable step-size
mechanism for code-constrained constant modulus stochastic gradient
algorithms applied to cdma interference suppression,” IEEE Trans.
Signal Process., vol. 57, no. 1, pp. 313–323, January 2009.
[48] B. Widrow and S. D. Stearns, Adaptive Signal Processing. Englewood
Cliffs, NJ, USA: Prentice-Hall, 1985.
[49] E. Eweda, “Comparison of RLS, LMS, and sign algorithms for tracking
randomly time-varying channels,” IEEE Trans. Signal Process., vol. 42,
p. 29372944, November 1994.
[50] F. Cattivelli, C. Lopes, and A. Sayed, “Diffusion recursive least-squares
for distributed estimation over adaptive networks,” IEEE Trans. Signal
Process., vol. 56, no. 5, pp. 1865–1877, May 2008.
[51] X. Zhao and A. H. Sayed, “Single-link diffusion strategies over adaptive
networks,” in Proc. IEEE International Conference on Acoustics, Speech
and Signal Processing, pp. 3749–3752, Kyoto, Japan, March 2012.
[52] A. Bose, “Smart transmission grid applications and their supporting
infrastructure,” IEEE Trans. Smart Grid, vol. 1, no. 1, pp. 11–19, Jun
2010.
| 3cs.SY
|
arXiv:1704.07106v4 [math.AC] 18 Dec 2017
COHEN–MACAULAY BINOMIAL EDGE IDEALS OF
CACTUS GRAPHS
GIANCARLO RINALDO
Abstract. We classify the Cohen-Macaulay binomial edge ideals of
cactus and bicyclic graphs.
Introduction
In 2010, binomial edge ideals were introduced in [6] and appeared independently also in [8]. Let S = K[x1 , . . . , xn , y1 , . . . , yn ] be the polynomial
ring in 2n variables with coefficients in a field K. Let G be a graph on vertex
set [n] and edges E(G). For each {i, j} ∈ E(G) with i < j, we associate a binomial fij = xi yj − xj yi . The ideal J(G) of S generated by fij = xi yj − xj yi
such that i < j , is called the binomial edge ideal of G. Any ideal generated
by a set of 2-minors of a 2 × n-matrix of indeterminates may be viewed as
the binomial edge ideal of a graph.
Algebraic properties of binomial edge ideals in terms of properties of the
underlying graph were studied in [6], [4] and [9]. In [4] and [9] the authors
considered the Cohen-Macaulay property of these graphs. Recently nice
results on Cohen-Macaulay bipartite graphs and blocks have been obtained
(see [7], [2] and [1]).
However, the classification of Cohen-Macaulay binomial edge ideals in
terms of the underlying graphs is still widely open and, as in the case of
monomial edge ideals introduced in [11], it seems rather hopeless to give a
full classification.
The aim of this paper is to extend the results of [10] where we classify
Cohen-Macaulay and unmixed binomial edge ideals J(G) with deviation,
namely the difference between the minimum number of the generators and
the height of J(G), less than or equal to 1. This invariant has an interesting combinatorial interpretation: if J(G) is unmixed than its deviation is
|E(G)| − n + c where c is the number of connected components of G (see
Remark 2.1 of [10]). Hence deviation represents the minimum number of
edges that must be removed from the graph to break all its cycles making
it into a forest (see Chapter 4 of [5]).
In section 2 we give a classification Cohen-Macaulay and unmixed binomial edge ideals J(G) when G is a cactus graph, a graph whose blocks are
cycles. This is a natural extension to the result obtained in [10] and useful
to study binomial edge ideals by the deviation invariant.
1
2
GIANCARLO RINALDO
In section 3 as an application of the results obtained in section 2 we
classify the Cohen-Macaulay and unmixed binomial edge ideals J(G) when
G is a bicyclic graph, that is the case of deviation 2.
1. Preliminaries
In this section we recall some concepts and notation on graphs and on
simplicial complexes that we will use in the article.
Let G be a simple graph with vertex set V (G) and edge set E(G). A
subset C of V (G) is called a clique of G if for all i and j belonging to C
with i 6= j one has {i, j} ∈ E(G). A vertex of a graph is called a cutpoint if
the removal of the vertex increases the number of connected components. A
connected subgraph of G that has no cutpoint and is maximal with respect
to this property is a block. A block graph B(G) is a graph whose vertices are
the blocks of G and two vertices are adjacent whenever the corresponding
blocks contain a common cutpoint of G. A connected graph is a cactus if
its blocks are cycles or edges.
Set V = {x1 , . . . , xn }. A simplicial complex ∆ on the vertex set V is a
collection of subsets of V such that: (i) {xi } ∈ ∆ for all xi ∈ V ; (ii) F ∈ ∆
and G ⊆ F imply G ∈ ∆. An element F ∈ ∆ is called a face of ∆. A
maximal face of ∆ with respect to inclusion is called a facet of ∆. A vertex
i of ∆ is called a free vertex of ∆ if i belongs to exactly one facet. The clique
complex ∆(G) of G is the simplicial complex whose faces are the cliques of
G. Hence a vertex v of a graph G is called free vertex if it belongs to only
one clique of ∆(G).
We need notation and results from [6] (section 3) that we recall for the
sake of completeness. Let T ⊆ [n], and let T = [n] \ T . Let G1 , . . . , Gc(T ) be
the connected components of the induced subgraph on T , namely GT . For
ei the complete graph on the vertex set V (Gi ). We set
each Gi , denote by G
[
e1 ), . . . , J(G
ec(T ) )),
(1)
PT (G) = ( {xi , yi }, J(G
i∈T
PT (G) is a prime ideal. Then J(G) is a radical ideal and
\
J(G) =
PT (G)
T ⊂[n]
is its primary decomposition (see Corollary 2.2 and Theorem 3.2 of [6]).
If there is no possible confusion, we write simply PT instead of PT (G).
Moreover, height PT = n + |T | − c(T ) (see [6, Lemma 3.1]). We denote by
M(G) the set of minimal prime ideals of J(G).
If each i ∈ T is a cutpoint of the graph GT ∪{i} , then we say that T is a
cutset for G. We denote by C(G) the set of all cutsets for G.
Lemma 1.1. [6] PT (G) ∈ M(G) if and only if T ∈ C(G).
Lemma 1.2. [9] Let G be a connected graph. Then J(G) is unmixed if and
only if for all T ∈ C(G) we have c(T ) = |T | + 1.
COHEN–MACAULAY BINOMIAL EDGE IDEALS OF CACTUS GRAPHS
3
The following observation gives motivation to consider the block graphs
in this context.
Proposition 1.3. Let J(G) be unmixed. Then the block graph B(G) is a
tree.
Proof. By Theorem 3.5 of [5] each block of the block graph is a complete
graph. We recall that two vertices in B(G) are adjacent when the corresponding blocks contain a common cutpoint of G. Moreover for each
cutpoint v, {v} ∈ C(G). By Lemma 1.2 the assertion follows.
We observe the following (see also Remark 3.1.(i) of [2])
Proposition 1.4. Let v be a vertex of G with neighbor set N (v) = {u ∈
V (G) | {u, v} ∈ E(G)} and let Gv the graph such that V (Gv ) = V (G) and
E(Gv ) = E(G) ∪ {{u1 , u2 } | u1 6= u2 , u1 , u2 ∈ N (v)}. Then
C(Gv ) = {T ∈ C(G) | v 6∈ T }.
Proof. If v is a free vertex then E(G) = E(Gv ). In fact by definition there
is only one clique containing v, that is all vertices adjacent to v are adjacent
to each other. Hence we assume v is not a free vertex in G. By definition
of the graph Gv v is a free vertex of Gv . This implies by Proposition 1.1 of
[9] that there are no cutsets of Gv containing v.
Let S ⊂ V (G) such that v ∈
/ S, and let
G1 , . . . , Gc1
be the connected components of the graph GS , with v ∈ V (G1 ). Each
{v, u} ∈ E(G) either intersects S, hence is not in E(GS ), or is an edge of
G1 since it is connected through v in G1 . In the same way, let H = Gv and
let
H1 , . . . , Hc2
be the connected components of the graph HS , with v ∈ V (H1 ). Each
{v, u} ∈ E(H) and each {u1 , u2 } ∈ E(H) with u1 , u2 ∈ N (v) either intersects S or is an edge of H1 . This implies that Hi = Gi for i = 2, . . . , c1 and
c1 = c2 . Moreover
(2)
V (Gi ) = V (Hi ),
∀i = 1, . . . , c1 .
In particular this is true for all the cutsets. That is for all T ∈ C(G) then
i ∈ T is a cutpoint of the graph GT ∪{i} , and the number of connected
components decreases when S = T \ {i}. The same happens for H, hence
T ∈ C(H). The same argument works in the other direction.
Corollary 1.5. Let v be a vertex of G that is not free vertex and let Gv
6
the graph such that V (Gv ) = V (G) and E(Gv ) = E(G) ∪ {{u1 , u2 } | u1 =
u2 , u1 , u2 ∈ N (v)}. Then
with Qv =
T
J(G) = J(Gv ) ∩ Qv
T ∈C(G),v∈T
PT (G).
4
GIANCARLO RINALDO
Proof. Let v ∈ V (G) then J(G) = Qv ∩ Qv with
\
Qv =
PT (G),
Qv =
T ∈C(G),v∈T
\
PT (G).
T ∈C(G),v∈T
/
Applying Proposition 1.4, we have that the cutsets in C(Gv ) are exactly
the cutsets in C(G) not containing v. Moreover the connected components
induced by any T in G and in Gv have the same set of vertices as stated in
the proof of Proposition 1.4 (see (2)). Using the notation introduced in the
ei = G
ei for each connected component, that is
mentioned proof H
J(Gv ) = Qv .
Example 1.6. Sometime the ideal Qv in Corollary 1.5 has a natural interpretation. For example let G = C4 with vertices V (C4 ) = {1, . . . , 4} and
edges {{1, 2}, {2, 3}, {3, 4}, {1, 4}}. Then we obtain
(3)
J(G) = J(G1̄ ) ∩ Q1 = J(D) ∩ (x1 , y1 , x3 , y3 )
where D, known as diamond graph, has V (D) = V (C4 ) and E(D) = E(C4 )∪
{{2, 4}} and the ring S/Qv is related to the isolated vertex 2 and 4. By
similar argument we obtain
(4)
J(D) = J(D2̄ ) ∩ Q2 = J(K4 ) ∩ (x2 , y2 , x4 , y4 )
where K4 is the complete graph with 4 vertices. If G = C5 such an interpretation fails. Let V (C5 ) = {1, . . . , 5} and edges {{1, 2}, {2, 3}, {3, 4}, {4, 5}, {1, 5}}.
If we consider J(G) = J(G1̄ ) ∩ Q1 , then G1̄ is the graph obtained adding the
edge {2, 5} to the cycle C5 , as expected by Corollary 1.5, but
Q1 = (x1 , y1 , f2,3 , f4,5 , x3 x4 , x3 y4 , y3 x4 , y3 y4 ).
Passing from a block that is not complete to a complete one (see (3),(4)
in Example 1.6 and Figure 2), is useful for our aim thanks to the following
nice result (see Theorem 1.1 of [4]) that we state using our notation
Theorem 1.7. Let G be a graph whose blocks are complete graphs. The
following conditions are equivalent:
(1) J(G) is Cohen-Macaulay;
(2) J(G) is unmixed;
(3) B(G) is a tree.
2. Classification of Cohen-Macaulay cactus graphs
In this section we provide a classification of Cohen-Macaulay binomial
edge ideal J(G) when G is a cactus graphs. Since a binomial edge ideal J(G)
is Cohen–Macaulay (resp. unmixed) if and only if J(H) is Cohen–Macaulay
(resp. unmixed) for each connected component H of G, we assume from
now on that the graph G is connected unless otherwise stated.
We start by the following
COHEN–MACAULAY BINOMIAL EDGE IDEALS OF CACTUS GRAPHS
5
Proposition 2.1. Let J(G) be an unmixed binomial edge ideal. If a cycle
Cl is a block of G then l ∈ {3, 4}.
Proof. We begin observing by [4] and [10] that unmixed binomial edge ideals
containing blocks that are cycles of length 3 and 4 exist. We assume that
exists a cycle Cl in G with l ≥ 5 and J(G) unmixed. Since Cl is a block we
represent G as
!
r
[
Gi
(5)
G = Cl ∪
i=1
where r ≥ 0, Gi are subgraphs of G for all i, |V (Gi ) ∩ V (Cl )| = 1. Since Cl
is a block we have that {v} = V (Gi ) ∩ V (Cl ) is a cutpoint and all the paths
between u ∈ V (Gi ) and w ∈ V (Gj ) with i 6= j pass through v (see Theorem
3.1 of [5]).
Let V (Cl ) = {i1 , i2 , . . . , il } such that {ij , ij+1 } is an edge of G with
j = 1, . . . , l − 1 and {i1 , il } is an edge of G, too. We observe that if T
is {ij , ik } such that {ij , ik } ∈
/ E(Cl ) then T ∈ C(G). Moreover GT has at
least two connected components induced by the remaining l − 2 vertices of
Cl . Assuming G is unmixed there are exactly 3 connected components in GT
and this implies that one of the vertices in T is a cutpoint and the other is
not a cutpoint. Thanks to this observation we easily obtain a contradiction.
We give the proof for the sake of completeness. We observe that in the cutset {i1 , i3 } we may assume without loss of generality that i1 is the cutpoint
and i3 is not a cutpoint. By the same argument focusing on {i1 , il−1 }, il−1
is not a cutpoint. If we consider {i2 , il−1 } and {i3 , il } we obtain that i2 and
il are cutpoints. Let T1 = {i2 , il } since GT 1 has 4 connected components we
obtain a contradiction.
Proposition 2.2. Let J(G) be an unmixed binomial edge ideal. If a cycle
C4 is a block of G then it satisfies the following (C4)-condition
• there are exactly two cutpoints in C4 and they are adjacent.
Proof. By Proposition 2 of [10] we know that a graph satisfying the thesis
exists. We prove that all the other cases are not unmixed. Let V (C4 ) =
{1, 2, 3, 4} and let E(C4 ) = {{1, 2}, {2, 3}, {3, 4}, {1, 4}}. We assume the
same representation of G given in (5). Suppose that either r = 0 or r = 1
with cutpoint 1. Then c({2, 4}) = 2. Hence it is not unmixed. Let r ≥ 2
and assume that 1 and 3 are cutpoints then c({1, 3}) = 4. Hence J(G) is
not unmixed.
Definition 2.3. A graph G is decomposable (resp. indecomposable) if exists
(resp. does not exist) a decomposition
(6)
G = G1 ∪ G2
with {v} = V (G1 ) ∩ V (G2 ) such that v is a free vertex of ∆(G1 ) and ∆(G2 ).
By a recursive decomposition (6) applied to each G1 and G2 , after a finite
6
GIANCARLO RINALDO
B1
B2
rv2
B3
w2r
v3
w1
...
r
w3
Bl
r
vl
Figure 1. B(G) is a path.
number of steps we obtain
(7)
G = G1 ∪ · · · ∪ Gr
where G1 , . . . , Gr are indecomposable and for 1 ≤ i < j ≤ r either V (Gi ) ∩
V (Gj ) = ∅ or V (Gi ) ∩ V (Gj ) = {v} and v is a free vertex of ∆(Gi ) and
∆(Gj ). The decomposition (7) is unique up to ordering and we say that G
is decomposable into indecomposable graphs G1 , . . . , Gr .
Lemma 2.4 (See [9]). Let G be a decomposable graph with decomposition
G = G1 ∪ G2 and V (G1 ) ∩ V (G2 ) = {v}. Then J(G) is Cohen-Macaulay
(resp. unmixed) if and only if J(G1 ) and J(G2 ) are Cohen-Macaulay (resp.
unmixed).
Our aim is to prove that every Cohen-Macaulay cactus graph G is decomposable into indecomposable graphs G1 , . . . , Gr such that the block graph
B(Gi ) is a path for 1 ≤ i ≤ r. Hence it is necessary as a first step to classify
the indecomposable Cohen-Macaulay cactus graphs whose block graph is a
path. To reach this goal we use from now on the following notation. Let
G be a graph such that B(G) is a path defined on the following sets (see
Figure 1)
(8)
vertices of B(G) {B1 , . . . , Bl }
edges of B(G)
{{Bi , Bi+1 } | i = 1, . . . , l − 1}
cutpoints of G
{wi = vi+1 | i = 1, . . . , l − 1}
and such that wi = vi+1 ∈ V (Bi ) ∩ V (Bi+1 ), with i = 1, . . . , l − 1.
Lemma 2.5. Let G be a graph such that B(G) is a path. We use notation
(8). If |V (Bi )| ≥ 3 for all 1 < i < l then the power set of {w1 , . . . , wl−1 } is
a subset of C(G).
Proof. We use induction on t the cardinality of T ∈ C(G). If t = 1 since
each wi is a cutpoint of G the claim follows. Let 2 ≤ t < l − 1 and
T = {wi1 , wi2 , . . . wit } with 1 ≤ i1 < i2 < · · · < it ≤ l − 1.
Let wj ∈ {w1 , . . . , wl−1 }\T and assume without loss of generality that j < i1
and i1 ≥ 2. We observe that the graph GT has the connected component
H = B1 ∪ . . . ∪ Bi1 −1 ∪ Bi′1 where Bi′1 is obtained removing the vertex wi1
from the block Bi1 . Since by hypothesis Bi1 has more than two vertices, Bi′1
COHEN–MACAULAY BINOMIAL EDGE IDEALS OF CACTUS GRAPHS
7
contains at least one vertex, vi′1 , that is not a cutpoint in G. Hence H{wj }
has two connected components: one containing a vertex in V (B1 ) \ {w1 }
and one containing the vertex vi′1 . That is wj is a cutpoint of GT and
T ∪ {wj } ∈ C(G).
By Proposition 2.1, Example 1.6 and Theorem 1.7 is useful for our aim
to compute the primary decompositions of J(G) whose blocks are complete
graphs, cycles C4 and diamond graphs D (see also Figure 2).
Proposition 2.6. Let G be a graph such that B(G) is a path. We use
notation (8). Let B1 = Km1 , Bl = Kml with m1 , ml ≥ 2 and
Bi ∈ {C4 , D, Kmi | mi ≥ 3} for 2 ≤ i ≤ l − 1
with the following labelling on the vertices of the blocks Bi ∈ {C4 , D}:
• V (C4 ) = {vi , vi′ , wi , wi′ }, E(C4 ) = {{vi , wi }, {vi , vi′ }, {vi′ , wi′ }, {wi , wi′ }};
• V (D) = {vi , vi′ , wi , fi }, E(D) = {{vi , wi }, {vi , vi′ }, {vi′ , fi }, {wi , fi }, {vi′ , wi }}.
Then T ∈ C(G) if and only if T ⊆ V ⊔ V ′ where V = {w1 , . . . , wl−1 },
[
[
{vi′ }
{vi′ , wi′ } ∪
V′ =
Bi =C4
Bi =D
and satisfying the following conditions:
(1) if vi′ ∈ T then wi ∈ T and vi ∈
/ T;
(2) if wi′ ∈ T then vi ∈ T and wi ∈
/ T.
Proof. For the sake of completeness we give the equivalent conditions to (1)
and (2) that are:
(1’) if vi ∈ T or wi ∈
/ T then vi′ ∈
/ T;
(2’) if wi ∈ T or vi ∈
/ T then wi′ ∈
/ T.
Suppose T is a cutset. Then v ∈ T is not a free vertex. Hence it follows T ⊆
V ⊔ V ′ . In fact a vertex fi in a diamond graph is a free vertex in the clique
{vi′ , wi , fi } and the same holds for all the vertices that are not cutpoints
in a block that is a complete graph. Moreover, if T ⊆ V then it satisfies
condition 1. and 2. in a trivial way. Hence we assume v ′ ∈ T with v ′ ∈ V ′ .
That is either v ′ = vi′ ∈ V (Bi ) with Bi ∈ {C4 , D} or v ′ = wi′ ∈ V (Bi ) with
Bi = C4 . We consider only the case v ′ = vi′ ∈ V (Bi ) with Bi = C4 since the
other cases follow by similar argument. Since T is a cutset, the graph GS
with S = T ∪ {vi′ } has a cutpoint in vi′ . Hence the two vertices vi and wi′
that are adjacent to vi′ in G must belong to GS , that is vi and wi′ are not in
T . By the same reason since vi′ is a cutpoint in GS and since wi is adjacent
to vi and wi′ in G, too, then wi ∈
/ S and wi ∈ T . That is condition (1) is
satisfied.
Now suppose T ⊆ V ⊔V ′ satisfying conditions (1) and (2) and let n′ = |V ′ |.
We prove that T is a cutset by induction on n′ . If n′ = 0 that is T ⊆ V ,
then it is a cutset by Lemma 2.5.
8
GIANCARLO RINALDO
Let n′ ≥ 1 and let v2′ ∈ V ′ then B2 ∈ {C4 , D} (see Figure 2). We
assume B2 = C4 since the other case is similar. By conditions (1) and (2’).
w2 ∈ T and neither v2 nor w2′ belongs to T . That is the graph GT contains
2 connected components, B1 and the isolated vertex w2′ , and a subgraph of
H = B3′ ∪ B4 ∪ · · · ∪ Bl , where B3′ is obtained removing the vertex w2 = v3
from B3 .
We claim that T ′ = T \ {v2′ , w2 } is a cutset of H. Moreover HT ′ has
a connected component containing a vertex in B3 that is adjacent to w2 .
By the claim it easily follows that T is a cutset of G and GT contains
the following connected components: B1 , the isolated vertex w2′ , and the
connected components of HT ′ .
If B3 is a complete graph then B3′ is a complete graph, too. Hence H
satisfies the hypothesis of our Proposition and T ′ is a cutset by induction
/T
hypothesis. If B3 = D then B3′ is the complete graph K3 . Moreover v3′ ∈
by condition (1’) applied to w2 = v3 ∈ T . Hence the set T ′ contains at most
the cutpoint w3 . Also in this case the graph H satisfies the hypothesis of
our Proposition and T ′ is a cutset by induction hypothesis. If B3 = C4 then
B3′ is a path of length 2 whose edges are
{v3′ , w3′ }, {w3′ , w3 }.
We note that also in this case v3′ ∈
/ T by condition (1’), and T contains at
most one of the vertices w3 and w3′ . In fact v3′ is the free vertex of the path
and only one of the cutpoints w3 , w3′ can appear in T ′ . Moreover we can
apply induction hypothesis on the subgraph H ′ obtained removing v3′ from
H.
By similar argument to the previous case we obtain the assertion under
the assumption w2′ ∈ V ′ . In this case B2 = C4 , v2 ∈ T and GT contains
the following connected components: the isolated vertex v2′ , B1′ , obtained
removing the vertex v2 = w1 from B1 , and the connected components of
HT ′ with H = B3 ∪ · · · ∪ Bl and T ′ = T \ {w2′ , v2 }.
Corollary 2.7. Let G be a graph that satisfies the hypothesis of Proposition
2.6. Then J(G) is unmixed.
Proof. By Lemma 1.2 it is sufficient to show that c(T ) = |T | + 1 for all
T ∈ C(G). Thanks to Lemma 2.5, T = {wi1 , wi2 , . . . wit } ⊆ {w1 , . . . , wl−1 }
with 1 ≤ i1 < i2 < · · · < it ≤ l − 1 is a cutset of G. We note that
(9)
GT = G1 ⊔ G2 ⊔ · · · ⊔ Gt+1
where G1 = B1 ∪ · · · ∪ Bi1 −1 ∪ Bi′′1 , Gt+1 = Bi′t +1 ∪ Bit +2 ∪ · · · ∪ Bl and
Gj = Bi′j +1 ∪ Bij +2 ∪ · · · ∪ Bij+1 −1 ∪ Bi′′j+1 with j = 2, . . . , t
and Bi′′ is the block obtained removing wi from Bi , Bi′ is the block obtained
removing wi−1 = vi from Bi . We easily observe that Gi is connected for
1 ≤ i ≤ t + 1. Hence c(T ) = t + 1. Now let T ⊔ T ′ ∈ C(G) with T =
{wi1 , wi2 , . . . wit } ⊆ {w1 , . . . , wl−1 } and T ′ ∩ {w1 , . . . , wl−1 } = ∅ with t′ =
COHEN–MACAULAY BINOMIAL EDGE IDEALS OF CACTUS GRAPHS
q
q
❅
❅
❅sq
q
q
q
❅
❅
q
❅q
B3 · · · Bl
q
v2
w2
Gw
2
v2′ q
qsw2′
′
=⇒
9
q
q
❅
B3 · · · Bl
B3 · · · Bl
❅
q
❅q
qs
q
w2
v2❅
w2
Gv ′
2
❅
=⇒
qf2
q
❅q
v2
v2′ qs
Figure 2. Cases B2 ∈ {C4 , D, K4 } of Theorem 2.8
|T ′ | > 0. This implies that v ′ ∈ T ′ is either vi′ or wi′ of Bi ∈ {C4 , D}. If we
focus on T the representation of GT in (9) holds in this case, too. Without
loss of generality let v ′ ∈ V (G1 ). Since the only cutpoint that induces the
connected component G1 is wi1 , by condition (1) of Proposition 2.6, v ′ = vi′1 .
Hence
(10)
GT ∪{v′
i1 }
where (G1 ){v′
i1 }
= (G1 ){v′
i1 }
⊔ G2 ⊔ · · · ⊔ Gt+1
= B1 ∪ · · · ∪ Bi1 −1 ⊔ {v} where v = fi1 if Bi1 = D or v = wi′ 1
if Bi1 = C4 and c(T ∪ {vi′1 }) = (t + 1) + 1 where the last summand is induced
by the isolated vertex v. The same argument holds for all v ′ ∈ T ′ .
Theorem 2.8. Let G be a graph that satisfies the hypothesis of Proposition
2.6. Then J(G) is Cohen-Macaulay.
Proof. We start observing that dim S/J(G) = n + 1. This follows by Corollary 2.7 and the formula (see [6])
(11)
dim S/J(G) = max{n − |T | + c(T )}.
Hence it is sufficient to prove that depth S/J(G) ≥ n + 1 using induction
on l, the number of blocks of G. Our strategy is to focus on the block
B2 . In particular when B2 ∈ {D, C4 } (see Figure 2) we consider the vertex
v ′ ∈ {v2′ , w2′ } and the following exact sequence
(12)
0 −→ S/J(G) −→ S/Qv′ ⊕ S/J(G′ ) −→ S/(Qv′ + J(G′ )) −→ 0
with G′ = Gv′ . By Corollary 2.5, J(G′ ) is a binomial edge ideal and the
graph G′ satisfies the hypothesis of our Theorem but with a second block
that has less cutsets than B2 (see Figure (2)). Moreover S/Qv′ (respectively
S/Qv′ + J(G′ )) is a tensor product of 3 (respectively 2) quotient rings whose
definining ideals are binomial edge ideals, and it is Cohen-Macaulay using
induction. To obtain this goal we have three cases to study: B2 = Km2 ,
B2 = D and B2 = C4 . We start induction with l = 3.
Case B2 = Km2 . S/J(G) is Cohen-Macaulay by Theorem 1.1 of [4].
Case B2 = D. We set G′ = Gv2 ′ . By Proposition 1.4 we have E(G′ ) =
E(G) ∪ {{v2 , f2 }}, that is the second block in G′ is the complete graph on
the vertices {v2 , v2′ , w2 , f2 }. By Corollary 1.5, J(G) = J(G′ ) ∩ Qv2′ , where
Qv2′ = P{v2′ ,w2 } = (xv2′ , yv2′ ) + (xw2 , yw2 ) + J(B1 ) + J(B3′ ) and B3′ is obtained
by removing the vertex w2 = v3 from the complete graph B3 . By Depth
10
GIANCARLO RINALDO
Lemma applied on the sequence (12) with v ′ = v2′ we obtain the assertion.
In fact G′ is Cohen-Macaulay by the case B2 = Km2 and its depth is n + 1.
Let S ′ = S/(xv2′ , yv2′ , xw2 , yw2 ), we obtain
(13) S/Qv2′ ∼
= S ′ /J(B1 ) + J(B3′ ) ∼
=
K[{xi , yi } : i ∈ V (B3′ )]
K[{xi , yi } : i ∈ V (B1 )]
∼
⊗ K[xf2 , yf2 ] ⊗
=
J(B1 )
J(B3′ )
where the 3 quotient rings associated to the complete graphs B1 , B3′ and
the isolated vertex f2 are Cohen-Macaulay. Using the formula (11) for each
ring, and adding the results thanks to (13), we obtain
depth S/Qv2′ = |(V (B1 ) + 1) + (1 + 1) + (|V (B3′ )| + 1) = n + 1.
By the same argument depth S/(Qv2′ + J(G′ )) = n. In fact
(14) S/Qv2′ + J(G′ ) ∼
= S ′ /J(B1 ∪ E ′ ) + J(B3′ ) ∼
=
K[{xi , yi } : i ∈ V (B1 ) ∪ {f2 }] K[{xi , yi } : i ∈ V (B3′ )]
∼
⊗
=
J(B1 ∪ E)
J(B3′ )
where E is the edge {w1 = v2 , f2 }. Hence J(B1 ∪ E) is Cohen-Macaulay by
Theorem 1.1 of [4]. By Depth Lemma the assertion follows.
Case B2 = C4 . We set v ′ = w2′ and G′ = Gw2 ′ . By Proposition 1.4
we have E(G′ ) = E(G) ∪ {{v2′ , w2 }}, that is the second block in G′ is a
diamond, hence J(G′ ) is Cohen-Macaualy by the case B2 = D. Moreover
depth S/Qv′ = n + 1 and depth S/Qv′ + J(G′ ) = n. In fact Qv′ = P{v2 ,w2′ } =
(xw2′ , yw2′ ) + (xv2 , yv2 ) + J(B1′ ) + J(B3 ), where B1′ is the complete graph
obtained removing the vertex v2 = w1 from the graph B1 . Using a representation of S/Qv′ similar to (13) we easily obtain the assertion. Moreover
Qv′ +J(G′ ) = (xw2′ , yw2′ )+(xv2 , yv2 )+J(B1′ )+J(B3 ∪E) where E is the edge
{v2′ , w2 }. Also in this case we obtain the assertion using a representation of
S/Qv′ + J(G′ ) equivalent to the one used in (14).
Let l > 3. Case B2 = Km2 . By Lemma 2.4 applied on the graphs
G1 = B1 and G2 = B2 ∪ · · · ∪ Bl , since G1 is a complete graph and G2
is Cohen-Macaulay by induction hypothesis, we obtain that G is CohenMacaulay.
Case B2 = D. As in the case B2 = D with l = 3, setting G′ = Gv2 ′ ,
the second block in G′ , namely B2′ , is the complete graph on the vertices
{v2 , v2′ , w2 , w2′ }, hence
G′ = B1 ∪ B2′ ∪ B3 ∪ · · · ∪ Bl .
Therefore J(G′ ) is Cohen-Macaulay using induction hypothesis
Tand Lemma
2.4 applied on B1 and B2′ ∪B3 ∪· · ·∪Bl . By Corollary 1.5 Qv2′ = T ∈C(G),v′ ∈T PT (G).
2
Thanks to condition (1) of Proposition 2.6 for each T ∈ C(G) with v2′ ∈ T
then w2 ∈ T and v2 ∈
/ T . Hence Qv2′ = J(B1 ) + (xv2′ , yv2′ , xw2 , yw2 ) + J.
We claim that J = J(H), where H = B3′ ∪ B4 ∪ · · · ∪ Bl , with B3′ obtained removing the vertex w2 = v3 from B3 . By the claim, and defining
COHEN–MACAULAY BINOMIAL EDGE IDEALS OF CACTUS GRAPHS
11
S ′ = S/(xv2′ , yv2′ , xw2 , yw2 ), we obtain a representation of S/Qv2′ by tensor
product similar to (13), that is
K[{xi , yi } : i ∈ V (B1 )]
K[{xi , yi } : i ∈ V (H)]
⊗ K[xf2 , yf2 ] ⊗
,
J(B1 )
J(H)
and using induction hypothesis on H we obtain the assertion. The claim
follows proving the condition
(15)
T ∈ C(H) if and only if T ∪ {v2′ , w2 } ∈ C(G).
Proof of (15). Case B3 = D. Let B3′ = K3 on the vertices {v3′ , f3 , w3 }, then
H satisfies Proposition 2.6. Let T ∈ C(H). To prove that T ∪{v2′ , w2 } ∈ C(G)
it is enough to check conditions (1) and (2) for all the vertices v ′ ∈ T ′ =
(T ∪{v2′ , w2 })\{w1 , . . . , wl−1 }. If v ′ ∈ V (B2 )∩T ′ then v ′ = v2′ and it satisfies
condition (1). If v ′ ∈ V (Bi ) ∩ T ′ with 4 ≤ i ≤ l then it satisfies condition 1.
and 2. with respect to G since it satisfies the same conditions with respect
to H. We end this implication observing that V (B3 ) ∩ T ′ = ∅. If T ⊆ V (H)
with T ∈
/ C(H) either there exists a block in {B4 , . . . , Bl } such that either
(1) or (2) is not satisfied or there exists v ∈ V (B3′ ) with v 6= w3 . In both
cases T ∪ {v2′ , w2 } does not satisfy (1) and (2) with respect to G.
The cases B3 ∈ {Km3 , C4 } follow by similar argument. We only point
out some facts when B3 = C4 . B3′ is a path defined on the vertex set
{v3′ , w3′ , w3 } and edges {v3′ , w3′ }, {w3′ , w3 }. In this case is useful to consider
H = E ∪ H ′ where E is the edge {v3′ , w3′ } and H ′ is obtained removing the
vertex v3′ from H. Then H ′ satisfies the hypothesis of the Proposition and
it is Cohen-Macaulay by induction. By Lemma 2.4 applied to H ′ and E we
obtain that J(H) is Cohen-Macaulay, too. Moreover by Lemma 2.3 of [9]
we have
C(H) = C(H ′ ) ∪ {T ∪ {w3′ } : T ∈ C(H ′ ) with w3 ∈
/ T }.
Observe that the graph studied in this case is decomposable with respect to
the vertex v2 . Hence also the graphs
(16)
B2 ∪ · · · ∪ Bl with B2 = D is Cohen-Macaulay.
Case B2 = C4 . Let G′ = Gw2 ′ . As in the case B2 = C4 and l = 3, the second
block of G′ is D. Then J(G′ ) is Cohen-Macaulay
by the case B2 = D with
T
l > 3, and J(G) = J(G′ ) ∩ Qw2′ with Qw2′ = T ∈C(G),w′ ∈T PT (G). We claim
2
that Qw2′ = J(B1′ ) + (xw2′ , yw2′ ) + (xv2 , yv2 ) + J(H) where B1′ is the complete
graph obtained removing v2 = w1 from B1 . Moreover, using the notation
introduced in Proposition 1.4,
H = (B3 ∪ B4 ∪ · · · ∪ Bl )v3 = (B3 )v3 ∪ B4 ∪ · · · ∪ Bl .
By the claim it follows, defining S ′ = S/(xv2 , yv2 , xw2′ , yw2′ ), the following
representation of S/Qw2′
K[{xi , yi } : i ∈ V (B1′ )]
K[{xi , yi } : i ∈ V (H)]
.
⊗ K[xv2′ , yv2′ ] ⊗
′
J(B1 )
J(H)
12
GIANCARLO RINALDO
q❝
q❝
❝q
✡❏
✡ ❏
✡
❝q
❏❝q
q❝
❏
❏
❏q
q
✡
✡
q✡
q❝
q
Figure 3. Tree and unicyclic Cohen-Macaulay indecomposable graphs.
We focus on the last factor since the other ones are exactly equivalent to
the ones already studied. The block (B3 )v3 is either a complete graph or a
diamond graph with v3 a free vertex. In both cases it is Cohen-Macaulay
by induction hypothesis and (16). The claim follows by the condition
(17)
T ∈ C(H) if and only if T ∪ {v2 , w2′ } ∈ C(G).
To prove (17) we use Proposition 2.6 and similar arguments to the ones used
to prove (15).
Lemma 2.9. Let G be an indecomposable cactus graph such that B(G) is a
path. We use notation (8). The following conditions are equivalent:
(1) J(G) is Cohen-Macaulay;
(2) J(G) is unmixed;
(3) One of the following 2 cases occurs
(a) G ∈ {K2 , C3 }.
(b) G has C4 subgraphs that satisfy the (C4)-condition, l ≥ 3 and
(i) B1 , Bl ∈ {C3 , K2 },
(ii) B2 and Bl−1 are C4 ,
(iii) Bi ∈ {C3 , C4 } for 3 ≤ i ≤ l − 2 and if Bi = C3 then
Bi+1 = C4 .
Proof. (1) ⇒ (2). It is a known fact. (2) ⇒ (3). By Proposition 2.1 only K2 ,
C3 and C4 are admissible blocks. Since every vertex of a complete graph is a
free vertex, we observe that every cactus graph whose blocks are K2 and C3
and not C4 is Cohen-Macaulay and is decomposable in single blocks. Now
suppose that the block graph is a path containing one or more blocks that
are C4 . By Proposition 2.2 a block C4 has two cutpoints, thus neither B1
nor Bl is C4 . Since by hypothesis G is indecomposable, two complete graphs
cannot be adjacent in B(G). We end observing that if a K2 is between two
cycles C4 it is not unmixed (see also Remark 4.7 of [2]). (3) ⇒ (1). The
implication follows applying Theorem 2.8.
Example 2.10. Thanks to Lemma 2.9 we obtain the Cohen-Macaulay indecomposable graph of trees and unicyclic graphs shown in Figure 3. The
bicyclic with the same properties are in Figure 4. From now on we underline
the free vertices by a circle around them.
COHEN–MACAULAY BINOMIAL EDGE IDEALS OF CACTUS GRAPHS
q❝
q
q❝
✡❏
✡ ❏
q
❏❝q
✡
q
q
q
q
❝q
❏
❏
❏q
q
q
❏
q
q
13
❏
❏q❝
Figure 4. Bicyclic Cohen-Macaulay indecomposable cactus graphs.
G3 q
G1 q
s
s
s
✡❏
✡ ❏
❏s
✡
s
q
q
q
q
G2
s
Figure 5. A non unmixed cactus graphs with B(G) that is a tree.
Theorem 2.11. Let G be a cactus graph. We use notation (8). The following conditions are equivalent
(1) J(G) is Cohen-Macaulay;
(2) J(G) is unmixed;
(3) B(G) is a tree, G is decomposable into indecomposable graphs G1 , . . . , Gr ,
and such that B(Gi ) is a path and Gi satisfies one of the equivalent
conditions of Lemma 2.9 for 1 ≤ i ≤ r.
Proof. (1) ⇒ (2). It is a known fact. (3) ⇒ (1). It follows by Lemma
2.4. (2) ⇒ (3). By Proposition 1.3 the block graph of G is a tree. Let
t be the number of vertices of B(G) having degree greater than two. We
make induction on t. If t = 0 then the block graph B(G) is a path and
the assertion follows by Lemma 2.4 and Lemma 2.9. Let t > 0 and let
B be a a vertex of B(G) whose degree is greater than 2. By Proposition
2.1 and Proposition 2.2, B = C3 . Suppose that C3 has a vertex v such
that {v} = V (G1 ) ∩ V (G2 ), G = G1 ∪ G2 and v is a free vertex of ∆(G1 )
and ∆(G2 ). Since G1 and G2 are cactus with number of blocks of degree
greater than 2 less than t, by induction hypothesis we are done. Suppose
by contradiction that in each of the 3 vertices of C3 the graph G is not
decomposable. Then C3 is adjacent to 3 blocks C4 as in figure 5, where G1 ,
G2 and G3 are intended as cactus subgraphs of G. We observe that the set
T containing the 6 vertices, indicated as filled dots in figure, is a cutset. But
GT has 6 components. Hence J(G) is not unmixed. Contradiction.
14
GIANCARLO RINALDO
q❝
q
q
❝q
q❝
❏
✡
✡
❏
❝q
❏✡
❏
❏
❏q
❝q
q
❏
✡
✡
❏
❏✡
❝q
✡❏
✡ ❏
❏q
✡
q
q
q❝
q
q
q
q
❏
❏
❏q❝
q
q
q
q
Figure 6. A Cohen-Macaulay cactus graphs.
Example 2.12. In figure 6 we have that G is the union of 4 indecomposable Cohen-Macaulay cactus graphs joined by free vertices (surrounded by a
circle). That are : a K2 ; a C3 ; one containing a K2 , a C3 and a C4 ; one
containing 2 K2 , one C3 and 3 C4 .
3. Classification of Cohen–Macaulay bicyclic graphs
In this section thanks to the classification of Cohen-Macaulay binomial
edge ideals of cactus graphs we classify the ones that are bicyclic, namely
the ideals of deviation 2. In fact a cactus graph having 2 cycles as blocks
is in particular a bicyclic graph. Hence we focus our attention on bicyclic
graphs G that are not cactus. In this case there exists one block B1 in G
such that
3
[
Pi
(18)
B1 =
i=1
where P1 , P2 and P3 are paths, V (Pi ) ∩ V (Pj ) = {a, b} for 1 ≤ i < j ≤ 3
and if B is a block of G with B 6= B1 then B is an edge.
Remark 3.1. The set T = {a, b}, where a and b are defined in (18), is a
cutset. Assume from now on that the ideal J(G) is unmixed. Let li be the
length of the path Pi in (18) such that
1 ≤ l1 ≤ l2 ≤ l3 .
If l1 = 1, that is {a, b} ∈ E(G), then a is a cutpoint and b is not a cutpoint.
In fact by Lemma 1.2, c(T ) = 3. But we have exactly two connected components induced by the paths P2 and P3 . Hence there exists another connected
component in GT that is not a subgraph of B1 and the assertion follows. By
a similar argument if l1 > 1 then neither a nor b is a cutpoint.
All over the section we use the notation defined in (18) and Remark 3.1.
We call these graphs non-cactus bicyclic graphs.
Lemma 3.2. Let G be a non-cactus bicyclic graph such that J(G) is unmixed. We use notation (18). Then
COHEN–MACAULAY BINOMIAL EDGE IDEALS OF CACTUS GRAPHS
15
(1) each path Pi has length less than 4;
(2) at most one path Pi has length 3.
Proof. (1). Suppose by contradiction that exists a path P ∈ {P1 , P2 , P3 }
of length l ≥ 4. Let a, a1 , c, b1 and b, 5 distinct vertices of V (P ), with
{a, a1 }, {b, b1 } ∈ E(P ). Let T = {a, b} and suppose that neither a nor b
is a cutpoint. We observe that T1 = {a, b1 } is a cutset. Hence c(T1 ) = 3.
But in GT 1 there are two connected components that are subgraphs of B1 ,
one containing the vertex c and one containing the vertex b. Hence there
exists another block B2 different from B1 that contains the vertex b1 . Let
E(B2 ) = {{b1 , b′1 }}. By the same argument there exists a block B3 such
that E(B3 ) = {{a1 , a′1 }}. The set T2 = {a1 , b1 } is a cutset and induces
4 components in GT 2 : The first containing a and b, the second containing
a′1 , the third containing b′1 , and the fourth containing c. Contradiction. By
similar argument is left to the reader to check that if a cutpoint and b is not
a cutpoint, then T1 = {b, c} is a cutset with c a cutpoint. Hence we have a
contradiction focusing on the cutset {a, c}.
(2). Suppose by contradiction that exist 2 paths P1 and P2 of length 3
such that
V (Pi ) = {a, ai , bi , b}, E(Pi ) = {{a, ai }, {ai , bi }, {bi , b}} for i = 1, 2.
By Remark 3.1 we assume the vertex b is not a cutpoint and we focus on the
cutset T = {b, a1 }. GT has two connected components that are subgraphs
of B1 , one containing the vertex a and one containing the isolated vertex
b1 . Hence there exists another block B2 different from B1 that contains the
vertex a1 . Let E(B2 ) = {{a1 , a′1 }}. By the same argument there exists a
block B3 such that E(B3 ) = {{a2 , a′2 }}. The set T1 = {b, a1 , a2 } is a cutset
with 5 components in GT 1 : the first containing a, the second (respectively
the third) containing a′1 (respectively a′2 ) and the 2 isolated vertices b1 and
b2 . Contradiction.
Lemma 3.3. Let G be a non-cactus bicyclic graph. Then J(G) is CohenMacaulay (respectively unmixed but not Cohen-Macaulay) if and only if
B(G) is a tree and G is decomposable into indecomposable graphs G1 , . . . , Gr ,
and such that G1 is one of the graphs in Figure 7 (respectively Figure 8).
Proof. Let (l1 , l2 , l3 ) ∈ N3 where li is the length of the paths Pi . By Remark
3.1 and Lemma 3.2 we have to study the following 4 cases:
(1, 2, 2), (1, 2, 3), (2, 2, 2), (2, 2, 3).
Case (1, 2, 2). Since l1 = 1 we obtain the graph on the left in Figure 7
that is Cohen-Macaulay. In fact, if we consider the cone from the vertex a
to the two connected components given by the isolated vertex f3 and the
path whose edges are {f1 , b} and {b, f2 }, by Theorem 3.8 of [9], we have the
assertion.
16
GIANCARLO RINALDO
f3 q❝
aq
q❝f2
❅
❅
❅qb
f1 q❝ ❅
fq❝1
✡❏
✡ ❏
q
b✡
❏qa
b1 q
qa1
❝qf2
❝qf3
Figure 7. Bicyclic Cohen-Macaulay non-cactus graphs.
Case (1, 2, 3). Since l1 = 1 we obtain the graph on the right in figure 7
removing the vertex f3 . Using similar argument to the one used in Proposition 2.2, we obtain that the only candidate is exactly the graph in figure
7 and
C(G) = {∅, {a}, {a1 }, {a, a1 }, {a, b}, {a, b1 }, {a1 , b}, {a, a1 , b}}.
We focus on the cutpoint a. By Corollary 1.5 we obtain J(G) = J(Ga ) ∩ Qa ,
moreover Qa = (xa , ya ) + J(P ) where J(P ) is the binomial edge ideal of the
path with
E(P ) = {{f1 , b}, {b, b1 }, {b1 , a1 }, {a1 , f3 }},
C(P ) = {∅, {a1 }, {b}, {b1 }, {a1 , b}}.
The ring S/Qa is Cohen-Macaulay of dimension 8 and it has two components: The path P on 5 vertices and the isolated vertex f2 . We also observe
that J(Ga ) is the cone from the vertex a1 to the vertex f3 and the graph
obtained attaching the edge {b1 , b} to the complete graph whose vertices are
{a, b, f1 , f2 }. Hence J(Ga ) is Cohen-Macaulay of dimension 8, too. Moreover Qa + J(Ga ) is a binomial edge ideal that is equal to the previous cone
removing the vertex a, hence it is Cohen-Macaulay of depth 7. By the Depth
Lemma applied on the following exact sequence the assertion follows
0 −→ S/J(G) −→ S/Qa ⊕ S/J(Ga ) −→ S/(Qa + J(Ga )) −→ 0.
Cases (2, 2, 2) and (2, 2, 3). These two cases are unmixed if and only if
we add the edges as in figure 8. Moreover we already found them in [10],
Example 3.2 and 3.3. In that paper by symbolic computation we observed
that they are not Cohen-Macaulay. In [2] there is an argument for the non
Cohen-Macaualyness of the bipartite one, the one on the right of figure 8.
Now we are ready to give the main result of the section.
Corollary 3.4. Let G be a bicyclic graph. Then J(G) is Cohen-Macaulay
if and only if B(G) is a tree, G is decomposable into indecomposable graphs
G1 , . . . , Gr , and such that one of the following cases occurs:
(1) G1 and G2 are in the set of unicyclic graphs in Figure 3;
(2) G1 is one of the bicyclic cactus graphs in Figure 4;
(3) G1 is one of the bicyclic non-cactus graphs in Figure 7.
COHEN–MACAULAY BINOMIAL EDGE IDEALS OF CACTUS GRAPHS
f1q❝
q
❅
❅
q ❅q
q
❅
❅
❅q
fq❝2
f1q❝
q
q
f❝q3
q
q
❅
❅
❅q
q
f❝q2
17
Figure 8. Bicyclic unmixed but non Cohen-Macaulay noncactus graphs.
Acknowledgements
The author was partially supported by GNSAGA of INdAM (Italy).
References
[1] A. Banerjee, L. Nunez-Betancourt, Graph Connectivity and Binomial edge ideals,
Proc. Amer. Math. Soc., Vol. 145 (2), 2017, pp. 487–499.
[2] D. Bolognini, A. Macchia, F. Strazzanti, Binomial edge ideals of bipartite graphs,
arXiv:1704.00152v1, 2017, pp. 1–29.
[3] CoCoATeam, CoCoA: a system for doing Computations in Commutative Algebra,
Available at http://cocoa.dima.unige.it
[4] V. Ene, J. Herzog, T. Hibi, Cohen-Macaulay binomial edge ideals, Nagoya Math. J.,
Vol. 204, 2011, pp. 57–68.
[5] F. Harary, Graph theory, Addison-Wesley series in Mathematics, 1972.
[6] J. Herzog, T. Hibi, F. Hreinsdottir, T. Kahle, J. Rauh, Binomial edge ideals and
conditional independence statements, Adv. in Appl. Math., Vol. 45, 2010, pp. 317–
333.
[7] D. Kiani, S. Saeedi Madani, Some Cohen-Macaulay and Unmixed Binomial Edge
Ideals, Comm. Algebra, Vol. 43, 2015, pp. 5434–5453.
[8] M. Ohtani, Graphs and ideals generated by some 2-minors, Comm. Algebra, Vol. 39,
2011, pp. 905–917.
[9] A. Rauf, G. Rinaldo, Construction of Cohen-Macaulay binomial edge ideals, Comm.
Algebra., Vol. 42, 2014, pp. 238–252.
[10] G. Rinaldo, Cohen-Macaulay binomial edge ideals of small deviation, Bull. Math. Soc.
Sci. Math. Roumanie, Tome 56(104) No. 4, 2013, 497–503.
[11] R.H. Villarreal, Cohen-Macaulay graphs, Manuscripta Math., Vol. 66, 1990, pp. 277–
293.
Department of Mathematics, University of Trento, via Sommarive, 14, 38123
Povo (Trento), Italy
| 0math.AC
|
arXiv:0711.4444v2 [cs.MS] 29 Nov 2007
INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET EN AUTOMATIQUE
Building the Tangent and Adjoint codes of the Ocean
General Circulation Model OPA with the Automatic
Differentiation tool TAPENADE
M.H. Tber — L. Hascoët — A. Vidard — B. Dauvergne
N° 6372
Novembre 2007
Thème NUM
apport
de recherche
Building the Tangent and Adjoint odes of the
O ean General Cir ulation Model OPA with the
Automati Dierentiation tool TAPENADE
∗
M.H. Tber
, L. Has oët
∗
†
, A. Vidard
, B. Dauvergne
∗
Thème NUM Systèmes numériques
Projet Tropi s
Rapport de re her he
Abstra t:
n° 6372 Novembre 2007 28 pages
The o ean general
ir ulation model OPA is developed by the
LODYC team at Paris VI university.
OPA has re ently undergone a major
rewriting, migrating to FORTRAN95, and its adjoint
ode needs to be re-
built. For earlier versions, the adjoint of OPA was written by hand at a high
development
ost.
We use the Automati
Dierentiation tool TAPENADE
to build me hani aly the tangent and adjoint
dierentiated
odes by
odes of OPA. We validate the
omparison with divided dieren es, and also with an
identi al twin experiment. We apply state-of-the-art methods to improve the
performan e of the adjoint
and Walther's binomial
ode.
In parti ular we implement the Griewank
he kpointing algorithm whi h gives us an optimal
trade-o between time and memory
onsumption. We apply a spe i
egy to dierentiate the iterative linear solver that
strat-
omes from the impli it time
stepping s heme.
Key-words:
OPA, general
ir ulation model, TAPENADE, Automati
ferentiation, reverse mode, Adjoint Code, Che kpointing
∗
†
INRIA Sophia Antipolis, Fran e
INRIA Grenoble, Fran e
Unité de recherche INRIA Sophia Antipolis
Dif-
Constru tion des odes Tangent et Adjoint du
modèle de ir ulation générale o éanique OPA
par l'outil de Diérentiation Automatique
TAPENADE
Résumé : Le modèle de
ir ulation générale o éanique OPA est développé
par l'équipe LODYC de l'université Paris VI. La nouvelle version 9 d'OPA
onstitue une évolution majeure, ave
TRAN95.
Les
en parti ulier une migration vers FOR-
odes Linéaire Tangent et Adjoint d'OPA, qui auparavant
étaient é rits à la main, doivent don
être redéveloppés. Nous utilisons l'outil
de Diérentiation Automatique TAPENADE pour
gent et Adjoint d'OPA 9.
raison ave
onstruire les
odes Tan-
Nous validons les dérivées obtenues par
ompa-
les Diéren es Divisées et sur deux appli ations test in luant des
expérien es jumelles. Nous utilisons le s héma de
he kpointing ré ursif bino-
mial de Griewank et Walther pour améliorer les performan es du
ode adjoint.
Nous utilisons une stratégie spé i que pour diérentier le solveur linéaire itératif provenant du s héma impli ite d'avan ement en temps.
montrent un
oût raisonnable, tant en
Nos résultats
onsommation mémoire que pour le
temps d'exé ution de l'adjoint.
Mots- lés :
OPA, Cir ulation O éanique, TAPENADE, Diérentiation Au-
tomatique, mode inverse, Code Adjoint, Che kpointing
Tangent and Adjoint dierentiation of OPA
3
1 Introdu tion
The development of tangent and adjoint models is an important step in addressing sensitivity analysis and variational data assimilation problems in O eanography. Sensitivity analysis is the study of how model output varies with
hanges in model inputs.
The sensitivity information given by the adjoint
model is used dire tly to gain an understanding of the physi al pro esses. In
data assimilation, one
onsiders a
ost fun tion whi h is a measure of the
model-data mis t. The adjoint sensitivities are used to build the gradient for
des ent algorithms. Similarly the tangent model is used in the
in remental algorithms [3℄ to linearize the
ontext of the
ost fun tion around a ba kground
ontrol. For the previous version 8 of the O ean General Cir ulation Model
OPA [17℄, Weaver et al [22℄ developed the numeri al tangent and adjoint
by hand using
odes
lassi al te hniques [5, 19℄. Sin e then, the OPA model has un-
dergone a major update. Parti ularly the new versions are fully rewritten in
FORTRAN95.
adjoint
In this paper, we report on the development of tangent and
odes of OPA using the Automati
Dierentiation (AD) tool TAPE-
NADE [12℄. A brief des ription of the OPA model and the
on guration used
in this work is given in the next se tion. In se tion 3 we present the prin iples of AD and how they are re e ted into the fun tionalities of the AD tool
TAPENADE. In se tion 4 we fo us on the most interesting di ulties that we
en ountered in the appli ation of AD to su h a large
ode. Se tion 5 shows
some experiments that validate our derivatives and presents two illustrative
appli ations, fo using on
omputational aspe ts rather than impli ations for
o eanography. An outlook of further work is given in the
on lusion.
2 The O ean General Cir ulation Model OPA
Developed by the LODYC team at Paris VI university, OPA is a exible o ean
ir ulation model that
on guration.
an be used either in a regional or in a global o ean
OPA is the o ean model
omponent of NEMO (Nu leus For
European Modelling of the O ean) and is widely used in the s ienti
om-
munity. Moreover it is be oming a major a tor in operational o eanography
(Mer ator, ECMWF, UK-Met o e) Its formulation is based on the so- alled
primitive equations for the temporal evolution of o ean velo ity
RR n° 6372
urrents, tem-
Tber et al.
4
perature and salinity in its three horizontal and verti al dimensions.
equations are derived from Navier-Stokes equations
These
oupled with a state equa-
tion for water density and heat equation, under Boussinesq and hydrostati
approximations.
U
Let us introdu e the following variables:
Uh + wk
(the subs ript
tial temperature,
S
h
the velo ity ve tor,
denotes the lo al horizontal ve tor),
the salinity,
p
the pressure and
ρ
T
U =
the poten-
the in-situ density. The
ve tor invariant form of the primitive equations in an orthogonal set of unit
ve tors linked to the earth are written as follows
∂Uh
∂t
∂p
∂z
∇·U
∂T
∂t
∂S
∂t
ρ
1
= − (∇ × U) × U + ∇ (U) 2
2
h
− f k × Uh −
1
∇h p + D U
ρ0
= −ρg
= 0
= −∇ · (T U) + D T
= −∇ · (SU) + D S
= ρ (T, S, p)
∇ is the generalized derivative ve tor operator, t the time, z the verti al
oordinate, ρ0 a referen e density, f the Coriolis a eleration, and g the gravity
U
T
S
a eleration. D , D and D are the parametrization of small s ale physi s for
where
momentum, temperature and salinity, in luding surfa e for ing terms. A full
des ription of the model basi s, dis retization, physi al and numeri al details
an be found in [17℄.
Through this paper, OPA is used in its global free surfa e
ORCA-2. In this
on guration
on guration the model uses a rotated grid with poles on
North Ameri a and Asia in order to avoid the singularity problem on the North
Pole. The spa e resolution is roughly equivalent to a geographi al mesh of 2°
by 1.3° with a meridional resolution of 0.5° near the Equator (see gure 1). The
Verti al domain, spreading from the surfa e to a depth of 5000m, is meshed
using 31 levels with levels 1 to 10 in the top 100 meters. The time step is 96
minutes so that there are 15 time steps per day. The model is for ed by heat,
freshwater, and momentum uxes from the atmosphere and/or the sea-i e.
INRIA
Tangent and Adjoint dierentiation of OPA
5
Figure 1: ORCA 2 Mesh
The solar radiation penetrates the upper layers of the o ean. Zero uxes of
heat and salt are applied through the bottom. On the lateral solid boundaries
a no-slip
ondition is also applied. Initialization of the model for temperature
and salinity is based on the Levitus et al. (1998)
limatology with a null initial
velo ity eld. For more details about the spa e time-domain and the o ean
physi s of ORCA-2, we refer to the page dedi ated to this
1
on guration in the
o ial website of NEMO-OPA .
The
to
on guration ORCA-2 is routinely used by MERCATOR/Meteo-Fran e
ompute the o eani
omponent of their seasonal fore asting system. The
size of OPA-9, 200 modules de ning 800 pro edures with over 100 000 lines of
FORTRAN95, makes it the largest appli ation dierentiated by TAPENADE
to date.
The
omputational kernel whi h is a tually dierentiated a
for 330 pro edures.
1 http://www.lody
RR n° 6372
.jussieu.fr/NEMO/general/des ription/ORCA_ on g.html
ounts
Tber et al.
6
3 Prin iples of AD and the tool TAPENADE
2
TAPENADE [12℄ is an AD tool developed by the Tropi s
team at INRIA.
Given the sour e of an original program that evaluates a mathemati al fun tion, and given a sele tion of input and output variables to be dierentiated,
TAPENADE produ es a new sour e program that
omputes the partial deriva-
tives of the sele ted outputs with respe t to the sele ted inputs.
Basi ally, TAPENADE does that by inserting additional statements into a
opy of the original program. Like other AD tools, TAPENADE is based on
the fundamental observation that the original program P, whatever its size and
m
n
run time, omputes a fun tion F, X∈IR 7→ Y ∈IR whi h is the omposition
of the elementary fun tions
words if
P
omputed by ea h run-time instru tion. In other
exe utes a sequen e of elementary statements
Ik , k ∈ [1..p],
then
P
a tually evaluates
F = fp ◦ fp−1 ◦ · · · ◦ f1 ,
where ea h
fk
is the fun tion implemented by
hain rule of derivative
derivatives of ea h
Calling
X0 = X
Y
Xk = fk (Xk−1 )
omponent of
and
variables, i.e. the su
Ik .
Therefore one an apply the
F ′ , i.e. the partial
al ulus to get the Ja obian matrix
essive
with respe t to ea h
the su
omponent of
X.
essive values of all intermediate
states of the memory throughout exe
ution of
P,
we get
′
F ′ (X) = fp′ (Xp−1 ) × fp−1
(Xp−2 ) × · · · × f1′ (X0 ) .
The derivatives
fk′
(1)
of ea h elementary instru tion are easily built, and must
be inserted in the dierentiated program so that ea h of them has the values
Xk−1
dire tly available for use. This pro ess yields analyti
are exa t up to numeri al a
derivatives, that
ura y.
In pra ti e, two sorts of derivatives are of parti ular importan e in s ienti
omputing: the tangent (or dire tional) derivatives, and the adjoint (or
reverse) derivatives.
In parti ular, tangent and adjoint are the two sorts of
derivative programs required for OPA, and TAPENADE provides both. The
′
tangent derivative is the produ t Ẏ = F (X) × Ẋ of the full Ja obian times a
dire tion
Ẋ
in the input spa e. From equation (1), we nd
′
Ẏ = F ′ (X) × Ẋ = fp′ (Xp−1) × fp−1
(Xp−2 ) × · · · × f1′ (X0 ) × Ẋ
2
(2)
http://www-sop.inria.fr/tropi s/
INRIA
Tangent and Adjoint dierentiation of OPA
whi h is most
u ts are mu h
7
heaply exe uted from right to left be ause matrix×ve tor prodheaper than matrix×matrix produ ts. This is also the most
onvenient exe ution order be ause it uses the intermediate values
P builds them.
X = F ′∗ (X) × Y of
Xk
in the
same order as the program
On the other hand the adjoint
derivative is the produ t
the
transposed Ja
Y in the output spa e. The resulting X
(Y · Y ). From equation (1), we nd
weight ve tor
dot produ t
obian times a
is the gradient of the
′∗
X = F ′∗ (X) × Y = f1′∗ (X0 ) × · · · × fp−1
(Xp−2 ) × fp′∗ (Xp−1 ) × Y
(3)
whi h is also most
heaply exe uted from right to left. However, this uses the
intermediate values
Xk
in the inverse of their building order in
Regarding the runtime
Ẏ
and adjoint
X
ost only a small multiple of the original program
slowdown fa tor is less than 4 in theory. In pra ti e it
the tangent, whereas it
below. Despite its higher
m runs
m
The
an be less than 2 for
ost, the adjoint
ode is still by large the
heapest
To get the gradient with the tangent mode would
of the tangent
is independent from
P.
an rea h up to 10 for the adjoint for a reason dis ussed
way to obtain gradients.
require
P.
ost for obtaining the derivatives, both tangent
ode, one per dimension of
X,
whereas this
ost
with the adjoint mode.
The di ulty of the adjoint mode lies in the fa t that it needs the intermediate values
Xk
in reverse order. To this end, TAPENADE basi ally uses a
two-sweeps strategy,
a
alled Store-All. In the rst sweep (the forward sweep),
opy of the original program
P
is run, together with Push statements that
store intermediate values on a sta k just before they get overwritten. In the
se ond sweep (the ba kward sweep), the derivative statements ompute the
′∗
elementary derivatives fk (Xk−1 ) for k = p down to 1, using Pop statements
to restore the intermediate values as they are required. This in urs a ost in
memory spa e as the maximum sta k size needed is attained at the end of
the forward sweep, and is thus proportional to the length of the program
P.
There is also a runtime penalty for these sta k manipulations. TAPENADE
implements a number of strategies [11℄ to mitigate this
data- ow analysis of the program's
values
Xk that need to be stored.
ontrol ow graph, redu ing the number of
However for very long programs su h as OPA,
involving unsteady simulations, Store-All
an not work alone.
ombines it with a storage/re omputation trade-o
RR n° 6372
ost, based on stati
TAPENADE
alled he kpointing.
Tber et al.
8
Che kpointing redu es the maximum sta k size at the
exe utions. Consider a pie e
C
of the original program
P.
ost of dupli ated
Che kpointing
illustrated on gure 2 means that during the main forward sweep,
C
C
as
pushes
no value on the sta k. When the ba kward sweep rea hes ba k to the pla e
where intermediate values are now missing on the sta k, it runs
C
a se ond
time, this time with the Store-All strategy i.e. pushing values on the sta k.
The ba kward sweep
an then resume safely.
To run
C
twi e requires that
enough of its input values, a snapshot, are stored but the size of a snapshot
is generally mu h less than the sta k size used by
down the adjoint program. When
C
is well
C.
hosen,
Obviously, this also slows
he kpointing
the peak size of the sta k by a fa tor of two. Che kpoints
whi h
an divide
an be nested, in
ase both the sta k's peak size and the adjoint runtime slowdown
grow as little as the logarithm of the size of
applies
he kpointing to ea h pro edure
an
P. In its default mode, TAPENADE
all.
{
C
successive
sweeps
Figure 2: Che kpointing applied to the program pie e
C.
Rightwards arrows
represent forward sweeps, thi k when they store intermediate values on the
sta k, thin otherwise.
Leftwards arrows represent ba kward sweeps.
Bla k
dots are stores, white dots are retrieves. Small dots are Push and Pops, big
dots are snapshots.
TAPENADE
apa ity to generate robust and e ient tangent and adjoint
odes has been demonstrated on several real-world test appli ations [15, 7, 1,
13, 16℄. Regarding the appli ation language, it
in FORTRAN. Taking into a
an handle programs written
ount the new programming
onstru ts provided
by FORTRAN95 has required an important programming eort in the past
few years, mostly to handle modules, stru tured data types, array notation,
pointers, and dynami
memory allo ation. Sin e the new OPA 9 is now written
in FORTRAN95, dierentiation of OPA is a very realisti
test for the new
TAPENADE 2.2.
INRIA
Tangent and Adjoint dierentiation of OPA
There exist several other AD tools.
9
Restri ting to the tools whi h, like
TAPENADE, operate by sour e transformation, provide tangent and adjoint
modes, use global program analysis to optimize the dierentiated
have demonstrated their appli ability on large industrial
odes, we
ode, and
an men-
tion TAF [4℄ a pioneer of AD for meteorology, now the standard AD tool for
the popular MIT Global Cir ulation Model. Unlike TAPENADE's, the adjoint
mode of TAF regenerates the intermediate values
an given initial point. This is
Xk
with Store-All strategy is getting blurred by nested
joint
su
by re omputation from
alled a Re ompute-All strategy. Comparison
odes grow more alike as more
he kpointing, as the ad-
he kpoints are inserted. OpenAD [20℄,
essor of ADIFOR and ADIC, uses the Store-All strategy. There are ex-
periments to also apply OpenAD to the MIT GCM. The tool Adol-C [10℄,
although using operator overloading instead of sour e transformation, is very
popular and has been applied su
adjoint mode
essfully to many industrial appli ations. Its
an be seen as an extension of the Store-All strategy: not only
the intermediate values are stored on the sta k, but also the
omputation graph
to be dierentiated. This allows the AD tool to perform further optimizations
on this graph, at the
ost of a higher memory
onsumption.
4 Applying TAPENADE to OPA
We generated working tangent and adjoint
odes for the
omputational kernel
of OPA, using TAPENADE. Depending on the nal appli ation (
f
se tion
5), the a tual fun tion to dierentiate as well as the input and output variables may be dierent, but the te hni al di ulties that we en ountered are
essentially the same. This se tion des ribes these points.
4.1
FORTRAN95
onstru ts
The new OPA 9 uses extensively the modular
had to extend the
all-graph internal representation of TAPENADE to handle
the nesting of modules and pro edures.
into the dierentiated
Be ause a module
Essentially this nesting is mirrored
ode.
an de ne private
ferentiated modules do not have a
RR n° 6372
onstru ts of FORTRAN95. We
omponents, subroutines in the dif-
ess to all variables of the original module.
Tber et al.
10
Therefore the dierentiated module must ontain its own opy of all the original
module's variables, types, and pro edures. This is a
dierentiation model: the dierentiated
the original
ode; it must
dierentiated
ode
ontain its own
hange in TAPENADE's
annot just
all or use parts of
opies of those. In other words, the
ode need not be linked with the original.
The interfa e me hanism of FORTRAN95 is a way to implement overloaded pro edures.
This is stati
overloading, whi h is resolved at
ompile
time. Therefore we had to extend TAPENADE type- he king phase to
pletely solve the
alls to interfa ed pro edures.
om-
Conversely, TAPENADE is
now able to generate interfa es on the dierentiated pro edures, so that the
general stru ture of the
ode is preserved.
The array notation of FORTRAN95 is used systemati ally in OPA. At the
same time, dierentiation requires that many
alls to intrinsi
fun tions be
split to propagate the derivatives. When these fun tions are used on arrays
("elemental" intrinsi s) TAPENADE must generate a
ode whi h is far from
trivial. For instan e the single statement from OPA:
zws(:,:,:) = SQRT(ABS(psal(:,:,:)))
generates in the adjoint mode
abs1 = ABS(psal(:,:,:))
mask = (psal(:,:,:) .GT. 0.0)
...
WHERE (abs1 .EQ. 0.0)
abs1b = 0.0
ELSEWHERE
abs1b = zwsb(:, :, :)/(2.0*SQRT(abs1))
END WHERE
WHERE (.NOT.mask(:, :, :))
psalb(:, :, :) = psalb(:, :, :) - abs1b
ELSEWHERE
psalb(:, :, :) = psalb(:, :, :) + abs1b
END WHERE
Without going in too mu h detail into the adjoint dierentiation model, we
observe that the test that is needed to prote t the dierentiated
ode against
INRIA
Tangent and Adjoint dierentiation of OPA
the non-dierentiability of
dierentiation of
ABS,
SQRT
11
at 0, as well as the test that
have been turned into
WHERE
ontrols the
onstru ts to keep the
runtime bene ts of array notation. Some temporary variables are introdu ed
automati ally to store
ontrol- ow de isions (e.g.
abs1
and
mask),
although
TAPENADE still doesn't do this in an optimal way on the example.
OPA uses pointers and dynami
DEALLOCATE).
memory allo ation ( alls to
ALLOCATE
and
This is an appli ation for the pointer analysis now available
in TAPENADE, nding whether a variable has a derivative, even when this
variable is a
essed through a pointer. Unfortunately, dynami
allo ation is
handled partly, i.e. only in the tangent mode of TAPENADE. In the adjoint
mode, we have no general strategy for memory allo ation and TAPENADE
sometimes
annot produ e a working
an allo ate should be a
ode. We understand that the adjoint of
DEALLOCATE,
and vi e-versa, but some
be made by hand on the dierentiated
4.2
hanges must
ode to make it work.
Che kpointing and hidden variables
OPA reads and writes several data les, not only during the pre- and postpro essing stages, but also during the
omputational kernel itself. Sour e terms
su h as the wind stress are being read at intermediate time steps. Also, some
modules and pro edures de ne private
but
annot be a
SAVE variables, whose value is preserved
essed from outside.
are just examples of a
Although unrelated, these two points
ommon problem: they
reentrant".
If a
alled pro edure modi es an internal
sible from the outside
alling
ontext to
an identi al result. Similarly if the
SAVE
an make a pro edure non
variable, it be omes impos-
all the pro edure a se ond time with
alled pro edure reads from a previously
opened le, and just moves the read pointer further in the le, then it be omes
impossible to
all the pro edure twi e and obtain the same values read.
Non reentrant pro edures are a problem for the
the adjoint mode. We saw in se tion 3 that
he kpointing strategy of
he kpointing relies on
alling the
he kpointed pie e twi e, in su h a way that the se ond
all is equivalent to the
rst. To this end, a su ient subset of the exe ution
ontext, the snapshot,
must be saved and restored. Hidden variables like an internal
the read pointer inside an opened le
RR n° 6372
SAVE variable or
annot be saved nor restored in general.
Tber et al.
12
When
then
he kpointing would require hidden variables to be put in the snapshot,
he kpointing should be forbidden.
Similarly, when a pro edure only allo ates some memory, the allo ation
must not be done twi e.
If this pro edure is
he kpointed, then one must
deallo ate the memory when restoring the snapshot before the dupli ate
all.
TAPENADE is not yet able to do this automati ally.
TAPENADE has some fun tionalities to
problem, but in all
ope with this hidden variables
ases intera tion with the user is ne essary. First, TAPE-
NADE issues a warning message when a subroutine
be ause of a private
SAVE
annot be
he kpointed
variable. The message is issued only when this vari-
able would be part of the snapshot for this pro edure. When this happened
for OPA, we just turned by hand the variables in question into publi
variables in the original
ode. In prin iple this
global
ould also be done automati-
ally. However there are only a handful su h variables, thus developing this is
not our priority.
When a subroutine is not reentrant be ause of I/O le pointers or be ause
of isolated memory allo ation or deallo ation, then TAPENADE lets the user
label the subroutine so that it must not be
he kpointed. For OPA, we took
another strategy: we modi ed the main I/O subroutines so that they always
rst make sure that the le is opened and then only use dire t read into the
le without using a read I/O pointer. Thus all I/O subroutines are reentrant.
4.3
Binomial Che kpointing
Automati
Dierentiation of OPA is one of the most ambitious appli ations
of TAPENADE so far. It means building the adjoint of a pie e of
ode that
performs an unsteady nonlinear simulation over a very large number of time
steps. Ea h time step
omputes a new state whose size ranges in the hundreds
of megabytes. In adjoint mode if no
he kpointing was applied, whi h means
that all intermediate values were to be stored on a sta k, we
ould exe ute
only a handful of time steps before we run out of memory even on our largest
workstation. Che kpointing is
ompulsory to
ompute the adjoint over several
thousands of time steps, whi h is our goal.
We saw in se tion 3 that TAPENADE applies
of subroutine
alls, i.e. ea h
all is
he kpointing at the level
he kpointed. This easy strategy is often
INRIA
Tangent and Adjoint dierentiation of OPA
far from optimal. On one hand several
13
alls are better not
TAPENADE now oers the option to mark sele ted
On the other hand,
he kpointed, and
alls for
not
he kpointing.
he kpointing should be applied at other lo ations. For
example at the top level of the simulation program is a loop over many time
steps.
We de nitely need an e ient
he kpointing s heme applied at this
level of time iterations.
One
lassi al solution used by TAF on the MIT GCM
multi level re ursive
he kpointing. Basi ally, it splits the
ode [14℄, is
alled
omplete time inter-
val into a small number of equidistant intervals, then apply the same strategy
to ea h of the sub-intervals.
For instan e 64 time steps
an be split into 4
large intervals of 4 small intervals of 4 time steps, as sket hed on gure 3. This
onsumes a maximum of 9 simultaneous snapshots, and the average number
of dupli ate exe utions for a time step is
1000 time steps
2.25.
In a more realisti
situation,
an be split into 10 large intervals of 10 small intervals of
10 time steps, and one
an gure out that this
onsumes a maximum of 27
simultaneous snapshots, and the average number of dupli ate exe utions for a
time step is
2.7.
0
16
32
48
52
56
60 61 62
57 58
53 54
49 50
36
40
44 45 46
41 42
37 38
33 34
20
24
28 29 30
25 26
21 22
17 18
4
8
12 13 14
9 10
5
1
6
2
Figure 3: Three-levels
ward
he kpointing with 64 time steps and 9 snapshots. For-
omputations go right, adjoint
omputations go left. Bla k
resent writing/taking a snapshot, white
snapshot.
RR n° 6372
ir les rep-
ir les represent reading an available
Tber et al.
14
However, it was shown in [21℄ that this strategy is not optimal. Under the
reasonable assumptions that all time steps
ost the same run time, and that
the snapshot needed to run again from time step
from step
n
to any later step
n + x,
the optimal distribution of nested
n to n+1 is the same as to run
Griewank and Walther have
hara terized
he kpoints, whi h follows a binomial law.
With this optimal strategy, both spatial and temporal omplexity of the adjoint
ode grow logarithmi ally with respe t to the number of time steps of the
original simulation.
In other words, both the slowdown fa tor whi h grows
like the number of times ea h time step is exe uted, and the memory whi h
grows like the number of simultaneous snapshots, grow logarithmi ally with
the total number of time steps.
In real appli ations, run-time and memory spa e do not behave symmetrially. One
an always wait a little longer for the result, whereas the memory
spa e is bounded. Therefore the maximum number of snapshots
d
that
an
be stored simultaneously is xed.
Then [8℄ shows that the optimal strategy
th
gives a slowdown fa tor that grows only like the d root of the total number of
time steps, whi h is still very good. Figure 4 shows the optimal
he kpointing
strategy for the same problem as gure 3 i.e. 64 time steps with memory for
9 snapshots.
only
2.
The average number of dupli ate exe utions for a time step is
For the more realisti
situation (1000 time steps and memory for 27
2.57.
snapshots) the average number of dupli ate exe utions is only
We implemented this optimal strategy in the adjoint
ode of OPA. We
made our rst experiments by hand modi ation of the adjoint
ode produ ed
by TAPENADE. Still, TAPENADE produ ed automati ally the pro edures
that store and retrieve the snapshot, and therefore the hand modi ation was
benign: given the number of time steps, a general pro edure
3
s hedules the
optimal sequen e of a tions (store snapshot, retrieve snapshot, run time step,
run adjoint time step) to dierentiate the
omplete simulation. Further ver-
sions of TAPENADE will fully automate this pro ess.
Figure 5 shows the
performan es on OPA. They are in good agreement with the theory. Noti e in
parti ular the two small in e tion points on the
and 800 iterations.
urve around 150 iterations
Going ba k to the optimality proof in [8℄, we see that
the optimal strategy is parti ularly e ient when the number of time steps is
3A
FORTRAN95 implementation of this s heduling pro edure an be found in www.inriasop/tropi s/ftp/Hi ham_Tber/
INRIA
Tangent and Adjoint dierentiation of OPA
0
16
30
15
38
45
51
56
60
62
57 58
52 53 54
46 47 48 49
39 40 41 42 43
31 32 33 34 35 36
19
22
24 25 26 27 28
20
17
3
6
9
10 11 12 13 14
7
4
1
Figure 4: Optimal binomial
he kpointing with 64 time steps and 9 snapshots
7.5
Slowdown factor
7
6.5
6
5.5
5
4.5
200
400
600
800
1000
1200
Total number of time steps
Figure 5: Optimal binomial he kpointing with 15 snapshots: slowdown fa tor
as a fun tion of the total length of the initial simulation. The slowdown fa tor
is the run-time ratio of the adjoint
RR n° 6372
ode
ompared to the original
ode.
Tber et al.
16
exa tly
(d + t)!
d!t!
where d is the number of snapshots and t is the number of dupli ate exe utions
allowed per time step. For our target ma hine d = 15 and we nd η(15, 2) =
136 and η(15, 3) = 816, whi h orresponds to the in e tion points of gure 5.
η(d, t) =
For the previous version OPA 8, the adjoint was written by hand. Nevertheless, even a hand-written adjoint must implement strategies to retrieve intermediate states in reverse order that is, something very
lose to
Looking at this hand-written adjoint, we rst observe that the
he kpointing.
he kpointing
strategy is neither multi level nor optimal binomial. It is more like a single level
strategy, with one snapshot stored every xed number of time steps. During the
reverse sweep, states between two stored snapshots are rebuilt approximately
using linear interpolation. The advantage is that few time steps are evaluated
twi e, and therefore the slowdown fa tor remains well below 4. We
an see at
least two drawba ks. First, this hand manipulation requires deep knowledge of
the original program and of the underlying equations. This method does not
blend easily with Automati
Dierentiation. It is not yet automated in any
AD tool and therefore tedious and error-prone
ode manipulations would still
be ne essary. Se ond, this introdu es approximation errors into the
omputed
derivatives, whose mathemati al behavior is un lear. The gradient obtained in
the end is used in
omplex optimizations or data-assimilation loops, and small
errors may result in poor
onvergen e. In any
ase, for very large numbers of
time steps, we believe a trade-o between exa t binomial
he kpointing and
approximate interpolation is worth experimenting. Interpolation is probably
good enough for many variables that vary very slowly, and whi h
ould be
designated by the end-user, and only the other variables would need to be
stored.
4.4
Iterative linear solver
The OPA model solves an ellipti
equation at the end of ea h time step, using
an iterative method that generates a sequen e of approximations of the exa t
solution. The me hani al appli ation of AD on this kind of methods gives a
sequen e of derivatives of the approximate solutions with the same number
INRIA
Tangent and Adjoint dierentiation of OPA
17
of iterations as the original solver. The reason is that AD keeps the ow of
ontrol of the original program in the dierentiated program. In parti ular the
onvergen e tests are still based only on the non-dierentiated variables. Naturally, one may ask whether and how AD-produ ed derivatives are reasonable
approximations to the desired derivative of the exa t solution. The issues of
derivative
onvergen e for iterative solvers in relation to AD are dis ussed in
[6, 9, 2℄.
OPA provides two alternative algorithms to solve the ellipti
PCG for Pre onditioned Conjugate Gradient, and SOR for Su
Relaxation method. Both algorithms give
equation:
essive Over-
orre t results for the original
ode,
but PCG is generally preferred thanks to its e ien y and ve torization properties. However, the AD-dierentiated
algorithms. Figure 6
tives obtained by divided dieren es.
with the SOR algorithm remain
reases.
be ome
On the
ode gives dierent results using the two
ompares the AD-derivatives with approximate derivaWe see that the derivatives obtained
orre t when the number of time steps in-
ontrary, the derivatives obtained with the PCG algorithm
ompletely wrong after 80 time steps. Noti e that this o
urs in tan-
gent mode as well as in adjoint mode: the derivatives obtained with PCG,
although wrong, remain identi al in tangent and adjoint. Our explanation is
that ea h iteration of PCG involves the
omputation of s alar produ ts of vari-
ables that depend on the state ve tor, thus making the numeri al algorithm
nonlinear even though the ellipti
equation is linear. In [6℄, Gilbert has shown
that the appli ation of AD to a xed point iteration gives a derivative xed
point iteration that
in the
onverges R-linearly to the desired derivative in parti ular
ase of a large
this is not the
ontra tive iterate or se ant updating.
Unfortunately
ase for quasi-Newton iterative solvers su h as PCG, for whi h
there is no similar
onvergen e result to our knowledge.
To solve this problem for the tangent-dierentiated OPA we exploit the
linearity of the ellipti
system, and for the adjoint-dierentiated OPA we ex-
ploit the self adjointness property of the ellipti
operator [22℄. We
an thus
use the original PCG routine itself to solve for the dierentiated linear systems. Pra ti ally, we do this using the so- alled bla k-box feature provided
in TAPENADE. Figure 6 shows that (here for the tangent mode) the PCG
gives the same a
RR n° 6372
ura y as the SOR solver.
18
Tber et al.
Figure 6: evolution of the relative error between tangent derivative and divided
dieren es, for the three strategies: SOR and straightforward AD, PCG and
straightforward AD, PCG with the bla k box strategy
INRIA
Tangent and Adjoint dierentiation of OPA
19
In another experiment, we tried to use straightforward AD with the PCG
solver, but this time xing the number of PCG iterations to some very high
value. We observed that the derivatives be ome
dieren es. This
expensive and the
oherent again with divided
ould be another way to solve our problem, but it is
ertainly
hoi e of the high iteration number is deli ate. This prob-
lem de nitely deserves further study, and
on rms the general re ommendation
not to dierentiate solvers of a nonlinear kind, and use a bla k-box strategy
instead.
5 Validation Experiments
5.1
The
Corre tness test
lassi al way to
gent and adjoint
he k for
orre tness of the automati ally generated tan-
odes is as follows:
1. Choose an arbitrary input
X
and and arbitrary dire tion
Ẋ .
Compute
the Divided Dieren e
DD =
for a good enough small
F (X + εẊ) − F (X)
ε
ε.
2. Using the tangent dierentiated program,
3. Using the adjoint dierentiated program,
and nally
test for the
derivative
he k that
Ẏ = F ′ (X) × Ẋ .
ompute
ompute
X = F ′∗ (X) × Ẏ
(DD · DD) = (Ẏ · Ẏ ) = (X · Ẋ).
We performed this
omplete global ORCA-2 simulation on 1000 time steps and its
odes.
The results are shown in table 1.
The values mat h, and
Table 1: Dot produ t test for 1000 time steps
(DD · DD) (ε = 10−7 )
(Ẏ · Ẏ )
(X · Ẋ)
RR n° 6372
4.405352760987440e+08
4.405346876439977e+08
4.405346876439867e+08
Tber et al.
20
(Ẏ · Ẏ )
and
(X · Ẋ)
mat h very well, up to the last few digits, whi h shows
that the tangent and adjoint
in a dierent
of
odes really
ompute the same derivatives, only
omputation order as shown by equations (2) and (3). The values
(DD · DD)
and
(Ẏ · Ẏ )
don't mat h so well, be ause of the weakness of the
Divided Dieren es approximation. Figure 7 shows this weakness: For a small
0
10
−2
10
−4
10
−6
10 −12
10
−10
−8
10
10
−6
ε
10
−4
−2
10
10
Figure 7: Relative error of Divided Dieren es with respe t to AD-generated
derivatives,
value of
of
ε,
ε,
omputed for various values of of the step size
the dominant error is due to ma hine a
ura y. For a large value
the dominant error is due to the se ond derivatives of
minimizes both errors, but
5.2
ε
annot eliminate them
F.
The best
ε
ompletely.
Sensitivity analysis on a long simulation
One of the main appli ation of adjoint models is the sensitivity analysis i.e. the
study of how model output varies with
hanges in model inputs. Using dire t
or statisti al methods would require many integration of the non linear model
while one adjoint model integration is enough to ompute this sensitivity. As an
example, gure 8 shows the output map of the sensitivity of the North Atlanti
meridional heat ux at 29°N to
hanges in the initial sea surfa e temperature
(SSTt0 ) over one year integration period, starting January 1, 1998.
This is
INRIA
Tangent and Adjoint dierentiation of OPA
done by
omputing the gradient respe t to
J=
Z
tN
t0
where
Ω
v
SSTt0
of
T.v dxdzdt
Ω
ross se tion at 29°N in the North Atlanti ,
is the zonal
temperature and
ZZ
21
is the meridional
T
is the
urrent velo ity.
Contours in gure 8 show where variation of initial SST would ee t the
most upon heat transport at 29°N. It shows large s ale patterns mainly lo ated
north of the 29°N parallel and in the Caribbean sea with a strong spot o
Moro
o. These results are
onsistent with those obtained by Marotzke et al.
([18℄)
This map was
omputed by the TAPENADE-generated adjoint of OPA on
the global ORCA2 grid, over 5475 time steps (1 year). This experiment was
done with the SOR algorithm as the iterative linear solver. The TAPENADEgenerated adjoint
omputed this sensitivity map in a time that is only 8.03
times that of the original simulation.
5.3
Data Assimilation
For further validation of the automati ally generated derivatives, we
out a data assimilation experiment.
arried
This was done in a so- alled twin ex-
periment framework whereby the dire t model traje tory is used to generate
syntheti
observations.
The initial sea surfa e temperature is perturbed by
a white noise and it has to be re overed using variational data assimilation
te hniques. Syntheti
observation are given by the sea surfa e height (SSH)
and the sea surfa e salinity (SSS) generated from the model's original outputs
starting from the unperturbed SST.
The
ost fun tion to be minimised is
J(SST (t0 )) =
Z
tN
k SSH(t) − SSH o(t) k2 + k SSS(t) − SSS o (t) k2 dt
(4)
t0
Where the supers ript
SSS(t)
For
o
stands for syntheti
observation and
SSH(t)
and
are model output.
omputing
ost issues, only the Antar ti
zoom of ORCA2 is
onsid-
ered, the minimisation is done an iterative gradient sear h algorithm where
RR n° 6372
Tber et al.
22
Figure 8: Sensitivity map of the North Atlanti
line), with respe t to
heat transport at 29°N (dotted
hanges in the initial surfa e temperature
INRIA
Tangent and Adjoint dierentiation of OPA
the gradient of
J
is
23
umputed using adjoint te hniques. Figure 9 illustrates
the performan e of the optimization loop for an integration period of 1 month
i.e. 450 time steps. The
ost fun tion de reases by two orders of magnitude.
Figure 10 indi ates that the true solution (top panel) is re overed with a
good approximation (bottom panel) from the randomly perturbed one(middle
panel), showing the quality of the derivatives obtained.
6
Cost function
10
5
10
4
10 0
10
1
2
10
10
Iterations
Figure 9: Twin experiment: Convergen e of the
ost fun tion
6 Con lusion and Outlook
The eort to build the tangent and adjoint
odes for the previous version 8
of the OPA o ean General Cir ulation Model has
ost several months devel-
opment from an experien ed resear her. For the new version OPA 9 written
in FORTRAN95, the use of the AD tool TAPENADE signi antly redu es
this eort. Our rst numeri al appli ations show the quality of the derivatives
obtained. This works validates the
hoi e of AD as the strategy to obtain the
tangent and adjoint for OPA 9, and for the versions to
ome.
At the same time, OPA is the largest FORTRAN95 appli ation dierentiated with TAPENADE. This work has pointed at a number of limitations of
RR n° 6372
24
Tber et al.
Figure 10: Twin experiment: True eld (top), Initial perturbed eld (middle)
and identi ed optimal sea surfa e temperatures (bottom)
INRIA
Tangent and Adjoint dierentiation of OPA
25
TAPENADE that have been lifted. Other limitations remain, su h as the nonreentrant pro edures, whi h need to be addressed in future work. Su
dierentiation of OPA de nitely in reases our
essful
on den e in TAPENADE.
This works is also an additional illustration of the superiority of the binomial
he kpointing strategy,
ompared to multi level
he kpointing. By the
standards of other appli ation elds, e.g. CFD, a slowdown of the adjoint
of only 7 for a nonsteady simulation on 1000 time steps would be
ode
onsidered
very good. By the standards of weather simulation or o ean modeling however, s ientists expe t yet faster adjoints, at the
tion. Even if we
ost of a radi al approxima-
onsider that these approximations
hange the mathemati al
nature of the optimization pro ess, we understand they are ne essary and we
shall study how they
an be proposed as an option by the AD tool.
This work has underlined several dire tions for further resear h in AD and
AD tools. Some of them are already being studied by resear hers in our groups.
Considering the appli ation language, two
onstru ts need to be dierentiated
better:
The next experiment to be made very soon is to apply TAPENADE to
the parallelized version of OPA. This is ne essary before the generated
tangent and adjoint
odes
an be used in produ tion
ontext.
The OPA sour e makes extensive use of the prepro essor dire tives su h
as
#IFDEF. TAPENADE does not deal with these dire
tives be ause they
do not respe t the synta ti stru ture of a ode. Handling these dire tives
in the AD tool is in our opinion hopeless. What might be done though,
is to generate dierentiated
odes for ea h possible prepro essed
and devise a tool to put the dire tives ba k into the dierentiated
This is made easier if the dierentiated
of the original, as is the
ode
ode,
odes.
losely follows the stru ture
ase with TAPENADE.
Considering spe i ally adjoint dierentiation, we hope to obtain more e ient
ode through a more systemati
ellipti
operator.
exploitation of self-adjointness, e.g.
We also hope to optimize the
its present version, TAPENADE applies
Using pro ling information, we believe we
whi h
he kpointing is useless or
he kpointing strategy.
he kpointing to ea h pro edure
an dete t several pro edure
In
all.
alls for
ounter-produ tive. TAPENADE is already
able to use this information to produ e a better adjoint.
RR n° 6372
of the
Tber et al.
26
Referen es
[1℄ W. Castaings, D. Dartus, M. Honnorat, F.-X. Le Dimet, Y. Loukili, and
J. Monnier. Automati
dierentiation: A tool for variational data assimi-
lation and adjoint sensitivity analysis for ood modeling. In Bü ker et al,
editor,
Automati Dierentiation: Appli ations, Theory, and Implemen-
tations, LNCSE, pages 249262. Springer, 2005.
[2℄ B. Christianson. Reverse a
Opti-
umulation and attra tive xed points.
mization Methods and Software, 3:311326, 1994.
[3℄ P. Courtier, J.-N. Thépaut, and A. Hollingsworth. A strategy for operational implementation of 4d-var, using an in remental approa h.
Meteorol. So ., 120:13671387, 1994.
[4℄ R. Giering.
Q. J. R.
Tangent linear and Adjoint Model Compiler, Users manual,
1997. http://www.autodi. om/tam .
[5℄ R. Giering and T. Kaminski. Re ipes for adjoint
ode
ACM
onstru tion.
Transa tions on Mathemati al Software, 24(4):437474, 1998.
[6℄ J.C. Gilbert. Automati
dierentiation and iterative pro esses.
tion Methods and Software, 1:1321, 1992.
[7℄ M.B. Giles, D. Ghate, and M.C. Duta. Using automati
adjoint CFD
ode development. In Uthup et al, editor,
Aerospa e Design and Optimization,
Optimiza-
dierentiation for
Re ent Trends in
pages 363373. Tata-M Graw Hill,
New Delhi, 2005. Post-SAROD-2005, Bangalore, India.
[8℄ A. Griewank. A hieving logarithmi
plexity in reverse automati
Software, 1:3554, 1992.
growth of temporal and spatial
dierentiation.
om-
Optimization Methods and
[9℄ A. Griewank, C. Bis hof, G. Corliss, A. Carle, and K. Williamson. Derivative
onvergen e for iterative equation solvers.
Software, 2:321355, 1993.
Optimization Methods and
INRIA
Tangent and Adjoint dierentiation of OPA
27
[10℄ A. Griewank, D. Juedes, J. Srinivasan, and C. Tyner. ADOL-C, a pa kage
for the automati
dierentiation of algorithms written in C/C++.
Trans. Math. Software, 22(2):131167, 1996.
[11℄ L. Has oët and M. Araya-Polo.
The adjoint data- ow analyses:
malization, properties, and appli ations.
Appli ations, Theory, and Tools, Le
In
ACM
For-
Automati Dierentiation:
ture Notes in Computational S ien e
and Engineering. Springer, 2005. Sele ted papers from AD2004 Chi ago.
[12℄ L. Has oët and V. Pas ual. Tapenade 2.1 user's guide. Te hni al Report
0300, INRIA, 2004.
[13℄ L. Has oët, M. Vázquez, and A. Dervieux. Automati
optimum design, applied to soni
dierentiation for
boom redu tion. In Kumar et al., editor,
Pro eedings of ICCSA'03, Montreal, Canada,
pages 8594. LNCS 2668,
Springer, 2003.
[14℄ Heimba h, P. and Hill, C. and Giering, R.
of the parallel MIT general
dierentiation.
An e ient exa t adjoint
ir ulation model, generated via automati
Future Generation Computer Systems,
21(8):13561371,
2005.
[15℄ J. Kim, E. Hunke, and W. Lips omb. Sensitivity analysis and parameter
tuning s heme for global sea-i e modeling.
O ean Modeling Journal, 14(1
2):6180, 2006.
[16℄ C. Lauvernet, F. Baret, L. Has oët, and F.-X. LeDimet. Improved estimates of vegetation biophysi al variables from meris toa images by using
spatial and temporal
Beijing, China, 2005.
onstraints.
In
pro eedings of the 9th ISPMSRS,
[17℄ G. Made , P. Dele luse, M. Imbard, and C. Levy. Opa8.1 o ean general
ir ulation model referen e manual. Te hni al report, Pole de Modelisation, IPSL, 1998.
[18℄ J. Marotzke, R. Giering, K.Q. Zhang, D. Stammer, C. Hill, and T. Lee.
Constru tion of the adjoint MIT o ean General Cir ulation Model and
RR n° 6372
Tber et al.
28
appli ation to atlanti
heat transport sensitivity.
J. Geophys. Res.,
(104(C12)):29,52929,548, 1999.
[19℄ O. Talagrand. The use of adjoint equations in numeri al modeling of the
Automati Dierentiation of Algorithms: Theory, Implementation and Appliation, pages 169180, 1991. Philadelphia, Penn: SIAM.
atmospheri
ir ulation. In A. Griewank and G. Corliss, editors,
[20℄ J. Utke, U. Naumann, M. Fagan, N. Tallent, Strout M., P. Heimba h,
C. Hill, and C. Wuns h.
Automati
OpenAD/F: A modular, open-sour e tool for
Dierentiation of Fortran
odes. Te hni al report ANL/MCS-
P1230-0205, Argonne National Laboratory, 2006.
Submitted to ACM
TOMS.
[21℄ A. Walther and A. Griewank. Advantages of binomial
he kpointing for
Numeri al
Mathemati s and Advan ed Appli ations, pages 834843. Springer, Berlin,
memory-redu ed adjoint al ulations. In Feistauer et al, editor,
2003. Pro eeding of ENUMATH 2003.
[22℄ A. Weaver, J. Vialard, and D. Anderson.
Three- and four-dimensional
variational assimilation with an o ean general
tropi al pa i
he ks.
ir ulation model of the
o ean: I. formulation, internal diagnosti s and
Monthly Weather Review, 131(7):13601378, 2003.
onsisten y
INRIA
Unité de recherche INRIA Sophia Antipolis
2004, route des Lucioles - BP 93 - 06902 Sophia Antipolis Cedex (France)
Unité de recherche INRIA Futurs : Parc Club Orsay Université - ZAC des Vignes
4, rue Jacques Monod - 91893 ORSAY Cedex (France)
Unité de recherche INRIA Lorraine : LORIA, Technopôle de Nancy-Brabois - Campus scientifique
615, rue du Jardin Botanique - BP 101 - 54602 Villers-lès-Nancy Cedex (France)
Unité de recherche INRIA Rennes : IRISA, Campus universitaire de Beaulieu - 35042 Rennes Cedex (France)
Unité de recherche INRIA Rhône-Alpes : 655, avenue de l’Europe - 38334 Montbonnot Saint-Ismier (France)
Unité de recherche INRIA Rocquencourt : Domaine de Voluceau - Rocquencourt - BP 105 - 78153 Le Chesnay Cedex (France)
Éditeur
INRIA - Domaine de Voluceau - Rocquencourt, BP 105 - 78153 Le Chesnay Cedex (France)
http://www.inria.fr
ISSN 0249-6399
| 5cs.CE
|
arXiv:1606.04074v2 [cs.AR] 18 Jun 2016
ENTRA: Whole-Systems Energy Transparency
Kerstin Edera , John P. Gallagherb,c,∗, Pedro López-Garcı́ac,e , Henk Mullerd ,
Zorana Bankovićc , Kyriakos Georgioua , Rémy Haemmerléc ,
Manuel V. Hermenegildoc,f , Bishoksan Kafleb , Steve Kerrisona ,
Maja Kirkebyb , Maximiliano Klemenc , Xueliang Lib , Umer Liqatc ,
Jeremy Morsea , Morten Rhigerb , Mads Rosendahlb
a University
of Bristol, UK
University, Denmark
c IMDEA Software Institute, Spain
d XMOS Ltd., Bristol, UK
e Spanish Council for Scientific Research, Spain
f Technical University of Madrid, Spain
b Roskilde
Abstract
Promoting energy efficiency to a first class system design goal is an important
research challenge. Although more energy-efficient hardware can be designed,
it is software that controls the hardware; for a given system the potential for
energy savings is likely to be much greater at the higher levels of abstraction
in the system stack. Thus the greatest savings are expected from energy-aware
software development, which is the vision of the EU ENTRA project. This
article presents the concept of energy transparency as a foundation for energyaware software development. We show how energy modelling of hardware is
combined with static analysis to allow the programmer to understand the energy
consumption of a program without executing it, thus enabling exploration of
the design space taking energy into consideration. The paper concludes by
summarising the current and future challenges identified in the ENTRA project.
1. Introduction
Energy efficiency is a major concern in systems engineering. The EU’s Future
and Emerging Technologies MINECC programme aims to “lay the foundations
for radically new technologies for computation that strive for the theoretical
limits in energy consumption.” The research objectives range from physics to
software; they include, among others, new elementary devices, as well as “software models and programming methods supporting the strive for the energetic
limit.” The ENTRA project, entraproject.eu, addresses the latter objective;
∗ Corresponding
author
Email address: [email protected] (John P. Gallagher)
Preprint submitted to MICPRO
June 21, 2016
WP7
Dissemination
WP1
WP4
Energy-aware
software engineering
tested on
Energy
optimization
enables
enables
Energy
Transparency
Case Studies
measured against
Benchmarks
WP6
WP5
relies on
relies on
Energy
Modelling
Program
Analysis
WP2
WP3
Project
Management
WP8
Figure 1: Overview of the ENTRA project work plan
we focus on energy transparency, which we regard as a key prerequisite for new
energy-aware system development methods and tools.
The ENTRA project ran from October 2012 to December 2015 (39 months)
and was funded by the European Commission under the 7th Framework Programme. The consortium contained three research institutions and one industrial partner specialising in the design of advanced multicore microcontrollers
(XMOS xCORE). The overview of the project structure is shown in Figure 1.
The foundations for the central concept of energy transparency were developed
in two work packages (WP2 and WP3) on energy modelling and energy analysis respectively. Energy transparency enables energy optimisations, studied in
WP4. WP1 concerned the development of tools and techniques applicable in
energy-aware software development. Finally there were work packages dealing
with benchmarking, evaluation, dissemination and project management. This
paper summarises mainly the outcomes of work packages WP1, WP2, WP3,
WP4 and WP6. The public deliverables of the project are all available on the
project website http://entraproject.eu.
After this introduction, we discuss the two main areas of research supporting
energy transparency. Section 2 presents approaches for building models of software energy consumption at different levels of abstraction. Section 3 contains an
overview of static resource analysis techniques, showing how an energy model
2
can be used in analysis of a program’s energy consumption. Section 4 summarises the role of energy transparency in energy-aware software development,
discusses the achievements in the project so far, and outlines current challenges
and directions for future research.
1.1. Energy-Aware Computing
Energy-aware computing is a research challenge that requires investigating
the entire system stack from application software and algorithms, via programming languages, compilers, instruction sets and micro architectures, to the design and manufacture of the hardware. This is because energy is consumed
by the hardware performing computations, but the control over the computation ultimately lies within the applications running on the hardware. While
hardware can be designed to save a modest amount of energy, the potential for
savings is far greater at the higher levels of abstraction in the system stack.
An estimate from Intel [1] is that energy-efficient software can realize savings of
a factor of three to five beyond what can be achieved through energy efficient
hardware. Roy and Johnson [2] list five objectives that “help make software design decisions consistent with the objectives of power minimization”: match the
algorithm to the hardware; minimize memory size and expensive memory accesses; optimise the performance, making maximum use of available parallelism;
take advantage of hardware support for power management; and select instructions, sequence them, and order operations in a way that minimizes switching
in the CPU and datapath. To achieve these objectives requires the programmer
and/or the tools to understand the relationship between code and energy usage.
Energy Transparency aims to enable exactly this.
1.2. Energy Transparency
The concept of energy transparency is at odds with the trend in modern
software engineering - the desire to abstract away machine-level details using
high-level languages, abstract data types and classes, libraries and layers of interpretation or compilation, in the interests of portability, programmer productivity, understandability and software reuse. By contrast, energy transparency
requires making visible how software impacts on energy consumption when executed on hardware. Availability of this information enables system designers
to find the optimal trade-off between performance, accuracy, and energy usage
of a computation. To achieve energy transparency, models of how energy is
consumed during a computation are required. As will be discussed in Section 2,
such models can be established at different levels of abstraction, ranging from
models that characterize individual functional hardware blocks [3], via Instruction Set Architecture (ISA) characterization models [4, 5, 6], to models based
on intermediate representations used by the compiler [7, 8]. The final energy
models provide information that feeds into static resource usage analysis algorithms [9, 10, 11, 12, 13, 14, 15, 16], where they represent the energy usage of
elementary parts of the computation. This is discussed in Section 3.
3
2. Energy Modelling
Energy models can rely on information at several possible abstraction levels,
from gate-level hardware description, through functional block and Instruction
Set Architecture (ISA), up to performance counter or transaction based abstractions. Energy models at higher levels tend to be faster to use, but have lower
accuracy than models at lower levels of abstraction. In ENTRA, the aim was
to provide accurate modelling that can be exploited through analysis that is
applied in order to estimate the energy consumed by software.
2.1. Defining and constructing an energy model
The ISA is a practical level of abstraction for energy modelling, because it
expresses underlying hardware operations and their relationship with the intent
of the software. Constructing a model at this level gives us the following benefits: energy costs can be attributed directly to individual machine instructions
as output by the back end of the compiler; instruction properties and energy
consumption are strongly correlated, e.g. energy consumption typically increases
with increasing numbers of operands; and machine instructions can be traced
back to the original source code statements written by the software developer,
as well as to various intermediate representations.
However, energy modelling at the ISA level requires additional effort in order to produce useful models: instruction costs must be captured through a
profiling suite and measurement of device power; in addition, indirect or statistical approaches are required to characterise instructions that cannot be profiled
through direct measurements. Furthermore, for multi-threaded architectures
other properties such as the cost of running multiple threads and the cost of
idle periods must be determined.
Our target architecture for energy modelling and analysis is the XMOS
xCORE embedded microcontroller [17]. Beyond offering timing-deterministic
instruction execution, the xCORE is hardware multi-threaded and comes in a
variety of multi-core configurations. The xCORE architecture is simple by design and, thus, ideal to investigate the advanced energy modelling and static
analysis techniques required to achieve energy transparency. The techniques we
developed are readily transferable to other deeply embedded, cache-less, IoTtype processors such as those in the ARM Cortex M series or the Atmel AVR.
The fact that the xCORE offers multi-threading made it a particularly interesting target for the ENTRA project.
We have shown that in the xCORE the number of active threads has an
impact upon energy consumption [6]. As such, the model must take this into
account. Traditional ISA-level models, such as that of [18], can attribute energy
costs simply to instructions, the transitions between instructions, and any additional effects that impact on energy consumption, such as cache hits and misses.
Although we build on this principle, parallelism has to be considered, yielding
4
a more complex model equation for Ep , the energy consumed by a program p:
Ep = Pb Nidl Tclk +
Nt X
X
((Mt Pi O + Pb ) Ni,t Tclk )
(1)
t=1 i∈ISA
In Eq. (1) the energy consumption is split into two parts, capturing idle and
active processor behaviour, respectively. For the former, we consider the base
processor idle power, Pb , that is present even when the device is waiting on
external events, multiplied by the the number of cycles with no active threads,
Nidl , and the clock period, Tclk . For the latter, individual instruction costs are
accounted for based on their costs Pi , as well as an aggregated inter-instruction
overhead, O, and a parallelism scaling factor, Mt , determined by the number of
active threads t. This is calculated for each ISA instruction i, and multiplied
by the number of occurrences in the target program at that particular level of
parallelism, Ni,t , as well as the clock period, Tclk .
Model parameters are separated into two groups. Values for the first group
of constants are obtained by profiling the processor for a fixed clock period
Tclk , yielding the base power Pb , inter-instruction overheads O, per-instruction
costs Pi and parallelism scaling Mt ; all measured in mW . The second group
must be determined through analysis of the target program. These include the
number of idle cycles, Nidl , the number of threads, Nt , in the program, and the
instruction counts, Ni,t , for each instruction i and number of active threads t. If
the analysis can produce these values, Eq. (1) can be used to estimate program
energy. We have demonstrated various simulation- and static analysis-based
methods that follow this principle.
To illustrate instruction profiling, an example heat map representing the device power from interleaving a selected subset of data manipulation instructions
is shown in Fig. 2. The profiling framework executes tightly coupled threads
through the xCORE pipeline, with random, valid operand values to produce
an average power estimate for each instruction. Random input data is shown
to cause higher power dissipation than more constrained data that would be
found in real-world programs [19], e.g. due to data dependencies, thus creating
a modest over-estimate in most cases.
Instruction profiling can only be used to determine the costs of instructions
that can be executed repeatedly and in succession. The cost of other instructions
can be estimated with a generic average or grouped by operand count [6], or
more accurately through a regression tree approach that identifies the most
significant of a set of features, including instruction length, whether memory is
accessed and others, to find the most similar directly profiled instruction [20].
The latter is the most accurate of the three approaches and adds no significant
modelling overhead.
This form of model can be used by various analysis methods. An Instruction
Set Simulation (ISS) can produce an instruction trace from which the instruction counts and thread activity can be determined. A cycle-accurate ISS will
achieve the best model accuracy. Alternatively, per-thread instruction execution statistics can be used instead to extrapolate the model terms. This is faster
5
195
165
150
Power (mW)
180
135
120
105
buf
u6
bu
u6
zext
rus
sext
rus
ldc
ru6
ldapf
u10
ldapb
u10
zext
2r
sext
2r
andnot
2r
mkmsk
rus
mkmsk
2r
clz
l2r
ldapf lu10
ldapb lu10
ldc lru6
neg
2r
not
2r
byterev
l2r
bitrev
l2r
eq 2rus
ldw
3rX
ldw
3r
ld16s
3rX
ld16s
3r
ld8u
3r
stw
3r
st16
3r
ld8u
3rX
st8
3r
ldaw
u6
shl
3r
shr
3r
shl 2rus
eq
3r
shr 2rus
lsu
3r
lss
3r
add 2rus
sub 2rus
ldaw lru6
and
3r
or
3r
sub
3r
add
3r
ashr l2rus
ldawb l2rus
ldawf l2rus
ashr
l3r
ldawb
l3r
crc32
l3r
xor
l3r
ldawf
l3r
lda16f
l3r
lda16b
l3r
mul
l3r
crc8
l4r
maccu
l4r
maccs
l4r
ladd
l5r
lsub
l5r
lmul
l6r
Even threads instruction (name & encoding)
210
lmul
l6r
lsub
l5r
ladd
l5r
maccs
l4r
maccu
l4r
crc8
l4r
mul
l3r
lda16b
l3r
lda16f
l3r
ldawf
l3r
xor
l3r
crc32
l3r
ldawb
l3r
ashr
l3r
ldawf l2rus
ldawb l2rus
ashr l2rus
add
3r
sub
3r
or
3r
and
3r
ldaw lru6
sub 2rus
add 2rus
lss
3r
lsu
3r
shr 2rus
eq
3r
shl 2rus
shr
3r
shl
3r
ldaw
u6
st8
3r
ld8u
3rX
st16
3r
stw
3r
ld8u
3r
ld16s
3r
ld16s
3rX
ldw
3r
ldw
3rX
eq 2rus
bitrev
l2r
byterev
l2r
not
2r
neg
2r
ldc lru6
ldapb lu10
ldapf lu10
clz
l2r
mkmsk
2r
mkmsk
rus
andnot
2r
sext
2r
zext
2r
ldapb
u10
ldapf
u10
ldc
ru6
sext
rus
zext
rus
bu
u6
buf
u6
Odd threads instruction (name & encoding)
Figure 2: Power dissipation of multi-threaded instruction interleaving in the XMOS xCORE
processor. Dashed lines denote a change in operand count, axis label colour indicates 16-bit
(green) and 32-bit (red) instruction encoding.
6
than producing a full trace, but can increase estimation error. It has been shown
to yield an acceptable ±10 % margin in benchmarks [6].
As we will see in Section 3, the program-dependent terms in the model can
be calculated by static analysis, removing the simulation step and allowing rapid
design-space exploration, as well as parameterised, bounded energy estimations.
The values of Ni,t and Nidl can be analysed as functions of the input state. This
allows the derivation of energy functions characterising the energy consumption
of all possible runs of a program, rather than a specific run or set of runs.
In our experiments, tracing was up to two orders of magnitude slower than
statistics-only simulation, the latter typically taking less than one minute. However, this can be mitigated by analysing single functions or blocks, ignoring other
parts of the trace, and terminating as soon as the blocks of interest have been
executed, making trace simulations take less than a minute in most of the cases
we observed. Static analysis can, of course, achieve faster results, by analysing
the same code blocks without simulation.
A key prerequisite for achieving high accuracy in energy modelling at the ISA
level is a predictable, time-deterministic architecture, where the instruction set
gives an accurate view of how the processor will behave. The xCORE’s thread
scheduling and cache-less SRAM memory subsystem, together with the absence
of performance enhancing complexity in the micro-architecture, enabled us to
achieve a very accurate model, which is essential to obtaining accurate energy
predictions. Thus, as for worst case execution time (WCET) analysis, the results
of energy modelling and energy consumption analysis are influenced by processor
architecture, and predictability determines the accuracy achievable as well as
the complexity of the modelling and analysis techniques [21].
2.2. Multi-core energy modelling
For a multi-core system the ISS must accurately simulate the network behaviour in order to capture the timing and link-traversal of data. This allows
accurate estimation of communicating multi-core processes [20]. Static analysis must provide similar characterisation, therefore instruction execution, network behaviour and the flow of communication between processes, must be predictable in order to enable energy transparency. The single-core, multi-threaded
processor model achieves an average error of 2.7 % over a suite of single- and
multi-threaded benchmarks, with a standard deviation of 4.4 %. The multicore model demonstrates an average energy estimation error of −4.9 % with a
standard deviation of 3.9 %. Models with less than 10 % error provide suitable
accuracy for energy estimations.
2.3. Energy modelling at higher levels of software abstraction
Performing static analysis at the ISA level can benefit from good accuracy
due to its closeness to the hardware, but it can suffer from a loss of useful
information such as program structure and types [22]. A good compromise is
found by modelling at the Intermediate Representation (IR) used by compilers,
where program information is preserved. Since the compiler is the natural place
7
for optimisations, modelling and predicting the energy consumption at IR level
could therefore enable energy specific optimisations.
Using a mapping technique, we lifted the energy model in Eq. (1) to the IR
level of the compiler [8], implemented within the LLVM tool chain [23]. ISA level
energy models can thus be propagated up to LLVM IR level, allowing energy
consumption estimation of programs at that level. This enables static analysis
to be performed at a higher level than ISA, thus making energy consumption
information accessible directly to the compiler and optimiser.
The mapping technique determines the energy characteristics of LLVM IR
instructions. It provides on-the-fly energy characterization that takes into consideration the compiler behavior, control flow graph (CFG) structure, types and
other aspects of instructions. Taking into account such information improved
the accuracy of the LLVM IR characterization significantly. The experimental
evaluation demonstrated that the mapping technique allowed for energy consumption transparency at the LLVM IR level, with accuracy keeping within 1%
of ISA-level estimations in most cases [8].
In principle, the same mapping technique may be used to map the energy
consumption of programs to even higher levels, such as the source code. However, a fine grained characterization for each line of source code using this
method is impractical due to the numerous transformations and optimisations
introduced by the compiler and the loss of accuracy resulting from the difficulty
of associating energy consumption costs obtained at lower levels to source code
lines.
An alternative approach to building a source-level energy model was investigated in [24]. The target language here was Java on the Android platform; any
attempt to map a lower-level energy model up to the source code would need
to deal explicitly with the Java virtual machine as well as operating system
layers, a highly impractical strategy. Instead, the basic energy-consuming operations from the source code are identified and the correlations to energy costs
are found by measuring energy consumption in a large number of execution
cases and analyzing the results using techniques based on regression analysis.
The resulting energy model of the basic operations implicitly includes the effect
of all the layers of the software stack down to the hardware. The approach is
inherently approximate; nevertheless such an approach may be the only feasible
one in complex software stacks, when source-level energy models are needed,
for instance to give the source-code programmer an energy profile of the code
under development.
3. Static Analysis of Energy Consumption
Static analysis is the other key component of energy transparency. It infers information about energy consumed by programs without actually running
them. As with energy models, analysis can be performed on program representations at different levels in the software stack, ranging from source code
(in different programming languages) to intermediate compiler representations
(such as LLVM IR [23]) or ISA.
8
Static analysis at a given level consists of reasoning about the execution
traces of a program at that level, in order to infer information (among other
things) about how many times certain basic elements of the program will be
executed. The role of the energy model is to provide information about the
energy consumption of such basic elements; it is used by the analysis to infer information about energy consumption of higher-level entities such as procedures,
functions, loops, blocks and the whole program.
Analysis can also be performed at a given software level using energy models
for a lower level. Such a model needs to be mapped upwards to the higher
level, as described in Section 2.3. The information inferred by static analysis
at a lower level can also be reflected upwards to a higher level using suitable
mapping information.
In the ENTRA project, this approach has been applied to the static analysis
of XC programs running on xCORE architectures. However, our framework is
language- and architecture-neutral. We will return to this in Section 4.
3.1. Analysis/modelling trade-off
Our hypothesis was that the choice of level affects the accuracy of the energy
models and the precision of the analysis in opposite ways: energy models at
lower levels will be more precise than at higher levels while at lower levels
more program structure and data structure information is lost, which often
implies a corresponding loss of accuracy in the analysis. This hypothesis about
the analysis/modelling level trade-off (and potential choices) is illustrated in
Figure 3.
In ENTRA we have explored different points in this space of combinations of
analysis and modelling. Our experimental results [25] confirm that the expected
trade-off exists, but also suggest that performing the static analysis at the LLVM
IR level is a good compromise. LLVM IR is close enough to the source code
level to preserve most of the program information needed by the static analysis,
whilst close enough to the ISA level to allow the propagation of the ISA energy
model up to the LLVM IR level without significant loss of accuracy.
3.2. Information inferred by analysis
The information inferred by the analyzers is guided by its final use: program
optimisation, verification, helping energy-aware software developers to make
design decisions, and so on. For example, they can infer safe approximations,
namely upper and lower bounds, on the energy consumed by the program or
parts of it. These approximations can be functions parametrised by the sizes of
the input data and other hardware features such as clock frequency and voltage.
The analyzers can then infer concrete values of the parameters that yield the
worst-case energy consumption of the program or its parts.
Static energy profiling [26] determines the distribution of energy usage over
the parts of the code. This can be very useful to the developer, showing which
parts of the program are the most energy-critical. Some functions or blocks in
the program are perhaps not particularly expensive in energy in themselves but
9
Internal
representa>on
Transforma>on
(including
asser>ons)
Energy
Model
+
Program
(including
asser>ons)
Layer
1
Analysis
Informa>on
loss
Energy
Modelling
Precision
Loss
Energy
Model
Transforma>on
Source
code
Analysis
HC
IR
Compiler
LLVM
IR
Energy
Model
Layer
2
Energy
Model
Op>miza>ons
Transforma>on
Op>mized
LLVM
IR
HC
IR
Analyser
LLVM
Code
Generator
Transforma>on
ISA
HC
IR
Energy
consump>on
es>ma>ons
Layer
3
Hardware
Figure 3: Analysis/modelling level trade-off and potential choices [25]
are called many times. Such parts are natural targets for optimisation, since
there a small improvement can yield considerable savings.
Note that the safety of bounds depends on energy models giving safe bounds
for each instruction. This is a challenging problem which is discussed in Section 4. Safe bounds are vital for applying energy analysis to verifying or certifying energy consumption.
3.3. A generic resource analysis framework
The resource analysis framework that we have developed is parametric with
respect to resources and programming languages. Regarding resources, the common assertion language allows the definition of different resources and how basic
components of a program affect the use of such resources. More concretely, it allows the encoding of different energy models for specific hardware architectures,
and, in particular, the energy models developed for the xCORE architecture at
the LLVM IR and ISA levels, described in Section 2.
Regarding programming languages, we differentiate between the input language (which in ENTRA can currently be XC source, LLVM IR, or ISA) and
the intermediate semantic program representation, which is what the resource
analysis actually operates on. We use Horn Clauses as the intermediate program representation, referred to as the “HC IR.” A transformation is performed
from each supported input language into HC IR, which is then passed to the
10
resource analysis. We have explored different approaches for this transformation. One approach is to perform a direct transformation into HC IR, and this
has been applied both to ISA and LLVM IR code [22, 25]. Another approach
consists in producing the HC IR by applying partial evaluation techniques to
instrumented interpreters that directly implement the semantics of the language
to be analysed [27]. In both cases, the resulting HC IR programs are analysed
by the CiaoPP tool (see Section 3.5).
Horn Clauses offer several features that make them convenient for analysis [28]. For instance, this representation inherently supports Static Single
Assignment (SSA) and recursive forms. There is a current trend favouring the
use of Horn Clause programs as intermediate representations in analysis and
verification tools [29, 30, 31, 32].
Using the generic HC IR representation, the assertion language and the
CiaoPP analysis tools, we have instantiated the general framework to produce
a series of concrete energy analyzers which have allowed us to study the advantages and limitations of different techniques as well as the trade-offs implied by
different choices of analysis and energy modelling levels.
3.4. A common assertion language
We have defined a common assertion language as a vehicle for propagating
energy-related information throughout the system levels, and for communication among the different analysis, verification, and optimisation tools, and the
actors involved in software development. This assertion language allows expressing energy models for different architectures, writing energy consumption
specifications, describing energy consumption of components that are not available at analysis time, expressing analysis results, and ensuring interoperability.
We refer the reader to [33] and its references for a full description of the Ciao
assertion language that is the basis for the assertions used in the HC IR, and the
(internal) common assertion language of ENTRA [34]. The ENTRA common
assertion language also includes a front end to express energy specifications and
other energy related information in the XC source code.
3.5. Energy analysis using CiaoPP
The input to the CiaoPP parametric static resource usage analyzer [13, 14,
15] is the HC IR, along with assertions in the common assertion language expressing the energy model for LLVM IR blocks and/or individual ISA instructions, and possibly some additional (trusted) information. The analyzer is based
on an approach in which recursive equations (cost relations), representing the
cost of running the program, are extracted from the program and solved, obtaining upper- and lower-bound cost functions (which may be polynomial, exponential or logarithmic) in terms of the program’s inputs [9, 10, 11, 12, 16].
These output cost functions express the energy consumption for each block
in the HC IR, which can be mapped directly back to the language represented
by the HC IR.
The generic resource analysis engine is fully based on abstract interpretation [15], and defines the resource analysis itself as an abstract domain that
11
is integrated into the PLAI abstract interpretation framework [35] of CiaoPP.
This brings in features such as multivariance, efficient fixpoints, and assertionbased verification and user interaction. The setting up and solving of recurrence
relations for inferring closed-form functions representing bounds on the sizes of
output arguments and the resource usage of the predicates in the program are
integrated into the PLAI framework as an abstract operation.
3.6. Direct energy analysis of LLVM IR
As mentioned, LLVM IR offers a good trade-off between analyzability and
accuracy. In addition to using a generic approach based on CiaoPP and a
translation to HC IR, the ENTRA project has experimented with an approach
that uses similar analysis techniques but operates directly on the LLVM IR
representation [36]. The advantage is that this approach can be integrated more
directly in the LLVM toolchain; in principle it is applicable to any languages
targeting LLVM. The energy model used is exactly the same as the one applied
in [25] and described in Section 2.1.
3.7. WCET-inspired energy consumption static analysis
Since the underlying challenges of analysing the timing and energy consumption behaviour of a program appear to be quite similar, in [8], we have applied
well known WCET analysis techniques to retrieve energy consumption estimations. One of the most popular WCET techniques is implicit path enumeration
(IPET) [37], which retrieves the worst case control flow path of programs based
on a cost model that assigns a timing cost to each CFG basic block. We have
replaced the timing cost model by the ISA energy model given in Equation (1).
In the absence of architectural performance enhancing features, such as caches,
this technique can provide safe upper bounds for timing. Through our experimental evaluation we have demonstrated that this is not the case for energy, as
energy consumption, in contrast to time, is data sensitive (see Section 4).
In order to explore the value and limits of applying IPET for energy consumption estimations, we have also extended the analysis to the LLVM IR level,
using the LLVM IR energy characterization given by the mapping technique referred to in Section 2.3. Furthermore, we have extended the energy consumption
analysis to multi-threaded embedded programs from two commonly used concurrency patterns: task farms and pipelines. The experimental evaluation on a set
of mainly industrial programs demonstrates that, although the energy bounds
retrieved cannot be considered safe, they can still provide valuable information for energy aware development, delivering energy transparency to software
developers in the absence of widely accessible software energy monitoring.
3.8. Probabilistic resource analysis
Bounds on energy consumption are useful, but information about the distribution of consumption within those bounds is even more so. For example,
it may be that most execution cases of a program result in consumption close
to the lower bound, while the upper bound is reached only in a few outlying
12
cases, or vice versa. From the distribution, estimates of average energy consumption can be derived. One approach to obtaining this kind of information is
to perform probabilistic static analysis of a program with respect to its energy
consumption. This is a special case of probabilistic output analysis, whose aim
is to derive a probability distribution of possible output values for a program
from a probability distribution of its input. The output in this case is energy
consumption [38, 39].
4. Discussion
Energy-aware software development needs energy transparency; designers
and programmers have to understand energy consumption at an early stage
in the development lifecycle in order to explore the design space taking energy
consumption into consideration. Many decisions taken early in the process, such
as the hardware platform, the degree of parallelism, fundamental algorithms
and data structures, can determine the overall energy efficiency of the final
application. The energy-aware software development lifecycle includes activities
such as:
• Providing an energy specification or energy budget.
• Making initial rough estimates of energy consumption based on high-level
models of the application, allowing exploration of design space with respect
to energy consumption.
• Choosing and configuring a hardware platform which suits the application, for example reducing the energy cost of frequent communications or
memory accesses.
• Developing code with constant reference to the energy consumption of
program parts, allowing energy “bugs” to be identified early.
• Performing energy optimisation of critical code sections using more precise
energy models and taking into account the compiler generated machine
instructions.
• For energy critical applications, providing guarantees in the form of tight
upper and lower bounds on energy consumption.
It is important to note that the development platform is seldom the same as
the final deployment platform, emphasising the importance of energy modelling
of the final target hardware. The alternative to energy transparency is to wait
until the application is run on the final intended platform and then measure its
energy usage. At this stage it is likely to be too late to do much about excessive
energy consumption.
13
4.1. Software energy modelling challenges
Verification of worst-case energy consumption requires the development of
a worst-case energy model. This is difficult, since the energy cost of executing
an instruction depends on the operands. To obtain the worst-case consumption
for an instruction we must therefore measure its execution with the operands
that induce it. The energy models built in ENTRA are based on averages
obtained from measuring the energy consumed when random, valid data is being
processed. We demonstrated that the variation due to data can range from 5 to
25% [8]. In [19] we examine the impact of operand values on instruction level
energy consumption and propose a probabilistic approach to developing worstcase energy models suitable for safe worst-case energy consumption analysis.
Data-sensitive energy modelling is a serious challenge. Determining the maximum amount of circuit switching between instruction data tends to take an
exponential amount of time to evaluate, making it difficult to determine or
guarantee the worst case data for an instruction sequence [40]. Even if a model
is built where the energy consumed by an instruction is given as a function of
its operands, there would still be further challenges in encoding these functions
in a suitable way to be exploited effectively by static analysis algorithms.
In [41] we have explored a technique that models both, upper and lower
bounds on the energy of “branchless” blocks of instructions, in order to take
into account the inter-instruction switching costs within a block. It uses an
Evolutionary Algorithm for a faster exploration of the search space, which is
also reduced by the fact that the algorithm does not have to deal with the
control-flow of the program. Then, such block-level model is fed into a static
analysis, which takes into account the program control-flow, and infers energy
information for the whole program and its procedures.
4.2. Software energy analysis challenges
Static analysis always involves a trade-off of precision against complexity of
the analysis. Obtaining tight bounds of energy usage depends on several factors, including accurate propagation of data size measures and extraction and
solution of the relations expressing energy consumption in terms of data sizes.
Both of these problems are solvable for a large class of useful programs, but
if program structure departs from standard patterns, precision may be rapidly
lost. For instance, our realisation of the general framework described in Section 3.3 using the CiaoPP resource analyser described in Section 3.5 (which uses
the HC IR) can deal with recursive programs, including multiple and mutual
recursion (e.g., divide-and-conquer programs), iterative programs with nested
loops, numeric/arithmetic programs, or programs operation on complex data
structures such as nested lists and arrays. The analysis produces parametrised
energy bounds (which depend on input data sizes) expressed by a large class
of functions (e.g., polynomial, factorial, exponential, logarithmic, and summations), in contrast to other approaches that are limited to polynomial functions,
or to non-parametric (i.e., absolute) bounds.
The experiments reported in [25] with this realisation of the analysis framework perform the analysis at the ISA and LLVM IR levels, and compare the
14
energy values obtained by evaluating the inferred energy functions for different
input data sizes, with actual hardware measurements (on the xCORE platform). The results show that our LLVM IR level analysis is reasonably accurate
(less than 6.4% average deviation vs. hardware measurements) and more powerful than the analysis at the ISA level, in the sense that it can deal with a
larger class of programs (e.g., programs involving structured types). The average deviation for the smaller set of benchmarks for which the ISA-level analysis
produced non-trivial results was 3.9%. Although we have tested our prototype
tools with relatively small programs, they exhibit features that are also present
in bigger real programs, and could be analysed at a bigger scale too, since we
have designed our analysis tools to enable scalability. Thus, we interpret our experimental results as very promising, and encourage us to continue our research
following an incremental approach. Making our prototype tools more robust,
powerful and scalable, as well as evaluating them with bigger real programs, is
an implementation (and research) challenge in itself.
As already said, our approach to developing energy analysers is architecture
neutral, a claim that is supported by the experimental results in [36], performed
with the direct energy analysis of LLVM IR described in Section 3.6, for both the
ARM Cortex-M and XMOS xCORE platforms. The benchmarks also contain
nested loops (some of them with complex control flow predicates) and perform
bitwise operations, as well as operations on arrays and matrices. Overall, the
final deviation vs. hardware measurements is typically less than 10% and 20%
on the XMOS and ARM platforms respectively, showing that the general trend
of the static analysis results can be relied upon to give an estimate of the
energy consumption. Instantiating our general approach to more platforms
(which include the hardware and operating system) and assessing it in different
application domains is another challenge that we intend to address in future
work.
Static analysis of multi-threaded code is difficult since precision is easily lost
due to thread interleaving. Accurate analysis of the timing and synchronisation
behaviour of threads is a pre-requisite for energy analysis using the model given
in Eq. (1), in which the energy of an instruction depends on how many threads
are active simultaneously.
4.3. Extending results to less predictable architectures
The ENTRA approach is generic and architecture-neutral; that is, it consists of a framework parametrised by an energy model and generic static analysis
tools. Front-ends that translate code into a common, analysable form such as
HC IR or LLVM IR can in principle be constructed for any programming language. However, much experimentation and investigation remains to be done
to apply the approach effectively to architectures that contain sources of unpredictability such as caches, complex pipelines and interrupts that are not present
in the xCORE architecture. A likely path of research is to follow the approach in
WCET analysis of such unpredictable architectures, employing supporting analyses to permit more accurate energy analyses. An example is approximation of
cache contents at specific program points, leading to more accurate models when
15
it can be guaranteed that a memory access will definitely hit or definitely miss
the cache.
4.4. Energy optimisations enabled by energy transparency.
Different types of optimisations at different levels of the software stack can
be performed by taking advantage of the information provided by the multilevel and multi-language ENTRA tools. Both static and dynamic energy optimisations are enabled by energy transparency, which have been investigated
in ENTRA. For dynamic optimisations, a framework for energy-aware stochastic scheduling based on evolutionary algorithms (EAs) has been developed, for
the cases where tasks are independent [42] and dependent [43]. In the latter
case, dependence has been modelled by using copula theory [44], in particular
Archimedean copulas [45]. EAs have also been used to improve energy-aware
allocation and scheduling for DVFS-enabled multicore environments. For example, the algorithms described in [46, 47] are able to deal with task migration
and preemption; and the ones in [48] allow decreasing program accuracy (by
performing loop perforation) in order to save energy.
Other optimisations include the use of energy analysis to choose software
parameters in order to transform programs to ensure that an energy target
is met while minimizing the loss in quality of service. Energy performance has
also been improved by optimisation techniques for task placement on an xCORE
network after identifying communication patterns among tasks. Using Swallow,
an experimental open platform of many xCOREs [49] as a source of model data,
it has been demonstrated that incorporating network-level energy consumption
and timing into the energy modelling process can aid in identifying the impact
of sub-optimal task placement in communicating multi-core applications [50].
Some energy optimisations have already been incorporated into the recently
released XC compiler by XMOS Ltd. These optimisations in the object code
are obtained by aggressively applying global dataflow analyses inspired by ENTRA project research. Results on case studies, showing a power reduction of
approximately 25% by using the global optimiser, can be found in an ENTRA
project report [51].
An experiment on energy optimisation of Android code was carried out using
the source code energy modelling mentioned in Section 2.3. The energy consumption of battery-driven mobile devices such as smartphones is of increasing
concern to developers of software for such devices. The study concerned the
optimisation of code for interactive games. The energy model combined with
execution profiling, enabled the developer to discover that 10 code blocks (out
of several hundred blocks in the considered code) consumed over 50% of the
overall energy. Aggressive source code optimisation and refactoring of these
blocks enabled energy savings of 6% to 50% in various use-case scenarios [52].
5. Conclusions
The goal of the ENTRA project was to provide techniques and tools supporting energy transparency at the software level. The results obtained include
16
energy models for the xCORE processor at different levels of abstraction, from
ISA level to LLVM IR, as well as preliminary energy models for less predictable
architectures. A model incorporating multi-core execution on the xCORE has
also been developed. These models are incorporated in static analyses at the
corresponding level of code. Experiments have compared predictions of energy
consumption for single-threaded programs with actual measured consumption,
and encouraging results with only a few percentage of error were obtained. In
addition, the trade-offs between accuracy and analysability at different levels
were explored, leading to a preliminary conclusion that analysis at LLVM level
provides a good compromise. The project identified challenging problems for
future research, extending the analyses to multi-threaded code, and building
data-sensitive energy models.
Acknowledgements
This research has received funding from the European Union 7th Framework
Programme under grant agreement no. 318337, ENTRA - Whole Systems Energy Transparency, grant agreement no. 611004, the ICT-Energy Coordination
and Support Action, Spanish MINECO TIN’12-39391 StrongSoft project, and
the Madrid M141047003 N-GREENS program.
References
[1] C. Edwards., Lack of software support marks the low power scorecard at
DAC, in: Electronics Weekly., 2011.
[2] K. Roy, M. C. Johnson, Software Design for Low Power, in: W. Nebel,
J. P. Mermet (Eds.), Low Power Design in Deep Submicron Electronics,
Vol. 337, Kluwer Academic, 1997, pp. 433–460.
URL http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.
44.4398
[3] D. Brooks, V. Tiwari, M. Martonosi, Wattch: a framework for
architectural-level power analysis and optimizations, in: Proceedings of
the 27th annual international symposium on Computer architecture, ISCA
’00, ACM, New York, NY, USA, 2000, pp. 83–94. doi:http://doi.acm.
org/10.1145/339647.339657.
URL http://doi.acm.org/10.1145/339647.339657
[4] V. Tiwari, S. Malik, A. Wolfe, M. Tien-Chien Lee, Instruction level power
analysis and optimization of software, The Journal of VLSI Signal Processing 13 (1996) 223–238, 10.1007/BF01130407.
URL http://dx.doi.org/10.1007/BF01130407
[5] S. Steinke, M. Knauer, L. Wehmeyer, P. Marwedel, An Accurate and Fine
Grain Instruction-level Energy Model Supporting Software Optimizations,
in: Proceedings of PATMOS, 2001.
17
[6] S. Kerrison, K. Eder, Energy modeling of software for a hardware multithreaded embedded microprocessor, ACM Trans. Embedded Comput. Syst.
14 (3) (2015) 56. doi:10.1145/2700104.
URL http://doi.acm.org/10.1145/2700104
[7] C. Brandolese, S. Corbetta, W. Fornaciari, Software energy estimation
based on statistical characterization of intermediate compilation code, in:
Low Power Electronics and Design (ISLPED) 2011 International Symposium on, 2011, pp. 333–338. doi:10.1109/ISLPED.2011.5993659.
[8] K. Georgiou, S. Kerrison, K. Eder, On the Value and Limits of Multilevel Energy Consumption Static Analysis for Deeply Embedded Single
and Multi-threaded Programs, Tech. rep. (Oct. 2015). arXiv:1510.07095.
URL http://arxiv.org/abs/1510.07095
[9] B. Wegbreit, Mechanical program analysis, Commun. ACM 18 (9) (1975)
528–539.
[10] M. Rosendahl, Automatic Complexity Analysis, in: 4th ACM Conference on Functional Programming Languages and Computer Architecture
(FPCA’89), ACM Press, 1989, pp. 144–156.
[11] S. K. Debray, N.-W. Lin, M. Hermenegildo, Task Granularity Analysis
in Logic Programs, in: Proc. of the 1990 ACM Conf. on Programming
Language Design and Implementation, ACM Press, 1990, pp. 174–188.
[12] S. K. Debray, P. López-Garcı́a, M. Hermenegildo, N.-W. Lin, Lower Bound
Cost Estimation for Logic Programs, in: 1997 International Logic Programming Symposium, MIT Press, Cambridge, MA, 1997, pp. 291–305.
[13] J. Navas, E. Mera, P. López-Garcı́a, M. Hermenegildo, User-Definable Resource Bounds Analysis for Logic Programs, in: International Conference
on Logic Programming (ICLP’07), Lecture Notes in Computer Science,
Springer, 2007, pp. 348–363.
[14] J. Navas, M. Méndez-Lojo, M. Hermenegildo, Safe Upper-bounds Inference
of Energy Consumption for Java Bytecode Applications, in: The Sixth
NASA Langley Formal Methods Workshop (LFM 08), 2008, pp. 29–32,
extended Abstract.
[15] A. Serrano, P. Lopez-Garcia, M. Hermenegildo, Resource Usage Analysis of Logic Programs via Abstract Interpretation Using Sized Types,
Theory and Practice of Logic Programming, 30th Int’l. Conference on
Logic Programming (ICLP’14) Special Issue 14 (4-5) (2014) 739–754.
doi:10.1017/S147106841400057X.
URL http://arxiv.org/abs/1405.4256
[16] E. Albert, P. Arenas, S. Genaim, G. Puebla, Closed-Form Upper Bounds
in Static Cost Analysis, Journal of Automated Reasoning 46 (2) (2011)
161–203.
18
[17] D. May,
The XMOS XS1 architecture. available
http://www.xmos.com/published/xmos-xs1-architecture (2013).
online:
[18] V. Tiwari, S. Malik, A. Wolfe, Compilation Techniques for Low Energy: An
Overview, in: Low Power Electronics, 1994. Digest of Technical Papers.,
IEEE Symposium, IEEE, 1994, pp. 38–39.
[19] J. Pallister, S. Kerrison, J. Morse, K. Eder, Data dependent energy modeling for worst case energy consumption analysis, Tech. rep. (May 2015).
arXiv:1505.03374.
URL http://arxiv.org/abs/1505.03374
[20] S. Kerrison, K. Eder, Modeling and visualizing networked multi-core embedded software energy consumption, Tech. rep. (Sep. 2015). arXiv:
1509.02830.
URL http://arxiv.org/abs/1509.02830
[21] R. Heckmann, M. Langenbach, S. Thesing, R. Wilhelm, The influence of
processor architecture on the design and the results of WCET tools, Proceedings of the IEEE 91 (7) (2003) 1038–1054. doi:10.1109/JPROC.2003.
814618.
[22] U. Liqat, S. Kerrison, A. Serrano, K. Georgiou, P. Lopez-Garcia, N. Grech,
M. Hermenegildo, K. Eder, Energy Consumption Analysis of Programs
based on XMOS ISA-level Models, in: G. Gupta, R. Pea (Eds.), LogicBased Program Synthesis and Transformation, 23rd International Symposium, LOPSTR 2013, Revised Selected Papers, Vol. 8901 of Lecture
Notes in Computer Science, Springer, 2014, pp. 72–90. doi:10.1007/
978-3-319-14125-1_5.
URL http://dx.doi.org/10.1007/978-3-319-14125-1_5
[23] C. Lattner, V. Adve, LLVM: A compilation framework for lifelong program
analysis and transformation, in: Proc. of the 2004 International Symposium
on Code Generation and Optimization (CGO), IEEE Computer Society,
2004, pp. 75–88.
URL http://dl.acm.org/citation.cfm?id=977395.977673
[24] X. Li, J. P. Gallagher, A top-to-bottom view: Energy analysis for mobile
application source code, Tech. rep., submitted to a conference (October
2015). arXiv:1510.04165.
URL http://arxiv.org/pdf/1510.04165v1.pdf
[25] U. Liqat, K. Georgiou, S. Kerrison, P. Lopez-Garcia, M. V. Hermenegildo,
J. P. Gallagher, K. Eder, Inferring Parametric Energy Consumption Functions at Different Software Levels: ISA vs. LLVM IR, in: M. V. Eekelen, U. D. Lago (Eds.), Foundational and Practical Aspects of Resource
Analysis. Fourth International Workshop FOPARA 2015, Revised Selected
Papers, Lecture Notes in Computer Science, Springer, 2016, in press.
19
arXiv:1511.01413.
URL http://arxiv.org/abs/1511.01413
[26] R. Haemmerlé, P. Lopez-Garcia, U. Liqat, M. Klemen, J. P. Gallagher, M. V. Hermenegildo, A Transformational Approach to Parametric Accumulated-cost Static Profiling, in: O. Kiselyov, A. King (Eds.),
13th International Symposium on Functional and Logic Programming
(FLOPS 2016), Vol. 9613 of LNCS, Springer, 2016, pp. 163–180. doi:
10.1007/978-3-319-29604-3_11.
URL http://dx.doi.org/10.1007/978-3-319-29604-3_11
[27] K. S. Henriksen, J. P. Gallagher, Abstract interpretation of PIC programs
through logic programming, in: Sixth IEEE International Workshop on
Source Code Analysis and Manipulation (SCAM 2006), IEEE Computer
Society, 2006, pp. 184–196.
[28] M. Méndez-Lojo, J. Navas, M. Hermenegildo, A Flexible (C)LP-Based Approach to the Analysis of Object-Oriented Programs, in: 17th International
Symposium on Logic-based Program Synthesis and Transformation (LOPSTR 2007), no. 4915 in LNCS, Springer-Verlag, 2007, pp. 154–168.
[29] S. Grebenshchikov, A. Gupta, N. P. Lopes, C. Popeea, A. Rybalchenko,
HSF(C): A Software Verifier Based on Horn Clauses - (Competition Contribution), in: C. Flanagan, B. König (Eds.), TACAS, Vol. 7214 of LNCS,
Springer, 2012, pp. 549–551.
[30] H. Hojjat, F. Konecný, F. Garnier, R. Iosif, V. Kuncak, P. Rümmer, A
Verification Toolkit for Numerical Transition Systems - Tool Paper, in:
D. Giannakopoulou, D. Méry (Eds.), FM 2012: Formal Methods - 18th
International Symposium, Paris, France, August 27-31, 2012. Proceedings,
Vol. 7436 of Lecture Notes in Computer Science, Springer, 2012, pp. 247–
251. doi:10.1007/978-3-642-32759-9_21.
URL http://dx.doi.org/10.1007/978-3-642-32759-9_21
[31] L. M. de Moura, N. Bjørner, Z3: An Efficient SMT Solver, in: C. R.
Ramakrishnan, J. Rehof (Eds.), Tools and Algorithms for the Construction
and Analysis of Systems, 14th International Conference, TACAS 2008, Vol.
4963 of Lecture Notes in Computer Science, Springer, 2008, pp. 337–340.
[32] N. Bjørner, F. Fioravanti, A. Rybalchenko, V. Senni (Eds.), Proceedings of
First Workshop on Horn Clauses for Verification and Synthesis, Vol. 169
of EPTCS, 2014. doi:10.4204/EPTCS.169.
URL http://dx.doi.org/10.4204/EPTCS.169
[33] M. V. Hermenegildo, F. Bueno, M. Carro, P. López, E. Mera, J. Morales,
G. Puebla, An Overview of Ciao and its Design Philosophy, Theory and
Practice of Logic Programming 12 (1–2) (2012) 219–252. doi:doi:10.
1017/S1471068411000457.
URL http://arxiv.org/abs/1102.5497
20
[34] K. Eder, N. Grech (Eds.), Common Assertion Language, ENTRA Project:
Whole-Systems Energy Transparency (FET project 318337), 2013, deliverable 2.1, http://entraproject.eu.
[35] K. Muthukumar, M. Hermenegildo, Compile-time Derivation of Variable
Dependency Using Abstract Interpretation, Journal of Logic Programming
13 (2/3) (1992) 315–347.
[36] N. Grech, K. Georgiou, J. Pallister, S. Kerrison, J. Morse, K. Eder, Static
analysis of energy consumption for LLVM IR programs, in: Proceedings of
the 18th International Workshop on Software and Compilers for Embedded
Systems, SCOPES 2015, ACM, New York, NY, USA, 2015, pp. 12–21.
doi:10.1145/2764967.2764974.
URL http://dl.acm.org/citation.cfm?doid=2764967.2764974
[37] Y.-S. Li, S. Malik, Performance analysis of embedded software using implicit path enumeration, Computer-Aided Design of Integrated Circuits
and Systems, IEEE Transactions on 16 (12) (1997) 1477–1487. doi:
10.1109/43.664229.
[38] M. Rosendahl, M. H. Kirkeby, Probabilistic output analysis by program
manipulation, in: N. Bertrand, M. Tribastone (Eds.), Proceedings Thirteenth Workshop on Quantitative Aspects of Programming Languages and
Systems, QAPL 2015, London, UK, 11th-12th April 2015., Vol. 194 of
EPTCS, 2015, pp. 110–124, qAPL, London. doi:10.4204/EPTCS.194.8.
URL http://dx.doi.org/10.4204/EPTCS.194.8
[39] M. Kirkeby, M. Rosendahl, Probabilistic Resource Analysis by Program
Transformation, in: Proc. of the Foundational and Practical Aspects of
Resource Analysis, LNCS, Springer, 2015, in press.
URL
http://entraproject.eu/wp-content/uploads/2016/02/
prob-res-fopara.pdf
[40] J. Morse, S. Kerrison, K. Eder, On the infeasibility of analysing worst-case
dynamic energy, Tech. rep. (Mar. 2016). arXiv:1603.02580.
URL http://arxiv.org/abs/1603.02580
[41] U. Liqat, Z. Banković, P. Lopez-Garcia, M. V. Hermenegildo, Inferring Energy Bounds Statically by Evolutionary Analysis of Basic Blocks, in: Workshop on High Performance Energy Efficient Embedded Systems (HIP3ES
2016), 2016, arXiv:1601.02800.
URL http://arxiv.org/abs/1601.02800
[42] Z. Banković, P. Lopez-Garcia, Stochastic vs. Deterministic Evolutionary
Algorithm-based Allocation and Scheduling for XMOS Chips, Neurocomputing 150 (2015) 82–89. doi:http://dx.doi.org/10.1016/j.neucom.
2014.06.077.
URL
http://www.sciencedirect.com/science/article/pii/
S0925231214012211
21
[43] Z. Banković, P. López-Garcı́a, Improved Energy-aware Stochastic Scheduling based on Evolutionary Algorithms via Copula-based Modeling of Task
Dependences, in: A. Herrero, J. Sedano, B. Baruque, H. Quintián, E. Corchado (Eds.), International Conference on Soft Computing Models in Industrial and Environmental Applications (SOCO 2015), Vol. 368 of Advances
in Intelligent Systems and Computing, Springer International Publishing,
2015, pp. 153–163. doi:10.1007/978-3-319-19719-7_14.
URL http://dx.doi.org/10.1007/978-3-319-19719-7_14
[44] R. B. Nelsen, Properties and applications of copulas: A brief survey, in:
First Brazilian Conference on Statistical Modelling in Insurance and Finance, 2003, pp. 10–28.
[45] A. J. McNeil, et al., Multivariate archimedean copulas, d-monotone functions and l1 -norm symmetric distributions (2009).
[46] Z. Banković, P. López-Garcı́a, Energy Efficient Allocation and Scheduling
for DVFS-enabled Multicore Environments using a Multiobjective Evolutionary Algorithm, in: Genetic and Evolutionary Computation Conference,
GECCO 2015, Companion Material Proceedings, ACM, 2015, pp. 1353–
1354. doi:10.1145/2739482.2764645.
URL http://doi.acm.org/10.1145/2739482.2764645
[47] Z. Banković, U. Liqat, P. López-Garcı́a, A Practical Approach for Energy
Efficient Scheduling in Multicore Environments by combining Evolutionary and YDS Algorithms with Faster Energy Estimation, in: The 11th
International Conference on Artificial Intelligence Applications and Innovations (AIAI’15), Vol. 458 of IFIP Advances in Information and Communication Technology, Springer International Publishing, 2015, pp. 478–493.
doi:10.1007/978-3-319-23868-5_35.
URL http://dx.doi.org/10.1007/978-3-319-23868-5_35
[48] Z. Banković, U. Liqat, P. López-Garcı́a, Trading-off Accuracy vs. Energy in Multicore Processors via Evolutionary Algorithms Combining
Loop Perforation and Static Analysis-based Scheduling, in: E. Onieva,
I. Santos, E. Osaba, H. Quintián, E. Corchado (Eds.), Hybrid Artificial
Intelligent Systems (HAIS 2015), Vol. 9121 of Lecture Notes in Computer Science, Springer International Publishing, 2015, pp. 690–701.
doi:10.1007/978-3-319-19644-2_57.
URL
http://link.springer.com/chapter/10.1007/
978-3-319-19644-2_57
[49] S. J. Hollis, S. Kerrison, Swallow: Building an Energy-Transparent
Many-Core Embedded Real-Time System, in: 2016 Design, Automation &
Test in Europe, IEEE, 2016.
URL
http://entraproject.eu/wp-content/uploads/2016/02/
KerrisonSwallow15.pdf
22
[50] S. Kerrison, K. Eder, Modeling and visualizing networked multi-core embedded software energy consumption, CoRR abs/1509.02830.
URL http://arxiv.org/abs/1509.02830
[51] H. Muller (Ed.), Evaluation Results, ENTRA Project: Whole-Systems
Energy Transparency (FET project 318337), 2016, deliverable 6.2,
http://entraproject.eu.
URL http://entraproject.eu/wp-content/uploads/2016/03/deliv_
6.2.pdf
[52] X. Li, J. P. Gallagher, An energy-aware programming approach for mobile
application development guided by a fine-grained energy model, Tech.
rep., Roskilde University, submitted for publication (February 2016).
URL
http://entraproject.eu/wp-content/uploads/2016/02/
LiGallagher-TSE-2016.pdf
23
| 6cs.PL
|
End-to-end 3D shape inverse rendering of different classes
of objects from a single input image
Shima Kamyab1 and S. Zohreh Azimifar1
arXiv:1711.05858v1 [cs.CV] 11 Nov 2017
1
Computer Science and Engineering and Information Technology, Shiraz
university, Shiraz, Iran
November 17, 2017
Abstract
In this paper a semi-supervised deep framework is proposed for the problem of 3D shape
inverse rendering from a single 2D input image. The main structure of proposed framework
consists of unsupervised pre-trained components which significantly reduce the need to labeled
data for training the whole framework. using labeled data has the advantage of achieving to
accurate results without the need to predefined assumptions about image formation process.
Three main components are used in the proposed network: an encoder which maps 2D input
image to a representation space, a 3D decoder which decodes a representation to a 3D structure
and a mapping component in order to map 2D to 3D representation. The only part that needs
label for training is the mapping part with not too many parameters. The other components
in the network can be pre-trained unsupervised using only 2D images or 3D data in each case.
The way of reconstructing 3D shapes in the decoder component, inspired by the model based
methods for 3D reconstruction, maps a low dimensional representation to 3D shape space with
the advantage of extracting the basis vectors of shape space from training data itself and is
not restricted to a small set of examples as used in predefined models. Therefore, the proposed
framework deals directly with coordinate values of the point cloud representation which leads
to achieve dense 3D shapes in the output. The experimental results on several benchmark
datasets of objects and human faces and comparing with recent similar methods shows the
power of proposed network in recovering more details from single 2D images.
1
Introduction
Inverse rendering stands for a specific class of machine learning techniques aim at recovering the
properties of a 3D scene like camera’s extrinsic parameters, scene lighting and the shape of the
scene from some existing measurements (e.g. 2D images)[1, 21]. It has attracted high interest in
the recent academic research due to wide applications in computer vision [15, 1, 33, 35, 38, 10, 26,
31, 32, 25, 27, 30, 13, 21, 7].
Inverse rendering is not a well posed problem because of existing infinite number of 3D structure which can result the same 2D image [10, 25]. Facing with such problems needs additional
assumptions and prior knowledge about the class of 3D structures to be recovered from 2D images.
For instance a statistical model can be used to determine which solution is likely and which is not
for a specific class of objects [25].
The aim of Inverse rendering problem is actually inverting the image formation process called
rendering. in Computer graphics, Rendering is expressed as an equation formulated as in 1 [19].
F = K(h + Gf )
(1)
where, F is the reflected radiance (light) from the surface (e.g. intensity of the pixels in the
image), K represents the way the object surface reflects the lights in various spatial and angular
positions, h describes the lighting of the scene, G is some measurement showing how light is
transferred between other existing surfaces in the scene and f is the light which is reflected by
other surfaces in scene affecting F. Knowing the quantities K, h, G, f means that the properties
of the scene are known and F can be computed using these quantities and this is a direct problem
1
to solve. The inverse rendering problem arises when F or some information about it are available
and the aim is to compute some values in the right-hand side of 1. In this paper, our focus is on
recovering shape geometry of an object from a single 2D input image having different poses which
is implicitly embedded in variable K as spatial properties of the surface in light reflection.
Inverse rendering problem as a machine learning task, can be faced through 3 main steps to be
solved:
• Selection of a model (i.e. assuming a specific class of hypotheses)
• Defining a score criterion
• Defining a Search strategy
Selection of a model: Selection of a model is one of the main necessities for solving machine
learning problems due to “no free lunch theorem” [34]. In the case of recovering 3D structures
from 2D images, this step is more important because in such problems we deal with an ill-posed
problem based on Hadmard definition for well-posed problems: there might be more than one
solutions that give the same 2D image and the solution does not depends continuously on the
data, which means that small errors in measurements may cause large errors in the solutions [24].
For solving such kinds of problems, some assumptions are usually made about the properties of
the solutions to restrict the solution space to feasible regions. Selection of a suitable model for
defining the target scene serves as some prior knowledge about the solution space, by which the
search can be performed more reliable. Like using any other assumptions, this method may have
some costs like losing some promising regions that are not spanned by the selected model. For
inverse rendering problems, there are various types of models and prior knowledge for the target
scene in the literature that make the existing methods different from each other: some approaches
use additional information, like multiple images or video sequence or landmarks, about the scene
to find a unique solution [38, 23, 9, 18], some of them work with a certain statistical model to
determine the feasibility of the solutions [12] and some other perform some regressing techniques
that relates the output measurements to 3D structures [10, 26, 31, 32, 25, 27, 13, 17, 30].
In this paper our focus is on the pose invariant recovering 3D structure of an object from a
single 2D image with the help of a model describing the properties of that object. The main
structure of our proposed framework is based on finding a relation between measurements, i.e. a
single 2D image for each scene, and 3D structure of the scene constrained with the used model.
Defining a score criterion:After selection of a suitable model describing the solution space,
the optimum solution should be searched in that space. For this aim, first a criterion should be
defined which guides the search toward finding the optimum solution satisfying the constraints of
the problem at hand. For instance, in the case of inverse rendering of human face from 2D image(s)
for face recognition task, first, the main objective is to recover the distinctiveness characteristics
of a face in an image and so a good solution should not be too generic so or too over determined
and there should be some tradeoff between these two properties in the objective criterion [32]. In
this paper, since the objective is to minimize the reconstruction error, we used Root Mean Square
Error (RMSE) as loss function for training and evaluation.
Defining a search strategy:There exist different types of search mechanisms to search for a
solution based on the defined score criterion. In the case of convex criteria, a closed form solution
exists and the search can be done in linear time. However, the nature of inverse rendering problem is
highly nonlinear and the inference and finding the relation between 2D image and corresponding 3D
structure is intractable. If some assumptions can be incorporated to the problem, like deterministic
models formulating behavior of all variables in the rendering equation to be defined accurately, so
the solution can be computed in closed form [1]. However we cannot have such strong assumptions
in all inverse rendering problems. In such cases we can use learning based methods by training some
models and then use them for test samples. The most popular learning framework in recent years
are deep networks which show their power in solving nonlinear problems. Deep networks like other
learning based structures can be trained in 3 ways: Supervised, unsupervised and semi-supervised.
Supervised methods use a set of labeled data to guide the training. Their result will be accurate
because of availability of ground truth in the training procedure but suffer from the lack of enough
labeled training data in some domains. In the case of 3D Inverse rendering applications, providing
enough realistic training data for a deep network may be impractical. One Solution for this problem
is to use a generative model for generating synthetic training data to trained the network [26, 27].
2
Figure 1: Main configuration of proposed framework
As another solution, There can be found some attempts in the recent literature to Unsupervised
training the network for 3D reconstruction[25, 30] in which a 3D structure is generated in a part
of the network and some rendering mechanism is used to transform the generated 3D structure to
something in the form of input measurement to the network and then try to reduce the difference
between input and rendered output. So the network does not need to labeled data for training. In
the existing unsupervised methods in the literature, they used an analytical form for rendering the
3D structure with some assumptions about image formation process. This characteristic although
is a significant improvement in reducing the need for training data but may affects the quality of
solutions due to the existing assumptions about rendering mechanism (such as camera properties,
scene illumination, reflectance properties and so on). In the case of 3D Inverse rendering problem,
there are well known 3D models that can help us to generate enough data for training the deep
networks.
In this paper, we propose a mechanism in which some parts of network can be trained using
unlabeled data and then the whole network will be fine-tuned using labeled training data for
aggregation of the components. We believe this properties reduces the nedd of deep network
for large set of training data. Therefore it is a semi-supervised framework for inverse rendering
problem.
The paper is organized as follows: we first briefly describe our framework in section 2. In
section 3 we review and analyze some recent related works. The proposed framework and network
structure will be define in more details in section 4. Some experimental results to verify the
performance of our proposed structure on several benchmarks will be demonstrated in section 5
and finally some discussions and conclusions will be found in sections 6 and 7.
2
Overview
Our proposed framework is composed of 3 main components based on the fact that in inverse
rendering problem we look for a way of transforming the 2D input image to its corresponding 3D
structure. We do this by mapping the 2D input image and corresponding 3D structure to suitable
representations and then finding a transformation between obtained representations. We will show
experimentally that nonlinear mapping of suitable representations results more accurate solutions.
finding suitable representations can be done using autoencoders as unsupervised tools for finding
representations or by pre-trained convolutional networks for finding suitable representations. We
did this either by different types of autoencoder or by some existing well known networks from
literature.
In the case of using autoencoders for finding representations, inspired by 3D Morphable Models
(3DMMs) [3] the network extracts the basis vectors from training data of each size and uses these
basis functions to reconstruct more accurate 3D shape of an object rather that using a pre-defined
3DMM. Figure 1 shows the overall structure of the proposed framework. More details about the
proposed idea can be found in section 4
3
3
Related works
In this section we review and analyze some of the most recent attempts for 3D inverse rendering
from single 2D images using deep frameworks as related works to ours concentrating on their search
strategies.
In [26], Richardson et al. proposed a deep iterative framework for 3D inverse rendering of
human faces. In each iteration of this framework, the output of the network will be fed into
the input as additional information to fine-tune the previous result. The training process in this
method is supervised by using a 3DMM for generating synthetic labeled training data. The final
output will be more fine-tuned using some form of Shape From Shading (SFS) method to achieve
more details.The authors improve their work in [27] where, the SFS component of their network
is implemented as another deep network. The output of this framework is the coefficients of a
predefined 3DMM computed using a few hundred 3D samples and so this framework does not
directly deal with 3D shape structures.
In [32], a supervised framework for 3D reconstruction of human faces from a single 2D image
is proposed where they also used a 3DMM for generating training data. For this aim they fitted
a 3DMM to 2D images via an optimization method and used that model as ground truth for the
training data. This strategy of finding ground truth may cause the found solutions to be dependent
on the power of the optimization method so that some powers of new work may be ignored. A
loss function is also used for the training the network for human faces which can be tuned to make
trade-off between generic face generation and overdetermined one, which is suitable for the purpose
of face recognition.
In [25] an unsupervised framework for 3D reconstruction of objects from single or multiple
input images or volumes is proposed. The authors suggested an inference network for encoding
the input data to a low dimensional representation which will be fed into a generative network for
generating 3D objects using the encoded data. The obtained 3D object will be rendered as a 2D
image using a fixed differentiable rendering method. The objective of training of this network is to
minimize the difference between rendered and input images. Using a fixed rendering component
may restrict the accuracy of generated 3D shapes if the input image rendering mechanism differs
from network rendering mechanism.The authors also suggested using a convolutional network as
rendering component. Using another deep network as rendering component may overcome the
mentioned limitation.
In [30] an autoencoder network is proposed for 3D reconstruction of human faces using an
unsupervised training mechanism. In this network the encoder part of the autoencoder is a convolutional network like well-known alexnet [16] or VGG-face [20] networks and the decoder part is
an analytical differentiable formula which uses the encoded data and some assumptions about the
image formation process to form to a 3D face and then render it to a 2D image. In this paper the
objective is also to minimize the difference between input image and rendered 3D shape using fix
rendering method.
In [8] a Generative Adversarial Network (GAN) is trained for generating 3D shapes of different
objects. In this method, in an unsupervised training approach, a random vector from a distribution
will be fed to the network and a 3D shape of an object will be generated and then will be rendered
in the form of a 2D image as input to a discriminator. In this way, the discriminator does not need
to 3D shape to control the process of learning 3D shapes. The rendering method here is fixed and
this network can be used to generate training data for inverse rendering frameworks of 3D objects.
In [12] a CNN is designed to recover the 3D geometry and texture of human faces from a single
2D image. This framework directly works with 3D coordinates in the form of proposed binary
volumes and maps 2D images directly to the coordinates of binary volume. Using binary volume
as 3D coordinates for representing human face may result in blur shapes.
In this paper, our aim is to design a deep framework for 3D shape inverse rendering of different
types of 3D shape representations such as point clouds and volume representations from a single
2D input image. The training process utilizes from both labeled and unlabeled realistic data for
improving the quality of resulted 3D shapes. The following section demonstrates the proposed
approach in more details.
4
4
Proposed framework
The main idea in our work is to end-to-end mapping of some representation of a single 2D input
image to some representation of corresponding 3D shape structure. Defining such structure is
actually based on the main functionality of deep networks, i. e. using different layers of nonlinear
computations in order to feature extraction and transformation [37]. Our proposed framework is
also made of components for feature extraction and transformation.
As stated in section 2, we used three main components in our framework: the first component
is trained to compute some representations for 2D input images. the second component is trained
for reconstructing a 3D shape from some other representations and the third components is used
to map the representation computed in the first component to one used in second component. The
representations in each component should satisfy its objective. for instance, the component for
computing representation from 2D image, should computes a representation that is suitable for
being mapped to 3D representation. On the other hand, the 3D representation should be suitable
for reconstructing the 3D shape in the form of point cloud or binary volume. We observed in our
experiments that more complex representations for 2D representation component achieve better
performance and decoder part of a linear autoencoder as 3D reconstruction component is suitable
for reconstructing the 3D shape in he form of point cloud.
The third component is a structure for mapping 2D low-dim representation to corresponding
3D representation. We actually proposed this component because the manifold of 2D images
may be different from the manifold of 3D shapes especially in the case of complicated shapes like
human faces and therefore a nonlinear mapping function should be found to obtain an accurate
3D representation from 2D image representation.
Note that our proposed frameworks deals directly with dense 3D shapes of the objects instead
of using predefined 3D models for representing a face. This characteristic has the advantage of not
being restricted to a spanning obtained by a few samples by using larger set of training data for
extracting suitable basis vectors for 3D and 2D shape spaces.
The main contributions of this paper are as follows:
• Using deep structures as different components in an interpretable framework for mapping
the representations from a manifold to another and directly reconstructing 3D results. Using
deep structures in each component improves handling the non-linearity of each stage.
• Possibility of unsupervised pre-training of each component using unlabeled data so that the
representations can be improved using realistic data in-the-wild and therefore reducing the
need for labeled data for training because of suitable initialization for training. In fact the
labeled data is needed only for the final aggregation of pre-trained components.
• Using the training data itself to extract the bases of each representation space instead of
using a predefined analytical models.
• Ability of training the framework for inverse rendering of different object classes from a single
2D input image.
In the following subsection, the proposed idea in this paper is demonstrated in the simplest case
(linear) in terms of Singular Value Decomposition (SVD).
4.1
Describing proposed idea in a simple case (i.e. linear) in terms of
Singular Value Decomposition (SVD)
In this section we analytically describe the proposed idea in this paper in the simplest case, i.e.
the linear case. These calculations gives a mathematical insight into the motivation of proposing
this framework. In section 5, we will evaluate our idea using different datasets and compare it
with other possible method in the case of solving 3D shape inverse rendering problem. For this
aim, assume the linear case for each component, i. e. linear autoencoder for 2D encoder and 3D
decoder components and linear feedforward network as the mapping component.
In [2] it is proven that, the optimal representation space resulted by a linear autoencoder, using
L2-norm as loss function, has the same spanning as the sub-space resulted by performing PCA on
the data. Therefore we can consider the use of a linear autoencoder equivalent to performing PCA
on training data. It is therefore can be inferred that the weights of hidden to output layer of the
5
trained linear autoencoder are in fact the bases of the resulted representation subspace which are
found using the training data of autoencoder. This characteristic means that, by using a linear
autoencoder, we can obtain the bases of a representation space for data of any size without the
concern of basis vector decomposition of a large data matrix.
Let X denotes the matrix which is formed by concatenating n1 samples of 2D input images
vectorized as D-dimensional column vectors as 2:
X = [x1 , ..., xn1 ]D×n1
(2)
By applying SVD on X, after subtracting its mean value, the bases (Eigen vectors) of the PCA
subspace can be obtained using SVD as 3:
[Ux , σx , Vx ] = SV D(X)
(3)
where, Ux is the matrix with the columns representing the eigenvectors of XX T , σx is a diagonal
matrix with eigenvalues of XX T as diagonal elements and Vx denotes the matrix with the eigenvectors of X T X on its columns. Therefore we can use the columns of Ux as the basis vectors of
the resulted representation subspace for 2D images, and the k-dimensional representation of X in
the resulting sub-space, spanned by the first k eigenvectors in Ux with largest eigen values in σx ,
can be expressed as Y in 4:
Yk×n1 = UxT XD×n1
(4)
Similarly, considering Z as the column matrix containing n2 p-dimensional vectorized training
samples of 3D shape structures, represented as 3D point clouds or binary volumes, the resulted
k’-dimensional representation of Z using another autoencoder can be expressed as Bk0 ×n2 in 7 :
Z = [z1 , ..., zn2 ]p×n2
(5)
[Uz , σz , Vz ] = SV D(Z)
(6)
Bk0 ×n2 = UzT Zk0 ×p
(7)
To find a transformation between two obtained representations Y and B for the inverse rendering
problem, if a linear feedforward net is used, it is equivalent to performing a least square mechanism
to find a linear relation between Y and B as 8:
Tk×k0 = argminT |B − T Y |
(8)
and the result can be used for reconstructing the 3D shape, say ẑ , as 9:
ẑ = Uz T Y
(9)
On the other hand, we can also solve a direct optimization problem that linearly tries to find the
linear relation between 2D images and corresponding 3D shapes as 10-11.
B̂ = argminB |Z − BX|
(10)
ẑ = B̂X
(11)
These two approaches look for a subspace that best describes the 3D shape of the objects in 2D
image. The first approach first tries to find a low dimensional representation of 2D input image
and corresponding 3D structure for a set of training data and then a linear solution is found for
transforming the 2D representation to 3D one via least squares. Note that the matrices b̂and Uz T
are used as mapping functions for the test data. And the examples used for calculating Ux do not
necessarily correspond to those which were used to obtain Uz and Y. Such mechanism shows the
possibility of unsupervised training of different components in this method.
The second closed form approach in 11 directly finds a linear mapping between 2D input image
and corresponding 3D structure. We believe that the first approach significantly improves the
process of finding a solution for inverse rendering problem and in Table 2 in the section 5.4, we
verified this claim experimentally compared with second approach in 10.
6
4.2
Loss function
As we mentioned before, the purpose of our proposed framework in this paper is to recover the shape
of an object. So we deal with a regression problem. Therefore, we used Root Mean Squared Error
(RMSE) as in 12 as the loss criterion for training the components in the proposed framework. This
criterion is the standard deviation of prediction error and is a standard way to regression analysis
[6, 14].
s
PN
i
i 2
i=1 |xˆo − xgt |
(12)
ERM SE =
N
Where, xˆo i and xigt denote the prediction and ground truth of the ith training sample and N stands
for the sample size. Here, the sample size stands for the size of the mini batch. There are different
loss functions are proposed to be used for inverse rendering problems in different researches [12, 32]
for different applications like reconstruction or recognition. In [12] a binary voxel representation
is used for representing the faces and sigmoid cross entropy loss function is used for training the
network. In our framework we deal with both mesh representation and voxel representations. So
we choose the standard RMSE for training the networks.
5
Experiments
In this section we perform some experiments to evaluate the performance of our proposed framework in different 3D inverse rendering scenarios. We used 2 types of datasets, objects and human
faces having different 3D shape representations, like point clouds and binary volumes, in order
to report and compare the reconstruction results. we also analyse some structures as different
components of the framework.
5.1
Datasets
We report the results of the performance of proposed framework on two types of datasets: objects
and human faces. In each case we trained a separate structure using that type of data representation. we also used some unlabeled 2D and 3D datasets in order to unsupervised pre-training the
components.
In the case of human faces, we used LFW dataset [11] as 2D unlabeled data for unsupervised
pre-training and Besel Face Model [22] for generating synthetic data as labeled training data along
with the Bosphorus dataset [28] as realistic labeled data for training and fine-tuning respectively.
Note that for aggregating all components using supervised training, we used faces generated by
BESEL Face Model (BFM) having natural expression rendered with 3 poses, i.e. -45, 0, 45 degrees
about z axis. Therefore the objective of human face shape inverse rendering in our work is to pose
invariant identity recovering of faces from a single 2D image. The 3D faces in our framework are
represented as 3D point clouds with 54903 vertices in 3D space which are normalized between 0,1
in each dimension. We generated 4000 training and 1000 test samples in natural expression and
rendered 2D images of size 32*32 with mentioned poses as 2D input images.
In the case of objects, we used some categories of imagenet [5] dataset as unlabeled data for pretraining and ModelNet10 [36] and ShapeNet [4] datasets as labeled training data which are to the
best of our knowledge, well-known datasets for 3D object reconstructions. we used 10 categories
for training the networks in the form of 30 × 30 × 30 binary voxel grid and rendered each training
sample as a 32 × 32 gray scale image from 8 viewpoints (between 0 and 180 about z axis). In the
case of linear autoencoders we vectorized the voxel grid as a binary column vector of size 27000.
We incorporate well-known alexnet in our experiments for evaluating the use of a deep CNN for
computing 2D representations for the face dataset. In this case we rendered and used 2D colored
images of the faces in each existing dataset with the size 227 × 227 × 3 which is the standard input
size to alexnet framework.
5.2
Parameter setting
Table 1 includes the configuration of each structure used for each type of dataset for the type of
used structure. Since we applied inverse rendering on different data types, we had to use different
structures for the network components in each case.
7
Encoder
Mapping
Decoder
Object (binary Volume)
Linear
Nonlinear
Convolutional
3 × 3 × 1−
3 × 3 × ×32−
1024-60
1024 - 100- 60
3 × 3 × 16
1 × 3 × 1 × 100−
60-100-400
60-100-400
400
3 × 3 × 3 × 32
3 × 3 × 3 × 16
400-27000 400-800-27000
30 × 30 × 30
Human Face (Point cloud)
Nonlinear
Convolutional
3 × 3 × 1−
3 × 3 × 32
1024 - 60 1024 - 100 - 60
3 × 3 × 16
1 × 3 × 100−
60-100-98
60-100-98
98
Linear
98-160470
98-200-160470
Table 1: Network configuration for each components using different structures
Besel face Dataset
ModelNet Dataset
ShapeNet Dataset
LS using low dim representation
0.8727
0.0970
0.0712
Directly finding Ls between 2D and 3D
2.2869
0.1469
0.2455
Table 2: Comparison of average test RMSE obtained by different datasets using closed form linear
solutions in section 4.1
Note that in the case of object datasets, since there are 10 classes of objects, there were need
to layers with larger sizes for representation units.
The learning rate was set for each configuration so that it gives better performance. In most
cases we uses 0.001 for 1000 epochs and then reduce it to 0.00001 for another 1000 epochs. In the
case of mesh representation the learning was slower and needed smaller learning rate for longer
training time. Batch size was set to a constant, in most cases equal to 40 for training. We trained
our networks from scratch on waterloo university servers: gpu-pr1-01 and gpu-pt1-01, saspc73, and
sharcnet-graham machines which are well-known servers having several gpus suitable for training
deep networks.
5.3
Evaluation mechanism
We used RMSE as the evaluation criterion on test and validation sets in different parts of our
experiments. In the case of face datasets we also showed the RMSE in the form of Heat maps,
showing point-wise RMSE of reconstructed face on the ground truth face, as a suitable tool for
error visualization in different regions of the reconstructed face.
We start the experiments by first reporting the evaluation results of closed form linear solutions
described in section 4.1 using different data types. in the next step we will use equivalent deep
structure by changing the mapping component to nonlinear structure and show that the results
can be significantly improved. As a nonlinear component for encoder part of our framework, we
fine-tuned alexnet using our labeled human face dataset and report the results. We then show
some reconstruction results on realistic human face dataset and compare our results with some of
the most recent methods for 3D reconstruction.
5.4
Results of solving linear closed form formula for inverse rendering
Regarding the analytical description of proposed idea in section 4.1, here we show that finding a
linear mapping between low dimensional representations of 2D and 3D data achieves significantly
better results rather that finding a direct linear mapping between them (2D and 3D data). Table
2 and figures ?? include the numerical and visual results of applying these methods on different
datasets, where “LS using low dim representation” stands for finding the mapping between low
dimensional representations of 2D and 3D data using least squares and “Directly finding ls between
2D and 3D” stands for finding a least squares solution for direct mapping 2D input images to 3D
data.
The numerical results in Table 2 indicate the effectiveness of the proposed idea of mapping low
dimensional representations instead of direct mapping in terms of RMSE for all existing datasets.
Looking at Figure 2, first, it is clear that using linear methods is not powerful enough for
addressing 3D shape inverse rendering problem from a single 2d input image and there is need
8
-
input image (32 × 32)
Ground Truth
LS using low dim
representation
Directly finding Ls
between 2D and 3D
Figure 2: Visual results of using linear methods for finding a mapping between 2D to 3D spaces
for BESE face dataset
to use nonlinear tools in this fields especially in the case of human face inverse rendering. And
second, we can see that using the low dimensional representation mapping works better that direct
mapping visually.
input images (32 × 32each)
Ground Truth
LS using low dim
representation
Directly finding Ls
between 2D and 3D
Figure 3: Visual results of comparing linear least squares for finding a mapping between 2D and
3D space for Modelnet10 dataset
In the next step of our experiments, we set up a deep structure consist of linear autoencoders
for finding representations and a nonlinear feedforward network for mapping the representations
as a strat point for using nonlinearity in the proposed framework.
5.5
Using linear encoders and decoders with nonlinear feedforward network as mapping component:
As a start point of using nonlinear deep structure in our framework, we only used a nonlinear
feedforward network as the mapping component and trained the network using 2000 samples of
BESEL face model dataset and training set of Modelnet10 dataset for 1000 epochs. Figures 4,5
show the results on test data for face and object datasets, respectively.
It can be observed that using nonlinearity only in the mapping component in our framework
results in significant improvement Visually and numerically. For instance, In the case of face
datasets, we have face-like reconstructions. this observation also shows that the relation between
subspaces of 2D and 3D representations have nonlinear relation.
In the next subsection, after fixing the nonlinear feedforward network as mapping component,
we consider the use of more complicated representations for 2D and 3D data. To do this, we
fine-tuned alexnet as encoder of 2D input image.
5.6
Transfer learning from Alexnet as 2D encoder
Figure 6 shows the proposed framework which uses Alexnet pre-trained model as encoder component. This model, trained on more than one million images of about 1000 categories of images from
ImageNet database [16] like human, chair, pencil and many other categories, is a well-known model
for extracting rich features from images for classification of them into 1000 object categories. By
9
input images (32 × 32)
Ground Truth
Network Reconstruction
Heat Map (RMSE)
Average RMSE (test error)= 0.54611
Figure 4: Results of using nonlinear feedforward network as mapping component on BESEL face
dataset
input images (32 × 32)
Ground Truth
Network Reconstruction
Average RMSE (test error) = 0.0625
Figure 5: Results of using nonlinear feedforward network as mapping component on Modelnet
dataset
using this model as encoder component, we first verify the possibility of using pre-trained components in our framework and second analyze the use of more complicated representations for solving
inverse rendering problem in our proposed framework. We fine-tuned alexnet using a set of colored
Figure 6: alexnet as encoder component in the proposed framework
2D input images of about 2000 samples generated by BESEL face dataset, and 10000 samples from
Modelnet10 and the results are shown in Figures 7, 8 respectively. in the case of Modelnet10, Since
there were no texture associated with the object to be rendered as a color image, we just represent
the input images with the size 227 × 227 × 3 to feed to the network.
From Figures 7, 8 it is observable that using this configuration and with computing more
complex representations for 2D input images, the accuracy of reconstruction will be improved
compared with the linear autoencoder. Actually, the objective of encoder component is to compute
representations which are more suitable for being mapped to 3D representation and the results show
that more complicated representations are appropriate for this aim.
In the case of decoder component, the aim is recovering sharp 3D shapes from 3D represen10
input image
(227 × 227 × 3)
Ground Truth
Network Reconstruction
Heat Map ( RMSE)
Average RMSE = 0.37124
Figure 7: Visual and numerical results of using alexnet as encoder in proposed framework (BESEL
face Dataset)
input image
(227 × 227 × 3 each)
Ground Truth
Network Reconstruction
Average RMSE = 0.0220
Figure 8: Visual and numerical results of using alexnet as encoder in proposed framework (Modelnet10 Dataset)
tations. therefore, the representation in this component should be so that the 3D shape can be
inferred from. in the case of point cloud 3d shapes, where we directly deal with real 3D coordinate
values the linear decoder gives the most promising result compared with nonlinear and convolutional structures with blur results. in the case of binary volume 3D shapes, where the output can
be digitized to 0 and 1 using some threshold value, nonlinear structures show promising performances. figure 9 shows the digaram of RMSE vs epoch using different mentioned structures for
different types of 3D shapes.
5.7
Pose invariant reconstruction
Figure 10 shows the power of proposed framework in reconstruction of a sample face input image
captured from 3 different poses.
5.8
Reconstruction results on realistic datasets
We evaluated the final configuration of our framework (Alexnet encoder – nonlinear mapping –
linear decoder) on Bosphorus Face dataset as a realistic 3D face dataset. Figure 8 shows the
results and their heat map in terms of RMSE on Bosphorus dataset.We also compared the visual
reconstruction results of proposed framework with [12] and [30] on some realistic images.
11
(a) ModelNet Dataset
(b) BESEL Face Dataset
Figure 9: Validation error vs Epoch for Different structures used as components in proposed
frameworks for (a) ModelNet dataset and (b) Face dataset where Lin: Linear encoder or decoder,
FF: Nonlinear FeedForward network, Conv: Convolutional encoder or decoder
Input image
Network Reconstruction
Heat map (RMSE)
Figure 10: Result of proposed framework on Bosphorus face dataset
In the case of realistic images we first manually cropped the face from input image and then
aligned tha intensity of the image with a sample image from training samples of our framework
using Procrustes analysis used in [29] and then use them as input image to our framework. Note
that in order to compute RMSE for heat map we computed the minimum RMSE from Ground
truth to network reconstruction.
The results in Figure 11 show that the network could successfully recover the main features in
each face. Note that using the linear decoder in the case of face datasets is equivalent to performing
PCA on all of training samples (about 2000) and this causes achieving better results compared with
a predefined 3DMM using small set of training data. Figure 9 shows some visual reconstruction
results, using some realistic images, obtained by proposed framework compared with two similar
recent reconstruction methods [30] and [12] which where described in section 3. We can see that
the results obtained by proposed framework are comparable with unsupervised raining methods
and can detect more shape details compared with the binary volume representation for 3D shapes.
We believe this is because of dealing directly with point clouds and extracting basis vectors from
a large set of training data.
6
Discussion
As we observed in the experiments, using a convolutional network as 2D encoder and mapping component in our framework significantly improves its performance. However in the case of 3D decoder
using a non-convolutional network (linear or non-linear) results in better reconstruction results.
We believe this is because a in convolutional Autoencoder, reconstruction of the representation
12
Figure 11: Result of proposed framework on Bosphorus face dataset
obtained from data may be inaccurate because of the information loss resulted by convolution and
pooling units.
Another Property of our proposed framework is the use of autoencoder to directly reconstruct
the 3D shape structure by learning from large set of examples. this structure Uses the training
data itself to extract bases of the shape space. We believe and this characteristic leads to have
unbiased reconstructions rather than using a predefined 3D model which is biased toward mean
shape of the training samples. On the other hand, working directly with point cloud representation
for representing the 3D shape helps the framework to achieve more detailed results.
As a disadvantage of regressing process in our framework, we can mention the existing loss
functions. in the case of situation which the output of the network should be some exact values
for regression, there is need to define a suitable loss function to guide the learning process toward
finding desired solution. If we can turn the problem of regression into the form of some classification
problem using losses like cross entropy and so on, we will achieve more desirable solutions like the
results obtained by Generator Adversarial Netwok (GAN) frameworks.
7
Conclusions and future works
In this paper, a semi-supervised interpretable framework is proposed for pose invariant 3D shape
inverse rendering from a single 2D input image. Using autoencoder based on mapping suitable
low dimensional representations computed from 2D and 3D data components for extracting representations and reconstruction of final 3D shapes in the proposed framework, on one hand, makes
it possible to use unlabeled data for unsupervised pre-training and therefore reducing the need to
large sets of labeled data for training the framework and on the other hand, helps to achieve more
promising and detailed reconstructions because of extraction of bases vector of 3D shape space
from the large set of training data compared with predefined 3D models.
in this proposed framework two types of 3D shape representations were used for reconstruction:
point cloud and binary volumes. The obtained results show that binary volumes are more suitable
representations for more complicated deep networks. However they express blur shapes with high
dimensional representation and can express less details compared with point cloud representations.
on the other hand, working with exact 3D coordinates of point clouds is not too easy for deep
networks and linear structures are preferred for such shape reconstructions.
In the future attempts one can unify the Inverse rendering framework for working with both
point cloud and volume shape structures for different class of objects.
13
Input image
Tewari et al. [30]
Jackson [12]
Ours
Figure 12: Comparing viual reconstruction results with [12, 30]
References
[1] Aldrian, O., and Smith, W. A. Inverse rendering of faces with a 3d morphable model.
IEEE transactions on pattern analysis and machine intelligence 35, 5 (2013), 1080–1093.
[2] Baldi, P., and Hornik, K. Neural networks and principal component analysis: Learning
from examples without local minima. Neural networks 2, 1 (1989), 53–58.
[3] Blanz, V., and Vetter, T. A morphable model for the synthesis of 3d faces. In Proceedings
of the 26th annual conference on Computer graphics and interactive techniques (1999), ACM
Press/Addison-Wesley Publishing Co., pp. 187–194.
[4] Chang, A. X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z.,
Savarese, S., Savva, M., Song, S., Su, H., et al. Shapenet: An information-rich 3d
model repository. arXiv preprint arXiv:1512.03012 (2015).
[5] Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. Imagenet: A
large-scale hierarchical image database. In Computer Vision and Pattern Recognition, 2009.
CVPR 2009. IEEE Conference on (2009), IEEE, pp. 248–255.
[6] Dou, P., Shah, S. K., and Kakadiaris, I. A. End-to-end 3d face reconstruction with
deep neural networks. arXiv preprint arXiv:1704.05020 (2017).
[7] Fouhey, D. F., Gupta, A., and Zisserman, A. Understanding higher-order shape via 3d
shape attributes. arXiv preprint arXiv:1612.06836 (2016).
[8] Gadelha, M., Maji, S., and Wang, R. 3d shape induction from 2d views of multiple
objects. arXiv preprint arXiv:1612.05872 (2016).
[9] Gao, Y., and Yuille, A. L. Exploiting symmetry and/or manhattan properties for 3d
object structure estimation from single and multiple images. arXiv preprint arXiv:1607.07129
(2016).
[10] Garrido, P., Zollhöfer, M., Casas, D., Valgaerts, L., Varanasi, K., Pérez, P.,
and Theobalt, C. Reconstruction of personalized 3d face rigs from monocular video. ACM
Transactions on Graphics (TOG) 35, 3 (2016), 28.
[11] Huang, G. B., Ramesh, M., Berg, T., and Learned-Miller, E. Labeled faces in the
wild: A database for studying face recognition in unconstrained environments. Tech. Rep.
07-49, University of Massachusetts, Amherst, October 2007.
14
[12] Jackson, A. S., Bulat, A., Argyriou, V., and Tzimiropoulos, G. Large pose 3d
face reconstruction from a single image via direct volumetric cnn regression. arXiv preprint
arXiv:1703.07834 (2017).
[13] Jiang, L., Zhang, J., Deng, B., Li, H., and Liu, L. 3d face reconstruction with geometry
details from a single image. arXiv preprint arXiv:1702.05619 (2017).
[14] Kim, H., Zollhöfer, M., Tewari, A., Thies, J., Richardt, C., and Theobalt, C.
Inversefacenet: Deep single-shot inverse face rendering from a single image. arXiv preprint
arXiv:1703.10956 (2017).
[15] Kim, K., Torii, A., and Okutomi, M. Multi-view inverse rendering under arbitrary
illumination and albedo. In European Conference on Computer Vision (2016), Springer,
pp. 750–767.
[16] Krizhevsky, A., Sutskever, I., and Hinton, G. E. Imagenet classification with deep
convolutional neural networks. In Advances in neural information processing systems (2012),
pp. 1097–1105.
[17] Liu, J., Yu, F., and Funkhouser, T. Interactive 3d modeling with a generative adversarial
network. arXiv preprint arXiv:1706.05170 (2017).
[18] Lun, Z., Gadelha, M., Kalogerakis, E., Maji, S., and Wang, R. 3d shape reconstruction from sketches via multi-view convolutional networks. arXiv preprint arXiv:1707.06375
(2017).
[19] Marschner, S. R., and Greenberg, D. P. Inverse rendering for computer graphics.
Cornell University, 1998.
[20] Parkhi, O. M., Vedaldi, A., Zisserman, A., et al. Deep face recognition. In BMVC
(2015), vol. 1, p. 6.
[21] Patow, G., and Pueyo, X. A survey of inverse rendering problems. In Computer graphics
forum (2003), vol. 22, Wiley Online Library, pp. 663–687.
[22] Paysan, P., Knothe, R., Amberg, B., Romdhani, S., and Vetter, T. A 3d face
model for pose and illumination invariant face recognition. In Advanced Video and Signal
Based Surveillance, 2009. AVSS’09. Sixth IEEE International Conference on (2009), Ieee,
pp. 296–301.
[23] Piotraschke, M., and Blanz, V. Automated 3d face reconstruction from multiple images
using quality measures. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition (2016), pp. 3418–3427.
[24] Poggio, T., and Koch, C. Ill-posed problems in early vision: from computational theory
to analogue networks. Proceedings of the Royal Society of London B: Biological Sciences 226,
1244 (1985), 303–323.
[25] Rezende, D. J., Eslami, S. A., Mohamed, S., Battaglia, P., Jaderberg, M., and
Heess, N. Unsupervised learning of 3d structure from images. In Advances in Neural Information Processing Systems (2016), pp. 4996–5004.
[26] Richardson, E., Sela, M., and Kimmel, R. 3d face reconstruction by learning from
synthetic data. In 3D Vision (3DV), 2016 Fourth International Conference on (2016), IEEE,
pp. 460–469.
[27] Richardson, E., Sela, M., Or-El, R., and Kimmel, R. Learning detailed face reconstruction from a single image. arXiv preprint arXiv:1611.05053 (2016).
[28] Savran, A., Alyüz, N., Dibeklioğlu, H., Çeliktutan, O., Gökberk, B., Sankur,
B., and Akarun, L. Bosphorus database for 3d face analysis. Biometrics and identity
management (2008), 47–56.
[29] Skoglund, K. Three-dimensional face modelling and analysis. Master’s thesis, Technical
University of Denmark, DTU, DK-2800 Kgs. Lyngby, Denmark, 2003.
15
[30] Tewari, A., Zollhöfer, M., Kim, H., Garrido, P., Bernard, F., Pérez, P., and
Theobalt, C. Mofa: Model-based deep convolutional face autoencoder for unsupervised
monocular reconstruction. arXiv preprint arXiv:1703.10580 (2017).
[31] Thies, J., Zollhofer, M., Stamminger, M., Theobalt, C., and Nießner, M.
Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition (2016), pp. 2387–2395.
[32] Tran, A. T., Hassner, T., Masi, I., and Medioni, G. Regressing robust and discriminative 3d morphable models with a very deep neural network. arXiv preprint arXiv:1612.04904
(2016).
[33] Wang, J., Che, C., Galeotti, J., Horvath, S., Gorantla, V., and Stetten, G.
Ultrasound tracking using probesight: Camera pose estimation relative to external anatomy
by inverse rendering of a prior high-resolution 3d surface map. In Applications of Computer
Vision (WACV), 2017 IEEE Winter Conference on (2017), IEEE, pp. 825–833.
[34] Wolpert, D. H., and Macready, W. G. No free lunch theorems for optimization. IEEE
transactions on evolutionary computation 1, 1 (1997), 67–82.
[35] Wood, E., Baltrušaitis, T., Morency, L.-P., Robinson, P., and Bulling, A. A 3d
morphable eye region model for gaze estimation. In European Conference on Computer Vision
(2016), Springer, pp. 297–313.
[36] Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., and Xiao, J. 3d shapenets:
A deep representation for volumetric shapes. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition (2015), pp. 1912–1920.
[37] Yu, D., and Deng, L. Automatic speech recognition: A deep learning approach. Springer,
2014.
[38] Zhu, X., Lei, Z., Liu, X., Shi, H., and Li, S. Z. Face alignment across large poses: A 3d
solution. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
(2016), pp. 146–155.
16
| 2cs.AI
|
1
Computation of Optimal Control Problems with
Terminal Constraint via Variation Evolution
Sheng ZHANG, Bo LIAO, and Fei LIAO
(2017.12)
Abstract: Enlightened from the inverse consideration of the stable continuous-time dynamics evolution, the Variation Evolving Method
(VEM) analogizes the optimal solution to the equilibrium point of an infinite-dimensional dynamic system and solves it in an
asymptotically evolving way. In this paper, the compact version of the VEM is further developed for the computation of Optimal Control
Problems (OCPs) with terminal constraint. The corresponding Evolution Partial Differential Equation (EPDE), which describes the
variation motion towards the optimal solution, is derived, and the costate-free optimality conditions are established. The explicit analytic
expressions of the costates and the Lagrange multipliers adjoining the terminal constraint, related to the states and the control variables,
are presented. With the semi-discrete method in the field of PDE numerical calculation, the EPDE is discretized as finite-dimensional
Initial-value Problems (IVPs) to be solved, with common Ordinary Differential Equation (ODE) numerical integration methods.
Key words: Optimal control, dynamics stability, variation evolution, initial-value problem, costate-free optimality condition.
I. INTRODUCTION
Optimal control theory aims to determine the inputs to a dynamic system that optimize a specified performance index while
satisfying constraints on the motion of the system. It is closely related to engineering and has been widely studied [1]. Because of
the complexity, Optimal Control Problems (OCPs) are usually solved with numerical methods. Various numerical methods are
developed and generally they are divided into two classes, namely, the direct methods and the indirect methods [2]. The direct
methods discretize the control or/and state variables to obtain the Nonlinear Programming (NLP) problem, for example, the
widely-used direct shooting method [3] and the classic collocation method [4]. These methods are easy to apply, whereas the
results obtained are usually suboptimal [5], and the optimal may be infinitely approached. The indirect methods transform the OCP
to a Boundary-value Problem (BVP) through the optimality conditions. Typical methods of this type include the well-known
indirect shooting method [2] and the novel symplectic method [6]. Although be more precise, the indirect methods often suffer
from the significant numerical difficulty due to the ill-conditioning of the Hamiltonian dynamics, that is, the stability of costates
dynamics is adverse to that of the states dynamics [7]. The recent development, representatively the Pseudo-spectral (PS) method
[8], blends the two types of methods, as it unifies the NLP and the BVP in a dualization view [9]. Such methods inherit the
advantages of both types and blur their difference.
Theories in the control field often enlighten strategies for the optimal control computation, for example, the non-linear variable
transformation to reduce the variables [10]. Recently, a new Variation Evolving Method (VEM), which is enlightened by the states
evolution within the stable continuous-time dynamic system, is proposed for the optimal control computation [11][12][13]. The
The first author is with the Computational Aerodynamics Institution, China Aerodynamics Research and Development Center, Mianyang, 621000, China. (e-mail:
[email protected]).
2
VEM also synthesizes the direct and indirect methods, but from a new standpoint. The Evolution Partial Differential Equation
(EPDE), which describes the evolution of variables towards the optimal solution, is derived from the viewpoint of variation motion,
and the optimality conditions will be asymptotically met under this frame. In Refs. [11] and [12], besides the states and the controls,
the costates are also employed in developing the EPDE, and this increases the complexity of the computation. Ref. [13] proposed
the compact version of the VEM that uses only the original variables, but it can only handles a class of OCPs with free terminal
states. In this paper, the VEM is further developed to accommodate the OCPs with terminal constraint, and the corresponding
evolution equations are derived.
Throughout the paper, our work is built upon the assumption that the solution for the optimization problem exists. We do not
describe the existing conditions for the purpose of brevity. Relevant researches such as the Filippov-Cesari theorem are
documented in [14]. In the following, first the principle of the VEM and the results regarding the OCPs without terminal constraint
are reviewed. Then the VEM for OCPs with terminal constraint is developed. During this course, the equivalent costate-free
optimality conditions are established, and the explicit analytic solution of costates and Lagrange multipliers in the classic treatment
are obtained. Later illustrative examples are solved to verify the effectiveness of the method.
II. PRELIMINARIES
A. Principle of VEM
The VEM is a newly developed method for the optimal solutions. It originates from the dynamics stability theory in the control
field.
Lemma 1 [15] (with small adaptation): For a continuous-time autonomous dynamic system like
x = f ( x )
where x ∈ \ n is the state, x =
(1)
dx
is its time derivative, and f : \ n → \ n is a vector function. Let x̂ , contained within the
dt
domain D , be an equilibrium point that satisfies f ( xˆ ) = 0 and D . If there exists a continuously differentiable function
V : D → \ such that
i) V ( xˆ ) = c and V ( x ) > c in D / { xˆ } .
ii) V ( x ) ≤ 0 in D and V ( x ) < 0 in D / { xˆ } .
where c is a constant. Then x = xˆ is an asymptotically stable point in D .
Lemma 1 aims to the dynamic system with finite-dimensional states, and it may be directly generalized to the
infinite-dimensional case as
Lemma 2: For an infinite-dimensional dynamic system described by
δ y ( x)
= f ( y, x)
δt
(2)
∂y ( x, t )
= f ( y, x)
∂t
(3)
or presented equivalently in the PDE form as
where “ δ ” denotes the variation operator and “ ∂ ” denotes the partial differential operator. x ∈ \ is the independent
variable, y ( x) ∈ \ n ( x) is the function vector of x , and f : \ n ( x) × \ → \ n ( x) is a vector function. Let yˆ ( x) , contained
3
within a certain function set D ( x) , is an equilibrium function that satisfies f ( yˆ ( x), x) = 0 . If there exists a continuously
differentiable functional V : D ( x) → \ such that
i) V ( yˆ ( x) ) = c and V ( y ( x) ) > c in D ( x) / { yˆ ( x)} .
ii) V ( y ( x) ) ≤ 0 in D ( x) and V ( y ( x) ) < 0 in D ( x) / { yˆ ( x)} .
where c is a constant. Then y ( x) = yˆ ( x) is an asymptotically stable solution in D ( x ) .
In the system dynamics theory, from the stable dynamics, we may construct a monotonously decreasing function (or functional)
V , which will achieve its minimum when the equilibrium is reached. Inspired by it, now we consider its inverse problem, that is,
from a performance index function to derive the dynamics that minimize this performance index, and optimization problems are
just the right platform for practice. Under this thought, the optimal solution is analogized to the equilibrium of a dynamic system
and is anticipated to be obtained in an asymptotically evolving way. Accordingly, a virtual dimension, the variation time τ , is
introduced to implement the idea that a variable x (t ) evolves to the optimal solution to minimize the performance index within the
dynamics governed by the variation dynamic evolution equations. Fig. 1 illustrates the variation evolution process of the VEM in
solving the OCP. Through the variation motion, the initial guess of variable will evolve to the optimal solution.
Fig. 1. The illustration of the variable evolving along the variation time
τ
in the VEM.
The VEM bred under this idea is demonstrated for the unconstrained calculus-of-variations problems and the OCPs with free
terminal states [11][13]. The variation dynamic evolution equations, derived under the frame of the VEM, may be reformulated as
the EPDE and the Evolution Differential Equation (EDE), by replacing the variation operator “ δ ” with the partial differential
operator “ ∂ ” and differential operator “ d ”. Since the right function of the EPDE only depends on the time t , it is suitable to be
solved with the well-known semi-discrete method in the field of PDE numerical calculation [16]. Through the discretization along
the normal time dimension, the EPDE is transformed to the finite-dimensional Initial-value Problem (IVP) to be solved, with the
mature Ordinary Differential Equation (ODE) integration methods. Note that the resulting IVP is defined with respect to the
variation time τ , not the normal time t .
B. Results for OCPs with free terminal states
To clear the development, first the results for the OCPs with free terminal states, obtained under the frame of the VEM, is
recalled.
Problem 1：Consider performance index of Bolza form
J = ϕ ( x (t f ), t f ) + ∫ L ( x (t ), u(t ), t ) dt
tf
t0
subject to the dynamic equation
(4)
4
x = f ( x,u, t )
(5)
where t ∈ \ is the time. x ∈ \ n is the state vector and its elements belong to C 2 [t0 , t f ] . u ∈ \ m is the control vector and its
elements belong to C1[t0 , t f ] . The function L : \ n × \ m × \ → \ and its first-order partial derivatives are continuous with
respect to x , u and t . The function ϕ : \ m × \ → \ and its first-order and second-order partial derivatives are continuous with
respect to x and t . The vector function f : \ n × \ m × \ → \ n and its first-order partial derivatives are continuous and Lipschitz
in x , u and t . The initial time t0 is fixed and the terminal time t f is free. The initial boundary conditions are prescribed as
x (t0 ) = x0
(6)
and the terminal states are free. Find the optimal solution ( xˆ , uˆ ) that minimizes J , i.e.
( xˆ , uˆ ) = arg min( J )
(7)
For this class of OCPs, presuming we already have a feasible initial solution of the states and control variables that satisfies Eqs.
(5) and (6), the variation dynamic evolution equations derived from the VEM are
t
δx
δu
= ∫ Φo (t , s ) f u ( s )
(s) d s
t
δτ
δτ
(8)
0
δu
= − Kpu (t )
δτ
δtf
= −kt L + ϕt + ϕ x T f
δτ
f
(
(9)
)
(10)
tf
where
pu (t ) = Lu + fu Tϕ x + fu T
(∫
tf
t
(
)
Φo T (σ , t ) Lx (σ ) + ϕtx (σ ) + ϕ xx T (σ ) f (σ ) + f x (σ )T ϕ x (σ ) d σ
)
(11)
K is the m × m dimensional positive-definite matrix and kt f is a positive constant. Φo (t , s ) is the n × n dimensional state
transition matrix from time point s to time point t , which satisfies
∂
Φo (t , s ) = f x (t )Φo (t , s)
∂t
(12)
Use the partial differential operator “ ∂ ” and the differential operator “ d ”, the variation dynamic evolution equations may be
reformulated to get the following EPDE and EDE as
∂
∂τ
∂u
⎡ t
⎤
⎡ x ⎤ ⎢ ∫ Φo (t , s ) f u ( s ) ( s ) d s ⎥
t0
=
∂τ
⎢u⎥ ⎢
⎥
⎣ ⎦ ⎢
− Kpu
⎣
⎦⎥
dt f
dτ
(
= − kt f L + ϕ t + ϕ x T f
with the definite conditions including the initial guess of t f , i.e., t f
τ =0
)
(13)
(14)
tf
= tf , and,
⎡ x (t ,τ ) ⎤
⎡ x (t ) ⎤
=⎢
⎢ u(t ,τ ) ⎥
⎥
⎣
⎦ τ = 0 ⎣ u (t ) ⎦
(15)
where x (t ) and u (t ) are the initial feasible solution. The solution at τ = ∞ , determined by Eqs. (13), (14) and (15), will satisfy
5
pu = 0
(16)
L(t f ) + φt (t f ) + ϕ x T (t f ) f (t f ) = 0
(17)
Eqs. (16) and (17) are the first-order optimality conditions for Problem 1 without the employment of costates. They are proved
equivalent to the traditional ones with costates [17].
III. VEM FOR OCPS WITH TERMINAL CONSTRAINT
A. Problem definition
In this paper, we consider the OCPs with terminal constraint that is defined as
Problem 2：Consider performance index of Bolza form
J = ϕ ( x (t f ), t f ) + ∫ L ( x (t ), u(t ), t ) dt
tf
(18)
t0
subject to the dynamic equation
x = f ( x,u, t )
(19)
where t ∈ \ is the time. x ∈ \ n is the state vector and its elements belong to C 2 [t0 , t f ] . u ∈ \ m is the control vector and its
elements belong to C1[t0 , t f ] . The function L : \ n × \ m × \ → \ and its first-order partial derivatives are continuous with
respect to x , u and t . The function ϕ : \ m × \ → \ and its first-order and second-order partial derivatives are continuous with
respect to x and t . The vector function f : \ n × \ m × \ → \ n and its first-order partial derivatives are continuous and Lipschitz
in x , u and t . The initial time t0 is fixed and the terminal time t f is free. The initial and terminal boundary conditions are
respectively prescribed as
x (t0 ) = x0
(
(20)
)
g x (t f ), t f = 0
(21)
where g : \ n × \ → \ q is a q dimensional vector function with continuous first-order partial derivatives. Find the optimal
solution ( xˆ , uˆ ) that minimizes J , i.e.
( xˆ , uˆ ) = arg min( J )
(22)
Compared with the OCPs with free terminal states as defined in Problem 1, Problem 2 covers them because Problem 1 may be
considered as a special case with g = 0 .
B. Derivation of variation dynamic evolution equations
Instead of circumventing Problem 2 by constructing an equivalent unconstrained functional problem that has the same extremum
[11], we will address Problem 2 in the way similar to Ref. [13]. We also consider the problem within the feasible solution domain
D o , in which any solution satisfies Eqs. (19), (20) and (21). First we transform the Bolza performance index to the equivalent
Lagrange type, i.e.
J =∫
tf
t0
(ϕ + ϕ
t
x
T
)
f ( x,u, t ) + L( x , u, t ) dt
(23)
where ϕt and ϕ x are the partial derivatives notated as before. Differentiating Eq. (23) with respect to the variation time τ gives
6
δtf
t
δJ
= (ϕt + ϕ x T f + L)
+∫
t
t δτ
δτ
f
0
f
δu⎞
⎛
T
T
T
T δx
dt
+ (ϕ x T fu + Lu T )
⎜ (ϕtx + f ϕ xx + ϕ x f x + Lx )
δτ
δτ ⎟⎠
⎝
(24)
where ϕtx and ϕ xx are second-order partial derivatives in the form of (column) vector and matrix, and f x and f u are the Jacobi
matrixes. For the solutions in D o ,
δx
δu
and
are related because of Eq. (19), and they need to satisfies the following variation
δτ
δτ
equation as
δ x
δx
δu
= fx
+ fu
δτ
δτ
δτ
with the initial condition
(25)
δx
= 0 . Note that f x and f u are time-dependent matrixes linearized at the feasible solution x (t ) and
δτ t
0
u(t ) . Eq. (25) is a linear time-varying equation and has a zero initial value. Thus according to the linear system theory [18], its
solution may be explicitly expressed by Eq. (8). Use Eq. (8) and follow the same derivation as Ref. [13], we again obtain
δtf
t
δJ
δu
= (ϕt + ϕ x T f + L)
+ ∫ pu T
dt
t
t
δτ
δτ
δτ
f
(26)
0
f
Different from the OCPs with free terminal states, we cannot use the variation dynamic evolution equations (9) and (10) to achieve
δJ
≤ 0 . This is because in that way the terminal constraint (21) is not guaranteed, and we have to find a solution that not only
δτ
guarantees
δJ
≤ 0 but also satisfies the variation equation of Eq. (21) as
δτ
δtf
t
δu
g x ∫ Φo (t f , t ) f u (t ) (t ) d t + ( g x f + gt )
=0
t
δτ
δτ
f
f
f
0
(27)
f
Theorem 1: The following variation dynamic evolution equations guarantees that the solution stays in the feasible domain and the
performance index
δJ
≤0
δτ
δu
= − K pu + f uT Φ T (t f , t ) g x T π
δτ
(
δtf
δτ
f
(
)
= − kt f L + ϕt + ϕ x T f + π T ( g x f f + gt f )
(28)
)
(29)
tf
where K is the m × m dimensional positive-definite matrix and kt f is a positive constant, pu is defined in Eq. (11) and the
parameter vector π ∈ \ q is the solution of the linear matrix equation
Mπ = − r
(30)
where the q × q dimensional matrix M and the q dimension vector r are
M = gx f
(∫
tf
t0
)
Φ (t f , t ) f u Kf uT Φ T (t f , t ) d t g x f T + kt f ( g x f f + gt f )( g x f f + gt f )T
r = gxf
(∫
tf
t0
)
Φ (t f , t ) f u Kpu d t + kt f ( g x f f + gt f ) (ϕt + ϕ x T f + L)
Moreover, under the evolution equations (28) and (29),
δJ
= 0 occurs only when
δτ
tf
tf
(31)
(32)
7
pu + f u T Φ T (t f , t ) g x f T π = 0
(ϕt + ϕ x T f + L + π T ( g x f f + gt f ))
(33)
tf
=0
(34)
and the optimal value of π satisfies
⎡ M s1 ⎤
⎡ rs1 ⎤
⎢ M ⎥ π = − ⎢r ⎥
⎣ s2 ⎦
⎣ s2 ⎦
(35)
where the q × q dimensional matrixes M s1 , M s 2 and the q dimension vectors rs1 , rs 2 are
M s1 = g x f
(∫
tf
t0
)
Φ (t f , t ) f u f uT Φ T (t f , t ) d t g x f T
M s 2 = ( g x f f + gt f )( g x f f + gt f )T
rs1 = g x f
tf
∫t
tf
Φ (t f , t ) f u pu d t
(36)
(37)
(38)
0
rs 2 = ( g x f f + gt f ) (ϕt + ϕ x T f + L)
tf
(39)
Proof: We will derive Eqs. (28) and (29) though the optimization theory. Reformulate Eq. (26) as a constrained optimization
problem subject to Eq. (27) as
J t1 = (ϕt + ϕ x T f + L)
Note that now
δtf
tf
δτ
tf
+ ∫ pu T
t0
δu
dt
δτ
(40)
δtf
δu
is the decision variable and
is the decision parameter. Since the minimum of this optimization problem
δτ
δτ
may be negative infinity, to penalize too large decision variable (parameter), we introduce another performance index J t 2 to
formulate a Multi-objective Optimization Problem (MOP) as
min( J t1 , J t 2 )
(41)
s.t.
gxf
tf
∫t
Φo (t f , t ) f u (t )
0
δtf
δu
(t ) d t + ( g x f + gt )
=0
δτ
δτ
f
f
where
Jt 2 =
tf 1 δ u T
δu
1 δtf 2
(
) +∫
( ) K −1
dt
t0 2 δτ
δτ
2kt f δτ
(42)
We use the weighting method to solve the Pareto optimal solution of this MOP, and the resulting performance index is
J t 3 = aJ t1 + bJ t 2
(43)
where a ≥ 0 , b ≥ 0 and a + b = 1 . When a = 1, b = 0 , we get a solution that minimizes J t1 . When a = 0, b = 1 , we get a solution
that minimize J t 2 . Otherwise, we get a compromising solution. For this MOP, obviously in the case of a = 0, b = 1 , the Pareto
optimal solution is that
δtf
δτ
= 0 and
δu
= 0 , and now the value of performance indexes are J t1 = 0 and J t 2 = 0 . For any other
δτ
cases, the compromising solution guarantees that J t1 ≤ 0 .
8
Set a =
1
1
, b = and introduce the Lagrange multiplier π ∈ \ q to adjoin the constraint, we may get the unconstrained
2
2
optimization problem as
Jt 4 =
1
1
1 ⎛
J t1 + J t 2 + π T ⎜ g x f
2
2
2 ⎝
Use the fist-order optimality conditions, i.e.,
tf
∫t
Φo (t f , t ) f u (t )
0
δtf
δtf ⎞
δu
+ gt
(t ) d t + g x f
⎟
δτ ⎠
δτ
δτ
f
(44)
f
∂J t 4
∂J t 4
= 0 and
= 0 , we may get Eqs. (28) and (29). Substitute Eqs. (28)
⎛δu⎞
⎛δtf ⎞
∂⎜
∂⎜
⎟
⎟
⎝ δτ ⎠
⎝ δτ ⎠
and (29) into Eq. (27), we have
gxf
tf
∫t
0
(
)
(
)
Φ (t f , t ) f u K pu + f uT Φ T (t f , t ) g x f T π d t + kt f g x f f + gt f (ϕt + ϕ x T f + L + πq×1T g x f f )
tf
=0
(45)
Further deduction gives
⎛
⎜ gxf
⎝
= − gxf
( ∫ Φ(t , t) f Kf Φ (t , t) d t ) g + k ( g f + g )( g
( ∫ Φ(t , t) f Kp d t ) − k ( g f + g ) (ϕ + ϕ f + L)
tf
t0
u
f
u
T
T
xf
f
T
xf
tf
tf
tf
u
f
t0
u
xf
tf
tf
t
x
xf
f + gt f )T
tf
⎞
⎟π
⎠
(46)
T
tf
Thus, with the definition of M in Eq. (31) and r in Eq. (32), Eq. (30) that determines π is obtained.
Furthermore, Eq. (26) may be reformulated as
δtf
t
δJ
δu
= (ϕt + ϕ x T f + L)
+ ∫ pu T
dt
t
t
δτ
δτ
δτ
δtf
t
= ϕt + ϕ x T f + L + π T ( g x f + gt )
+ ∫ pu + f u T Φ T (t f , t ) g x T π
t
t δτ
f
0
f
(
f
f
)
f
0
f
⎛
− πT ⎜ gx f
⎜
⎝
tf
∫t
Φ (t f , t ) f u
0
(
f
)
T
δu
dt
δτ
(47)
δtf ⎞
δu
⎟
d t + ( g x f + gt )
⎟
δτ
t δτ
f
f
f
⎠
With Eqs. (28) and (29), and because now Eq. (27) holds, then
δJ
= −kt ϕt + ϕ x T f + L + π T ( g x f + gt )
δτ
f
This means
(
f
f
)
2
tf
−∫
tf
t0
(p
u
) (
T
)
+ f u T Φ T (t f , t ) g x f T π K pu + f u T Φ T (t f , t ) g x f T π d t
(48)
δJ
δJ
≤ 0 and
= 0 occurs only when Eqs. (33) and (34) hold.
δτ
δτ
For the optimal value of π , since K may be arbitrary right-dimensional positive-definite matrix and kt f may be arbitrary
positive constant, we consider three case i) K = 1 , kt f = 1 , ii) K = 21 , kt f = 1 , and iii) K = 1 , kt f = 2 , where 1 is the m × m
dimensional identity matrix. By comparing the three cases of substituting the specific values into Eq. (30), we may obtain Eq. (35),
which is irrelevant to the specific value of K and kt f .
■
Regarding the linear equation (30), assuming that the control satisfies the controllability requirement [19], then the solution is
guaranteed. When M is invertible, the parameter π may be calculated as
π = − M −1r
(49)
9
C. Equivalence to the classic optimality conditions
Actually, Eqs. (33) and (34) are the first-order costate-free optimality conditions for Problem 2. We will show that they are
equivalent to the traditional ones with costates [17]. By the adjoining method [14], we may constructed the functional as
(
)
J = ϕ ( x (t f ), t f ) + π T g x (t f ), t + ∫
tf
t0
(L + λ
T
)
( f − x ) dt
(50)
where λ ∈ \ n is the costate variable vector and π ∈ \ q is the Lagrange multiplier parameter. Then the first-order variation may
be derived as
(
δ J = ϕt + π T gt + H
f
f
)
tf
(
)
δ t f + λ(t f ) − ϕ x (t f ) − g x T π δ x (t f ) + ∫
f
tf
t0
( (H
λ
)
− x )T δ λ + ( H x + λ)T δ x + H uTδ u dt
(51)
where H = L + λT f is the Hamiltonian. Through δ J = 0 , we have
λ + H x = λ + Lx + f x T λ = 0
(52)
H u = Lu + f u T λ = 0
(53)
and the transversality conditions
H (t f ) + ϕt f + π T gt f = 0
(54)
λ(t f ) − ϕ x (t f ) − g x f T π = 0
(55)
Theorem 2: For Problem 2, the optimality conditions given by Eqs. (33) and (34) are equivalent to the optimality conditions given
by (52)-(54).
Proof: Define a quantity γ (t ) as
(
tf
)
γ (t ) = ϕ x (t ) + Φo T (t f , t ) g x f T π + ∫ Φo T (σ , t ) Lx (σ ) + ϕtx (σ ) + ϕ xx T (σ ) f (σ ) + f x (σ )T ϕ x (σ ) d σ
t
(56)
Obviously, when t = t f , there is
γ (t f ) = ϕ x (t f ) + g x f T π
(57)
Lu + fu T γ = 0
(58)
Then Eq. (33) is simplified as
Differentiate γ (t ) with respect to t . In the process, we will use the Leibniz rule [20]
d
dt
(∫
a (t )
b (t )
)
h(σ , t ) d σ = h ( a (t ), t )
a (t )
d
d
a (t ) − h ( b(t ), t ) b(t ) + ∫ ht (σ , t ) d σ
(t )
b
dt
dt
(59)
and the property of Φo (σ , t ) [18]
∂Φo (σ , t )
= −Φo (σ , t ) f x (t )
∂t
(60)
Φo (t , t ) = 1
(61)
where 1 is the n × n dimensional identity matrix. Then we have
(
)
(
)
tf
d
γ (t ) = ϕtx + ϕxxT f − f xTΦoT (t f , t ) gx f T π − Lx + ϕtx + ϕxxT f + f xTϕx − f xT ∫ ΦoT (σ , t ) Lx (σ ) + ϕtx (σ ) + ϕxxT (σ ) f (σ ) + f x (σ )T ϕx (σ ) dσ
t
dt
(
tf
(
)
= −Lx − f xT ϕx (t ) + ΦoT (t f , t ) gx f T π + ∫ ΦoT (σ , t ) Lx (σ ) + ϕtx (σ ) + ϕxxT (σ ) f (σ ) + f x (σ )T ϕx (σ ) dσ
t
)
(62)
= −Lx − f xTγ (t )
This means γ (t ) conforms to the same dynamics as the costates λ(t ) in Eq. (52). Furthermore, use Eqs. (47) and (57), we have
10
(
δ J = L + ϕt + π T gt + γ T f
f
)
tf
δtf + ∫
tf
t0
(p
u
)
T
+ f uT Φ T (t f , t ) g x f T π δ u d t
(63)
which hold in the feasible solution domain D o . Compare Eq. (63) with Eq. (51), because δ t f may be arbitrary small, to achieve
the extremal condition, Eqs. (34) and (54) should be same, i.e.
(L +ϕ
tf
+ π T gt f + λT f
) = (L +ϕ + π
tf
t
T
gt f + γ T f
)
(64)
tf
Since Eq. (64) is generally hold for arbitrary g and f , we can conclude that
π=π
(65)
γ (t f ) = λ(t f )
(66)
which implies the sameness of Eq. (57) and Eq. (55). With Eq. (62), and because γ (t ) also conforms to the same boundary
conditions as the costates λ(t ) , the relation that γ (t ) = λ(t ) is established. Therefore Eqs. (58) and (53) are identical.
■
Again, by investigating the optimality condition (33), it is found that the optimal control are related to the future state, and thus
the optimal feedback control law in the analytic form does not exists for the OCPs. From Theorem 2, we actually have got the
explicit analytic expression of the costates λ and the Lagrange multipliers π for the classic treatment in Eq. (50), which formerly
can only be obtained numerically by solving the BVP. After the proof of Theorem 2, now the variables evolving direction using the
VEM is easy to determine.
Theorem 3: Solving the IVP with respect to τ , defined by the variation dynamic evolution equations (8), (28) and (29) from a
feasible initial solution, when τ → +∞ , ( x , u) will satisfy the optimality conditions of Problem 2.
Proof: By Lemma 2 and with Eq. (18) as the Lyapunov functional, we may claim that the minimum solution of Peoblem 2 is an
asymptotically stable solution within the feasibility domain D o for the infinite-dimensional dynamics governed by Eqs. (8), (28)
and (29). From a feasible initial solution, any evolution under these dynamics maintains the feasibility of the variables, and they
also guarantee
δJ
δJ
≤ 0 . The functional J will decrease until
= 0 , which occurs when τ → +∞ due to the asymptotical
δτ
δτ
approach. When
δJ
=0, this determines the optimal conditions, namely, Eqs. (33) and (34).
δτ
■
Presume that we already have a feasible initial solution x (t ) , u (t ) and tf that satisfy Eqs. (19), (20) and (21), Theorem 3
guarantees the infinite-dimensional variation dynamic evolution equations (8), (28) and (29) may be used to obtain the optimal
solution that minimizes Eq. (18).
D. Formulation of EPDE
Use the partial differential operator “ ∂ ” and the differential operator “ d ” to reformulate the variation dynamic evolution
equations, we may get the EPDE and EDE as
t
∂u
⎡
∂ ⎡ x ⎤ ⎢ ∫t0 Φo (t , s ) f u ( s ) ∂τ ( s ) d s
=⎢
∂τ ⎢⎣ u ⎥⎦ ⎢
− K pu + f u T Φ T (t f , t ) g x f T π
⎣
(
)
⎤
⎥
⎥
⎥
⎦
(67)
11
dt f
dτ
(
= − kt f L + ϕt + ϕ x T f + π T ( g x f f + gt f )
Put into this perspective, the definite conditions are t f
τ =0
)
(68)
tf
⎡ x (t ,τ ) ⎤
⎡ x (t ) ⎤
= tf and ⎢
=⎢
⎥
⎥ , where x (t ) and u (t ) are the initial
⎣ u(t ,τ ) ⎦ τ = 0 ⎣ u (t ) ⎦
feasible solution. Eqs. (67) and (68) realize the anticipated variable evolving along the variation time τ as depicted in Fig. 1. The
initial conditions of x (t ,τ ) and u(t ,τ ) at τ = 0 belong to the feasible solution domain and their value at τ = +∞ represents the
optimal solution of the OCP. The right part of the EPDE (67) is also only a vector function of time t . Thus we may apply the
semi-discrete method to discretize it along the normal time dimension and further use ODE integration methods to get the
numerical solution.
As already pointed out previously that Problem 2 contains Problem 1, thus the results developed in this paper are of more general
meaning. See the evolution equations, when the terminal states are free, i.e., g = 0 , they are degraded to the case in Ref. [13].
Moreover, this paper considers the OCPs defined with free terminal time, yet the results obtained are also applicable to the simpler
case with fixed terminal time. For those OCPs, the equation regarding the terminal time t f is not necessary anymore, while the
evolution equations regarding state variable x and control variable u are still similar with M , r in Eq. (30) simplified as
M = gxf
(∫
tf
t0
)
Φ (t f , t ) f u Kf u T Φ T (t f , t ) d t g x f T
r = gxf
(∫
tf
t0
Φ (t f , t ) f u Kpu d t
)
IV. ILLUSTRATIVE EXAMPLES
First a linear example taken from Xie [21] is considered.
Example 1: Consider the following dynamic system
x = Ax + bu
⎡x ⎤
⎡0 1 ⎤
⎡0⎤
, b = ⎢ ⎥ . Find the solution that minimizes the performance index
where x = ⎢ 1 ⎥ , A = ⎢
⎥
⎣0 0⎦
⎣1 ⎦
⎣ x2 ⎦
J=
1 tf 2
u dt
2 ∫t0
with the boundary conditions
⎡1⎤
⎡0⎤
x (t0 ) = ⎢ ⎥ , x (t f ) = ⎢ ⎥
1
⎣⎦
⎣0⎦
where the initial time t0 = 0 and the terminal time t f = 2 are fixed.
In solving this example using the VEM, the EPDE derived is
t A( t − s ) ∂u
⎡
⎤
e
b ( s) d s ⎥
∂τ
∂ ⎡ x ⎤ ⎢ ∫t0
⎥
=⎢
T ⎫
∂τ ⎢⎣ u ⎥⎦ ⎢
⎧
A(t f −t )
⎥
T
−
+
K
u
e
b
π
⎨
⎬⎥
⎢
⎩
⎭
⎣
⎦
(
In particular, for this problem, π is a constant that may be calculated as
)
(69)
(70)
12
⎡ 3.0 ⎤
π = − M −1r = ⎢
⎥
⎣ −2.5 ⎦
tf
where M = K ∫ e
t0
A(t f −t )
(
bb T e
)
A(t f −t ) T
tf
d t and r = K ∫ e
A(t f −t )
t0
bu d t = − Ke
A ( t f − t0 )
x0 . The one-dimensional matrix K was set as
K = 0.1 . The definite conditions of the EPDE, i.e., the initial guess of the states x (t ) and the control u (t ) , were obtained by
numerical integration. To achieve the feasibility, we designed the following control law as
u (t ) = −ωn 2 x1 (t ) − 2ωnξ x2 (t )
where the damp parameter is ξ =0.707 and the frequency parameter is set to be time-varying as ωn = 5t . Using the semi-discrete
method, the time horizon [t0 , t f ] was discretized uniformly with 41 points. Thus, a dynamic system with 123 states was obtained
and the OCP was transformed to a finite-dimensional IVP. The ODE integrator “ode45” in Matlab, with default relative error
tolerance 1×10-3 and default absolute error tolerance 1×10-6, was employed to solve the IVP. For comparison, the analytic solution
by solving the BVP is presented.
xˆ1 = 0.5t 3 − 1.75t 2 + t + 1
xˆ2 = 1.5t 2 − 3.5t + 1
λˆ1 = 3
λˆ = −3t + 3.5
2
uˆ = 3t − 3.5
Figs. 2, 3 and 4 show the evolving process of x1 (t ) , x2 (t ) and u (t ) solutions to the optimal, respectively. At τ = 300s, the
numerical solutions are indistinguishable from the optimal, and this shows the effectiveness of the VEM. Fig. 5 plots the profile of
performance index value against the variation time. It declines rapidly at first and almost reaches the minimum when τ = 40s. Then
it keeps approaching the analytic minimum of 3.25 monotonously. In addition, from Eq. (56), we may compute that
(
γ (t ) = Φo T (t f , t )π = e
)
A(t f −t ) T
0 ⎤ ⎡ 3.0 ⎤ ⎡
⎡ 1
3
⎤
=⎢
π=⎢
⎥
⎢
⎥
⎥
1
t
t
−
2.5
3
3.5
−
−
+
t
f
⎦ ⎣
⎦
⎣
⎦⎣
⎡ λˆ ⎤
It is exactly identical to ⎢ 1 ⎥ , as proved in Theorem 2.
ˆ
⎣⎢ λ2 ⎦⎥
1.4
The analytic solution
Numerical solutions with VEM
1.2
1
x1(t)
0.8
τ = 300s
0.6
0.4
τ = 20.1s
τ = 10.4s
0.2
τ = 6.0s
τ = 2.2s
τ = 0s
0
-0.2
0
0.5
1
1.5
2
t
Fig. 2 The evolution of numerical solutions of x1 to the optimal solution.
13
1
The optimal solution
Numerical solutions with VEM
0.5
0
x2(t)
τ = 300s
-0.5
-1
-1.5
τ = 20.1s
-2
τ = 6.0s
τ = 10.4s
τ = 2.2s
τ = 0s
-2.5
0
0.5
1
1.5
2
t
Fig. 3 The evolution of numerical solutions of
6
x2
to the optimal solution.
The optimal solution
Numerical solutions with VEM
4
u (t)
2
0
τ = 300s
τ = 20.1s
-2
τ = 10.4s
τ = 6.0s
-4
τ = 2.2s
τ = 0s
-6
0
0.5
1
1.5
2
t
Fig. 4 The evolution of numerical solutions of
u
to the optimal solution.
12
Minimum of J
Funcitonal value of numerical solution
10
J
8
6
4
2
0
50
100
150
τ (s)
200
250
300
Fig. 5 The approach to the minimum of performance index.
Now we consider a nonlinear example with free terminal time t f , the Brachistochrone problem [22], which describes the motion
curve of the fastest descending.
Example 2: Consider the following dynamic system
14
x = f ( x, u )
⎡x⎤
⎡ V sin(u ) ⎤
⎢
⎥
where x = ⎢ y ⎥ , f = ⎢⎢ −V cos(u ) ⎥⎥ , g = 10 is the gravity constant. Find the solution that minimizes the performance index
⎢⎣ g cos(u ) ⎦⎥
⎣⎢V ⎦⎥
J = tf
with the boundary conditions
⎡x⎤
⎡0⎤
⎡ x⎤
⎡2⎤
⎢ y⎥
= ⎢⎢ 0 ⎥⎥ , ⎢ ⎥ = ⎢ ⎥
⎢ ⎥
y
−2
⎢⎣V ⎥⎦
⎢⎣ 0 ⎥⎦ ⎣ ⎦ t f ⎣ ⎦
t0 = 0
This example has fixed terminal position boundary conditions and free terminal velocity V (t f ) . Thus there is
⎡1 0 0 ⎤
gxf = ⎢
⎥
⎣0 1 0⎦
In the specific form of the EPDE (67) and the EDE (68), the parameters K and kt f were set to be 0.1 and 0.05, respectively. The
⎡ x (t ,τ ) ⎤
⎢
⎥
definite conditions, i.e., ⎢u (t ,τ ) ⎥
⎢ t f (τ ) ⎥
⎣
⎦
, were obtained from a physical motion along a straight line that connects the initial position to
τ =0
the terminal position, that is
tf = 0.8
x = 2.5t
2
u =
π
4
y = −2.5t 2
V = 5 2t
We also discretized the time horizon [t0 , t f ] uniformly, with 101 points. Thus, a large IVP with 405 states (including the terminal
time) is obtained. We still employed “ode45” in Matlab for the numerical integration. In the integrator setting, the default relative
error tolerance and the absolute error tolerance are 1×10-3 and 1×10-6, respectively. For comparison, we computed the optimal
solution with GPOPS-II [23], a Radau PS method based OCP solver.
Fig. 6 gives the states curve in the x y coordinate plane, showing that the numerical results starting from a straight line approach
the optimal solution over time. The control solutions are plotted in Fig. 7, and the asymptotical approach of the numerical results
are demonstrated. In Fig. 8, the terminal time profile against the variation time τ is plotted. The result of t f declines rapidly at
first and then gradually approaches to the minimum decline time, and it only changes slightly after τ = 40s. At τ = 300s, we
compute that t f = 0.8168s from the VEM, very close to the result of 0.8165s from GPOPS-II. Fig. 9 presents the evolution profiles
⎡ −0.1477 ⎤
of the Lagrange multipliers π . They also approach the optimal value of ⎢
⎥ rapidly.
⎣ 0.0564 ⎦
15
0
The optimal solution
Numerical solutions with VEM
-0.5
τ = 0s
τ = 2.0s
y(t)
τ = 6.5s
τ = 10.9s
τ = 19.6s
-1
τ = 300s
-1.5
-2
0
0.5
1
1.5
2
2.5
x(t)
Fig. 6 The evolution of numerical solutions in the
xy
coordinate plane to the optimal solution.
1.4
1.2
The optimal solution
Numerical solutions with VEM
τ = 19.6s
u (t)
1
τ = 10.9s
τ = 6.5s
0.8
τ = 2.0s
τ = 0s
0.6
0.4
τ = 300s
0.2
0
0
0.2
0.4
0.6
0.8
1
t
Fig. 7. The evolution of numerical solutions of
u
to the optimal solution.
0.9
Minimum decline time
Evolving profile of tf
0.86
f
t (s)
0.88
0.84
0.82
0.8
0
50
100
150
τ (s)
200
250
300
Fig. 8 The evolution profile of t f to the minimum decline time.
16
0.2
π corresponding to x( tf )= 2
0.15
π corresponding to y( tf )= -2
0.1
π
0.05
0
-0.05
-0.1
-0.15
-0.2
0
50
100
150
τ (s)
200
250
Fig. 9 The evolution profiles of the Lagrange multipliers
300
π.
V. CONCLUSION
This paper further developes the compact Variation Evolving Method (VEM) to address the computation of Optimal Control
Problems (OCPs) with terminal constraint. A set of more general evolution equations is derived, and the costate-free optimality
conditions are established for OCPs with prescribed terminal boundary conditions. Especially, from the equivalence proof of the
optimality conditions, even if the costates are completely not considered during the derivation, its analytic expressions related to
the states and the control variables are obtained. The analytic relations between Lagrange multipliers, which adjoin the terminal
constraints, and the original variables are also explicitly got. These results may help us to deepen the understanding towards the
optimal control theory. However, currently the proposed method requires the initial solution to be feasible. This is inflexible,
especially for the problem with terminal constraint, and further studies will be carried out to address this issue.
REFERENCES
[1] H. J. Pesch and M. Plail, “The maximum principle of optimal control: A history of ingenious ideas and missed opportunities,” Control &
Cybernetics, vol. 38, no. 4, pp. 973-995, 2009.
[2] J. T. Betts, “Survey of numerical methods for trajectory optimization,” J. Guid. Control Dynam., vol. 21, no. 2, pp. 193-206, 1998.
[3] Q. Lin, R. Loxton, and K. L. Teo, "The control parameterization method for nonlinear optimal control: a survey," Journal of Industrial and
Management Optimization, vol. 10, no. 1, pp. 275- 309, 2014
[4] C. Hargraves and W. Paris, “Direct trajectory optimization using nonlinear programming and collocation,” J. Guid. Control Dynam., vol. 10,
no. 4, pp. 338-342, 1987.
[5] O. V. Stryk and R. Bulirsch, “Direct and indirect methods for trajectory optimization,” Ann. Oper. Res., vol. 37, no. 1, pp. 357-373, 1992.
[6] H. J. Peng, Q. Gao, Z. G. Wu, and W. X. Zhong, “Symplectic approaches for solving two-point boundary-value problems,” J. Guid. Control
Dynam., vol. 35, no. 2, pp. 653-658, 2012.
[7] A. V. Rao, “A survey of numerical methods for optimal control,” in Proc. AAS/AIAA Astrodynam. Specialist Conf., Pittsburgh, PA, 2009, AAS
Paper 09-334.
[8] D. Garg, M. A. Patterson, W. W. Hager, A. V. Rao, et al, A Unified framework for the numerical solution of optimal control problems using
pseudospectral methods,” Automatica, vol. 46, no. 11, pp. 1843-1851, 2010.
17
[9] I. M. Ross and F. Fahroo, “A perspective on methods for trajectory optimization,” in Proc. AIAA/AAS Astrodynam. Conf., Monterey, CA, 2002,
AIAA Paper No. 2002-4727.
[10] I. M. Ross and F. Fahroo, “Pseudospectral methods for optimal motion planning of differentially flat systems,” IEEE Trans. Autom. Control,
vol. 49, no. 8, pp. 1410-1413, 2004.
[11] S. Zhang, E. M. Yong, W. Q. Qian, and K. F. He, “A Variation Evolving Method for Optimal Control ArXiv,” arXiv: 1703.10263 [cs.SY].
[12] S. Zhang, and W. Q. Qian, “Computation of time-optimal control problem with variation evolution principle,” arXiv:1711.02998 [cs.SY].
[13] S. Zhang, and K. F. He, “Variation evolving for optimal control computation, a compact way,” arXiv:1709.02242 [cs.SY].
[14] R. F. Hartl, S. P. Sethi, and R. G. Vickson, “A survey of the maximum principles for optimal control problems with state constraint,” SIAM
Rev., vol. 37, no. 2, pp. 181-218, 1995.
[15] H. K. Khalil, Nonlinear Systems. New Jersey, USA: Prentice Hall, 2002, pp. 111-181.
[16] H. X. Zhang and M. Y. Shen. Computational Fluid Dynamics — Fundamentals and Applications of Finite Difference Methods. Beijing, China:
National Defense Industry Press, 2003, pp. 76-78.
[17] A. E. Bryson and Y. C. Ho, Applied Optimal Control: Optimization, Estimation, and Control. Washington, DC, USA: Hemisphere, 1975, pp.
42-125.
[18] D.Z. Zhen. Linear System Theory. Beijing, China: Tsinghua University Press, 2002, pp. 85-134.
[19] D. G. Hull, Optimal Control Theory for Applications, New York, USA: Springer, 2003, pp. 89-71.
[20] H. Wang, J. S. Luo, and J. M. Zhu. Advanced Mathematics. Changsha, China: National University of Defense Technology Press, 2000, pp.
203-210.
[21] X. S. Xie, Optimal Control Theory and Application. Beijing, China: Tsinghua Univ. Press, 1986, pp. 167-215.
[22] Sussmann, H. J., & Willems, J. C. (2000). The Brachistochrone problem and modern control theory. In Anzaldo-Meneses, A., Bonnard, B.,
Gauthier, J. P., & Monroy-Perez, F. (Eds.), Contemporary Trends in Nonlinear Geometric Control Theory and its Applications, Singapore:
World Scientific.
[23] M. A. Patterson and A. V. Rao, “GPOPS-II: AMATLAB software for solving multiple-phase optimal control problems using hp-adaptive
Gaussian quadrature collocation methods and sparse nonlinear programming,” ACM Trans. Math. Software, vol. 41, no. 1, pp. 1-37, 2014.
| 3cs.SY
|
A Type System Theory for Higher-Order Intensional Logic
Support for Variable Bindings in Hybrid Intensional-Imperative
Programs in GIPSY
Joey Paquet
Concordia University, Montreal, Canada
[email protected]
arXiv:0906.3919v1 [cs.LO] 22 Jun 2009
Serguei A. Mokhov
Concordia University, Montreal, Canada
[email protected]
Abstract
We describe a type system for a platform called the General Intensional Programming System
(GIPSY), designed to support intensional programming languages built upon intensional logic and
their imperative counter-parts for the intensional execution model. In GIPSY, the type system glues
the static and dynamic typing between intensional and imperative languages in its compiler and runtime environments to support the intensional evaluation of expressions written in various dialects of
the intensional programming language Lucid. The intensionality makes expressions to explicitly take
into the account a multidimensional context of evaluation with the context being a first-class value
that serves a number of applications that need the notion of context to proceed. We describe and
discuss the properties of such a type system and the related type theory as well as particularities of
the semantics, design and implementation of the GIPSY type system.
Keywords: Intensional Programming, Imperative Programming, Type System, Type Theory, General Intensional Programming System (GIPSY), Context
1
Introduction
The GIPSY is an ongoing effort for the development of a flexible and adaptable multi-lingual programming language development framework aimed at the investigation on the Lucid family of intensional
programming languages [2, 2, 22, 1, 12, 11, 17, 18]. Using this platform, programs written in many
flavors of Lucid can be compiled and executed. The framework approach adopted is aimed at providing the possibility of easily developing compiler components for other languages of intensional nature,
and to execute them on a language-independent run-time system. Lucid being a functional “data-flow”
language, its programs can be executed in a distributed processing environment. However, the standard
Lucid algebra (i.e. types and operators) is extremely fine-grained and can hardly benefit from parallel
evaluation of their operands. Adding granularity to the data elements manipulated by Lucid inevitably
comes through the addition of coarser-grained data types and their corresponding operators. Lucid semantics being defined as typeless, a solution to this problem consists in adding a hybrid counterpart to
Lucid to allow an external language to define an algebra of coarser-grained types and operators. In turn,
this solution raises the need for an elaborated type system to bridge the implicit types of Lucid semantics
with the explicit types and operators (i.e. functions) defined in the hybrid counterpart language. This
paper presents such a type system that is used in the GIPSY at compile time for static type checking, as
well as at run-time for dynamic type checking.
After our problem statement and short presentation of our proposed solution, this section presents
a brief introduction to the GIPSY project [14, 6, 13, 21, 10, 25, 16], and the beginnings of the GIPSY
Type System [10] in the GIPSY to support hybrid and object-oriented intensional programming [9, 24],
and Lucx’s context type extension known as context calculus [20, 15, 19] for contexts to act as firstclass values. Then, we present the complete GIPSY type system as used by the compiler (General
Intensional Programming Compiler – GIPC) and the run-time system (General Eduction Engine – GEE)
when producing and executing a binary GIPSY program (called General Eduction Engine Resources –
GEER) respectively.
1
A Type System for HOIL Support in GIPSY
1.1
Mokhov and Paquet
Problem Statement and Proposed Solution
Problem Statement Data types are implicit in Lucid (as well as in its dialects or many functional languages). As such, the type declarations never appear in the Lucid programs at the syntactical level. The
data type of a value is inferred from the result of evaluation of an expression. In most imperative languages, like Java, C++ and the like, the types are explicit and the programmers must declare the types of
the variables, function parameters and return values before they are used in evaluation. In the GIPSY, we
want to allow any Lucid dialect to be able to uniformly invoke functions/methods written in imperative
languages and the other way around and perform semantic analysis in the form of type assignment and
checking statically at compile time or dynamically at run time, perform any type conversion if needed,
and evaluate such hybrid expressions. At the same time, we need to allow a programmer to specify, or
declare, the types of variables, parameters, and return values for both intensional and imperative functions as a binding contract between inter-language invocations despite the fact that Lucid is not explicitly
typed. Thus, we need a general type system, well designed and implemented to support such scenarios.
Proposed Solution The uniqueness of the type system presented here is that it is above a specific programming language model of either the Lucid family of languages or imperative languages. It is designed
to bridge programming language paradigms, the two most general of them would be the intensional and
imperative paradigms. GIPSY has a framework designed to support a common run-time environment
and co-existence of the intensional and imperative languages. Thus, the type system is that of a generic
GIPSY program that can include code segments written in a theoretically arbitrary number of intensional
and imperative dialects supported by the system vs. being a type system for a specific language. What
follows is the details of the proposed solution and the specification of the type system.
1.2
Introduction to the GIPSY Type System
The introduction of JLucid, Objective Lucid, and the General Imperative Compiler Framework (GICF) [10,
9, 8, 4] prompted the development of the GIPSY Type System as implicitly understood by the Lucid language and its incarnation within the GIPSY to handle types in a more general manner as a glue between
the imperative and intensional languages within the system. Further evolution of different Lucid variants such as Lucx introducing contexts as first-class values and JOOIP (Java Object-Oriented Intensional
Programming) highlighted the need of the further development of the type system to accommodate the
more general properties of the intensional and hybrid languages.
1.2.1
Matching Lucid and Java Data Types
Here we present a case of interaction between Lucid and Java. Allowing Lucid to call Java methods
brings a set of issues related to the data types, especially when it comes to type checks between Lucid
and Java parts of a hybrid program. This is pertinent when Lucid variables or expressions are used as
parameters to Java methods and when a Java method returns a result to be assigned to a Lucid variable
or used in an intensional expression. The sets of types in both cases are not exactly the same. The basic
set of Lucid data types as defined by Grogono [3] is int, bool, double, string, and dimension.
Lucid’s int is of the same size as Java’s long. GIPSY and Java double, boolean, and String are
equivalent. Lucid string and Java String are simply mapped internally through StringBuffer; thus,
one can think of the Lucid string as a reference when evaluated in the intensional program. Based on
this fact, the lengths of a Lucid string and Java String are the same. Java String is also an object
in Java; however, at this point, a Lucid program has no direct access to any String’s properties (though
internally we do and we may expose it later to the programmers). We also distinguish the float data
2
A Type System for HOIL Support in GIPSY
Mokhov and Paquet
Table 1: Matching data types between Lucid and Java.
Return Types of Java Methods
Types of Lucid Expressions
Internal GIPSY Types
int, byte, long
float
double
boolean
char
String
Method
Method
[]
Object
Object
void
int, dimension
float
double
bool
char
string, dimension
function
operator
[]
class
URL
bool::true
GIPSYInteger
GIPSYFloat
GIPSYDouble
GIPSYBoolean
GIPSYCharacter
GIPSYString
GIPSYFunction
GIPSYOperator
GIPSYArray
GIPSYObject
GIPSYEmbed
GIPSYVoid
Parameter Types Used in Lucid
Corresponding Java Types
Internal GIPSY Types
string
float
double
int
dimension
bool
class
URL
[]
operator
function
String
float
double
int
int, String
boolean
Object
Object
[]
Method
Method
GIPSYString
GIPSYFloat
GIPSYDouble
GIPSYInteger
Dimension
GIPSYBoolean
GIPSYObject
GIPSYEmbed
GIPSYArray
GIPSYOperator
GIPSYFunction
type for single-precision floating point operations. The dimension index type is said to be an integer
or string (as far as its dimension tag values are concerned), but might be of other types eventually, as
discussed in [20]. Therefore, we perform data type matching as presented in Table 1. Additionally, we
allow void Java return type which will always be matched to a Boolean expression true in Lucid as an
expression has to always evaluate to something. As for now our types mapping and restrictions are as
per Table 1. This is the mapping table for the Java-to-IPL-to-Java type adapter. Such a table would exist
for mapping between any imperative-to-intensional language and back, e.g. the C++-to-IPL-to-C++ type
adapter.
2
Simple Theory of GIPSY Types
Our simple theory of the GIPSY types (STGT) is based on the “Simple theory of types” (STT) by
Mendelson [7]. The theoretical and practical considerations are described in the sections that follows.
The STT partitions the qualification domain into an ascending hierarchy of types with every individual
value assigned a type. The type assignment is dynamic for the intensional dialects as the resulting
type of a value in an intensional expression may not be known at compile time. The assignment of
the types of constant literals is done at compile-time, however. In the hybrid system, which is mostly
statically typed at the code-segment boundaries, the type assignment also occurs at compile-time. On
the boundary between the intensional programming languages (IPLs) and imperative languages, prior to
3
A Type System for HOIL Support in GIPSY
Mokhov and Paquet
an imperative procedure being invoked, the type assignment to the procedure’s parameters from IPL’s
expression is computed dynamically and matched against a type mapping table similar to that of Table 1.
Subsequently, the when the procedure call returns back to the IPL, the type of the imperative expression
is matched back through the table and assigned to the intensional expression that expects it. The same
table is used when the call is made by the procedure to the IPL and back, but in the reverse order.
Further in STT, all quantified variables range over only one type making the first-order logic applicable as the underlying logic for the theory. This also means the all elements in the domain and all
co-domains are of the same type. The STT states there is an atomic type that has no member elements
in it, and the members of the second-high from the basic atomic type. Each type has a next higher type
similarly to succ in Peano arithmetic and the next operator in Lucid. This is also consistent to describe
the composite types, such as arrays and objects as they can be recursively decomposed (or “flattened”,
see [10]) into the primitive types to which the STT applies.
Symbolically, the STT uses primed and unprimed variables and the infix set notation of ∈. The
formulas Φ(x) rely on the fact that the unprimed variables are all of the same type. This is similar to the
notion of a Lucid stream with the point-wise elements of the stream having the same type. The primed
variables (x0 ) in STT range over the next higher type. There two atomic formulas in STT of the form of
identity, x = y, and set membership, y ∈ x0 .
The STT defines the four basic axioms for the variables and the types they can range over: Identity,
Extensionality, Comprehension, and Infinity. We add the Intensionality on as the fifth axiom. The variables in the definition of the Identity relationship and in the Extensionality and Comprehension axioms
typically range over the elements of one of the two nearby types. In the set membership, only the unprimed variables that range over the lower type in the hierarchy can appear on the left of ∈; conversely,
the primed ones that range over higher types can only appear on the right of ∈. The axioms are defined
as:
1. Identity: x = y ↔ ∀z0 [x ∈ z0 ↔ y ∈ z0 ]
2. Extensionality: ∀x[x ∈ y0 ↔ x ∈ z0 ] → y0 = z0
3. Comprehension: ∃z0 ∀x[x ∈ z0 ↔ Φ(x)]. This covers objects and arrays as any collection of elements here may form an object of the next, higher type. The STT states the comprehension axiom
is schematic with respect to Φ(x) and the types.
4. Infinity: ∀x, y[x 6= y → [xRy yRx]]. There exists a non-empty binary relation R over the elements
of the atomic type that is transitive, irreflexive, and strongly connected.
W
5. Intensionality: the intensional types and operators are based on the intensional logic and context
calculus. These are extensively described in the works [15, 20, 23, 12] and related. This present
type system accommodates the two in a common hybrid execution environment of the GIPSY. A
context c is a finite subset of the relation: c ⊂ {(d, x)|d ∈ DIM ∧ x ∈ T }, where DIM is the set of
all possible dimensions, and T is the set of all possible tags.
2.1
Types of Types
Types of types are generally referred to as kinds. Kinds provide categorization to the types of similar
nature. While some type systems provide kinds as first class entities available to programmers, in GIPSY
we do not expose this functionality in our type system at this point. However, at the implementation level
there are provisions to do so that we may later decide to expose for the use of programmers. Internally,
we define several broad kinds of types, presented the the sections that follow.
4
A Type System for HOIL Support in GIPSY
2.1.1
Mokhov and Paquet
Numeric Kind
The primitive types under this category are numerical values, which are represented by GIPSYInteger,
GIPSYFloat, and GIPSYDouble. They provide implementation of the common arithmetic operators,
such as addition, multiplication and so on, as well as logical comparison operators of ordering and
equality. Thus, for a numerical type T , the following common operators are provided. The resulting type
of any arithmetic operator is the largest of the two operands in terms of length (the range of double of
length say k covers the range of int with the length say m and if both appear as arguments to the operator,
then the resulting value’s type is that of without loss of information, i.e. largest in length double). The
result of the logical comparison operators is always Boolean B regardless the length of the left-hand-side
and right-hand-side numerical types.
1. Tmax : T1 ≥ T2 → T1 | T1 < T2 → T2
2. Tmultiply : Tk × Tm → Tmax(k,m)
3. Tdivide : Tk /Tm → Tmax(k,m)
4. Tadd : Tk + Tm → Tmax(k,m)
5. Tsubtract : Tk − Tm → Tmax(k,m)
6. Tmod : Tk %Tm → Tmax(k,m)
7. Tpow : Tk ˆ Tm → Tmax(k,m)
8. T> : T > T → B
9. T< : T < T → B
10. T≥ : T ≥ T → B
11. T≤ : T ≤ T → B
12. T= : T = T → B
A generalized implementation of the arithmetic operators is done by realization of the interface
called IArithmeticOperatorsProvider and its concrete implementation developed in the general
delegate class GenericArithmeticOperatorsDelegate. This design and implementation allow not
only further exposure of kinds-as-first-class values later on after several iterations of refinement, but
also will allow operator and type overloading or replacement of type handling implementation altogether if some researchers wish to do so. The implementation of the engine, GEE, is thus changed,
to only refer to the interface type implementation when dealing with these operators. Equivalently for
the logic comparison operators we have the ILogicComparisonOperatorsProvider class and the
corresponding GenericLogicComparisonOperatorsDelegate class. The latter relies on the comparator implemented for the numerical kind, such as NumericComparator. Using comparators (i.e.
classes that implement the standard Comparator interface) allows Java to use and to optimize buildin sorting and searching algorithms for collections of user-defined types. In our case, the class called
GenericLogicComparisonOperatorsDelegate is the implementation of the delegate class that also
relies on it. The example for the numeric types for the described design is in Figure 1.
It is important to mention, that grouping of the numeric kind of integers and floating-point numbers
does not violate the IEEE 754 standard [5], as these kinds implementation-wise wrap the corresponding
Java’s types (which are also grouped under numeric kind) and their semantics including the implementation of IEEE 754 by Java in accordance with the Java Language Specification.
5
A Type System for HOIL Support in GIPSY
Mokhov and Paquet
Figure 1: Example of Provider Interfaces and Comparators
2.1.2
Logic Kind
Similarly to numeric types, the primitive type GIPSYBoolean fits the category of the types that can be
used in the Boolean expressions. The operations the type provides expect the arguments to be of the
same type – Boolean. The following set of operators on the logic type B we provide in the GIPSY type
system:
1. Band : B B → B
V
2. Bor : B B → B
W
3. Bnot : ¬B → B
4. Bxor : B
L
B→B
5. Bnand : ¬(B B) → B
V
6. Bnor : ¬(B B) → B
W
L
Note that the logical XOR operator (denoted as ) is different from the corresponding bitwise operator in Section 2.1.3 similarly to as the bitwise vs. logical are for AND and OR. Again, similarly to the generalized implementation of arithmetic operators, logic operator providers implement the
ILogicOperatorsProvider interface, with the most general implementation of it in the delegate class
GenericLogicOperatorsDelegate.
2.1.3
Bitwise Kind
Bitwise kind of types covers all the types that can support bitwise operations on the the entire bit length
of a particular type T . Types in this category include the numerical and logic kinds described earlier in
6
A Type System for HOIL Support in GIPSY
Mokhov and Paquet
Section 2.1.2 and Section 2.1.1. The parameters on both sides of the operators and the resulting type are
always the same. There are no implicit compatible type casts performed unlike for the numeric kind.
1. Tbit−and : T &T → T
2. Tbit−or : T |T → T
3. Tbit−not : !T → T
4. Tbit−xor : T
N
T →T
5. Tbit−nand : !(T &T ) → T
6. Tbit−nor : !(T |T ) → T
The generalized implementation of this kind’s operators is done through the interface that we called
IBitwiseOperatorsProvider, which is generically implemented in the corresponding delegate class
GenericBitwiseOperatorsDelegate.
2.1.4
Composite Kind
As the name suggests, the composite kind types consist of compositions of other types, possibly basic types. The examples of this kind are arrays, structs, and abstract data types and their realization
such as objects and collections. In the studied type system these are GIPSYObject, GIPSYArray, and
GIPSYEmbed. This kind is characterized by the constructors, dot operator to decide membership as well
as to invoke member methods and define equality. The design of of these types, just like the entire
type system, adheres to the Composite design pattern. The most prominent examples of this kind are in
Figure 2, including GIPSYContext which is composed of Dimensions and indirectly of the TagSets.
Figure 2: Composite Types.
2.1.5
Intensional Kind
The intentional types kind primarily has to do with encapsulation dimensionality, context information and
their operators. These are represented by the types GIPSYContext, Dimension, and TagSet1 types. The
1 Not
mentioned here; please refer to [19, 15].
7
A Type System for HOIL Support in GIPSY
Mokhov and Paquet
common operators on these types include the context switching and querying operators @ and # as well
as context calculus operators. Additional operators can be included depending on the intensional dialect
used, but the mentioned operators are said to be the baseline operators that any intensional language
can be translated to use. Implementation-wise there is a IContextSetOperatorsProvider interface
and its general implementation in GenericContextSetOperatorsDelegate. The context calculus
operators on simple contexts include standard set operators, such as union, difference, intersection, and
Lucx-specific isSubContext, projection, hiding, and override [19, 15].
2.1.6
Function Kind
The function types represent either “functional” functions, imperative procedures, and binary and unary
operators that, themselves, can be passed as parameters or returned as results. In our type system these
are represented by GIPSYFunction, GIPSYOperator, and GIPSYEmbed. The common operators on this
include equality and evaluation.
3
Describing Some GIPSY Types’ Properties
To show the overview of the properties of GIPSY types to get a better and more complete understanding
of the spectrum of their behavior we cover them in light of description and comparison to existential,
union, intersection, and linear types.
3.1
Existential Types
All of the presented GIPSY types are existential types because they represent and encapsulate modules
and abstract data types and separate their implementation from the public interface specified by the
abstract GIPSYType. These are defined as follows:
T = ∃GIPSYType{Object a; Object getEnclosedTypeObject(); Object getValue(); }
The above is implemented in the following manner, e.g.:
boolT {Boolean a; Object getEnclosedTypeObject(); Boolean getValue(); }
intT {Long a; Object getEnclosedTypeObject(); Long getValue(); }
doubleT {Double a; Object getEnclosedTypeObject(); Double getValue(); }
... and so on.
All these correspond to be subtypes of the the more abstract and general existential type T . Assuming
the value t ∈ T , then t.getEnclosedTypeObject() and t.getValue() are well typed regardless the what
the GIPSYType may be.
3.2
Union Types
A union of two types produces another type with the valid range of values that is valid in either of
the two; however, the operators defined on the union types must be those that are valid for both of the
types to remain type-safe. A classical example of that is in C or C++ the range for the signed char is
−128 . . . 127 and the range of the unsigned char is 0 . . . 255, thus:
signed char ∪ unsigned char = −128 . . . 255
In C and C++ there is a union type that roughly corresponds to this notion, but it does not enforce the
operations that are possible on the union type that must be possible in both left and right types of the
8
A Type System for HOIL Support in GIPSY
Mokhov and Paquet
uniting operator. In the class hierarchy, such as in GIPSY, the union type among the type and its parent
is the parent class; thus, in our specific type system the following holds:
∀T ∈ G : T ∪ GIPSYType = GIPSYType
GIPSYArray ∪ GIPSYObject = GIPSYObject
GIPSYVoid ∪ GIPSYBoolean = GIPSYBoolean
GIPSYOperator ∪ GIPSYFunction = GIPSYFunction
where T is any concrete GIPSY type and G is a collection of types in the GIPSY type system we are
describing. Equivalently, the union of the two sibling types is their common parent class in the inheritance hierarchy. Interestingly enough, while we do not explicitly expose kinds of types, we still are able
to have the following union type relationships defined based on the kind of operators they provide as
siblings under a common interface:
∀T
∀T
∀T
∀T
∀T
∀T
∈ A : T ∪ IArithmeticOperatorProvider = IArithmeticOperatorProvider
∈ L : T ∪ ILogicOperatorProvider = ILogicOperatorProvider
∈ B : T ∪ IBitwiseOperatorProvider = IBitwiseOperatorProvider
∈ C : T ∪ IContextOperatorProvider = IContextOperatorProvider
∈ D : T ∪ ICompositeOperatorProvider = ICompositeOperatorProvider
∈ F : T ∪ IFunctionOperatorProvider = IFunctionOperatorProvider
where T is any concrete GIPSY type and A is a collection of types that provide arithmetic operators,
L – logic operators providers, B – bitwise operators providers, C – context operators providers, D –
composite operator providers, and F – function operator providers. Thus:
{GIPSYInteger, GIPSYFloat, GIPSYDouble} ∈ A
{GIPSYBoolean} ∈ L
{GIPSYInteger, GIPSYFloat, GIPSYDouble, GIPSYBoolean} ∈ B
{GIPSYContext, Dimension} ∈ C
{GIPSYObject, GIPSYArray, GIPSYEmbed}, GIPSYString} ∈ D
{GIPSYFunction, GIPSYOperator, GIPSYEmbed} ∈ F
Another particularity of the GIPSY type system is the union of the string and integer types under dimension:
GIPSYInteger ∪ GIPSYString = Dimension
and this is because in our dimension tag values we allow them to be either integers or strings. While not
a very common union in the majority of type system, they do share a common set of tag set operators
defined in [20] for ordered finite tag sets (e.g. next(), etc.).
3.3
Intersection Types
Intersection type of given two types is a range where the sets of valid values overlap. Such types are
safe to pass to methods and functions that expect either of the types as the intersection types are more
restrictive and compatible in both original types. A classical example of an intersection type if it were
implemented in C or C++ would be:
signed char ∩ unsigned char = 0 . . . 127
The intersection types are also useful in describing the overloaded functions. Sometimes the are called
as refinement types. In the class hierarchy, the intersection between the parent and child classes is the
most derived type, and the intersection of the sibling classes is empty. While the functionality offered
9
A Type System for HOIL Support in GIPSY
Mokhov and Paquet
by the intersection types is promising, it is not currently explicitly or implicitly considered in the GIPSY
type system, but planned for the future work.
3.4
Linear Types
Linear (or “uniqueness”) types are based on linear logic. The main idea of these types is that values
assigned to them have one and only one reference to them throughout. These types are useful to describe
immutable values like strings or hybrid intensional-imperative objects (see [24] for details). These are
useful because most operations on such an object “destroys” it and creates a similar object with the
new values, and, therefore, can be optimized in the implementation for the in-place mutation. Implicit
examples of such types in the GIPSY type system are GIPSYString that internally relies on Java’s
StringBuffer that does something very similar as well as the immutable GIPSYObject is in JOOIP
[24] and immutable GIPSYFunction. Since we either copy or retain the values in the warehouse, and,
therefore, one does not violate referential transparency or create side effects in, and at the same time be
more efficient as there is no need to worry about synchronization overhead.
4
Conclusion
Through a series of discussions, specification, design, and implementation details we presented a type
system used by the GIPSY for static and dynamic type checking and evaluation of intensional and hybrid
intensional-imperative languages. We highlighted the particularities of the system that does not attribute
to a particular specific language as traditionally most type systems do, but to an entire set of languages
and hybrid paradigms that are linked through the type system. We pointed out some of the limitations
of the type system and the projected work to remedy those limitations. Overall, this is a necessary
contribution to GIPSY-like systems to have a homogeneous environment to statically and dynamically
type-check intensional and hybrid programs.
4.1
Future Work
The type system described in this paper has been implemented in the GIPSY. However, due to recent
changes including the introduction of contexts as first-class values in Generic Lucid, as well as the
development of the fully-integrated hybrid language JOOIP [24], our implementation still needs thorough
testing using complex program examples testing the limits of the type system. Additional endeavours,
noted in the previous sections of this paper, include:
• Exposing more operators for composite types to Lucid code segments.
• Allowing custom user-defined types and extension of the existing operators and operator overloading.
• Expose kinds as first class entities, allowing programs more explicit manipulation of types.
• Consider adding intersection types for more flexibility in the future development of the type system, allowing more type casting possibilities at the programming level.
4.2
Acknowledgments
This work was funded by NSERC and the Faculty of Computer Science and Engineering, Concordia
University.
10
A Type System for HOIL Support in GIPSY
Mokhov and Paquet
References
[1] E. Ashcroft, A. Faustini, R. Jagannathan, and W. Wadge. Multidimensional, Declarative Programming.
Oxford University Press, London, 1995.
[2] E. A. Ashcroft and W. W. Wadge. Erratum: Lucid - a formal system for writing and proving programs. SIAM
J. Comput., 6((1):200), 1977.
[3] P. Grogono. GIPC increments. Technical report, Department of Computer Science and Software Engineering,
Concordia University, Montreal, Canada, Apr. 2002.
[4] P. Grogono, S. Mokhov, and J. Paquet. Towards JLucid, Lucid with embedded Java functions in the GIPSY. In
Proceedings of the 2005 International Conference on Programming Languages and Compilers (PLC 2005),
Las Vegas, USA, pages 15–21. CSREA Press, June 2005.
[5] IEEE. 754-2008: IEEE standard for floating-point arithmetic. [online], Aug. 2008. http://ieeexplore.
ieee.org/servlet/opac?punumber=4610933.
[6] B. Lu. Developing the Distributed Component of a Framework for Processing Intensional Programming
Languages. PhD thesis, Department of Computer Science and Software Engineering, Concordia University,
Montreal, Canada, Mar. 2004.
[7] E. Mendelson. Introduction to Mathematical Logic. Chapman & Hall, 4 edition, 1997.
[8] S. Mokhov and J. Paquet. General imperative compiler framework within the GIPSY. In Proceedings of the
2005 International Conference on Programming Languages and Compilers (PLC 2005), Las Vegas, USA,
pages 36–42. CSREA Press, June 2005.
[9] S. Mokhov and J. Paquet. Objective Lucid – first step in object-oriented intensional programming in the
GIPSY. In Proceedings of the 2005 International Conference on Programming Languages and Compilers
(PLC 2005), Las Vegas, USA, pages 22–28. CSREA Press, June 2005.
[10] S. A. Mokhov. Towards hybrid intensional programming with JLucid, Objective Lucid, and General Imperative Compiler Framework in the GIPSY. Master’s thesis, Department of Computer Science and Software
Engineering, Concordia University, Montreal, Canada, Oct. 2005. ISBN 0494102934.
[11] C. B. Ostrum. The Luthid 1.0 Manual. Department of Computer Science, University of Waterloo, Ontario,
Canada, 1981.
[12] J. Paquet. Scientific Intensional Programming. PhD thesis, Department of Computer Science, Laval University, Sainte-Foy, Canada, 1999.
[13] J. Paquet, P. Grogono, and A. H. Wu. Towards a framework for the general intensional programming compiler in the GIPSY. In Proceedings of the 19th Annual ACM Conference on Object-Oriented Programming,
Systems, Languages, and Applications (OOPSLA 2004), Vancouver, Canada, Oct. 2004. ACM.
[14] J. Paquet and P. Kropf. The GIPSY architecture. In Proceedings of Distributed Computing on the Web,
Quebec City, Canada, 2000.
[15] J. Paquet, S. A. Mokhov, and X. Tong. Design and implementation of context calculus in the GIPSY environment. In Proceedings of the 32nd Annual IEEE International Computer Software and Applications
Conference (COMPSAC), pages 1278–1283, Turku, Finland, July 2008. IEEE Computer Society.
[16] J. Paquet and A. H. Wu. GIPSY – a platform for the investigation on intensional programming languages. In
Proceedings of the 2005 International Conference on Programming Languages and Compilers (PLC 2005),
pages 8–14, Las Vegas, USA, June 2005. CSREA Press.
[17] J. Plaice, B. Mancilla, G. Ditu, and W. W. Wadge. Sequential demand-driven evaluation of eager TransLucid.
In Proceedings of the 32nd Annual IEEE International Computer Software and Applications Conference
(COMPSAC), pages 1266–1271, Turku, Finland, July 2008. IEEE Computer Society.
[18] T. Rahilly and J. Plaice. A multithreaded implementation for TransLucid. In Proceedings of the 32nd Annual
IEEE International Computer Software and Applications Conference (COMPSAC), pages 1272–1277, Turku,
Finland, July 2008. IEEE Computer Society.
[19] X. Tong. Design and implementation of context calculus in the GIPSY. Master’s thesis, Department of
Computer Science and Software Engineering, Concordia University, Montreal, Canada, Apr. 2008.
[20] X. Tong, J. Paquet, and S. A. Mokhov. Context Calculus in the GIPSY. Unpublished, 2007.
[21] E. Vassev and J. Paquet. A generic framework for migrating demands in the GIPSY’s demand-driven execu-
11
A Type System for HOIL Support in GIPSY
[22]
[23]
[24]
[25]
Mokhov and Paquet
tion engine. In Proceedings of the 2005 International Conference on Programming Languages and Compilers
(PLC 2005), pages 29–35, Las Vegas, USA, June 2005. CSREA Press.
W. Wadge and E. Ashcroft. Lucid, the Dataflow Programming Language. Academic Press, London, 1985.
K. Wan. Lucx: Lucid Enriched with Context. PhD thesis, Department of Computer Science and Software
Engineering, Concordia University, Montreal, Canada, 2006.
A. Wu, J. Paquet, and S. A. Mokhov. Object-Oriented Intensional Programming: A New Concept in ObjectOriented and Intensional Programming Domains. Unpublished, 2007.
A. H. Wu and J. Paquet. Object-oriented intensional programming in the GIPSY: Preliminary investigations. In Proceedings of the 2005 International Conference on Programming Languages and Compilers
(PLC 2005), pages 43–47, Las Vegas, USA, June 2005. CSREA Press.
12
| 6cs.PL
|
arXiv:1612.03735v1 [cs.DS] 4 Dec 2016
A Note on Testing Intersection of Convex Sets
in Sublinear Time
Israela Solomon
February 27, 2018
We present a simple sublinear time algorithm for testing the following geometric property. Let P1 , ..., Pn be n convex sets in Rd (n ≫ d), such as polytopes, balls, etc. We assume that the complexity of each set depends only on d
(and not on the number of sets n). We test the property that there exists a
common point in all sets, i.e. that their intersection is nonempty. Our goal
is to distinguish between the case where the intersection is nonempty, and the
case where even after removing many of the sets the intersection is empty. In
particular, our algorithm returns PASS if all of the n sets intersect, and returns
α
n sets, for any
FAIL with probability at least 1 − ǫ if no point belongs to d+1
given 0 < α, ǫ < 1.
We call a collection of q convex sets Pi1 , ..., Piq a ’q-tuple’ (the order of
elements in the tuple is not important).
The motivation for our algorithm comes from the following well known theorem by Eduard Helly [1][4][3].
Helly’s Theorem. Let X1 , ..., Xn be a finite collection of convex sets in Rd ,
with n > d. The intersection of every d + 1 of these sets is nonempty, if and
only if the whole collection has a nonempty intersection.
Helly’s theorem inspires us to consider the following algorithm.
Algorithm 1: Tester for nonempty intersection
for i = 1, 2, ..., t do
Pick (d + 1)-tuple Pi1 , Pi2 , ..., Pid+1 with uniform distribution
if ∩d+1
j=1 Pij = ∅ then
return FAIL
end
end
return PASS
Theorem 1. For t = logα ǫ, the algorithm returns PASS if the intersection
of all sets is nonempty, and returns FAIL with probability at least 1 − ǫ if the
α
n sets is empty. The running time of the algorithm
intersection of every d+1
does not depend on n.
1
α
In order to prove the theorem, we should show that if any d+1
n sets have
an empty intersection, then the probability of picking a (d + 1)-tuple whose
intersection is empty and returning FAIL is high. For that purpose, we use the
following theorem by Katchalski and Liu [2].
Fractional Helly Theorem. Let
q > d > 0 be integers and let F be a family
of n convex sets in Rd with α nq of the q-tuples intersecting. Then there is a
1
q−d
α
.
point in at least βn of the sets, where β =
(dq )
Setting q = d + 1, we obtain the following corollary.
Corollary 1. If the maximal number
of sets from P1 , ..., Pn that intersect is
n
α
less than d+1
n, then less than α d+1
of the (d + 1)-tuples intersect. Namely,
n
at least (1 − α) d+1 of the (d + 1)-tuples have empty intersection.
Now we are ready to prove the correctness of the algorithm.
Proof of Theorem 1. If the intersection of all of the n sets is nonempty, then the
intersection of every d + 1 sets is nonempty, and the algorithm returns PASS.
α
Assume that the intersection of every d+1
n sets is empty. Then, by Corol
n
lary 1, there are less than α d+1 choices of d+1 sets such that their intersection
is nonempty, which means that the probability to pick such sets (and not returning FAIL) in each round is smaller than α. Therefore, the probability that
our algorithm returns FAIL is at least 1 − αt . Setting t = logα ǫ, the probability
to return FAIL is at least 1 − ǫ.
The running time of the algorithm is logα ǫ times the time it takes to compute
the intersection of d + 1 convex sets. Note that in general the complexity of
a convex set is not bounded. Since we assume that the complexity of each set
depends only on d, then the running time is O(f (d) · logα ǫ), which does not
depend on n.
Acknowledgments
I would like to thank Ronitt Rubinfeld for her interest and for suggesting me to
write this note.
References
[1] E. Helly. ber Mengen konvexer Krper mit gemeinschaftlichen Punkte.
Jahresbericht der Deutschen Mathematiker-Vereinigung, 32:175–176, 1923.
URL http://eudml.org/doc/145659.
[2] M. Katchalski and A. Liu. A problem of geometry in Rn . Proceedings of the American Mathematical Society, 75(2):284–288, 1979. URL
http://www.jstor.org/stable/2042758.
2
[3] Wikipedia.
Helly’s
Theorem.
https://en.wikipedia.org/wiki/Helly%27s_theorem.
URL
[4] T.
Zhang.
Notes
on
Helly’s
Theorem.
URL
http://www.math.uchicago.edu/~may/VIGRE/VIGRE2007/REUPapers/FINALAPP/ZhangT.pdf.
3
| 8cs.DS
|
JOURNAL OF COMPUTER SCIENCE AND ENGINEERING, VOLUME 1, ISSUE 1, MAY 2010
42
A Metaheuristic Approach for IT Projects
Portfolio Optimization
Shashank Pushkar, Abhijit Mustafi and Akhileshwar Mishra
Abstract— Optimal selection of interdependent IT Projects for implementation in multi periods has been challenging in the
framework of real option valuation. This paper presents a mathematical optimization model for multi-stage portfolio of IT
projects. The model optimizes the value of the portfolio within a given budgetary and sequencing constraints for each period.
These sequencing constraints are due to time wise interdependencies among projects. A Metaheuristic approach is well suited
for solving this kind of a problem definition and in this paper a genetic algorithm model has been proposed for the solution. This
optimization model and solution approach can help IT managers taking optimal funding decision for projects prioritization in
multiple sequential periods. The model also gives flexibility to the managers to generate alternative portfolio by changing the
maximum and minimum number of projects to be implemented in each sequential period.
.Index Terms— IT Projects portfolio Management, optimization, Financial Evaluation, Genetic Algorithm , Real Option.
—————————— ——————————
1 INTRODUCTION
IT Projects investment decisions are crucial for any firm to
implement e-Business. Widely used Discounted Cash Flow
(DCF) method is not appropriate to evaluate IT investments
as it is designed for the projects with no option features
([1],[2]). The inability of Discounted Cash Flow analysis to
take care of impact of flexibility underlying IT investment
decisions have forced IT managers to rely on their gut instinct.[3].
Real Option analysis has been an alternative approach
that incorporates impact of flexibility while evaluating IT
projects. This project evaluation method has been used
based on quantification of projects benefits and costs [4] as
well as the risk and volatility of cash flows ([5],[6]).But very
few research has scrutinized the relevance of real option
analysis of IT investments for optimizing a portfolio of
projects. Dickinson et. al. [7] introduces an optimization
model for interdependent technology projects but it is not in
the framework of real option valuation. Portfolio of IT
projects involve interdependencies which can create multiple options ( [8],[9]).Cobb and channes [10] propose an
approach for the Real Option Valuation of the portfolio of
real investment projects. Bardhan et.al. [11] proposes an integer programming model to get the optimal sequence for
implementation of IT projects.Costa et. al.[12] proposes an
approach for evaluating the software project risk.Fang and
Chen [13] propose a portfolio selection model for a mixed
R&D projects.Liang et. al. (2008)[14] give a framework for
Shashank Pushkar is with Birla Institute of Technology, Mesra , Ranchi,
India-835215.Phone Number: 91-651-227554(O), 91-9470141130.
A Mustafi is with Birla Institute of Technology, Mesra , Ranchi, India835215. .Phone Number: 91-651-227554(O), 91-9431382747.
.Akhileshwar Mishra is with National Institute of Technology, Department of Computer Application , Jamshedpur Jharkhand, India-831014.
the IT investment on the basis of Real Option and Mean –
Variance theory perspective.Peters and Verhoef[15] quantifies the yields of risk of IT portfolio. Shashank et.al. [16]
proposes a dynamic programming solution to this class of
problem,but it works for the smaller portfolio as with increaded number of projects the variable becomes very large
because of curse of dimensionality.
This paper gives a simplified mathematical model to optimize an IT projects portfolio where projects can be sequentially interdependent (which is an important characteristic of
IT projects. The model dynamically calculates option values
for each project due to its dependent projects implemented
in the period subsequent to it. Genetic Algorithms have a
proven track record in handling such large search field problems. The simple act of crossover and mutation allows the
algorithm to search a large search space and converge quickly to an optimal solution. In case of exceptionally large
search spaces the algorithm provides the option of being
terminated after a fixed number of iterations . This may not
provide the best solution in all cases but usually gives a very
good approximation of the solution. The proposed solution
yields the optimum sequence for implementation of IT
projects to have the maximum overall portfolio value
(DCF+ROV) across multiple time periods. This can help IT
managers taking optimal funding decisions. The rest of the
paper is organized as follows.
Section 2 specifies the problem The mathematical model
is defined in section 3. A Genetic Algorithm solution is given
in section 4 and an illustrative numerical example is given in
section 5. Finally we conclude in section 6.
© 2010 JCSE
http://sites.google.com/site/jcseuk/
43
2
PROBLEM SPECIFICATION
The Problem undertaken comes under the category of IT
project valuation and investment decisions for project portfolio management. This problem applies to any firm that
decides to implement e-Business or wants to invest in multiple IT projects. Here the collection of IT projects is a portfolio of projects to be implemented in sequential periods. The
projects are implemented separately but they have two types
of dependency. They are as follows:
1. Total dependency
2. Partial dependency
3. No dependency
Total dependency of project i on project j indicates that the
capabilities developed for project j is also required by project
i i.e. j creates an option to implement i and I can be implemented either together with j or after the implementation of j
but not before j‖s implementation. If i is implemented after
implementation of j, Call Option value of i should be added
to the DCF value of project j(according to real option analysis).
Partial dependency of project i on j indicates that the capabilities developed for project j supports or enhances the
capabilities required by project i, but there is no strict requirement of project i to be implemented together with j or
after j‖s implementation. But if i is implemented without
implementation of j, the benefit level of i would reduce by
some fraction depending on the level of dependency.
α1
α5
α2
α3
α8
α4
α7
α6
Fig.1 Possible dependencies among the IT
projects
In the fig. 1 α1, α2 .etc. .are all investment projects. The
solid arrow indicates total dependency (e.g. α2 is totally dependent on α1). Dashed arrow indicates partial dependency
(e.g. α6 is partially dependent on α5). The project that has no
dependency on any other projects are independent and
without arrow.
All these IT projects in the portfolio could not be implemented simultaneously due to the constraints of the budget
of the firm and uncertainty regarding success of IT projects
(customer response, implementation success within the firm
etc.).Therefore these projects are generally implemented
stage-wise in sequential periods. Each period may have dif-
ferent budget. The projects implemented in initial stage or
period provides opportunities (option) for remaining unimplemented projects which are dependent on implemented
ones. Thus the initial implementation of projects provide
flexibility to the managers to decide about whether to implement the remaining projects in the portfolio (seeing the
response from the customers and the success of projects implementation within the firm ).Since these projects generally
share technology and the firm , they share risk. So, the
projects implemented in early stages (periods) relax the uncertainties for the remaining unimplemented projects.
According to real option valuation methodologies,
projects create option for dependent projects to be implemented in subsequent periods and hence option value of
dependent projects must be added as Real Option value to it.
So, the total value of a project will be its DCF value plus Real
Option values due to its dependent projects to be implemented in subsequent periods. Here it is important to note
that the Real Option Value for these dependent projects will
not be added to Option generating project which are either
implemented before it or together with it.
The problem requires to be formulated as a mathematical
model for the optimization of such IT projects portfolio
across multiple sequential periods so that the overall value
of the portfolio is maximized.
3 MATHEMATICAL MODEL
3.1 Notation
p
np
N
Bk
r
Cik
Di
Set of given projects
Number of project.
Total number of periods for completing the
implementation of the projects
Budget for period k (k=1,2,..,N)
rate of interest
Present Value of cost of project i if i is implemented
in period k
( i = 1,2,..,n p ; k = 1,2,..,N).
cost of project i funded in period k /(1 r ) k-1
( here if k =1, it means the project is funded in the
beginningof first period)
Set of projects directly dependent on project i
Di
dij
P
Level of dependence of project i on project j, defined
as follows:
dij 0 , if j is not directly dependent on i;
0 dij 1 , j is partially dependent on i;
dij 1 , , if j is totally dependent on i.
Rik
Present value of return from project i , if it is
implemented in period k.
44
x
rit /(1 r ) t
Qk min
t k
where,
return from project i at the end of t th period.
rit
x
Xik
Xik
the expected number of periods up to which the
project is going to give the return .
k period of implementation of project i
Binary value indicating implementation period for
project i , where,
and ,
Xik
1
k 1
i.e. each project i is implemented only once.
Wik
Vij
Discounted cash flow(DCF) value of project i if i is
implemented in period k .
The Net Present Option value for project i attributed
to its dependent projects if i is implemented in
period k.
Option value for project i due to project j , where
Yijk
Vij 0 if j Di Di.
1 , if i is implemented prior to implementation of j
Vik
0 , otherwise.
Qk min
minimum number of projects to be
implemented in period k
Qk max =Maximum number of periods to be implemented
in period k
3.2 Problem Statement
It follows from notation above that
np
Vik
Vij * dij * Xik * (1
j 1
Wik
k
Xik )(i
P, j
k 1
( Rik Cik ) * Xik , (i
P)
The problem is to maximize
N
i P
( Number of
Here the objective function indicates that the real option
value of project i due to project j will only be added if i is
implemented prior to the implementation of j. Observe that
no additional sequencing constraint is required as maximizing the above objective function will automatically take care
of the sequencing.
4 GENETIC ALGORITHM BASED SOLUTION
1, if i is implemented in period k;
0, otherwise.
N
i X i k ≤ Qk max
2)
≤
projects to be implemented in a period)
(Vik Wik )
k 1
Subject to the constraints
1) Σ C i k . X i k - B k ≤ 0 ( Budget Constraint)
i
Di )
Genetic algorithms[18] are meta-heuristic optimization techniques based on natural theory and survival of the fittest .
The operators involved in GA tend to be heavily inspired by
natural selection and consequently successive generations of
the algorithm continue to propagate the best traits of the
population.
This leads to rapid convergence of the
search[25]. Also the introduction of the mutation operator
ensures that diversity is not neglected and the search is not
trapped in a local maximum. A flow chart illustrating the
basics steps of GA based optimization is given in fig 2.
The chromosome structure chosen to represent the problem
is a sequence of ―b‖ bits where
b =(no of projects*no of periods)
A representative bit sequence for a portfolio of seven
projects to be completed over three(3) periods would consequently be a 21 bit sequence. It is easy to visualize the above
case in terms of integer numbers where any number in the
range 0 – 2b-1 would have its binary representation as one of
the possible chromosomes.
Once defined the chromosome is divided into N equilength sequences where N is the number of periods under
consideration. Every set bit in these subsequences would
represent a project to be completed in that particular period.
The GA is then introduced to search over the search space
to optimize the NPV of the portfolio
45
α2
1
0
0
0
0
0
α3
1
0
0
0
0
0
α4
0
0.25
1
0
0
0
α5
0
0
0
0
0
0
α6
0
0
0
0
0.25
α7
0
0
1
0
0
0
0
TABLE 3 – OPTION VALUES DUE TO DEPENDENCY AMONG
PROJECTS
α1
α2
α3
α4
α5
α6
α7
Fig. 2 Flowchart for GA based Optimization
.
5 NUMERICAL EXAMPLE
A seven IT projects portfolio is taken as an illustrative example. The interdependencies among projects are shown in
Figure 1 in Section 2 of the paper. The results were simulated using MATLAB® genetic algorithm tollbox.
Here the project planning horizon is taken as N= 3 periods. i.e. k = 1,2,3 and n p = 7.
Other data about the portfolio is listed in the tables below.
TABLE 1- COST-RETURN OF PROJECTS
Project
i
Present Value
of cost in
period k
15
30
70
60
15
50
125
α1
α2
α3
α4
α5
α6
α7
Present Value of
return in period k
13
35
65
100
20
150
150
Note: Present value of costs and returns for projects to be implemented in period k is taken same for k= 1, 2, 3 for the shake of simplicity.
Table 2 – DEPENDENCY LEVEL AMONG PROJECTS
α1
α1
α2
α3
α4
α5
α6
α7
0
0
0
0
0
0
α1 α2 α3 α4 α5 α6 α7
10
0
0
0
0
0
0
0
0
0
0
0
10
0
0
0
0
0
0
5
10
0
0
0
0
0
0
0
0
0
0
0
0
0
5
0
0
10
0
0
0
0
Note: Option values for dependent projects are taken
same for simplicity but they can be calculated using
nested option model option model( Benaroach et.al. [20])
for each project due to its each dependent project.
Budgets for each three periods are as under:
B1 = 90 , B2 = 125, B3 = 175.
Qkmin = 2 and Qkmax = 3 for each k = 1, 2, 3.
TABLE 4-OPTIMIZATION RESULTS
Selected Projects
for funding
Costs of projects
Budget
PORTFOLIO
VALUE
K=1
α1, α3
85
90
K=2
α2, α4, α5
105
125
616.75
K=3
α6, α7
175
175
The GA solution given in section IV of the paper is applied to the above example portfolio. The results obtained
are tabulated above in Table 4 and the convergence of the
algorithm is shown in Fig. below.
The results show that those projects that provide infrastructure to many other projects(thus having high option
values) are selected in early periods for funding and those
having no option value are deferred to the later periods. This
also indicates that the maximization of the option component of portfolio value selects the projects ,which have maximum number of dependent projects ,earlier than those
projects which are having less dependent projects or no dependent. Since the projects of these types of portfolio shares
risk for being successful, early implementation of high op-
46
tion projects would lower the overall risk of success for the
portfolio.
The model yields the optimum value for the overall portfolio along with the periods for the funding of projects in
the portfolio. The method calculates the option values of
each project due to its dependent projects dynamically and
thus represents a significant improvement over the existing
models for prioritization of IT projects in real option valuation framework .This approach can help IT managers taking
optimal funding decisions .
As a final demonstration the best chromosome for the
representative case used in the paper is presented below.
The chromosome represents 3 periods of activity and each
column has the set bits representing the projects to be completed ideally in that particular period. The solution clearly
demonstrates the findings of Table 4.
100
010
100
010
010
001
001
Fig. 3 The best chromosome for the case study
Fig.4 The Convergence and best solution for the GA
6 CONCLUDING REMARK
We have developed a multi period portfolio optimization
model. The model uses GA to calculate option values for
each project due to its dependent projects implemented subsequently. This is the improvement over the existing works
where the option values are calculated statically for each
projects due to its dependent projects. and not dynamically
at the time of implementation decision making. The meta
heuristic nature of the solution yields the optimal sequence
of implementation of IT projects in multiple periods to get
the maximum overall portfolio value .The proposed algorithm is very suitable to solve this problem as it is modeled
as a multistage optimization problem and it makes possible
to calculate option values for each project due to its subsequently implemented dependent projects to maximize the
overall portfolio value. The model also gives flexibility to the
managers to generate alternative portfolio by changing the
maximum and minimum number of projects to be funded in
a period.
The research work can be further extended by incorporating fuzziness in the model as the terms like dependency level among projects, level of benefits etc. are uncertain and
may change with the changing decision time
REFERENCES
[1] Benaroach, M. , and Kauffman, R.J.A Case for using real options
pricing analysis to evaluate information technology project investments.Information system Research, 10,1(1999),70-86.
[2] Benaroach, M., and Kauffman, R.J.Justifying electronic network
expansion using real option analysis . MIS Quarterly, 24, 2
(2000), 197-225.
[3] Fitchman, R. Real options and IT platform adoption : Implications
for theory and practice. Info.Sys.Res., 15, 2 (2004), 132-154.
[4] Kambil, A. , Henderson, C. J. , and Mohsenzadeh, H. Strategic
management of information technology investments . In R.
Banker, R. J. Kauffman and M. A. Mahmood (ed.), Strategic Information Technology Management: Perspective on Organizational Growth and Competitive Advantage. Harrisburg, PA,
Idea Group Publishing,1993,161-178.
[5] Benaroach, M. Managing information technology investment
risk: A real options perspective . Journal of Management Information System, 19,2(2002), 43-84.
[6] Schwartz, E. S., and Zozaya-Gorostiza, C.Investment under uncertainty in information technology: acquisition and development projects. Mgmt.Sci.,49,1(2003),57-70.
[7] Dickinson, M., Thomton, A. , and Graves, S. Technology portfolio
management: Optimizing interdependent projects over multiple time periods. IEEE Transactions on Engineering Management 48,4(2001),518-527.
[8] Bardhan, I.R., Sougtad, R, and Bagchi, S. Prioritization of a portfolio of information technology projects. Journal of Management
Information Systems, 21, 2(2004),33-60.
[9] Bardhan, I.R., Sougtad, R, and Bagchi, S., A real option approach
for prioritizing a portfolio of information technology projects: A
case study of a utility company.2004b.In R Sprague(ed.), Proc.
37th Hawaii Intl. Conf. Sys. Sci., IEEE Comp.Soc.Press, Los Alamitos, CA.
[10] Cobb, B. R. And Charnes, J. MSimulations and Optimizations for
Real Option Valuation. Proceedings of Winter simulation Conference,2003.
[11] Bardhan, I.R., Kauffman,R.J., and Narapanawe,S. Optimizing an
IT project portfolio with time wise interdependencies IEEE computing Society Press, Los Alamitos, CA, 2006.
[12] Costa. H. R,Barros, M. D., Travassos, G. H. Evaluating Software
Project Portfolio Risks. The Journal of Systems and Software
80(2007)16-31.
[13] Fang , Y. , Chen, L., Fukusima , M. A mixed R& D Projects and
Scurities portfolio selection model. European Journal of Operation research 185(2008) 700-715
.[14] Liang-Chuan Wu, Chorng-Shyong Ong(2008) Management of
47
information technology investment: A framework based on a
Real Options and Mean–Variance theory perspective. Technovation ,28(3)122-134.
[15] Peters , R. J. , Verhoef, C. Quantifying the yield of risk-bearing IT
portfolios. Science of Computer Programming 71(2008) 17-56.
[16] Shashank Pushkar, Sharma R.R. and A. Mishra. Dynamic Programming approach to optimize portfolio of Interdependent IT
projects.CiiT International Journal of Artificial Intelligent Systems and Machine learning, 1,4(2009)127-131.
[17] Trigeorrgis, L. Real Options, Cambridge, MA, MIT Press, 1996.
[18] Goldberg, D E, “Genetic Algorithms in Search, Optimization, and
Machine Learning”, Addison-Wesley Professional, Ist Edition,
1989
[19] Paulinas M. ,Usinskas A. , “A survey of Genetic Algorithms
Applications for Image Enhancement and Segmentation”, Information Technology and Control, Vol. 36 No. 3, 2007.
[20] Benaroch, M., Shah, S., and M. Jeffery. On the valuation of multistage investments embeddingnested (compound) real options.
2005. Workingpaper, Syracuse University, Syracuse, NY.
Shashank Pushkar is a Lecturer in the Department of Computer
Science and Engineering, Birla Institute of Technology , Mesra ,
Ranchi. His research interest is in the field of Information Technology
project Management and Optimization Technique.
Abhijit Mustafi is a MCA from the University of North Bengal, India.
He is currently a Senior Lecturer in the Department of CSE, BIT
mesra, India. His research interests include image processing, meta
heuristic algorithms and web mining.
Dr. Akhileshwar Mishra is PhD from IIT Kharagpur. He is also a
professor of Computer Applications , National Institute of Technology, Jamshedpur. He specializes in computer applications and Optimizations in the field of Industrial Management
| 5cs.CE
|
Optimal Design and Optimal Control of Elastic Structures
Undergoing Finite Rotations and Elastic Deformations
A. Ibrahimbegović1, C. Knopf-Lenoir2 , A. Kučerová1 and P. Villon2
arXiv:0902.1037v2 [cs.NE] 10 Feb 2009
1
Ecole Normale Supérieure - Cachan, LMT, GCE, 61, avenue du président Wilson,
94235 Cachan, France, email: [email protected]
2
Université de Technologie - Compiegne, Lab. Roberval, GSM,
60200 Compiegne, France
Abstract
In this work we deal with the optimal design and optimal control of structures undergoing large rotations. In other words, we show how to find the corresponding initial
configuration and the corresponding set of multiple load parameters in order to recover a desired deformed configuration or some desirable features of the deformed
configuration as specified more precisely by the objective or cost function. The model
problem chosen to illustrate the proposed optimal design and optimal control methodologies is the one of geometrically exact beam. First, we present a non-standard formulation of the optimal design and optimal control problems, relying on the method
of Lagrange multipliers in order to make the mechanics state variables independent
from either design or control variables and thus provide the most general basis for developing the best possible solution procedure. Two different solution procedures are
then explored, one based on the diffuse approximation of response function and gradient method and the other one based on genetic algorithm. A number of numerical
examples are given in order to illustrate both the advantages and potential drawbacks
of each of the presented procedures.
Keywords: structures, finite rotation, optimization, control
1 Introduction
Modern structures should often be designed to withstand very large displacements and
rotations and remain fully operational. Moreover, the construction phase has also to
be mastered and placed under control, with trying to precisely guide the large motion
of a particular component of the total structural assembly. Ever increasing demands
to achieve more economical design and construction thus require that the problems of
this kind be placed on a sound theoretical and computational basis, such as the one
explored in this work. Namely, optimization methods can be called upon to guide
the design procedure and achieve desired reduction of mechanical and/or geometric
properties. Similarly, the control methods are employed to provide an estimate of the
loads and the minimal effort in placing the structure, or its component, directly into
an optimal (desired) shape. Either of these tasks, optimal design or optimal control,
can formally be presented as the minimization of the chosen cost or objective function
1
specifying precisely the desired goal. The main difference between two procedures
concerns the choice of the variables defining the cost function: the design variables
are typically related to the mechanical properties (e.g. Young’s modulus) or geometry
of the structure (e.g. particular coordinates in the initial configuration), whereas the
control variables are related to the actions (e.g. forces) applied on the structure in order
to place it into desired position. Rather then insisting upon this difference and treating
the optimal design and optimal control in quite different manners (as done in a number
of traditional expositions on the subject), we focus in this work on their common
features which allow a unified presentation of these two problems and the development
of a novel solution procedure applicable to both problems. The latter implies that the
nonlinear mechanics model under consideration of geometrically exact beam has to
be placed on the central stage and one should show how to fully master the variation
in chosen system properties or loads in order to achieve the optimal goal. The main
contributions put forward in presenting this unconventional approach can be stated as
follows:
i) We first present the theoretical framework for treating nonlinear structural mechanics and optimization and/or control as a coupled (nonlinear) problem; In
any such problem of optimization or control, the nonlinear mechanics equilibrium equations are reduced to a mere constraint with respect to the admissibility of a given state of the structure, i.e. its displacements and rotations. By
using the classical method of Lagrange multipliers (e.g. see [19] or [29] the
mechanics equilibrium equations can be promoted from constraint to a one the
governing equations to be solved in a coupled problem of this kind, and the intrinsic dependence on state variables (displacements and rotations) with respect
to optimal design or control variables can be eliminated turning all variables
into the independent variables. For clarity, this idea is also developed within
the framework of a discrete, finite-element-based approximation, thus providing the finite element model including the degrees of freedom pertinent not only
to displacements and rotations, but also to optimal design and/or control variables. A detailed development is presented for the chosen model problem of 3D
geometrically exact beam (e.g. see [27] or [10]). We note in passing that the
proposed approach is quite opposite from the traditional ones (e.g. see [17] for
a very recent review), where the two fields directly concerned by the coupled
problems of this kind, nonlinear mechanics on one side and optimal design or
control on the other, are studied and developed separately and resulting solution
procedures for one or another are applied in a sequential manner. One typically
employs two different computer programs, one for mechanics and another for
optimization or control (e.g. the toolbox of MATLAB), so that the communication requirements are reduced to a bear minimum: so-called sensitivity (e.g.
[30] or [25]) for optimization code, and design or control variables for the finite element code for mechanics. It is clear that such a traditional approach to
analysis and design or control will largely sacrifice the computational efficiency
for the cases of practical interest where both the cost function and mechanics
problem are nonlinear and require iterative procedures to be solved.
ii) The second aspects which is elaborated upon in this work pertains to an alternative method for solving a coupled problem for analysis and either design
2
or control, where two sub-problems, once brought up to the same level by the
Lagrange multiplier method, are solved simultaneously. In other words, once
the interdependence of the state variables, i.e. the displacements and rotations
on one side and the design or control variables or another, is no longer enforced, one can iterate simultaneously to solve for all of them. In particular,
the sensitivity analysis needs no longer be performed separately, but it is naturally integrated as a part of the simultaneous iterative procedure. It is important
to note that the iterative intermediate values are no longer consistent with the
equilibrium equations constraint, except at the convergence, where basically the
same solution is obtained as for the standard sequential solution procedure but
with a (significantly) reduced number of iterations. In other words the simultaneous and sequential solution procedures will always yield the same result
providing the solution is unique. The problems of this kind concern the optimization and/or control of geometrically nonlinear response of structures where
no bifurcation phenomena occur. Solving the nonlinear problems in structural
mechanics can be quite a demanding task for the structures where the stiffness
may differ significantly in different deformation modes (i.e. beam bending versus stretching). Adding the control or optimization problem on top makes such
a nonlinear problem so much more challenging. Different modifications of the
basic solution procedure are developed and tested, including genetic algorithms
[7] to explore the initial phase of the solution procedure, gradient based acceleration near solution points and response surface for a part of the solution
constructed by diffuse approximation (e.g. see [2]).
The outline of the paper is as follows. In the next section we briefly present the
chosen model of geometrically exact 3D beam, capable of representing very large
displacements and rotations. The theoretical formulation of the optimal control and
optimal design of the chosen mechanical model is presented in Section 3, along with
their discrete approximations constructed by the finite element method given in Section 4. The proposed solution procedure is described in detail in Section 5. Several
numerical examples are presented in Section 6 in order to illustrate the performance of
various algorithms used in computations. Some closing remarks are stated in Section
8.
2 Model problem: geometrically exact 3D beam
In this section, we briefly review the governing equations of the chosen model problem
of a structure undergoing large rotations, a 3D curved beam. For a more thorough
discussion of the chosen model we refer to [27], [26] or [13], among others. We
assume that the initial configuration of the beam is internal force free and that it can
be described by a 3D position vector ϕ0 (s) identifying the position of each point of the
neutral fiber (an inextensible fiber in pure bending) and the corresponding placement
of the cross-section of the rod, which is carried out by choosing a local ortho-normal
triad of vectors. The vector triad of this kind can be obtained by simply rotating the
global triad by an orthogonal tensor.For a usual choice of so-called normal coordinates
with the first vector of the triad being orthogonal to the cross-section and the remaining
3
s
φ = (ϕ, Λ)
φ0 = (ϕ0 , Λ0 )
✄
✂
ξ
✁
Figure 1: Initial and deformed configuration of the 3D geometrically exact beam.
two placed in the plane of the cross section, this orthogonal tensor becomes a known
function of the initial configuration, Λ0 (s). For the case of a curved beam studied
here, ’s’ is chosen as the arc length parameter.
By applying the external loading f(t), parameterized by pseudo-time ’t’ (where
’pseudo’ implies that the inertia effects are neglected) we obtain the beam deformed
configuration defined by the position vector ϕ(s, t) and the orthogonal tensor Λ(s, t).
The latter is an accordance with the usual kinematic hypothesis that the cross-section
of the beam would not deform, which, along with the hypothesis that the first vector in
the triad remains orthogonal to it (with other two within the plane of the cross-section)
fully determines Λ(s, t). See Figure 1 where the initial and deformed configuration of
the beam are presented.
In short, one can state that the configuration space of the described model of 3D
beam can be written as
C := {φ = (ϕ, Λ) | ϕ ∈ R3 , Λ ∈ SO(3)}
(1)
where R3 and SO(3) are spaces of 3D vectors and special orthogonal tensors, respectively.
The main difficulty in numerical solution of the structural mechanics problems
featuring the beams of this kind stems from the presence of SO(3) group in its configuration space (e.g. see [1], [13], [11] for a more thorough discussion of these issues).
In short, in performing the standard task of computing the virtual work principle or the
consistent linearization, where a small rotation described by a skew-symmetric tensor
δΘ ∈ so(3), ought to be superposed on a large rotation described by an orthogonal
tensor Λ ∈ SO(3) one must first make use of exponential mapping
Λε = Λ exp[ε δΘ] ;
exp[δΘ] = cos δθ I +
sin δθ
1 − cos δθ
δΘ +
δθ ⊗ δθ
δθ
δθ2
(2)
where δθ is the axial vector of the skew-symmetric tensor δΘ i.e. δΘ v = δθ × v,
∀v ∈ R3 .
The complexity of the last expression is in sharp contrast, with respect to a simple additive update of virtual displacement field δϕ ∈ R3 when superposed on the
deformed configuration ϕ ∈ R3
ϕε = ϕ + ε δϕ
4
(3)
The results in (2) and (3) can be presented in an equivalent form by stating that the
tangent space of the chosen beam model is defined by
T C := {δφ := (δϕ, δθ) |δϕ ∈ R3 , δθ ∈ R3 }
(4)
The strain measures employed in this beam theory (e.g. [10]) can be written in
direct tensor notation form as
ǫ = Λϕ′
(5)
Ωv = ω × v ; ∀v ∈ R3
(6)
for the axial and shear strains, and
Ω = ΛT Λ′ ;
for bending and torsional strains. In (5), (6) and subsequent equations we denote with
superposed prime the derivative with respect to arc-length coordinate in the initial
configuration, i.e.
∂
(·) = (·)′
(7)
∂s
We consider the simplest case of linear elastic material model for the beam which
allows us to express the constitutive equations in terms of stress resultants as
n = C(ǫ − ǫ0 ) ;
m = D(ω − ω 0 ) ;
C = diag(EA, GA, GA)
(8)
D = diag(GJ, EI, EI)
(9)
For illustration, we also consider the simplest case of a circular cross section, with
section diameter ’d’, and
A=
d2 π
;
4
I=
d4 π
;
64
J=
d4 π
32
(10)
as the section area, moment of inertia and polar moment.
In order to complete the description of the chosen beam model we state the equilibrium equations in the weak form as
G(φ; δφ) :=
Z
(δǫ · n + δω · m) ds − Gext (δφ) = 0
(11)
where Gext (δφ) is the external virtual work and δǫ and δω are the virtual strains. The
latter can be obtained as the Gâteaux derivative of the real strains in (5) and (6), by
taking the results in (2) and (3) into consideration. In particular, this leads to
δb
ǫ(φ; δφ) = Dφ [b
ǫ(φ)]
d T ′
[Λ ϕ ] |ε=0
=
dε t,ε t,ε
= ΛT δϕ′ + ǫ × δθ
5
(12)
and
b
b t,ε )]
δ Ω(φ;
δφ) = Dφ [Ω(φ
d T ′
=
[Λ Λ ] |ε=0
dε t,ε t,ε
= δΘ′ + δΘT Ω + ΩδΘ
(13)
which can also be written in an equivalent form in terms of the corresponding axial
vectors
δω = δθ ′ + ω × δθ
(14)
For the external load which derives from a given potential, such that
Gext (δφ) := Dφ Πext (φ)
(15)
one can also define the governing equilibrium equation of the chosen model problem
of the geometrically exact elastic beam from the principle of minimum of the total
potential energy defined according to
Z
1
Π(φ) :=
{(ǫ − ǫ0 ) · n + (ω − ω 0 ) · m} ds − Πext (φ) → min
(16)
l 2
which implies that
Dφ [Π(φ)] ≡ G(φ; δφ) = 0
∗
Dφ [Dφ Π(φ)] ≡ Dφ G(φ; δφ) > 0
Π(φ) = min∗ Π(φ ) =⇒
{z
}
|
∀φ
δφ·Kδφ
(17)
In (17) above we denoted by K the second variation of the total potential energy
or the tangent operator, which is obtained by the consistent linearization procedure
(e.g. see [20]). For pertinent details of the consistent linearization we refer to [10] for
quaternion parameterization of finite rotations or [13] for rotation vector-like parameters for finite rotations.
An alternative case of external loading where the potential may not be defined and
where the weak form in (11) is rather the starting point of the solution procedure, can
also be encountered in applications. For example, in the case of the follower force p0
and follower moment l0 which follow the motion of a particular cross-section of the
beam at node ’a’, we can express their contribution to virtual work according to
Gext (Λa ; δϕa , δθ a ) := δϕa · Λp0 +δθ a · Λl0
|{z}
|{z}
p
(18)
l
The follower force and moment will also contribute to the tangent operator according to
∆ϕa
= δϕa · P∆θ a + δθ a · L∆θ a
(19)
Dφ Gext (Λ; δϕa , δθa ) ·
∆θ a
where Pv = p × v; Lv = l × v; ∀v ∈ R3 . This contribution should also be taken
into account when computing the solution to (11) and trying to ensure the quadratic
convergence rate.
6
3 Coupled optimality problem
The presented beam model provides an excellent basis to master the optimization
problem as well as the control problem of geometrically nonlinear elastic structures.
Although the former deals with geometric characteristics of the beam and the latter
with the external loading sequence, the two problems can be formulated and solved in
quite an equivalent manner, as shown next.
3.1 Optimal design
The optimal design problem addressed herein pertains to selecting the desired features
of the mechanical model by leaving the free choice of the geometric properties of the
beam model, e.g. the thickness variation or the chosen initial shape. This task is often
referred to as the shape optimization. From the mathematical standpoint the shape
optimization can be formulated as the problem of minimization of so-called objective
or cost function J(·), specifying the desired features. The latter is considered as a
functional which depends not only on mechanics state variables φ but also on design
variables d (e.g. the beam thickness or its shape).
The shape optimization procedure is interpreted herein as minimization of so-called
ˆ which can be written as
cost or objective functional J(·),
ˆ φ̂(d), d) =
J(
min
ˆ∗
ˆ φ̂(d∗ ), d∗ )
J(
(20)
G(φ (d∗ );·)=0
Contrary to the minimization of the total potential energy functional in (17), not all
mechanical and design variables are admissible candidates, but only those for which
the weak form of the equilibrium equations is satisfied. In other words, we now need
to deal with a constrained minimization problem.
The classical shape optimization procedure of solving this constrained minimization problem is carried out in a sequential manner, where for each iterative value of
design variables d(i) , a new iterative procedure must be completed leading to φ(d(i) )
verifying the equilibrium equations. The considerable computational cost of such a
procedure, most of it waisted on iterating to convergence on equilibrium equations
even for non-converged values of design variables, can be drastically reduced by formulating the minimization problem in (20) using the method of Lagrange multipliers
with
max min
L(φ∗ , d∗ ; λ) ;
∀λ ∀(φ∗ , d∗ )
L(φ∗ , d∗ ; λ) = J(φ∗ , d∗ ) + G(φ∗ , d∗ ; λ)
(21)
In (21) above λ are the Lagrange multipliers inserted into the weak form of equilibrium equations instead of virtual displacements and rotations. In accordance to the
results presented in the previous section we can write explicitly
G(φ, d; λ) =
Z
{λ′ϕ · Λn + λθ · (ET n + ΩT m) + λ′θ · m}ds − Gext (λ)
l
7
(22)
The main difference of (21) with respect to constrained minimization problem in
(20) pertains to the fact that state variables φ and design variables d are now considered independent and they can be iterated upon (and solved for) simultaneously.
The Kuhn-Tucker optimality condition (e.g. [19]) associated with the minimization
problem in (18) can be written as
0 = Dφ [L(φ, d; λ)] = Dφ [J(φ, d)] + Dφ [G(φ, d; λ)]
(23)
with the explicit form of the last term which can be written as
R
Dφ [G(φ, d; λ)] · δφ = l λ′ϕ · (ΛCΛT δϕ′ + ΛCEδθ)
T
+λ
CΛT δϕ′ + (ET CE + ΩT DΩ)δθ + ΩT Dδθ ′ ] + λ′θ · (DΩδθ + Dδθ ′ ) ds
R θ· [E
+ l λ′ϕ · (ΛNT δθ) + λθ · [NΛT δϕ′ + [Ξ (ǫ × n) + Ξ (ω × m)] δθ + Mδθ ′ ] ds
(24)
where we denoted Ξ(a × b) = (a ⊗ b) − (a · b)I, as well as Mv = m × v and
Nv = n × v, ∀v ∈ R3 ; Moreover, we have
0 = Dd [L(·)] = Dd J(·) + Dd G(·)
(25)
where
Dd [G(·)] · δd =
Z
λ′ϕ
l
∂m
∂n
′
T ∂m
T ∂n
+ λθ ·
· δdds
·Λ
+ λθ · E
+Ω
∂d
∂d
∂d
∂d
(26)
Finally, we also have
0 = Dλ [L(·)] · δλ =
Z
l
′
δλϕ · Λn + δλθ · ET n + ΩT m + δλ′θ · m ds
(27)
For illustration we can consider further that the diameter of a circular cross-section
is chosen as the design variable, which allows us to express explicitly the result in (23)
as
∂n
∂C
∂C
∂A ∂A ∂A
d2 π
=
(ε − ε0 ) ;
= diag[E
,G ,G
]; A=
;
∂d
∂d
∂d
∂d
∂d
∂d
4
∂m
∂D
∂D
∂J
∂I
∂I
d4 π
=
(ω − ω0 ) ;
= diag[G , E , E ] ; J = 2I =
∂d
∂d
∂d
∂d ∂d ∂d
32
(28)
(29)
In order to provide a similar explicit result for directional derivative of the cost function, we consider a simple choice given as the beam mass (or equivalently its volume
for a constant density),
J(φ, d) ≡ V =
Z
A ds ;
A=
d2 π
4
(30)
l
In such a case the contribution of the cost function to the Kuhn-Tucker optimality
conditions can be written as
8
Dφ J(φ, d) · δφ = 0
Z
∂A
δd ds ;
Dd J(φ, d) · δd =
∂d
∂A
dπ
=
∂d
2
(31)
l
Dλ J(φ, d) · δλ = 0
We can also consider a more complex case of practical interest where the shape optimization is carried out with respect to the beam axis form in the initial configuration.
A reference configuration is selected in such a case (see Figure 1) and the design variable is given in terms of the position vector describing the beam initial configuration
with respect to this reference configuration d ≡ ϕ0 (ξ). The cost function in (25) can
now be described as
J(φ, d) :=
Z
l
A ds =
Zξ2
A j(ξ) dξ ;
j(ξ) =
∂d(ξ)
∂ξ
(32)
ξ1
In this case all the integrals in (23) to (27) must be recomputed with the same
change of variables like the one presented in (32) above typical of the isoparametric
parent element mapping (e.g. see Zienkiewicz and Taylor [31]). We also note in
passing that the derivatives with respect to arc-length coordinate ought to be computed
by making use of the chain rule
1 ∂
d
(·) =
(·)
ds
j(ξ) ∂ξ
(33)
For example, the contribution of the cost function to the Kuhn-Tucker optimality
conditions for such a choice of design variables can be written as
Dd [J(φ, d)] · δd =
Zξ2
A
1
∂d(ξ)
d(ξ) ·
dξ
j(ξ)
∂ξ
(34)
ξ1
3.2 Optimal control
The optimal control problem studied herein concerns the quasi-static external loading
sequence which is chosen to bring the structure directly towards an optimal or desired
final state, which may involve large displacements and rotations. More precisely, we
study the mechanics problems where introducing the pseudo-time parameter ’t’ to
describe a particular loading program is not enough and one also needs to employ the
control variables ν. The latter contributes towards the work of external forces, which
can be written as
Z
ext
G (ν; δφ) := {δφ · F0 ν}ds
(35)
l
where F0 contains the (fixed) distribution of the external loading to be scaled by the
chosen control.
9
Optimal control can be presented in the following form: find the value of the control variables ν, such that the final value of the state variables φ be as close as possible
to the desired optimal (fixed) values φd ; This can be formulated in terms of the constrained minimization of the chosen cost function J(φ̂(ν), ν) which can be written
as
ˆ φ̂(ν), ν) =
ˆ φ̂(ν ∗ ), ν ∗ )
J(
min
(36)
J(
ˆ ∗
φ̂d
G(φ(ν );·)=0
where the role of the constraint, as indicated by the last expression, is to fix the values
of the state variables with respect to the chosen value of the control through the given
set of equilibrium equations. In other words, for a given control we will finally obtain
the configuration φ the same as the desired state φd if the latter verifies equilibrium
equations, or, in the opposite case simply the solution closest to φd which also verifies
the equilibrium equations.
In order to remove this constraint and to be able to consider the state variables independently from the control variables, we resort to the classical method of Lagrange
multipliers; Namely, by introducing the Lagrange multipliers λ we can rewrite the optimal control problem by making use of the Lagrangian functional L(·) which allows
to obtain the corresponding form of the unconstrained optimization problem.
max min L(φ, ν, λ)
∀λ ∀(φ,ν )
;
L(φ, ν, λ) = J(φ, ν) + G(φ, ν; λ)
(37)
φ̄d
One can readily obtain the Kuhn-Tucker optimality conditions for this problem
according to
∂L
· δλ := G(φ, ν; δλ)
∂λ
∂J(·)
∂G(φ; λ)
∂L
· δφ :=
· δφ +
· δφ
0 =
∂φ
∂φ
∂φ
0 =
(38)
(39)
and
∂L
∂J
∂Gext
δν :=
δν +
δν
(40)
∂ν
∂ν
∂ν
The first of these equations is precisely the weak form of equilibrium equation, whereas
the second two will provide the basis for computing the control ν and the Lagrange
multipliers λ. The explicit form of the equilibrium equation is similar to the one in
(11) only with the external load term (35), but with the variation of the Lagrange multiplier δλ replacing the virtual displacement δϕ. For writing the explicit form of other
two Kuhn-Tucker equations we choose a particular form of the objective function such
that
Z
1
1
2
2
(41)
J(φ, ν) =
kφ − φ̄d k + αkνk ds
2
2
l
0=
where φ̄d is the desired beam shape and α is a scalar parameter specifying the weighted
contribution of the chosen control. With such a choice we seek to minimize the ”distance” between the desired shape and the computed admissible shape (the one which
satisfies equilibrium) as well as the force or control needed to achieve that state. The
explicit form of the first term in (39) and (40) can thus be written as
Z
∂J
· δφ =
(42)
δφ · φ − φ̄d ds
∂φ
l
10
and
Z
∂J
· δν = {δν · (αν)} ds
(43)
∂ν
l
Explicit result for the second term of Kuhn-Tucker equations in (39) is identical to the
one in (24) potentially modified according to (19) for the case of the follower load.
Finally, the second term of the Kuhn-Tucker equation in (40) can be written as
Z
∂Gext
· δν =
δν · FT0 λ ds
(44)
∂ν
l
which concludes with the description of all the problem ingredients.
4 Finite element discrete approximations
In this section we discuss several important aspects of numerical implementation of
the presented theory for analysis and design and related issues which arise in numerical simulations.
The analysis part of the problem, i.e. the state variables are represented by using the
standard isoparametric finite element approximations (e.g. see [31]). In particular, this
implies that the element initial configuration is represented with respect to its parent
element placed in the natural coordinate space, corresponding to a fixed interval, −1 =
ξ1 ≤ ξ ≤ ξ2 = +1, by using
nen
X
Na (ξ) xa
ϕ0 (s) ≡ x (ξ) =
(45)
a=1
In (45) above x (ξ) is the position vector field with respect to the reference configuration, xa are nodal values of an element with nen nodes and Na (ξ) are the corresponding shape functions. The latter can easily be constructed for beams by using the
Lagrange polynomials, which for an element with nen nodes can be written by using
the product of monomial expressions
nen
Y
ξ − ξb
Na (ξ) =
ξa − ξb
b=1,b6=a
(46)
where ξa , a ∈ [1, nen ] are the nodal values of natural coordinates.
With isoparametric interpolations one chooses the same shape functions in order
to approximate the element displacement field, which allows us to construct the finite
element representation of the element deformed configuration as
ϕ (ξ) =
nen
X
Na (ξ) ϕa
(47)
a=1
where ϕa are the nodal values of the position vector in the deformed configuration.
The virtual and incremental displacement field are also represented by isoparametric
finite element interpolations
δϕ (ξ) =
nen
X
Na (ξ) δϕa ;
∆ϕ (ξ) =
nen
X
a=1
a=1
11
Na (ξ) ∆ϕa
(48)
The latter enables that a new (iterative) guess for the deformed configuration be easily
obtained with the corresponding additive updates of the nodal values
ϕa ←− ϕa + ∆ϕa
The finite element approximation of the incremental displacement field in (48)
above, where at each point ξ ∈ [ξ1 , ξ2 ] the corresponding value is a linear combination
of the nodal values, are referred to as the continuum consistent (e.g. see [9, 10]), since
they allow to commute the finite element interpolation and the consistent linearization of nonlinear problem (e.g. [20]). Therefore, we also choose the isoparametric
interpolations for virtual and incremental rotation field with
δθ (ξ) =
nen
X
Na (ξ) δθ a ;
∆θ (ξ) =
nen
X
Na (ξ) ∆θ a
(49)
a=1
a=1
so that the commutativity of the finite element discretization and consistent linearization would also apply to the rotational state variables. The only difference from
the displacement field concerns the multiplicative updates of the rotation parameters,
which can be written for any nodal point ′ a′ as
Λa ← Λa exp [∆θ a ]
(50)
In the combined analysis and design procedure proposed herein one must also interpolate the Lagrange multipliers, which is also done by using the isoparametric interpolations according to
n
en
P
Na (ξ) λϕa
λϕ (ξ) =
a=1
Na (ξ) λa ⇐⇒
λ (ξ) =
n
en
P
λθ (ξ) =
Na (ξ) λθa
a=1
nen
X
(51)
a=1
The corresponding integrals appearing in governing Lagrangian functional in (21)
or Kuhn-Tucker optimality conditions in (23), (25) and (27) are computed by numerical integration (e.g. Gauss quadrature, see [31]). To illustrate these ideas we further
state a single element contribution to the analysis part of the governing Lagrangian
functional in (22) given in the discrete approximation setting by
G(φa , d, λa ) :=
λ ϕa
λ θa
Λ(ξl ) E(ξl ) 0
0
Ω(ξl ) I
"
X
nip
dNa (ξl )
I
ds
l=1
T
n(ξl )
m(ξl )
0
0
Na (ξl )I
0
dNa (ξl )
I
ds
j(ξ)wl − Gext (λ(ξl ))
#T
(52)
where ξl and wl are the abscissas and weights of the chosen numerical integration rule
(e.g. see [31]) and nip is the total number of integration points for a single element.
In order to complete the discretization procedure one must also specify the interpolations of the design variables. If the latter is the thickness or the diameter for the
chosen case of a circular cross-section, or the element nodal coordinates, it is possible
12
to use again the isoparametric finite element approximations. However, the best results
are obtained by reducing the number of design variables as opposed to those chosen
at the element level, by using the concept of a design element. This implies increasing
the degree of polynomial by employing, for example, Bézier and B-spline curves for
representation of beam shape and reducing significantly the number of design parameters (e.g. see [16] for a detailed discussion of these ideas). What is important to note
from the standpoint of the simultaneous solution procedure presented further on is that
the design variable at any point is given as a linear combination of the design element
interpolation parameters
d(ξl ) =
ndn
X
Ba (ξl ) da
(53)
a=1
Consequently, the finite or rather design element discretization and consistent linearization will commute again. This observation was already made earlier for linear
analysis problem by Chenais and Knopf-Lenoir [3]. With these results on hand we
can write the discrete approximation of equilibrium equations as
Dλ L(φ, d; λ) := 0 7→
"
#T
nel
nint
dNa (ξe )
X
1
0
0
ds
rλ,e
a =
dNa (ξe )
e=1
0
Na (ξe )1
1
ds
e=1
)
T
T
n(ξe )
Λ (ξe ) E(ξe ) 0
j(ξe ) − faext = 0 (54)
m(ξe )
0
Ω(ξe ) I
A
where ’A’ denotes the finite element assembly operator.
By the same token the discrete approximation of the Kuhn-Tucker optimality conditions in (23) can be presented as
Dφ L(φ, d; λ) := 0 7→
"
#T
nel
T
nint
dNa (ξe )
T
X
1
0
0
Λ
(ξ
)
E(ξ
)
0
e
e
φ,e
ds
ra =
a (ξe )
e=1
0
Ω(ξe ) I
0
Na (ξe )1 dNds
1
e=1
T
C 0
Λ (ξe ) E(ξe ) 0
+
0 D
0
Ω(ξe ) I
0
ΛT (ξe )N T (ξe )
0
+ N(ξe )ΛT (ξe ) Ξ(ǫ × n) + Ξ(ω × m) M
0
0
0
)
#
"
nen
dNb (ξe )
X
1
0
0
λ
ϕb
ds
j(ξe )wl = 0 (55)
λθb
0
Nb (ξe )1 dNds(ξe ) 1
A
b=1
Similarly, for the simple choice of the objective function in (30) the discrete ap13
proximation of the optimality condition in (25) can be written as
Dd L(φ, d; λ) := 0 7→
"
#T
nel
T
nint
dNa (ξe )
X
1
0
0
ΛT (ξe ) E(ξe ) 0
d,e
ds
ra =
a (ξe )
e=1
0
Ω(ξe ) I
1
0
Na (ξe )1 dNds
e=1
#T
ndn
ndn
X
X
∂A(ξe )
Bb (ξe )db j(ξe )wl = 0
(56)
Bb (ξe )db j(ξe )wl +
∂d b=1
b=1
A
In summary, the discrete approximation of the optimal design problem reduces to
the following set of nonlinear algebraic equations where the unknowns are the nodal
values of displacements and rotations, the corresponding values of Lagrange multipliers and the chosen values of thickness parameters which support Bézier curve
thickness interpolations,
λ
r := f int − f ext
φ
=0
r(φ, d; λ) := r := Kλ
(57)
T
int
∂f
∂V
d
r :=
+ ∂d
∂d
Similar procedure leads to the discrete approximation of the optimal control problem. In fact, the first of the nonlinear algebraic equations in (54) is only slightly
modified to include the control variables in accordance with the expression in (35)
rλ := f int − F0 ν = 0
(58)
where ν are the chosen control parameters.
For the particular choice of the objective function of the control problem given in
(41), we can further obtain the discrete approximation of the optimality condition in
(39) according to
Dφ L(φ, ν, λ) = 0
⇒
rφ := φ − φ̄d + Kλ = 0
(59)
The discrete approximation of the last optimality condition in (40) can also be
written explicitly as
Dν L(φ, ν, λ) = 0
⇒
rν := αν − FT0 λ = 0
(60)
We note in passing that the last two optimality conditions can be combined to eliminate the Lagrange multipliers leading to
αν + FT0 K−1 φ = FT0 K−1 φ̄d
(61)
The last result can be combined with the equilibrium equations in (58) providing
a reduced set of equations with nodal displacements, rotations and control variables
as the only unknowns. This kind of form is fully equivalent to the arc-length solution
procedure, which is used for solving the nonlinear mechanics problems in the presence
of the critical points (e.g. see [24], [4] or [23], among others). One can recognize
that (58) is only a particular choice of the supplementary condition which is used to
stabilize the system.
14
∂n(ξe )
∂d
∂m(ξe )
∂d
!
5 Solution procedure
Two novel solution procedures are developed for solving this class of problems of
optimal design and optimal control as described next.
5.1 Diffuse approximation based gradient methods
The first solution procedure is a sequential one, where one first computes grid values
of the cost function and then carry out the optimization procedure by employing the
approximate values interpolated from the grid. It is important to note that all the grid
values provide the design or control variables along with the corresponding mechanical state variables of displacements and rotations which must satisfy the weak form of
equilibrium equation. In other to ensure this requirement, for any grid value of design
or control variables one also has to solve associated nonlinear problem in structural
mechanics.
The main goal of the subsequent procedure is to avoid solving these nonlinear mechanics problems for other but grid values, and simply assume that the interpolated
values of the cost function will be ”sufficiently” admissible with respect to satisfying
the equilibrium equations. Having relaxed the equilibrium admissibility requirements
we can pick any convenient approximation of the cost function, which will simplify
the subsequent computation of the optimal value and thus make it much more efficient.
These interpolated values of the cost function can be visualized as a surface (yet referred to as the response surface) trying to approximate sufficiently well the true cost
function. The particular method which is used to construct the response surface of
this kind is the method of diffuse approximations (see [22] or [2]). By employing the
diffuse approximations the approximate value of the cost function Jˆappr is constructed
as the following quadratic form
1
Jˆappr (x) = c + xT b + xT Hx
2
where c is a constant reference value, b = (bi ) ; bi =
∂ 2 Jˆappr
∂xi ∂xj
∂ Jˆappr
∂xi
(62)
is the gradient and H =
is the Hessian of the approximate cost function. In (62) above
[Hij ] ; Hij =
variables x should be replaced by either design variables d for the case of an optimal
design problem or by control variables ν in the case when we deal with an optimal
control problem.
We further elaborate on this idea for a simple case where only 2 design or control
variables are used, such that x = (x1 , x2 )T . For computational proposes in such a
case one uses the polynomial approximation typical of diffuse approximation (see
[2]) employing the chosen quadratic polynomial basis p(x) and a particular point
dependent coefficient values a(x)
Jappr (x) = pT (x)a(x) ; pT (x) = [1, x1 , x2 , x21 , x1 x2 , x22 ] ; a(x) = [a1 (x), a2 (x), . . . , a6 (x)]
(63)
By comparing the last two expressions one can easily conclude that
a4 a5
a2
(64)
; H=
c = a1 ; b =
a5 a6
a3
15
The approximation of this kind is further fitted to the known, grid values of the cost
function; J(xi ), i = 1, 2, . . . , n, trying to achieve that the point-dependent coefficients
remain smooth when passing from one sub-domain to another. This can be stated as
the following minimization problem:
n
1X
T
∗ 2
a(x) = arg min
f
(a
,
x)
;
f
(a
,
x)
:=
W
(x,
x
)
J(x
)
−
p
(x
)a
(65)
i
i
i
∀a∗
2 i=1
∗
∗
where W (x, xi ) are the weighting functions associated with a particular data point
x, which are constructed by using a window function ρ(·) based on cubic splines
according to
k x − xi k
W (x, xi ) = ρ
; ρ(s) := 1−3s2 +2s3 ; r(x) = max[dist(x, xk )] (66)
k
r(x)
with xk (x), k = 1, . . . n + 1 (= 7 for the present 2-component case) as the closest
grid nodes of the given point x. We can see that the weighting functions W (·) in
(65) and (66) above take a unit value at any of the closest grid nodes xi and vanish
outside the given domain of influence. While the former assures the continuity of the
coefficients a(x), the latter ensures that the approximation remains local in character.
Similar construction can be carried out for higher order problems, which requires an
increased number of closest neighbors in the list.
By keeping the chosen point x fixed and considering the coefficients a of diffuse
approximation as constants, the minimization of f (·) amounts to using the pseudoderivative of diffuse approximation (see [22]) in order to compute x yielding the minimum of Japp (x) according to
0=
n
X
p(xi ) W (x, xi )(Ji − pT (xi )a)
(67)
i=1
which allows us to write a set of linear equations
a(x) = (PWPT )−1 PWj
P = [p(x1 ), p(x2 ) . . . p(xn )] ; j = (J(x1 ), J(x2 ), . . . , J(xn ))
W = diag (W (x1 , x), W (x2 , x), . . . , W (xn , x))
(68)
We note in passing that the computed minimum value of Jˆappr does not necessarily
provide the minimum of the true cost function, which also has to satisfy the equilibrium equations; however, for a number of applications this solution can be quite
acceptable. If the latter is not sufficient, we ought explore an alternative solution procedure capable of providing the rigorously admissible value of any computed minima
of the cost function, by carrying out the simultaneous solution of the cost function
minimization and the equilibrium equations. The proposed procedure is based on genetic algorithm as described next.
5.2 Genetic algorithm based method
Genetic algorithms belong to the most popular optimization methods nowadays. They
follow up an analogy of processes that run in nature within the evolution of living
16
organisms over a period of many millions of years. Unlike the classic gradient optimization methods, genetic algorithms operate on so called population which is a
set of possible solutions, applying the genetic operators, such as cross-over, mutation
and selection. The principles of genetic algorithm was first proposed by Holland [6].
Ever since, the genetic algorithms have reached wide application domain (e.g. see the
books of Goldberg [5] and Michalewicz [21] for extensive review).
Genetic algorithms in the original form operate on population of so-called chromosomes. These are binary strings which represent possible solutions in a certain way.
In engineering problems we are usually working with real variables, which in the kind
of applications described further are optimized values of load control or design variables. Adaptation of the genetic algorithm idea to this problem is made possible by
Storn [28] by considering the chromosomes as vectors instead of binary strings and
using differential operators which can affect the distance between the chromosomes.
In this work we employ an improved version of this kind of algorithm referred to as
’Simplified Atavistic Differential Evolution’, or SADE ([18]). It is shown [7, 8] that
the algorithm is well suitable for dealing with a fairly large number of variables which
is of particular interest for the problem on hands. The SADE algorithm is designed
to explore all possible minima and thus find the global minimum, even for the case
where the cost function may have very steep gradients and isolated peak values. Only
a short description of the corresponding procedure is given subsequently (see [7] for a
more elaborate presentation).
In the tradition of evolutionary methods, the first step is to generate a starting generation of chromosomes by choosing the random values of all state variables along
with those for control or design. Subsequently we repeat until convergence the cycles
containing: the creation of a new generation of chromosomes, by mutation, local mutation or cross-over and the evaluation and selection, which reduces the actual number
of chromosomes to the initial number.
In the computations to follow we will work with the population of ’10×n’ chromosomes, where ’n’ is the total number of unknowns in the problem. This population can
evolve through the following operations; Mutation: let xi (g) be the i-th chromosome
in a generation g,
xi (g) = (xi1 (g), xi2 (g), ..., xin (g)),
(69)
where n is the number of variables of the cost function. If a certain chromosome
xi (g) is chosen to be mutated, a random chromosome RP is generated from function
domain and a new one xk (g + 1) is computed using the following relation:
xk (g + 1) = xi (g) + MR(RP − xi (g)).
(70)
Parameter MR is the constant of algorithm equal to 0.5. Number of new chromosomes
created by the operator ’mutation’ is defined by ’radioactivity’, which is another parameter of the algorithm, with a constant value set to 0.1 for all our calculations. If
a certain chromosome is chosen to be locally mutated, all its coordinates have to be
altered by a random value from a given (usually very small) range. Its aim is to utilize the near neighborhood of existing chromosomes and make search for improved
solutions faster. It is useful only for cost functions with steep gradients in the case,
where near the optimal function value, a small change in value of variable may introduce a large change in function value. The aim of the cross operator is to create
17
as many new chromosomes as there were in the last generation. The operator creates
new chromosome xi (g +1) according to the following sequential scheme: choose randomly two chromosomes xq (g) and xr (g), compute their difference vector, multiply it
by a coefficient CR and add it to the third, also randomly selected chromosome xp (g),
i.e,
xi (g + 1) = xp (g) + CR(xq (g) − xr (g)).
(71)
Every component exceeding the defined interval is changed to the appropriate
boundary value of domain. The parameter CR has probably the most important effect
on the algorithm’s behavior; it seems that if the speed up of the convergence is needed,
this parameter should be set to some lower value (from 0.05 to 0.1). In the opposite
case, the higher values of this parameter could improve the ability to solve for any
local minimum. In our computations this value was set to 0.3.
Selection represents the kernel of the genetics algorithm. The goal is to provide
a progressive improvement of the whole population, which is achieved by reducing the
number of ”living” chromosomes together with conservation of better ones. Modified
tournament strategy is used for this purpose: two chromosomes are chosen randomly
from a population, they are compared and the worse of them is cast off. This conserves
population diversity thanks to a good chance of survival even for badly performing
chromosomes.
It was observed that the SADE algorithm can be quite inefficient in the later stage
of the analysis, where the solution with all or a large number of components is almost converged. For that reason we have tried to further improve the performance
by forcing the algorithm to stick better to these converged values. In particular we
have experimented with a modified form of the cross operator, which contrary to the
one in (71) will produce the new chromosome by building on top of the best possible
previous value according to
xi (g + 1) = max(xq (g); xr (g)) + SG.CR(xq (g) − xr (g)).
(72)
where SG is the sign change parameter which is supposed to get the correct orientation
of the increase with respect to the gradient. Moreover, the parameter CR is no more
constant, but its values are chosen randomly from interval (0, CL), where CL is new
parameter of the algorithm, with a smaller influence at its behavior than CR. Operator
local mutation is now switched off and parameter MR in operator mutation is also no
more constant, but chosen randomly from interval (0, 1). This algorithm modification
is subsequently referred to as GRADE.
6 Numerical examples
In this section we present several illustrative examples dealing with the coupled problems of mechanics and either optimal control or optimal design. The computations
are carried out by using a mechanics model of 2D geometrically exact beam (see [12]
or [14]) developed either within the MATLAB environment for diffuse approximation
based solution procedure or within the C ++ SADE computer code for genetic algorithms.
18
6.1 Optimal control of a cantilever structure in the form of letter
T
20
15
Final shape
10
EA = 12000
5
cGA = 5000
Initial shape
EI = 1000
F
M
0
-5
-10
0
-5
5
10
15
20
Figure 2: T letter cantilever: Initial, final and intermediate configurations
In this example we study the optimal control problem of deploying initially curved
cantilever beam in the final configuration which takes the form of letter T. See Figure
2 for initial and final configurations indicated by thick lines and a number of intermediate deformed configurations indicated by thin lines. The chosen geometric and
material properties are as follows: The diameter of the circular curved part and the
length of the flat part of the cantilever are both equal to 10; the beam cross-section
is a unit square; the chosen values of Young’s and shear moduli are 12000 and 6000,
respectively.
The deployment is carried out by applying a vertical force F and a moment M
at the end of the curved part of the cantilever. In other words the chosen control is
represented by a vector ν = (F, M)T . The desired shape of the cantilever which takes
the form of letter T corresponds to the values of force F = 40 and moment M = 205.
The optimal control problem is then defined as follows. The objective or cost function
is defined by imposing the desired shape only on displacement degrees of freedom
with
Z
1
kφ̂(ν) − φd k2 ds
(73)
J(ν) =
2 L
which is further recast in the discrete approximation setting as
n
2
el X
T e
1X
J (ν) =
le uea (ν) − uda
ua (ν) − uda
4 e=0 a=1
h
(74)
where uea (ν) are computed and uda are desired nodal displacements. We note in passing that no condition is imposed through the cost function on either rotational degrees
19
of freedom or control vector, which nevertheless introduces no difficulties in solving
this problem.
The first solution is obtained by the diffuse approximation based gradient method.
The calculation of the cost function is first carried out for all the ’nodes’ of the following grids: 5 × 5; 10 × 10; 15 × 15 and 20 × 20. The gradient type procedure is
then started on the grid and, thanks to the smoothness of the diffuse approximation
base representation of the approximate value of the cost function, converged with no
difficulty in roughly 20-40 iterations. The iterative values obtained in the gradient
method computations are for different grids shown in Figure 3. Grid is constructed for
the following interval of values of force and moment:
F ∈ h10, 60i ;
M ∈ h175, 225i.
We note that all different choices of the grid can result in different solutions, since
none of the solutions of this kind will satisfy the equilibrium equations.
Quite large difference between known optimal solution and solutions of response
surface could be explained by different influence of each control variable to value of
cost function, see Figure 4. All used grids weren’t able to describe small influence of
force F .
The second solution method used for this problem employs the genetic algorithm
based computation. In this computation we have used the same admissible intervals
like in the previous case for the control variables, force and moment. The computations are carried out starting from the random values chosen in this interval and
stopped when the first value of the cost function J ≤ 10−7 is found.
In order to be able to look into statistics, one hundred runs are performed with each
one converging to the exact solution. Three types of procedures were tried, with either
tuning on or off the parameters controlling local mutation (LM) and the gradient-type
acceleration of sign change (SG). The results are presented in Table 1.
Computation
with SADE
LM-on SG-off
LM-off SG-off
LM-off SG-on
Letter T problem
Number of fitness calls
Minimum Maximum Mean Value
240
2860
648.8
280
1680
625.4
220
1640
591.2
Table 1: T-letter cantilever: SADE algorithm performance
We can see that the best performance is achieved with the simple gradient-like
modification of SADE genetic algorithm to accelerate convergence in the latest stage.
The results obtained with this simple gradient-like modification of SADE algorithm
can further be improved by using described GRADE algorithm, with a special role of
the radioactivity parameter. See Table 2.
The second solution procedure applied to the same example is so-called simultaneous solution of mechanics equilibrium and optimal control equations, written explicitly in (58) to (60) with value α = 0. The total numb er of unknowns in this case
20
F approchée
o
o
225 o
o
o
X0
220
225 o
o
o
o
220 o
o
o
X0 o
215
X
o
o
o
o
o
o
o
o
o
o
o
X
205
X
X
X
X
X
X
o
o
o
X
X
X
O
205
o
200
o
o
o
o
o
o
o
X
210
o
X
X
o
o
o
o
X
210
200 o
o
o
X
X
215
o
F approchée
o
o
o
o
X o
X
X
o
X
o
X
X
o
X
o
X
X
o
X
X
o
X
O
X
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
195
195
190
190
o
o
o
o
o
185
185
o
o
o
o
o
o
o
o
o
o
180
180 o
o
o
o
o
o
o
o
o
o
175 o
10
o
15
o
50
o
55
o
60
175 o
10
15
o
20
25
o
35
30
40
o
45
50
o
60
55
Grid (5 × 5)
Solution : F = 60.000
M = 205.26
38 evaluations of AD
o
o
o
o
o
F approchée
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
X0
220
o
o
o
o
o
215 o
o
o
o
o
o
o
o
210
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
X
o
o
o
o
o
o
o
o
o
X
o
o
o
o
o
o
o
o
o
o
o
o
205
X
X
o
o
o
o
o
X
X
X
X
X
X
o
o
o
25
o
30
o
35
o
40
45
Grid (10 × 10)
Solution : F = 59.073
M = 204.91
38 evaluations of AD
225 o
o
20
o
X
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
220 o
o
o
o
o
o
X0
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
215 o
o
o
o
o
o
oX
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
Xo
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
oX
o
o
o
o
o
o
o
o
o
o
o
o
X X
O
o
o
o
o
o
o
210
o
XO
X
X
205
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
200 o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
195
F approchée
o
o
o
225 o
200
195
X
o
o
o
o
o
o
o
o
o
o
o
o
X
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
X
X
o
X
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
190 o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
190
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
185 o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
180 o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
175 o
10
o
o
15
o
o
20
o
o
o
o
o
o
o
o
40
o
45
o
o
50
o
o
55
o
o
60
185
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
180
o
o
175 o
10
o
o
15
o
20
o
o
o
o
o
25
o
30
o
o
o
o
o
35
o
o
40
o
Grid (15 × 15)
Solution : F = 51.218
M = 204.95
19 evaluations of AD
45
o
o
o
o
o
50
o
55
o
o
o
o
60
25
30
35
Grid (20 × 20)
Solution : F = 47.444
M = 204.97
15 evaluations of AD
Figure 3: T letter cantilever: Gradient method iterative computation on a grid.
is equal to 44, with 2 control variables (the force and the moment), 21 components
of nodal displacements and rotations and 21 Lagrange multiplier. For computational
convenience, the problem of solving the set of nonlinear algebraic equations in (58)
to (60) is recast as the minimization statement which allows direct application of the
genetic algorithm with
λ
r
(75)
r := rφ = 0 ⇔ min rT r
(ϕ,ν,λ)
rν
The solution efficiency of the proposed simultaneous procedure depends on the
chosen upper and lower bounds of the admissible interval and the initial guess for the
21
0
−5
−0.005
negative control function J
netative control function J
0
−10
−15
−20
−25
230
−0.01
−0.015
−0.02
−0.025
230
220
30
180
20
170
50
200
40
190
60
210
50
200
moment M
220
60
210
moment M
40
190
30
180
force F
20
170
10
Whole scale of values.
force F
10
More detailed about value of optimum.
Figure 4: T letter cantilever: contour of the cost function.
Computation
with GRADE
radioactivity = 0.0
radioactivity = 0.1
radioactivity = 0.2
radioactivity = 0.3
Letter T problem
Number of fitness calls
Minimum Maximum Mean Value
180
880
410.0
240
1200
440.8
280
1180
512.4
280
1300
588.4
Table 2: T letter cantilever: GRADE algorithm performance
solution. For example, the mechanics state variables are chosen as these featuring in
the desired beam shape, ϕd , and the bounds are controlled by the chosen parameter
EP according to
(76)
ϕ ∈ (1 − EP )ϕd , (1 + EP )ϕd
The equivalent bounds on control variables are obtained from the precious result
obtained by solving the grid minimization problem which results with the minimum of
the response surface. Finally, the Lagrange multipliers is solved for from (59) where
the adopted values for ϕ and ν are chosen.
One hundred computations is performed with the indicated bounds on the unknowns. Value of EP parameter is set to 0.0001. Choice of parameters of GRADE
algorithm were: P R = 30, CL = 2 and ’radioactivity’ equal to 0.2. Table 3 summarizes the statistics of this computation.
Minimum Maximum Mean Value
F
39.957
40.048
40.000
M
204.86
205.15
205.00
number of evaluations of J(·)
100740
298080
172176
Standard deviation
0.0199
0.0548
———
Table 3: T letter cantilever : solution statistics
The last part of the analysis carried out in this examples concerns an attempt to
further increase the efficiency of the simultaneous solution procedure. In that sense,
22
we employ the GRADE version of genetic algorithm with the choice of parameters
P R = 20, CL = 1 and ’radioactivity’ equal to 0.2. A very small value of bounds
is chosen as well with EP = 0.00001. The computations we have performed with a
hundred runs of the genetic algorithm can be summarized with the statistics given in
Table 4.
Minimal Maximal Mean Value
F
39.973
40.034
40.000
M
204.96
205.05
205.00
number of evaluations of J(·)
14720
201480
37701
Standard deviation
0.0135
0.0192
———
Table 4: T Letter cantilever : solution statistics
We can see from Table 4 that the proper choice of the bounds can force the algorithm to always converge to the same solution. The latter is the consequence of
using the simultaneous solution procedure which assures that the computed solution
also satisfies the equilibrium equations. Moreover, the total cost of the simultaneous
solution procedure can be reduced beyond the one needed for response-surface-based
approximate solution computations, by either reducing the interval as done herein or
by making gradient-type modification of this algorithm in order to accelerate the convergence rate.
6.2 Optimal control of a cantilever structure in form of letter I
In the second example we deal with a problem which has a multiple solution and
its regularized form which should restore the solution uniqueness. To that end, a
cantilever beam is used very much similar to the one studied in the previous example,
except for a shorter straight bar with length equal 2. The cantilever is controlled
by a moment M and a couple of follower forces which follow the rotation of the
cross-sections to which they are attached. The initial and final configuration, which
is obtained for a zero couple and a moment M = 205.4 are shown in Figure 5, along
with a number of intermediate configurations.
The first computation is performed with the cost function identical to the one in
(74), imposing only the minimum of difference between the desired and the computed
deformed shape, with no restriction on control variables. The computation is carried
out by using the GRADE genetic algorithm by starting with random values within the
selected admissible intervals for the force couple and moment according to
F ∈ h0, 20i ;
M ∈ h0, 230i
The algorithm performance is illustrated in Table 5.
A number of different solutions have been obtained with different computer runs
which were performed (see Figure 6). However, all these solutions remain in fact
clearly related considering that the applied moment and the force couple play an equivalent role in controlling the final deformed shape; It can be shown for this particular
problem that any values of force and moment which satisfy a slightly perturbed version (because of the straight bar flexibility) of the force equilibrium
F · h + M = M̄ = 205.4
23
20
15
Final shape
el.4
Shape after 50% of loading
10
5
F
el.3
M
F
EA = 12000
cGA = 5000
EI = 1000
el.2
Initial shape
el.1
F
0
F
-5
-10
0
-5
5
M
10
15
20
Figure 5: I letter cantilever: initial, final and intermediate configurations
Letter I problem
Number of fitness calls
Minimum Maximum Mean Value
180
640
359.8
Table 5: I letter cantilever: GRADE algorithm performance
will be admissible solution, thus we have infinitely many solutions for the case where
only the final shape is controlled by the choice of the cost function.
In order to eliminate this kind of problem we can perform a regularization of the
cost function, by requiring not only that the difference between computed and final
shape be minimized but also that control variables be as small as possible; Namely,
with a modified form of the cost function
n
2
el X
T
1X
J (ûa (ν), ν) =
le ûa (ν) − uda
ûa (ν) − uda + αν T ν
4 e=0 a=1
h
(77)
where α is a chosen weighting parameter specifying the contribution of the control.
We set a very small value α = 10−9 and choose the convergence tolerance to 10−12
and carry out the computation for yet a hundred times. Whereas a more stringent
value of the tolerance requires somewhat larger number of cost function evaluation,
the result in each of the runs remains always the same, given as
F = 68.466 ;
M = 68.526
and the found optimal value of the cost function is J = 1.4078631.10−5.
24
results of GRADE algorithm
linear regression for results values: F = 102.63 - 0.49989.M
force F
100
50
0
0
50
100
moment M
200
150
Figure 6: I letter cantilever: 100 different solutions
Letter I problem extended
Number of fitness calls
Minimum Maximum Mean Value
820
16320
1758.8
Table 6: I letter cantilever: GRADE algorithm performance
6.3 Optimal control of deployment of a multibody system
The optimal control procedure of the deployment problem of a multibody system is
studied in this example. In its initial configuration the multibody system consists of
two flexible component (with EA = 0.1, GA = 0.05 and EI = 103 ) each 5 units long
connected by the revolute joints (see [15]) to a single stiff component (with EA = 1,
GA = 0.5 and EI = 105 ) of the length equal to 10, which are all placed parallel
to horizontal axis. In the final deployed configuration the multibody system should
take the form of a letter B with the stiff component being completely vertical and two
flexible component considerably bent. The deployment of the system is controlled by
five control variables, three moments M1 , M2 and M3 , a vertical V and a horizontal
force H. See Figure 7.
The cost function in this problem is again chosen as the one in 74, which controls
that the system would find the configuration as close as possible to the desired configuration. The desired configuration of the system corresponds to the values of forces
H = 0.04, V = −0.05 and moments M1 = 0.782, M2 = −0.792 and M3 = 0.792.
The solution is computed by using the SADE and GRADE genetics algorithms and
starting with the random choice in the interval of interest defined as
H ∈ h0.025; 0.05i,
V ∈ h−0.06; −0.035i,
25
M1 ∈ h0.6; 0.9i,
12
EA1 = 1.0 EA2 = 0.1
cGA1 = 1.0 cGA2 = 5.0
EI1 = 1.05 EI2 = 1000.0
10
8
Final shape
6
4
}
EA1
cGA1
EI1
{
EA2
cGA 2
EI2
Shape after 20% of loading
M3
M1
2
M2
H
0
Initial shape
V
-2
0
5
10
Figure 7: Multibody system deployment: initial, final and intermediate configurations.
M2 ∈ h−0.9; −0.65i,
M3 ∈ h0.6; 0.85i.
The solution of the problem is typically more difficult to obtain with an increase in
the number of control variables, one of the reasons being a more irregular form of the
cost function. In that sense,we refer to illustrative representation of the cost function
contours in different subspaces of control variables as shown in Figures 8 and 9.
The convergence tolerance on cost function is chosen equal to 10−6 . The SADE
algorithm performance for the simplest choice of algorithm parameters is presented
in Table 7 and the GRADE algorithm performance for modified value of radioactivity
parameter is presented in Table 8.
Computation
with SADE
LM-on SG-off
LM-off SG-off
LM-off SG-on
Letter B problem
Number of fitness calls
Minimum Maximum Mean Value
2600
165800
46887.5
2350
177150
42494.0
2400
177850
34612.1
Table 7: Results of SADE algorithm for 5-variable control problem
One can notice the order of magnitude of increase in cost function evaluation,
which is brought about by a more elaborate form of the cost function (see Figures
8 and 9). However, the latter is not the only reason. In this particular problem the role
of moments in the list of control variables is much more important than the role of the
horizontal and vertical forces in bringing the system in the desired shape. This affects
the conditioning of the equations to be solved and the slow convergence rate of the
complete system is in reality only the slow convergence of a single or a couple of control components. The latter is illustrated in Figure 10, where we provide the graphic
26
0
0
−0.001
−0.1
−0.002
−0.003
−0.3
control function J
control function J
−0.2
−0.4
−0.5
−0.6
−0.004
−0.005
−0.006
−0.007
−0.7
−0.008
−0.8
−0.035
−0.009
0.05
−0.04
−0.01
−0.035
0.05
−0.045
−0.045
0.04
−0.05
force V
0.045
0.04
−0.04
0.045
0.03
−0.06
−0.06
0.025
force H
force V
H − V subspace
H − V subspace
0
0
−5
−0.01
−10
control function J
control function J
0.03
−0.055
force H
0.025
0.035
−0.05
0.035
−0.055
−15
−20
−0.02
−0.03
−0.04
−25
−30
1
−0.05
1
0.05
0.05
0.045
0.8
0.045
0.8
0.04
0.04
0.035
moment M1
0.035
0.03
0.6
0.025
0.03
moment M1
0.6
force H
H − M1 subspace
0.025
force H
H − M1 subspace
0
−0.01
0
−0.02
−0.03
control function J
control function J
−5
−10
−15
−20
−0.04
−0.05
−0.06
−0.07
−0.08
−25
−0.09
−0.1
1
−30
1
−0.035
−0.04
0.8
−0.035
0.8
−0.04
−0.045
−0.045
−0.05
moment M1
−0.055
moment M1
0.6
−0.06
force V
−0.05
0.6
−0.055
−0.06
force V
V − M1 subspace
V − M1 subspace
Figure 8: Multibody system deployment: contours of the cost function in different
subspaces.
representation of iterative values for computed chromosomes, where every chromosome is represented by a continuous line. We can note that the population of optimal
values of moments converges much more quickly than the force values which seek
a large number of iteration in order to stabilize. Another point worthy of further ex27
0
0
−0.2
−10
−0.4
negative control function J
−20
control function J
−30
−40
−50
−60
−0.6
−0.8
−1
−1.2
−1.4
−1.6
−70
−1.8
−80
−2
−90
−0.65
−100
−0.65
−0.7
−0.75
−0.7
−0.75
−0.8
−0.8
−0.85
moment M2
−0.9
0.55
0.6
0.65
0.7
0.8
0.75
0.9
0.85
moment M2
0.95
−0.85
−0.9
0.55
0.65
0.6
0.7
0.8
0.75
0.95
0.9
0.85
moment M1
moment M1
M1 − M2 subspace
M1 − M2 subspace
0
−10
−20
0
−0.2
−30
negative control function J
control function J
−0.4
−40
−50
−60
−70
−0.6
−0.8
−1
−1.2
−1.4
−1.6
−80
0.9
−1.8
−90
−2
0.55
1
−100
0.55
0.6
0.65
0.7
0.8
0.75
0.8
0.85
0.9
moment M1
0.95
0.6
0.8
0.7
0.6
moment M3
0.65
0.7
0.75
0.6
0.8
0.85
moment M1
M1 − M3 subspace
0.9
0.95
0.5
moment M3
M1 − M3 subspace
0
control function J
−20
−40
0
−60
−0.2
control function J
−80
−100
0.9
0.8
−0.4
−0.6
0.55
−0.8
0.6
0.65
0.7
moment M3
−1
−0.65
0.7
0.75
−0.7
0.6
0.5
−0.9
−0.85
−0.8
−0.75
−0.7
0.8
−0.75
−0.65
−0.8
moment M2
0.85
−0.85
−0.9
0.9
moment M3
moment M2
M2 − M3 subspace
M2 − M3 subspace
Figure 9: Multibody system deployment: contours of the cost function in different
subspaces.
ploration is the best way to accelerate the convergence rate in the final computational
phase.
28
Computation
with GRADE
radioactivity = 0.0
radioactivity = 0.1
radioactivity = 0.2
radioactivity = 0.3
Letter B problem
Number of fitness calls
Minimum Maximum Mean Value
—–
—–
—–
1500
114050
23862.5
1900
117850
20632.0
3050
122550
34520.0
Table 8: Results of GRADE algorithm for 5D task
G1:J=0.0091823
G4:J=0.0071823
G26:J=1.3302e-05
G43:J=1.1839e-05
Figure 10: Multibody system deployment: convergence of iterative chromosome populations
6.4 Optimal design of shear deformable cantilever
In this last example we study an optimal design problem which considers the thickness
optimization of a shear deformable cantilever beam, shown in Figure 11.
The beam axis in the initial configuration of the cantilever and the thickness is
considered as the variable to be chosen in order to assure the optimal design specified by a cost function. In the setting of discrete approximation, we choose 4 beam
elements each with a constant thickness hi , which results with 4 design variables
29
40
F = 1000
20
h1
h2
h4
h3
0
-20
-40
-60
E = 75000
G = 50000
c = 5/6
ρ = 1/30
he
b = 30
imposed mass: Mo = 30000
-80
-100
-200
0
200
400
600
800
1000
1200
Figure 11: Shear deformable cantilever beam optimal design : initial and deformed
shapes
d ≡ h = (h1 , h2 , h3 , h4 ). The beam mechanical and geometric properties are:
Young’s modulus E = 75000, shear modulus G = 50000, rectangular cross section
b × hi with width b = 30 and mass density
ρ = 1/30. The latter is needed for computR
ing the total mass of the beam M = L ρbh(s)ds to be used as the corresponding limitation on the computed solution assuring reasonable values of the optimal thickness
under the free-end vertical force F = 1000. In order to assure a meaningful result the
computations are performed under chosen value of mass limitation is M0 = 30000.
Other limitations are also placed on the admissible values of the thickness for each
element.
The first computation is performed by using the diffuse approximation based response function and the sequential solution procedure. The cost function is selected
as the shear energy of the beam and problem is cast as maximization of the shear
energy, with
Z
1
∗
∗
∗
J(ϕ(d)) =
max
J(ϕ̂ (d )) ; J(ϕ ) =
GAγ 2 ds
(78)
∗
∗
G(ϕ (d ),·)=0
L 2
where γ is the shear strain component. The bounds on thickness values are chosen
as shown in Table 9. The diffuse approximation computations on the grid are started
Thickness h1
Min
30
Max
60
h2 h3 h4
30 15 15
60 35 35
Table 9: Shear deformable cantilever optimal design : thickness admissible values
from an initial guesses for thickness h0 = (55, 50, 30, 20). It took 11 iterations to
converge to solution given as
h = (45.094, 40.074, 19.832, 15.000)
30
The corresponding value of shear energy for this solution is Jappr = 16.3182; we recall
it is only an approximate solution, since the computed value does not correspond to
any of the grid nodes.
The same solution is next sought by using GRADE algorithm; The chosen values
for GRADE parameters are: P R = 10, CL = 1.0 and ’radioactivity’= 0.2. The genetic algorithm is executed 100 times, leading to the computational statistics reported
in Table 10.
nb. of comput. J(·)
Minimum Maximum Mean Value
120
3400
674.8
Table 10: Shear deformable cantilever optimal design : computation statistics
The algorithm yielded two different solutions, both essentially imposed by the chosen bounds; Namely, out of 100 runs, 57 converged to h = (60, 30, 15, 15), with the
corresponding value of J = 17.974856, whereas 43 settled for h = (30, 60, 15, 15)
with a very close value of J = 17.926995. Hence, each of two solutions leads to an
improved value of the cost function.
The second part of this example is a slightly modified version of the first one, in
the sense that the mechanics part of the problem is kept the same and only a new
cost function is defined seeking to minimize the Euclidean norm of the computed
displacement vector, i.e.
Z
1
∗
∗
∗
J(ϕ(d)) =
max
J(ϕ̂ (d )) ; J(ϕ ) =
kϕ − xk2 ds
(79)
G(ϕ∗ (d∗ ),·)=0
2 L
Such a choice of the cost function is made for being well known to result with a
well-conditioned system expressing optimality conditions. Indeed, the same type of
sequential solution procedure using diffuse approximation of cost function now needs
only a few iterations to find the converged solution, starting from a number initial
guesses. The final solution value is given as
h = (42.579, 35.480, 26.941, 15.000)
.
In the final stage of this computation we recompute the solution of this problem by
using the genetic algorithm. The admissible value of the last element thickness is also
slightly modified by reducing the lower bound to 5 (instead of 15) and higher bound
to 25 (instead of 35) in order to avoid the optimal value which is restricted by a bound.
The first solution to this problem is obtained by using again the sequential procedure,
where the GRADE genetic algorithm is employed at the last stage. The computed
value of the displacement vector norm for found solution is 623808 and mass M is
30062. The computations are carried out a hundred times starting from random initial
values. The statistics of these computations are given in Table 11.
The same kind of problem is now repeated by using the simultaneous solution procedure, where all the optimality condition are treated as equal and solved simultaneously resulting with 4 thickness variables, 15 displacement and rotation components
and as many Lagrange multipliers as unknowns. The latter, in fact, is eliminated prior
31
h1
h2
h3
h4
nb. of comput. J(·)
Minimum Maximum Mean Value Standard deviation
43.772
43.807
43.790
0.0094
35.914
35.949
35.932
0.0088
26.313
26.346
26.328
0.0082
14.184
14.210
14.197
0.0064
1440
9960
3497
——–
Table 11: Shear deformable cantilever optimal design : computation statistics
to solution by making use of optimality condition in (57)2 . The chosen upper and
lower bounds of the admissible interval are chosen as
ϕ ∈ [(1 − EP )ϕp , (1 + EP )ϕp ]
(80)
where the guess for the displacement ϕp is obtained by solving mechanics problem
with the values of thickness parameters given in Table 11. The limitation on total mass
is added to the cost function. The choice of GRADE algorithm parameters is given
as P R = 20, CL = 2 and ’radioactivity’ equal to 0.1. The computation is stopped
with a fairly loose tolerance, which allows to accelerate the algorithm convergence but
does not always lead a unique solution. Yet, the results in Table 12 show that standard
deviation indeed remains small, or that the solution is practically unique.
Minimum Maximum Mean Value Standard deviation
h1
43.782
43.794
43.789
0.0026
h2
35.925
35.935
35.930
0.0021
26.315
26.324
26.319
0.0019
h3
h4
14.197
14.202
14.200
0.0010
T
nb. of comput. r r
111340
968240
313006
——–
Table 12: Shear deformable cantilever optimal design : simultaneous computation
statistics
7 Acknowledgements
This work was supported by the French Ministry of Research and the European Student Exchange Program ERASMUS. This support is gratefully acknowledged.
8 Conclusions
The approach advocated herein for dealing with a coupled problem of nonlinear structural mechanics and optimal design or optimal control, which implies bringing all the
optimality conditions at the same level and treating all variables as independent rather
than considering equilibrium equations as a mere constraint and state variables as dependent on design or control variables, is fairly unorthodox and rather unexplored.
For a number of applications the proposed approach can have a great potential. In
32
particular, the problems of interest to this work concern the large displacements and
rotations of a structural systems. The key ingredient of such an approach pertains to
geometrically exact formulation of a nonlinear structural mechanics problem, which
makes dealing with nonlinearity description or devising the solution schemes much
easier then for any other model of this kind. The model problem of the geometrically
exact beam explored in detail herein is not the only one available in this classl; We
refer to work in [9] for shells or to [10] for 3D solids, with both models sharing the
same configuration space for mechanics variables as 3D beam. The latter also allows
to directly exploit the presented formulation and the solution procedures of a coupled
problem of nonlinear mechanics, for either shells or 3D solids, and optimal control or
optimal design.
Two different solution procedures are presented herein; the first one, which exploits
the response surface representation of the true cost function followed by a gradient
type solution step, leads to only an approximate solution. Although the quality of
such a solution can always be improved by further refining the grid which serves
to construct the response surface, the exact solution is never computed unless the
minimum corresponds to one of the grid points. The second solution procedure, which
solves simultaneously the optimality conditions and nonlinear mechanics equilibrium
equations, does deliver the exact solution, although often only after the appropriate
care is taken to choose the sufficiently close initial guess and to select the admissible
intervals of all variables accordingly. Probably the best method in that sense is the
combination of sequential and simultaneous procedure, where the first serves to reduce
as much as possible the admissible interval and provide the best initial guess, whereas
the second furnishes the exact solution.
A number of further improvements can be made for the proposed methods of this
kind, to help increase both the convergence rates and accuracy of the computed solution. One has to remember that even only mechanics component equilibrium equations
for a problem of this kind can be very difficult to solve, since the large difference in
stiffness in different modes (e.g. bending versus stretching) can result with a poorly
conditioned set of equations. Adding the optimality conditions on top only further
increases the difficulty. Finding the best possible way to deal with this problem is
certainly worthy of further explorations.
References
[1] J.H. Argyris. An excursion into large rotations. Comput. Methods Appl. Mech.
Eng., 32:85–155, 1982.
[2] P. Breitkopf, C. Knopf-Lenoir, A. Rasineux, and P. Villon. Efficient optimization
strategy using hermite diffuse approximation. In H. Mang, editor, Proceedings
Fifth WCCM, Vienna, 2002.
[3] D. Chenais and C. Knopf-Lenoir. From design sensitivity to code structure in
arch optimization. In Opti 89, Southampton, U.K.
[4] M. A. Crisfield. A fast incremental/iterative solution procedure that handles snap
through. Computers & Structures, 13:55–62, 1981.
33
[5] D.E. Goldberg. Genetic algorithms in search, optimization and machine learning. Addison-Wesley, 1989.
[6] J. H. Holland. Adaptation in natural and artificial systems. University of Michigan, Ann Arbor, MI, Internal report, 1975.
[7] O. Hrstka and A. Kučerová. Improvements of the different types of binary and
real coded genetic algorithms preventing the premature convergence. Advances
in Engineering Software in press, 2003.
[8] O. Hrstka, A. Kučerová, M. Lepš, and J. Zeman. A competitive comparison of
different types of evolutionary algorithms. Computers and Structures, 81/1819:1979–1990, 2003.
[9] A. Ibrahimbegović. Stress resultant geometrically nonlinear shell theory with
drilling rotations - part 1: a consistent formulation. Comput. Methods Appl.
Mech.Eng., 118:265–284, 1994.
[10] A. Ibrahimbegović. Finite element implementation of reissner’s geometrically
nonlinear beam theory: three dimensional curved beam finite elements. Comput.
Methods Appl. Eng., 122:10–26, 1995.
[11] A. Ibrahimbegović. On the choice of finite rotation parameters. Comput. Methods Appl. Mech. Engng., 149:49–71, 1997.
[12] A. Ibrahimbegović and F. Frey. Finite element analysis of linear and non-linear
planar deformations of elastic initially curved beams. International Journal for
Numerical Methods in Engineering, 36:3239–3258, 1993.
[13] A. Ibrahimbegović, F. Frey, and I. Kozar. Computational aspects of vector-like
parameterization of three-dimensional finite rotations. International Journal for
Numerical Methods in Engineering, 38:3653–3673, 1995.
[14] A. Ibrahimbegović and S. Mamouri. Nonlinear dynamics of flexible beams in
planar motions formulation and time-stopping scheme for stiff problems. Comp.
Struct., 70:1–22, 1999.
[15] A. Ibrahimbegović and S. Mamouri. On rigid components and joint constraints
in nonlinear dynamics of flexible multibody systems imployin 3d geometrically
exact beam model. Comp. Methods Appl. Mech. Eng., 188:805–831, 2000.
[16] M. Kegl. Shape optimal design of structures: an efficient shape representation
concept. Int. J. Numer. Meth. Engng., 49:1571–1588, 2000.
[17] M. Kleiber, H. Antunez, T.D. Hein, and P. Kowalczyk. Parameter sensitivity
in nonlinear mechanics; theory and finite element computations. John Wiley &
Sons, 1997.
[18] A. Kučerová and O. Hrstka. Homepage of SADE.
http://klobouk.fsv.cvut.cz/∼ondra/sade/sade.html.
34
[19] D.G. Luenberger. Linear and nonlinear programming. Addison-Wesley Publ.,
1984.
[20] J.E. Marsden and T.J.R. Hughes.
Prentice-Hall, 1983.
Mathematical foundations of elasticity.
[21] Z. Michalewicz. Genetic Algorithms+Data Structures=Evolution Programs.
Springer-Verlag, 1992.
[22] B. Nayroles, G. Touzot, and P. Villon. Generalizing the finite element method:
diffuse approximation and diffuse elements. Computational Mechanics, 10:307–
318, 2002.
[23] E. Ramm. Strategies for tracing the non-linear response near limit points. In
W. Wunderlich, editor, Nonlinear Finite Element Analysis in Structural Mechanics, pages 63–89, Berlin, 1988. Springer-Verlag.
[24] E. Riks. The application of newtons method to the problem of elastic stability.
Journal of Applied Mechanics, 39:1060–1066, 1972.
[25] B. Rousselet. A finite strain rod model and its design sensitivity. Mechanics of
Structures and Machines, 20:415–432, 1992.
[26] J.C. Simo. The (symmetric) hessian for geometrically nonlinear models in solid
mechanics: intrinsic definition and geometric interpretation. Comput. Methods
Appl. Mech. Engng., 96:183–200, 1992.
[27] J.C. Simo and L. Vu-Quoc. Three-dimensional finite strain model - part ii: computational aspects. Comput. Methods Appl. Mech. Eng., 38:79–116, 1986.
[28] R. Storn. On the usage of differential evolution for function optimization. In Biennial Conference of the North American Fuzzy Information Processing Society,
pages 519–523, 1996.
[29] G. Strang. Introduction to applied mathematics. Wellesley-Cambridge Press,
1986.
[30] D.A. Tortorelli and P. Michaleris. Design sensitivity analysis: overview and
review. Inverse Problems in Engineering, 1:71–105, 1994.
[31] O.C. Zienkiewicz and R.L. Taylor. The finite element method, volume vols I, II
& III. Butterworth, London, 2000.
35
| 5cs.CE
|
arXiv:1709.08666v1 [cs.CV] 25 Sep 2017
Fast Vehicle Detection in Aerial Imagery
Jennifer Carlet
KeyW Corp.
Bernard Abayowa
Sensors Directorate, Air Force Research Lab
Beavercreek, OH
WPAFB, OH
Abstract
modified is described in Section 4. Section 5 gives the results. The paper is concluded in Section 6.
In recent years, several real-time or near real-time object detectors have been developed. However these object
detectors are typically designed for first-person view images where the subject is large in the image and do not
directly apply well to detecting vehicles in aerial imagery.
Though some detectors have been developed for aerial imagery, these are either slow or do not handle multi-scale
imagery very well. Here the popular YOLOv2 detector is
modified to vastly improve it’s performance on aerial data.
The modified detector is compared to Faster RCNN on several aerial imagery datasets. The proposed detector gives
near state of the art performance at more than 4x the speed.
2. Deep learning object detection algorithms
Since a deep CNN easily won the 2012 ImageNet competition, CNNs have become the state of the art in object
detection in images. Whereas before hand crafted features
such gradients, color, etc., were used to detect objects; now
CNNs can automatically learn which features are relevant
for detection. This work focused on version 2 of the you
only look once (YOLO) detector[10], however it is compared to Faster RCNN, which is considered by most to be
state of the art. For testing, open-source TensorFlow[1] versions of the detectors were used[12, 2].
2.1. Faster RCNN
1. Introduction
Faster RCNN is the follow on to Fast RCNN[4] and
RCNN[5]. Faster RCNN starts with a CNN adds a Region Proposal Network (RPN) to create proposals (bounding boxes) from the features given by the CNN. Then ROI
pooling and a classifier is used to classify and score each
bounding box. A diagram of the net from the original paper
is given in Figure 1. Due to its speed and accuracy, Faster
RCNN has been heavily used since its inception in 2015.
Object detection from ground view is a popular problem with a lot of interest from the academic computer vision community. Detection from aerial views, while there
is some interest, is significantly less studied. Consequently
recent advancements are primarily large in image object detection and classification, mostly using deep convolutional
neural networks (CNNs). Often these neural networks do
not work well when directly applied to small in image objects. However the networks can often be modified to improve their performance on this type of data.
Of particular interest is detecting vehicles from aerial
platforms at near real-time speeds. While Faster RCNN[11]
has been proven to be effective at detecting vehicles in aerial
imagery, it is unable to reach anywhere near the real time
speeds desired for many applications. Other methods use
sliding window techniques which can also be slow. Newer
detectors that can run on whole images are much faster but
have yet to be proven on aerial imagery. In this paper, an
open-source fast deep CNN is modified into a near real-time
multi-scale detector for aerial imagery.
The rest of this paper is organized as follows. Section
2 introduces the deep CNNs used in this paper. Section 3
covers the aerial imagery datasets used. How the net was
Figure 1. Faster RCNN
1
2.2. YOLOv2
DLR3k Data from the German Aerospace Center
(DLR),there are 20 images taken from a camera on an airplane about 1000 feet over Munich, Germany.
YOLOv2 not surprisingly is the improved version
YOLO[9], a net that reached near Faster RCNN accuracy
at much greater speeds. YOLOv2 also starts with a CNN,
and then is followed by two more CNNs that simultaneously produce bounding boxes, object confidence scores,
and class scores. Additionally YOLOv2 includes a route
and reorg layers that allow the net to use features from earlier in the net similar to SSD[8].
NeoVision2 - Helicopter A set of 32 video clips taken
from an HD camera on helicopter flying in Los Angeles,
California area.
AFVID Video clips taken from a small UAV flying about
1000 feet over Avon Park, Florida or flying about 700 feet
over Camp Atterbury, Indiana.
3. Datasets
To train a deep neural network object detector requires
a vast amount of data; hence the coupling of deep learning with big data. The neural nets in the previous section
all provide pretrained network weights, usually started on
ImageNet and then on PASCAL VOC[3] or MS COCO[7].
Detailed in Table 1, these datasets provide thousands of images, and the detectors trained on them are good general
purpose object detectors. However these datasets largely
contain imagery taken from personal cameras at ground
level and contain very little to no aerial data. Additionally these images are relatively low resolution compared to
aerial imagery.
AF Building Camera Handful of images taken from
cameras on a tower at Wright Patterson Air Force Base.
3.2. Dataset segmentation
Much of aerial imagery is given as video clips, meaning
it can not be considered independent and identically distributed(IID), unlike the the large detection image datasets.
This means even though there may be thousands of images,
some of it may be so close in appearance that effectively
there is less data and none of our datasets are anywhere
near the size of even PASCAL VOC. Ideally each dataset
should be separated into train (60%), validation (20%), and
test (20%) sets. Typically this should only be done with
datasets that have independent images. Some of the datasets
consist of a short video sequences where at least two videos
are kept out, one for validation and one for testing.
Table 1. Public large scale detection datasets
Dataset
ImageNet
MS COCO
PASCAL VOC (2007+2012)
Number of Training Images
450000+
200000+
~20000
4. Modifying YOLOv2
There are several publicly available aerial/elevated imagery detection datasets. Additionally AFRL has some in
house aerial imagery, referred to as Air Force aerial vehicle
imagery dataset (AFVID), that has been truthed.
YOLO provides models and pretrained weights for the
MS COCO and PASCAL VOC datasets, both which are capable of detecting vehicles that are relatively large in the
image. To detect objects that are small relative to the size
of the image and to detect different numbers of classes the
YOLO net needs to be modified and fine-tuned using appropriate data.
3.1. Dataset descriptions
Aerial data can be surprisingly diverse. It can be from
different view points, have different ground sampling distances (gsd), different image sizes, aspect ratios, color, etc.
Vehicles in different data may be significantly different in
size and appearance. Figure 2 shows different images from
four different datasets. It is obvious how different the VOC
data is from any of the elevated data, while the VEDAI and
AFVID data are somewhat similarities, and the AF Building Camera data is more similar to the aerial and satellite
data. A more detailed look at the aerial datasets is available
in Table 2.
4.1. Number of classes
In the configuration file that defines the YOLO net model
there is a ’classes’ definition that is used to define the number of classes. To change the number of classes used, more
the just the classes setting will need to be changed; the number of filters in the last convolutional layer must be altered
to reflect the changed number of classes. The number of
filters is set by num(classes + coords + 1), where num is
number of anchor boxes, and coords is four corresponding
to the four coordinates used to define a bounding box.
VEDAI Vehicle Detection in Aerial Imagery (VEDAI)
dataset consists of satellite imagery taken over Utah in
2012. There are 1,210 orthonormal images in RGB and IR.
Only RGB is used here.
4.2. Net resolution and depth
The standard YOLOv2 net has input resolution of
416x416 and after pass-through and max-pooling has an
2
(a) PASCAL VOC
(b) VEDAI
(c) AFVID
(d) AF Building Camera
Figure 2. Sample imagery
Table 2. Aerial imagery datasets
Dataset
VEDAI
DLR3k
NeoVision2-Helicopter
AFVID 1
AFVID 2
AFVID 3
AFVID 4
AF Building Camera
Image Width,
Height
1024, 1024
5616, 3744
1920, 1080
1600, 1200
1600, 1200
1600, 1200
1600, 1200
1920, 1080
Mean Vehicle
Width,Height
41.2, 40.8
30.4, 30.0
124.6, 85.8
39.4, 34.2
55.6, 63.7
43.5, 38.6
67.7, 63.5
51.7, 35.5
output feature resolution of 26x26 and 13x13. For a
1024x1024 image that would create feature resolutions corresponding to about 40 and 80 pixels, which is approximately the size of and twice the size of the vehicles in the
VEDAI imagery, meaning a vehicle may correspond to a
single point on a feature map. Ideally there should be multiple points per vehicle. To increase net feature resolution
and consequently the number of feature points per vehicle,
there are two methods: increase the input resolution or to
decrease the net depth.
Ratio Target
Area/Image Area
0.0016
0.00004
0.0069
0.0008
0.0019
0.0009
0.0026
0.0012
Percent of Targets
Overlapping
5
14
20
33
54
40
34
53
ing the width and height of the first layer of the net. In the
previous example doubling the width and height leads to
output sizes of 52x52 and 26x26 and 20/40 pixels per feature, so that there can be about four feature points for the
average vehicle. However, this significantly increases GPU
memory usage and decreases the speed of the net.
To decrease the net depth, convolutional and max pooling layers are removed so the net is downsampled less, making the net shallower. This goes against the now conventional wisdom that deeper nets are typically better, but has
been shown to work well on aerial data[6, 13]. Remov-
Increasing the net input resolution simply means increas3
ing one max pooling layer and associated CNN layers gives
the same output resolution as doubling the input resolution
without having as great as an effect the memory usage or
speed of the net. A sample shallower YOLO net is given in
Appendix A Table 8b alongside a typical net.
large and somewhat diverse they make good general purpose object detectors. The results for these detectors tested
on aerial imagery is given in Table 3. While Faster RCNN
gives the best precision and recall, it is the slowest of the
tested detectors. All of the detectors performed quite poorly
without fine-tuning for aerial data.
4.3. Net shape
5.2. YOLO modifications
YOLO provide pretrained square shaped nets, however
most of our data is not square. In YOLOv2, the shape of the
net can be changed to closer match the aspect ratio of the
input data. Faster RCNN does this automatically.
The biggest increases in performance come from altering the net’s size and depth. Table 4 compares the performance of results of several modified YOLO nets and a
shallow Faster RCNN net based on [13]. All the nets were
fine-tuned using AFVID and VEDAI data. The YOLO nets
were based off of the YOLO-VOC net but Faster RCNN
was from COCO. The first YOLO listed is the standard
YOLO modified for one class (vehicles). While it is five
times faster than Faster RCNN, the performance across the
board is much weaker. Doubling the size as described in
4.2, gives a performance boost, particularly on the AFVID
dataset, but halves the speed. Removing several convolutional layers and a max pooling layer greatly improves precision, recall, and speed. With the standard input resolution
but shallower, YOLO is approximately seven times faster
the shallower Faster RCNN, but still has lower recall and
lower precision on the AFVID and VEDAI datasets. Increasing the size of the shallower net decreases the speed of
the net again, but still boosts precision and recall. Changing
the input shape of the shallow YOLO net to better match
the aspect ratio of the AFVID data causes a large increase
in the precision for that data, but causes a small decrease on
the square VEDAI data. On the AF Building Camera data,
which none of the nets were trained on, the shallow YOLO
nets do better than Faster RCNN.
One application is to use this CNN network on live data
from the AF Building Cameras, which has a larger variance in vehicle sizes than most aerial datasets. The nets
were trained on the AFVID data and a few images from the
AF Building Cameras. The results in Table 5, show that
rectangular shallow YOLO outperforms Faster RCNN on
this data, while the shallow Faster RCNN is still better on
AFVID and VEDAI data.
4.4. Anchors
In [13], the Faster RCNN anchors were fixed, as the
bounding boxes were refined to be of fixed shape and size.
For multi-scale detection, it should be better to have multiple anchor box sizes. The five anchor sizes in YOLOv2
were determined from a k-means approach using the bounding boxes in the VOC dataset. Due to the difference in orientation and scale of vehicles, most of the anchors were
kept, with only the largest being removed, since it is not
expected that there will be large objects in the image.
5. Experiments and results
To calculate how the detectors are performing, two metrics are used:
precision = tp/(f p + tp)
(1)
recall = tp/nObjects.
(2)
and
In defense applications, the false alarm rate (FAR) is often
used instead of precision but
F AR = 1 − precision,
(3)
and recall is sometimes referred to the detection rate. The
another metric used here is frames per second (FPS). To
match detections to ground truth, intersection over union
(IOU) is used. IOU is ratio between the area in which the
two boxes overlap and the total area of each box not including the overlap.
A detector was desired for aerial imagery with small vehicles. Only a single class, vehicle, was used in training
and testing. Therefore average precision (AP) and average
recall (AR) refer to the average over all the test images at
an IOU threshold of 0.5.
5.3. Deeper look at results
For cases where the input to the net is an entire image, the object size relative the size of image is more important to than the size of the object itself. For example,
in the MSCOCO dataset images are typically 640x480 and
the benchmark define a small object to be less than 32x32
a medium object between 32x32 and 96x96. While most
of the mean object sizes in the aerial imagery fall in the
medium category by this definition, the image sizes are
larger so when resized to the input resolution of the nets
they equivalently fall into the small category. Hence the cat-
5.1. Pretrained object detectors
Most open source detectors provide their best trained for
PASCAL VOC and/or MS COCO. Since these datasets are
4
Table 3. Pretrained object detectors tested on aerial data
Net - trained
YOLO - VOC+COCO
YOLO - VOC
Faster RCNN - COCO
AFVID
AP: 0.0 AR: 0.03
AP: 0.0 AR: 0.0
AP: 0.04 AR: 0.18
VEDAI
AP: 0.0 AR: 0.03
AP: 0.0 AR: 0.01
AP: 0.01 AR: 0.12
AF Building Camera
AP: 0.05 AR: 0.07
AP: 0.04 AR: 0.04
AP: 0.13 AR: 0.19
FPS
16
15
3
Table 4. Results fine-tuned on AFVID and VEDAI data
Net
Shallow Faster RCNN
YOLO
YOLO
Shallow YOLO
Shallow YOLO
Shallow YOLO
Input Size
600xA
416x416
832x832
416x416
832x832
864x480
AFVID
AP: 0.88 AR: 0.89
AP: 0.0 AR: 0.03
AP: 0.01 AR: 0.13
AP: 0.1 AR: 0.37
AP: 0.21 AR: 0.79
AP: 0.76 AR: 0.87
VEDAI
AP: 0.78 AR: 0.91
AP: 0.0 AR: 0.01
AP: 0.0 AR: 0.05
AP: 0.46 AR: 0.72
AP: 0.66 AR: 0.91
AP: 0.61 AR: 0.81
AF Building Camera
AP: 0.05 AR: 0.08
AP: 0.01 AR: 0.02
AP: 0.0 AR: 0.04
AP: 0.14 AR: 0.21
AP: 0.15 AR: 0.3
AP: 0.21 AR: 0.3
FPS
3
15
6
25
8
13
G. Irving, M. Isard, Y. Jia, R. Jozefowicz, L. Kaiser,
M. Kudlur, J. Levenberg, D. Mané, R. Monga,
S. Moore, D. Murray, C. Olah, M. Schuster, J. Shlens,
B. Steiner, I. Sutskever, K. Talwar, P. Tucker, V. Vanhoucke, V. Vasudevan, F. Viégas, O. Vinyals, P. Warden, M. Wattenberg, M. Wicke, Y. Yu, and X. Zheng.
TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. Software available from tensorflow.org.
egories were redefined in terms of the image size, such that
small objects are 0.1% or less, medium objects are between
0.1% and 0.3% and large objects are greater than 0.3% of
the original image. Looking at the results in terms of size
on AF Building Camera data in Table 6 shows a surprising
result for Faster RCNN, the medium size is actually has the
worst performance. Shallow YOLO performs as expected.
The irregularity in the medium size objects in Faster
RCNN may be due to a quirk in the small AF Building
Camera dataset. Shown in Figure 3 is a set images taken
by the same camera at different times, with the last two sets
of images taken only a few minutes apart. Both detectors
do well when the cars are sparse and most have visible separation. However when the parking lot is full and the cars
are densely packed the Faster RCNN sometimes fails to detect some of the vehicles while YOLO does a much better
job, irregardless of illumination. Therefore the poor performance on medium sized objects by this Faster RCNN net
may be due to its inability handle heavily overlapping objects in some images.
[2] X. Chen and A. Gupta. An implementation of
faster RCNN with study for region sampling. CoRR,
abs/1702.02138, 2017.
[3] M. Everingham, L. Van Gool, C. K. I. Williams,
J. Winn, and A. Zisserman. The pascal visual object
classes (voc) challenge. International Journal of Computer Vision, 88(2):303–338, June 2010.
[4] R. Girshick. Fast r-cnn. In International Conference
on Computer Vision (ICCV), 2015.
[5] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich
feature hierarchies for accurate object detection and
semantic segmentation. In Computer Vision and Pattern Recognition, 2014.
6. Conclusions
While the out of the box YOLO performs poorly on
aerial imagery, a few modifications can greatly improve its
performance. First making the net shallower to increase its
output resolution. Second changing the net shape to more
closer match the aspect ratio of the data. While the modified
YOLO’s precision and recall is still typically a bit worse
than Faster RCNN, it’s increased speed makes a it good
option when near real-time vehicle detectors for aerial imagery are required.
[6] T. Huster and N. C. Gale. Deep learning for pedestrian
detection in aerial imagery. In MSS Passive Sensors,
2016.
[7] T. Lin, M. Maire, S. J. Belongie, L. D. Bourdev, R. B.
Girshick, J. Hays, P. Perona, D. Ramanan, P. Dollár,
and C. L. Zitnick. Microsoft COCO: common objects
in context. CoRR, abs/1405.0312, 2014.
[8] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. E.
Reed, C. Fu, and A. C. Berg. SSD: single shot multibox detector. CoRR, abs/1512.02325, 2015.
References
[9] J. Redmon, S. K. Divvala, R. B. Girshick, and
A. Farhadi. You only look once: Unified, real-time
object detection. CoRR, abs/1506.02640, 2015.
[1] M. Abadi, A. Agarwal, P. Barham, E. Brevdo,
Z. Chen, C. Citro, G. S. Corrado, A. Davis, J. Dean,
M. Devin, S. Ghemawat, I. Goodfellow, A. Harp,
5
Table 5. Results fine-tuned on AFVID and AF Building Camera data
Net
Shallow Faster RCNN
Shallow YOLO
Shallow YOLO
Input Size
600xA
416x416
864x480
AFVID
AP: 0.88 AR: 0.89
AP: 0.13 AR: 0.47
AP: 0.74 AR: 0.82
VEDAI
AP: 0.1 AR: 0.46
AP: 0.03 AR: 0.22
AP: 0.05 AR: 0.27
AF Building Camera
AP: 0.42 AR: 0.58
AP: 0.25 AR: 0.39
AP: 0.5 AR: 0.66
Table 6. AF building camera object size
Detector
Shallow Faster RCNN
Shallow YOLO (864x480)
mAP
0.080
0.124
mAP Large
0.171
0.404
[10] J. Redmon and A. Farhadi. Yolo9000: Better, faster,
stronger. arXiv preprint arXiv:1612.08242, 2016.
[11] S. Ren, K. He, R. Girshick, and J. Sun. Faster R-CNN:
Towards real-time object detection with region proposal networks. In Advances in Neural Information
Processing Systems (NIPS), 2015.
[12] T. H. Trieu. Darkflow. https://github.com/
thtrieu/darkflow, 2017.
[13] G. K. W. A. Sakla, N. T. Mundhenk. Deep multimodal vehicle detection in aerial isr imagery. In IEEE
Winter Conference on Applications of Computer Vision, 2017.
Appendix
6
mAP Medium
0.080
0.329
mAP Small
0.117
0.102
FPS
3
25
13
(a) Faster RCNN Low Density
(b) YOLO Low Density
(c) Faster RCNN High Density, High Illumination
(d) YOLO High Density, High Illumination
(e) Faster RCNN High Density, Low Illumination
(f) YOLO High Density, Low Illumination
Figure 3. Object density and illumination
7
Table 7. YOLOv2 nets
layer
Input
Convolutional
Max Pooling
Convolutional
Max Pooling
Convolutional
Convolutional
Convolutional
Max Pooling
Convolutional
Convolutional
Convolutional
Max Pooling
Convolutional
Convolutional
Convolutional
Convolutional
Convolutional
Max Pooling
Convolutional
Convolutional
Convolutional
Convolutional
Convolutional
Convolutional
Convolutional
Route
Convolutional
Reorg
Route
Convolutional
Convolutional
filters
32
64
128
64
128
256
128
256
512
256
512
256
512
1024
512
1024
1024
1024
1024
1024
64
1024
125
kernal size
3x3/1
2x2/2
3x3/1
2x2/2
3x3/1
1x1/1
3x3/1
2x2/2
3x3/1
1x1/1
3x3/1
2x2/2
3x3/1
1x1/1
3x3/1
1x1/1
3x3/1
2x2/2
3x3/1
1x1/1
3x3/1
1x1/1
3x3/1
3x3/1
3x3/1
1x1/1
3x3/1
1x1/1
net size
416 x 416 x 3
416 x 416 x 32
208 x 208 x 32
208 x 208 x 64
104 x 104 x 64
104 x 104 x 128
104 x 104 x 64
104 x 104 x 128
52 x 52 x 128
52 x 52 x 256
52 x 52 x 128
52 x 52 x 256
26 x 26 x 256
26 x 26 x 512
26 x 26 x 256
26 x 26 x 512
26 x 26 x 256
26 x 26 x 512
13 x 13 x 512
13 x 13 x 1024
13 x 13 x 512
13 x 13 x 1024
13 x 13 x 1024
13 x 13 x 1024
13 x 13 x 1024
13 x 13 x 1024
26 x 26 x 512
26 x 26 x 64
13 x 13 x 256
13 x 13 x 1280
13 x 13 x 1024
13 x 13 x 125
layer
Input
Convolutional
Max Pooling
Convolutional
Max Pooling
Convolutional
Convolutional
Convolutional
Max Pooling
Convolutional
Convolutional
Convolutional
Max Pooling
Convolutional
Convolutional
Convolutional
Convolutional
Convolutional
Route
Convolutional
Reorg
Route
Convolutional
Convolutional
filters
32
64
128
64
128
256
128
256
512
256
512
256
512
64
1024
125
kernal size
3x3/1
2x2/2
3x3/1
2x2/2
3x3/1
1x1/1
3x3/1
2x2/2
3x3/1
1x1/1
3x3/1
2x2/2
3x3/1
1x1/1
3x3/1
1x1/1
3x3/1
1x1/1
3x3/1
1x1/1
(b) Shallow YOLOv2 net
(a) Standard YOLOv2 net
8
net size
416 x 416 x 3
416 x 416 x 32
208 x 208 x 32
208 x 208 x 64
104 x 104 x 64
104 x 104 x 128
104 x 104 x 64
104 x 104 x 128
52 x 52 x 128
52 x 52 x 256
52 x 52 x 128
52 x 52 x 256
26 x 26 x 256
26 x 26 x 512
26 x 26 x 256
26 x 26 x 512
26 x 26 x 256
26 x 26 x 512
52 x 52 x 256
52 x 52 x 64
26 x 26 x 256
26 x 26 x 768
26 x 26 x 1024
26 x 26 x 30
| 1cs.CV
|
Sparse Coding by Spiking Neural Networks:
Convergence Theory and Computational Results
arXiv:1705.05475v1 [cs.LG] 15 May 2017
Ping Tak Peter Tang, Tsung-Han Lin, and Mike Davies
Intel Corporation
{peter.tang, tsung-han.lin, mike.davies}@intel.com
Abstract
In a spiking neural network (SNN), individual neurons operate autonomously and only communicate
with other neurons sparingly and asynchronously via spike signals. These characteristics render a massively
parallel hardware implementation of SNN a potentially powerful computer, albeit a non von Neumann one.
But can one guarantee that a SNN computer solves some important problems reliably? In this paper, we
formulate a mathematical model of one SNN that can be configured for a sparse coding problem for feature
extraction. With a moderate but well-defined assumption, we prove that the SNN indeed solves sparse
coding. To the best of our knowledge, this is the first rigorous result of this kind.
1
Introduction
A central question in computational neuroscience is to understand how complex computations emerge from
networks of neurons. For neuroscientists, a key pursuit is to formulate neural network models that resemble the
researchers’ understanding of physical neural activities and functionalities. Precise mathematical definitions
or analysis of such models is less important in comparison. For computer scientists, on the other hand, a key
pursuit is often to devise new solvers for specific computational problems. Understanding of neural activities
serves mainly as an inspiration for formulating neural network models; the actual model adopted needs not
be so much faithfully reflecting actual neural activities as to be mathematically well defined and possesses
provable properties such as stability or convergence to the solution of the computational problem at hand.
This paper’s goal is that of a computer scientist. We formulate here two neural network models that can
provably solve a mixed `2 -`1 optimization problem (often called a LASSO problem). LASSO is a workhorse
for sparse coding, a method applicable across machine learning, signal processing, and statistics. In this work,
we provide a framework to rigorously establish the convergence of firing rates in a spiking neural network to
solutions corresponding to a LASSO problem. This network model, namely the Spiking LCA, is first proposed
in [16] to implement the LCA model [15] using analog integrate-and-fire neuron circuit. We will call the LCA
model in [15] the Analog LCA (A-LCA) for clarity. In the next section, we introduce the A-LCA model and
its configurations for LASSO and its constrained variant CLASSO. A-LCA is a form of Hopfield network,
but the specific (C)LASSO configurations render convergence difficult to establish. We will outline our recent
results that use a suitable generalization of the LaSalle principle to show that A-LCA converges to (C)LASSO
solutions.
In A-LCA, neurons communicate among themselves with real numbers (analog values) during certain time
intervals. In Spiking LCA (S-LCA), neurons communicate among themselves via “spike” (digital) signals
that can be encoded with a single bit. Moreover, communication occurs only at specific time instances.
Consequently, S-LCA is much more communication efficient. Section 3 formulates S-LCA and other auxiliary
variables such as average soma currents and instantaneous spike rates. The section subsequently provides a
proof that the instantaneous rates converge to CLASSO solutions. This proof is built upon the results we
obtained for A-LCA and an assumption that a neuron’s inter-spike duration cannot be arbitrarily long unless
it stops spiking altogether after a finite time.
Finally, we devise a numerical implementation of S-LCA and empirically demonstrate its convergence to
CLASSO solutions. Our implementation also showcases the potential power of problem solving with spiking
1
neurons in practice: when an approximate implementation of S-LCA is ran on a conventional CPU, it is able
to converge to a solution with modest accuracy in a short amount of time. The convergence is even faster than
FISTA [4], one of the fastest LASSO solvers. This result suggests that a specialized spiking neuron hardware
is promising, as parallelism and sparse communications between neurons can be fully leveraged in such an
architecture.
2
Sparse Coding by Analog LCA Neural Network
We formulate the sparse coding problem as follows. Given N vectors in RM , Φ = [φ1 , φ2 , . . . , φN ], N > M , (Φ
is usually called a redundant—due to N > M —dictionary) and a vector s ∈ RM (consider s an input signal),
try to code (approximate well) s as Φa where a ∈ RN contains as many zero entries as possible. Solving a
sparse coding problem has attracted a tremendous amount of research effort [9]. One effective way is to arrive
at a through solving the LASSO problem [19] where one minimizes the `2 distance between s and Φa with a
`1 regularization on the a parameters. For reasons to be clear later on, we will consider this problem with the
additional requirement that a be non-negative: a ≥ 0. We call this the CLASSO (C for constrained) problem:
1
ks − Φak22 + λkak1
(1)
2
Rozell, et al., presented in [15] the first neural network model aims at solving LASSO. N neurons are used
to represent each of the N dictionary atoms φi . Each neuron receives an input signal bi that serves to increase
a “potential” value ui (t) that a neuron keeps over time. When this potential is above a certain threshold,
neuron-i will send inhibitory signals that aim to reduce the potential values of the list of receiving neurons
with which neuron-i “competes.” The authors called this kind of algorithms expressed in this neural network
mechanism Locally Competitive Algorithms (LCAs). In this paper, we call this as analog LCA (A-LCA).
Mathematically, an A-LCA can be described as a set of ordinary differential equations (a dynamical system)
of the form
X
u̇i (t) = bi − ui (t) −
wij T (uj (t)), i = 1, 2, . . . , N.
argmina≥0
j6=i
The function T is a thresholding (also known as an activation) function that decides when and how an inhibition
signal is sent. The coefficients wij further weigh the severity of each inhibition signal. In this general form,
A-LCA is an instantiation of the Hopfield network proposed in [10, 11].
Given a LASSO or CLASSO problem, A-LCA is configured by bi = φTi s, wij = φTi φj . For LASSO, the
thresholding function T is set to T = T±λ , and for CLASSO it is set to T = Tλ : Tλ (x) is defined as 0 when
def
x ≤ λ and x − λ when x > λ; and T±λ (x) = Tλ (x) + Tλ (−x). Note that if all the φi s are normalized to
φTi φi = 1, then the dynamical system in vector notation is
u̇ = b − u − (ΦT Φ − I)a,
a = T(u).
(2)
The vector function T : RN → RN simply applies the same scalar function T to each of the input vector’s
component. We say A-LCA solves (C)LASSO if a particular solution of the dynamical system converges to
a vector u∗ and that a∗ = T(u∗ ) is the optimal solution for (C)LASSO. This convergence phenomenon was
demonstrated in [15].
LCA needs not be realized on a traditional computer via some classical numerical differential equation
solver; one can realize it using, for example, an analog circuit which may in fact be able to solve (C)LASSO
faster or with less energy. From the point of view of establishing A-LCA as a robust way to solve (C)LASSO,
rigorous mathematical results on A-LCA’s convergence is invaluable. Furthermore, any convergence theory
here will bound to have bearings on other neural network architectures, as we will see in Section 3. Had the
thresholding function T in A-LCA be strictly increasing and unbounded above and below, standard Lyapunov
theory can be applied to establish convergence of the dynamical system. This is already pointed out in
Hopfield’s early work for both graded neuron model [11] and spiking neuron model [12]. Nevertheless, such
an A-LCA does not correspond to (C)LASSO where the thresholding functions are not strictly increasing.
Furthermore, the CLASSO thresholding function is bounded below as well. While Rozell, et al., demonstrated
some convergence phenomenon [15], it is in two later works [1, 2] that Rozell and other colleagues attempted
to complement the original work with convergence analysis and proofs. Among other results, these works
2
stated that for any particular A-LCA solution u(t), T(u(t)) with T = T±λ converges to a LASSO optimal
solution. Unfortunately, as detailed in [18], there are major gaps in the related proofs and thus the convergence
claims are in doubt. Moreover, the case of T = Tλ for the CLASSO problem was not addressed. In [18], one
of our present authors established several convergence results which we now summarize so as to support the
development of Section 3. The interested reader can refer to [18] for complete details.
A-LCA is a dynamical system of the form u̇ = F(u), F : RN → RN . In this case, the function F is
defined as F(x) = b − x − (ΦT Φ − I)T(x). Given any “starting point” u(0) ∈ RN , standard theory of ordinary
differential equations shows that there is a unique solution u(t) such that u(0) = u(0) and u̇(t) = F(u(t)) for
all t ≥ 0. Solutions are also commonly called flows. The two key questions are (1) given some (or any) starting
point u(0) , whether and in what sense the flow u(t) converges, and (2) if so, what relationships exist between
the limiting process and the (C)LASSO solutions.
The LaSalle invariance principle [14] is a powerful tool to help answer the first question. The gist of the
principle is that if one can construct a function V : RN → R such that it is non-increasing along any flow, then
one can conclude that all flows must converge to a special set1 M which is the largest positive invariant set2
inside the set of points at which the Lie derivative of V is zero. The crucial technical requirements on V are that
V possesses continuous partial derivatives and be radially unbounded3 . Unfortunately, the natural choice of
V for A-LCA does not have continuous first partial derivatives everywhere, and not radially unbounded in the
case of CLASSO. Both failures are due to the special form of T with V (u) = (1/2)ks − ΦT(u)k22 + λ kT(u)k1 .
Based on a generalized version of LaSalle’s principle proved in [18], we establish that any A-LCA flow u(t)
(LASSO or CLASSO) converges to M, the largest positive invariant set inside the “stationary” set S = { u |
(∂V /∂un )Fn (u) = 0 whenever |T (un )| > 0. }.
Having established u(t) → M, we further prove in [18] that M is in fact the inverse image under T of the
set C of optimal (C)LASSO solutions. The proof is based on the KKT [6] condition that characterizes C and
properties particular to A-LCA.
Theorem 1. (A-LCA convergence results from [18]) Given the A-LCA
u̇ = F(u),
F(u) = b − u − (ΦT Φ − I)T(u).
T is based on T±λ if one wants to solve LASSO and Tλ , CLASSO. Let u(0) be an arbitrary starting point and
u(t) be the corresponding flow. The following hold:
1. Let C be the set of (C)LASSO optimal solutions and F̂ = T−1 (C) be C’s inverse image under the
corresponding thresholding function T. Then any arbitrary flow u(t) always converges to the set F̂.
2. Moreover, limt→∞ E(a(t)) = E ∗ where E ∗ is the optimal objective function value of (C)LASSO, E(a) =
(1/2)ks − Φak22 + λkak1 and a(t) = T(u(t)).
3. Finally, when the (C)LASSO optimal solution a∗ is unique, then there is a unique u∗ such that F(u∗ ) = 0.
Furthermore u(t) → u∗ and T(u(t)) → T(u∗ ) = a∗ as t → ∞.
3
Sparse Coding by Spiking LCA Neural Network
A-LCA is inherently communication efficient: Neuron-i needs to communicate to others only when its internal
state ui (t) exceeds a threshold, namely |T (ui (t))| > 0. In a sparse coding problem, it is expected that the
internal state will eventually stay perpetually below the threshold for many neurons. Nevertheless, for the
entire duration during which a neuron’s internal state is above threshold, constant communication is required.
Furthermore, the value to be sent to other neurons are real valued (analog) in nature. In this perspective, a
spiking neural network (SNN) model holds the promise of even greater communication efficiency. In a typical
SNN, various internal states of a neuron are also continually evolving. In contrast, however, communication in
the form of a spike—that is one bit—is sent to other neurons only when a certain internal state reaches a level
(a firing threshold). This internal state is reset right after the spiking event, thus cutting off communication
1 u(t)
→ M if dist(u(t), M) → 0 where dist(x, M) = inf y∈M kx − yk2 .
set is positive invariant if any flow originated from the set stays in that set forever.
3 The function V is radially unbounded if |V (u)| → ∞ whenever kuk → ∞
2
2A
3
immediately until the time when the internal state is “charged up” enough. Thus communication is necessary
only once in a certain time span and then a single bit of information carrier suffices.
While such a SNN admits mathematical descriptions [16, 3], there is hitherto no rigorous results on the
network’s convergence behavior. In particular, it is unclear how a SNN can be configured to solve specific
problems with some guarantees. We present now a mathematical formulation of a SNN and a natural definition
of instantaneous spiking rate. Our main result is that under a moderate assumption, the spiking rate converges
to the CLASSO solution when the SNN is suitably configured. To the best of our knowledge, this is the first
time a rigorous result of this kind is established.
In a SNN each of the N neurons maintains, over time t, an internal soma current µi (t) configured to receive a
Rt
constant input bi and an internal potential vi (t). The potential is “charged” up according to vi (t) = 0 (µi − λ)
where λ ≥ 0 is a configured bias current. When vi (t) reaches a firing threshold νf at a time ti,k , neuroni resets its potential to νr but simultaneously fires an inhibitory signal to a preconfigured set of receptive
neurons, neuron-js, whose soma current will be diminished according to a weighted exponential decay function:
µj (t) ← µj (t) − wji α(t − ti,k ), where α(t) = e−t for t P
≥ 0 and zero otherwise. Let {ti,k } be the ordered time
sequence of when neuron-i spikes and define σi (t) = k δ(t − ti,k ), then the soma current satisfies both the
algebraic and differential equations below (the operator ∗ denotes convolution):
X
X
µi (t) = bi −
wij (α ∗ σj )(t), µ̇i (t) = bi − µi (t) −
wij σj (t).
(3)
j6=i
j6=i
Equation 3 together with the definition of the spike trains σi (t) describe our spiking LCA (S-LCA).
An intuitive definition of spike rate of a neuron is clearly the number of spikes per unit time. Hence we
define the instantaneous spiking rate ai (t) and average soma current ui (t) for neuron-i as:
Z t
Z t
1
1
def
def
σi (s) ds and ui (t) =
µi (s) ds, t0 ≥ 0 is a parameter.
(4)
ai (t) =
t − t0 t0
t − t 0 t0
Rt
Apply the operator (t − t0 )−1 t0 ds to the differential equation portion in (3), using also the relationship
u̇i (t) = (µi (t) − ui (t))/(t − t0 ), and we obtain
X
u̇i (t) = bi − ui (t) −
wij aj (t) − (ui (t) − ui (t0 ))/(t − t0 ).
(5)
j6=i
Consider now a CLASSO problem where the dictionary atoms are non-negative and normalized to unit
Euclidean norm. Configure S-LCA with λ and wij = φTi φj from Equation 1, and set νf ← 1, νr ← 0. So
configured, it can be shown that the soma currents’ magnitudes (and thus that of the average currents as well)
are bounded: there is a B such that |µi (t)|, |ui (t)| ≤ B for all i = 1, 2, . . . , N and all t > t0 . Consequently,
lim u̇i (t) = lim (µi (t) − ui (t))/(t − t0 ) = 0,
t→∞
t→∞
for i = 1, 2, . . . , N .
The following relationship between ui (t) and ai (t) is crucial:
Z ti,k
Z t
1
1
(µi − λ) +
(µi − λ) = ai (t) + vi (t)/(t − t0 ).
ui (t) − λ =
t − t 0 t0
t − t0 ti,k
(6)
(7)
From this equation and a moderate assumption that inter-spike duration ti,k+1 − ti,k cannot be arbitrarily
long unless neuron-i stops spiking altogether, one can prove that
Tλ (ui (t)) − ai (t) → 0
as t → ∞.
(8)
The complete proof for this result is left in the Appendix.
We can derive convergence of S-LCA as follows. Since the average soma currents are bounded, Bolzanodef
T
Weierstrass theorem shows that u(t) = [u1 (t), u2 (t), . . . , uN (t)] has at least one limit point, that is, there is
∗
N
a point u ∈ R and a time sequence t1 < t2 < · · · , tk → ∞ such that u(tk ) → u∗ as k → ∞. By Equation 8,
def
T(u(tk )) → T(u∗ ) = a∗ . By Equations 5 and 6, we must therefore have
0 = b − u∗ − (ΦT Φ − I)a∗ .
(9)
Since S-LCA is configured for a CLASSO problem, the limit u∗ is in fact a fixed point of A-LCA, which is
unique whenever CLASSO’s solution is. In this case, the limit point of the average currents {u(t) | t ≥ 0} is
unique and thus indeed we must have u(t) → u∗ and a(t) → a∗ , the CLASSO solution.
4
Potential
Neuron 2
Neuron ID
Neuron 1
1
Neuron 3
0
0
0
2
4
6
8
10
12
Spike count
1
0
-1
2
4
6
8
10
4
6
8
10
12
8
10
12
15
2
0
2
(c)
(a)
Current
01
02
03
12
9
6
3
0
12
SNN time
0
2
4
6
SNN time
(b)
(d)
Figure 1: Detail dynamics of a simple 3-neuron spiking network: In the beginning before any neuron fires,
the membrane potentials (see 1-a) of the neurons grow linearly with a rate determined by the initial soma
currents (see 1-b). This continues until Neuron 3 becomes the first to reach the firing threshold; an inhibitory
spike is sent to Neurons 1 and 2, causing immediate drops in their soma currents. Consequently, the growths
of Neurons 1 and 2’s membrane potentials slow down, and the neurons’ instantaneous spike rates decrease.
The pattern of membrane integration, spike, and mutual inhibition repeats; the network rapidly moves into a
steady state where stable firing rates can be observed. The convergent firing rates yield the CLASSO optimal
solution of [0.684, 0, 1.217] (this solution is also verified by running the LARS algorithm [8]). The four subfigures: (a) Evolution of membrane potential. (b) Evolution of soma current. (c) Spike raster plot. (d) Solid
Rt
lines are the cumulative spike count of each neuron, and dashed line depicts the value of 0 Tλ (ui (s))ds in the
corresponding A-LCA. The close approximation indicates a strong tie between the two formulations.
4
Numerical Simulations
To simulate the dynamics of S-LCA on a conventional CPU, one can precisely solve the continuous-time spiking
network formulation by tracking the order of firing neurons. In between consecutive spiking events, the internal
variables, vi (t) and µi (t), of each neuron follow simple differential equations and permit closed-form solutions.
This method, however, is likely to be slow that it requires a global coordinator that looks ahead into the future
to determine the next firing neuron. For efficiency, we instead take an approximate approach that evolves the
network state in constant-sized discrete time steps. At every step, the internal variables of each neuron are
updated and a firing event is triggered if the potential exceeds the firing threshold. The simplicity of this
approach admits parallel implementations and is suitable for specialized hardware designs. Nevertheless, this
constant-sized-time-step approach introduces errors in spike timings: the time that a neuron sends out a spike
may be delayed by up to a time step. As we will see in this section, the timing error is the major factor that
limits the accuracy of the solutions from spiking networks. However, such an efficiency-accuracy trade-off may
in fact be desirable for certain applications such as those in machine learning.
4.1
Illustration of SNN dynamics
We solve a simple CLASSO problem: mina 12 ks − Φak22 + λ kak1 subject to a ≥ 0, where
0.5
0.3313 0.8148 0.4364
s = 1 , Φ = [φ1 φ2 φ3 ] = 0.8835 0.3621 0.2182 , λ = 0.1.
1.5
0.3313 0.4527 0.8729
We use a 3-neuron network configured with bi = φTi s, wij = φTi φj , the bias current as λ = 0.1 and firing
threshold set to 1. Figure 1 details the dynamics of this simple 3-neuron spiking network. It can be seen from
this simple example that the network only needs very few spike exchanges for it to converge. In particular, a
weak neuron, such as Neuron 2, is quickly rendered inactive by inhibitory spike signals from competing neurons.
This raises an important question: how many spikes are in the network? We do not find this question easy to
answer theoretically. However, empirically we see the number of spikes in σi (t) in S-LCA can be approximated
5
1e2
1e0
1e-1
1e-2
1e-3
1e-4
1e0
1e-1
1e-2
1e-3
1e-4
1e-5
1e-6
1e-1
Average from t = 0
Average from t = 0.5
Exponential kernel
Thresholded average current
1e1
(E(t)-E*) / E*
(E(t)-E*) / E*
1e2
step = 0.01
step = 0.001
step = 0.0001
1e1
1e0
1e1
1e2
1e3
1e4
1e-5
1e-1
1e5
SNN time
1e0
1e1
1e2
1e3
1e4
1e5
SNN time
(a)
(b)
Figure 2: (a) The convergence of a 400-neuron spiking network to a CLASSO solution. (b) Comparing the
convergence of different formulations to read out solutions from a spiking neural network. Using a positive t0
for Equation 4 gives the fastest initial convergence, while using the thresholded average current reaches the
highest accuracy the quickest. Despite a lack of theoretical guarantee, the exponential kernel method yields an
acceptable, though less accurate, solution. This kernel is easy to implement in hardware and thus attractive
when a SNN “computer” is to be built.
from the state variable ui in A-LCA, that is the ui (s) in Equation 10 below are solutions to Equation 2:
Z
t
Z
σi (s)ds ≈
0
t
Tλ (ui (s))ds,
i = 1, 2, . . . , N.
(10)
0
Figure 1(d) shows the close approximation of spike counts using (10) in the example. We observe that such
approximation consistently holds in large-scale problems, suggesting a strong tie between S-LCA and A-LCA.
Since in an A-LCA configured for a sparse coding problem, we expect Tλ (ui (t)) for most i’s to converge to
zero, (10) suggests that the total spike count in S-LCA is small.
4.2
Convergence of spiking neural networks
We use a larger 400-neuron spiking network to empirically examine the convergence of spike rates to CLASSO
solution. The neural network is configured to perform feature extraction from a 8×8 image patch, using a 400atom dictionary learned from other image datasets.4 With the chosen λ, the optimal solution has 8 non-zeros
entries. Figure 2(a) shows the convergence of the objective function value in the spiking network solution,
comparing to the true optimal objective value obtained from a conventional CLASSO solver. Indeed, with a
small step size, the spiking network converges to a solution very close to the true optimum.
The relationships among step size, solution accuracy and total computation cost are noteworthy. Figure 2(a)
shows that increasing the step size from 10−3 to 10−2 sacrifices two digits of accuracy in the computed E ∗ .
The total computation cost is reduced by a factor of 103 : It takes 102 times fewer time units to converge, and
each time unit requires 10 times fewer iterations. This multiplication effect on cost savings is highly desirable
in applications such as machine learning where accuracy is not paramount. We note that a large-step-size
configuration is also suitable for problems whose solutions are sparse: The total number of spikes are fewer
and thus total timing errors are correspondingly fewer.
There are several ways to “read out” a SNN solution. Most rigidly, we can adhere to ai (t) in Equation 4
with t0 = 0. In practice, picking some t0 > 0 is better when we expect a sparse solution: The resulting
ai (t) will be identically zero for those neurons that only spike before time t0 . Because Tλ (ui (t)) − ai (t) → 0
(Equation 8), another alternative is to use Tλ (ui (t)) as the solution, which is more likely to deliver a truly
R t t−s
sparse solution. Finally, one can change ai (t)’s definition to τ −1 0 e− τ σi (s)ds, so that the impact of the
spikes in the past decays quickly. Figure 2(b) illustrates these different “read out” methods and shows that
the exponential kernel is as effective empirically, although we must point out that the previous mathematical
convergence analysis is no longer applicable in this case.
6
1e-2
1e-4
0.4
4
2
1672 spikes
(avg 0.052 spike/neuron)
1e-3
0
0.05
0.1
0.15
0.2
0.25
Execution time (second)
(a) 52×52 image
0.3
0.1
0
0
0.05
0.1
0.15
0.2
1e1
0.25
0.3
S-LCA
FISTA
5th iteration
1e0
6
Sparsity (%)
1e-1
L2 error
(E(t)-E*) / E*
1e0
8
S-LCA (l2 error)
FISTA (l2 error)
S-LCA (sparsity)
FISTA (sparsity)
S-LCA
FISTA
5th iteration
(E(t)-E*) / E*
0.7
1e1
1e-1
1e-2
10922 spikes
(avg 0.019 spike/neuron)
1e-3
1e-4
0
0.5
(b) Breakdown of (a)
1
1.5
2
2.5
3
Execution time (second)
Execution time (second)
(c) 208×208 image
Figure 3: CPU execution time for spiking neural networks, with a step size of 0.01. There are 32,256 unknowns
in the 52×52 image case shown in (a), and 582,624 unknowns in the 208×208 image case shown in (c). (b)
shows the breakdown of the objective function in the 52×52 image experiment. The `2 error is defined as
ks−Φak2
ksk2 , and the sparsity is the percentage of entries with values greater than 0.01. Note that the spiking
network finds the optimal solution by gradually increasing the sparsity, rather than decreasing as in FISTA.
This results in the spare spiking activities of the neurons.
4.3
CPU benchmark of a spiking network implementation
Our earlier discussions suggest that the spiking network can solve CLASSO using very few spikes. This
property has important implications to a SNN’s computational efficiency. The computation cost of a N neuron spiking network has two components: neuron states update and spiking events update. Neuron states
update includes updating the internal potential and current values of every neuron, and thus incurs an O(N )
cost at every time step. The cost of spiking events update is proportional to N times the average number
of inter-neuron connections because a spiking neuron updates the soma currents of those neurons to which
it connects. Thus this cost can be as high as O(N 2 ) (for networks with all-to-all connectivity, such as in
the two previous examples) or as low as O(N ) (for networks with only local connectivity, such as in the
example below). Nevertheless, spiking-event cost is incurred only when there is a spike, which may happen
far fewer than once per time step. In practice we observe that computation time is usually dominated by
neuron-states update, corroborating the general belief that spiking events are relatively rare, making spiking
networks communication efficient.
We report the execution time of simulating the spiking neural network on a conventional CPU, and compare
the convergence time with FISTA [4], one of the fastest LASSO solvers. We solve a convolutional sparse coding
problem [20] on a 52x52 image and a 208x208 image.5 The experiments are ran on 2.3GHz Intel R Xeon R CPU
E5-2699 using a single core. SIMD is enabled to exploit the intrinsic parallelism of neural network and matrix
operations. As shown in Figure 3, the spiking network delivers much faster early convergence than FISTA,
despite its solution accuracy plateauing due to spike timing errors. The convergence trends in both figures
are similar, demonstrating that spiking networks can solve problems of various sizes. The fast convergence of
spiking networks can be attributed to their ability to fully exploit the sparsity in solutions to reduce the spike
counts. The fine-grain asynchronous communication can quickly suppress most neurons from firing. In FISTA
or in any other conventional solvers, communications between variables is similarly needed, but is realized
through matrix-vector multiplications performed in an iteration-to-iteration basis. The only way to exploit
sparsity is to avoid computations involving variables that have gone to zero during one iteration. A comparison
of how the sparsity in solutions evolves in S-LCA and FISTA can be found in Figure 3(b).
5
Discussion
Our work is closely related to the recent progress on optimality-driven balanced network [7, 3, 5]. The SNN
model in [3, 5] differs slightly from ours in that only one internal state is used in the former. Using our
4 The
5 We
input has 128 dimensions by splitting the image into positive and negative channels.
use 8×8 patches, a stride of 4, and a 128×224 dictionary.
7
language here, neuron-i’s spike is generated by µi (t) reaching a threshold and not by vi (t), whose role is
eliminated altogether. Despite the differences in the details of neuron models, spikes in both networks occur
from a competitive process between neurons, and serve to minimize a network-level energy function. This
work furthers the understanding of the convergence property in such spiking networks. Additionally, it is
argued that in a tightly balanced excitatory/inhibitory network, spike codes are highly efficient that each
spike is precisely timed to keep the network in optimality. This work provides evidence of the high coding
efficiency even before the network settles into steady-state. By utilizing well-timed spikes, the neurons are able
to collectively solve optimization problems with minimum communications. We demonstrate that this insight
can be translated into practical value through an approximate implementation on conventional CPU.
We observe that mathematical rigor was not a focus in [3]: The statement that in a tightly balanced
network the potential converges to zero is problematic when taken literally as all spiking events will eventually
cease in that case. The stationary points of the loss function (Equation 6 in [3]) are no longer necessarily the
stationary points when the firing rates are constrained to be non-negative. The general KKT condition has
to be used in this situation. The condition E(no spike) > E(spike) does not affect the behavior of the loss
function in between spikes. In essence, there is no guarantee that the trajectory of the r(t) variable generated
d
E(r(t)) ≤ 0.
by the SNN is descending the loss function, that is, dt
Our SNN formulation and the established convergence properties can be easily extended to incorporate an
additional `2 -penalty term, the so-called elastic-net problem [21]
argmina≥0
1
ks − Φak22 + λ1 kak1 + λ2 kak22
2
(11)
The elastic-net formulation can be handled by modifying the slope of the activation function T in A-LCA as
follows
0
if x ≤ λ1
def
def
Tλ (x) =
, T±λ (x) = Tλ (x) + Tλ (−x).
x−λ1
if
x
>
λ
1
2λ2 +1
In S-LCA, this corresponds to setting the bias current to λ1 and modifying the firing thresholds νf of the
neurons to 2λ2 + 1.
There are several other works studying the computation of sparse representations using spiking neurons.
Zylberberg et al. [22] show the emergence of sparse representations through local rules, but do not provide a
network-level energy function. Hu et al. [13] derive a spiking network formulation that minimizes a modified
time-varying LASSO objective. Shapero et al. [16, 17] are the first to propose the S-LCA formulation, but
yet to provide an in-depth analysis. We believe the S-LCA formulation can be a powerful primitive in future
spiking network research.
The computational power of spikes enables new opportunities in future computer architecture designs. The
spike-driven computational paradigm motivates an architecture composed of massively parallel computation
units. Unlike the von Neumann architecture, the infrequent but dispersed communication pattern between the
units suggests a decentralized design where memory should be placed close to the compute, and communication
can be realized through dedicated routing fabrics. Such designs have the potential to accelerate computations
without breaking the energy-density limit.
8
Appendices
A
Governing Algebraic and Differential Equations
Consider a neural networking consisting of N neurons. The only independent variables are the N soma
currents µi (t) for i = 1, 2, . . . , N . There are another N variables of potentials vi (t) which are depedent on the
currents to be described momentarily. Consider the following configurations. Each neron receives a positive
constant input current bi . A nonnegative current bias λ and a positive potential threshold ν are set a priori.
At any given time t0 such that vi (t0 ) < ν, the potential evolves according to
Z t
vi (t) =
(µi (s) − λ) ds
t0
until the time ti,k > t0 when vi (t) = ν. At this time, a spike signal is sent from neuron-i to all the nerons
that are connected to it, weighted by a set of pre-configured weights wj,i . The potential vi (t) is reset to zero
immediately afterwards. That is, for t > ti,k but before the next spike is generated,
Z t
vi (t) =
(µi (s) − λ) ds.
ti,k
Moreover, for any consecutive spike times ti,k and ti,k+1 ,
Z ti,k+1
(µi (s) − λ) ds = ν.
ti,k
Finally, when neuron-i receives a spike from neuron-j at time tj,k with a weight wi,j , the soma current µi (t)
is changed by the additive signal −wi,j α(t − tj,k ) where
α(t) = H(t)e−t/τ ,
H(t) being the Heaviside function that is 1 for t ≥ 0 and 0 otherwise. The sign convention used here means
that a positive wi,j means that a spike from neuron-j always tries to inhibit neuron-i.
Suppose the initial potentials vi (0) are all set to be below the spiking threshold ν, then the dynamics of
the system can be succintly described by the set of algebraic equations
X
µi (t) = bi −
wi,j (α ∗ σj )(t), i = 1, 2, . . . , N
(AE)
j6=i
where ∗ is the convolution operator and σj (t) is the sequence of spikes
X
σj (t) =
δ(t − tj,k ),
k
δ(t) being the Dirac delta function. The spike times are determined in turn by the evolution of the soma
currents that govern the evolutions of the potentials.
One can also express the algebraicR equations AE as a set of differential equations. Note that the Heaviside
t
function can be expressed as H(t) = −∞ δ(s) ds. Hence
Z t
d
d
α(t) =
e−t/τ
δ(s) ds
dt
dt
−∞
1
= − α(t) + δ(t).
τ
Thus, differentiating Equation AE yields
X
1
µ̇i (t) = (bi − µi (t)) −
wi,j σj (t).
(DE)
τ
j6=i
Note that Equations AE and DE are given in terms of the spike trains σj (t) that are governed in turn by
the soma currents themselves as well as the configuartions of initial potentials, the spiking threshold ν and
bias current λ.
9
B
Defining Spike Rates and Average Currents
Suppose the system of spiking neurons are initialized with sub-threshold potentials, that is, vi (0) < ν for
all i = 1, 2, . . . , N . Thus at least for finite time after 0, all soma currents remain constant at bi and that
no neurons will generate any spikes. Furthermore, consider for now that wi,j ≥ 0 for all i, j. That is, only
inhibitory signals are present. Let the spike times for each neuron i be 0 < ti,1 < ti,2 < · · · . This sequence
could be empty, finite, or infinite. It is empty if the potential vi (t) never reaches the threshold. It is finite if
the neuron stop spiking from a certain time onwards. We will define the spike rate, ai (t), and average current,
ui (t), for each neuron as follows.
1 Rt
def
t 0 σi (s) ds t > 0, ,
ai (t) =
0
t=0
and
def
ui (t) =
R
1 t
t 0
bi
µi (s) ds
t > 0,
.
t=0
With these definitions, the section presents the following results.
• The inhibition assumption leads to the fact that all the soma currents are bounded above. This in turns
shows that none of the neurons can spike arbitrarily rapidly.
• The fact that neurons cannot spike arbitrarily rapidly implies the soma currents are bounded from below
as well.
• The main assumption needed (that is, something cannot be proved at this point) is that if a neron spikes
infinitely often, then the duration between consecutive spikes cannot be arbitrarily long.
• Using this assumption and previous established properties, one can prove an important relationship
between the spike rate and average current in terms of the familiar thresholding function T
Proposition 1. There exists bounds B− and B+ such that µi (t) ∈ [B− , B+ ] for all i and t ≥ 0. With the
def
convention that ti,0 = 0, then there is a positive value R > 0 such that ti,k+1 −ti,k ≥ 1/R for all i = 1, 2, . . . , N
and k ≥ 0, whenever these values exist.
Proof. Because all spike signals are inhibitory, clearly from Equation AE, we have µi (t) ≤ bi for all t ≥ 0.
def
Thus, defining B+ = maxi bi leads to µi (t) ≤ B+ for all i and t ≥ 0.
Given any two consecutive ti,k and ti,k+1 that exist,
Z ti,k+1
+
ν = vi (ti,k ) +
(µi (s) − λ) ds
ti,k
≤
vi (t+
i,k )
+ (ti,k+1 − ti,k )(B+ − λ).
Note that vi (t+
i,k ) = 0 if k ≥ 1. For the special case when k = 0, this value is vi (0) < ν. Hence
n
o
ti,k+1 − ti,k ≥ min min{ν − vi (0)}, ν (B+ − λ)−1 .
i
Thus there is a R > 0 so that ti,k+1 − ti,k ≥ 1/R whenever these two spike times exist.
Finally, because of duration between spikes cannot be arbitrarily small, it is easy to see that
def
γ =
∞
X
−`
e Rτ ≥ (α ∗ σ)(t).
`=0
Therefore,
def
B− = min{−γ
X
wi,j } ≤ µi (t)
j6=i
for all i = 1, 2, . . . , N and t ≥ 0. So indeed, there are B− and B+ such that µi (t) ∈ [B− , B+ ] for all i and
t ≥ 0.
10
Proposition 1 shows that among other things, there is a lower bound of the duration of consecutive spikes.
The following is an assumption.
Assumption 1. Assume that there is a positive number r > 0 such that whenever the numbers ti,k and ti,k+1
exist, ti,k+1 − ti,k ≤ 1/r.
In simple words, this assumption says that unless a neuron stop spiking althogether after a certain time,
the duration between consecutive spike cannot become arbitrarily long. With this assumption and the results
in Proposition 1, the following important relationship between u(t) and a(t) can be established.
Theorem 2. Let T (x) be the thresholding function where T (x) = 0 for x ≤ λ, and T (x) = x − λ for x > λ.
For each neuron i, there is a function ∆i (t) such that
T (ui (t)) = ai (t) ν + ∆i (t)
and that ∆i (t) → 0 as t → ∞.
Proof. Let
A = { i | neuron-i spikes infinitely often }
(A stands for “active”), and
I = { i | neuron-i stop spiking after a finite time }
(I stands for “inactive”). First consider i ∈ I. Let ti,k be the time of the final spike. For any t > ti,k ,
ui (t) − λ
=
1
t
Z
1
t
Z
ti,k
(µi (s) − λ) ds +
0
1
t
Z
t
(µi (s) − λ) ds
ti,k
ti,k
1
(µi (s) − λ) ds + vi (t)
t
0
1
= ai (t) ν + vi (t),
t
1
ui (t) = ai (t) ν + λ + vi (t).
t
=
Note that vi (t) ≤ ν always. If vi (t) ≥ 0, then
0 ≤ T (ui (t)) − ai (t) ≤ ν/t.
If vi (t) < 0,
−ai (t) ν ≤ T (ui (t)) − ai (t) ν ≤ 0.
Since i ∈ I, ai (t) → 0 obviously. Thus
T (ui (t)) − ai (t) ν → 0.
Consider the case of i ∈ A. For any t > 0, let ti,k be the largest spike time that is no bigger than t. Because
i ∈ A, ti,k → ∞ as t → ∞.
ui (t) − λ
=
1
t
Z
ti,k
(µi (s) − λ) ds +
0
1
= ai (t) ν +
t
Z
1
t
Z
t
(µi (s) − λ) ds
ti,k
t
(µi (s) − λ) ds.
ti,k
Furthermore, note that because of the assumption ti,k+1 − ti,k ≤ 1/r always, where r > 0, lim inf ai (t) ≥ r.
In otherwords, there is a time T large enough such that ai (t) ≥ r/2 for all i ∈ A and t ≥ T . Moreover,
0 ≤ t − ti,k ≤ ti,k+1 − ti,k ≤ 1/r and µi (t) − λ ∈ [B− − λ, B+ − λ]. Thus
1
t
Z
t
(µi (s) − λ) ds ∈
ti,k
1
[B− − λ, B+ − λ]/r → 0.
t
11
When this term is eventually smaller in magnitude than ai (t) ν,
Z
1 t
T (ui (t) = ai (t) ν +
(µi (s) − λ) ds
t ti,k
and we have
T (ui (t)) − ai (t) ν → 0.
C
Spiking Neural Nets and LCA
This section shows that for a spiking neural net (SNN) that corresponds to a LCA, the limit points of the
SNN necessarily are the fixed points of the LCA. In particular, when the LCA corresponds to a constrained
LASSO, that is LASSO where the parameters are constrained to be nonnegative, whose solution is unique,
then SNN necessarily converges to this solution. The proof for all these is surprisingly straightforward.
The following differential equation connecting u̇i (t) to ui (t) and all other spiking rates aj (t) is crucial.
u̇i (t) =
X
1
1
(bi − ui (t)) −
wi,j aj (t) − (ui (t) − bi ) .
τ
t
(rates-DE)
j6=i
Derivation of this relationship is straightforward. First, apply the operation (1/t)
Z
X
1
1 t
µ̇i (s) ds = (bi − ui (t)) −
wi,j aj (t).
t 0
τ
Rt
0
to Equation DE:
j6=i
To find an expression for the left hand side above, note that
Z
d
d 1 t
ui (t) =
µi (s) ds
dt
dt t 0
Z
1 t
1
µi (t) − 2
=
µi (s) ds
t
t 0
1
(µi (t) − ui (t)) .
=
t
Therefore
Z
1 t
1
(µi (t) − bi )
µ̇i (s) ds =
t 0
t
1
1
(µi (t) − ui (t)) + (ui (t) − bi )
=
t
t
d
1
=
ui (t) + (ui (t) − bi ) .
dt
t
Consequently, Equation rates-DE is established.
Observe that because µi (t) is bounded (Proposition 1), so is the average current ui (t). This means that
u̇i (t) → 0 as t → ∞ because it was shown just previously that u̇i (t) = (µi (t) − ui (t))/t.
Since µi (t) and ai (t) are all bounded, the vectors u(t) must have a limit point (Bolzano-Weierstrass) u∗ .
By Theorem 2, there is a correpsonding a∗ such that T(u∗ ) = a∗ ν. Moreover, we must have
1
(b − u∗ ) − W a∗
τ
= 0. Hence
0=
where the matrix W has entries wi,j and wi,i
1
1
(b − u∗ ) − W T (u∗ ).
τ
ν
∗
∗
∗
Indeed, u , a = T (u ) correspond to a fixed point of LCA. In the case when this LCA corresponds to a
LASSO with unique solution, there is only one fixed point, which implies that there is also one possible limit
point of SNN, that is, the SNN must converge, and to the LASSO solution.
0=
12
References
[1] A. Balavoine, J. Romberg, and C. J. Rozell. Convergence and rate analysis of neural networks for sparse approximation. IEEE Trans. Neural Netw., 23(9):1377–1389, September 2012.
[2] A. Balavoine, C. J. Rozell, and J. Romberg. Convergence of a neural network for sparse approximation using
nonsmooth Lojasiewicz inequality. In Proceedings of the International Joint Conference on Neural Networks,
Dalla, TX, August 2013.
[3] D. G. T. Barrett, S. Denève, and C. K. Machens. Firing rate predictions in optimal balanced networks. In NIPS,
2013.
[4] A. Beck and M. Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM
Journal on Imaging Sciences, 2(1):183–202, 2009.
[5] M. Boerlin, C. Machens, and S. Deneve. Predictive coding of dynamical variables in balanced spiking networks.
PLoS Comput Biol, 9(11), 2013.
[6] S. Boyd and L. Vandenberghe. Convex optimization. Cambridge University Press, Cambridge, 2004.
[7] S. Denève and C. K. Machens. Efficient codes and balanced networks. Nature neuroscience, 19(3):375–382, 2016.
[8] B. Efron, T. Hastie, I. Johnstone, and R. Tibshirani. Least angle regression. The Annals of Statistics, 32(2):407–
499, 2004.
[9] M. Elad. Sparse and Redundant Representations: From Theory to Applications in Signal and Image Processing.
Springer, 2010.
[10] J. J. Hopfield. Neural networks and physical systems with emergent collective computational abilities. Proc. Natl.
Acad. Sci., 79(8):2554–2558, 1982.
[11] J. J. Hopfield. Neurons with graded response have collective computational properties like those of two-state
neurons. Proc. Natl. Acad. Sci., 1:3088–3092, 1984.
[12] J. J. Hopfield and A. V. Herz. Rapid local synchronization of action potentials: Toward computation with coupled
integrate-and-fire neurons. Proc. Natl. Acad. Sci., 92(15):6655–6662, 1995.
[13] T. Hu, A. Genkin, and D. B. Chklovskii. A network of spiking neurons for computing sparse representations in an
energy-efficient way. Neural Comput., 24(11):2852–2872, 2012.
[14] J. P. LaSalle. Some extensions of Liapunov’s second method. IRE Trans. Circuit Theory, 7(4):520–527, December
1960.
[15] C. J. Rozell, D. H. Johnson, R. G. Baraniuk, and B. A. Olshausen. Sparse coding via thresholding and local
competition in neural circuits. Neural Comput., 20(10):2526–2563, 2008.
[16] S. Shapero, C. Rozell, and P. Hasler. Configurable hardware integrate and fire neurons for sparse approximation.
Neural Netw., 45:134–143, 2013.
[17] S. Shapero, M. Zhu, J. Hasler, and C. Rozell. Optimal sparse approximation with integrate and fire neurons.
International journal of neural systems, 24(05):1440001, 2014.
[18] P. T. P. Tang. Convergence of LCA Flows to (C)LASSO Solutions. ArXiv e-prints, Mar. 2016, 1603.01644.
[19] R. Tibshirani. Regression shrinkage and selection via the Lasso. J. Royal Statist. Soc B., 58(1):267–288, 1996.
[20] M. D. Zeiler, D. Krishnan, G. W. Taylor, and R. Fergus. Deconvolutional networks. In IEEE CVPR, 2010.
[21] H. Zou and T. Hastie. Regularization and variable selection via the elastic net. J. Royal Statist. Soc B., 67:301–320,
2005.
[22] J. Zylberberg, J. T. Murphy, and M. R. DeWeese. A sparse coding model with synaptically local plasticity
and spiking neurons can account for the diverse shapes of v1 simple cell receptive fields. PLoS Comput Biol,
7(10):e1002250, 2011.
13
| 9cs.NE
|
Under consideration for publication in Theory and Practice of Logic Programming
1
arXiv:cs/0505085v1 [cs.PL] 31 May 2005
Improving PARMA trailing
TOM SCHRIJVERS∗ and BART DEMOEN
Dept. of Computer Science, K.U.Leuven, Belgium
(e-mail: {toms,bmd}@cs.kuleuven.ac.be)
MARIA GARCIA DE LA BANDA
School of Computer Science and S.E, Monash University, Australia
(e-mail: [email protected])
PETER J. STUCKEY
NICTA Victoria Laboratory
Department of Computer Science and S.E.
University of Melbourne, Australia
(e-mail: [email protected])
submitted 20 November 2003; revised 22 October 2004; accepted 30 May 2005
Abstract
Taylor introduced a variable binding scheme for logic variables in his PARMA system,
that uses cycles of bindings rather than the linear chains of bindings used in the standard
WAM representation. Both the HAL and dProlog languages make use of the PARMA representation in their Herbrand constraint solvers. Unfortunately, PARMA’s trailing scheme
is considerably more expensive in both time and space consumption. The aim of this paper
is to present several techniques that lower the cost.
First, we introduce a trailing analysis for HAL using the classic PARMA trailing scheme
that detects and eliminates unnecessary trailings. The analysis, whose accuracy comes
from HAL’s determinism and mode declarations, has been integrated in the HAL compiler
and is shown to produce space improvements as well as speed improvements. Second, we
explain how to modify the classic PARMA trailing scheme to halve its trailing cost. This
technique is illustrated and evaluated both in the context of dProlog and HAL. Finally,
we explain the modifications needed by the trailing analysis in order to be combined with
our modified PARMA trailing scheme. Empirical evidence shows that the combination is
more effective than any of the techniques when used in isolation.
To appear in Theory and Practice of Logic Programming.
KEYWORDS: constraint logic programming, program analysis, trailing
1 Introduction
The logic programming language Mercury (Somogyi et al. 1996) is considerably
faster than traditional implementations of Prolog due to two main reasons. First,
∗ Research Assistant of the fund for Scientific Research - Flanders (Belgium)(F.W.O. - Vlaanderen)
Mercury requires the programmer to provide type, mode and determinism declarations whose information is used to generate efficient target code. And second,
variables can only be ground (i.e., bound to a ground term) or new (i.e., first time
seen by the compiler and hence unconstrained). Since neither aliased variables nor
partially instantiated structures are allowed, Mercury does not need to support full
unification; only assignment, construction, deconstruction and equality testing for
ground terms are required. Furthermore, it does not need to perform trailing, a
technique that allows an execution to resume computation from a previous program state: information about the old state is logged during forward computation
and used to restore it during backtracking. This usually means recording the state
of unbound variables right before they become aliased or bound. Since Mercury’s
new variables have no run-time representation they do not need to be trailed.
HAL (Demoen et al. 1999; de la Banda et al. 2002) is a constraint logic language
designed to support the construction, extension and use of constraint solvers. HAL
also requires type, mode and determinism declarations and compiles to Mercury
so as to leverage from its sophisticated compilation techniques. However, unlike
Mercury, HAL includes a Herbrand constraint solver which provides full unification. This solver uses Taylor’s PARMA scheme (Taylor 1991; Taylor 1996) rather
than the standard WAM representation (Aı̈t-Kaci 1991). This is because, unlike
the WAM, the PARMA representation of ground terms does not contain reference
chains and, hence, it is equivalent to that of Mercury. Thus, calls to the Herbrand
constraint solver can be replaced by calls to Mercury’s more efficient routines whenever ground terms are being manipulated.
Unfortunately, the increased expressive power of full unification comes at a cost,
which includes the need to perform trailing. Furthermore, trailing in the PARMA
scheme is more expensive than in the WAM, both in terms of time and space. We
present here two techniques to counter the trailing penalty of the PARMA scheme.
The first is a trailing analysis that detects and eliminates at compile-time unnecessary trailings and is suitable for any system based on the classic PARMA trailing
scheme. Without other supporting information such analysis is rather inaccurate,
since little is known at compile-time about the way predicates are used. However,
when mode and determinism information is available at compile-time, as in HAL,
significant accuracy improvements can be obtained. The second technique is a modified PARMA trailing scheme which considerably reduces the required trail stack
size. This technique can be applied to any PARMA-based system and has been
implemented by us in both dProlog (Demoen and Nguyen 2000) and the Mercury
back-end of the HAL system. Finally, we detail the modifications required by our
trailing analysis in order to be combined with our modified trailing scheme. The
empirical evaluation of each technique indicates that the combination of the modified trailing scheme with the trailing analysis results in a significant reduction of
trail size at a negligible time cost.
The rest of the paper proceeds as follows. The next section provides a quick background on trailing, the classic PARMA scheme, and when trailing can be avoided.
Section 3 summarizes the information used by our analyzer to improve its accuracy.
Section 4 presents the notrail analysis domain. Section 5 shows how to analyze
2
HAL’s body constructs. Section 6 shows how to use the analysis information to
avoid trailing. Section 7 presents the modified trailing scheme. Section 8 shows
the changes required by the analysis to deal with this modified scheme. Section 9
presents the results from the experimental evaluation of each technique. Finally,
future work is discussed in Section 10.
2 Background
We begin by setting some terminology. A bound variable is a variable that is bound
to some nonvariable term. An aliased variable is unbound and equated with some
other variable. A free variable is unbound and unaliased. We will also refer to a
new variable, which is a variable in HAL (and Mercury) which has no run-time
representation, since it is yet to be constrained.
In the WAM, an unbound variable is represented by a linear chain. If the variable
is free the chain has length one (a cell containing a self-reference). When two free
variables are unified, the younger cell is made to point to the older cell (see Section
2.2 for a discussion of relative cell age). These two variables are now aliased. A series
of unifications of free variables thus results in a linear chain of references of which
the last one is a self-reference or, in case the variable becomes instantiated, a bound
term. This representation implies that testing whether a (source level) variable is
bound or unbound, requires dereferencing. Such dereferencing is necessary during
each unification and it is thus performed quite often.
Example 1
Consider the execution of the goal X = Y, Z = W, X = Z, X = a when each variable is initially represented by a self-reference. Using the WAM representation, the
first unification points X at Y. The second unification points Z at W. In the third
unification we must first dereference X to get Y, dereference Z to give W, and then
point Y at W. In the last unification we dereference X and set W to a. The changes in
heap states are shown in Figure 1.
X
Y
X
Y
X
Y
X
Y
X
Z
W
Z
W
Z
W
Z
W
Z
(a) Initially
(b) X = Y
(c) Z = W
(d) X = Z
Y
a
W
(e) X = a
Fig. 1. Example of binding chains using the WAM representation.
In his PARMA-system (Taylor 1996), Taylor introduced a different variable representation scheme that does not suffer from this dereferencing need. In this scheme
an unbound variable is represented by a circular chain. If the variable is free the
3
chain has length one (a self-reference as in the WAM). Unifying two variables in
this scheme consists of cutting their circular chains and combining them into one
big circular chain. When the variable is bound, each cell in the circular chain is
replaced by the value to which it is bound. No dereferencing is required to verify
whether a cell is bound, because the tag in a cell immediately identifies the cell as
being bound or not. However, as we will see later, other costs are incurred by the
scheme.
Example 2
Consider the execution of the same goal X = Y, Z = W, X = Z, X = a when again
each variable is initially represented by a self-reference. Using the PARMA representation, the first unification points X at Y and Y at X. The second unification
points Z at W and W at Z. In the third unification we must point X at W and Z at
Y. In the final unification each variable in the chain of X is set to a. The changes
in heap states are shown in Figure 2. Notice how no references remain in the final
state, as opposed to Figure 1(e).
X
Y
X
Y
X
Y
X
Y
X
Z
W
Z
W
Z
W
Z
W
Z
(a) Initially
(b) X = Y
(c) Z = W
(d) X = Z
a
a
a
a
Y
W
(e) X = a
Fig. 2. Example of binding chains using the PARMA representation.
Another difference between the WAM and PARMA binding schemes becomes
apparent when constructing a new term containing an unbound variable X. Effectively, we are aliasing a new variable with X and, hence, this new variable must be
added into the variable chain of X.
Example 3
Consider the execution of the goal X = Y, Z = f(X) when each variable is initially
represented by a self-reference.
Using the WAM representation, the first unification points X at Y. The second
unification constructs a heap term f(X) with the content of X, namely Y, and points
Z at this.
Using the PARMA representation, the first unification chains X and Y together.
The second unification has to add the copy of X in f(X), to the chain for X. The
resulting heap states are shown in Figure 3.
As mentioned before, trailing is a technique that stores enough information regarding the representation state of a variable before each choice-point, to be able to
reconstruct such state upon backtracking. For both WAM and PARMA chains the
4
X
Y
X
Z
Z
(a) X = Y
Y
X
f
Y
Z
(b) Z=f(X)
X
Z
(c) X = Y
Y
f
(d) Z=f(X)
Fig. 3. Example of constructing a term containing an unbound variable using both
WAM (a)(b) and PARMA (c)(d) representations.
change of representation state occurs at the cell level: from being a self-reference
(when the variable represented by the cell – the associated variable – is unbound
and unaliased), to pointing to another cell in the chain (when the associated variable gets aliased), to pointing to the final bound structure (when the variable is
bound directly or indirectly). Thus, what needs to be trailed are the cells.
In the rest of the section we will discuss the PARMA trailing scheme in greater
detail, the orthogonal issue of conditional/unconditional trailing, and a possible
improvement based on compile-time detection of unnecessary trailings.
2.1 The classic PARMA Scheme: Value trailing
The classic PARMA trailing scheme uses value trailing, described by the following
C-like code:1
valuetrail(p) {
*(tr++) = *p; /* store the contents of the cell p */
*(tr++) = p; /* store the address of the cell p */
}
which takes the address p of a cell in a PARMA chain and stores in the trail stack
first the (old) contents of the cell and then its address. Here, tr is a global pointer
to the top of the trail stack.
The untrail operation for value trailing is straightforwardly defined by:
untrail_valuetrail() {
address = *(--tr); /* retrieve the cell address */
*address = *(--tr); /* recover the cell contents */
}
which first pops the address of a cell and then its contents.
In contrast, trailing in the WAM stores only the address of the cell. The reasons
are twofold. First, a cell is updated at most once, from a self-reference to a pointer to
1
All code in this paper is pseudo-C code. Implementation details that obfuscate rather than
clarify the concepts at hand, have been omitted.
5
either another cell in a linear chain or a structure. And second, for a self-referencing
cell the address and the content of the cell are the same. Therefore, when a cell
is updated the old content of the cell (which is the one stored during trailing) is
always the same as its address. This allows the WAM value trailing to be optimized
by only storing the address of the cell, reducing by half the space cost of a single
cell trailing.
Let us now discuss when cells need to be trailed in the classic PARMA scheme.
We have seen before that trailing is only needed when the representation state of a
variable changes, and that this can only happen when the variable is unbound and,
due to a unification, it becomes either aliased or bound. Therefore, we only need
to trail cells when their associated variables are involved in a unification or when
creating a new term which contains an unbound variable. The following discussion
distinguishes three cases: cells associated to variables involved in a variable–variable
unification, in a variable–nonvariable unification, and in new term construction.
Trailing during variable–variable unification: The result of aliasing two unbound
variables belonging to separate chains is the merging of the two chains into a single
one. This can be done by changing the state of only two cells: those associated to
each of the variables. Since each associated cell appears in a different chain, the
final chain can be formed by simply interchanging their respective successors. One
can then reconstruct the previous situation by remembering which two cells have
been changed and what their initial value was. This is achieved for unification X =
Y by the following (simplified) code:
valuetrail(X);
valuetrail(Y);
tmp = *X;
*X = *Y;
*Y = tmp;
Notice that X and Y are trailed independently. As only their associated cells need
to be trailed, we will refer to this kind of trailing as shallow trailing.
In contrast, for this kind of unification the WAM will update and trail the last
cell in just one of the two linear chains. Hence, the space cost is four times lower
(one value as opposed to four).
Example 4
Consider the PARMA trailing that occurs during the first three unifications of the
goal X = Y, Z = W, X = Z, X = a from Example 2, when each variable is initially
represented by a self-reference. From the first unification we trail X together with
its initial value (which, since X is a self-reference, is also) X, and Y together with its
initial value Y. Similarly, for the second unification we trail Z together with its value
Z, and W together with its value W. For the third unification, we trail X together with
its value Y, and Z together with its value W. The resulting trail is
X
X
Y
Y
Z
Z
W
6
W
Y
X
W
Z
The WAM trail for the same goal illustrated in Figure 1 trails first X, then Z and
finally Y. The resulting trail is X Z Y .
Trailing during variable–nonvariable unification: When an unbound variable becomes bound, every single cell in its chain is set to point to the nonvariable term.
Thus, we can only reconstruct the chain if all cells in the chain are trailed. The
combined unification-trailing (simplified) code for unification X = Term is as follows:
start = X;
do {
next = *X;
valuetrail(X);
*X = Term;
X = next;
} while (X != start);
Since all cells in the chain of the unbound variable are trailed, we will refer to
this kind of trailing as deep trailing.
In contrast, for this kind of unification, the WAM will trail again just one cell in
the linear chain. Hence, the space complexity for WAM is just O(1) compared to
O(n) for PARMA, where n is the number of cells in the chain. However, the time
complexity is O(n) for both, due to the dereferencing in the WAM.
Example 5
Consider the PARMA trailing that happens in the last unification X = a of the goal
from Example 2. The binding of all variables in the chain adds the trail elements
W
X
X
Y
Y
Z
Z
W
In contrast the WAM trailing adds a single trail element Y .
Trailing during new term construction: As mentioned before, when a new term is
constructed on the heap with a copy of an unbound variable X, the cell containing
this copy must be added into the chain for X. This means we must trail X since
its value (i.e., its successor in the chain) is going to change. We do not need to trail
the new cell since it clearly has no previous value we need to recover. The combined
construction-trailing (simplified) code for constructing f(X) where X is an unbound
variable and th is the current top of heap pointer, is:
*(++th) = *X;
valuetrail(X);
*X = th;
In contrast, for this construction the WAM need not trail at all since it simply
points the new cell at the old unbound variable.
If X is either a bound or a new variable, this complexity does not arise: X will
be placed in the new structure pointing to either the nonvariable term or to itself,
with no trailing required in any case.
7
Summary: The major advantage of the PARMA binding scheme is that it requires
no dereferencing, while its major disadvantages are (for a detailed account see
(Lindgren et al. 1995)):
1. PARMA trails more cells per unification: two in variable-variable unifications
and all in variable-nonvariable, versus one.
2. Trailing of an individual cell is more expensive: two slots used versus one.
3. Unlike in the WAM, cells can be trailed more than once: every time a cell is
updated which can happen more than once.
4. Copying an unbound variable into a structure involves trailing a cell.
As a result, the trail stack usage is expected to be much higher in the PARMA
scheme than in the WAM. Demoen and Nguyen (Demoen and Nguyen 2000) have
indeed observed in the dProlog system maximal trail sizes for the PARMA scheme
that are on average twice as large as with the WAM scheme. The techniques we
present in this paper attempt to counter the disadvantages. The trailing analysis
reduces the number of trailings and thereby counters disadvantages 1, 3 and 4,
while the modified trailing scheme counters disadvantage 2.
2.2 Conditional versus unconditional trailing
A cell that is changed only requires trailing if the cell did exist before the most
recent choice point since, otherwise, there is no previous state that has to be restored during backtracking. This property applies equally to the WAM and PARMA
schemes.
In some systems a simple run-time test can be used to verify whether a cell is
older than the most recent choice point. Younger cells require no trailing. If all
cells on the heap are kept in order of allocation, the test simply checks whether the
address of the cell is smaller than that of bh, the address of the top of the heap at
the beginning of the most recent choice point. Systems, such as dProlog, which take
advantage of this property use what is known as conditional trailing. Let us assume
the existence of function is older(p,bh) which succeeds if p < bh. Conditional
trailing is then described by the following code:
cond_valuetrail(p, bh) {
if (is_older(p,bh))
valuetrail(p);
}
thus avoiding the trailing of cells which are newer than the most recent choice point.
The code for variable–variable and variable–nonvariable unification described in the
previous sections using the unconditional valuetrail operation can be rewritten
to use conditional trailing by simply substituting each call to valuetrail by a call
to cond valuetrail. The untrail operation remains unchanged.
In systems where the order of cells on the heap is not guaranteed, unconditional
trailing is required. The Mercury back-end of the HAL system, for example, is
such a system since Mercury uses the Boehm garbage collector which does not
8
preserve the order of the cells on the heap between garbage collections. Other
systems use unconditional trailing at least during some unifications (see for instance (Van Roy and Despain 1992)). In (Demoen and Nguyen 2000) it is shown
that global performance is hardly affected by the choice between conditional or
unconditional trailing, since the savings made on avoided trailings are balanced by
the overhead of the run-time tests.
The differences between conditional and unconditional trailing do not affect the
proposed analysis. Thus, the same analysis can still be used if at some point conditional trailing becomes available in Mercury.
2.3 Unnecessary trailing in the classic PARMA scheme:
When considering the trailing of an unbound variable appearing in a unification,
there are at least two cases in which its trailing can be avoided:
• If the variable is new there is no previous value to remember and, therefore,
trailing is not required. This is in fact a subset of the cases exploited by
conditional trailing.
• The cells that need to be trailed (the associated cell in the case of variable–
variable, all cells in the case of variable–nonvariable) have already been trailed
since the most recent choice-point. Upon backtracking only the earliest trailing
after the choice-point is important, since that is the one which enables the
reconstruction of the state of the variable before the choice-point.
In the following sections we will see how compile-time analysis information can
be obtained to detect the above two cases and can therefore be used to (a) eliminate
unnecessary trailing in the classical PARMA trailing scheme, and (b) eliminate runtime tests performed by conditional trailing on variables known at compile-time to
have no representation and thus be younger than the most recent choice point.
3 Language Requirements
The analysis presented in this paper was designed for the HAL language. However,
it can be useful for any language that uses PARMA representation and that provides
accurate information regarding the following properties:
• Instantiation state: trailing analysis can gain accuracy by taking into account
the instantiation state of a program variable, i.e. whether the variable is new,
ground or old. State new corresponds to program variables with no internal
representation (equivalent to Mercury’s free instantiation). State ground corresponds to program variables known to be bound to ground terms. In any
other case the state is old, corresponding to program variables which might
be unbound but do have a representation (a chain of length one or more) or
bound to a term not known to be ground. Program variables with instantiation state new, ground or old will be called new, ground or old variables,
respectively. Note that once a new variable becomes old or ground, it can never
9
become new again. And once it is known to be ground, it remains ground.
Thus, the three states can be considered mutually exclusive. The information
should be available at each program point p as a table associating with each
variable in scope of p its instantiation state.
We will represent the instantiation table information at program point p as
follows. Let V arp denote the set of all program variables in scope at program
point p. The function instp : V arp → {new, ground, old} defines the instantiation state of program variable X at point p. This function allows us to
partition V arp into three disjoint sets: N ewp , Groundp and Oldp containing
the set of new, ground and old variables, respectively.
• Determinism: trailing analysis can also gain accuracy from the knowledge
that particular predicates have at most one solution. This information should
be available as a table associating with each predicate (procedure to be more
precise) its determinism. Herein we will refer to six main kinds of determinism:
semidet (minimum-maximum set of solutions: 0-1), det (1-1), multi (1-∞),
nondet (0-∞), erroneous (1,0), and failure (0-0).
For our purposes we will only be interested in whether a predicate can return
more than one answer. We will represent the determinism table by a function
det : P red → {0, 1, ∞} which maps each predicate q to its maximum number
of solutions.
• Sharing: trailing analysis can exploit sharing information to increase accuracy.
This information should be available at each program point p as a table
associating with each variable in scope of p the set of variables which possibly
share with it. Clearly, any variables that may be aliased together must possibly
share.
We will represent the sharing table at program point p by the function sharep :
Oldp → P(Oldp ) which assigns to each program variable in Oldp the set of
program variables in Oldp that share with it. Note that program variables in
N ewp and Groundp cannot share by definition.
4 The notrail Analysis Domain
The aim of the notrail domain is to keep enough information to be able to decide
whether the run-time variables in a unification need to be trailed or not, so that if
possible, optimized versions which do not perform the trailing can be used instead.
In order to do this, we must remember that only run-time variables which are
unbound and have a representation (i.e., are not new) need to be trailed. This
suggests making use of the instantiation information mentioned in the previous
section. Note that, since the analysis works on the level of program variables, some
indirection will be required.
We have already established that program variables in N ewp and Groundp represent run-time variables which do not need to be trailed. Thus, only variables in
Oldp need to be represented in the notrail domain. the set of new, ground and
old program variables, respectively. Assuming that V arp contains n variables and
10
the tree we have used to implement the underlying table is sufficiently balanced,
then, the size of the Oldp is O(n) and the complexity of instp is O(log n).
Recall that Oldp contains all program variables representing not only run-time
variables which are unbound and have a representation, but also run-time variables
bound to terms which the analysis cannot ensure to be ground. This is necessary
to ensure correctness: even though run-time variables which are bound do not need
to be trailed, the nonvariable terms to which they are bound might contain one
or more unbound run-time variables. It is the trailing state of these unbound runtime variables that is represented through the domain representation of the bound
program variable.
Now that we have decided which program variables need to be represented by
our domain, we have to decide how to represent them. We saw before that it is
unnecessary to trail a run-time variable in a variable–variable unification if its
associated cell has already been trailed, i.e., if the run-time variable has already
been shallow trailed since the most recent choice-point. For the case of variable–
nonvariable unification this is not enough, we need to ensure all cells in the chain
have already been trailed, i.e, the run-time variable has already been deep trailed.
This suggests a domain which distinguishes between shallow and deep trailed runtime variables. This can be easily done by partitioning Oldp into three disjoint
sets of program variables with a different trailing state: those representing runtime variables which might not have been trailed yet, those representing run-time
variables which have at least been shallow trailed, and those representing run-time
variables which have been deep trailed. It is sufficient to keep track of only two
sets to be able to reconstruct the third. Hence, the type of the elements of our
notrail domain Lnotrail will be P(Oldp ) × P(Oldp ), where the first component
contains the set of program variables representing run-time variables which have
already been shallow trailed, and the second component contains the set of program
variables representing run-time variables which have already been deep trailed.
In the following we will use l1 , l2 , . . . to denote elements of Lnotrail at program
points 1, 2, . . ., and s1 , s2 , . . . and d1 , d2 , . . . for the already shallow and deep trailed
components of the corresponding elements. Also, the elements of the domain will be
referred to as descriptions, with descriptions before and after a goal being referred
to as the pre- and post-descriptions, respectively.
Note that, by definition, we can state that if a run-time variable has already been
deep trailed, then it has also been shallow trailed (i.e., if all cells in the chain have
already been trailed, then the cell associated to the variable has also been trailed).
The partial ordering relation ⊑ on Lnotrail is thus defined as follows:
∀(s1p , d1p ), (s2p , d2p )
∈ Lnotrail :
(s1p , d1p )
⊑
(s2p , d2p )
⇔
s2p
d2p
⊆
⊆
d1p
d1p
∪
s1p
This implies that deep trailing is stronger information than shallow trailing, and
shallow trailing is stronger than no trailing at all. Also note that descriptions are
compared at the same program point only (so that the instantiation and sharing
information is identical). An example of a trailing lattice is shown in Fig. 4. Clearly
11
(∅, ∅) SS
SSSS
kk
kkkk
({Y }, ∅)
({X}, ∅)
SSSS
RRRR
kkkk
llll
(∅, {Y })
({X, Y }, ∅)
(∅, {X})
RRRR
SSSS
kkkk
llll
({Y }, {X})
({X}, {Y })
SSSS
kkkk
(∅, {X, Y })
Fig. 4. Notrail lattice Hasse diagram for variables {X, Y } where if l1 ⊑ l2 then l1
is below l2 in the diagram
(Lnotrail , ⊑) is a complete lattice with top description ⊤p = (∅, ∅) and bottom
description ⊥p = (∅, Oldp ).
There are two important points that need to be taken into account when considering the above domain. The first point is that the dp component of a description will
be used not only to represent already deep trailed variables but any variable in Oldp
which, for whatever reason (e.g., it has been initialized since the last choicepoint),
does not need to have any part of it trailed
The second point is that as soon as a deeply trailed program variable X is made
to share with a shallow trailed program variable Y , X also must become shallow
trailed since some cell in some newly merged chain might come from Y and thus
might not have been trailed. The sharing information at each program point is used
to define the following function which makes trailing information consistent with
its associated sharing information:
consistp ((s, d)) = (s ∪ x, d \ x)
where
x = {X ∈ d|(sharep (X) \ d) 6= ∅}
Intuitively, the function eliminates from d every program variable X which shares
with other variables not in d, and adds them to s. From now on we will assume that
∀(s, d) ∈ Lnotrail : consistp ((s, d)) = (s, d) and use the consist function to preserve
this property.2
Given HAL’s implementation of the sharing analysis domain ASub (Søndergaard 1986)
the time complexity of determining sharep (X) for a variable X is O(n2 ). Furthermore, since ASub explicitly carries the set of ground variables at each program point
(gp ), we will use this set rather than computing a new one (Groundp ) from the instantiation information, thus increasing efficiency. The major cost of consistp is the
computation of x: for each of the O(n) variables the sharep set has to be computed.
All other set operations are negligible in comparison. Hence, the overall time complexity is O(n3 ). We will see that the complexity of this function determines the
2
Note that the notrail domain can be seen as a “product domain” that also includes the mode
and sharing information. However, for simplicity, we will consider the different elements separately, relating them only via their associated program point.
12
complexity of all the operations that use it. Thus, we will use it only when strictly
necessary.
In summary, each element lp = (sp , dp ) in our domain can be interpreted as
follows. Consider a program variable X. If X ∈ dp , this means that all cells in
all chains represented by X have already been trailed (if needed). Therefore, X
does not need to be trailed in any unification for which lp is a pre-description.
Note that X could be a bound variable which includes many different variable
chains. If X ∈ sp we have two possibilities. If X is known to be unbound, then its
associated cell has been shallow trailed. Therefore, it does not need to be trailed
in any unification for which lp is a pre-description (although, in practice, we will
only consider optimizing variable-variable unifications). If X might be bound, then
a cell of one of its chains might not be trailed. As a result, no optimization can be
performed in this case.
We could, of course, represent bound variables more accurately, by requiring the
domain to keep track of the different chains contained in the structures to which
the program variables are bound, their individual trailing state and how these are
affected by the different program constructs. Known techniques (see for instance
(Janssens and Bruynooghe 1993; Van Hentenryck et al. 1995; Mulkers et al. 1994;
Lagoon and Stuckey 2001)) based on type information could be used to keep track
of the constructor that a variable is bound to and of the trailing state of the different
arguments, thereby making this approach possible.
5 Analyzing HAL Body Constructs with Lnotrail
This section defines the notrail operations required by HAL’s analysis framework (Bueno et al. 2001; Nethercote 2001) to analyze the different body constructs.
This framework is quite similar to the well known framework of (Bruynooghe 1991)
when analyzing a single module. While the analysis framework handles analysis
of multiple module programs, it makes no extra demands on the analysis domain.
Thus, for this paper we will simply treat the program to be analyzed as a single
module. For each body construct in HAL, we will show how to obtain the postdescription from the information contained in the pre-description.
Variable initialization init(X)
In HAL a variable X transits from its initial instantiation new to instantiation old
by being initialized. Since a new variable does not need to be trailed, we can simply
add X to the d component of the pre-description (recall that d not only represents
already deep trailed variables, but also any other old variable which does not need
to be trailed). Formally, let l1 = (s1 , d1 ) be the pre-description, the post-description
l2 can be obtained as:
l2 = (s1 , d1 ∪ {X})
Variable–variable unification: X = Y . There are several cases to consider:
• If one of the variables (say X) is new, it will simply be assigned a copy of
13
•
•
•
•
the pointer of Y . After the unification is performed, the trailing state of X
becomes that of Y . Thus, the trailing state of X in the post-description should
be that of Y in the pre-description. Note that this will never require a call to
consist since a new variable cannot introduce any sharing.
If one of the variables is ground, the other one will be ground after the unification. Hence, neither of them will appear in the post-description.
If both variables are deep trailed, all cells in their associated chains are trailed
and will remain trailed after unification (which is obtained by simply merging
the chains). Hence, all variables retain their current trailing state and the
pre-description will remain unchanged.
If both variables are already aliased (they belong to the same chain) nothing is
done by unification. Hence, they will retain the current trailing state. Hence,
all variables retain their current trailing state and the pre-description will
remain unchanged.
Otherwise, at least one of the variables is not deep trailed and two unaliased
variables are being considered. If both variables are unbound, unification will
merge both chains while at the same time performing shallow trailing if necessary. Thus, after the unification both variables will be shallow trailed. If
at least one variable is bound, the other one will become bound after the
unification. As stated earlier, bound variables can be treated in the same
way.
Note that if either variable was deep trailed before the unification, all shared
variables must become shallow trailed as well after the unification. This requires applying the consist function.
Formally, let l1 = (s1 , d1 ) be the pre-description and g2 be the set of ground
variables at program point 2 after the unification. Its post-description l2 can be
obtained as:
same(X, Y, l1 )
X is new
remove ground(l1 , g2 ) X is ground
l2 = unify(X, Y ) =
min(X, Y, l1 )
X and Y are old
unify(Y, X)
otherwise
with
(s1 ∪ {X}, d1 ) Y ∈ s1
same(X, Y, (s1 , d1 )) =
(s , d ∪ {X}) Y ∈ d1
1 1
(s1 , d1 )
otherwise
remove ground(li , vi ) = (si \ vi , di \ vi )
{X, Y } ⊆ d1
(s1 , d1 )
min(X, Y, (s1 , d1 )) =
consist2 ((s1 ∪ {X, Y }, d1 \ {X, Y })) X 6∈ share1 (Y )
(s1 , d1 )
otherwise
Here same(X, Y, li ) gives X the same trailing state as Y , remove ground(li , vi )
removes all variables in vi from li , and min(X, Y, li ) distinguished between three
cases. If X and Y are both deep trailed, nothing has to be changed. If X and Y
are definitely not aliased (they do not share) it ensures that they move to a shallow
14
X
f/1
X
f/1
(a) Before.
(b) After.
Fig. 5. Term construction example: f (X). The dashed line represents a choicepoint.
trailed state. Otherwise, the description must remain unchanged since unification
might have done nothing (and thus they might still be untrailed, so adding them
to s1 would be a mistake). Note that there is no need to apply consist here since
X and Y already share in the pre-description and, although sharing information
might have changed, it can only create sharing among variables already connected
(through X and Y ) by the closure under union performed by consist.
The worst case time complexity, O(n3 ), is again due to consist.
Variable–term unification: Y = f (X1 , . . . , Xn ). There are two cases to consider: If
Y is new, the unification simply constructs the term in Y . Otherwise, we can treat
this for the purpose of the analysis as two unifications, Y ′ = f (X1 , . . . , Xn ), Y = Y ′
where Y ′ is a new variable. Since unifications of the form Y ′ = Y have been
discussed above, here we only focus on the construction into a new variable. In the
following we assume that the Y in the variable-term unification is new.
When a term, e.g. f (X), is constructed with X being represented by a PARMA
chain, the argument cell in the structure representation of f /1 is inserted in the
chain of X (see Fig.5). While X requires shallow trailing, the cell of the term
requires no trailing at all as it is newly created.
The generalization of this to an n-ary variable term unification is as follows. If
all arguments are deep trailed, then Y becomes deep trailed and the arguments
remain deep trailed. Otherwise, Y and all its arguments become shallow trailed
(since each argument is at least shallow trailed by the operation). Note that if at
least one argument was deep trailed, and since each argument shares with Y after
the unification, we must apply consist to maintain the information consistent.
Formally, let l1 = (s1 , d1 ) be the pre-description of the unification, x be the set
of variables {X1 , . . . , Xn } and g2 the set of ground variables after the unification.
Its post-description l2 can be obtained as:
(s1 , d1 ∪ {Y })
x ⊆ d1
l2 =
consist2 (remove ground((s1 ∪ x ∪ {Y }, d1 \ x), g2 )) otherwise
The worst case time complexity is O(n3 ). This definition can be combined with
the previous one for the overall definition of variable–term unification. The implementation can be more efficient, but the complexity will still be O(n3 ).
15
Predicate call: q(X1 . . . Xn ). Let l1 be the pre-description of the predicate call and
x the set of variables {X1 , . . . , Xn }. The first step will be to project l1 onto x
resulting in description lproj . Note that onto-projection is trivially defined as:
onto proj(l, v) = (s ∩ v, d ∩ v)
The second step consists in extending lproj onto the set of variables local to the
predicate call. Since these variables are known to be new (and thus they do not
appear in Old1 ), the extension operation in our domain is trivially defined as the
identity. Thus, from now on we will simply disregard the extension steps required
by HAL’s framework.
Let lanswer be the answer description resulting from analyzing the predicate’s
definition for calling description lproj . We will assume that the set v of variables
local to q/n has already been projected out from lanswer , where out-projection is
identical to remove ground, which has time complexity O(n).
In order to obtain the post-description, we will make use of the determinism
information. Thus, the post-description l2 can be derived by combining the lanswer
and l1 , using the determinism of the predicate call as follows:
• If the predicate has determinism multi or nondet (which can have more than
one answer), then all variables not in x become not trailed by the (possible)
introduction of a new choice point. Hence, l2 is equal to lanswer except for the
fact that we have to apply the consist function in order to take into account
the changes in sharing involving variables not in x.
• Otherwise, we know the trailing state of variables in l1 is unchanged except by
possibly new introduced sharing. Thus, l2 is the result of combining lanswer
and l1 as follows: the trailing state of variables in x is taken from lanswer ,
while that of other variables is taken from l1 . Any deep trailed variables that
share with non-deep trailed variables must, of course, become shallow trailed.
Formalized, the combination3 function is defined as:
l2
=
=
comb(l1 , lanswer )
consist2 (((s1 \ x) ∪ sanswer , (d1 \ x) ∪ danswer ))
consist2 (lanswer )
det(q) ≤ 1
otherwise
Obviously, the complexity is O(n3 ) because of consist.
Example 6
Assume that the call q(X) has pre-description ({X, Y }, ∅) and the predicate q/1
has answer description ({X}, ∅). The post-description of the call depends on the
determinism of the predicate. If the predicate q/1 has at most one solution, the postdescription will be (({X, Y } \ {X}) ∪ {X}, (∅ \ {X}) ∪ ∅) = ({X, Y }, ∅). Otherwise
the post-description will be equal to the answer description, ({X}, ∅).
3
Note that the combination is not the meet of the two descriptions. It is the “specialized combination” introduced in (de la Banda et al. 1998) which assumes that lanswer contains the most
accurate information about the variables in x, the role of the combination being just to propagate this information to the rest of variables in the clause.
16
Disjunction: (G1 ; G2 ; . . . ; Gn ). Disjunction is the reason why trailing becomes necessary. As mentioned before, trailing might be needed for all variables which were
already old before the disjunction. Thus, let l0 be the pre-description of the entire
disjunction. Then, ⊤ will be the pre-description of each Gi except for Gn whose
pre-description is simply l0 (since the disjunction implies no backtracking over the
last branch).
Let li = (si , di ), 1 ≤ i ≤ n be the post-description of goal Gi . We will again
assume that the set vi of variables local to each Gi has already been projected out
from li . The end result ln+1 of the disjunction is the least upper bound (lub) of all
branches,4 which is defined as:
l1 ⊔ . . . ⊔ ln = consistn+1 (remove ground((s, d), gn+1 ))
where
s
d
s′i
d′i
=
=
=
=
(s′1 ∩ . . . ∩ s′n ) \ d
(d′1 ∩ . . . ∩ d′n )
si ∪ d′i
di ∪ gi
Intuitively, all variables which are deep trailed in all descriptions are ensured
to remain deep trailed; all variables which are trailed in all descriptions but have
not always been deep trailed (i.e., are not in d) are ensured to have already been
(at least) shallow trailed. Note that variables which are known to be ground in all
descriptions (those in gn+1 ) are eliminated. This is consistent with the view that
only old variables are represented by the descriptions and avoids adding overhead
to the abstract operations.
HAL also includes switches, which are disjunctions where the compiler has detected that only one branch needs to be executed. Switches are treated identically
to disjunctions except for the fact that the pre-description for each Gi is l0 rather
than ⊤.
Example 7
Let l0 = (∅, {X, Y, Z}) be the pre-description of the code fragment:
( A = a, X = Y ; A = b, X = f(Y, Z) )
Let us assume there is no sharing at that program point. Assuming that A is old,
then this is simply a disjunction. Then, the pre-descriptions of the first branch is
(∅, ∅), the ⊤ element of our domain. The pre-description of the second branch is
(∅, {X, Y, Z}), i.e., since this is the last branch in the disjunction, its pre-description
is identical to the pre-description of the entire disjunction. Their post-descriptions
are ({X, Y }, ∅) and (∅, {X, Y, Z}), respectively. Finally, the lub of the two postdescriptions results in ({X, Y }, ∅).
Now assume A is ground. Then this code fragment is a switch on A. The predescription for the first branch becomes (∅, {X, Y, Z}) and the post description
4
Note that this is not the lub of the notrail domain alone, but that of the product domain
which includes sharing (and groundness) information.
17
is the same. Finally the lub of the two post-descriptions for the two branches is
(∅, {X, Y, Z}).
The time complexity of the joining of the branches is simply that of the lub
operator (O(n3 )) for a fixed maximum number of branches, and it is completely
dominated by the consistn+1 function.
If-then-else: I → T ; E . Although the if-then-else could be treated as (I, T ; E),
this is rather inaccurate since (as in the case of switches) only one branch will ever
be executed and, thus, there is no backtracking between the two branches.
Hence, we can do better if no old variable that exists before the if-then-else is
bound or aliased, i.e. possibly requiring trailing and backtracking if the condition
fails. This is not a harsh restriction, since it is ensured whenever the if-condition
is used in a logical way, i.e., it simply inspects existing variables and does not
change any non-local variable. However, in general it is not possible to statically
determine this property. Instead a safe approximation is used: the if-then-else is
treated as (I, T ; E) if the condition contains any pre-existing old variables, otherwise
the following stronger treatment is used.
Let l1 be the pre-description to the if-then-else. Then l1 will also be the predescription to both I and E. Let lI be the post-description obtained for I. Then lI
will also be the pre-description of T . Finally, let lT and lE be the post-descriptions
obtained for T and E, respectively. Then, the post-description for the if-then-else
can be obtained as the lub lT ⊔ lE .
The time complexity of the joining of the branches is again O(n3 ), just like the
operation over the disjunction.
Example 8
Let l0 = (∅, ∅) be the pre-description of the following if-then-else where N is known
to be ground:
( N = 1 -> X = Y
; X = f(Y, Z) )
Assume no variables share before the if-then-else. Then, l0 is equal to the predescription
of
both
the
thenand
else-branch.
The
post-description of the then-branch is ({X, Y }, ∅) and that of the else-branch is
({X, Y, Z}, ∅). The post-description finally is obtained as their lub: ({X, Y }, ∅).
If the pre-description was l0 = (∅, {X, Y, Z}) as in Example 7, then the postdescription would be (∅, {X, Y, Z}), since no additional trailing will be required.
Higher-order term construction: Y = p(X1 , . . . , Xn ). This involves the creation of
a partially evaluated predicate, i.e., we are assuming there is a predicate with name
p and arity equal or higher than n for which the higher-order construct Y is being
created. In HAL, Y is required to be new. Also, it is often too difficult or even
impossible to know whether Y will be actually called or not and, if so, where.
Thus, HAL follows a conservative approach and requires that the instantiation of
the “captured” arguments (i.e., X1 , . . . , Xn ) remain unchanged after calling Y . It
18
also guarantees (through type and mode checking) that no higher-order terms are
ever unified.
The above requirements allow us to follow a simple (although conservative) approach: Only after a call to Y will the trailing of the captured variables be affected.
If the call to Y might have more than one solution and thus may involve backtracking, then the involved variables will be treated safely in the analysis at the
call location if they are still statically live there.
If the call to Y does not involve backtracking but does involve unifications, then
trailing information might not be inferred correctly at the call location. This is
because the captured variables are generally not known at the call location. To
keep the trailing information safe, any potential unifications have to be accounted
for in the higher-order unification. Since the construction of the higher-order term
involves no backtracking and all unifications leave the variables they involve at
least shallow trailed, it is sufficient to demote all captured deep trailed variables to
shallow trailed status, together with all sharing deep trailed variables.
Formally, let l1 = (s1 , d1 ) be the pre-description of the higher-order term construction and x be the set of variables {X1 , . . . , Xn }. Then its post-description l2
can be obtained with a time complexity of O(n3 ) as:
consist2 ((s1 ∪ (x ∩ d1 ), d1 \ x)) x ∩ d1 6= ∅
l2 =
l1
otherwise
Higher-order call: call(P, X1 , . . . , Xn ). The exact impact of a higher-order call is
difficult to determine in general. Fortunately, even if the exact predicate associated
to variable P is unknown, the HAL compiler still knows its determinism. This can
help us improve accuracy. If the predicate might have more than one solution, all
variables must become not trailed. Since the called predicate is typically unknown,
no answer description is available to improve accuracy.
Otherwise, the worst that can happen is that the deep trailed arguments of the
call become shallow trailed. So in the post-description we move all deep trailed
arguments to the set of shallow trailed variables, together with all variables they
share with. Recall that for this case the captured variables have already been taken
care of when constructing the higher-order term.
The sequence of steps is much the same as that for the predicate call. First, we
project the pre-description l1 onto the set x of variables {X1 , . . . , Xn }, resulting in
lproj . Next, the answer description lanswer of the higher-order call is computed as
indicated above:
(s ∪ d, ∅) det(P ) ≤ 1
lanswer =
(∅, ∅)
otherwise
The combination of lanswer and l1 is computed to obtain the post-description l2 .
6 Trailing Optimization
The optimization phase consists of deciding for each unification in the body of
a clause which variables need to be trailed. This decision is based on the pre19
description of the unification, inferred by the trailing analysis. If some variables do
not need to be trailed, the general unification predicate is replaced with a variant
that does not trail those particular variables. Thus, we will need a different variant
for each possible combination of variables that do and do not need to be trailed.
• For the unification of two unbound variables, trailing is omitted for either
variable if it is shallow trailed or deep trailed in the pre-description.
• For the binding of an unbound variable X, trailing of X is omitted if it is
deep trailed in the pre-description.
• In the construction of a term containing an old unbound variable X, trailing
of X is omitted if X is either shallow or deep trailed in the pre-description.
• For the unification of two bound variables, the trailing for chains in the structure of either is omitted if it is deep trailed in the pre-description.
Often it is not known at compile time whether a variable is bound or not, so a
general variable-variable unification predicate is required that performs run-time
boundness tests before selecting the appropriate kind of unification. Various optimized variants of this general predicate are needed as well.
Experimental results for the analysis are presented in Section 9.
7 The improved trailing scheme
Let us now present a trailing scheme which is more sophisticated than the classic
PARMA value trailing discussed in Section 2. We will start by considering the
improvements that apply to each kind of unification (variable–variable and variable–
nonvariable) and finish by showing how to combine them.
Our modified scheme must be able to apply different untrail operations depending
on the kinds of trailing that was performed. A simple tagging scheme (explained
in detail in Section 7.3) is used to indicate the kind of untrailing required in each
case.
7.1 Variable–variable unification: swap trailing
In the classic scheme the value trailing of both cells takes up four trail stack slots
(two for the addresses of each variable plus another two for their contents) when
trailing is unconditional. Undoing such variable–variable unification consists of simply restoring the old values of the cells separately. However, there is a more economic
inverse operation that undoes the swapping that happened during unification: simply swapping back. This swapping only requires the addresses of the involved cells
and not their respective old contents. We introduce a new kind of trailing named
swap trailing which exploits this and also the corresponding untrailing operation.
Swap trailing is defined by the following code:
swaptrail(p, q) {
*(tr++) = p;
*(tr++) = set_tag(q,SWAP_TRAIL);
}
20
where p and q are the addresses of the two cells, tr is a pointer to the top of
the trailing stack, SWAP TRAIL is a tag, and the function set tag(c,t) tags cell c
with tag t. Note that swap trailing only consumes two slots in the trail stack, as
opposed to the four used by (unconditional) value trailing in the classical scheme.
The untrail operation for swap trailing is:
untrail_swaptrail() {
q = untag(*(--tr)); /* recover address q */
p = *(--tr);
/* recover address p */
tmp = *q;
*q = *p;
/* swap contents of p with q */
*p = tmp;
}
The above improvement assumes that both cells are unconditionally trailed. If conditional value trailing is available, the classic scheme would either consume zero,
two or four slots if respectively none, only one or both variables are older than the
most recent choice point. Swap trailing can only be used in conjunction with conditional trailing to replace the four slot case, with value trailing still needed for the
two slot case. As a result the code for conditional variable–variable trailing looks
like:
cond_varvartrail(p, q, bh) {
if (is_older(p,bh)) {
if (is_older(q,bh)) {
swaptrail(p,q); /* trail both using swaptrail */
} else {
valuetrail(p); /* only trail p */
}
} else if (is_older(q,bh)) {
valuetrail(q);
/* only trail q */
}
}
It is important to note that the potential gain in space on the trail obtained by the
above operations comes at a cost in execution time (more run-time operations are
needed) and that the gain in space is not guaranteed.
7.2 Variable–nonvariable unification: chain trailing
As seen before, variable-nonvariable unification pulls the entire chain of the variable
apart by setting every cell in the chain to the nonvariable. In the case of classic
value trailing, every address of a cell is stored twice: once as the address of a cell
and once as the contents of the predecessor cell. This means that there is quite
some redundancy. The obvious improvement is to store each address only once. We
name this chain trailing. Because the length of the chain is not known, a marker is
needed to indicate, for the untrailing operation, where chain trailing ends. The last
21
entry of the chain encountered during untrailing, is the first one actually trailed.
We use the CHAIN END tag to mark this entry.
The last address put on the trail is tagged with CHAIN BEGIN to indicate the kind
of trailing. For chains of length one, the last and first cell coincide. The CHAIN END
tag is used to mark this single address.
Chain trailing is defined by the code:
chaintrail(p) {
start = p;
*(tr++) = set_tag(p,CHAIN_END);
p = *p;
only_one = TRUE;
while (p != start) {
/*trail each cell address*/
only_one = FALSE;
*(tr++) = p;
p = *p;
}
if (!only_one) {
/* if more than one cell */
last = tr - 1;
/* tag last one as CHAIN_BEGIN*/
*last = set_tag(*last,CHAIN_BEGIN);
}
}
The untrail operation for reconstructing the chain is straightforward: it dispatches
to the appropriate untrailing action depending on the tag of the first cell encountered during untrailing. If this is CHAIN BEGIN, meaning n ≥ 1, the corresponding
code is:
untrail_chaintrail() {
head = untag(*(--tr));
previous = head;
current = *(--tr);
while (get_tag(current) != CHAIN_END) {
*current = previous;
previous = current;
current = *(--tr);
}
current = untag(current);
*current = previous;
*head = current;
}
If the first tag is CHAIN END, then n = 1 and the code for untrailing is:
untrail_shortchain() {
cell = untag(*(--tr));
*cell = cell;
}
22
Example 9
Consider the trailing that occurs using the improved scheme for the goal X = Y, Z
= W, X = Z, X = a from Example 4. The first unification is a swaptrail, trailing X
and Y, similarly the second unification swaptrails Z and W and the third unification
swap trails X and Z. Finally the last unification chain trails X. The resulting trail
looks like:
X
Y
sw
Z
W
sw
X
Z
sw
X
ce
W
Z
Y
cb
where we use superscripts sw, cb and ce to represent the SWAP TRAIL, CHAIN BEGIN
and CHAIN END tags respectively. This uses 10 trail entries compared to the 24
entries in Examples 4 and 5
The above improvement assumes that all cells are unconditionally trailed. Let us
assume that the chain consists of n cells, k of which are older than the most recent
choice point. If conditional trailing is available and 2 ∗ k < n, our unconditional
chain trailing will consume more space than the classic conditional value trailing.
Fortunately, a conditional variant of chain trailing is also possible:
cond_chaintrail(p, bh) {
start = p;
first = TRUE;
only_one = TRUE;
do {
if (is_older(p,bh))
/* trail each older cell in chain*/
if (first) {
*(tr++) = set_tag(p,CHAIN_END); /*tag if first*/
first = FALSE;
} else {
only_one = FALSE;
*(tr++) = p;
}
p = *p;
} while (p != start);
if (!only_one) {
/* if more than one older cell */
last = tr - 1;
/* tag last one as CHAIN_BEGIN*/
*last = set_tag(*last,CHAIN_BEGIN);
}
}
This conditional variant uses only k slots of the stack trail, so it is clearly an
improvement over conditional value trailing whenever k > 0.
Note that the untrail operation used is the same as for the unconditional chain
trailing. This might look wrong at first since the cond chaintrail might not trail
all cells in the chain. However, this is simply exploiting the fact that the objective
of trailing is to be able to reconstruct the bindings that existed at the creation time
of a choice point. Thus, the final state of younger cells and the state of any cell
23
during the intermediate steps of untrailing are irrelevant. In fact, the more general
– and better with respect to stack trail consumption – principle behind this is that
only the old cells (older than the most recent choice point) in the chain pointing
to other old cells have to be trailed (an old cell must have been made to point to a
new cell after the last choice-point). The kind of trailing suitable for this insight is a
special kind of value trailing, where the successive equal slots on the trail stack are
overlapped. The above cond chaintrail operation only approximates this, since
an implementation would incur an undue time overhead because of the extra runtime tests needed to test the age of the successors. Thus, we store the addresses of
old cells even if they neither point to nor are pointed to by old cells.
Example 10
Figure 6 illustrates with a small example how the above specified conditional chain
trailing, together with previous trailings, safely restores the state of all variables
older than the most recent choice point. Consider the following goal X = Z, Z =
Y, X = a, fail and let us assume that both X and Y are older than the most
recent choice point, Z is newer, and all three are chains of length 1 as depicted
in Figure 6(a). The successive forward steps are shown in the Figures 6(b), 6(c)
and 6(d). X is value trailed during X = Z, as is Y during Z = Y. The addresses of X
and Y are stored on the trail stack with conditional chain trailing during X = a5 .
The cb and ce to the side of the stack trail entries represent the CHAIN BEGIN and
CHAIN END tags respectively.
The execution fails immediately after X = a, and backtracks to the initial state in
three steps. First (Figure 6(e)), the conditional chain trailing is untrailed, creating
a chain of X and Y. Next (Figure 6(f)), the value trailing of Y is undone and finally
(Figure 6(g)), the value trailing of X is reversed too. The final state corresponds
to the initial state, except for Z, which is still bound to a. However, as Z did
not exist before the most recent choice point, its content is irrelevant at that point
because it is inaccessible and will be reclaimed from the heap anyway when forward
execution resumes. Note the illegal intermediate state illustrated in Figure 6(f) is
not important since it only occurs in the middle of untrailing, and never during
execution.
7.3 Combining the improvements
Let us first consider the combination in the context of the modified unconditional
trailing scheme of the Mercury back-end of HAL. In this context, in addition to
swap and unconditional chain trailing, function trailing is used to allow custom
trailings for constraint solvers. Function trailing stores a pointer to an untrailing
function and to untrailing data. Thus, we need four different tags to distinguish
the different trailing information that can appear on the trail. Fortunately, there
are two tag bits available (because of the aligned addressing for 32 bit machines).
There is one constraint on the allocation of the four different tags to the kinds of
5
This could be avoided if X is known to have been trailed already.
24
X
Y
X
Z
Y
X
X
X
Z
(a) Initially
X
Z
a
Y
Y
X
X
(e) Untrail X = a
Y
Z
(b) X = Z
Y
X
Y
a
(f) Untrail Z = Y
a
Y
a
Y
X
Y
Y
X
X
cb
ce
(d) X = a
X
X
X
a
Z
(c) Z = Y
X
Z
Y
Y
X
X
Y
Z
a
(g) Untrail X = Z
Fig. 6. Conditional chain trailing example.
trailing: the CHAIN END tag should not look the same as the tag of the intermediate
addresses in a chain trail.
The general untrail operation then simply looks like:
untrail(tr_cp) {
while (tr > tr_cp) {
switch (get_tag(*tr)) {
case FUNCTION_TRAIL:
untrail_functiontrail();
break;
case SWAP_TRAIL:
untrail_swaptrail();
break;
case CHAIN_BEGIN:
untrail_chaintrail();
break;
case CHAIN_END:
untrail_shortchain();
}
}
}
Note that, since we are assuming we are in a modified unconditional trailing scheme,
value trailing is never used. This is because value trailing is only needed in the
modified scheme whenever only one of the two variables involved in a variablevariable unification is newer than the most recent choice point, and thus only that
25
one was trailed. Otherwise swap trailing will be used. Since no conditional trailing
is allowed, swap trailing is always used for variable-variable unifications.
Let us now consider the combination in the context of the modified conditional
trailing scheme of dProlog. In this context only value, swap and conditional chain
trailing are used. The remarks on the application and allocation of tags is the
same as for the unconditional case and the general conditional untrail operation
looks identical except for the fact that the FUNCTION TRAIL case is substituted by
a VALUE TRAIL case, and the call to untrail functiontrail() is substituted by a
call to untrail valuetrail().
When looking at the value trailings of chains of length one in the example in
the previous section (see Figure 6), there is an obvious trailing alternative in the
conditional system that stores no redundant information: chain trailing. Indeed, if
such a variable would be chain trailed instead of value trailed, only one instead of
two slots would be used on the stack. However, this would require more run-time
tests and we have not implemented this.
value_trail(p) {
if (*p == p) /* self pointer */
*(tr++) = set_tag(p,CHAIN_END);
else
*(tr++) = *p;
*(tr++) = p;
}
Experimental results for both the conditional and unconditional trailing scheme
are presented in Section 9.
8 Analysis for the improved trailing scheme
Trailing analyses heavily depend on the details of the trailing scheme. The analysis
presented in Section 4 was defined for the classic PARMA trailing scheme. In this
section we present the modifications needed by that analysis in order to be applied
to our improved trailing scheme. As we will see, the improved scheme gives rise to
fewer opportunities for trail savings.
8.1 Unnecessary trailing in the improved trailing scheme
The main difference between the two schemes in terms of unnecessary trailing appears when considering cells that have been trailed since the most recent choicepoint. In the case of value- and chain-trailing, these cells do not need to be trailed
again since the information stored the first time allows us to reconstruct the state
right before the choice-point.6 As we will see later in the experimental evaluation,
this allows our previous analysis to detect many spurious trailings.
6
This is assuming that the semantics of function trailing is such that it does not rely on the
intermediate state of any Herbrand variable during untrailing.
26
In the case of swap trailing, however, cells need to be trailed even if they have
already been trailed since the most recent choice-point. This is because swap trailing
is an incremental kind of trailing (the content of the cells is not stored during
the trailing, but only the incremental change) and thus relies on future trailings
for proper untrailing of cells. As a result, during the untrailing process in our
improved scheme, all later chain and swap trailings have to be undone before the
swap trailing can be untrailed correctly. Thus, there is no opportunity here to
avoid future trailings between two choice points, after the first trailing has been
performed. Let us illustrate this with a counterexample.
Counterexample. Let us not trail variables a second time between two choice points.
Consider then the following code:
X = Y, Z = W, X = Z, fail
where all variables are older than the most recent choice point and, initially, they
are represented as chains of length one, as depicted in Figure 7(a). In the first two
steps the four variables are aliased and swap trailed pairwise, creating two chains
of length two (see Figure 7(b)). The s’s represent SWAP TRAIL tags.
X
Z
Y
W
(a) Initially
W
Z
Y
W X
X
Y
Z
(b)
Z = W
s
X
s
Z
W
Z
Y
W X
Y
s
(c) X = Z
X = Y,
X
Y
Z
W
s
(d) Untrail
Fig. 7. Counterexample of incremental behavior of swap trailing: it does not eliminate the need for further trailing of the same cells.
Next X and Z are aliased, creating one large chain (see Figure 7(c)). During
this step X and Z are not (swap) trailed since they have already been swap trailed
after the most recent choice point (and we are assuming this means trailing is not
needed). Finally, the execution fails and untrailing tries to restore the situation at
the most recent choice point. However, Figure 7(d) shows that the omission of the
last swap trailing was invalid, as untrailing fails to restore the correct situation.
Thus, a cell involved in swap trailing still needs trailing later in the same segment
of the execution.
8.2 The Ltrail analysis domain
The implications for the Ltrail analysis domain are simple: it only needs to distinguish between variables that do not have to be trailed again (deep trailed) and those
which have to (rest). In other words, variables can only have one of two possible
27
states at a particular program point: deep trailed or not trailed at all. Hence, the
type of elements of our Ltrail domain will be P(Oldp ). The ordering ⊑ is simply
⊇.
All the operations we have defined for the Lnotrail domain have to be adapted to
this simplification. This adaptation is rather straightforward: every description l in
Ltrail is treated as if it were the description (∅, l) in Lnotrail , and new descriptions l′
in Ltrail are obtained by first calculating the (s′ , d′ ) descriptions using the Lnotrail
operations and then setting l′ = d′ .
8.3 Optimization based on the analysis
Again, the pre-description of every unification is used to improve that unification.
The possible optimizations based on the Ltrail domain are more limited than those
for the Lnotrail domain, as only deep trailed variables are represented in the descriptions:
• For the unification of two variables, a variant without (swap) trailing can be
used if both variables are in the pre-description (i.e. deep-trailed).
• For the binding of an unbound variable Y to a term f (X1 , ..., Xn ), a variant
of the unification without (chain) trailing can be used if Y is in the predescription. In addition, no (swap) trailing is required for any of the Xi that
appear in the pre-description.
• For the unification of two bound variables, if both variables are in the predescription, or if one is in the pre-description and the other is known to be
ground, then no trailing is needed at runtime. This means that if during the
recursive unification process of the bound variables, unbound variables are
unified or bound, nothing will need to be trailed for these unbound variables.
9 Experimental Results
We first examine the effect of the trailing analysis Lnotrail and its associated optimizations on the classic PARMA trailing scheme for HAL. We then look at the
effect of the improved PARMA trailing scheme, and at the effect of the use of the
trailing analysis Ltrail on the improved PARMA trailing scheme. Finally, we examine the improved PARMA trailing scheme in the context of dProlog. All timing
results were obtained on an Intel Pentium 4 2.00 GHz 512 MB.
9.1 Effect of trailing analysis using Lnotrail in HAL
The Lnotrail analyzer has been implemented in the analysis framework of HAL and
applied to six HAL benchmarks that use the Herbrand solver: icomp, hanoi, qsort,
serialize, warplan and zebra. Table 1 gives a summary of these benchmarks. All
benchmarks make use of the Herbrand solver and cannot be executed as Mercury
programs (without significantly modifying the algorithm and representation).
The pre-descriptions inferred for the unifications of these benchmarks have been
28
Table 1. HAL Benchmark descriptions and lines of code
Benchmark
icomp
hanoi
qsort
serialize
warplan
zebra
Description
a cut down version of the interactive BIM compiler
the Hanoi puzzle using difference lists
the quick sort algorithm using difference lists
the classic Prolog palindrome benchmark
war planner for robot control
the classic five houses puzzle
Lines
294
31
43
74
316
82
Table 2. Compilation statistics for notrail analysis
Benchmark
icomp
hanoi
qsort
serialize
warplan
zebra
Compilation Time
Analysis Total Relative
1.170 2.110
55.5 %
0.030 0.350
8.6 %
0.020 0.810
2.5 %
0.040 0.430
9.3 %
1.080 2.590
41.7 %
0.090 0.560
16.1 %
Old unifications
Improved Total Relative
314 1,542
20.4 %
13
13 100.0 %
7
7 100.0 %
1
20
5.0 %
93 1,347
6.9 %
40
177
22.6 %
Size
Relative
120.5 %
100.0 %
100.0 %
100.2 %
156.2 %
108.6 %
used to optimize the generated Mercury code by avoiding unnecessary trailing, as
explained in Section 6.
Table 2 shows, for each benchmark, the analysis time in seconds compared to
the total compilation time, the number of improved unifications compared to the
total number of unifications involving old variables, and the size of generated binary
executable. The binary size of the optimized program is expressed as the number
of bytes relative to the unoptimized program.
The high compilation times obtained for some benchmarks are due to the existence of predicates with many different pre-descriptions, something the analysis
has not been optimized for yet. The deterministic nature of both hanoi and qsort
benchmarks, allows the analysis to infer that all unifications should be replaced
by a non-trailing alternative. In the other benchmarks a much smaller fraction of
unifications can be improved due to the heavy use of non-deterministic predicates.
The last table shows that due to the multi-variant specialization, there may
be a considerable size blow-up. In particular, for icomp and warplan the size is
substantially increased. Various approaches to limit the number generated variants,
explored in other work, apply to this work as well. For example, one approach is
to use profiling information to only retain the most performance-critical variants
(see (Ferreira and Damas 2003)). Another approach, taken in (Mazur 2001), is to
only generate the most and least optimized variants. The latter would reproduce
the optimal result for hanoi and qsort.
Table 3 presents the execution times in seconds obtained by executing each benchmark a number of times in a loop; the iteration number in the table gives that loop
29
Table 3. Benchmark timings for classic PARMA: unoptimized (cparma) and optimized with trailing analysis (caparma)
Benchmark
icomp
hanoi
qsort
serialize
warplan
zebra
Iterations
10,000
10
10,000
10,000
10
200
cparma
0.834
0.990
0.363
0.901
1.293
1.239
Time
caparma
0.790
0.707
0.303
0.884
1.407
1.254
relative
94.7 %
71.4 %
83.5 %
98.1 %
108.8 %
101.2 %
Table 4. Benchmark trail sizes for classic PARMA: unoptimized (cparma) and optimized with trailing analysis (caparma)
Benchmark
icomp
hanoi
qsort
serialize
warplan
zebra
Maximum trail
cparma caparma relative
5,545
4,217
76.1 %
61,441
0
0.0 %
11,801
0
0.0 %
16,569
12,657
76.4 %
17
9
52.9 %
209
185
88.5 %
Trailing operations
cparma
caparma relative
1,110
860
77.5 %
7,864,300
0
0.0 %
1,510
0
0.0 %
2,120
1,620
76.4 %
102,290
101,820
99.5 %
5,153,800 4,920,600
95.5 %
count. This execution process (and the iteration number) is also used to obtain all
other results shown for these HAL benchmarks.
The significant speed-up obtained for both the hanoi and qsort benchmarks
is explained by the effects of replacing all unifications with a non-trailing version
on the maximum size of the trail stack (in kilobytes), and on the total number of
trailing operations, as shown in Table 4. In the non-deterministic benchmarks, a
much smaller fraction of the trailing operations is removed. This results in a smaller
speed-up or even a slight slow-down. The slow-down shows that the optimization
does not come without a cost.
The larger active code size due to the multi-variant specialization has an impact on the instruction cache behavior. Table 5 shows the impact on instruction
references and instruction cache misses, obtained with the cachegrind skin of the
valgrind memory debugger (see (Nethercote and Seward 2003)). The number of instruction references is the number of times an instruction is retrieved from memory
and the instruction cache miss rate is the percentage of instruction references in
main memory instead of cache.
The table clearly shows that the elimination of all trailing operations results in a
considerable reduction of executed instructions. On the other side of the spectrum,
the multi-variant specialization has a negative effect on the instruction cache miss
rate, which explains the slow-down of the warplan benchmark.
30
Table 5. Benchmark instruction cache misses for classic PARMA: unoptimized
(cparma) vs. optimized with trailing analysis (caparma)
Benchmark
icomp
hanoi
qsort
serialize
warplan
zebra
I1 instruction cache miss rate
cparma caparma
relative
0.85 %
1.79 % 210.6 %
0.00 %
0.00 %
-%
0.00 %
0.00 %
-%
0.00 %
0.70 %
∞%
1.55 %
4.44 % 286.5 %
0.40 %
0.10 %
25.0 %
Instruction references
cparma
caparma
relative
716 ×106
709 ×106
99.0 %
991 ×106
839 ×106
84.7 %
427 ×106
397 ×106
93.0 %
6
912 ×10
899 ×106
98.6 %
1,559 ×106 1,560 ×106 100.1 %
1,300 ×106 1,291 ×106
99.3 %
9.2 Effect of the improved trailing scheme in the Mercury back-end of
HAL
The improved unconditional PARMA trailing scheme has also been implemented in
the Mercury back-end of HAL. Since Mercury already has a tagged trail, this was
not too difficult. Aside from the discussed trailings for unification, this system also
requires trailing when a term is constructed with an old variable as an argument.
In this term construction, the argument cell in the term structure is inserted in
the variable chain. This modifies one cell in the old variable chain. In the classic
scheme this cell is trailed with value trailing. To avoid value trailing altogether this
has been replaced with swap trailing in the improved trailing scheme.
Table 6 presents the timing and maximal trail for both the classic and improved
trailing scheme for the six HAL benchmarks used before.
Table 6. Timing and maximal trail for the classic (cparma) and improved (iparma)
unconditional PARMA trailing scheme for the Mercury back-end of HAL.
Benchmark
icomp
hanoi
qsort
serialize
warplan
zebra
cparma
0.834
0.990
0.363
0.901
1.293
1.239
Time
iparma
0.809
0.944
0.350
0.836
1.284
1.171
relative
97.0 %
95.4 %
96.4 %
92.8 %
99.3 %
94.5 %
Maximal trail
cparma iparma relative
5,545
3,049
55.0 %
61,441
40,961
66.7 %
11,801
7,857
66.6 %
16,569
10,233
61.8 %
17
9
52.9 %
209
105
50.2 %
In all benchmarks the improved trailing scheme is faster than the classic scheme.
The differences are a few percentages though, with a maximum difference of slightly
more than 7% for the serialize benchmark. Much more important are the effects
of the improved trailing scheme on the maximal trail size. The maximal trail is
at least 30% and up to 50% smaller for the improved scheme than for the classic
scheme.
31
9.3 Effect of the improved trailing scheme combined with trailing
analysis Ltrail in the Mercury back-end of HAL
The trailing analysis presented in Section 4 and implemented in HAL, was modified,
as proposed in Section 8, to deal with the improved trailing scheme. Table 7 presents
the timing and maximal trail for the HAL benchmarks obtained under the improved
scheme with the information inferred by the modified analysis, and compares the
results obtained under the same scheme without any analysis information.
Table 7. Timing and maximal trail for the improved unconditional PARMA scheme
without (iparma) and with (iaparma) Ltrail trailing analysis, relative to the classic
scheme without trailing.
Benchmark
icomp
hanoi
qsort
serialize
warplan
zebra
iparma
97.0 %
95.4 %
96.4 %
92.8 %
99.3 %
94.5 %
Time
iaparma
93.3 %
71.6 %
83.5 %
92.8 %
99.7 %
91.9 %
relative
96.2 %
75.1 %
86.6 %
100.0 %
100.4 %
97.3 %
iparma
55.0 %
66.7 %
66.6 %
61.8 %
52.9 %
50.2 %
Maximal trail
iaparma
relative
47.9 %
87.1 %
0.0 %
0.0 %
0.0 %
0.0 %
61.8 % 100.0 %
52.9 % 100.0 %
46.4 %
92.4 %
For the serialize and warplan benchmarks the analysis was not able to reduce
the number of actual trailing operations. For the other four benchmarks the combination of the improved scheme with analysis yields better results, both for time and
maximal trail. For the hanoi and qsort benchmarks there is again a drastic improvement: all trailings have been avoided, with a distinctive time improvement of
25% and 15 % respectively. For the other two benchmarks, icomp and zebra, there
is a maximal trail improvement of about 10% together with a slightly reduced time,
4% and 3% better respectively. Overall, the combination of the improved scheme
with the trailing analysis never makes the results worse. Since it drastically improves some benchmarks and shows a modest improvement of others, it is fair to
conclude that the combination is superior to the improved system without analysis.
9.4 Effect of the improved trailing scheme in dProlog
Let us now present the experimental results of the improved conditional PARMA
trailing scheme in dProlog for several small benchmarks and one bigger program,
comp. Table 8 shows the timing and maximal trail use for each benchmark. Time is
given in seconds and applies to the number of runs (iterations) given. The maximal
trail size is given in kilobytes and applies to a single run.
The time difference between the classic and the improved scheme is negligible.
The improved scheme is at most 8.8% slower, for the zebra benchmark, but on
32
Table 8. PARMA in dProlog: classic (cparma) vs. improved trailing (iparma)
Benchmark
Iterations
boyer
10
browse
10
cal
100
chat
50
crypt
2,000
ham
20
meta qsort
1,250
nrev
50,000
100
poly 10
queens 16
20
queens
100
reducer
200
sdda
12,000
send
100
tak
100
zebra
300
relative average
comp
1
comp relative
Time
cparma iparma
.950
.920
1.010
1.010
1.800
1.800
1.020
1.040
1.160
1.170
1.160
1.130
1.070
1.090
.900
.860
.630
.650
1.810
1.790
3.310
3.300
.440
.430
1.000
1.010
.800
.800
1.620
1.520
2.510
2.730
100%
99.9%
1.930
1.890
100%
97.9%
Maximal trail
cparma iparma
450.6
225.3
5.2
4.5
0.4
0.2
3.6
1.9
0.5
0.2
0.8
0.4
12.6
7.4
0.4
0.2
52.6
26.3
0.7
0.3
0.7
0.3
18.9
10.0
1.3
0.8
0.5
0.2
373.1
186.6
1.6
0.8
100%
51.7%
2516.3
1319.8
100%
52.4%
average both are about equally fast. The price for the lower trail usage is an increase
in instructions executed and that is why there is no net speedup.
The differences in maximal trail use however are substantial. While swap trail
and chain trail halve the trail stack consumption, value trailing is still used for
some cases of variable–variable trailing. Yet experimental results show that that
kind of variable–variable trailing does not occur very often in most benchmarks, as
the maximal trail stack is effectively halved in eleven benchmarks and on average
the maximal trail use is 51.7% of the classical scheme.
The results for the smaller benchmarks are confirmed by the larger comp program.
Execution time is nearly the same for the classic and improved trailing scheme and
the maximal trail shows a similar improvement of almost 50%.
10 Related and future work
As far as we know, the modifications suggested to the classic PARMA trailing
scheme are new.
A somewhat similar analysis for detecting variables that do not have to be trailed
is presented by Debray in (Debray 1992) together with corresponding optimizations.
Debray’s analysis however is for the WAM variable representation and in a traditional Prolog setting, i.e., without type, mode and determinism declarations. Also
33
in (Van Roy and Despain 1992) trailing is avoided, but only for variables that are
new in our terminology and, again, the setting is basically the WAM representation.
Taylor too keeps track of a trailing state of variables in the global analysis of
his PARMA system with the classic PARMA trailing scheme (see (Taylor 1991;
Taylor 1989)). As opposed to the Lnotrail analysis we have presented here, Taylor’s
analysis is less precise and closer to the Ltrail analysis presented here: the trailing
state of a variable can only be that it has to be trailed or not, i.e. there is no
intermediary shallow trailing state.
There exist also two run-time technique for preventing the multiple value trailing
between two choice points. The first, described in (Noyé 1994), only works in the
WAM scheme, because it introduces linear reference chains that PARMA does
not allow. The second, described in (Aggoun and Beldiceanu 1990), maintains a
timestamp for every cell that corresponds to the choicepoint before the last update.
However, such a timestamp requires additional space, even in the case that the cell
is never updated. In the context of PARMA, timestamps would likely consume more
space than is actually saved by avoiding trailing.
Finally, there are other approaches to the reconstruction of state on backtracking
other than trailing, using either copying (Schulte 1999) or recomputation (Van Hentenryck and Ramachandran 1995
While PARMA (and for that matter WAM) bindings do not keep enough information to allow recomputation on backtracking, a copying approach to backtracking in
PARMA is quite feasible. This remains as an interesting question for future work.
There is little room left for optimization of the trailing analysis for the improved
unconditional trailing scheme. Of course, the analysis itself can be improved by
adopting a more refined representation for bound variables. Currently, all PARMA
chains in the structure of a bound variable are represented by the same trailing
state. Bound variables could be represented more accurately, by requiring the domain to keep track of the different chains contained in the structures to which
the program variables are bound, their individual trailing state and how these are
affected by the different program constructs. Known techniques (see for instance
(Janssens and Bruynooghe 1993; Van Hentenryck et al. 1995; Mulkers et al. 1994;
Lagoon and Stuckey 2001; Lagoon et al. 2003)) based on type information could
be used to keep track of the constructor that a variable is bound to and the trailing state of the different arguments, thereby making this approach possible. This
applies equally to the analysis of the classical scheme.
Additionally, it would be interesting to see how much extra gain analysis can add
to the improved conditional trailing scheme as implemented in dProlog or in the
Mercury back-end of HAL that supports conditional trailing. Such analysis would
certainly not improve the maximal trail, but it would remove the overhead of the
run-time test. This will most likely also result in a small speed-up.
Though experimental results show that the improved scheme with analysis is
better than the classic scheme with analysis, this need not be true for all programs.
Recall that between two choice points all value trailings of a cell but the first can be
eliminated in the classic scheme, while no swap trailings could be eliminated in the
improved scheme. A hybrid scheme would be possible using analysis to decide on a
single unification basis if either swap trailing or value trailing is better at minimizing
34
the amount of trailing and the cost of untrailing. This analysis would require a more
global view of all the trailings in between two choice points. Moreover, some trailings
could be common to different pairs of choice points and optimality would depend
on where execution spends most of its time.
Also the untrailing operation can be improved: when analysis is able to determine
for instance that the only trailing that happened was a swap trailing, no tags need
to be set and tested.
Acknowledgements
We would like to thank the referees for their detailed and insightful reports which
have significantly improved the paper.
References
Aggoun, A. and Beldiceanu, N. 1990. Time Stamps Techniques for the Trailed Data in
Constraint Logic Programming Systems. In SPLT’90: 8ème Séminaire Programmation
en Logique, S. Bourgault and M. Dincbas, Eds. CNET, Trégastel, France, 487–510.
Aı̈t-Kaci, H. 1991. Warren’s Abstract Machine: A Tutorial Reconstruction. MIT Press.
Bruynooghe, M. 1991. A Practical Framework for the Abstract Interpretation of Logic
Programs. Journal of Logic Programming 10, 1/2/3&4, 91–124.
Bueno, F., de la Banda, M. J. G., Hermenegildo, M. V., Marriott, K., Puebla,
G., and Stuckey, P. J. 2001. A Model for Inter-module Analysis and Optimizing
Compilation. In LOPSTR ’00: Selected Papers form the 10th International Workshop
on Logic Based Program Synthesis and Transformation, K.-K. Lau, Ed. Lecture Notes
in Computer Science, vol. 2042. Springer Verlag, London, UK, 86–102.
de la Banda, M. J. G., Demoen, B., Marriott, K., and Stuckey, P. J. 2002. To the
Gates of HAL: A HAL Tutorial. In FLOPS 2002: Proceedings of the 6th International
Symposium on Functional and Logic Programming, Z. Hu and M. Rodrı́guez-Artalejo,
Eds. Lecture Notes in Computer Science, vol. 2441. Springer Verlag, Aizu, Japan, 47–66.
de la Banda, M. J. G., Marriott, K., Stuckey, P. J., and Søndergaard, H. 1998.
Differential Methods in Logic Program Analysis. Journal of Logic Programming 35, 1,
1–37.
Debray, S. 1992. A Simple Code Improvement Scheme for Prolog. Journal of Logic
Programming 13, 1 (May), 57–88.
Demoen, B., de la Banda, M. J. G., Harvey, W., Marriott, K., and Stuckey, P. J.
1999. An Overview of HAL. In CP’99: Proceedings of the 5th International Conference
on Principles and Practice of Constraint Programming, J. Jaffar, Ed. Lecture Notes in
Computer Science, vol. 1713. Springer Verlag, Alexandria, Virginia, USA, 174–188.
Demoen, B. and Nguyen, P.-L. 2000. So Many WAM Variations, so Little Time.
In CL2000: Proceedings of the 1st International Conference on Computational Logic,
J. Lloyd, V. Dahl, U. Furbach, M. Kerber, K.-K. Lau, C. Palamidessi, L. Moniz Pereira,
Y. Sagiv, and P. J. Stuckey, Eds. Lecture Notes in Artificial Intelligence, vol. 1861. ALP,
Springer Verlag, London, UK, 1240–1254.
Ferreira, M. and Damas, L. 2003. Controlling Code Expansion in a Multiple Specialization Prolog Compiler. In Proceedings of CICLOPS 2003: Colloquium on Implementation
of Constraint and LOgic Programming Systems, R. Lopes and M. Ferreira, Eds. University of Porto, Mumbai, India, 75–87. Technical Report DCC-2003-05, DCC - FC &
LIACC, Univeristy of Porto, December 2003.
35
Janssens, G. and Bruynooghe, M. 1993. Deriving Descriptions of Possible Value of
Program Variables by means of Abstract Interpretation. Journal of Logic Programming 13, 205–258.
Lagoon, V., Mesnard, F., and Stuckey, P. J. 2003. Termination Analysis with Types
Is More Accurate. In ICLP 2003: Proceedings of the 19th International Conference on
Logic Programming, C. Palamidessi, Ed. Lecture Notes in Computer Science, vol. 2916.
Springer Verlag, Mumbai, India, 254–268.
Lagoon, V. and Stuckey, P. 2001. A Framework for Analysis of Typed Logic Programs.
In FLOPS 2001: Proceedings of the 5th International Symposium on Functional and
Logic Programming, H. Kuchen and K. Ueda, Eds. Lecture Notes in Computer Science,
vol. 2024. Springer Verlag, Tokyo, Japan, 296–310.
Lindgren, T., Mildner, P., and Bevemyr, J. 1995. On Taylor’s Scheme for Unbound
Variables. Tech. rep., Computer Science Department, Uppsala University. Oct.
Mazur, N. 2001. Compile-time Garbage Collection for the Declarative Language Mercury.
Ph.D. thesis, Department of Computer Science, K.U.Leuven, Leuven, Belgium.
Mulkers, A., Winsborough, W. H., and Bruynooghe, M. 1994. Live-Structure
Dataflow Analysis for Prolog. ACM Transactions on Programming Languages and Systems 16, 2, 205–258.
Nethercote, N. 2001. The Analysis Framework of HAL. M.S. thesis, University of
Melbourne.
Nethercote, N. and Seward, J. 2003. Valgrind: A Program Supervision Framework.
Electronic Notes in Theoretical Computer Science 89, 2, 1–23.
Noyé, J. 1994. Elagage de contexte, retour arrière superficiel, modifications réversibles
et autres: une étude approfondie de la WAM. Ph.D. thesis, Université de Rennes I.
Schulte, C. 1999. Comparing Trailing and Copying for Constraint Programming. In
Proceedings of the Sixteenth International Conference on Logic Programming, D. De
Schreye, Ed. MIT Press, Las Cruces, NM, USA, 275–289.
Somogyi, Z., Henderson, F., and Conway, T. 1996. The Execution Algorithm of
Mercury, an Efficient Purely Declarative Logic Programming Language. Journal of
Logic Programming 29, 1–3, 17–64.
Søndergaard, H. 1986. An Application of Abstract Interpretation of Logic Programs:
Occur Check Reduction. In ESOP 86: Proceedings of the European Symposium on
Programming, B. Robinet and R. Wilhelm, Eds. Lecture Notes in Computer Science,
vol. 213. Springer Verlag, Saarbrücken, Germany, 327–338.
Taylor, A. 1989. Removal of Dereferencing and Trailing in Prolog Compilation. In
Proceedings of the 6th Internation Conference on Logic Programming, G. Levi and
M. Martelli, Eds. MIT Press, Lisbon, Portugal, 48–60.
Taylor, A. 1991. High Performace Prolog Implementation. Ph.D. thesis, Basser Department of Computer Science.
Taylor, A. 1996. Parma - Bridging the Performance GAP Between Imperative and Logic
Programming. Journal of Logic Programming 29, 1-3, 5–16.
Van Hentenryck, P., Cortesi, A., and Charlier, B. L. 1995. Type analysis of Prolog
using Type Graphs. Journal of Logic Programming 22, 3, 179–209.
Van Hentenryck, P. and Ramachandran, V. 1995. Backtracking without Trailing in
CLP(Rlin). ACM Transactions on Programming Languages and Systems 17, 4 (July),
635–671.
Van Roy, P. and Despain, A. 1992. High-Performance Logic Programming with the
Aquarius Prolog Compiler. IEEE Computer 25, 1, 54–68.
36
| 6cs.PL
|
An Optimal Algorithm for ℓ1-Heavy Hitters in Insertion Streams
and Related Problems
arXiv:1603.00213v1 [cs.DS] 1 Mar 2016
Arnab Bhattacharyya⋆ , Palash Dey⋆ , and David P. Woodruff†
⋆
Indian Institute of Science, Bangalore
†
⋆
IBM Research, Almaden
{arnabb,palash}@csa.iisc.ernet.in,† [email protected]
Abstract
We give the first optimal bounds for returning the ℓ1 -heavy hitters in a data stream of insertions,
together with their approximate frequencies, closing a long line of work on this problem. For a stream
of m items in {1, 2, . . . , n} and parameters 0 < ε < ϕ 6 1, let fi denote the frequency of item i,
i.e., the number of times item i occurs in the stream. With arbitrarily large constant probability, our
algorithm returns all items i for which fi > ϕm, returns no items j for which fj 6 (ϕ − ε)m, and
returns approximations f˜i with |f˜i − fi | 6 εm for each item i that it returns. Our algorithm uses
O(ε−1 log ϕ−1 + ϕ−1 log n + log log m) bits of space, processes each stream update in O(1) worst-case
time, and can report its output in time linear in the output size. We also prove a lower bound, which
implies that our algorithm is optimal up to a constant factor in its space complexity. A modification of
our algorithm can be used to estimate the maximum frequency up to an additive εm error in the above
amount of space, resolving Question 3 in the IITK 2006 Workshop on Algorithms for Data Streams
for the case of ℓ1 -heavy hitters. We also introduce several variants of the heavy hitters and maximum
frequency problems, inspired by rank aggregation and voting schemes, and show how our techniques can
be applied in such settings. Unlike the traditional heavy hitters problem, some of these variants look at
comparisons between items rather than numerical values to determine the frequency of an item.
1
Introduction
The data stream model has emerged as a standard model for processing massive data sets. Because of the
sheer size of the data, traditional algorithms are no longer feasible, e.g., it may be hard or impossible to
store the entire input, and algorithms need to run in linear or even sublinear time. Such algorithms typically
need to be both randomized and approximate. Moreover, the data may not physically reside on any device,
e.g., if it is internet traffic, and so if the data is not stored by the algorithm, it may be impossible to recover
it. Hence, many algorithms must work given only a single pass over the data. Applications of data streams
include data warehousing [HSST05, BR99, FSG+ 98, HPDW01], network measurements [ABW03, CKMS08,
DLOM02, EV03], sensor networks [BGS01, SBAS04], and compressed sensing [GSTV07, CRT06]. We refer
the reader to recent surveys on the data stream model [Mut05, Nel12, Cor12].
One of the oldest and most fundamental problems in the area of data streams is the problem of finding the
ℓ1 -heavy hitters (or simply, “heavy hitters”), also known as the top-k, most popular items, frequent items,
elephants, or iceberg queries. Such algorithms can be used as subroutines in network flow identification at
IP routers [EV03], association rules and frequent itemsets [AS94, SON95, Toi96, Hid99, HPY00], iceberg
queries and iceberg datacubes [FSG+ 98, BR99, HPDW01]. The survey [CH08] presents an overview of the
state-of-the-art for this problem, from both theoretical and practical standpoints.
We now formally define the heavy hitters problem that we focus on in this paper:
1
Definition 1. ((ε, ϕ)-Heavy Hitters Problem) In the (ε, ϕ)-Heavy Hitters Problem, we are given parameters 0 < ε < ϕ 6 1 and a stream a1 , . . . , am of items aj ∈ {1, 2, . . . , n}. Let fi denote the number of
occurrences of item i, i.e., its frequency. The algorithm should make one pass over the stream and at the end
of the stream output a set S ⊆ {1, 2, . . . , n} for which if fi > ϕm, then i ∈ S, while if fi 6 (ϕ − ε)m, then
i∈
/ S. Further, for each item i ∈ S, the algorithm should output an estimate f˜i of the frequency fi which
satisfies |fi − f˜i | 6 εm.
Note that other natural definitions of heavy hitters
Pn are possible and sometimes used. For example, ℓ2 heavy hitters are those items i for which fi2 > ϕ2 j=1 fj2 , and more generally, ℓp -heavy hitters are those
Pn
items i for which fip > ϕp j=1 fjp . It is in this sense that Definition 1 corresponds to ℓ1 -heavy hitters.
While ℓp -heavy hitters for p > 1 relax ℓ1 -heavy hitters and algorithms for them have many interesting
applications, we focus on the most direct and common formulation of the heavy hitters notion.
We are interested in algorithms which use as little space in bits as possible to solve the (ε, ϕ)-Heavy
Hitters Problem. Further, we are also interested in minimizing the update time and reporting time of such
algorithms. Here, the update time is defined to be the time the algorithm needs to update its data structure
when processing a stream insertion. The reporting time is the time the algorithm needs to report the answer
after having processed the stream. We allow the algorithm to be randomized and to succeed with probability
at least 1 − δ for 0 < δ < 1. We do not make any assumption on the ordering of the stream a1 , . . . , am .
This is desirable as often in applications one cannot assume a best-case or even a random order. We are also
interested in the case when the length m of the stream is not known in advance, and give algorithms in this
more general setting.
The first algorithm for the (ε, ϕ)-Heavy Hitters Problem was given by Misra and Gries [MG82], who
achieved O(ε−1 (log n + log m)) bits of space for any ϕ > ε. This algorithm was rediscovered by Demaine et
al. [DLOM02], and again by Karp et al. [KSP03]. Other than these algorithms, which are deterministic,
there are also a number of randomized algorithms, such as the CountSketch [CCFC04], Count-Min sketch
[CM05], sticky sampling [MM02], lossy counting [MM02], space-saving [MAE05], sample and hold [EV03],
multi-stage bloom filters [CFM09], and sketch-guided sampling [KX06]. Berinde et al. [BICS10] show that
using O(kε−1 log(mn)) bits of space, one can achieve the stronger guarantee of reporting, for each item i ∈ S,
res(k)
res(k)
f˜i with |f˜i − fi | 6 ε/kF1
, where F1
< m denotes the sum of frequencies of items in {1, 2, . . . , n}
excluding the frequencies of the k most frequent items.
We emphasize that prior to our work the best known algorithms for the (ε, ϕ)-Heavy Hitters Problem
used O(ε−1(log n + log m)) bits of space. Two previous lower bounds were known. The first is a lower bound
n
of log( 1/ϕ
) = Ω(ϕ−1 log(ϕn)) bits, which comes from the fact that the output set S can contain ϕ−1 items
and it takes this many bits to encode them. The second lower bound is Ω(ε−1 ) which follows from a folklore
reduction from the randomized communication complexity of the Index problem. In this problem, there are
two players, Alice and Bob. Alice has a bit string x of length (2ε)−1 , while Bob has an index i. Alice creates
a stream of length (2ε)−1 consisting of one copy of each j for which xj = 1 and copies of a dummy item
to fill the rest of the stream. She runs the heavy hitters streaming algorithm on her stream and sends the
state of the algorithm to Bob. Bob appends (2ε)−1 copies of the item i to the stream and continues the
execution of the algorithm. For ϕ = 1/2, it holds that i ∈ S. Moreover, fi differs by an additive εm factor
depending on whether xi = 1 or xi = 0. Therefore by the randomized communication complexity of the
Index problem [KNR99], the (ε, 1/2)-heavy hitters problem requires Ω(ε−1 ) bits of space. Although this proof
was for ϕ = 1/2, no better lower bound is known for any ϕ > ε.
Thus, while the upper bound for the (ε, ϕ)-Heavy Hitters Problem is O(ε−1 (log n + log m)) bits, the
best known lower bound is only Ω(ϕ−1 log n + ε−1 ) bits. For constant ϕ, and log n ≈ ε−1 , this represents a
nearly quadratic gap in upper and lower bounds. Given the limited resources of devices which typically run
heavy hitters algorithms, such as internet routers, this quadratic gap can be critical in applications.
A problem related to the (ε, ϕ)-Heavy Hitters Problem is estimating the maximum frequency in a
data stream, also known as the ℓ∞ -norm. In the IITK 2006 Workshop on Algorithms for Data Streams,
Open Question 3 asks for an algorithm to estimate the maximum frequency of any item up to an additive εm
error using as little space as possible. The best known space bound is still O(ε−1 log n) bits, as stated in the
original formulation of the question (note that the “m” in the question there corresponds to the “n” here).
2
Space complexity
Problem
(ε, ϕ)-Heavy Hitters
ε-Maximum and ℓ∞ -approximation
ε-Minimum
ε-Borda
ε-Maximin
Upper bound
O ε
−1
−1
−1
log ϕ + ϕ log n + log log m
[Theorem 1 and 7]
O ε−1 log ε−1 + log n + log log m
[Theorem 1 and 7]
−1
O ε log log ε−1 + log log m
[Theorem 4 and 8]
O n(log ε−1 + log n) + log log m
[Theorem 5 and 8]
O nε−2 log2 n + log log m
[Theorem 6 and 8]
Lower bound
log ϕ + ϕ−1 log n + log log m
[Theorem 9 and 14]
Ω ε−1 log ε−1 + log n + log log m
[Theorem 9 and 14]
Ω ε−1 + log log m
[Theorem 11 and 14]
Ω n(log ε−1 + log n) + log log m
[Theorem 12 and 14]
Ω n(ε−2 + log n) + log log m
[Theorem 13]
Ω ε
−1
−1
Table 1: The bounds hold for constant success probability algorithms and for n sufficiently large in terms of ε. For
the (ε, ϕ)-Heavy Hitters problem and the ε-Maximum problem, we also achieve O(1) update time and reporting
time which is linear in the size of the output. The upper bound for ε-Borda (resp. ε-Maximin) is for returning
every item’s Borda score (resp. Maximin score) up to an additive εmn (resp. additive εm), while the lower bound
for ε-Borda (resp. ε-Maximin) is for returning only the approximate Borda score (resp. Maximin score) of an
approximate maximum.
Note that, if one can find an item whose frequency is the largest, up to an additive εm error, then one can
solve this problem. The latter problem is independently interesting and corresponds to finding approximate
plurality election winners in voting streams [DB15]. We refer to this problem as the ε-Maximum problem.
Finally, we note that there are many other variants of the (ε, ϕ)-Heavy Hitters Problem that one
can consider. One simple variant of the above is to output an item of frequency within εm of the minimum
frequency of any item in the universe. We refer to this as the ε-Minimum problem. This only makes sense
for small universes, as otherwise outputting a random item typically works. This is useful when one wants
to count the “number of dislikes”, or in anomaly detection; see more motivation below. In other settings,
one may not have numerical scores associated with the items, but rather, each stream update consists of
a “ranking” or “total ordering” of all stream items. This may be the case in ranking aggregation on the
web (see, e.g., [MBG04, MYCC07]) or in voting streams (see, e.g., [CS05, CP11, DB15, Xia12]). One may
consider a variety of aggregation measures, such as the Borda score of an item i, which asks for the sum, over
rankings, of the number of items j 6= i for which i is ranked ahead of j in the ranking. Alternatively, one may
consider the Maximin score of an item i, which asks for the minimum, over items j 6= i, of the number of
rankings for which i is ranked ahead of j. For these aggregation measures, one may be interested in finding
an item whose score is an approximate maximum. This is the analogue of the ε-Maximum problem above.
Or, one may be interested in listing all items whose score is above a threshold, which is the analogue of the
(ε, ϕ)-Heavy Hitters Problem.
We give more motivation of these variants of heavy hitters in this section below, and more precise
definitions in Section 2.
1.1
Our Contributions
Our results are summarized in Table 1. We note that independently of this work and nearly parallelly,
there have been improvements to the space complexity of the ℓ2 -heavy hitters problem in insertion streams
[BCIW16] and to the time complexity of the ℓ1 -heavy hitters problem in turnstile1 streams [LNNT16]. These
works use very different techniques.
1 In a turnstile stream, updates modify an underlying n-dimensional vector x initialized at the zero vector; each update is of
the form x ← x + ei or x ← x − ei where ei is the i’th standard unit vector. In an insertion stream, only updates of the form
x ← x + ei are allowed.
3
Our first contribution is an optimal algorithm and lower bound for the (ε, ϕ)-Heavy Hitters Problem.
Namely, we show that there is a randomized algorithm with constant probability of success which solves this
problem using
O(ε−1 log ϕ−1 + ϕ−1 log n + log log m)
bits of space, and we prove a lower bound matching up to constant factors. In the unit-cost RAM model with
O(log n) bit words, our algorithm has O(1) update time and reporting time linear in the output size, under
the standard assumptions that the length of the stream and universe size are at least poly(ε−1 log(1/ϕ)).
Furthermore, we can achieve nearly the optimal space complexity even when the length m of the stream is
not known in advance. Although the results of [BICS10] achieve stronger error bounds in terms of the tail,
which are useful for skewed streams, here we focus on the original formulation of the problem.
Next, we turn to the problem of estimating the maximum frequency in a data stream up to an additive
εm. We give an algorithm using
O(ε−1 log ε−1 + log n + log log m)
bits of space, improving the previous best algorithms which required space at least Ω(ε−1 log n) bits, and show
that our bound is tight. As an example setting of parameters, if ε−1 = Θ(log n) and log log m = O(log n),
our space complexity is O(log n log log n) bits, improving the previous Ω(log2 n) bits of space algorithm.
We also prove a lower bound showing our algorithm is optimal up to constant factors. This resolves Open
Question 3 from the IITK 2006 Workshop on Algorithms for Data Streams in the case of insertion streams,
for the case of “ℓ1 -heavy hitters”. Our algorithm also returns the identity of the item with the approximate
maximum frequency, solving the ε-Maximum problem.
We then focus on a number of variants of these problems. We first give nearly tight bounds for finding an
item whose frequency is within εm of the minimum possible frequency. While this can be solved using our
new algorithm for the (ε, ε)-Heavy Hitters Problem, this would incur Ω(ε−1 log ε−1 + log log m) bits of
space, whereas we give an algorithm using only O(ε−1 log log(ε−1 ) + log log m) bits of space. We also show a
nearly matching Ω(ε−1 + log log m) bits of space lower bound. We note that for this problem, a dependence
on n is not necessary since if the number of possible items is sufficiently large, then outputting the identity
of a random item among the first say, 10ε−1 items, is a correct solution with large constant probability.
Finally, we study variants of heavy hitter problems that are ranking-based. In this setting, each stream
update consists of a total ordering of the n universe items. For the ε-Borda problem, we give an algorithm
using O(n(log ε−1 + log log n) + log log m) bits of space to report the Borda score of every item up to an
additive εmn. We also show this is nearly optimal by proving an Ω(n log ε−1 + log log m) bit lower bound for
the problem, even in the case when one is only interested in outputting an item maximum Borda score up to
an additive εmn. For the ε-Maximin problem, we give an algorithm using O(nε−2 log2 n + log log m) bits
of space to report the maximin score of every item up to an additive εm, and prove an Ω(nε−2 + log log m)
bits of space lower bound even in the case when one is only interested in outputting the maximum maximin
score up to an additive εm. This shows that finding heavy hitters with respect to the maximin score is
significantly more expensive than with respect to the Borda score.
1.2
Motivations for Variants of Heavy Hitters
While the (ε, ϕ)-Heavy Hitters and ε-Maximum problem are very well-studied in the data stream literature, the other variants introduced are not. We provide additional motivation for them here.
For the ε-Minimum problem, in our formulation, an item with frequency zero, i.e., one that does not
occur in the stream, is a valid solution to the problem. In certain scenarios, this might not make sense, e.g.,
if a stream containing only a small fraction of IP addresses. However, in other scenarios we argue this is
a natural problem. For instance, consider an online portal where users register complaints about products.
Here, minimum frequency items correspond to the “best” items. That is, such frequencies arise in the context
of voting or more generally making a choice: in cases for which one does not have a strong preference for
an item, but definitely does not like certain items, this problem applies, since the frequencies correspond to
“number of dislikes”.
4
The ε-Minimum problem may also be useful for anomaly detection. Suppose one has a known set of
sensors broadcasting information and one observes the “From:” field in the broadcasted packets. Sensors
which send a small number of packets may be down or defective, and an algorithm for the ε-Minimum
problem could find such sensors.
Finding items with maximum and minimum frequencies in a stream correspond to finding winners under
plurality and veto voting rules respectively in the context of voting2 [BCE+ 15]. The streaming aspect of
voting could be crucial in applications like online polling [KTW11], recommender systems [RV97, HKTR04,
AT05] where the voters are providing their votes in a streaming fashion and at every point in time, we would
like to know the popular items. While in some elections, such as for political positions, the scale of the
election may not be large enough to require a streaming algorithm, one key aspect of these latter votingbased problems is that they are rank-based which is useful when numerical scores are not available. Orderings
naturally arise in several applications - for instance, if a website has multiple parts, the order in which a user
visits the parts given by its clickstream defines a voting, and for data mining and recommendation purposes
the website owner may be interested in aggregating the orderings across users. Motivated by this connection,
we define similar problems for two other important voting rules, namely Borda and maximin. The Borda
scoring method finds its applications in a wide range of areas of artificial intelligence, for example, machine
learning [HHS94, CP06, VZ14, PPR15], image processing [LN06, MG09], information retrieval [Li14, AM01,
NC06], etc. The Maximin score is often used when the spread between the best and worst outcome is very
large (see, e.g., p. 373 of [MR91]). The maximin scoring method also has been used frequently in machine
learning [WL04, JMP+ 14], human computation [MPC12, MPC13], etc.
2
Preliminaries
We denote the disjoint union of sets by ⊔. We denote the set of all permutations of a set U by L(U). For
a positive integer ℓ, we denote the set {1, . . . , ℓ} by [ℓ]. In most places, we ignore floors and ceilings for the
sake of notational simplicity.
2.1
Model of Input Data
The input data is an insertion-only stream of elements from some universe U. In the context of voting, the
input data is an insertion-only stream over the universe of all possible rankings (permutations).
2.2
Communication Complexity
We will use lower bounds on communication complexity of certain functions to prove space complexity lower
bounds for our problems. Communication complexity of a function measures the number of bits that need to
be exchanged between two players to compute a function whose input is split among those two players [Yao79].
In a more restrictive one-way communication model, Alice, the first player, sends only one message to Bob,
the second player, and Bob outputs the result. A protocol is a method that the players follow to compute
certain functions of their input. Also the protocols can be randomized; in that case, the protocol needs to
output correctly with probability at least 1 − δ, for δ ∈ (0, 1) (the probability is taken over the random
coin tosses of the protocol). The randomized one-way communication complexity of a function f with error
probability δ is denoted by R1-way
(f ). [KN97] is a standard reference for communication complexity.
δ
2.3
Model of Computation
Our model of computation is the unit-cost RAM model on words of size O(log n), capable of generating
uniformly random words and of performing arithmetic operations in {+, −, log2 } in one unit of time. We
note that this model of computation has been used before [DHKP97]. We store an integer C using a variable
length array of [BB08] which allows us to read and update C in O(1) time and O(log C) bits of space.
2 In
fact, the first work [Moo81] to formally pose the heavy hitters problem couched it in the context of voting.
5
2.4
Universal Family of Hash Functions
Definition 2. (Universal family of hash functions)
A family of functions H = {h|h : A → B} is called a universal family of hash functions if for all a 6= b ∈ A,
Pr{h(a) = h(b)} = 1/|B|, where h is picked uniformly at random from H.
We know that there exists a universal family of hash functions H from [k] to [ℓ] for every positive integer
ℓ and every prime k [LRSC01]. Moreover, |H|, the size of H, is O(k 2 ).
2.5
Problem Definitions
We now formally define the problems we study here. Suppose we have 0 < ε < ϕ < 1.
Definition 3. (ε, ϕ)-List heavy hitters
Given an insertion-only stream of length m over a universe U of size n, find all items in U with frequency
more than ϕm, along with their frequencies up to an additive error of εm, and report no items with frequency
less than (ϕ − ε)m.
Definition 4. ε-Maximum
Given an insertion-only stream of length m over a universe U of size n, find the maximum frequency up to
an additive error of εm.
Next we define the minimum problem for 0 < ε < 1.
Definition 5. ε-Minimum
Given an insertion-only stream of length m over a universe U of size n, find the minimum frequency up to
an additive error of εm.
Next we define related heavy hitters problems in the context of rank aggregation. The input is a stream
of rankings (permutations) over an item set U for the problems below. The Borda score of an item i is the
sum, over all rankings, of the number of items j 6= i for which i is ranked ahead of j in the ranking.
Definition 6. (ε, ϕ)-List borda
Given an insertion-only stream over a universe L(U) where |U| = n, find all items with Borda score more
than ϕmn, along with their Borda score up to an additive error of εmn, and report no items with Borda
score less than (ϕ − ε)mn.
Definition 7. ε-Borda
Given an insertion-only stream over a universe L(U) where |U| = n, find the maximum Borda score up to
an additive error of εmn.
The maximin score of an item i is the minimum, over all items j 6= i, of the number of rankings for which
i is ranked ahead of j.
Definition 8. (ε, ϕ)-List maximin
Given an insertion-only stream over a universe L(U) where |U| = n, find all items with maximin score more
than ϕm along with their maximin score up to an additive error of εm, and report no items with maximin
score less than (ϕ − ε)m.
Definition 9. ε-maximin
Given an insertion-only stream over a universe L(U) where |U| = n, find the maximum maximin score up
to an additive error of εm.
Notice that the maximum possible Borda score of an item is m(n − 1) = Θ(mn) and the maximum
possible maximin score of an item is m. This justifies the approximation factors in Definition 6 to 9.
We note that finding an item with maximum Borda score within additive εmn or maximum maximin score
within additive εm corresponds to finding an approximate winner of an election (more precisely, what is
known as an ε-winner) [DB15].
6
3
Algorithms
In this section, we present our upper bound results. All omitted proofs are in Appendix B. Before describing
specific algorithms, we record some claims for later use. Lemma 1 follows by checking whether we get all
heads in log m tosses of a fair coin.
Lemma 1. Suppose m is a power of two3 . Then there is an algorithm A for choosing an item with probability
1/m that has space complexity of O(log log m) bits and time complexity of O(1) in the unit-cost RAM model.
Proof. We generate a (log2 m)-bit integer C uniformly at random and record the sum of the digits in C.
Choose the item only if the sum of the digits is 0, i.e. if C = 0.
We remark that the algorithm in Lemma 1 has optimal space complexity as shown in Proposition 2 in
Appendix B.
Our second claim is a standard result for universal families of hash functions.
2
Lemma 2. For S ⊆ A, δ ∈ (0, 1), and universal family of hash functions H = {h|h : A → [⌈|S| /δ⌉]}:
Pr [∃i 6= j ∈ S, h(i) = h(j)] 6 δ
h∈U H
Proof. For every i 6= j ∈ S, since H is a universal family of hash functions, we have Prh∈U H [h(i) = h(j)] 6
|S|2
1
⌈|S|2 /δ⌉ . Now we apply the union bound to get Prh∈U H [∃i 6= j ∈ S, h(i) = h(j)] 6 ⌈|S|2 /δ⌉ 6 δ
Our third claim is folklore and also follows from the celebrated DKW inequality [DKW56]. We provide
a simple proof here that works for constant δ.
Lemma 3. Let fi and fˆi be the frequencies of an item i in a stream S and in a random sample T of
size r from S, respectively. Then for r > 2ε−2 log(2δ −1 ), with probability 1 − δ, for every universe item i
simultaneously,
fˆi
fi
−
6 ε.
r
m
Proof for constant δ. This follows by Chebyshev’s inequality and a union bound. Indeed, consider a given i ∈
[n] with frequency fi and suppose we sample each of its occurrences pairwise-independently with probability
r/m, for a parameter r. Then the expected number E[fˆi ] of sampled occurrences is fi · r/m and the variance
Var[fˆi ] is fi · r/m(1 − r/m) 6 fi r/m. Applying Chebyshev’s inequality,
h
4fi r
rε i Var[fˆi ]
6
.
6
Pr fˆi − E[fˆi ] >
2
(rε/2)2
mr2 ε2
4fi
. By the union bound, if we sample
Setting r = εC2 for a constant C > 0 makes this probability at most Cm
r
each element in the stream independently with probability m , then the probability there exists an i for which
Pn 4fi
4
1
|fˆi − E[fˆi ]| > rε
i=1 Cm 6 C , which for C > 400 is at most 100 , as desired.
2 is at most
For now, assume that the length of the stream is known in advance; we show in Section 3.5 how to remove
this assumption.
3 In all our algorithms, whenever we pick an item with probability p > 0, we can assume, without loss of generality, that 1/p
is a power of two. If not, then we replace p with p′ where 1/p′ is the largest power of two less than 1/p. This does not affect
correctness and performance of our algorithms.
7
3.1
List Heavy Hitters
For the (ε, ϕ)-List heavy hitters problem, we present two algorithms. The first is slightly suboptimal, but
simple conceptually and already constitutes a very large improvement in the space complexity over known
algorithms. We expect that this algorithm could be useful in practice as well. The second algorithm is more
complicated, building on ideas from the first algorithm, and achieves the optimal space complexity upto
constant factors.
We note that both algorithms proceed by sampling O(ε−2 ln(1/δ)) stream items and updating a data
structure as the stream progresses. In both cases, the time to update the data structure is bounded by
O(1/ε), and so, under the standard assumption that the length of the stream is at least poly(ln(1/δ)ε),
the time to perform this update can be spread out across the next O(1/ε) stream updates, since with large
probability there will be no items sampled among these next O(1/ε) stream updates. Therefore, we achieve
worst-case4 update time of O(1).
3.1.1
A simpler, near-optimal algorithm
Theorem 1. Assume the stream length is known beforehand. Then there is a randomized one-pass algorithm A for the (ε, ϕ)-List heavy hitters problem
which succeeds with probability at least 1 − δ using
O ε−1 (log ε−1 + log log δ −1 ) + ϕ−1 log n + log log m bits of space. Moreover, A has an update time of O(1)
and reporting time linear in its output size.
Overview The overall idea is as follows. We sample ℓ = O(ε−2 ) many items from the stream uniformly at
random as well as hash the id’s (the word “id” is short for identifier) of the sampled elements into a space
of size O(ε−4 ). Now, both the stream length as well as the universe size are poly(ε−1 ). From Lemma 3,
it suffices to solve the heavy hitters problem on the sampled stream. From Lemma 2, because the hash
function is chosen from a universal family, the sampled elements have distinct hashed id’s. We can then feed
these elements into a standard Misra-Gries data structure with ε−1 counters, incurring space O(ε−1 log ε−1 ).
Because we want to return the unhashed element id’s for the heavy hitters, we also use log n space for
recording the ϕ−1 top items according to the Misra-Gries data structure and output these when asked to
report.
Proof of Theorem 1. The pseudocode of our (ε, ϕ)-List heavy hitters algorithm is in Algorithm 1. By
Lemma 3, if we select a subset S of size at least ℓ = 6ε−2 log(6δ −1 ) uniformly at random from the stream,
then Pr[∀i ∈ U, |(fˆi/|S|) − (fi/n)| 6 ε] > 1 − δ/3, where fi and fˆi are the frequencies of item i in the input
stream and S respectively. First we show that with the choice of p in line 10 in Algorithm 1, the number
of items sampled is at least ℓ and at most 11ℓ with probability at least (1 − δ/3). Let Xi be the indicator
random
Pmvariable of the event that the item xi is sampled for i ∈ [m]. Then the total number of items sampled
X = i=1 Xi . We have E[X] = 6ℓ since p = 6ℓ/m. Now we have the following.
Pr[X 6 ℓ or X > 11ℓ] 6 Pr[|X − E[X]| > 5ℓ] 6 δ/3
The inequality follows from the Chernoff bound and the value of ℓ. From here onwards we assume that the
number of items sampled is in [ℓ, 11ℓ].
We use (a modified version of) the Misra-Gries algorithm [MG82] to estimate the frequencies of items
in S. The length of the table in the Misra-Gries algorithm is ε−1 . We pick a hash function h uniformly at
2
random from a universal family H = {h|h : [n] → ⌈4ℓ /δ ⌉} of hash functions of size |H| = O(n2 ). Note that
picking a hash function h uniformly at random from H can be done using O(log n) bits of space. Lemma 2
shows that there are no collisions in S under this hash function h with probability at least 1 − δ/3. From here
onwards we assume that there is no collision among the ids of the sampled items under the hash function h.
We modify the Misra-Gries algorithm as follows. Instead of storing the id of any item x in the Misra-Gries
table (table T1 in line 5 in Algorithm 1) we only store the hash h(x) of the id x. We also store the ids (not
4 We
emphasize that this is stronger than an amortized guarantee, as on every insertion, the cost will be O(1).
8
Algorithm 1 for (ε, ϕ)-List heavy hitters
Input: A stream S of length m over U = [n]; let f (x) be the frequency of x ∈ U in S
Output: A set X ⊆ U and a function fˆ : X → N such that if f (x) > ϕm, then x ∈ X and f (x) − εm 6
fˆ(x) 6 f (x) + εm and if f (y) 6 (ϕ − ε)m, then y ∈
/ X for every x, y ∈ U
1: Initialize:
2: ℓ ← 6 log(6/δ)/ε2
2
3: Hash function h uniformly at random from a universal family H ⊆ {h : [n] → ⌈4ℓ /δ⌉}.
−1
4: An empty table T1 of (key, value) pairs of length ε . Each key entry of T1 can store an
2
integer in [0, ⌈400ℓ /δ⌉] and each value entry can store an integer in [0, 11ℓ].
⊲ The table T1
will be in sorted order by value throughout.
5: An empty table T2 of length 1/ϕ. Each entry of T2 can store an integer in [0, n]. ⊲ The entries
of T2 will correspond to ids of the keys in T1 of the highest 1/ϕ values
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
procedure Insert(x)
With probability p = 6ℓ/m, continue. Otherwise, return .
Perform Misra-Gries update using h(x) maintaining T1 sorted by values.
if The value of h(x) is among the highest 1/ϕ valued items in T1 then
if xi is not in T2 then
if T2 currently contains 1/ϕ many items then
For y in T2 such that h(y) is not among the highest 1/ϕ valued items in T1 , replace y with
x.
else
We put x in T2 .
Ensure that elements in T2 are ordered according to corresponding values in T1 .
procedure Report( )
return items in T2 along with their corresponding values in T1
the hash of the id) of the items with highest 1/ϕ values in T1 in another table T2 . Moreover, we always
maintain the table T2 consistent with the table T1 in the sense that the ith highest valued key in T1 is the
hash of the ith id in T2 .
Upon picking an item x with probability p, we create an entry corresponding to h(x) in T1 and make
its value one if there is space available in T1 ; decrement the value of every item in T1 by one if the table is
already full; increment the entry in the table corresponding to h(x) if h(x) is already present in the table.
When we decrement the value of every item in T1 , the table T2 remains consistent and we do not need to
do anything else. Otherwise there are three cases to consider. Case 1: h(x) is not among the 1/ϕ highest
valued items in T1 . In this case, we do not need to do anything else. Case 2: h(x) was not among the 1/ϕ
highest valued items in T1 but now it is among the 1/ϕ highest valued items in T1 . In this case the last item
y in T2 is no longer among the 1/ϕ highest valued items in T1 . We replace y with x in T2 . Case 3: h(x) was
among the 1/ϕ highest valued items in T1 . When the stream finishes, we output the ids of all the items in
table T2 along with the values corresponding to them in table T1 . Correctness follows from the correctness
of the Misra-Gries algorithm and the fact that there is no collision among the ids of the sampled items.
3.1.2
An optimal algorithm
Theorem 2. Assume the stream length is known beforehand. Then there is a randomized one-pass algorithm A for the (ε, ϕ)-List heavy
hitters problem which succeeds with constant probability using
O ε−1 log ϕ−1 + ϕ−1 log n + log log m bits of space. Moreover, A has an update time of O(1) and reporting
time linear in its output size.
Note that in this section, for the sake of simplicity, we ignore floors and ceilings and state the results for
9
a constant error probability, omitting the explicit dependence on δ.
Algorithm 2 for (ε, ϕ)-List heavy hitters
Input: A stream S of length m over universe U = [n]; let f (x) be the frequency of x ∈ U in S
Output: A set X ⊆ U and a function fˆ : X → N such that if f (x) > ϕm, then x ∈ X and f (x) − εm 6
fˆ(x) 6 f (x) + εm and if f (y) 6 (ϕ − ε)m, then y ∈
/ X for every x, y ∈ U
1: Initialize:
2: ℓ ← 105 ε−2
3: s ← 0
4: Hash functions h1 , . . . , h200 log(12ϕ−1 ) uniformly at random from a universal family
H ⊆ {h : [n] → [100/ε]}.
5: An empty table T1 of (key, value) pairs of length 2ϕ−1 . Each key entry of T1 can store an
element of [n] and each value entry can store an integer in [0, 10ℓ].
6: An empty table T2 with 100ε−1 rows and 200 log(12ϕ−1 ) columns. Each entry of T2 can
store an integer in [0, 100εℓ].
7: An empty 3-dimensional table T3 of size at most 100ε−1 × 200 log(12ϕ−1 ) × 4 log(ε−1 ). Each
entry of T3 can store an integer in [0, 10ℓ].
⊲ These are upper bounds; not all the allowed
cells will actually be used.
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
procedure Insert(x)
With probability ℓ/m, increment s and continue. Else, return
Perform Misra-Gries update on T1 with x.
for j ← 1 to 200 log(12ϕ−1 ) do
i ← hj (x)
With probability ε, increment T2 [i, j]
t ← ⌊log(10−6 T2 [i, j]2 )⌋ and p ← min(ε · 2t , 1)
if t > 0 then
With probability p, increment T3 [i, j, t]
procedure Report( )
X←∅
for each key x with nonzero value in T1 do
for j ← 1 to 200 log(12ϕ−1 ) do
P4 log(ε−1 )
T3 [h(x), j, t]/ min(ε2t , 1)
fˆj (x) ← t=0
fˆ(x) ← median(fˆ1 , . . . , fˆ10 log ϕ−1 )
if fˆ(x) > (ϕ − ε/2)s then
X ← X ∪ {x}
return X, fˆ
Overview As in the simpler algorithm, we sample ℓ = O(ε−2 ) many stream elements and solve the (ε/2, ϕ)List heavy hitters problem on this sampled stream. Also, the Misra-Gries algorithm for (ϕ/2, ϕ)-List
heavy hitters returns a candidate set of O(ϕ−1 ) items containing all items of frequency at least ϕℓ. It
remains to count the frequencies of these O(ϕ−1 ) items with upto εℓ/2 = O(ε−1 ) additive error, so that we
can remove those whose frequency is less than (ϕ − ε/2)ℓ.
Fix some item i ∈ [n], and let fi be i’s count in the sampled stream. A natural approach to count fi
approximately is to increment a counter probabilistically, instead of deterministically, at every occurrence of
i. Suppose that we increment a counter with probability 0 6 pi 6 1 whenever
h i item i arrives in the stream.
ˆ
Let the value of the counter be ĉi , and let fi = ĉi /pi . We see that E fˆi = fi and Var[fˆi ] 6 fi /pi . It
follows that if pi = Θ(ε2 fi ), then Var[fˆi ] = O(ε−2 ), and hence, fˆi is an unbiased estimator of fi with additive
10
error O(ε−1 ) with constant probability. We call such a counter an accelerated counter as the probability of
incrementing accelerates with increasing counts. For each i, we can maintain O(log ϕ−1 ) accelerated counters
independently and take their median to drive the probability of deviating by more than O(ε−1 ) down to
O(ϕ). So, with constant probability, the frequency for each of the O(ϕ−1 ) items in the Misra-Gries data
structure is estimated within O(ε−1 ) error, as desired.
However, there are two immediate issues with this approach. The first problem is that we may need to
keep counts for Ω(ℓ) = Ω(ε−2 ) distinct items, which is too costly for our purposes. To get around this, we
use a hash function from a universal family to hash the universe to a space of size u = Θ(ε−1 ), and we work
throughout with the hashed id’s. We can then show that the space complexity for each iteration is O(ε−1 ).
Also, the accelerated counters now estimate frequencies of hashed id’s instead of actual items, but because
of universality, the expected frequency of any hashed id is ℓ/u = O(ε−1 ), our desired error bound.
The second issue is that we need a constant factor approximation of fi , so that we can set pi to Θ(ε2 fi ).
But because the algorithm needs to be one-pass, we cannot first compute pi in one pass and then run the
accelerated counter in another. So, we divide the stream into epochs in which fi stays within a factor of 2,
and use a different pi for each epoch. In particular, set pti = ε · 2t for 0 6 t 6 log(pi /ε). We want to keep a
running estimate of i’s count to within a factor of 2 to know if the current epoch should be incremented. For
this, we subsample each element of the stream with probability ε independently and maintain exact counts
for the observed hashed id’s. It is easy to see that this requires only O(ε−1 ) bits in expectation. Consider
any i ∈ [u] and the prefix of the stream upto b 6 ℓ, and let fi (b) be i’s frequency in the prefix, let c̄i (b) be i’s
frequency among the samples in the prefix, and f¯i (b) = c̄iε(b) . We see that E f¯i (b) = fi (b), Moreover, we
show that for any b ∈ [ℓ], f¯i (b) is a 4-factor approximation of fi (b) with constant probability. By repeating
O(log ϕ−1 ) times independently and taking the median, the error probability can be driven down to O(ϕ).
Now, for every hashed id i ∈ [u], we need not one accelerated counter but O(log(εfi )) many, one corresponding to each epoch t. When an element with hash id i arrives at position b, we decide, based on f¯i (b),
the epoch t it belongs to and then increment the t’th accelerated counter with probability pti . The storage
cost over all i is still O(1/ε). Also, we iterate the whole set of accelerated counters O(log ϕ−1 ) times, making
the total storage cost O(ε−1 log ϕ−1 ).
P
Let ĉi,th bei the count in the accelerated counter for hash id i and epoch t. Then, let fˆi = t ĉi,t /pti .
Clearly, E fˆi = fi . The variance is O(ε−2 ) in each epoch, and so, Var[fˆi ] = O(ε−2 log ε−1 ), not O(ε−2 )
which we wanted. This issue is fixed by a change in how the sampling probabilities are defined. We now go
on to the formal proof.
Proof of Theorem 2. Pseudocode appears in Algorithm 2. Note that the numerical constants are chosen for
convenience of analysis and have not been optimized. Also, for the sake of simplicity, the pseudocode does
not have the optimal reporting time, but it can be modified to achieve this; see the end of this proof for
details.
By standard Chernoff bounds, with probability at least 99/100, the length of the sampled stream ℓ/10 6
s 6 10ℓ. For x ∈ [n], let fsamp (x) be the frequency of x in the sampled stream. By Lemma 3, with probability
at least 9/10, for all x ∈ [n]:
fsamp (x) f (x)
ε
−
6
s
m
4
P
−1
Now, fix j ∈ [10 log ϕ ] and x ∈ [n]. Let i = hj (x) and fi = y:hj (y)=hj (x) fsamp (y). Then, for a random
f
(x)
ε
hj ∈ H, the expected value of fsi − samp
is 100
, since H is a universal mapping to a space of size 100ε−1.
s
Hence, using Markov’s inequality and the above:
1
f (x) fsamp
fsamp (x) fi
1
3
ε
ε
ε
f (x) fi
6 Pr
+ Pr
<
−
−
−
+
<
(1)
>
>
>
Pr
m
s
2
m
s
4
m
s
4
10 25
20
In Lemma 4 below, we show that for each j ∈ [200 log(12ϕ−1 )], with error probability at most 3/10, fˆj (x)
(in line 23) estimates fi with additive error at most 5000ε−1, hence estimating fsi with additive error at
most 2ε . Taking the median over 200 log(12ϕ−1 ) repetitions (line 24) makes the error probability go down
11
to ϕ6 using standard Chernoff bounds. Hence, by the union bound, with probability at least 2/3, for each of
(x)
the 2/ϕ keys x with nonzero values in T1 , we have an estimate of f m
within additive error ε, thus showing
correctness.
Lemma 4. Fix x ∈ [n] and j ∈ [200 log 12ϕ−1 ], and let i = hj (x). Then, Pr[|fˆj (x) − fi | > 5000ε−1] 6 3/10,
where fˆj is the quantity computed in line 23.
Proof. Index the sampled stream elements 1, 2, . . . , s, and for b ∈ [s], let fi (b) be the frequency of items with
hash id i restricted to the first b elements of the sampled stream. Let f¯i (b) denote the value of T2 [i, j] · ε−1
after the procedure Insert has been called for the first b items of the sampled stream.
Claim 1. With probability at least 9/10, for all b ∈ [s] such that fi (b) > 100ε−1, f¯i (b) is within a factor of
4 of fi (b).
Proof. Fix b ∈ [s]. Note that E f¯i (b) = fi (b) as T2 is incremented with rate ε. Var[f¯i (b)] 6 fi /ε, and so by
Chebyshev’s inequality:
4
Pr[|f¯i (b) − fi (b)| > fi (b)/2] <
fi (b)ε
We now break the stream into chunks, apply this inequality to each chunk and then take a union bound to
conclude. Namely, for any integer t > 0, define bt to be the first b such that 100ε−1 2t 6 fi (b) < 100ε−12t+1
if such a b exists. Then:
X
4
Pr[∃t > 0 : |f¯i (bt ) − fi (bt )| > fi (bt )/2] <
100
·
2t−1
t
<
1
10
So, with probability at least 9/10, every f¯i (bt ) and fi (bt ) are within a factor of 2 of each other. Since for
every b > b0 , fi (b) is within a factor of 2 from some fi (bt ), the claim follows.
Assume the event in Claim 1 henceforth. Now, we are ready to analyze T3 and in particular, fˆj (x). First
of all, observe that if t < 0 in line 15, at some position b in the stream, then T2 [i, j] at that time must be at
most 1000, and so by standard Markov and Chernoff bounds, with probability at least 0.85,
(
< 4000ε−1, if t < 0
fi (b)
(2)
> 100ε−1 ,
if t > 0
i
h
Assume this event. Then, fi − 4000ε−1 6 E fˆj (x) 6 fi .
Claim 2.
Var(fˆj (x)) 6 20000ε−2
Proof. If the stream element at position b causes an increment in T3 with probability ε2t (in line 17), then
1000 · 2t/2 6 T2 [i, j] 6 1000 · 2(t+1)/2 , and so, f¯i (b) 6 1000ε−12(t+1)/2 . This must be the case for the highest
b = b̄t at which the count for i in T3 increments at the t’th slot. The number of such occurrences of i is at
most fi (b̄t ) 6 4f¯i (b̄t ) 6 4000ε−12(t+1)/2 by Claim 1 (which can be applied since fi (b) > 100ε−1 by Equation
2). So:
Var[fˆj (x)] 6
X fi (b̄t )
t>0
ε2t
6
X 4000
t>0
ε2
2−t/3 6 20000ε−2
Elements inserted with probability 1 obviously do not contribute to the variance.
12
So, conditioning on the events mentioned, the probability that fˆj (x) deviates from fi by more than
5000ε−1 is at most 1/50. Removing all the conditioning yields what we wanted:
Pr[|fˆj (x) − fi | > 5000ε−1] 6
1
3
1
+
+
6 0.3
50 20 10
We next bound the space complexity.
Claim 3. With probability at least 2/3, Algorithm 2 uses O(ε−1 log ϕ−1 + ϕ−1 log n + log log m) bits of
storage, if n = ω(ε−1 ).
Proof. The expected length of the sampled stream is ℓ = O(ε−2 ). So, the number of bits stored in T1 is
O(ϕ−1 log n). For T2 , note that in lines 13-15, for any given j, T2 is storing a total of εℓ = O(ε−1 ) elements
in expectation. So, for k > 0, there can be at most O((ε2k )−1 ) hashed id’s with counts between 2k and 2k+1 .
Summing over all k’s and accounting for the empty cells gives O(ε−1 ) bits of storage, and so the total space
requirement of T2 is O(ε−1 log ϕ−1 ). .
The probability that a hashed id i gets counted in table T3 is at most 10−6 ε3 f¯i2 (s) from line 15 and our
definition of f¯i above. Moreover, from Claim 1, we have that this is at most 16 · 10−6 ε3 fi2 (s) if fi > 100ε−1.
Therefore, if fi = 2k · 100ε−1 with k > 0, then the expected value of a cell in T3 with first coordinate i is at
most 1600 · 22k ε = 2O(k) . Taking into account that there are at most O((ε2k )−1 ) many such id’s i and that
the number of epochs t associated with such an i is at most log(16 · 10−6 ε2 fi2 ) = O(log(εfi )) = O(k) (from
line 15), we get that the total space required for T3 is:
O(log ϕ−1 )
X
O(ε−1 ) +
j=1
= O(ε−1 log ϕ−1 )
∞
X
k=0
!
O((ε2k )−1 ) · O(k) · O(k)
where the first O(ε−1 ) term inside the summation is for the i’s with fi < 100ε−1. Since we have an expected
space bound, we obtain a worst-case space bound with error probability 1/3 by a Markov bound.
The space required for sampling is an additional O(log log m), using Lemma 1.
We note that the space bound can be made worst case by aborting the algorithm if it tries to use more
space.
The only remaining aspect of Theorem 2 is the time complexity. As observed in Section 3.1, the update
time can be made O(1) per insertion under the standard assumption of the stream being sufficiently long.
The reporting time can also be made linear in the output by changing the bookkeeping a bit. Instead of
computing fˆj and fˆ at reporting time, we can maintain them after every insertion. Although this apparently
makes INSERT costlier, this is not true in fact because we can spread the cost over future stream insertions.
The space complexity grows by a constant factor.
3.2
Maximum
By tweaking Algorithm 1 slightly, we get the following result for the ε-Maximum problem.
Theorem 3. Assume the length of the stream is known beforehand. Then there is a randomized onepass algorithm A for the ε-Maximum problem which succeeds with probability at least 1 − δ using
O (min{1/ε, n}(log 1/ε + log log 1/δ) + log n + log log m) bits of space. Moreover, the algorithm A has an update time of O(1).
Proof. Instead of maintaining the table T2 in Algorithm 1, we just store the actual id of the item with
maximum frequency in the sampled items.
13
3.3
Minimum
Theorem 4. Assume the length of the stream is known beforehand. Then there is a randomized
one-pass algorithm A for the ε-Minimum problem which succeeds with probability at least 1 − δ using
O ((1/ε) log log(1/εδ) + log log m) bits of space. Moreover, the algorithm A has an update time of O(1).
Overview Pseudocode is provided in Algorithm 3. The idea behind our ε-Minimum problem is as follows.
It is most easily explained by looking at the REPORT(x) procedure starting in line 13. In lines 14-15 we
ask, is the universe size |U | significantly larger than 1/ε? Note that if it is, then outputting a random item
from |U | is likely to be a solution. Otherwise |U | is O(1/ε).
The next point is that if the number of distinct elements in the stream were smaller than 1/(ε log(1/ε)),
then we could just store all the items together with their frequencies with O(1/ε) bits of space. Indeed,
we can first sample O(1/ε2 ) stream elements so that all relative frequencies are preserved up to additive ε,
thereby ensuring each frequency can be stored with O(log(1/ε) bits. Also, since the universe size is O(1/ε),
the item identifiers can also be stored with O(log(1/ε) bits. So if this part of the algorithm starts taking
up too much space, we stop, and we know the number of distinct elements is at least 1/(ε log(1/ε)), which
means that the minimum frequency is at most O(mε log(1/ε)). This is what is being implemented in steps
9-10 and 18-19 in the algorithm.
We can also ensure the minimum frequency is at least Ω(mε/ log(1/ε)). Indeed, by randomly sampling
O((log(1/ε)/ε) stream elements, and maintaining a bit vector for whether or not each item in the universe
occurs - which we can with O(1/ε) bits of space since |U | = O(1/ε) - any item with frequency at least
Ω(εm/ log(1/ε)) will be sampled and so if there is an entry in the bit vector which is empty, then we can
just output that as our solution. This is what is being implemented in steps 8 and 16-17 of the algorithm.
Finally, we now know that the minimum frequency is at least Ω(mε/ log(1/ε)) and at most O(mε log(1/ε)).
At this point if we randomly sample O((log6 1/ε)/ε) stream elements, then by Chernoff bounds all item
frequencies are preserved up to a relative error factor of (1 ± 1/ log2 (1/ε)), and in particular the relative
minimum frequency is guaranteed to be preserved up to an additive ε. At this point we just maintain the
exact counts in the sampled stream but truncate them once they exceed poly(log(1/ε))) bits, since we know
such counts do not correspond to the minimum. Thus we only need O(log log(1/ε)) bits to represent their
counts. This is implemented in step 11 and step 20 of the algorithm.
Proof of Theorem 4. The pseudocode of our ε-Minimum algorithm is in Algorithm 3. If the size of the
universe |U| is at least 1/((1−δ)ε), then we return an item x chosen from U uniformly at random. Note that
there can be at most 1/ε many items with frequency at least εm. Hence every item x among other remaining
δ/((1−δ)ε) many items has frequency less than εm and thus is a correct output of the instance. Thus the
probability that we answer correctly is at least (1 − δ). From here on, let us assume |U| < 1/((1−δ)ε).
Now, by the value of pj , it follows from the proof of Theorem 1 that we can assume ℓj < |Sj | < 11ℓj
for j = 1, 2, 3 which happens with probability at least (1 − (δ/3)). We first show that every item in U with
frequency at least εm is sampled in S1 with probability at least (1 − (δ/6)). For that, let Xij be the indicator
random variable for the event that the j th sample in S1 is item i where i ∈ U is an item with frequency at
least εm. Let H ⊂ U be the set of items with frequencies at least εm. Then we have the following.
Pr[Xij = 0] = 1 − ε
⇒ Pr[Xij = 0 ∀j ∈ S1 ] 6 (1 − ε)ℓ1 6 exp{−εℓ1 } = εδ/6
Now applying union bound we get the following.
Pr[∃i ∈ H, Xij = 0 ∀j ∈ S1 ] 6 (1/ε)εδ/6 6 δ/6
Hence with probability at least (1 − (δ/3) − (δ/6)) > (1 − δ), the output at line 17 is correct. Now we show
below that if the frequency of any item x ∈ U is at most ε ln(6/δ)/ln(6/εδ), then x ∈ S1 with probability at least
(1 − (δ/6)).
ln(6/εδ)/ε
Pr[x ∈
/ S1 ] = (1 − ε ln(6/δ)/ln(6/εδ))
6 δ/6
14
Algorithm 3 for ε-Minimum
Input: A stream S = (xi )i∈[m] ∈ U m of length m over U; let f (x) be the frequency of x ∈ U in S
Output: An item x ∈ U such that f (x) 6 f (y) + εm for every y ∈ U
1: Initialize:
2:
3:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
4:
5:
ℓ1 ← log(6/εδ)/ε, ℓ2 ← log(6/δ)/ε2 , ℓ3 ← log
p1 ← 6ℓ1/m, p2 ← 6ℓ2/m, p3 ← 6ℓ3/m
S1 , S2 , S3 ← ∅
B1 ← the bit vector for S1
6 6
( /δε)
/ε
procedure Insert(x)
Put x in S1 with probability p1 by updating the bit vector B1
if the number of distinct items in the stream so far is at most 1/(ε log(1/ε)) then
Pick x with probability p2 and put the id of x in S2 and initialize the corresponding counter to 1
if x ∈
/ S2 and increment the counter corresponding to x by 1.
Pick x with probability p3 , put the id of x in S3 and initialize the corresponding counter to 1 if
xi ∈
/ S3 and increment the counter corresponding to xi by 1. Truncate counters of S3 at 2 log7 (2/εδ).
procedure Report( )
if |U| > 1/((1−δ)ε) then
return an item x from the first 1/((1−δ)ε) items in U (ordered arbitrarily) uniformly at random
if S1 6= U then
return any item from U \ S1
if the number of distinct items in the stream is at most 1/(ε log(1/ε)) then
return an item in S2 with minimum counter value in S2
return the item with minimum frequency in S3
Hence from here onwards we assume that the frequency of every item in U is at least εm ln(6/δ)/ln(6/εδ).
If the number of distinct elements is at most 1/(ε ln(1/ε)), then line 19 outputs the minimum frequency
item up to an additive factor of εm due to Chernoff bound. Note that we need only O(ln(1/((1−δ)ε))) bits of
space for storing ids. Hence S2 can be stored in space O((1/ε ln(1/ε)) ln(1/((1−δ)ε) ln ln(1/δ )) = O(1/ε ln ln(1/δ)).
Now we can assume that the number of distinct elements is at least 1/(ε ln(1/ε)). Hence if f (t) is the
frequency of the item t with minimum frequency, then we have mε/ln(1/ε) 6 f (t) 6 mε ln(1/ε).
Let fi be the frequency of item i ∈ U, ei be the counter value of i in S3 , and fˆi = ei m/ℓ3 . Now again by
applying Chernoff bound we have the following for any fixed i ∈ U.
Pr[|fi − fˆi | > fi/ln2 (1/ε)] 6 2 exp{−ℓ3 fi/(m ln4 (1/ε))}
2 6
( /εδ )
6 2 exp{−fi ln
6 εδ/6.
/(εm)}
Now applying union bound we get the following using the fact that |U| 6 1/ε(1−δ).
Pr[∀i ∈ U, |fi − fˆi | 6 fi/ln2 (1/ε)] > 1 − δ/6
Again by applying Chernoff bound and union bound we get the following.
Pr[∀i ∈ U with fi > 2mε ln(1/ε), |fi − fˆi | 6 fi/2] > 1 − δ/6
Hence the items with frequency more than 2mε ln(1/ε) are approximated up to a multiplicative factor of 1/2
from below in S3 . The counters of these items may be truncated. The other items with frequency at most
2mε ln(1/ε) are be approximated up to (1 ± 1/ln2 (1/ε)) relative error and thus up to an additive error of εm/3.
15
The counters of these items would not get truncated. Hence the item with minimum counter value in S3 is
the item with minimum frequency up to an additive εm.
We need O(ln(1/εδ )) bits of space for the bit vector B1 for the set S1 . We need O(ln2 (1/εδ )) bits of space
for the set S2 and O((1/ε) ln ln(1/εδ )) bits of space for the set S3 (by the choice of truncation threshold). We
need an additional O (ln ln m) bits of space for sampling using Lemma 1. Moreover, using the data structure
of Section 3.3 of [DLOM02] Algorithm 3 can be performed in O(1) time. Alternatively, we may also use the
strategy described in Section 3.1 of spreading update operations over several insertions to make the cost per
insertion be O(1).
3.4
Borda and Maximin
Theorem 5. Assume the length of the stream is known beforehand. Then there is a randomized onepass algorithm A for (ε, ϕ)-List
Borda
problem which succeeds with probability at least 1 − δ using
O n log n + log 1ε + log log δ1 + log log m bits of space.
Proof. Let ℓ = 6ε−2 log(6nδ −1 ) and p = 6ℓ/m. On each insertion of a vote v, select v with probability p and
store for every i ∈ [n], the number of candidates that candidate i beats in the vote v. Keep these exact
counts in a counter of length n.
Then it follows from the proof of Theorem 1 that ℓ 6 |S| 6 11ℓ with probability at least (1 − δ/3).
Moreover, from a straightforward application of the Chernoff bound (see [DB15]), it follows that if ŝ(i)
denotes the Borda score of candidate i restricted to the sampled votes, then:
m
ŝ(i) − s(i) < εmn > 1 − δ
Pr ∀i ∈ [n],
|S|
The space complexity for exactly storing the counts is O(n log(nℓ)) = O(n(log n + log ε−1 + log log δ −1 ))
and the space for sampling the votes is O(log log m) by Lemma 1.
Theorem 6. Assume the length of the stream is known beforehand. Then there is a randomized onepass algorithm A for (ε, ϕ)-List maximin problem
which succeeds with probability at least 1 − δ using
O nε−2 log2 n + nε−2 log n log δ −1 + log log m bits of space.
Proof. Let ℓ = (8/ε2 ) ln(6n/δ ) and p = 6ℓ/m. We put the current vote in a set S with probability p. Then
it follows from the proof of Theorem 1 that ℓ 6 |S| 6 11ℓ with probability at least (1 − δ/3). Suppose
|S| = ℓ1 ; let S = {vi : i ∈ [ℓ1 ]} be the set of votes sampled. Let DE (x, y) be the total number of votes
in which x beats y and DS (x, y)) be the number of such votes in S. Then by the choice of ℓ and the
Chernoff bound (see [DB15]), it follows that |DS (x, y)m/ℓ1 − DE (x, y)| 6 εm/2 for every pair of candidates
x, y ∈ U. Note that each vote can be stored in O(n log n) bits of space. Hence simply finding DS (x, y)
for every x, y ∈ U by storing S and returning all the items with maximin score at least (ϕ − ε/2)ℓ1 in S
requires O nε−2 log n(log n + log δ −1 ) + log log m bits of memory, with the additive O(log log m) due to
Lemma 1.
3.5
Unknown stream length
Now we consider the case when the length of the stream is not known beforehand. We present below an
algorithm for (ε, ϕ)-List heavy hitters and ε-Maximum problems in the setting where the length of the
stream is not known beforehand.
Theorem 7. There is a randomized one-pass algorithm for (ε, ϕ)-List
heavy hitters and ε-Maximum
problems with space complexity O ε−1 log ε−1 + ϕ−1 log n + log log m bits and update time O(1) even when
the length of the stream is not known beforehand.
16
Proof. We describe below a randomized one-pass algorithm for the (8ε, ϕ)-List heavy hitters problem.
We may assume that the length of the stream is at least 1/ε2 ; otherwise, we use the algorithm in Theorem 1
and get the result. Now we guess the length of the stream to be 1/ε2 , but run an instance I1 of Algorithm 1
with ℓ = log(6/δ)/ε3 at line 2. By the choice of the size of the sample (which is Θ(log(1/δ)/ε3 )), I1 outputs
correctly with probability at least (1 − δ), if the length of the stream is in [1/ε2 , 1/ε3 ]. If the length of the
stream exceeds 1/ε2 , we run another instance I2 of Algorithm 1 with ℓ = log(6/δ)/ε3 at line 2. Again by
the choice of the size of the sample, I2 outputs correctly with probability at least (1 − δ), if the length
of the stream is in [1/ε3 , 1/ε4 ]. If the stream length exceeds 1/ε3 , we discard I1 , free the space it uses, and
run an instance I3 of Algorithm 1 with ℓ = log(6/δ)/ε3 at line 2 and so on. At any point of time, we have
at most two instances of Algorithm 1 running. When the stream ends, we return the output of the older
of the instances we are currently running. We use the approximate counting method of Morris [Mor78]
to approximately count the length of the stream. We know that the Morris counter outputs correctly with
k
probability (1 − 2− /2 ) using O(log log m+ k) bits of space at any point in time [Fla85]. Also, since the Morris
counter increases only when an item is read, it outputs correctly up to a factor of four at every position if it
outputs correctly at positions 1, 2, 4, . . . , 2⌊log2 m⌋ ; call this event E. Then we have Pr(E) > 1 − δ by choosing
k = 2 log2 (log2 m/δ) and applying union bound over the positions 1, 2, 4, . . . , 2⌊log2 m⌋ . The correctness of the
algorithm follows from the correctness of Algorithm 1 and the fact that we are discarding at most εm many
items in the stream (by discarding a run of an instance of Algorithm 1). The space complexity and the O(1)
update time of the algorithm follow from Theorem 1, the choice of k above, and the fact that we have at
most two instances of Algorithm 1 currently running at any point of time.
The algorithm for the ε-Maximum problem is same as the algorithm above except we use the algorithm
in Theorem 3 instead of Algorithm 1.
Note that this proof technique does not seem to apply to our optimal Algorithm 2. Similarly to Theorem 7,
we get the following result for the other problems.
Theorem 8. There are randomized one-pass algorithms for ε-Minimum, (ε, ϕ)-Borda, and (ε, ϕ)-Maximin
problems with space complexity O ((1/ε) log log(1/εδ ) + log log m), O n log n + log 1ε + log log 1δ + log log m ,
and O nε−2 log2 n + nε−2 log n log(1/δ) + log log m bits respectively even when the length of the stream is
not known beforehand. Moreover, the update time for ε-Minimum is O(1).
4
Hardness
In this section, we prove space complexity lower bounds for the ε-Heavy hitters, ε-Minimum, ε-Borda,
and ε-maximin problems. We present reductions from certain communication problems for proving space
complexity lower bounds. Let us first introduce those communication problems with necessary results.
4.1
Communication Complexity
Definition 10. ( Indexingm,t )
Let t and m be positive integers. Alice is given a string x = (x1 , · · · , xt ) ∈ [m]t . Bob is given an index i ∈ [t].
Bob has to output xi .
The following is a well known result [KN97].
Lemma 5. R1-way
(Indexingm,t ) = Ω(t log m) for constant δ ∈ (0, 1).
δ
[SW15] defines a communication problem called Perm, which we generalize to ε-Perm as follows.
Definition 11. (ε-Perm)
Alice is given a permutation σ over [n] which is partitioned into 1/ε many contiguous blocks. Bob is given an
index i ∈ [n] and has to output the block in σ where i belongs.
17
Our lower bound for ε-Perm matches the lower bound for Perm in Lemma 1 in [SW15] when ε = 1/n.
For the proof, the reader may find useful some information theory facts described in Appendix A.
Lemma 6. R1-way
(ε − Perm) = Ω(n log(1/ε)), for any constant δ < 1/10.
δ
Proof. Let us assume σ, the permutation Alice has, is uniformly distributed over the set of all permutations.
Let τj denotes the block the item j is in for j ∈ [n], τ = (τ1 , . . . , τn ), and τ<j = (τ1 , . . . , τj−1 ). Let M (τ ) be
Alice’s message to Bob, which is a random variable depending on the randomness of σ and the private coin
tosses of Alice. Then we have R1−way (ε − Perm) > H(M (τ )) > I(M (τ ); τ ). Hence it is enough to lower
bound I(M (τ ); τ ). Then we have the following by chain rule.
I(M (τ ); τ ) =
=
>
n
X
j=1
n
X
j=1
n
X
j=1
I(M (τ ); τj |τ<j )
H(τj |τ< j) − H(τj |M (τ ), τ< j)
H(τj |τ< j) −
= H(τ ) −
n
X
j=1
n
X
j=1
H(τj |M (τ ))
H(τj |M (τ ))
n
The number of ways to partition n items into 1/ε blocks is n!/((εn)!)(1/ε) which is Ω((n/e) /(εn/e)n ). Hence we
have H(τ ) = n log(1/ε). Now we consider H(τj |M (τ )). By the correctness of the algorithm, Fano’s inequality,
we have H(τj |M (τ )) 6 H(δ) + (1/10) log2 ((1/ε) − 1) 6 (1/2) log(1/ε). Hence we have the following.
I(M (τ ); τ ) > (n/2) log(1/ε)
Finally, we consider the Greater-than problem.
Definition 12. ( Greater-thann )
Alice is given an integer x ∈ [n] and Bob is given an integer y ∈ [n], y 6= x. Bob has to output 1 if x > y and
0 otherwise.
The following result is due to [Smi88, MNSW98]. We provide a simple proof of it that seems to be
missing5 in the literature.
Lemma 7. R1-way
(Greater-thann ) = Ω(log n), for every δ < 1/4.
δ
Proof. We reduce the Augmented-indexing2,⌈log n⌉+1 problem to the Greater-thann problem thereby
proving the result. Alice runs the Greater-thann protocol with its input number whose representation in
binary is a = (x1 x2 · · · x⌈log n⌉ 1)2 . Bob participates in the Greater-thann protocol with its input number
whose representation in binary is b = (x1 x2 · · · xi−1 1
0| ·{z
· · 0}
)2 . Now xi = 1 if and only if a > b.
(⌈log n⌉−i+1) 0′ s
4.2
Reductions
We observe that a trivial Ω((1/ϕ) log n) bits lower bound for (ε, ϕ)-List heavy hitters, (ε, ϕ)-List borda,
(ε, ϕ)-List maximin follows from the fact that any algorithm may need to output 1/ϕ many items from the
universe. Also, there is a trivial Ω(n log n) lower bound for (ε, ϕ)-List borda and (ε, ϕ)-List maximin
because each stream item is a permutation on [n], hence requiring Ω(n log n) bits to read.
We show now a space complexity lower bound of Ω( 1ε log ϕ1 ) bits for the ε-Heavy hitters problem.
5A
similar proof appears in [KNR99] but theirs gives a weaker lower bound.
18
Theorem 9. Suppose the size of universe n is at least 1/4ε(ϕ−ε) and that ϕ > 2ε. Any randomized one
pass (ε, ϕ)-Heavy hitters algorithm with success probability at least (1 − δ) must use Ω((1/ε) log 1/ϕ) bits
of space, for constant δ ∈ (0, 1).
Proof. Consider the Indexing1/2(ϕ−ε),1/2ε problem where Alice is given a string x = (x1 , x2 , · · · , x1/ε ) ∈
1
[1/2(ϕ−ε)] /2ε and Bob is given an index i ∈ [1/2ε]. We assume ϕ > 2ε. The stream we generate is over
[1/2(ϕ−ε)] × [1/2ε] ⊆ U (this is possible since |U| > 1/(4ε(ϕ−ε))).
Let m be a large positive integer. Alice generates a stream of length m/2 by inserting εm copies of (xj , j)
for each j ∈ [1/2ε]. Alice now sends the memory content of the algorithm to Bob. Bob resumes the run of
the algorithm by generating another stream of length m/2 by inserting (ϕ − ε)m copies of (j, i) for each
j ∈ [1/2(ϕ−ε)]. The length of the stream is m, the frequency of the item (xi , i) is ϕm, while the frequency of
every other item is (ϕ − ε)m or εm. Hence from the output of the (ε, ϕ)-Heavy hitters algorithm, Bob
1
knows i with probability at least (1−δ). Now the result follows from Lemma 5, since 1ε log ϕ−ε
> 1ε log ϕ1 .
We now use the same idea as in the proof of Theorem 9 to prove an Ω( 1ε log 1ε ) space complexity lower
bound for the ε-Maximum problem.
Theorem 10. Suppose the size of universe n is at least ε12 . Any randomized one pass ε-Maximum algorithm
with success probability at least (1 − δ) must use Ω( 1ε log 1ε ) bits of space, for constant δ ∈ (0, 1).
Proof. Consider the Indexing1/ε,1/ε problem where Alice is given a string x = (x1 , x2 , · · · , x1/ε ) ∈ [1/ε] /ε
and Bob is given an index i ∈ [1/ε]. The stream we generate is over [1/ε] × [1/ε] ⊆ U (this is possible since
|U| > ε12 ). Let m be a large positive integer. Alice generates a stream of length m/2 in such a way that the
frequency of every item in {(xj , j) : j ∈ [1/ε]} is at least ⌊εm/2⌋ and the frequency of any other item is 0.
Alice now sends the memory content of the algorithm to Bob. Bob resumes the run of the algorithm by
generating another stream of length m/2 in such a way that the frequency of every item in {(j, i) : j ∈ [1/ε]}
is at least ⌊εm/2⌋ and the frequency of any other item is 0. The frequency of the item (xi , i) is at least ⌊εm⌋
where as the frequency of every other item is at most ⌊εm/2⌋. Hence the ε/5-Maximum algorithm must
output (xi , i) with probability at least (1 − δ). Now the result follows from Lemma 5.
1
For ε-Minimum, we prove a space complexity lower bound of Ω(1/ε) bits.
Theorem 11. Suppose the universe size n is at least 1/ε. Then any randomized one pass ε-Minimum
algorithm must use Ω(1/ε) bits of space.
Proof. We reduce from Indexing2,5/ε to ε-Minimum thereby proving the result. Let the inputs to Alice and
Bob in Indexing2,5/ε be (x1 , . . . , x5/ε ) ∈ {0, 1}5/ε and an index i ∈ [5/ε] respectively. Alice and Bob generate
a stream S over the universe [(5/ε) + 1]. Alice puts two copies of item j in S for every j ∈ U with xj = 1 and
runs the ε-Minimum algorithm. Alice now sends the memory content of the algorithm to Bob. Bob resumes
the run of the algorithm by putting two copies of every item in U \ {i, (5/ε) + 1} in the stream S. Bob also
puts one copy of (5/ε)+1 in S. Suppose the size of the support of (x1 , . . . , x5/ε ) be ℓ. Since 1/(2ℓ+(2/ε)−1) > ε/5,
we have the following. If xi = 0, then the ε-Minimum algorithm must output i with probability at least
(1 − δ). If xi = 1, then the ε-Minimum algorithm must output (5/ε) + 1 with probability at least (1 − δ).
Now the result follows from Lemma 5.
We show next a Ω(n log(1/ε)) bits space complexity lower bound for ε-Borda.
Theorem 12. Any one pass algorithm for ε-Borda must use Ω(n log(1/ε)) bits of space.
Proof. We reduce ε-Perm to ε-Borda. Suppose Alice has a permutation σ over [n] and Bob has an index
i ∈ [n]. The item set of our reduced election is U = [n] ⊔ D, where D = {d1 , d2 , . . . , d2n }. Alice generates a
vote v over the item set U from σ as follows. The vote v is B1 ≻ B2 ≻ · · · ≻ B1/ε where Bj for j = 1, . . . , 1/ε
is defined as follows.
Bj = d(j−1)2εn+1 ≻ d(j−1)2εn+2 ≻ · · · ≻ d(2j−1)εn
≻ σjεn+1 ≻ · · · ≻ σ(j+1)εn ≻ d(2j−1)ε+1 ≻ · · · ≻ d2jεn
19
Alice runs the ε-Borda algorithm with the vote v and sends the memory content to Bob. Let D−i =
−−→
←−−
−−→
D \ {i}, D−i be an arbitrary but fixed ordering of the items in D−i , and D−i be the reverse ordering of D−i .
−−→
←−−
Bob resumes the algorithm by generating two votes each of the form i ≻ D−i and i ≻ D−i . Let us call the
resulting election E. The number of votes m in E is 5. The Borda score of the item i is at least 12n. The
Borda score of every item x ∈ U is at most 9n. Hence for ε < 1/15, the ε-Borda algorithm must output the
item i. Moreover, it follows from the construction of v that an εmn additive approximation of the Borda
score of the item i reveals the block where i belongs in the ε-Perm instance.
We next give a nearly-tight lower bound for the ε-maximin problem.
Theorem 13. Any one-pass algorithm for ε-maximin requires Ω(n/ε2 ) memory bits of storage.
Proof. We reduce from Indexing. Let γ = 1/ε2 . Suppose Alice has a string y of length (n−γ)·γ, partitioned
into n − γ blocks of length γ each. Bob has an index ℓ = i + (j − γ − 1) · γ where i ∈ [γ], j ∈ {γ + 1, . . . , n}.
The Indexing problem is to return yℓ for which there is a Ω(|y|) = Ω(n/ε2 ) lower bound (Lemma 5).
The initial part of the reduction follows the construction in the proof of Theorem 6 in [VWWZ15], which
we encapsulate in the following lemma.
Lemma 8 (Theorem 6 in [VWWZ15]). Given y, Alice can construct a matrix P ∈ {0, 1}n×γ using public
randomness, such that if P i and P j are the i’th and j’th rows of P respectively, then with probability at least
√
√
2/3, ∆(P i , P j ) > γ2 + γ if yℓ = 1 and ∆(a, b) 6 γ2 − γ if yℓ = 0.
Let Alice construct P according to Lemma 8 and then adjoin the bitwise complement of the matrix P
below P to form the matrix P ′ ∈ {0, 1}2n×γ ; note that each column of P ′ has exactly n 1’s and n 0’s. Now,
we interpret each row of P as a candidate and each column of P as a vote in the following way: for each
′
v ∈ [γ], vote v has the candidates in {c : Pc,v
= 1} in ascending order in the top n positions and the rest of
the candidates in ascending order in the bottom n positions. Alice inserts these γ votes into the stream and
sends the state of the ε-Maximin algorithm to Bob as well as the Hamming weight of each row in P ′ . Bob
inserts γ more votes, in each of which candidate i comes first, candidate j comes second, and the rest of the
2n − 2 candidates are in arbitrary order.
Note that because of Bob’s votes, the maximin score of j is the number of votes among the ones casted
by Alice in which j defeats i. Since i < j, in those columns v where Pi,v = Pj,v , candidate i beats
candidate j. Thus, the set of votes in which j defeats i is {v | Pi,v = 0, Pj,v = 1}. The size of this set is
√
1
i
j
j
i
Therefore, if Bob can estimate the maximin score of j upto γ/4 additive error,
2 ∆(P , P ) + |P | − |P | . √
he can find ∆(P i , P j ) upto γ/2 additive error as Bob knows |P i | and |P j |. This is enough, by Lemma 8,
to solve the Indexing problem with probability at least 2/3.
Finally, we show a space complexity lower bound that depends on the length of the stream m.
Theorem 14. Any one pass algorithm for ε-Heavy hitters, ε-Minimum, ε-Borda, and ε-maximin must
use Ω(log log m) memory bits, even if the stream is over a universe of size 2, for every ε < 14 .
Proof. It is enough to prove the result only for ε-Heavy hitters since the other three problems reduce
to ε-Heavy hitters for a universe of size 2. Suppose we have a randomized one pass ε-Heavy hitters
algorithm which uses s(m) bits of space. Using this algorithm, we will show a communication protocol for
the Greater-thanm problem whose communication complexity is s(2m ) thereby proving the statement.
The universal set is U = {0, 1}. Alice generates a stream of 2x many copies of the item 1. Alice now sends
the memory content of the algorithm. Bob resumes the run of the algorithm by generating a stream of 2y
many copies of the item 0. If x > y, then the item 1 is the only ε-winner; whereas if x < y, then the item 0
is the only ε-winner.
Acknowledgments
We would like to thank Jelani Nelson for a helpful conversation which led us to discover an error in a previous
version of the paper.
20
References
[ABW03] Arvind Arasu, Shivnath Babu, and Jennifer Widom. CQL: A language for continuous queries
over streams and relations. In Database Programming Languages, 9th International Workshop,
DBPL 2003, Potsdam, Germany, September 6-8, 2003, Revised Papers, pages 1–19, 2003.
[AM01] Javed A Aslam and Mark Montague. Models for metasearch. In Proceedings of the 24th annual
international ACM SIGIR conference on Research and development in information retrieval,
pages 276–284. ACM, 2001.
[AS94] Rakesh Agrawal and Ramakrishnan Srikant. Fast algorithms for mining association rules in
large databases. In VLDB’94, Proceedings of 20th International Conference on Very Large Data
Bases, September 12-15, 1994, Santiago de Chile, Chile, pages 487–499, 1994.
[AT05] Gediminas Adomavicius and Alexander Tuzhilin. Toward the next generation of recommender
systems: A survey of the state-of-the-art and possible extensions. Knowledge and Data Engineering, IEEE Transactions on, 17(6):734–749, 2005.
[BB08] Daniel K Blandford and Guy E Blelloch. Compact dictionaries for variable-length keys and data
with applications. ACM Transactions on Algorithms (TALG), 4(2):17, 2008.
[BCE+ 15] Felix Brandt, Vincent Conitzer, Ulle Endriss, Jérôme Lang, and Ariel Procaccia. Handbook of
computational social choice, 2015.
[BCIW16] Vladimir Braverman, Stephen R. Chestnut, Nikita Ivkin, and David P. Woodruff.
Beating countsketch for heavy hitters in insertion streams.
2016.
Preprint at
http://arxiv.org/abs/1511.00661. To appear in STOC ‘16.
[BGS01] Philippe Bonnet, Johannes Gehrke, and Praveen Seshadri. Towards sensor database systems. In
Mobile Data Management, Second International Conference, MDM 2001, Hong Kong, China,
January 8-10, 2001, Proceedings, pages 3–14, 2001.
[BICS10] Radu Berinde, Piotr Indyk, Graham Cormode, and Martin J. Strauss. Space-optimal heavy
hitters with strong error bounds. ACM Trans. Database Syst., 35(4):26, 2010.
[BR99] Kevin S. Beyer and Raghu Ramakrishnan. Bottom-up computation of sparse and iceberg cubes.
In SIGMOD 1999, Proceedings ACM SIGMOD International Conference on Management of
Data, June 1-3, 1999, Philadelphia, Pennsylvania, USA., pages 359–370, 1999.
[CCFC04] Moses Charikar, Kevin Chen, and Martin Farach-Colton. Finding frequent items in data streams.
Theoretical Computer Science, 312(1):3–15, 2004.
[CFM09] Yousra Chabchoub, Christine Fricker, and Hanene Mohamed. Analysis of a bloom filter algorithm via the supermarket model. In 21st International Teletraffic Congress, ITC 2009, Paris,
France, September 15-17, 2009, pages 1–8, 2009.
[CH08] Graham Cormode and Marios Hadjieleftheriou. Finding frequent items in data streams. Proceedings of the VLDB Endowment, 1(2):1530–1541, 2008.
[CKMS08] Graham Cormode, Flip Korn, S. Muthukrishnan, and Divesh Srivastava. Finding hierarchical
heavy hitters in streaming data. TKDD, 1(4), 2008.
[CM05] Graham Cormode and S Muthukrishnan. An improved data stream summary: the count-min
sketch and its applications. Journal of Algorithms, 55(1):58–75, 2005.
21
[Cor12] Graham Cormode. Sketch techniques for massive data. In Graham Cormode, Minos Garofalakis,
Peter J. Haas, and Chris Jermaine, editors, Synopses for Massive Data: Samples, Histograms,
Wavelets, Sketches, volume 4 of Foundations and Trends in Databases, pages 1–294. Now Publishers Inc., Hanover, MA, USA, January 2012.
[CP06] Ben Carterette and Desislava Petkova. Learning a ranking from pairwise preferences. In Proceedings of the 29th annual international ACM SIGIR conference on Research and development
in information retrieval, pages 629–630. ACM, 2006.
[CP11] Ioannis Caragiannis and Ariel D. Procaccia. Voting almost maximizes social welfare despite
limited communication. Artificial Intelligence, 175(910):1655 – 1671, 2011.
[CRT06] E. J. Candes, J. Romberg, and T. Tao. Stable signal recovery from incomplete and inaccurate
measurements. Communications on Pure and Applied Mathematics, 59:1207–1223, 2006.
[CS05] Vincent Conitzer and Tuomas Sandholm. Communication complexity of common voting rules.
In Proceedings of the 6th ACM conference on Electronic commerce, pages 78–87. ACM, 2005.
[CT12] Thomas M Cover and Joy A Thomas. Elements of information theory. John Wiley & Sons,
2012.
[DB15] Palash Dey and Arnab Bhattacharyya. Sample complexity for winner prediction in elections.
In Proceeding of the 14th International Conference on Autonomous Systems and Multiagent
Systems (AAMAS-15), 2015.
[DHKP97] Martin Dietzfelbinger, Torben Hagerup, Jyrki Katajainen, and Martti Penttonen. A reliable
randomized algorithm for the closest-pair problem. Journal of Algorithms, 25(1):19–51, 1997.
[DKW56] A Dvoretzky, J Kiefer, and J Wolfowitz. Asymptotic minimax character of the sample distribution function and of the classical multinomial estimator. Annals of Mathematical Statistics,
27(3):642 – 669, 1956.
[DLOM02] Erik D Demaine, Alejandro López-Ortiz, and J Ian Munro. Frequency estimation of internet
packet streams with limited space. In AlgorithmsESA 2002, pages 348–360. Springer, 2002.
[EV03] Cristian Estan and George Varghese. New directions in traffic measurement and accounting:
Focusing on the elephants, ignoring the mice. ACM Trans. Comput. Syst., 21(3):270–313, 2003.
[Fla85] Philippe Flajolet. Approximate counting: a detailed analysis. BIT Numerical Mathematics,
25(1):113–134, 1985.
[FSG+ 98] Min Fang, Narayanan Shivakumar, Hector Garcia-Molina, Rajeev Motwani, and Jeffrey D.
Ullman. Computing iceberg queries efficiently. In VLDB’98, Proceedings of 24rd International
Conference on Very Large Data Bases, August 24-27, 1998, New York City, New York, USA,
pages 299–310, 1998.
[GSTV07] Anna C. Gilbert, Martin J. Strauss, Joel A. Tropp, and Roman Vershynin. One sketch for all:
fast algorithms for compressed sensing. In Proceedings of the 39th Annual ACM Symposium on
Theory of Computing, San Diego, California, USA, June 11-13, 2007, pages 237–246, 2007.
[HHS94] Tin Kam Ho, Jonathan J Hull, and Sargur N Srihari. Decision combination in multiple classifier
systems. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 16(1):66–75, 1994.
[Hid99] Christian Hidber. Online association rule mining. In SIGMOD 1999, Proceedings ACM SIGMOD
International Conference on Management of Data, June 1-3, 1999, Philadelphia, Pennsylvania,
USA., pages 145–156, 1999.
22
[HKTR04] Jonathan L Herlocker, Joseph A Konstan, Loren G Terveen, and John T Riedl. Evaluating collaborative filtering recommender systems. ACM Transactions on Information Systems (TOIS),
22(1):5–53, 2004.
[HPDW01] Jiawei Han, Jian Pei, Guozhu Dong, and Ke Wang. Efficient computation of iceberg cubes
with complex measures. In Proceedings of the 2001 ACM SIGMOD international conference on
Management of data, Santa Barbara, CA, USA, May 21-24, 2001, pages 1–12, 2001.
[HPY00] Jiawei Han, Jian Pei, and Yiwen Yin. Mining frequent patterns without candidate generation.
In Proceedings of the 2000 ACM SIGMOD International Conference on Management of Data,
May 16-18, 2000, Dallas, Texas, USA., pages 1–12, 2000.
[HSST05] John Hershberger, Nisheeth Shrivastava, Subhash Suri, and Csaba D. Tóth. Space complexity
of hierarchical heavy hitters in multi-dimensional data streams. In Proceedings of the Twentyfourth ACM SIGACT-SIGMOD-SIGART Symposium on Principles of Database Systems, June
13-15, 2005, Baltimore, Maryland, USA, pages 338–347, 2005.
[JMP+ 14] Albert Jiang, Leandro Soriano Marcolino, Ariel D Procaccia, Tuomas Sandholm, Nisarg Shah,
and Milind Tambe. Diverse randomized agents vote to win. In Advances in Neural Information
Processing Systems, pages 2573–2581, 2014.
[KN97] Eyal Kushilevitz and Noam Nisan. Communication Complexity. Cambridge University Press,
New York, NY, USA, 1997.
[KNR99] Ilan Kremer, Noam Nisan, and Dana Ron. On randomized one-round communication complexity.
Computational Complexity, 8(1):21–49, 1999.
[KSP03] Richard M Karp, Scott Shenker, and Christos H Papadimitriou. A simple algorithm for finding
frequent elements in streams and bags. ACM Transactions on Database Systems (TODS),
28(1):51–55, 2003.
[KTW11] Peter Kellner, Joe Twyman, and Anthony Wells. Polling voting intentions. In Political Communication in Britain, pages 94–108. Springer, 2011.
[KX06] Abhishek Kumar and Jun (Jim) Xu. Sketch guided sampling - using on-line estimates of flow
size for adaptive data collection. In INFOCOM 2006. 25th IEEE International Conference
on Computer Communications, Joint Conference of the IEEE Computer and Communications
Societies, 23-29 April 2006, Barcelona, Catalunya, Spain, 2006.
[Li14] Hang Li. Learning to rank for information retrieval and natural language processing. Synthesis
Lectures on Human Language Technologies, 7(3):1–121, 2014.
[LN06] Alessandra Lumini and Loris Nanni. Detector of image orientation based on borda count.
Pattern Recognition Letters, 27(3):180–186, 2006.
[LNNT16] Kasper Green Larsen,
Jelani Nelson,
Huy Le Nguyen,
and Mikkel Thorup.
Optimal space heavy hitters with fast update and query.
Abstract at
http://ita.ucsd.edu/workshop/16/files/abstract/abstract_3738.txt, 2016.
[LRSC01] Charles E Leiserson, Ronald L Rivest, Clifford Stein, and Thomas H Cormen. Introduction to
algorithms. The MIT press, 5:2–2, 2001.
[MAE05] Ahmed Metwally, Divyakant Agrawal, and Amr El Abbadi. Efficient computation of frequent
and top-k elements in data streams. In Proceedings of the 10th International Conference on
Database Theory, ICDT’05, pages 398–412, Berlin, Heidelberg, 2005. Springer-Verlag.
23
[MBG04] Amélie Marian, Nicolas Bruno, and Luis Gravano. Evaluating top-k queries over web-accessible
databases. ACM Trans. Database Syst., 29(2):319–362, 2004.
[MG82] Jayadev Misra and David Gries. Finding repeated elements. Sci. Comput. Program., 2(2):143–
152, 1982.
[MG09] Md Maruf Monwar and Marina L Gavrilova. Multimodal biometric system using rank-level
fusion approach. Systems, Man, and Cybernetics, Part B: Cybernetics, IEEE Transactions on,
39(4):867–878, 2009.
[MM02] Gurmeet Singh Manku and Rajeev Motwani. Approximate frequency counts over data streams.
In Proceedings of the 28th international conference on Very Large Data Bases, pages 346–357.
VLDB Endowment, 2002.
[MNSW98] Peter Bro Miltersen, Noam Nisan, Shmuel Safra, and Avi Wigderson. On data structures and
asymmetric communication complexity. J. Comput. Syst. Sci., 57(1):37–49, 1998.
[Moo81] J. Strother Moore. Journal of Algorithms, June 1981. p. 208–209.
[Mor78] Robert Morris. Counting large numbers of events in small registers. Communications of the
ACM, 21(10):840–842, 1978.
[MPC12] Andrew Mao, Ariel D. Procaccia, and Yiling Chen. Social Choice for Human Computation. In
Proceedings of the Fourth Workshop on Human Computation (HCOMP-12), 2012.
[MPC13] Andrew Mao, Ariel D. Procaccia, and Yiling Chen. Better Human Computation Through
Principled Voting. In Proceedings of of the 27th Conference on Artificial Intelligence (AAAI’13),
2013.
[MR91] D.J. Mullen and B. Roth. Decision making: Its logic and practice. Savage, MD: Rowman and
Littlefield Publishers, Inc., 1991.
[Mut05] Shanmugavelayutham Muthukrishnan. Data streams: Algorithms and applications. Now Publishers Inc, 2005.
[MYCC07] Nikos Mamoulis, Man Lung Yiu, Kit Hung Cheng, and David W. Cheung. Efficient top-k
aggregation of ranked inputs. ACM Trans. Database Syst., 32(3):19, 2007.
[NC06] Rabia Nuray and Fazli Can. Automatic ranking of information retrieval systems using data
fusion. Information Processing & Management, 42(3):595–614, 2006.
[Nel12] Jelani Nelson. Sketching and streaming algorithms for processing massive data. XRDS: Crossroads, The ACM Magazine for Students, 19(1):14–19, 2012.
[PPR15] Adarsh Prasad, Harsh Pareek, and Pradeep Ravikumar. Distributional rank aggregation, and an
axiomatic analysis. In Proceedings of the 32nd International Conference on Machine Learning
(ICML-15), pages 2104–2112, 2015.
[RV97] Paul Resnick and Hal R Varian. Recommender systems. Communications of the ACM, 40(3):56–
58, 1997.
[SBAS04] Nisheeth Shrivastava, Chiranjeeb Buragohain, Divyakant Agrawal, and Subhash Suri. Medians
and beyond: new aggregation techniques for sensor networks. In Proceedings of the 2nd International Conference on Embedded Networked Sensor Systems, SenSys 2004, Baltimore, MD,
USA, November 3-5, 2004, pages 239–249, 2004.
[Smi88] D. Smirnov. Shannon’s information methods for lower bounds for probabilistic communication
complexity. Master’s thesis, Moscow University, 1988.
24
[SON95] Ashok Savasere, Edward Omiecinski, and Shamkant B. Navathe. An efficient algorithm for
mining association rules in large databases. In VLDB’95, Proceedings of 21th International
Conference on Very Large Data Bases, September 11-15, 1995, Zurich, Switzerland., pages
432–444, 1995.
[SW15] Xiaoming Sun and David P. Woodruff. Tight bounds for graph problems in insertion streams. In
Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques,
APPROX/RANDOM 2015, August 24-26, 2015, Princeton, NJ, USA, pages 435–448, 2015.
[Toi96] Hannu Toivonen. Sampling large databases for association rules. In VLDB’96, Proceedings
of 22th International Conference on Very Large Data Bases, September 3-6, 1996, Mumbai
(Bombay), India, pages 134–145, 1996.
[VWWZ15] Dirk Van Gucht, Ryan Williams, David P Woodruff, and Qin Zhang. The communication
complexity of distributed set-joins with applications to matrix multiplication. In Proceedings of
the 34th ACM Symposium on Principles of Database Systems, pages 199–212. ACM, 2015.
[VZ14] Maksims N Volkovs and Richard S Zemel. New learning methods for supervised and unsupervised preference aggregation. The Journal of Machine Learning Research, 15(1):1135–1176,
2014.
[WL04] Gang Wang and Frederick H Lochovsky. Feature selection with conditional mutual information
maximin in text categorization. In Proceedings of the thirteenth ACM international conference
on Information and knowledge management, pages 342–349. ACM, 2004.
[Xia12] Lirong Xia. Computing the margin of victory for various voting rules. In Proceedings of the
13th ACM Conference on Electronic Commerce, pages 982–999. ACM, 2012.
[Yao79] Andrew Chi-Chih Yao. Some complexity questions related to distributive computing (preliminary report). In Proceedings of the eleventh annual ACM symposium on Theory of computing,
pages 209–213. ACM, 1979.
25
Appendix A
Information Theory Facts
For a discrete P
random variable X with possible values {x1 , x2 , . . . , xn }, the Shannon entropy of X is defined
as H(X) = − ni=1 Pr(X = xi ) log2 Pr(X = xi ). Let Hb (p) = −p log2 p−(1−p) log2 (1−p) denote the binary
entropy function when p ∈ (0, 1). For two random variables X and Y with possible values {x1 , x2 ,P
. . . , xn } and
{y1 , y2 , . . . , ym }, respectively, the conditional entropy of X given Y is defined as H(X | Y ) = i,j Pr(X =
Pr(Y =y )
xi , Y = yj ) log2 Pr(X=xi ,Y j=yj ) . Let I(X; Y ) = H(X) − H(X | Y ) = H(Y ) − H(Y | X) denote the mutual
information between two random variables X, Y . Let I(X; Y | Z) denote the mutual information between
two random variables X, Y conditioned on Z, i.e., I(X; Y | Z) = H(X | Z) − H(X | Y, Z). The following
summarizes several basic properties of entropy and mutual information.
Proposition 1. Let X, Y, Z, W be random variables.
1. If X takes value in {1, 2, . . . , m}, then H(X) ∈ [0, log m].
2. H(X) > H(X | Y ) and I(X; Y ) = H(X) − H(X | Y ) > 0.
3. If X and Z are independent, then we have I(X; Y | Z) > I(X; Y ). Similarly, if X, Z are independent
given W , then I(X; Y | Z, W ) > I(X; Y | W ).
4. (Chain rule of mutual information) I(X, Y ; Z) = I(X; Z) +
PnI(Y ; Z | X). More generally, for
any random variables X1 , X2 , . . . , Xn , Y , I(X1 , . . . , Xn ; Y ) =
i=1 I(Xi ; Y | X1 , . . . , Xi−1 ). Thus,
I(X, Y ; Z | W ) > I(X; Z | W ).
5. (Fano’s inequality) Let X be a random variable chosen from domain X according to distribution µX ,
and Y be a random variable chosen from domain Y according to distribution µY . For any reconstruction
function g : Y → X with error δg ,
Hb (δg ) + δg log(|X | − 1) > H(X | Y ).
We refer readers to [CT12] for a nice introduction to information theory.
Appendix B
We remark that the algorithm in Lemma 1 has optimal space complexity as shown in Proposition 2 below
which may be of independent interest.
Proposition 2. [⋆] Any algorithm that chooses an item from a set of size n with probability p for 0 < p 6
in unit cost RAM model must use Ω(log log m) bits of memory.
1
n,
Proof. The algorithm generates t bits uniformly at random (the number of bits it generates uniformly at
random may also depend on the outcome of the previous random bits) and finally picks an item from the
say x. Consider a run R of the algorithm where it chooses the item x with smallest number of random bits
getting generated; say it generates t random bits in this run R. This means that in any other run of the
algorithm where the item x is chosen, the algorithm must generate at least t many random bits. Let the
random bits generated in R be r1 , · · · , rt . Let si be the memory content of the algorithm immediately after
it generates ith random bit, for i ∈ [t], in the run R. First notice that if t < log2 n, then the probability
with which the item x is chosen is more than n1 , which would be a contradiction. Hence, t > log2 n. Now
we claim that all the si ’s must be different. Indeed otherwise, let us assume si = sj for some i < j. Then
the algorithm chooses the item x after generating t − (j − i) many random bits (which is strictly less than
t) when the random bits being generated are r1 , · · · , ri , rj+1 , · · · , rt . This contradicts the assumption that
the run R we started with chooses the item x with smallest number of random bits generated.
26
| 8cs.DS
|
Beyond Log-concavity: Provable Guarantees for Sampling
Multi-modal Distributions using Simulated Tempering Langevin
Monte Carlo
arXiv:1710.02736v2 [cs.LG] 6 Nov 2017
Rong Ge*, Holden Lee†, Andrej Risteski‡
November 7, 2017
Abstract
In the last several years, provable guarantees for iterative optimization algorithms like gradient descent and expectation-maximization in non-convex settings have become a topic of intense
research in the machine learning community. These works have shed light on the practical success of these algorithms in many unsupervised learning settings such as matrix completion,
sparse coding, and learning latent variable Bayesian models.
Another elementary task at inference-time in Bayesian settings, besides model learning, is
sampling from distributions that are only specified up to a partition function (i.e., constant
of proportionality). As a concrete example, in latent-variable models, sampling the posterior
on the latent variables is how a model is used after it has been learned. Similar worst-case
theoretical issues plague this task as do the learning one: without any assumptions, sampling
(even approximately) can be #P-hard. However, few works have provided “beyond worst-case”
guarantees for such settings.
The analogue of “convexity” for inference is “log-concavity”: for log-concave distributions,
classical results going back to [BÉ85] show that natural continuous-time Markov chains called
Langevin diffusions mix in polynomial time. The most salient feature of log-concavity violated
in practice is uni-modality: commonly, the distributions we wish to sample from are multimodal. In the presence of multiple deep and well-separated modes, Langevin diffusion suffers
from torpid mixing.
We address this problem by combining Langevin diffusion with simulated tempering. The
result is a Markov chain that mixes more rapidly by transitioning between different temperatures
of the distribution. We analyze this Markov chain for the canonical multi-modal distribution:
a mixture of gaussians (of equal variance). The algorithm based on our Markov chain provably
samples from distributions that are close to mixtures of gaussians, given access to the gradient
of the log-pdf. For the analysis, we use a spectral decomposition theorem for graphs [GT14]
and a Markov chain decomposition technique [MR02].
*
Duke University, Computer Science Department [email protected]
Princeton University, Mathematics Department [email protected]
‡
Massachusetts Institute of Technology, Applied Mathematics and IDSS [email protected]
†
1
Contents
1 Introduction
1.1 Our results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Prior work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
4
5
2 Preliminaries
2.1 Markov chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Overview of Langevin dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Overview of simulated tempering . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
6
8
8
3 Our Algorithm
9
4 Overview of proof
4.1 Decomposing the simulated tempering chain
4.2 Existence of partitions . . . . . . . . . . . .
4.3 Mixing at highest temperature . . . . . . .
4.4 Discretizing the Langevin diffusion . . . . .
4.5 Estimating the partition functions . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5 Spectral gap of simulated tempering
11
12
13
13
13
14
14
6 Defining the partitions
6.1 Proving the eigenvalue gap . . . . . . . . . . . . . . .
6.2 Scaled temperature approximates mixture of gaussians
6.3 Poincaré inequalities on small subsets . . . . . . . . .
6.4 Existence of partition . . . . . . . . . . . . . . . . . .
6.5 Making arbitrarily fine partitions . . . . . . . . . . . .
6.6 Restriction to large compact set . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
16
17
18
19
23
24
28
7 Mixing at the highest temperature
29
8 Discretizing the continuous chains
31
9 Proof of main theorem
35
10 Acknowledgements
40
A Background on Markov chains
A.1 Discrete time Markov chains . . . . . . .
A.2 Restricted and projected Markov chains
A.3 Conductance and clustering . . . . . . .
A.4 Continuous time Markov processes . . .
B Examples
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
42
42
44
44
45
46
2
C Pertubation tolerance
C.1 Mixing time of the tempering chain .
C.2 Mixing time at highest temperature
C.3 Discretization . . . . . . . . . . . . .
C.4 Putting things together . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
D Another lower bound for simulated tempering
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
47
48
49
49
50
51
Introduction
In recent years, one of the most fruitful directions of research has been providing theoretical guarantees for optimization in non-convex settings. In particular, a routine task in both unsupervised and
supervised learning is to use training data to fit the optimal parameters for a model in some parametric family. Theoretical successes in this context range from analyzing tensor-based approaches
using method-of-moments, to iterative techniques like gradient descent, EM, and variational inference in a variety of models. These models include topic models [Ana+12; AGM12; Aro+13; AR15],
dictionary learning [Aro+15; Aga+14], gaussian mixture models [HK13], and Bayesian networks
[Aro+17].
Finding maximum likelihood values of unobserved quantities via optimization is reasonable in
many learning settings, as when the number of samples is large maximum likelihood will converge to
the true values of the quantities. However, for Bayesian inference problems (e.g. given a document,
what topics is it about) the number of samples can be limited and maximum likelihood may not be
well-behaved [SR11]. In these cases we would prefer to sample from the posterior distribution. In
more generality, the above (typical) scenario is sampling from the posterior distribution over the
latent variables of a latent variable Bayesian model whose parameters are known. In such models,
the observable variables 𝑥 follow a distribution 𝑝(𝑥) which has a simple and succinct form given the
values of some latent variables ℎ, i.e., the joint 𝑝(ℎ, 𝑥) factorizes as 𝑝(ℎ)𝑝(𝑥|ℎ) where both factors
are explicit. Hence, the posterior distribution 𝑝(ℎ|𝑥) has the form 𝑝(ℎ|𝑥) = 𝑝(ℎ)𝑝(𝑥|ℎ)
. Even though
𝑝(𝑥)
the numerator is easy to evaluate, without structural assumptions such distributions are often
hard to sample from (exactly or approximately). The difficulty is in evaluating the denominator
P
𝑝(𝑥) = ℎ 𝑝(ℎ)𝑝(𝑥|ℎ), which can be NP-hard to do even approximately even for simple models like
topic models [SR11].
The sampling analogues of convex functions, which are arguably the widest class of real-valued
functions for which optimization is easy, are log-concave distributions, i.e. distributions of the form
𝑝(𝑥) ∝ 𝑒−𝑓 (𝑥) for a convex function 𝑓 (𝑥). Recently, there has been renewed interest in analyzing
a popular Markov Chain for sampling from such distributions, when given gradient access to 𝑓 —a
natural setup for the posterior sampling task described above. In particular, a Markov chain called
Langevin Monte Carlo (see Section 2.2), popular with Bayesian practitioners, has been proven to
work, with various rates depending on the properties of 𝑓 [Dal16; DM16; Dal17].
Log-concave distributions are necessarily uni-modal: their density functions have only one local
maximum, which must then be a global maximum. This fails to capture many interesting scenarios. Many simple posterior distributions are neither log-concave nor uni-modal, for instance, the
posterior distribution of the means for a mixture of gaussians. In a more practical direction, complicated posterior distributions associated with deep generative models [RMW14] and variational
auto-encoders [KW13] are believed to be multimodal as well.
3
Figure 1: Mixture of two gaussians. Left: the two gaussians are well-separated, local sampling
algorithm cannot move between modes. Right: Same distribution at high temperature, it is now
possible to move between modes.
The goal of this work is to initiate an exploration of provable methods for sampling “beyond
log-concavity,” in parallel to optimization “beyond convexity”. As worst-case results are prohibited
by hardness results, we must again make assumptions on the distributions we will be interested in.
As a first step, in this paper we consider the prototypical multimodal distribution, a mixture of
gaussians.
1.1
Our results
We formalize the problem of interest as follows. We wish to sample from a distribution 𝑝 : R𝑑 → R,
such that 𝑝(𝑥) ∝ 𝑒−𝑓 (𝑥) , and we are allowed to query ∇𝑓 (𝑥) and 𝑓 (𝑥) at any point 𝑥 ∈ R𝑑 .
To start with, we focus on a problem where 𝑒−𝑓 (𝑥) is the density function of a mixture of gausP
sians. That is, given centers 𝜇1 , 𝜇2 , . . . , 𝜇𝑛 ∈ R𝑑 , weights 𝑤1 , 𝑤2 , . . . , 𝑤𝑛 ( 𝑛𝑖=1 𝑤𝑖 = 1), variance 𝜎 2
(all the gaussians are spherical with same covariance matrix 𝜎 2 𝐼), the function 𝑓 (𝑥) is defined as1
𝑛
X
‖𝑥 − 𝜇𝑖 ‖2
𝑤𝑖 exp −
𝑓 (𝑥) = − log
2𝜎 2
𝑖=1
Ç
å!
.
(1)
Furthermore, suppose that 𝐷 is such that ‖𝜇𝑖 ‖ ≤ 𝐷, ∀𝑖 ∈ [𝑛]. We show that there is an efficient
algorithm that can sample from this distribution given just access to 𝑓 (𝑥) and ∇𝑓 (𝑥).
Theorem 1.1 (main,
𝑓 (𝑥) as defined in Equation (1), there is an algorithm with
Ä informal). Given
ä
1 1
running time poly 𝑤min , 𝐷, 𝑑, 𝜀 , 𝜎2 that outputs a sample from a distribution within TV-distance
𝜀 of 𝑝(𝑥).
Note that because the algorithm does not have direct access to 𝜇1 , 𝜇2 , . . . , 𝜇𝑛 , even sampling
from this mixture of gaussians distribution is very non-trivial. Sampling algorithms that are based
on making local steps (such as the ball-walk [LS93; Vem05] and Langevin Monte Carlo) cannot
move between different components of the gaussian mixture when the gaussians are well-separated
(see Figure 1 left). In the algorithm we use simulated tempering (see Section 2.3), which is a
technique that considers the distribution at different temperatures (see Figure 1 right) in order to
move between different components.
In Appendix B, we give a few examples to show some simple heuristics cannot work and the
assumption that all gaussians have the same covariance cannot be removed. In particular, we show
1
Note that the expression inside the log is essentially the probability density of a mixture of gaussians, except the
normalization factor is missing. However, the normalization factor can just introduce a constant shift of 𝑓 and does
not really change ∇𝑓 .
4
random initialization is not enough to find all the modes. We also give an example where for a
mixture of two gaussians, even if the covariance only differs by a constant multiplicative factor,
simulated tempering is known to take exponential time.
Of course, requiring the distribution to be exactly a mixture of gaussians is a very strong assumption. Our results can be generalized to all functions that are “close” to a mixture of gaussians.
More precisely, the function 𝑓 satisfies the following properties:
∃𝑓˜ :
R𝑑 → R where
𝑓˜ − 𝑓
∞
≤Δ,
∇𝑓˜ − ∇𝑓
∞
≤ 𝜏 and ∇2 𝑓˜(𝑥) ⪯ ∇2 𝑓 (𝑥) + 𝜏 𝐼, ∀𝑥 ∈ R𝑑
(2)
and 𝑓˜(𝑥) = − log
𝑛
X
𝑖=1
Ç
𝑤𝑖 exp −
‖𝑥 − 𝜇𝑖 ‖
2𝜎 2
å!
2
(3)
Intuitively, these conditions show that the density of the distribution is within a 𝑒Δ multiplicative factor to an (unknown) mixture of gaussians. Our theorem can be generalized to this
case.
Theorem 1.2 (general case, informal). ForÄ function 𝑓 (𝑥) that äsatisfies Equations (2) and (3),
there is an algorithm that runs in time poly 𝑤min , 𝐷, 𝑑, 1𝜀 , 𝜎12 , 𝑒Δ that outputs a sample 𝑥 from a
distribution that has TV-distance at most 𝜀 from 𝑝(𝑥).
1.2
Prior work
Our algorithm will use two classical techniques in the theory of Markov chains: Langevin diffusion,
a chain for sampling from distributions in the form 𝑝(𝑥) ∝ 𝑒−𝑓 (𝑥) given only gradient access to
𝑓 and simulated tempering, a heuristic technique used for tackling multimodal distributions. We
recall briefly what is known for both of these techniques.
For Langevin dynamics, convergence to the stationary distribution is a classic result [Bha78].
Understanding the mixing time of the continuous dynamics for log-concave distributions is also
a classic result: [BÉ85; Bak+08] show that log-concave distributions satisfy a Poincaré and logSobolev inequality, which characterize the rate of convergence. Of course, algorithmically, one can
only run a “discretized” version of the Langevin dynamics, but results on such approaches are
much more recent: [Dal16; DM16; Dal17] obtained an algorithm for sampling from a log-concave
distribution over R𝑑 , and [BEL15] gave a algorithm to sample from a log-concave distribution
restricted to a convex set by incorporating a projection step. [RRT17] give a nonasymptotic analysis
of Langevin dynamics for arbitrary non-log-concave distributions with certain regularity and decay
properties. Of course, the mixing time is exponential in general when the spectral gap of the chain
is small; furthermore, it has long been known that transitioning between different modes can take
an exponentially long time, a phenomenon known as meta-stability [Bov+02; Bov+04; BGK05]. It
is a folklore result that guarantees for mixing extend to distributions 𝑒−𝑓 (𝑥) where 𝑓 (𝑥) is a “nice”
function that is close to a convex function in 𝐿∞ distance; however, this does not address more
global deviations from convexity.
It is clear that for distributions that are far from being log-concave and many deep modes,
additional techniques will be necessary. Among many proposed heuristics for such situations is
simulated tempering, which effectively runs multiple Markov chains, each corresponding to a different temperature of the original chain, and “mixes” between these different Markov chains. The
5
intuition is that the Markov chains at higher temperature can move between modes more easily,
and if one can “mix in” points from these into the lower temperature chains, their mixing time
ought to improve as well. Provable results of this heuristic are however few and far between.
[WSH09; Zhe03] lower-bound the spectral gap for generic simulated tempering chains. The crucial
technique our paper shares with theirs is a Markov chain decomposition technique due to [MR02].
However, for the scenario of Section 1.1 we are interested in, the spectral gap bound in [WSH09] is
exponentially small as a function of the number of modes. Our result will remedy this.
2
Preliminaries
In this section we first introduce notations for Markov chains. More details are deferred to Appendix A. Then we briefly discuss Langevin Monte Carlo and Simulated Tempering.
2.1
Markov chains
In this paper, we use both discrete time and continuous time Markov chains. In this section we
briefly give definitions and notations for discrete time Markov chains. Continuous time Markov
chains follow the same intuition, but we defer the formal definitions to Appendix A.
Definition 2.1. A (discrete time) Markov chain is 𝑀 = (Ω, 𝑃 ), where Ω is a measure space and
𝑃 (𝑥, 𝑦) 𝑑𝑦 is a probability measure for each 𝑥. It defines a random process (𝑋𝑡 )𝑡∈N0 as follows. If
𝑋𝑠 = 𝑥, then
P(𝑋𝑠+1 ∈ 𝐴) = 𝑃 (𝑥, 𝐴) : =
R
Z
𝑝(𝑥, 𝑦) 𝑑𝑦.
(4)
𝐴
A stationary distribution is 𝑝(𝑥) such that if 𝑋0 ∼ 𝑝, then 𝑋𝑡 ∼ 𝑝 for all 𝑡; equivalently,
Ω 𝑝(𝑥)𝑃 (𝑥, 𝑦) 𝑑𝑥 = 𝑝(𝑦).
A chain is reversible if 𝑝(𝑥)𝑃 (𝑥, 𝑦) = 𝑝(𝑦)𝑃 (𝑦, 𝑥).
If a Markov chain has a finite number of states, then it can be represented as a weighted graph
where the transition probabilities are proportional to the weights on the edges. A reversible Markov
chain can be represented as a undirected graph.
Variance, Dirichlet form and Spectral Gap An important quantity of the Markov chain is
the spectral gap.
Definition 2.2. For a discrete-time Markov chain 𝑀 = (Ω, 𝑃 ), let 𝑃 operate on functions as
Z
(𝑃 𝑔)(𝑥) = E𝑦∼𝑃 (𝑥,·) 𝑔(𝑦) =
𝑔(𝑥)𝑃 (𝑥, 𝑦) 𝑑𝑦.
Suppose 𝑀 = (Ω, 𝑃 ) has unique stationary distribution 𝑝. Let ⟨𝑔, ℎ⟩𝑝 :=
define the Dirichlet form and variance by
ℰ𝑀 (𝑔, ℎ) = ⟨𝑔, (𝐼 − 𝑃 )ℎ⟩𝑝
Var𝑝 (𝑔) = 𝑔 −
R
Ω 𝑔(𝑥)ℎ(𝑥)𝑝(𝑥) 𝑑𝑥
and
(6)
2
Z
𝑔𝑝 𝑑𝑥
Ω
6
(5)
Ω
(7)
𝑝
Write ℰ𝑀 (𝑔) for ℰ𝑀 (𝑔, 𝑔). Define the eigenvalues of 𝑀 , 0 = 𝜆1 ≤ 𝜆2 ≤ · · · to be the eigenvalues
of 𝐼 − 𝑃 with respect to the norm ‖·‖𝑝 .
Define the spectral gap by
Gap(𝑀 ) =
inf
𝑔∈𝐿2 (𝑝)
ℰ𝑀 (𝑔)
.
Var𝑝 (𝑔)
(8)
In the case of a finite, undirected graph, the function just corresponds to a vector 𝑥 ⊥ ⃗1. The
Dirichlet form corresponds to 𝑥⊤ ℒ𝑥 where ℒ is the normalized Laplacian matrix, and the variance
is just the squared norm ‖𝑥‖2 .
The spectral gap controls mixing for the Markov chain. Define the 𝜒2 distance between 𝑝, 𝑞 by
Z Ç
𝜒2 (𝑝||𝑞) =
Ω
𝑞(𝑥) − 𝑝(𝑥)
𝑝(𝑥)
å2
Z Ç
𝑝(𝑥) 𝑑𝑥 =
Ω
𝑞(𝑥)2
𝑝(𝑥)
å
− 1.
(9)
Let 𝑝0 be any initial distribution and 𝑝𝑡 be the distribution after running the Markov chain for 𝑡
steps. Then
𝜒2 (𝑝||𝑝𝑡 ) ≤ (1 − 𝐺′ )𝑡 𝜒(𝑝||𝑝0 )
(10)
where 𝐺′ = min(𝜆2 , 2 − 𝜆max ).
Restrictions and Projections Later we will also work with continuous time Markov chains
(such as Langevin dynamics, see Section 2.2). In the proof we will also need to consider restrictions
and projections of Markov chains. Intuitively, restricting a Markov chain 𝑀 to a subset of states
𝐴 (which we denote by 𝑀 |𝐴) removes all the states out of 𝐴, and replaces transitions to 𝐴 with
¯ 𝒫 ) “merges” all
self-loops. Projecting a Markov chain 𝑀 to partition 𝒫 (which we denote by 𝑀
parts of the partition into individual states. For formal definitions see Appendix A.
Conductance and clustering Finally we define conductance and clusters for Markov chains.
These are the same as the familiar concepts as in undirected graphs.
Definition 2.3. Let 𝑀 = (Ω, 𝑃 ) be a Markov chain with unique stationary distribution 𝑝. Let
𝑄(𝑥, 𝑦) = 𝑝(𝑥)𝑃 (𝑥, 𝑦)
x
𝑄(𝐴, 𝐵) =
𝑄(𝑥, 𝑦) 𝑑𝑥 𝑑𝑦.
(11)
(12)
𝐴×𝐵
(I.e., 𝑥 is drawn from the stationary distribution and 𝑦 is the next state in the Markov chain.)
Define the (external) conductance of 𝑆, 𝜑𝑀 (𝑆), and the Cheeger constant of 𝑀 , Φ(𝑀 ), by
𝑄(𝑆, 𝑆 𝑐 )
𝑝(𝑆)
𝜑𝑀 (𝑆).
Φ(𝑀 ) =
min
𝜑𝑀 (𝑆) =
𝑆⊂Ω,𝑝(𝑆)≤ 12
(13)
(14)
The clustering of a Markov chain is analogous to a partition of vertices for undirected graphs.
For a good clustering, we require the inner-conductance to be large and the outer-conductance to
be small.
7
Definition 2.4. Let 𝑀 = (Ω, 𝑃 ) be a Markov chain on a finite state space Ω. We say that 𝑘
disjoint subsets 𝐴1 , . . . , 𝐴𝑘 of Ω are a (𝜑in , 𝜑out )-clustering if for all 1 ≤ 𝑖 ≤ 𝑘,
Φ(𝑀 |𝐴𝑖 ) ≥ 𝜑in
𝜑𝑀 (𝐴𝑖 ) ≤ 𝜑out .
2.2
(15)
(16)
Overview of Langevin dynamics
Langevin diffusion is a stochastic process, described by the stochastic differential equation (henceforth SDE)
√
𝑑𝑋𝑡 = −∇𝑓 (𝑋𝑡 ) 𝑑𝑡 + 2 𝑑𝑊𝑡
(17)
where 𝑊𝑡 is the Wiener process. The crucial (folklore) fact about Langevin dynamics is that
Langevin dynamics converges to the stationary distribution given by 𝑝(𝑥) ∝ 𝑒−𝑓 (𝑥) . Substituting
𝛽𝑓 for 𝑓 in (17) gives the Langevin diffusion process for inverse temperature 𝛽, which has stationary
distribution ∝ 𝑒−𝛽𝑓 (𝑥) . Equivalently it is also possible to consider the temperature as changing the
magnitude of the noise:
»
𝑑𝑋𝑡 = −∇𝑓 (𝑋𝑡 )𝑑𝑡 + 2𝛽 −1 𝑑𝑊𝑡 .
Of course algorithmically we cannot run a continuous-time process, so we run a discretized
version of the above process: namely, we run a Markov chain where the random variable at time 𝑡
is described as
p
𝑋𝑡+1 = 𝑋𝑡 − 𝜂∇𝑓 (𝑋𝑡 ) + 2𝜂𝜉𝑘 , 𝜉𝑘 ∼ 𝑁 (0, 𝐼)
(18)
√
where 𝜂 is the step size. (The reason for the 𝜂 scaling is that running Brownian motion for 𝜂 of
√
the time scales the variance by 𝜂.)
The works [Dal16; DM16; Dal17] have analyzed the convergence properties (both bias from the
stationary distribution, and the convergence rate) for log-concave distributions, while [RRT17] give
convergence rates for non-log-concave distributions. Of course, in the latter case, the rates depend
on the spectral gap, which is often exponential in the dimension.
2.3
Overview of simulated tempering
Simulated tempering is a technique that converts a Markov chain to a new Markov chain whose
state space is a product of the original state space and a temperature. The new Markov chain allows
the original chain to change “temperature” while maintaining the correct marginal distributions.
Given a discrete time Markov chain, we will consider it in 𝐿 temperatures. Let [𝐿] denote the set
{1, 2, ..., 𝐿}, we define the simulated tempering chain as follows:
Definition 2.5. Let 𝑀𝑖 , 𝑖 ∈ [𝐿] be a sequence of Markov chains with state space Ω and unique
stationary distributions 𝑝𝑖 . Let 𝑟1 , . . . , 𝑟𝐿 be such that
𝑟𝑖 > 0,
𝐿
X
𝑟𝑖 = 1.
𝑖=1
Define the simulated tempering Markov chain with relative probabilities 𝑟𝑖 as follows. The
states of 𝑀st are Ω × [𝐿]. Suppose the current state is (𝑥, 𝑘).
8
1. With probability
transition.
1
2,
keep 𝑘 fixed, and update 𝑥 according to 𝑀𝑘 . We will call this a Type 1
2. With probability 21 , do the following Metropolis-Hastings step: draw 𝑘 ′ randomly from {0, . . . , 𝐿−
1}. Then transition to (𝑥, 𝑘 ′ ) with probability
®
𝑟𝑘′ 𝑝𝑘′ (𝑥)
min
,1
𝑟𝑘 𝑝𝑘 (𝑥)
´
and stay at (𝑥, 𝑘) otherwise. We will call this a Type 2 transition.
Remark 2.6. For the type two transitions, we can instead just pick 𝑘 ′ from {𝑘 − 1, 𝑘, 𝑘 + 1}. This
𝑝 (𝑥)
will slightly improve our bounds on mixing time, because the ratio 𝑝𝑘𝑘′(𝑥) for 𝑘 ′ ∈ {𝑘 − 1, 𝑘 + 1} is
bounded, and can be exponential otherwise. For simplicity, we stick with the traditional definition
of the simulated tempering Markov chain.
The typical setting is as follows. The Markov chains come from a smooth family of Markov
chains with parameter 𝛽 ≥ 0, and 𝑀𝑖 is the Markov chain with parameter 𝛽𝑖 , where 0 ≤ 𝛽1 ≤
· · · 𝛽𝐿 = 1. (Using terminology from statistical physics, 𝛽 = 𝜏1 is the inverse temperature.) We are
interested in sampling from the distribution when 𝛽 is large (𝜏 is small). However, the chain suffers
from torpid mixing in this case, because the distribution is more peaked. The simulated tempering
chain uses smaller 𝛽 (larger 𝜏 ) to help with mixing. For us, the stationary distribution at inverse
temperature 𝛽 is ∝ 𝑒−𝛽𝑓 (𝑥) .
Of course, the Langevin dynamics introduced in previous section is a continuous time Markov
chain. In the algorithm we change it to a discrete time Markov chain by fixing a step size. Another
difficulty in running the simulated tempering chain directly is that we don’t have access to 𝑝𝑘
(because we do not know the partition function). We make use of the flexibility in 𝑟𝑖 ’s to fix this
issue. For more details see Section 3.
The crucial fact to note is that the stationary distribution is a “mixture” of the distributions
corresponding to the different temperatures. Namely:
Proposition 2.7 (folklore). If the 𝑀𝑘 are reversible Markov chains with stationary distributions
𝑝𝑘 , then the simulated tempering chain 𝑀 is a reversible Markov chain with stationary distribution
𝑝(𝑥, 𝑖) = 𝑟𝑖 𝑝𝑖 (𝑥).
3
Our Algorithm
Our algorithm will run a simulated tempering chain, with a polynomial number of temperatures,
while running discretized Langevin dynamics at the various temperatures. The full algorithm is
specified in Algorithm 2.
As we mentioned before, an obstacle in running the simulated tempering chain is that we do not
have access to the partition function. We solve this problem by estimating the partition function
from high temperature to low temperature, adding one temperature at a time (see Algorithm 2).
Note that if the simulated tempering chain mixes and produce good samples, by standard reductions
it is easy to estimate the (ratios of) partition functions.
Our main theorem is the following.
9
Algorithm 1 Simulated tempering Langevin Monte Carlo
“1 , . . . , 𝑍
“ℓ ; step size 𝜂, time interINPUT: Temperatures 𝛽1 , . . . , 𝛽ℓ ; partition function estimates 𝑍
val 𝑇 , number of steps 𝑡.
OUTPUT: A random sample 𝑥 ∈ R𝑑 (approximately from the distribution 𝑝ℓ (𝑥) ∝ 𝑒𝛽ℓ 𝑓 (𝑥) ).
Let (𝑥, 𝑘) = (𝑥0 , 1) where 𝑥0 ∼ 𝑁 (0, 𝛽1 𝐼).
for 𝑠 = 0 → 𝑡 − 1 do
√
(1) With probability 12 , keep 𝑘 fixed. Update 𝑥 according to 𝑥 ←[ 𝑥 − 𝜂𝛽𝑘 ∇𝑓 (𝑥) + 2𝜂𝜉𝑘 ,
𝜉𝑘 ∼ 𝑁 (0, 𝐼). Repeat this 𝑇𝜂 times.
1
(2) ßWith probability
2 , make a type 2 transition, where the acceptance ratio is
™
min
b𝑘′
𝑒−𝛽𝑘′ 𝑓 (𝑥) /𝑍
,1
b𝑘
𝑒−𝛽𝑘 𝑓 (𝑥) /𝑍
.
end for
If the final state is (𝑥, 𝑙) for some 𝑥 ∈ R𝑑 , return 𝑥. Otherwise, re-run the chain.
Algorithm 2 Main algorithm
INPUT: A function 𝑓 : R𝑑 , satisfying assumption (2), to which we have gradient access.
OUTPUT: A random sample 𝑥 ∈ R𝑑 .
Let 0 ≤ 𝛽1 < · · · < 𝛽𝐿 = 1 be a sequence of inverse temperatures satisfying (194) and (195).
“1 = 1.
Let 𝑍
for ℓ = 1 → 𝐿 do
Run the simulated tempering chain in Algorithm 1 with temperatures 𝛽1 , . . . , 𝛽ℓ , estimates
“1 , . . . , 𝑍
“𝑖 , step size 𝜂, time interval 𝑇 , and number of steps 𝑡 given by Lemma 9.4.
𝑍
If ℓ = 𝐿, return the sample.
Ä ä
Ä
ä
’
c 1 P𝑚 𝑒(−𝛽ℓ+1 +𝛽ℓ )𝑓 (𝑥𝑗 ) .
If ℓ < 𝐿, repeat to get 𝑚 = 𝑂(𝐿2 ln 1𝛿 ) samples, and let 𝑍
ℓ+1 = 𝑍ℓ 𝑚
𝑗=1
end for
10
P
2
‖𝑥−𝜇𝑖 ‖
𝑛
on R𝑑 where
Theorem 3.1 (Main theorem). Suppose 𝑓 (𝑥) = − ln
𝑖=1 𝑤𝑖 exp − 2𝜎 2
P𝑛
𝑖=1 𝑤𝑖 = 1, 𝑤min = min1≤𝑖≤𝑛 𝑤𝑖 > 0, and 𝐷 = max1≤𝑖≤𝑛 ‖𝜇𝑖 ‖. Then Algorithm 2 with parameters
given by Lemma
9.4 produces a sample from a distribution 𝑝′ with ‖𝑝 − 𝑝′ ‖1 ≤ 𝜀 in time
Ä
ä
poly 𝑤min , 𝐷, 𝑑, 𝜎1 , 1𝜀 .
For simplicity, we stated the theorem for distributions which are exactly mixtures of gaussians. The theorem is robust to 𝐿∞ perturbations as in (2), we give the more general theorem in
Appendix C.
4
Overview of proof
We will first briefly sketch the entire proof, and in the subsequent sections expand on all the
individual parts.
The key part of our proof is a new technique for bounding the spectral gap for simulated
tempering chain using decompositions (Section 4.1): for each temperature we make a partition
into “large” pieces that are well-connected from the inside. If this partition can be done at every
temperature, the difference in temperature is small enough, and the chain mixes at the highest
temperature, we show the simulated tempering chain also mixes quickly. This is a general theorem
for the mixing of simulated tempering chains that may be useful in other settings.
Figure 2: For a mixture of gaussians, we can partition space into regions where Langevin Monte
Carlo mixes well.
We then show that if 𝑓 (𝑥) is a mixture of gaussians, then indeed the partition exists (Section 4.2).
Here we use spectral clustering techniques developed by [GT14] for finite graphs. The main technical
difficulty is in transferring between the discrete and continuous cases.
Finally we complete the proof by showing that
1. the Markov chain mixes at the highest temperature (Section 4.3);
2. the discretized Markov chain approximates the continuous time Markov chain (Section 4.4);
11
3. the partition functions are estimated correctly which allows us to run the simulated tempering
chain (Section 4.5).
At last, in Appendix C we prove the arguments are tolerant to 𝐿∞ perturbations, i.e. that the
algorithm works for distributions that are not exactly mixtures of gaussians.
4.1
Decomposing the simulated tempering chain
First we show that if there exists a partition 𝒫𝑖 for each temperature such that
1. the Markov chain mixes rapidly within each set of the partition (i.e., Gap(𝑀𝑖 |𝐴 ) is large),
and
2. the sets in the partition are not too small,
and the chain mixes at the highest temperature, then the simulated tempering chain mixes rapidly.
Theorem (Theorem 5.2). Let 𝑀𝑖 = (Ω, 𝑃𝑖 ), 𝑖 ∈ [𝐿] be a sequence of Markov chains with state space
Ω and stationary distributions 𝑝𝑖 . Consider the simulated tempering chain 𝑀 = (Ω × [𝐿], 𝑃st ) with
min(𝑟𝑖 )
probabilities (𝑟𝑖 )𝐿
𝑖=1 . Let 𝑟 = max(𝑟𝑖 ) .
Let 𝒫𝑖 be a partition of the ground set Ω, for each 𝑖 ∈ [𝐿], with 𝒫1 = {Ω}.
Define the overlap parameter of (𝒫𝑖 )𝐿
𝑖=1 to be
𝛿((𝒫𝑖 )𝐿
𝑖=1 )
ïZ
=
ò
min
1<𝑖≤𝐿,𝐴∈𝒫
min{𝑝𝑖−1 (𝑥), 𝑝𝑖 (𝑥)} 𝑑𝑥 /𝑝𝑖 (𝐴).
𝐴
Define
𝑝min = min 𝑝𝑖 (𝐴).
𝑖,𝐴∈𝒫𝑖
Then the spectral gap of the tempering chain satisfies
Gap(𝑀st ) ≥
𝑟4 𝛿 2 𝑝2min
32𝐿4
min
1≤𝑖≤𝐿,𝐴∈𝒫𝑖
(Gap(𝑀𝑖 |𝐴 )).
(19)
To prove this, we use techniques similar to existing work on simulated tempering. More precisely,
similar to the proof in [WSH09], we will apply a “decomposition” theorem (Theorem A.5) for
analyzing the mixing time of the simulated tempering chain.
Note that as we are only using this theorem in the analysis, we only need an existential, not a
algorithmic result. In order to apply this theorem, we will show that there exist good partitions
𝒫𝑖 , such that the spectral gap Gap(𝑀𝑖 |𝐴 ) within each set is large, and each set in the partition has
size poly(𝑤min ).
Remark 4.1. In Appendix D, Theorem D.3, we also give a different (incomparable) criterion
for lower-bounding the spectral gap that improves the bound in [WSH09], in some cases by an
exponential factor. Theorem D.3 requires that the partitions 𝒫𝑖 be successive refinements, but has
the advantage of depending on a parameter 𝛾((𝒫𝑖 )𝐿
𝑖=1 ) that is larger than 𝑝min , and unlike 𝑝min , can
even be polynomial when the 𝒫𝑖 have exponentially many pieces. Theorem D.3 will not be necessary
for the proof of our main theorem.
12
4.2
Existence of partitions
We will show the existence of good partitions 𝒫𝑖 for 𝑖 ≥ 2 using a theorem of [GT14]. The theorem
shows if the 𝑘-th singular value is large, then it is possible to have a clustering with at most 𝑘 parts
which has high “inside” conductance within the clusters and low “outside” conductance between
the clusters (Definition A.7).
Theorem (Spectrally partitioning graphs, Theorem A.8). Let 𝑀 = (Ω, 𝑃 ) be a reversible Markov
chain with |Ω| = 𝑛 states. Let 0 = 𝜆1 ≤ 𝜆2 ≤ · · · ≤ 𝜆𝑛 be the eigenvalues of the Markov chain.
For any 𝑘 ≥ 2, if 𝜆𝑘 > 0, then √
there exists 1 ≤ ℓ ≤ 𝑘 − 1 and a ℓ-partitioning of Ω into sets
𝑃1 , . . . , 𝑃ℓ that is a (Ω(𝜆𝑘 /𝑘 2 ), 𝑂(ℓ3 𝜆ℓ ))-clustering.
For a mixture of 𝑛 gaussians, using the Poincaré inequality for gaussians we can show that
𝜆𝑛+1 for the continuous Langevin chain is bounded away from 0. Therefore one would hope to
use Theorem A.8 above to obtain a clustering. However, there are some difficulties, especially that
Theorem A.8 only holds for a discrete time, discrete space Markov chain.
To solve the discrete time problem, we fix a time 𝑇 , and consider the discrete time chain where
each step is running Langevin for time 𝑇 . To solve the discrete space problem, we note that we can
apply the theorem to the Markov chain 𝑀𝑖 projected to any partition (See Definition A.4 for the
definition of a projected Markov chain). A series of technical lemmas will show that the eigenvalues
and conductances do not change too much if we pass to the discrete time (and space) chain.
Another issue is that although the theorem guarantees good inner-conductance, it does not
immediately give a lowerbound for the size of the clusters. Here we again use Poincaré inequality
to show any small set must have a large outer-conductance, therefore the clustering guaranteed by
the theorem cannot have small clusters. Thus the assumptions of Theorem A.8 are satisfied, and
we get a partition with large internal conductance and small external conductance for the projected
chain (Lemma 6.8).
By letting the size of the cells in the partition go to 0, we show that the gap of the projected
chain approaches the gap of the continuous chain (Lemma 6.10). Because this lemma only works
for compact sets, we also need to show that restricting to a large ball doesn’t change the eigenvalues
too much (Lemma 6.11).
4.3
Mixing at highest temperature
Next, we need to show mixing at the highest temperature. Over bounded domains, we could set the
highest temperature to be infinite, which would correspond to uniform sampling over the domain.
Since we are working over an unbounded domain, we instead compare it to Langevin dynamics
for a strictly convex function which is close in 𝐿∞ distance (Lemma 7.3). We use the fact that
Langevin dynamics mixes rapidly for strictly convex functions, and an 𝐿∞ perturbation of 𝜀 affects
the spectral gap by at most a factor of 𝑒𝜀 .
4.4
Discretizing the Langevin diffusion
Up to this point, though we subdivided time into discrete intervals of size 𝑇 , in each time interval
we ran the continuous Langevin chain for time 𝑇 . However, algorithmically we can only run a
discretization of Langevin diffusion – so we need to bound the drift of the discretization from the
continuous chain.
13
For this, we follow the usual pattern of discretization arguments: if we run the continuous chain
for some time 𝑇 with step size 𝜂, the drift of the discretized chain to the continuous will be 𝜂𝑇 . If
we have a bound on 𝑇 , this provides a bound for the drift. More precisely, we show:
Lemma (Lemma 8.1). Let 𝑝𝑡 , 𝑞 𝑡 : R𝑑 × [𝐿] → R be the distributions after running the simulated
tempering chain for 𝑡 steps, where in 𝑝𝑡 , for any temperature 𝑖 ∈ 𝐿, the Type 1 transitions are taken
according to the (discrete time) Markov kernel 𝑃𝑇 : running Langevin diffusion for time 𝑇 ; in 𝑞 𝑡 ,
the Type 1 transitions are taken according to running 𝑇𝜂 steps of the discretized Langevin diffusion,
using 𝜂 as the discretization granularity, s.t. 𝜂 ≤
KL(𝑝𝑡 ||𝑞 𝑡 ) .
𝜎2
2 .
Then,
𝜂2
𝜂2
𝜂
2
(𝐷
+
𝑑)𝑇
𝑡
+
max E𝑥∼𝑝0 (·,𝑖) ‖𝑥 − 𝑥* ‖22 + 4 𝑑𝑡𝑇 𝜂
6
6
𝑖
𝜎
𝜎
𝜎
To prove this, consider the two types of steps separately. The Type 2 steps in the tempering
chains do not increase the KL divergence between the continuous and discretized version of the
chains, and during the Type 1 steps, the increase in KL divergence on a per-step basis can be
bounded by using existing machinery on discretizing Langevin diffusion (see e.g. [Dal16]) along
with a decomposition theorem for the KL divergence of a mixture of distributions (Lemma 8.6).
We make a brief remark about max𝑖 E𝑥∼𝑝0 (·,𝑖) ‖𝑥 − 𝑥* ‖22 : since the means 𝜇𝑖 satisfy ‖𝜇𝑖 ‖ ≤ 𝐷,
it’s easy to characterize the location of 𝑥* and conclude that bounding this quantity essentially
requires that most of the mass in the initial distributions should be concentrated on a ball of size
𝑂(𝐷). Namely, the following holds:
√
Lemma 4.2. Let 𝑥* = argmin𝑥∈R𝑑 𝑓˜(𝑥). Then, ‖𝑥* ‖ ≤ 2𝐷.
4.5
Estimating the partition functions
Finally, the Metropolis-Hastings filter in the Type 2 step of the simulated tempering chain requires
us to estimate the partition functions at each of the temperatures. It is sufficient to estimate the
partition functions to within a constant factor, because the gap of the tempering chain depends on
the ratio of the maximum-to-minimum probability of a given temperature.
For this, we run the chain for temperatures 𝛽1 , . . . , 𝛽ℓ , obtain good samples for 𝑝ℓ , and use them
to estimate 𝑍ℓ+1 . We use Lemma 9.1 to show that with high probability, this is a good estimate.
5
Spectral gap of simulated tempering
In this section we prove a lower bound for the spectral gap of simulated tempering given a partition.
Assumptions 5.1. Let 𝑀𝑖 = (Ω, 𝑃𝑖 ), 𝑖 ∈ [𝐿] be a sequence of Markov chains with state space Ω
and stationary distributions 𝑝𝑖 . Consider the simulated tempering chain 𝑀st = (Ω × [𝐿], 𝑃st ) with
min(𝑟𝑖 )
probabilities (𝑟𝑖 )𝐿
𝑖=1 . Let 𝑟 = max(𝑟𝑖 ) .
Let 𝒫𝑖 be a partition2 of the ground set Ω, for each 𝑖 ∈ [𝐿], with 𝒫1 = {Ω}.
Define the overlap parameter of (𝒫𝑖 )𝐿
𝑖=1 to be
𝛿((𝒫𝑖 )𝐿
𝑖=1 ) =
2
ïZ
min
1<𝑖≤𝐿,𝐴∈𝒫
ò
min{𝑝𝑖−1 (𝑥), 𝑝𝑖 (𝑥)} 𝑑𝑥 /𝑝𝑖 (𝐴).
𝐴
We allow overlaps on sets of measure 0.
14
Theorem 5.2. Suppose Assumptions 5.1 hold. Define
𝑝min = min 𝑝𝑖 (𝐴).
𝑖,𝐴∈𝒫𝑖
Then the spectral gap of the tempering chain satisfies
Gap(𝑀st ) ≥
𝑟4 𝛿 2 𝑝2min
32𝐿4
min
1≤𝑖≤𝐿,𝐴∈𝒫𝑖
(Gap(𝑀𝑖 |𝐴 )).
(20)
Proof. Let 𝑝st be the stationary distribution of 𝑃st . First note that we can easily switch between
𝑝𝑖 and 𝑝st using 𝑝st (𝐴 × {𝑖}) = 𝑤𝑖 𝑝𝑖 (𝐴). Note 𝑤𝑖 ≥ 𝐿𝑟 .
Define the partition 𝒫 on Ω × [𝐿] by
𝒫 = {𝐴 × {𝑖} : 𝐴 ∈ 𝒫𝑖 } .
By Theorem A.5,
Gap(𝑀st ) ≥
1
Gap(𝑀 st ) min Gap(𝑀st |𝐵 ).
𝐵∈𝒫
2
(21)
The second term Gap(𝑀st |𝐵 ) is related to (Gap(𝑀𝑖 |𝐴 )). We now lower-bound Gap(𝑀 st ). We will
abuse notation by considering the sets 𝐵 ∈ 𝒫 as states in 𝑀 st , and identify a union of sets in 𝒫
with the corresponding set of states for 𝑀 st .
We bound Gap(𝑀 st ) by bounding the conductance of 𝑀 st using Cheeger’s inequality (Theorem A.9).
Suppose first that 𝑆 ⊆ 𝒫, Ω × {1} ̸∈ 𝑆. Intuitively, this means the “highest temperature” is not
in the set. We will go from top to bottom until we find a partition that is in the set, the interaction
of this part and the temperature above it will already provide a large enough cut.
1
Let 𝑖 be minimal such that 𝐴 × {𝑖} ∈ 𝑆 for some 𝐴. There is 2𝐿
probability of proposing a
switch to level 𝑖 − 1, so
𝑃st (𝐴 × {𝑖}, Ω × {𝑖 − 1}) ≥
=
≥
≥
®
´
1
1
𝑝𝑖−1 (𝑥) 𝑟𝑖−1
𝑟𝑖 𝑝𝑖 (𝑥) min
, 1 𝑑𝑥
2𝐿 𝑟𝑖 𝑝𝑖 (𝐴) 𝐴
𝑝𝑖 (𝑥) 𝑟𝑖
Z
ß
™
1
𝑟𝑖−1
min 𝑝𝑖−1 (𝑥)
, 𝑝𝑖 (𝑥) 𝑑𝑥/𝑝𝑖 (𝐴)
2𝐿 𝐴
𝑟𝑖
Z
1 min 𝑟𝑗
min {𝑝𝑖−1 (𝑥), 𝑝𝑖 (𝑥)} 𝑑𝑥/𝑝𝑖 (𝐴)
2𝐿 max 𝑟𝑗 𝐴
1
𝑟𝛿.
2𝐿
Z
(22)
(23)
(24)
(25)
We have that (defining 𝑄 as in Definition A.6)
𝑄st (𝑆, 𝑆 𝑐 )
𝑝(𝑆)
𝑝st (𝐴 × {𝑖})𝑃st (𝐴 × {𝑖}, Ω × {𝑖 − 1})
≥
𝑝(𝑆)
2
𝑟𝑝min 1
𝑟 𝛿𝑝min
≥
𝑟𝛿 =
𝐿 2𝐿
2𝐿2
𝜑st (𝑆) =
15
(26)
(27)
(28)
Now consider the case that Ω × {1} ∈ 𝑆. This case the highest temperature is in the set. We
then try to find the part with highest temperature that is not in the set.
Note that 𝑝st (Ω × {1}) ≥ 𝐿𝑟 . Define 𝐴 × {𝑖} as above for 𝑆 𝑐 , then
𝑄st (𝑆, 𝑆 𝑐 )
𝑝(𝑆)
𝑝st (Ω × {1})𝑃st (𝐴 × {𝑖}, Ω × {𝑖 − 1})
≥
𝑝(𝑆)
𝑟 1
≥
𝑟𝛿.
𝐿 2𝐿
𝜑st (𝑆) =
(29)
(30)
(31)
Thus by Cheeger’s inequality A.9,
Gap(𝑀 st ) ≥
Φ(𝑀 st )2
𝑟4 𝛿 2 𝑝2min
=
2
8𝐿4
(32)
Therefore we have proved the projected Markov chain (between partitions) has good spectral
gap. What’s left is to prove that inside each partition the Markov chain has good spectral gap,
note that
Gap(𝑀st |𝐵×{𝑖} ) ≥
1
Gap(𝑀𝑖 |𝐴 )
2
(33)
because the chain 𝑀st , on a state in Ω × {𝑖}, transitions according to 𝑀𝑖 with probability
Plugging (32) and (33) into (21) gives the bound.
1
2.
Remark 5.3. Suppose in the type 2 transition we instead pick 𝑘 ′ as follows: With probability
let 𝑘 ′ = 𝑘 − 1, and with probability 12 , let 𝑘 ′ = 𝑘 + 1. If 𝑘 ′ ̸∈ [𝐿], let 𝑘 ′ = 𝑘 instead.
1
Then the 2𝐿
becomes 14 in the proof above so we get the improved gap
1
2,
Gap(𝑀st ) ≥
6
𝑟4 𝛿 2 𝑝2min
128𝐿2
min
1≤𝑖≤𝐿,𝐴∈𝒫𝑖
(Gap(𝑀𝑖 |𝐴 )).
(34)
Defining the partitions
In this section, we assemble all the ingredients to show that there exists a partition for the Langevin
chain such that min2≤𝑖≤𝐿,𝐴∈𝒫𝑖 (Gap(𝑀𝑖 |𝐴 )) is large, and each part in the partition also has significant probability. Hence the partition will be sufficient for the partitioning technique discussed in
previous section.
The high-level plan is to use Theorem A.8 to find the partitions for each temperature. Indeed,
if we have a mixture of 𝑛 gaussians, it is not hard to show that the (𝑛 + 1)-st eigenvalue is large:
Lemma 6.1 (Eigenvalue gap for mixtures). Let 𝑝𝑖 (𝑥) = 𝑒−𝑓𝑖 (𝑥) be probability distributions on Ω
P
P
and let 𝑝(𝑥) = 𝑛𝑖=1 𝑤𝑖 𝑝𝑖 (𝑥), where 𝑤1 , . . . , 𝑤𝑛 > 0 and 𝑛𝑖=1 𝑤𝑖 = 1. Suppose that for each 𝑝𝑖 , a
Poincaré inequality holds with constant 𝐶.
Then the (𝑛 + 1)-th eigenvalue of L satisfies
𝜆𝑛+1 (−L ) ≥
16
1
.
𝐶
We defer the proof to Section 6.1. However, there are still many technical hurdles that we need
to deal with before we can apply Theorem A.8 on spectral partitioning.
1. When the temperature is different, the distribution (which is proportional to 𝑒−𝛽𝑓 (𝑥) ) is no
longer a mixture of gaussians. We show that it is still close to a mixture of gaussians in the
sense that the density function is point-wise within a fixed multiplicative factor to the density
of a mixture of gaussians (Section 6.2). This kind of multiplicative guarantee allows us to
relate the Poincaré constants between the two distributions.
2. We then show (Section 6.3) a Poincaré inequality for all small sets. This serves two purposes
in the proof: (a) it shows that the inner-conductance is large. (b) it shows that if a set has
small conductance then it cannot be small. We also deal with the problem of continuous time
here by taking a fixed time 𝑡 and running the Markov chain in multiples of 𝑡.
3. Now we can prove Lemma 6.8, which shows that if we discretize the continuous-space Markov
chain, then there exists good partitioning in the resulting discrete-space Markov chain (Section 6.4).
4. We then show that if we restrict the Langevin chain to a large ball, and then discretize the
space in the large ball finely enough, then in the limit the spectral gap of the discretized chain
is the same as the spectral gap of the continuous-space Markov chain (Section 6.5).
5. Finally in Section 6.6 we show it is OK to restrict the Langevin chain restricted to a large
ball.
6.1
Proving the eigenvalue gap
Now we prove Lemma 6.1. The main idea is to use the variational characterization of eigenvalues,
and show that there can be at most 𝑛 “bad” directions.
Proof. We use the variational characterization of eigenvalues:
𝜆𝑛+1 (−L ) =
min
max
subspace 𝑆 ⊆ 𝐿2 (𝑝) 𝑔⊥𝑝 𝑆
dim 𝑆 = 𝑛
− ⟨𝑔, L 𝑔⟩
.
‖𝑔‖2𝑝
R
To lower-bound this, it suffices to produce a 𝑛-dimensional subspace 𝑆 and lower-bound
‖∇𝑔‖2 𝑝 𝑑𝑥
R𝑑
‖𝑔‖2𝑝
for 𝑔 ⊥ 𝑆. We choose
𝑝𝑖
:1≤𝑖≤𝑛 .
𝑆 = span
𝑝
ß
Take 𝑔 ⊥𝑝
𝑝𝑖
𝑝
™
(35)
for each 𝑖. Then, since a Poincaré inequality holds on 𝑝𝑖 ,
𝑝𝑖
𝑝 𝑑𝑥 = 0
R𝑑 𝑝
R
R
2
2
1
𝑑 ‖∇𝑔‖ 𝑝𝑖 𝑑𝑥
R𝑑 ‖∇𝑔‖ 𝑝𝑖 𝑑𝑥
=⇒
= R
≥ .
2
Var𝑝𝑖 (𝑔)
𝐶
‖𝑔‖𝑝
Z
𝑔
𝑖
17
(36)
(37)
Thus
R
R𝑑
‖∇𝑔‖2 𝑝 𝑑𝑥
‖𝑔‖2𝑝
2
𝑖=1 R𝑑 ‖∇𝑔‖ 𝑤𝑖 𝑝𝑖 𝑑𝑥
P𝑛
2
𝑖=1 𝑤𝑖 ‖𝑔‖𝑝𝑖
P𝑚 R
=
≥
1
,
𝐶
(38)
as needed.
6.2
Scaled temperature approximates mixture of gaussians
The following lemma shows that changing the temperature is approximately the same as changing
the variance of the gaussian. We state it more generally, for arbitrary mixtures of distributions in
the form 𝑒−𝑓𝑖 (𝑥) .
Lemma 6.2 (Approximately scaling
the temperature). Let 𝑝𝑖 (𝑥) = 𝑒−𝑓𝑖 (𝑥) be probability distribuR −𝛽𝑓 (𝑥)
tions on Ω such that for all 𝛽 > 0, 𝑒 𝑖 𝑑𝑥 < ∞. Let
𝑝(𝑥) =
𝑛
X
𝑤𝑖 𝑝𝑖 (𝑥)
(39)
𝑖=1
𝑓 (𝑥) = − ln 𝑝(𝑥)
(40)
where 𝑤1 , . . . , 𝑤𝑛 > 0 and 𝑛𝑖=1 𝑤𝑖 = 1. Let 𝑤min = min1≤𝑖≤𝑛 𝑤𝑖 .
Define the distribution at inverse temperature 𝛽 to be 𝑝𝛽 (𝑥), where
P
𝑔𝛽 (𝑥) = 𝑒−𝛽𝑓 (𝑥)
Z
𝑒−𝛽𝑓 (𝑥) 𝑑𝑥
𝑍𝛽 =
(41)
(42)
Ω
𝑝𝛽 (𝑥) =
𝑔𝛽 (𝑥)
.
𝑍𝛽
(43)
𝑤𝑖 𝑒−𝛽𝑓𝑖 (𝑥)
(44)
Define the distribution 𝑝e𝛽 (𝑥) by
𝑔e𝛽 (𝑥) =
𝑛
X
𝑖=1
‹𝛽 =
𝑍
Z X
𝑛
𝑤𝑖 𝑒−𝛽𝑓𝑖 (𝑥) 𝑑𝑥
(45)
.
(46)
Ω 𝑖=1
𝑝e𝛽 (𝑥) =
𝑔e𝛽 (𝑥)
‹𝛽
𝑍
Then for 0 ≤ 𝛽 ≤ 1 and all 𝑥,
ï
𝑔𝛽 (𝑥) ∈ 1,
1
ò
𝑔e𝛽
𝑤min
ï
ò
ï
ò
‹𝛽
𝑍
1
1
𝑝e𝛽
𝑝e𝛽 .
𝑝𝛽 (𝑥) ∈ 1,
⊂ 𝑤min ,
𝑤min
𝑍𝛽
𝑤min
18
(47)
(48)
Proof. By the Power-Mean inequality,
𝑔𝛽 (𝑥) =
≥
𝑛
X
!𝛽
𝑤𝑖 𝑒−𝑓𝑖 (𝑥)
(49)
𝑖=1
𝑛
X
𝑤𝑖 𝑒−𝛽𝑓𝑖 (𝑥) = 𝑔e𝛽 (𝑥).
(50)
𝑖=1
On the other hand, given 𝑥, setting 𝑗 = argmin𝑖 𝑓𝑖 (𝑥),
𝑔𝛽 (𝑥) =
This gives (47). This implies
6.3
e𝛽
𝑍
𝑍𝛽
−𝑓𝑖 (𝑥)
𝑤𝑖 𝑒
𝑖=1
−𝑓𝑗 (𝑥) 𝛽
≤ (𝑒
≤
!𝛽
𝑛
X
(51)
)
1
𝑤min
𝑛
X
(52)
𝑤𝑖 𝑒−𝛽𝑓𝑖 (𝑥) =
𝑖=1
1
𝑔e𝛽 (𝑥).
𝑤min
(53)
∈ [𝑤min , 1], which gives (48).
Poincaré inequalities on small subsets
In this section we prove Poincaré inequalities for small sets. In fact we need to prove that this
property is true robustly, in order to transform the continuous time Markov chain to a discrete
time Markov chain.
Definition 6.3. Given a measure 𝑝 on Ω, say that a Poincaré inequality with constant 𝐶 holds on
sets of measure ≤ 𝐷 if whenever 𝑝(Supp(𝑔)) ≤ 𝐷,
ℰ𝑝 (𝑔) =
Z
‖∇𝑔‖2 𝑝 𝑑𝑥 ≥
Ω
1
Var𝑝 (𝑔).
𝐶
This is robust in the following sense: If the above condition is satisfied, then 𝑔 still satisfies a
Poincaré inequality even if it is not completely supported on a small set, but just has a lot of mass
on a small set.
The main reason we need the robust version is that when we transform the continuous time
Markov chain to a discrete time Markov chain, even if we initialized in a small set after some time
the probability mass is going to spill over to a slightly larger set.
Lemma 6.4 (Robustly Poincaré on small sets). Let 𝐴 ⊆ R𝑑 be a subset. Suppose that for 𝑝, a
Poincaré inequality with constant 𝐶 holds on sets of measure ≤ 2𝑝(𝐴). Then if
Z
𝑔 2 𝑝 𝑑𝑥 ≥ 𝑘
Z
𝑔 2 𝑝 𝑑𝑥
(54)
R𝑑
𝐴
with 𝑘 > 2𝑝(𝐴), then
1
ℰ𝑝 (𝑔) ≥
20𝐶
Ç
2𝑝(𝐴)
1−
𝑘
19
å Z
𝑘
R𝑑
𝑔 2 𝑝 𝑑𝑥.
We lower-bound ℰ𝑝 (𝑔) by showing that (54) implies that not much of 𝑔’s mass comes from when
is small, so that much of 𝑔’s mass comes from the intersection of the set 𝐵 ′ where 𝑔 is large and
the set 𝐴. This means we can use the Poincaré inequality on a “sliced” version of 𝑔 on 𝐴.
𝑔2
Proof. By scaling we may assume
Let
R 2
1
𝑔 𝑝 𝑑𝑥 = 1. It suffices to show that ℰ(𝑔) ≥ 20𝐶
1 − 2𝑝(𝐴)
𝑘.
𝑘
®
𝑘
𝐵 = 𝑥 ∈ Ω : 𝑔(𝑥) ≥
2𝑝(𝐴)
ℎ(𝑥) =
2
0,
´
(55)
𝑘
𝑔(𝑥)2 ≤
𝑔(𝑥) −
q
𝑘
,
q 2𝑝
𝑘
𝑔(𝑥) +
,
2𝑝
𝑔(𝑥) >
q2𝑝(𝐴)
𝑘
(56)
2𝑝(𝐴)
q
𝑔(𝑥) < −
𝑘
2𝑝(𝐴) .
i.e., we “slice” out the portion where 𝑔(𝑥)2 is large and translate it to 0. (Note we cannot just take
ℎ(𝑥) = 𝑔(𝑥)1𝐵 (𝑥) because this is discontinuous. Compare with [BGL13, Proposition 3.1.17].)
ℰ𝑝 (𝑔) =
≥
Z
Z
‖∇𝑔‖2 𝑝 𝑑𝑥
(57)
‖∇ℎ‖2 𝑝 𝑑𝑥 = ℰ(ℎ).
(58)
By Cauchy-Schwarz, noting that Supp(ℎ) ⊆ 𝐵 and Vol𝑝 (𝐵) ≤
ã2
ÅZ
ℎ𝑝 𝑑𝑥
We can lower bound
R
≤
ÅZ
ã
ℎ2 𝑝 𝑑𝑥
2𝑝(𝐴)
𝑘 ,
2𝑝(𝐴)
.
𝑘
n
(59)
ℎ2 𝑝 𝑑𝑥 as follows. Let 𝐵 ′ = 𝑥 ∈ Ω : 𝑔(𝑥)2 ≥
Z
2𝑘
3𝑝(𝐴)
o
.
𝑔 2 𝑝 𝑑𝑥 ≥ 𝑘
(60)
𝐴
2𝑘
2𝑘
𝑝(𝐴) =
3𝑝(𝐴)
3
Z
2𝑘
𝑘
𝑔 2 𝑝 𝑑𝑥 ≥ 𝑘 −
=
3
3
𝐴∩𝐵 ′
Z
Z
2
ℎ
ℎ2 𝑝 𝑑𝑥 ≥
𝑔 2 𝑝 𝑑𝑥
2
Ω
𝐴∩𝐵 ′ 𝑔
Z
𝐴∩𝐵 ′ 𝑐
𝑔 2 𝑝 𝑑𝑥 ≤
≥ 0.15
Z
(61)
(62)
(63)
𝑔 2 𝑝 𝑑𝑥
(64)
𝐴∩𝐵 ′
≥
1
𝑘.
20
(65)
»
where we used the fact that when 𝑦 2 ≥
2𝑘
2𝑝(𝐴) ,
20
(𝑦−
𝑘
)2
2𝑝(𝐴)
2
𝑦
Ä»
≥
» ä2
2𝑘
−
3𝑝
2𝑘
3𝑝
𝑘
2𝑝
> 0.15. Putting
everything together,
ÅZ
Var(ℎ) =
ã
ℎ2 𝑝 𝑑𝑥 −
Ç 𝐵
ã2
ÅZ
ℎ𝑝 𝑑𝑥
å ÅZ 𝐵
(66)
2𝑝(𝐴)
ℎ2 𝑝 𝑑𝑥
𝑘
𝐵
Ç
å
2𝑝(𝐴) 𝑘
≥ 1−
𝑘
20
≥
1−
ã
(67)
(68)
ℰ𝑝 (𝑔) ≥ ℰ𝑝 (ℎ)
(69)
1
1
Var(ℎ) =
𝐶
𝐶
≥
Ç
1−
2𝑝(𝐴)
𝑘
å
𝑘
.
20
(70)
Lemma 6.5 (Conductance from small-set Poincaré). Let 𝐴 ⊆ Ω be a set. Suppose that a Poincaré
inequality with constant 𝐶 holds on sets of measure ≤ 2𝑝(𝐴). Let
Z
𝜑𝑡 (𝐴) =
𝑃𝑡 (𝑥, 𝐴𝑐 )
𝐴
𝑝(𝑥)
𝑑𝑥
𝑝(𝐴)
be the conductance of a set 𝐴 after running Langevin for time 𝑡. Then
1 1
,
(1 − 4𝑝(𝐴)) 𝑡 .
𝜑𝑡 (𝐴) ≥ min
2 80𝐶
™
ß
(71)
A Poincaré inequality can be thought of as giving a lower bound on “instantaneous” conductance. We show that this implies good conductance for a finite time 𝑇 . What could go wrong
is that the rate of mass leaving a set 𝐴 is large at time 0 but quickly goes to 0. We show using
Lemma 6.4 that this does not happen until significant mass has escaped.
Proof. We want to bound
1
𝑃𝑡 (𝑥, 𝐴)1𝐴 𝑝(𝑥) 𝑑𝑥
𝑝(𝐴)
Z
1
=1−
(𝑃𝑡 1𝐴 )1𝐴 𝑝(𝑥) 𝑑𝑥
𝑝(𝐴)
1
=1−
⟨𝑃𝑡 1𝐴 , 1𝐴 ⟩𝑝
𝑝(𝐴)
∂
1 ¨
=1−
𝑃𝑡/2 1𝐴 , 𝑃𝑡/2 1𝐴 𝑝
𝑝(𝐴)
𝜑𝑡 (𝐴) = 1 −
Z
(72)
(73)
(74)
(75)
since (𝑃𝑡 )𝑡≥0 is a one-parameter semigroup and is self-adjoint with respect to 𝑝. Now by definition
of L ,
𝑑
𝑑
⟨𝑃𝑡 𝑔, 𝑃𝑡 𝑔⟩𝑝 = 2 𝑃𝑡 𝑔, 𝑃𝑡 𝑔
𝑑𝑡
𝑑𝑡
≠
∑
𝑝
= 2 ⟨𝑃𝑡 𝑔, L 𝑃𝑡 𝑔⟩𝑝 = −2ℰ(𝑃𝑡 𝑔)
(76)
Let 𝑡0 be the minimal 𝑡 such that
1
‖𝑃𝑡 1𝐴 ‖ 𝑝 𝑑𝑥 ≤
2
𝐴
Z
2
1
‖1𝐴 ‖2 𝑝 𝑑𝑥 = 𝑝(𝐴).
2
𝐴
Z
21
(77)
(Let 𝑡 = ∞ if this never happens; however, (83) will show that 𝑡 < ∞.) For 𝑡 < 2𝑡0 , by Lemma 6.4,
𝑑
𝑑𝑡
Z
‖𝑃𝑡 1𝐴 ‖2 𝑝 𝑑𝑥 ≤ −2ℰ(𝑃𝑡 1𝐴 )
(78)
1
𝑑
‖𝑃𝑡 1𝐴 ‖2 𝑝 𝑑𝑥
(1 − 4𝑝(𝐴))
40𝐶
𝑑𝑡
Z
1
𝑑
≤−
‖𝑃𝑡 1𝐴 ‖2 𝑝 𝑑𝑥.
(1 − 4𝑝(𝐴))
20𝐶
𝑑𝑡
Z
≤ −2
(79)
(80)
This differential inequality implies exponential decay, a fact known as Gronwall’s inequality.
Z
‖𝑃𝑡 1𝐴 ‖2 𝑝 𝑑𝑥 ≤ 𝑒− 20 (1−4𝑝(𝐴))𝑡
𝐶
Z
‖1𝐴 ‖2 𝑝 𝑑𝑥
𝐴
|
𝜑𝑡 (𝐴) ≤ 1 −
{z
}
𝑝(𝐴)
1
𝑃 1𝐴
𝑃 (𝐴) 𝑡/2
2
(82)
𝑝
1
≤ 1 − 𝑒− 40𝐶 (1−4𝑝(𝐴))𝑡 ≤ max
where the last inequality follows because because
For 𝑡 ≥ 2𝑡0 , we have, because
R
1
2
𝐴 ‖𝑃𝑡 𝐴 ‖ 𝑝 𝑑𝑥
(81)
𝑑 −𝑥
𝑑𝑥 𝑒
1
2
≥
Å
1 1
,
(1 − 4𝑝(𝐴))𝑡
2 80𝐶
ã
(83)
when 𝑒−𝑥 ≥ 12 .
is decreasing, 𝑃𝑡/2 1𝐴
2
𝑝
≤ ‖𝑃𝑡0 1𝐴 ‖2𝑝 = 21 𝑝(𝐴) so
𝜑𝑡 (𝐴) ≥ 12 .
Lemma 6.6 (Easy direction of relating Laplacian of projected chain). Let (Ω, 𝑃 ) be a reversible
Markov chain and 𝑃 its projection with respect to some partition 𝒫 = {𝐴1 , . . . , 𝐴𝑛 } of Ω. Let
ℒ = 𝐼 − 𝑃, ℒ = 𝐼 − 𝑃.
𝜆𝑛 (ℒ) ≤ 𝜆𝑛 (ℒ).
Proof. The action of ℒ on functions on [𝑛] is the same as the action of ℒ on the subspace of functions
that are constant on each set in the partition, denoted by 𝐿2 (𝒫). This means that 𝐿2 ([𝑛]) under
the action of ℒ embeds into 𝐿2 (𝑝) under the action of ℒ. Let 𝜋 = E[·|𝒫] : 𝐿2 (𝑝) → 𝐿2 (𝒫) be the
projection; note that for ℎ ∈ 𝐿2 (𝒫), ⟨ℎ, 𝜋𝑔⟩ = ⟨ℎ, 𝑔⟩.
By the variational characterization of eigenvalues, we have that for some 𝑆,
𝜆𝑛 (ℒ) =
min
𝑔 ⊥𝑝 𝑆
𝑔 ∈ 𝐿2 (𝑝)
⟨𝑔, ℒ𝑔⟩
≤
‖𝑔‖2𝑝
min
𝑓 ⊥𝑝 𝜋(𝑆)
𝑔 ∈ 𝐿2 (𝒫)
⟨𝑓, ℒ𝑔⟩
≤
‖𝑔‖2𝑝
max
𝐿2 (𝒫)
𝑆⊆
dim(𝑆) = 𝑛 − 1
⟨𝑔, ℒ𝑔⟩
= 𝜆𝑛 (ℒ).
‖𝑔‖2𝑝
(84)
Lemma 6.7 (Small-set Poincaré inequality for mixtures). Keep the setup of Lemma 6.2. Further
suppose that
Z
Z
𝑍𝛽 =
𝑒−𝛽𝑓1 𝑑𝑥 = · · · =
𝑒−𝛽𝑓𝑛 𝑑𝑥.
R𝑑
R𝑑
(E.g., this holds for gaussians of equal variance.) Suppose that a Poincaré inequality holds for each
−𝛽𝑓
𝑝e𝛽,𝑖 = 𝑒 𝑍𝛽 𝑖 with constant 𝐶𝛽 .
Then on 𝑝𝛽 , a Poincaré inequality holds with constant
22
2𝐶𝛽
𝑤min
on sets of size ≤
2
𝑤min
2 .
Proof. Let 𝐴 = Supp(𝑔). Then for all 𝑗, by Lemma 6.2,
𝑤min 𝑝e𝛽,𝑗 (𝐴) ≤ 𝑤𝑗 𝑝e𝛽,𝑗 (𝐴) ≤
𝑛
X
𝑝e𝛽,𝑖 (𝐴) = 𝑝e𝛽 (𝐴) ≤
𝑖=1
1
𝑤min
𝑝𝛽 (𝐴) ≤
𝑤min
2
(85)
1
=⇒ 𝑝e𝛽,𝑗 (𝐴) ≤ .
2
(86)
As in (66), using 𝑝e𝛽,𝑗 (𝐴) ≤ 12 ,
Var𝑝e𝛽,𝑗 (𝑔) ≥
≥
ÅZ
1
2
ã
Z𝐴
𝑔 2 𝑝e𝛽,𝑗 𝑑𝑥 −
𝐴
ã2
ÅZ
𝐴
(87)
𝑔 𝑝e𝛽,𝑗 𝑑𝑥
𝑔 2 𝑝e𝛽,𝑗 𝑑𝑥
(88)
Then
ℰ𝑝e𝛽 (𝑓 ) =
≥
Z
2
‖∇𝑓 ‖
𝐴
𝑛
X
𝑤𝑗 𝑝e𝛽,𝑗 (𝑥) 𝑑𝑥
(89)
𝑖=1
𝑛
1 X
𝑤𝑖 Var𝑝e𝛽,𝑗 (𝑓 )
𝐶𝛽 𝑖=1
(90)
Z
𝑛
1
1 X
𝑤𝑖
𝑓 2 𝑝e𝛽,𝑗 (𝑥) 𝑑𝑥
≥
𝐶𝛽 𝑖=1 2 𝐴
≥
1
Var𝑝e𝛽 (𝑓 ).
2𝐶𝛽
Using Lemma 6.2 and Lemma C.3(3), 𝑝𝛽 satisfies a Poincaré inequality with constant
6.4
(91)
(92)
2𝐶𝛽
𝑤min .
Existence of partition
Now we are ready to prove the main lemma of this section, which gives the partitioning for any
discrete space, discrete time Markov chain. In later subsections we will connect this back to the
continuous space Markov chain.
Lemma 6.8. Let 𝑝(𝑥) ∝ 𝑒−𝑓 (𝑥) be a probability density function on R𝑑 . Let 𝐶 and 𝜇 ≤ 1 be
such that the Langevin chain on 𝑓 (𝑥) satisfies 𝜆𝑛+1 (L ) ≥ 𝐶1 , and a Poincaré inequality holds with
constant 2𝐶 on sets of size ≤ 𝜇.
Let 𝒫 = {𝐴1 , . . . , 𝐴𝑚 } be any partition of 𝐵 ⊆ R𝑑 . Let (R𝑑 , 𝑃𝑇 ) be the discrete-time Markov
chain where each step is running continuous Langevin for time 𝑇 ≤ 𝐶2 , (𝐵, 𝑃𝑇 |𝐵 ) be the same
Markov chain restricted to 𝐵, and ([𝑚], 𝑃𝑇 |𝐵 ) is the projected chain with respect to 𝒫.
Suppose that 𝐵 satisfies the following.
1. For all 𝑥 ∈ 𝐵, 𝑃𝑇 (𝑥, 𝐵 𝑐 ) ≤
2. 𝜆𝑛+1 (𝐼 − 𝑃𝑇 |𝐵 ) ≥
3
4
Ä
𝑇
1000𝐶 .
𝜆𝑛+1 (𝐼 − 𝑃𝑇 ) −
𝑇
6𝐶
ä
.
Then there exists ℓ ≤ 𝑛 and a partition 𝒥 of [𝑚] into ℓ sets 𝐽1 , . . . , 𝐽ℓ such that the following
hold.
23
Ä Ä
1. 𝒥 is a Ω
𝑇2
𝐶 2 𝑚8
ä
,𝑂
Ä ää
𝑇
𝐶
-clustering.
2. Every set in the partition has measure at least
𝜇
2.
Proof. First we show that the (𝑛 + 1)th eigenvalue is large. Note that the eigenvalues of 𝑃𝑇 are
the exponentials of the eigenvalues of L .
𝜆𝑛+1 (𝐼 − 𝑃𝑇 |𝐵 ) ≥ 𝜆𝑛+1 (𝐼 − 𝑃𝑇 |𝐵 )
Å
ã
𝑇
3
𝜆𝑛+1 (𝐼 − 𝑃𝑇 ) −
≥
4
6𝐶
3
𝑇
3
𝑇
−𝜆𝑛+1 (−L )𝑇
= (1 − 𝑒
−
) ≥ (1 − 𝑒−𝑇 /𝐶 −
)
4Å
6𝐶
4
6𝐶
ß
™
ã
1 𝑇
𝑇
𝑇
3
min
,
−
=
≥
4
2 2𝐶
6𝐶
4𝐶
Lemma 6.6
(93)
assumption
(94)
(95)
(96)
Let 𝑐 be a small constant to be chosen. Let 𝑘 ≤ 𝑛 + 1 be the largest integer so that
𝜆𝑘−1 (𝐼 − 𝑃𝑇 |𝐵 ) ≤
Then 𝜆𝑘 (𝐼 − 𝑃𝑇 |𝐵 ) >
parameters
𝑐2 𝑇 2
.
𝐶 2 𝑛6
𝑐2 𝑇 2
𝐶 2 𝑛6
(97)
By Theorem A.8, for some 1 ≤ ℓ ≤ 𝑘 − 1, there exists a clustering with
p
𝜆𝑘
Ω 2 , 𝑂(ℓ3 𝜆ℓ ) =
𝑘
Å Å
ã
ã
Ç Ç 2 2å
𝑐 𝑇
Ω
𝐶 2 𝑛8
𝑐𝑇
,𝑂
𝐶
Å
ãå
(98)
S
Now consider a set 𝐽 in the partition. Let 𝐴 = 𝑗∈𝐽 𝐴𝑗 . Suppose by way of contradiction that
𝑝(𝐴) ≤ 𝜇2 . By Lemma 6.5 and noting 𝑝(𝐴) < 12 , the conductance of 𝐴 in R𝑑 is
𝜑𝑇 (𝐴) ≥
𝑇
1
(1 − 4𝑝(𝐴))𝑇 ≥
.
80(2𝐶)
320𝐶
(99)
The conductance of 𝐴 in 𝐵 satisfies
Å
𝑂
R
Now by assumption,
𝐴
𝑃 (𝑥,𝐵 𝑐 )𝑝(𝑥) 𝑑𝑥
𝑝(𝐴)
≤
Å
𝑂
𝑐𝑇
𝐶
ã
𝑇
1000𝐶 ,
𝑐𝑇
𝐶
ã
+
≥ 𝜑𝑇 |𝐵 (𝐴).
(100)
so
𝑇
≥ 𝜑𝑇 (𝐴)
1000𝐶
(101)
For 𝑐 chosen small enough, together (99) and (101) give a contradiction.
6.5
Making arbitrarily fine partitions
In this section we show when the discretization is fine enough, the spectral gap of the discrete
Markov chain approaches the spectral gap of the continuous-space Markov chain.
We will need the following fact from topology.
24
Lemma 6.9 (Continuity implies uniform continuity on compact set). If (Ω, 𝑑) is a compact metric
space and 𝑔 : Ω → R is continuous, then for every 𝜀 > 0 there is 𝛿 such that for all 𝑥, 𝑥′ ∈ Ω with
𝑑(𝑥, 𝑥′ ) < 𝛿,
|𝑔(𝑥) − 𝑔(𝑥′ )| < 𝜀.
We know that the gap of a projected chain is at least the gap of the original chain, Gap(𝑀 ) ≥
Gap(𝑀 ) (Lemma 6.6). We show that if the size of the cells goes to 0, then the reverse inequality
also holds. Moreover, the convergence is uniform in the size of the cells.
Lemma 6.10. Let 𝑀 = (R𝑑 , 𝑃 ′ ) be a reversible Markov chain where the kernel 𝑃 ′ : Ω × Ω → R is
continuous and > 0 everywhere and the stationary distribution 𝑝′ : Ω → R is a continuous function.
Fix a compact set Ω ⊂ R𝑛 . Then
𝒫
lim inf
𝛿↘0 𝐾,𝒫
Gap(𝑀 |𝐾 )
=1
Gap(𝑀 |𝐾 )
where the infimum is over all compact sets 𝐾 ⊆ Ω and all partitions 𝒫 of 𝐾 composed of sets 𝐴
with diam(𝐴) < 𝛿.
Proof. By Lemma 6.9 on ln 𝑃 ′ (𝑥, 𝑦) : Ω × Ω → R and 𝑝′ (𝑥), there exists 𝛿 > 0 such that for all
𝑥, 𝑦 ∈ Ω such that 𝑑(𝑥, 𝑥′ ) < 𝛿 and 𝑑(𝑦, 𝑦 ′ ) < 𝛿,
𝑃 ′ (𝑥′ , 𝑦 ′ )
∈ [𝑒−𝜀 , 𝑒𝜀 ]
𝑃 ′ (𝑥, 𝑦)
𝑝′ (𝑥)
∈ [𝑒−𝜀 , 𝑒𝜀 ].
𝑝′ (𝑥′ )
(102)
(103)
We also choose 𝛿 small enough so that for all sets 𝐴 with diameter < 𝛿, 𝑝(𝐴) < 𝜀.
Let 𝑃 and 𝑝 denote the kernel and stationary distribution for 𝑀 |𝐾 , and let
𝑃 (𝑥, 𝑑𝑦) = 𝛿𝑥 (𝑑𝑦)𝑃rej (𝑥) + 𝑃tr (𝑥, 𝑦) 𝑑𝑦,
(104)
where 𝑃rej (𝑥) = 𝑃 ′ (𝑥, 𝐾 𝑐 ) and 𝑃tr = 𝑃 ′ (the notation is to remind us that this is when a transition
succeeds). Let
𝑃acc (𝑥) = 1 − 𝑃rej (𝑥) =
Z
𝑃tr (𝑥, 𝑦) 𝑑𝑦.
(105)
𝑌
Consider a partition 𝒫 all of whose components have diameter < 𝛿. Let the projected chain be
𝑀 |𝐾 = (Ω, 𝑃 ) with stationary distribution 𝑝. We let capital letters 𝑋, 𝑌 denote the elements of Ω,
or the corresponding subsets of Ω, and 𝑓 denote a function Ω → R. Let 𝑄 be the probability distribution on Ω × Ω given by 𝑄(𝑥, 𝑑𝑦) = 𝑝(𝑥)𝑃 (𝑥, 𝑑𝑦), and similarly define 𝑄(𝑋, 𝑌 ) = 𝑝(𝑋)𝑃 (𝑋, 𝑌 ).
Also write
𝑃 (𝑋, 𝑌 ) = 1𝑋=𝑌 𝑃 rej (𝑋) + 𝑃 tr (𝑋, 𝑌 )
Z
𝑃 rej (𝑋) =
(106)
𝑝(𝑥)𝑃rej (𝑥) 𝑑𝑥 = 1 − 𝑃 acc (𝑋)
(107)
𝑝(𝑥)𝑃tr (𝑥, 𝑦) 𝑑𝑥.
(108)
Z𝑋
𝑃 tr (𝑋, 𝑌 ) =
𝑋
25
We have
s
(𝑔(𝑥) − 𝑔(𝑦))2 𝑝(𝑥)𝑃 (𝑥, 𝑑𝑦) 𝑑𝑥
E𝑥,𝑦∼𝑄 [𝑔(𝑥) − 𝑔(𝑦)]2
R
(109)
=
inf
𝑔
𝑔
2 (𝑔(𝑥) − E𝑝 𝑔(𝑥))2 𝑝(𝑥) 𝑑𝑥
E𝑥∼𝑝 [𝑔(𝑥) − E𝑝 𝑔]2
P
2
𝑋,𝑌 (𝑔(𝑋) − 𝑔(𝑌 )) 𝑝(𝑋)𝑃 (𝑋, 𝑌 ) 𝑑𝑥 𝑑𝑦
R P
(110)
Gap(𝑀 ) = inf
𝑔
2 ( 𝑋 𝑔(𝑋) − E𝑝 𝑔(𝑋))2 𝑝(𝑋) 𝑑𝑥
P
2
E𝑋,𝑌 ∼𝑄 [E𝑝 [𝑔|𝑋] − E𝑝 [𝑔|𝑌 ]]2
𝑋,𝑌 (E𝑝 [𝑔|𝑋] − E𝑝 [𝑔|𝑌 ]) 𝑝(𝑋)𝑃 (𝑋, 𝑌 )
R
= inf
=
inf
.
𝑔
𝑔
2 (E𝑝 [𝑔|𝒫](𝑥) − E𝑝 [𝑔])2 𝑝(𝑥) 𝑑𝑥
E𝑋∼𝑝 [E𝑝 [𝑔|𝑋] − E𝑝 [𝑔]]2
(111)
Gap(𝑀 ) = inf
We will relate these two quantities.
Consider the denominator of (109). By the Pythagorean theorem, it equals the variation between sets in the partition, given by the denominator of (111), plus the variation within sets in the
partition.
2
2
2
E [𝑔(𝑥) − E 𝑔] = E [E[𝑔|𝒫](𝑥) − E 𝑔] + E [𝑔(𝑥) − E[𝑔|𝒫](𝑥)] .
𝑥∼𝑝
𝑝
𝑥∼𝑝
𝑝
𝑥∼𝑝 𝑝
𝑝
(112)
We also decompose the numerator of (109). First we show that we can approximate 𝑝(𝑥)𝑃tr (𝑥, 𝑦)
with a distribution where 𝑦 is independent of 𝑥 given the set 𝑋 containing 𝑥. Using (102) and
(103),
s
𝑝(𝑥)𝑝(𝑦) 𝑋×𝑌 𝑝(𝑥′ )𝑃tr (𝑥′ , 𝑦 ′ ) 𝑑𝑥′ 𝑑𝑦 ′
𝑝(𝑦)
𝑝(𝑥)𝑃 tr (𝑋, 𝑌 )
=
(113)
𝑝(𝑌 )
𝑝(𝑋)𝑝(𝑌 )
s
𝑝(𝑥)𝑃tr (𝑥, 𝑦) 𝑋×𝑌 𝑝(𝑥′ )𝑝(𝑦 ′ ) 𝑑𝑥′ 𝑑𝑦 ′
2𝜀
≤𝑒
(114)
𝑝(𝑋)𝑝(𝑌 )
= 𝑒2𝜀 𝑝(𝑥)𝑃tr (𝑥, 𝑦).
(115)
Let 𝑅 be the distribution on Ω × Ω defined as follows:
𝑋, 𝑌 ∼ 𝑄,
𝑥 ∼ 𝑝|𝑋
𝑦 ∼ 𝑝|𝑌 .
We then have by (115) that
x
x
(𝑔(𝑥) − 𝑔(𝑦))2 𝑝(𝑥)𝑃 (𝑥, 𝑦) 𝑑𝑥 𝑑𝑦 =
(𝑔(𝑥) − 𝑔(𝑦))2 𝑝(𝑥)𝑃tr (𝑥, 𝑦) 𝑑𝑥 𝑑𝑦
Ω×Ω
Ω×Ω
≥ 𝑒−2𝜀
x
(𝑔(𝑥) − 𝑔(𝑦))2 𝑝(𝑥)𝑃 tr (𝑋, 𝑌 )
Ω×Ω
= 𝑒−2𝜀
h
E
(116)
(117)
𝑝(𝑦)
𝑑𝑥 𝑑𝑦
𝑝(𝑌 )
[𝑔(𝑥) − 𝑔(𝑦)]2
(118)
(119)
(𝑥,𝑦)∼𝑅
−
X x
[𝑔(𝑥) − 𝑔(𝑦)]2 𝑝(𝑥)𝑃 rej (𝑋)
𝑋 𝑋×𝑋
i
𝑝(𝑦)
𝑑𝑥 𝑑𝑦 .
𝑝(𝑋)
(120)
We use the Pythagorean Theorem: letting ℬ be the 𝜎-algebra of Ω,
E[𝑔(𝑥) − 𝑔(𝑦)|𝒫 × ℬ] = E[𝑔|𝑋] − 𝑔(𝑦),
(121)
E[𝑔(𝑥) − 𝑔(𝑦)|𝒫 × 𝒫] = E[𝑔|𝑋] − E[𝑔|𝑌 ].
(122)
𝑝
𝑅
𝑝
𝑅
26
𝑝
Then
(120) = 𝑒−2𝜀
h
E
(𝑥,𝑦)∼𝑅
−
X x
[(E[𝑔|𝑋] − E[𝑔|𝑌 ])2 + (𝑔(𝑥) − E[𝑔|𝒫](𝑥))2 + (𝑔(𝑦) − E[𝑔|𝒫](𝑦))2 ]
𝑝
𝑝
[(𝑔(𝑥) − E[𝑔|𝒫](𝑥))2 + (𝑔(𝑦) − E[𝑔|𝒫](𝑦))2 ]𝑝(𝑋)𝑃 rej (𝑋)
𝑋 𝑋×𝑋
= 𝑒−2𝜀
h
i
𝑝(𝑦)
𝑑𝑥 𝑑𝑦
𝑝(𝑌 )
2
E [𝑔(𝑥) − 𝑔(𝑦)]
(123)
(124)
(125)
𝑥,𝑦∼𝑄
+
x
[(𝑔(𝑥) − E[𝑔|𝒫](𝑥))2 + (𝑔(𝑦) − E[𝑔|𝒫](𝑦))2 ]𝑝(𝑋)𝑃 tr (𝑋, 𝑌 )
𝑋×𝑌
Thus, using
𝑎′ +𝑏′
𝑎+𝑏
≥ min
¶
𝑎′ 𝑏′
𝑎, 𝑏
©
i
𝑝(𝑦)
𝑑𝑥 𝑑𝑦
𝑝(𝑌 )
(126)
for 𝑎′ , 𝑏′ , 𝑎, 𝑏 > 0, and the decompositions (112) and (126),
2
nE
E𝑥,𝑦∼𝑄 [𝑔(𝑥) − 𝑔(𝑦)]2
𝑋,𝑌 ∼𝑄 [(E𝑝 [𝑔|𝑋] − E𝑝 [𝑔|𝑌 ]) ]
−2𝜀
,
(127)
≥
𝑒
min
2 E𝑥∼𝑝 [𝑔(𝑥) − E𝑝 𝑔]2
2 E𝑥∼𝑝 [E𝑝 [𝑔|𝒫](𝑥) − E𝑝 𝑔]2
s
P
𝑝(𝑦)
2
2
𝑋,𝑌 ( 𝑋×𝑌 [(𝑔(𝑥) − E[𝑔|𝒫](𝑥)) + (𝑔(𝑦) − E[𝑔|𝒫](𝑦)) ]𝑝(𝑋)𝑃 tr (𝑋, 𝑌 ) 𝑝(𝑌 ) 𝑑𝑥 𝑑𝑦) o
.
2 E𝑥∼𝑝 [𝑔(𝑥) − E𝑝 [𝑔|𝒫](𝑥)]2
(128)
The first ratio in the minimum is at least Gap(𝑃 ) by (111). We now bound the second ratio (128).
The numerator of (128) is
≥ min 𝑃 acc (𝑋)
𝑋
E
(𝑥,𝑦)∼𝑅
[(𝑔(𝑥) − E[𝑔|𝒫](𝑥))2 + (𝑔(𝑦) − E[𝑔|𝒫](𝑦))2 ]
= 2 min 𝑃 acc (𝑋) E [𝑔(𝑥) − E[𝑔|𝒫](𝑥)]2 .
𝑥∼𝑝
𝑋
(129)
(130)
We claim that 𝑃 acc (𝑋) ≥ (1 − 𝜀) Gap(𝑀 ). Consider 𝑔(𝑌 ) = 1𝑋=𝑌 . Then
Gap(𝑃 ) ≤
E𝑋,𝑌 ∼𝑄 [𝑔(𝑋) − 𝑔(𝑌 )]2
2[E𝑥∼𝑝 [𝑔(𝑋)2 ] − [E𝑥∼𝑝 𝑔(𝑋)]2 ]
2𝑄(𝑋, 𝑋 𝑐 )
2[𝑝(𝑋) − 𝑝(𝑋)2 ]
𝑃 acc (𝑋)
≤
1 − 𝑝(𝑋)
𝑃 acc (𝑋)
≤
1−𝜀
≤
(131)
(132)
(133)
(134)
Putting everything together,
Gap(𝑀 |𝐾 ) ≥ 𝑒−2𝜀 min{Gap(𝑀 |𝐾 ), (1 − 𝜀) Gap(𝑀 |𝐾 )}.
(135)
Combined with Lemma 6.6 (Gap(𝑀 |𝐾 ) ≥ Gap(𝑀 |𝐾 )) and letting 𝜀 ↘ 0, this finishes the proof.
27
6.6
Restriction to large compact set
Finally we show it is OK to restrict to a large compact set. Intuitively this should be clear as the
gaussian density functions are highly concentrated around their means.
Lemma 6.11. Let 𝑝𝛽 (𝑥) ∝ 𝑒−𝛽𝑓¶(𝑥) where 𝑓 (𝑥) = ©− ln 𝑛𝑖=1 𝑤𝑖 𝑝𝑖 (𝑥) and 𝑝𝑖 (𝑥) is the pdf of a
gaussian with mean 𝜇𝑖 . Let 𝐵𝑅 = 𝑥 ∈ R𝑑 : ‖𝑥‖ ≤ 𝑅 .
For any 𝑇 > 0, for any 𝜀1 > 0, there exists 𝑅 such that for any 𝑟 ≥ 𝑅,
P
1. For any 𝑥 ∈ 𝐵𝑅 , 𝑃𝑇 (𝑥, 𝐵 𝑐 ) ≤ 𝑒−𝛽𝑇 /2 .
2. For any 𝑓 with Supp(𝑓 ) ⊆ 𝐵𝑅 ,
ℰ𝑃𝑇 |𝐵 (𝑔)
𝑅
Var𝑝|𝐵 (𝑔)
Ç
≥ (1 − 𝜀1 )
𝑅
ℰ𝑃𝑇 (𝑔)
− 𝑒−𝛽𝑇 /2
Var𝑝 (𝑔)
å
(136)
3. For all 𝑚, 𝜆𝑚 (𝑃𝑇 |𝐵𝑅 ) ≥ (1 − 𝜀1 )(𝜆𝑚 (𝑃𝑇 ) − 𝑒−𝛽𝑇 /2 ).
Note that we can improve the 𝑒−𝛽𝑇 /2 to arbitrary 𝜀2 > 0 with a more careful analysis using
martingale concentration bounds, but this weaker version suffices for us.
Proof. Let 𝜇 be such that ‖𝜇𝑖 ‖ ≤ 𝐷 for all 1 ≤ 𝑖 ≤ 𝑚. Let 𝑌𝑡 = ‖𝑋𝑡 ‖2 . By It^o’s Lemma,
¨
∂
√
𝑑𝑌𝑡 = 2 𝑋𝑡 , −𝛽∇𝑓 (𝑋𝑡 ) 𝑑𝑡 + 2 𝑑𝐵𝑡 + 2𝑑 · 𝑑𝑡
*
=
−2 𝑋𝑡 ,
≤
≤
−‖𝑋𝑡 −𝜇𝑖 ‖2 /2
𝑖=1 𝑤𝑖 (𝑋𝑡 − 𝜇𝑖 )𝑒
P𝑛
−‖𝑋𝑡 −𝜇𝑖 ‖2 /2
𝑖=1 𝑤𝑖 𝑒
P𝑛
!
+ 2𝑑 𝑑𝑡 +
−‖𝑋𝑡 −𝜇𝑖 ‖2
𝑖=1 𝑤𝑖 [(𝑋𝑡 − 𝜇) + (𝜇 − 𝜇𝑖 )]𝑒
−2
𝑋𝑡 ,
P𝑛
−‖𝑋𝑡 −𝜇𝑖 ‖2
𝑖=1 𝑤𝑖 𝑒
√
(−2𝛽𝑌𝑡 + 𝛽𝐷 ‖𝑋𝑡 ‖ + 2𝑑)𝑑𝑡 + 8(𝑋𝑡 − 𝜇)* 𝑑𝐵𝑡
Ç
å
√
𝐷2 𝛽
−𝛽𝑌𝑡 +
+ 2𝑑 𝑑𝑡 + 8(𝑋𝑡 − 𝜇)* 𝑑𝐵𝑡
*
=
−𝛽
+
𝛽
P𝑛
+
√
8(𝑋𝑡 − 𝜇)* 𝑑𝐵𝑡
!
+ 2𝑑 𝑑𝑡 +
√
8(𝑋𝑡 − 𝜇)* 𝑑𝐵𝑡
(137)
(138)
(139)
(140)
(141)
4
2
Let 𝐶 = 𝐷4 𝛽 + 2𝑑 and consider the change of variable 𝑍𝑡 = 𝑒𝛽𝑡 (𝑌𝑡 − 𝐶𝛽 ). Since this is linear in 𝑌𝑡 ,
It^o’s Lemma reduces to the usual change of variables and
Å
ã
√
𝐶
𝛽𝑡
𝑑𝑍𝑡 ≤ 𝛽𝑒
𝑌𝑡 −
𝑑𝑡 + 𝑒𝛽𝑡 ((−𝛽𝑌𝑡 + 𝐶)𝑑𝑡 + 8(𝑋𝑡 − 𝜇* )𝑑𝐵𝑡 )
(142)
𝛽
√ 𝛽𝑡
≤ 8𝑒 (𝑋𝑡 − 𝜇)* 𝑑𝐵𝑡 .
(143)
Suppose 𝑍0 is a point with norm ≤ 𝑅. By the martingale property of the It^o integral and Markov’s
inequality,
ï
Å
E 𝑒𝛽𝑇 ‖𝑋𝑇 ‖2 −
𝐶
𝛽
ãò
= E𝑍𝑇 ≤ 𝑍0 = ‖𝑋0 ‖2 −
𝐶
𝛽
𝐶
(1 − 𝑒−𝛽𝑇 )
𝛽
𝑒−𝛽𝑇 (𝑅2 + 𝐶𝛽 (1 − 𝑒−𝛽𝑇 ))
(144)
=⇒ E ‖𝑋𝑇 ‖2 ≤ 𝑒−𝛽𝑇 ‖𝑋0 ‖2 +
(145)
P(‖𝑋𝑇 ‖ ≥ 𝑅) ≤
(146)
𝑅2
≤ 𝑒−𝛽𝑇 /2
28
(147)
for all 𝑅 large enough. This shows the first part.
Note that the restricted 𝑃𝑇 |𝐵𝑅 operates on functions 𝑔 with Supp(𝑔) ⊆ 𝐵𝑅 as
𝑃𝑇 |𝐵𝑅 𝑔(𝑥) =
Z
𝐵𝑅
Z
=
𝐵𝑅
𝑃𝑇 |𝐵𝑅 (𝑥, 𝑑𝑦)𝑔(𝑥)
(148)
𝑐
𝑃𝑇 (𝑥, 𝑦)𝑔(𝑥) 𝑑𝑦 + 𝑃𝑇 (𝑥, 𝐵𝑅
)𝑔(𝑥)
(149)
𝑐
= 1𝐵𝑅 [𝑃𝑇 𝑔(𝑥) + 𝑃𝑇 (𝑥, 𝐵𝑅
)𝑔(𝑥)]
(150)
Without loss of generality we may assume that E𝑔 = 0. (This is unchanged by whether we take
𝑥 ∼ 𝑝 or 𝑥 ∼ 𝑝|𝐵𝑅 .) Then for 𝑅 large enough,
⟨𝑃𝑇 |𝐵𝑅 𝑔, 𝑔⟩𝑝 ≤ ⟨𝑃𝑇 𝑔, 𝑔⟩𝑝 + 𝑒−𝛽𝑇 /2 ‖𝑔‖2𝑝
⟨(𝐼 − 𝑃𝑇 |𝐵𝑅 )𝑔, 𝑔⟩𝑝|𝐵 ≥
⟨(𝐼 − 𝑃𝑇 |𝐵𝑅 )𝑔, 𝑔⟩𝑝
⟨(𝐼 − 𝑃𝑇 )𝑔, 𝑔⟩ − 𝑒−𝛽𝑇 /2 ‖𝑔‖2
𝑝(𝐵𝑅 )
ℰ𝑝𝑇 |𝐵 (𝑔)
𝑅
Var𝑝|𝐵
𝑅
(152)
𝑝(𝐵𝑅 )
𝑅
≥
(151)
1
≥
𝑝(𝐵𝑅 )
(𝑔)
Ç
(153)
å
ℰ𝑝𝑇 (𝑔)
− 𝑒−𝛽𝑇 /2 .
Var𝑝 (𝑔)
(154)
Taking 𝑅 large enough, 𝑝(𝐵1𝑅 ) can be made arbitrarily close to 1. The inequality for eigenvalues
follows from the variational characterization as in the proof of Lemma C.3.
7
Mixing at the highest temperature
Definition 7.1. For a function 𝑓 : R𝑑 → R ∪ {∞}, define the 𝛼-strongly convex envelope
SCE𝛼 [𝑓 ] : R𝑑 → R ∪ {∞} by
SCE𝛼 [𝑓 ] =
sup
𝑔(𝑥)
(155)
𝑔≤𝑓
𝑔 is 𝛼-s.c.
where “s.c.” is an abbreviation for strongly convex.
Proposition 7.2. SCE𝛼 [𝑓 ] is 𝛼-strongly convex.
Note we use the following definition of strongly convex, valid for any (not necessarily differentiable) function 𝑓 : R𝑑 → R,
∀𝑡 ∈ [0, 1],
1
𝑓 (𝑡𝑥 + (1 − 𝑡)𝑦) ≤ 𝑡𝑓 (𝑥) + (1 − 𝑡)𝑓 (𝑦) − 𝛼𝑡(1 − 𝑡) ‖𝑥 − 𝑦‖22
2
29
(156)
Proof. Let 𝑡 ∈ [0, 1]. We check
𝑡SCE𝛼 [𝑓 ](𝑥) + (1 − 𝑡)SCE𝛼 [𝑓 ](𝑦) = 𝑡
𝑔(𝑥) + (1 − 𝑡)
sup
𝑔≤𝑓
𝑔 is 𝛼-s.c.
≤
sup
sup
𝑔(𝑦)
(157)
𝑔≤𝑓
𝑔 is 𝛼-s.c.
[𝑡𝑔(𝑥) + (1 − 𝑡)𝑔(𝑦)]
(158)
1
[𝑔(𝑡𝑥 + (1 − 𝑡)𝑦) + 𝛼𝑡(1 − 𝑡) ‖𝑥 − 𝑦‖22 ]
2
(159)
𝑔≤𝑓
𝑔 is 𝛼-s.c.
≤
sup
𝑔≤𝑓
𝑔 is 𝛼-s.c.
≤ SCE𝛼 [𝑓 ](𝑡𝑥 + (1 − 𝑡)𝑦).
Ç
−
𝑖=1 𝑤𝑖 𝑒
P𝑛
Lemma 7.3. Let 𝑓 (𝑥) = − ln
‖𝑥−𝜇𝑖 ‖2
(160)
å
where 𝑤1 , . . . , 𝑤𝛼 > 0,
2
P𝑚
𝑖=1 𝑤𝑖
= 1, and
𝑤min = min1≤𝑖≤𝑚 𝑤𝑖 . Suppose ‖𝜇𝑖 ‖ ≤ 𝐷 for all 1 ≤ 𝑖 ≤ 𝑚.
Then there exists a 12 -strongly convex function 𝑔(𝑥) such that ‖𝑓 − 𝑔‖∞ ≤ 𝐷2 .
Proof. We show that 𝑔 = SCE 1 [𝑓 ] works. It is
2
that 𝑔(𝑥) ≤ 𝑓 (𝑥) − 𝐷2 .
Let 𝑥, 𝑦 ∈ R𝑑 . We have
𝑓 (𝑥) = − ln
𝑛
X
−
𝑤𝑖 𝑒
‖𝑥−𝜇𝑖 ‖2
!
= − ln
2
𝑖=1
≥ − ln
𝑛
X
Ç
𝑤𝑖 𝑒
𝑖=1
𝑛 Ç
X
𝑤𝑖 𝑒
1
2 -strongly
−
‖𝑦−𝜇𝑖 ‖2
−
‖𝑦−𝜇𝑖 ‖2
2
𝑖=1
2
−
convex by Proposition 7.2. We show
‖𝑥−𝑦‖2
+⟨𝑥−𝑦,𝜇𝑖 −𝑦⟩
2
å
å
(161)
‖𝑥 − 𝑦‖2
+ ⟨𝑥 − 𝑦, 𝜇𝑖 − 𝑦⟩
− max −
𝑖
2
(162)
ô
ñ
‖𝑥 − 𝑦‖2
‖𝑥 − 𝑦‖2
= 𝑓 (𝑦) +
+ ⟨𝑥 − 𝑦, 𝑦⟩ + min
− ⟨𝑥 − 𝑦⟩ , 𝜇𝑖
𝑖
4
4
(163)
ñ
î
ó
‖𝑥 − 𝑦‖2
+ ⟨𝑥 − 𝑦, 𝑦⟩ + min − ‖𝜇𝑖 ‖2
≥ 𝑓 (𝑦) +
𝑖
4
2
‖𝑥 − 𝑦‖
≥ 𝑓 (𝑦) +
+ ⟨𝑥 − 𝑦, 𝑦⟩ − 𝐷2
4
ô
(164)
(165)
The RHS is a 12 -strongly convex function in 𝑥 that equals 𝑓 (𝑦) − 𝐷2 at 𝑥 = 𝑦, and is ≤ 𝑓 (𝑥)
everywhere. Therefore, SCE 1 [𝑓 ](𝑦) ≥ 𝑓 (𝑦) − 𝐷2 .
2
Lemma 7.4. Keep the setup of Lemma 7.3. Then Langevin diffusion on 𝛽𝑓 (𝑥) satisfies a Poincaré
inequality with constant
16𝑒2𝛽𝐷
𝛽
2
.
Proof. Let 𝑔(𝑥) be as in Lemma 7.3. Since 𝛽𝑔(𝑥) is 𝛽2 -strongly convex, by Theorem A.12 it satisfies
2
a Poincaré inequality with constant 16
𝛽 . Now ‖𝛽𝑓 − 𝛽𝑔‖∞ ≤ 𝛽𝐷 , so by Lemma C.3, 𝛽𝑓 satisfies a
Poincaré inequality with constant
16 2𝛽𝐷2
.
𝛽 𝑒
30
8
Discretizing the continuous chains
As a notational convenience, in the section to follow we will denote 𝑥* = argmin𝑥∈R𝑑 𝑓˜(𝑥).
Lemma 8.1. Let 𝑝𝑡 , 𝑞 𝑡 : R𝑑 × [𝐿] → R be the distributions after running the simulated tempering
chain for 𝑡 steps, where in 𝑝𝑡 , for any temperature 𝑖 ∈ 𝐿, the Type 1 transitions are taken according
to the (discrete time) Markov kernel 𝑃𝑇 : running Langevin diffusion for time 𝑇 ; in 𝑞 𝑡 , the Type 1
transitions are taken according to running 𝑇𝜂 steps of the discretized Langevin diffusion, using 𝜂 as
the discretization granularity, s.t. 𝜂 ≤
KL(𝑝𝑡 ||𝑞 𝑡 ) .
𝜎2
2 .
Then,
𝜂2
𝜂
𝜂2 2
2
(𝐷
+
𝑑)𝑇
𝑡
+
max E𝑥∼𝑝0 (·,𝑖) ‖𝑥 − 𝑥* ‖22 + 4 𝑑𝑡𝑇
𝜎6
𝜎6 𝑖
𝜎
Before proving the above statement, we make a note on the location of 𝑥* to make sense of
max𝑖 E𝑥∼𝑝0 (·,𝑖) ‖𝑥 − 𝑥* ‖22 Namely, we show:
√
Lemma 8.2 (Location of minimum). Let 𝑥* = argmin𝑥∈R𝑑 𝑓˜(𝑥). Then, ‖𝑥* ‖ ≤ 2𝐷.
Proof. Since 𝑓˜(0) =
𝐷2
,
𝜎2
it follows that min𝑥∈R𝑑 𝑓˜(𝑥) ≤ −𝐷2 /𝜎 2 . However, for any 𝑥, it holds that
2
‖𝜇𝑖 − 𝑥‖
𝑓˜(𝑥) ≥ min
𝑖
𝜎2
2
‖𝑥‖ − max𝑖 ‖𝜇𝑖 ‖2
≥
𝜎2
2
‖𝑥‖ − 𝐷2
≥
𝜎2
Hence, if ‖𝑥‖ >
√
2𝐷, 𝑓˜(𝑥) > min𝑥∈R𝑑 𝑓˜(𝑥). This implies the statement of the lemma.
We prove a few technical lemmas. First, we prove that the continuous chain is essentially
contained in a ball of radius 𝐷. More precisely, we show:
Lemma 8.3 (Reach of continuous chain). Let 𝑃𝑇𝛽 (𝑋) be the Markov kernel corresponding to evolving Langevin diffusion
𝑑𝑋𝑡
= −𝛽∇𝑓˜(𝑋𝑡 ) + 𝑑𝐵𝑡
𝑑𝑡
with 𝑓˜ and 𝐷 are as defined in 2 for time 𝑇 . Then,
E[‖𝑋𝑡 − 𝑥* ‖2 ] ≤ E[‖𝑋0 − 𝑥* ‖2 ] + (4𝛽𝐷2 + 2𝑑)𝑇
Proof. Let 𝑌𝑡 = ‖𝑋𝑡 − 𝑥* ‖2 . By It^
os Lemma, we have
‖𝑋𝑡 −𝜇𝑖 ‖2
𝜎2
𝑛
X
𝑤𝑖 𝑒−
*
𝑑𝑌𝑡 = −2⟨𝑋𝑡 − 𝑥 , 𝛽
P𝑛
𝑖=1
−
𝑖=1 𝑤𝑖 𝑒
(𝑋𝑡 − 𝜇𝑖 )
‖𝑋𝑡 −𝜇𝑖 ‖2
𝜎2
31
𝑑
√ X
⟩ + 2𝑑 𝑑𝑡 + 8 (𝑋𝑡 )𝑖 𝑑(𝐵𝑖 )𝑡
𝑖=1
(166)
We claim that
−⟨𝑋𝑡 − 𝑥* , 𝑋𝑡 − 𝜇𝑖 ⟩ ≤
𝐷
2
Indeed,
−⟨𝑋𝑡 − 𝑥* , 𝑋𝑡 − 𝜇𝑖 ⟩ ≤ −‖𝑋𝑡 ‖2 + ‖𝑋𝑡 ‖(‖𝜇𝑖 ‖ + ‖𝑥* ‖) + ‖𝑥* ‖‖𝜇𝑖 ‖
≤ 4𝐷2
where the last inequality follows from ‖𝜇𝑖 | ≤ 𝐷 and Lemma 8.2 Together with (166), we get
𝑑
√ X
𝑑𝑌𝑡 ≤ 𝛽4𝐷2 + 2𝑑 𝑑𝑡 + 8 (𝑋𝑡 )𝑖 𝑑(𝐵𝑖 )𝑡
𝑖=1
Integrating, we get
√ Z
𝑌𝑡 ≤ 𝑌0 + 𝛽4𝐷 𝑇 + 2𝑑𝑇 + 8
2
0
𝑇
𝑑
X
(𝑋𝑡 )𝑖 𝑑(𝐵𝑖 )𝑡
𝑖=1
Taking expectations and using the martingale property of the It^o integral, we get the claim of the
lemma.
Next, we prove a few technicall bound the drift of the discretized chain after 𝑇 /𝜂 discrete steps.
The proofs follow similar calculations as those in [Dal16].
We will first need to bound the Hessian of 𝑓˜.
Lemma 8.4 (Hessian bound).
∇2 𝑓˜(𝑥) ⪯
2
𝐼, ∀𝑥 ∈ R𝑑
𝜎2
Proof. For notational convenience, let 𝑝(𝑥) = 𝑛𝑖=1 𝑤𝑖 𝑒−𝑓𝑖 (𝑥) , where 𝑓𝑖 (𝑥) =
R
𝑍 = R𝑑 𝑒−𝑓𝑖 (𝑥) 𝑑𝑥. Note that 𝑓˜(𝑥) = − log 𝑝(𝑥). The Hessian of 𝑓˜ satisfies
P
∇2 𝑓˜ =
P
𝑖 𝑤𝑖 𝑒
−𝑓𝑖 ∇2 𝑓
𝑖
𝑝
⪯ max ∇2 𝑓𝑖 ⪯
𝑖
−
1
2
P
𝑖,𝑗
(𝑥−𝜇𝑖 )2
𝜎2
+ log 𝑍 and
𝑤𝑖 𝑤𝑗 𝑒−𝑓𝑖 𝑒−𝑓𝑗 (∇𝑓𝑖 − ∇𝑓𝑗 )⊗2
𝑝2
2
𝐼
𝜎2
as we need.
Lemma 8.5 (Bounding interval drift). In the setting of this section, let 𝑥 ∈ R𝑑 , 𝑖 ∈ [𝐿], and let
𝜎2
𝜂 ≤ 2𝛼
.
ä 𝑑𝑇 𝜂
4𝜂 2 Ä
”
KL(𝑃𝑇 (𝑥, 𝑖)||𝑃
(𝑥,
𝑖))
≤
‖𝑥 − 𝑥* ‖22 ) + 2𝑇 𝑑 + 4
𝑇
6
3𝜎
𝜎
32
d
Proof. Let 𝑥𝑗 , 𝑖 ∈ [0, 𝑇 /𝜂 − 1] be a random variable distributed as 𝑃
𝜂𝑗 (𝑥, 𝑖). By Lemma 2 in [Dal16]
and Lemma 8.4 , we have
”
KL(𝑃𝑇 (𝑥, 𝑖)||𝑃
𝑇 (𝑥, 𝑖)) ≤
𝑇 /𝜂−1
𝑑𝑇 𝜂
𝜂3 X
E[‖∇𝑓 (𝑥𝑘 )‖22 ] + 4
4
3𝜎 𝑘=0
𝜎
Similarly, the proof of Corollary 4 in [Dal16] implies that
𝑇 /𝜂−1
𝜂
X
E[‖∇𝑓 (𝑥𝑘 )‖22 ] ≤
𝑘=0
4
8𝑇 𝑑
‖𝑥 − 𝑥* ‖22 + 2
𝜎2
𝜎
Finally, we prove a convenient decomposition theorem for the KL divergence of two mixtures
of distributions, in terms of the KL divergence of the weights and the components in the mixture.
Concretely:
Lemma 8.6. Let 𝑤, 𝑤′ : 𝐼 → R be distributions over a domain 𝐼 with full support. Let 𝑝𝑖 , 𝑞𝑖 : ∀𝑖 ∈ 𝐼
be distributions over an arbitrary domain. Then:
ÅZ
KL
𝑤𝑖 𝑝𝑖 ||
Z
𝑖∈𝐼
𝑖∈𝐼
ã
𝑤𝑖′ 𝑞𝑖 ≤ KL(𝑤||𝑤′ ) +
Z
𝑤𝑖 KL(𝑝𝑖 ||𝑞𝑖 )
𝑖∈𝐼
Proof. Overloading notation, we will use 𝐾𝐿(𝑎||𝑏) for two measures 𝑎, 𝑏 even if they are not necessarily probability distributions, with the obvious definition.
ÅZ
KL
𝑖∈𝐼
𝑤𝑖 𝑝𝑖 ||
Z
𝑖∈𝐼
𝑤𝑖′ 𝑞𝑖
ã
𝑤′
𝑤𝑖 𝑝𝑖 ||
𝑤𝑖 𝑞𝑖 𝑖
= KL
𝑤𝑖
𝑖∈𝐼
𝑖∈𝐼
Ç
å
Z
′
𝑤
≤
𝑤𝑖 KL 𝑝𝑖 ||𝑞𝑖 𝑖
𝑤𝑖
𝑖∈𝐼
Ç å
Z
𝑤𝑖
=
𝑤𝑖 log
+ KL(𝑝𝑖 ||𝑞𝑖 )
𝑤𝑖′
𝑖∈𝐼
ÇZ
å
Z
′
Z
= KL(𝑤||𝑤 ) +
𝑤𝑖 KL(𝑝𝑖 ||𝑞𝑖 )
𝑖∈𝐼
where the first inequality holds due to the convexity of KL divergence.
With this in mind, we can prove the main claim:
Proof of 8.1. Let’s denote by 𝑃𝑇 (𝑥, 𝑖) : R𝑑 × [𝐿] → R, ∀𝑥 ∈ R𝑑 , 𝑖 ∈ [𝐿] the distribution on R𝑑 × [𝐿]
corresponding to running the Langevin diffusion chain for 𝑇 time steps on the 𝑖-th coordinate,
”
starting at 𝑥 × {𝑖}, and keeping the remaining coordinates fixed. Let us define by 𝑃
𝑇 (𝑥, 𝑖) :
𝑑
R × [𝐿] → R the analogous distribution, except running the discretized Langevin diffusion chain
for 𝑇𝜂 time steps on the 𝑖-th coordinate.
33
Let’s denote by 𝑅 (𝑥, 𝑖) : R𝑑 × [𝐿] → R the distribution on R𝑑 × [𝐿], running the Markov
transition matrix corresponding to a Type 2 transition in the simulated tempering chain, starting
at (𝑥, 𝑖).
We will proceed by induction. Towards that, we can obviously write
𝑝
𝑡+1
1
=
2
𝐿−1
X
!
1
𝑝 (𝑥, 𝑖)𝑃𝑇 (𝑥, 𝑖) +
2
𝑖=0
Z
𝑥∈R𝑑
𝐿−1
X
Z
𝑡
!
𝑡
𝑝 (𝑥, 𝑖)𝑅(𝑥, 𝑖)
𝑥∈R𝑑 𝑖=0
and similarly
𝑞
𝑡+1
1
(𝑥, 𝑖) =
2
𝐿−1
X
Z
!
𝑡
”
𝑞 (𝑥, 𝑖)𝑃
𝑇 (𝑥, 𝑖)
𝑥∈R𝑑 𝑖=0
1
+
2
𝐿−1
X
Z
!
𝑡
𝑞 (𝑥, 𝑖)𝑅(𝑥, 𝑖)
𝑥∈R𝑑 𝑖=0
(Note: the 𝑅 transition matrix doesn’t change in the discretized vs continuous version.)
By convexity of KL divergence, we have
KL(𝑝
𝑡+1
𝑡+1
||𝑞
1
) ≤ KL
2
1
+ KL
2
𝐿−1
X
Z
𝑥∈R𝑑 𝑖=0
Z
𝐿−1
X
Z
𝑡
𝑝 (𝑥, 𝑖)𝑃𝑇 (𝑥, 𝑖)||
𝑥∈R𝑑 𝑖=0
𝐿−1
X
Z
𝑡
𝑝 (𝑥, 𝑖)𝑅(𝑥, 𝑖)||
𝑥∈R𝑑 𝑖=0
𝐿−1
X
!
𝑡
”
𝑞 (𝑥, 𝑖)𝑃
𝑇 (𝑥, 𝑖)
!
𝑡
𝑞 (𝑥, 𝑖)𝑅(𝑥, 𝑖)
𝑥∈R𝑑 𝑖=0
By Lemma 8.6, we have that
𝐿−1
X
Z
KL
𝐿−1
X
Z
𝑡
𝑝 (𝑥, 𝑖)𝑅(𝑥, 𝑖)||
𝑥∈R𝑑 𝑖=0
!
𝑡
𝑞 (𝑥, 𝑖)𝑅(𝑥, 𝑖)
≤ KL(𝑝𝑡 ||𝑞 𝑡 )
𝑥∈R𝑑 𝑖=0
Similarly, by Lemma 8.5 together with Lemma 8.6 we have
𝐿−1
X
Z
KL
𝑥∈R𝑑 𝑖=0
KL(𝑝𝑡 ||𝑞 𝑡 ) +
Z
𝑡
𝑝 (𝑥, 𝑖)𝑃𝑇 (𝑥, 𝑖)||
𝐿−1
X
𝑥∈R𝑑 𝑖=0
!
”
𝑞 (𝑥, 𝑖)𝑃𝑇 (𝑥, 𝑖) ≤
𝑡
Å
4𝜂 2
3𝜎 6
ã
max E𝑥∼𝑝𝑡 (·,𝑖) ‖𝑥 − 𝑥* ‖22 + 2𝑇 𝑑 +
𝑖
𝑑𝑇 𝜂
𝜎4
By Lemmas 8.3 and 8.2, we have that for any 𝑖 ∈ [0, 𝐿 − 1],
E𝑥∼𝑝𝑡 (·,𝑖) ‖𝑥 − 𝑥* ‖22 ≤ E𝑥∼𝑝𝑡−1 (·,𝑖) ‖𝑥‖2 + (4𝐷2 + 2𝑑)𝑇
Hence, inductively, we have E𝑥∼𝑝𝑡 (·,𝑖) ‖𝑥 − 𝑥* ‖22 ≤ E𝑥∼𝑝0 (·,𝑖) ‖𝑥 − 𝑥* ‖22 + (4𝐷2 + 2𝑑)𝑇 𝑡
Putting together, we have
KL(𝑝
𝑡+1
||𝑞
𝑡+1
4𝜂 2
𝑑𝑇 𝜂
) ≤ KL(𝑝 ||𝑞 ) + 6 max E𝑥∼𝑝0 (·,𝑖) ‖𝑥 − 𝑥* ‖22 + (4𝐷2 + 2𝑑)𝑇 𝑡 + 2𝑇 𝑑 + 4
𝑖
3𝜎
𝜎
𝑡
𝑡
Å
ã
By induction, we hence have
KL(𝑝𝑡 ||𝑞 𝑡 ) .
𝜂2 2
𝜂2
𝜂
2
(𝐷
+
𝑑)𝑇
𝑡
+
max E𝑥∼𝑝0 (·,𝑖) ‖𝑥 − 𝑥* ‖22 + 4 𝑑𝑡𝑇 𝜂
𝜎6
𝜎6 𝑖
𝜎
as we need.
34
9
Proof of main theorem
Before putting everything together, we show how to estimate the partition functions. We will apply
the following to 𝑔1 (𝑥) = 𝑒−𝛽ℓ 𝑓 (𝑥) and 𝑔2 (𝑥) = 𝑒−𝛽ℓ+1 𝑓 (𝑥) .
Lemma 9.1 (Estimating the partition function to within a constant factor). Suppose that 𝑝1 (𝑥) =
𝑔2 (𝑥)
𝑔1 (𝑥)
e
𝑍1 and 𝑝2 (𝑥) = 𝑍2 are probability distributions on Ω. Suppose 𝑝1 is a distribution such that
𝑑𝑇 𝑉 (𝑝e1 , 𝑝1 ) <
variable
𝜀
,
2𝐶 2
and
𝑔2 (𝑥)
𝑔1 (𝑥)
∈ [0, 𝐶] for all 𝑥 ∈ Ω. Given 𝑛 samples from 𝑝e1 , define the random
𝑛
1X
𝑔2 (𝑥𝑖 )
𝑟=
.
𝑛 𝑖=1 𝑔1 (𝑥𝑖 )
(167)
Let
𝑟= E
𝑥∼𝑝1
and suppose 𝑟 ≥
1
𝐶.
𝑔2 (𝑥)
𝑍2
=
𝑔1 (𝑥)
𝑍1
(168)
𝑛𝜀2
Then with probability ≥ 1 − 𝑒− 2𝐶 4 ,
𝑟
− 1 ≤ 𝜀.
𝑟
(169)
𝑔2 (𝑥)
𝜀
𝑔2 (𝑥)
− E
.
≤ 𝐶𝑑𝑇 𝑉 (𝑝e1 , 𝑝1 ) ≤
2𝐶
𝑥∼𝑝
e1 𝑔1 (𝑥) 𝑥∼𝑝1 𝑔1 (𝑥)
(170)
Proof. We have that
E
The Chernoff bound gives
𝑔2 (𝑥)
𝜀
≥
𝑔
(𝑥)
2𝐶
𝑥∼𝑝
e1 1
!
𝑟− E
P
2
𝜀
)
( 2𝐶
2
𝑛𝜀2
2( 𝐶
2)
= 𝑒− 2𝐶 4 .
≤𝑒
−
𝑛
(171)
Combining (170) and (171) using the triangle inequality,
𝑛𝜀2
1
P |𝑟 − 𝑟| ≥ 𝐶 ≤ 𝑒− 2𝐶 4 .
𝜀
Å
Dividing by 𝑟 and using 𝑟 ≥
1
𝐶
ïZ
R
R𝑑
(172)
gives the result.
Lemma 9.2. Suppose that 𝑓 (𝑥) = − ln
𝑝𝛼 (𝑥) ∝ 𝑒−𝛼𝑓 (𝑥) , 𝑍𝛼 =
If 𝛼 < 𝛽, then
ã
ñ
P𝑛
𝑖=1 𝑤𝑖 𝑒
−
‖𝑥−𝜇𝑖 ‖2
ô
2
, 𝑝(𝑥) ∝ 𝑒−𝑓 (𝑥) , and for 𝛼 ≥ 0 let
𝑒−𝛼𝑓 (𝑥) 𝑑𝑥. Suppose that ‖𝜇𝑖 ‖ ≤ 𝐷 for all 𝑖.
ò
min{𝑝𝛼 (𝑥), 𝑝𝛽 (𝑥)} 𝑑𝑥 /𝑝𝛽 (𝐴) ≥ min
𝑥
𝐴
𝑍𝛽
𝑝𝛼 (𝑥)
≥
𝑝𝛽 (𝑥)
𝑍𝛼
𝑍𝛽 1 −2(𝛽−𝛼)
∈
𝑒
𝑍𝛼
2
35
𝐷+ √1𝛼
(173)
√
𝑑+
q Ä
ln
2
𝑤min
ä2
, 1 .
(174)
!
1 Ä
1
𝑑
+𝛼
ln 𝑤 1
𝐷2 + 𝛼
Choosing 𝛽 − 𝛼 = 𝑂
ä , this quantity is Ω(1).
min
2
−𝛼‖𝑥−𝜇𝑖 ‖ /2 .
Proof. Let 𝑝e𝛼 (𝑥) ∝ 𝑛𝑖=1 𝑤𝑖 𝑒q
Ä
ä
√
2
Let 𝐶 = 𝐷 + √1𝛼
𝑑 + 2 ln 𝑤min
. Then by Lemma 6.2,
P
1
P
(‖𝑥‖ ≥ 𝐶)
𝑤min 𝑥∼𝑝e𝛼
𝑛
1 X
𝑤𝑖 P𝑥∼e𝑔𝛼 (‖𝑥‖ ≥ 𝐶)
≤
𝑤min 𝑖=1
P𝑥∼𝑝 (‖𝑥‖ ≥ 𝐶) ≤
≤
1
𝑤min
𝑛
X
𝑖=1
𝑤𝑖 P
Ä
𝑥∼𝑁 0, √1𝛼 𝐼𝑑
=
ñ
(176)
ä (‖𝑥‖2 ≥ (𝐶 − 𝐷)2 )
1
‖𝑥‖2 ≥
P
𝑤min 𝑥∼𝑁 (0,𝐼𝑑 )
1
(175)
√
Ç
Å
𝑑+
Ç
2
2 ln
2
𝜀𝑤min
ïZ
𝐴
ò
ãå2
(178)
2
2
𝑑 ln
+ ln
𝜀𝑤min
𝜀𝑤min
Å
P
‖𝑥‖ ≥ 𝑑 + 2
𝑤min 𝑥∼𝑁 (0,𝐼𝑑 )
1 𝑤min
1
≤
=
𝑤min 2
2
√
using the 𝜒2𝑑 tail bound P𝑦∼𝜒2 (𝑦 ≥ 2( 𝑑𝑥 + 𝑥)) ≤ 𝑒−𝑥 from [LM00].
𝑑
Thus, using 𝑓 (𝑥) ≥ 0,
≤
(177)
ã
Å
ãåô
(180)
.
𝑝𝛼 (𝑥)
, 1}𝑝𝛽 (𝑥) 𝑑𝑥 𝑝𝛽 (𝐴)
𝑝𝛽 (𝑥)
𝐴
Z
.
𝑍𝛽
≥
min{ 𝑒(−𝛽+𝛼)𝑓 (𝑥) , 1}𝑝𝛽 (𝑥) 𝑑𝑥 𝑝𝛽 (𝐴)
𝑍𝛼
𝐴
𝑍𝛽
≥
𝑍𝛼
R −𝛽𝑓 (𝑥)
𝑒
𝑑𝑥
= R −𝛼𝑓 (𝑥)
𝑒
𝑑𝑥
min{𝑝𝛼 (𝑥), 𝑝𝛽 (𝑥)} 𝑑𝑥 /𝑝𝛽 (𝐴) ≥
Z
min{
Z
=
≥
(179)
𝑑
ZR
𝑒(−𝛽+𝛼)𝑓 (𝑥) 𝑝𝛼 (𝑥) 𝑑𝑥
‖𝑥‖≤𝐷+ √1𝛼
1
≥ 𝑒
2
√
−(𝛽−𝛼) max
1 −2(𝛽−𝛼)
≥ 𝑒
2
36
𝑑+
q
2 ln
Ä
‖𝑥‖≤𝐷+ √2
𝛼
(182)
(183)
(184)
(185)
2
𝑤min
ä 𝑒(−𝛽+𝛼)𝑓 (𝑥) 𝑝𝛼 (𝑥) 𝑑𝑥
Ä√ »
𝑑+
ln 𝑤 2
min
q Ä
ä 2
√
𝐷+ √1𝛼 ( 𝑑+
(181)
ln
(
2
𝑤min
)
(186)
ä (𝑓 (𝑥))
)
(187)
(188)
Lemma 9.3. If 𝑝 =
P𝑛
𝑖=1 𝑤𝑖 𝑝𝑖
where 𝑝𝑖 are probability distributions and 𝑤𝑖 > 0 sum to 1, then
𝜒2 (𝑝||𝑞) ≤
𝑛
X
𝑤𝑖 𝜒2 (𝑝𝑖 ||𝑞).
(189)
𝑖=1
Proof. We calculate
𝜒2 (𝑝||𝑞) =
𝑛
X
𝑞(𝑥)2
𝑑𝑥 − 1
𝑖=1 𝑤𝑖 𝑝𝑖 (𝑥)
𝑖=1
≤
𝑛
X
Z
𝑤𝑖
!Ç
X
𝑖=1
=
(190)
P𝑛
𝑛
X
𝑞(𝑥)2
𝑖𝑛𝑤𝑖
𝑝𝑖 (𝑥)
𝑞(𝑥)2
𝑑𝑥 − 1
𝑝𝑖 (𝑥)
å
ÇZ
𝑤𝑖
𝑖=1
=
å
𝑑𝑥 − 1
𝑛
X
(191)
𝑤𝑖 𝜒2 (𝑝𝑖 ||𝑞).
(192)
𝑖=1
Lemma 9.4. Suppose that Algorithm 1 is run on temperatures 0 < 𝛽1 < · · · < 𝛽ℓ ≤ 1, ℓ ≤ 𝐿 with
”1 , . . . , 𝑍
cℓ satisfying
partition function estimates 𝑍
c𝑖
”1
𝑍
𝑍
1
−
≤ 1+
𝑍𝑖 𝑍1
𝐿
Å
ã𝑖−1
(193)
for all 1 ≤ 𝑖 ≤ ℓ and with parameters satisfying
𝛽1 = 𝑂
Ç 2å
𝜎
(194)
𝐷2
Ñ
𝛽𝑖 − 𝛽𝑖−1 = 𝑂
Ä
𝐷2 𝑑 + ln
Å
𝑇 = Ω 𝐷2 ln
Ö
Ä
1
Å
(195)
ää
1
𝑤min
ãã
(196)
𝑤min
𝐷8 𝑑4 + ln
𝑡=Ω
Ä
ä4
1
𝑤min
𝜀𝜎 2
𝑑𝑡𝑇
Ç
ln
4
𝜎 8 𝑤min
Ç
𝜂=𝑂
é
𝜎2
1 𝐷2 𝑑 ln (1/𝑤min )
𝜖
𝜎 8 𝑤min
å
Ç
max
𝐷2
𝑚16
,
𝜎 2 ln(1/𝑤min )4
è
å
(197)
å
.
(198)
Ä
Ä
2
ä
ä
Let 𝑞 0 be the distribution 𝑁 0, 𝜎𝛽1 , 1 on R𝑑 × [ℓ]. The distribution 𝑞 𝑡 after running for 𝑡 steps
satisfies 𝑝 − 𝑞 𝑡 1Ä≤ ä𝜀.
Ä Ä ää
Setting 𝜀 = 𝑂 𝐿1 above and taking 𝑚 = Ω ln 1𝛿 samples, with probability 1 − 𝛿 the estimate
Ñ
“ℓ+1 = 𝑍
“ℓ
𝑍
𝑚
1 X
𝑒(−𝛽ℓ+1 +𝛽ℓ )𝑓𝑖 (𝑥𝑗 )
𝑚 𝑗=1
also satisfies (238).
37
é
(199)
Proof. First consider the case 𝜎 = 1.
Consider simulated tempering 𝑀st |𝐵𝑅 : the type 1 transitions are running continuous Langevin
for time 𝑇 , and if a type 1 transition would leave 𝐵𝑅 , then instead stay at the same location. Let
𝑝𝑡 |𝐵𝑅 be the distribution of 𝑀𝑠𝑡 |𝐵𝑅 after 𝑡 steps starting from 𝑝0 . By the triangle inequality,
𝑝 − 𝑞𝑡
1
≤ ‖𝑝 − 𝑝|𝐵𝑅 ‖1 + 𝑝|𝐵𝑅 − 𝑝𝑡 |𝐵𝑅
1
+ 𝑝𝑡 |𝐵𝑅 − 𝑝𝑡
+ 𝑝𝑡 − 𝑞 𝑡
1
1
(200)
Note that ‖𝑝 − 𝑝|𝐵𝑅 ‖1 and 𝑝𝑡 |𝐵𝑅 − 𝑝𝑡 1 approach 0 as 𝑅 → ∞, so we concentrate on the other
two terms.
Let 𝑀𝑖 be the chain at inverse temperature 𝛽𝑖 . Consider first 𝑖 ≥ 2. By Lemma 6.11, for any
𝜀 > 0 we can choose 𝑅 such that for all 𝑥 ∈ 𝐵𝑅 , 𝑃𝑇 (𝑥, 𝐵 𝑐 ) ≤ 𝑒−𝛽𝑇 /2 and
𝜆𝑛 (𝐼 − 𝑃𝑇 |𝐵𝑅 ) ≥ (1 − 𝜀)(𝜆𝑛 (𝐼 − 𝑃𝑇 ) − 𝑒−𝛽𝑇 /2 )
Ä
For 𝑇 = Ω ln
Ä
1
𝑤min
ä
ä
𝐷2 , we have 𝑒−𝛽𝑇 /2 = 𝑜
Ä
𝑇 𝑤min
𝐷2
ä
. Thus we can ensure
3
𝑇 𝑤min
𝜆𝑛 (𝐼 − 𝑃𝑇 |𝐵𝑅 ) ≥
𝜆𝑛 (𝐼 − 𝑃𝑇 ) − 𝑜
4
𝐷2
Å
Å
ãã
By Theorem A.12, a Poincaré inequality holds for 𝛽𝑖 𝑓𝑗 with constant
P
the generator for Langevin on 𝑔e𝛽𝑖 (𝑥) = 𝑛𝑗=1 𝑤𝑖 𝑒−𝛽𝑖 𝑓𝑗 (𝑥) ,
𝜆𝑛+1 (L ) ≥ 𝑤min 𝜆𝑛+1 (L ′ )
ã
Å
𝑤min
≥Ω
𝐷2
(201)
(202)
≤ 𝑂(𝐷2 ). Letting L ′ be
8
𝛽𝑖
by Lemma 6.2 and Lemma C.3
(203)
by Lemma 6.1
(204)
Ä
2
𝑤2
ä
By Lemma 6.7 on 𝑝𝛽 , a Poincaré inequality holds with constant 𝑂 𝑤𝐷min . on sets of size ≤ min
2 .
By Lemma 6.10, for each 𝑖, we can choose a partition 𝒬𝑖 of 𝐵𝑅 such that for every compact 𝐾
consisting of a union of sets in 𝒬𝑖 ,
Gap(𝑀𝑖 |𝐾 ) ≥ (1 − 𝜀) Gap(𝑀𝑖 |𝐾
All the conditions of Lemma 6.8 are satisfied with 𝐶 = 𝑂
Ä Ä
is a Ω
ä
Ä
ää
, 𝑂 𝑇 𝑤𝐷min
2
2
𝑤min
at least 4
2 (ln 1 )2 𝑚8
𝑤min
𝑤min
Ä
𝒬𝑖
).
𝐷2
𝑤min
(205)
ä
. We obtain a partition 𝒫𝑖 that
-clustering under the projected chain with each set in the
partition having measure
under 𝑝𝑖 . By having chose the partition fine enough, and by
Cheeger’s inequality A.9, for each set 𝐴 in the 𝒫𝑖 ,
1
(ln 𝑤min
)4
𝒫𝑖
Gap(𝑀𝑖 |𝐴 ) ≥ (1 − 𝜀) Gap(𝑀𝑖 |𝐴 ) ≥ Ω
𝑚16
!
.
(206)
For the highest temperature, by Lemma 7.4, we have
Gap(𝑀1 ) = Ω 𝛽1 𝑒−2𝛽1 𝐷
Ä
38
2
ä
Å
=Ω
1
.
𝐷2
ã
(207)
By Lemma 9.2, since the condition on 𝛽𝑖 − 𝛽𝑖−1 is satisfied, 𝛿((𝒫𝑖 )ℓ𝑖=1 ) = Ω(1). By assumption
c𝑖 , 𝑟 = Ω(1) (𝑟 is defined in Assumption 5.1). By Theorem 5.2, the spectral gap of the simulated
on 𝑍
tempering chain is
𝑟4 𝛿 2 𝑝2min
𝐺 := Gap(𝑀st ) = Ω
min
ℓ4
1
)4
(ln 𝑤min
(
!
𝑚16
1
, 2
𝐷
)!
Ñ
𝑤4
= min
min
ℓ4
1
ln 𝑤min
4
é
𝑚16
1
, 2 .
𝐷
(208)
ln
Choosing 𝑡 = Ω
𝑝 − 𝑞𝑡
1
Ä
ℓ
𝜀𝑤min
ä!
, we get by Cauchy-Schwarz and (225) that
𝐺
≤ 𝜒2 (𝑝||𝑞 𝑡 ) ≤ (1 − 𝐺)𝑡 𝜒2 (𝑝||𝑞 0 ) ≤ 𝑒−𝐺𝑡 𝜒2 (𝑝||𝑞 0 ) = 𝑂
Å
𝜀𝑤min
𝜒2 (𝑝||𝑞 0 )
ℓ
ã
(209)
1
(Note that 𝐺 < 2 − 𝜆max because the chain is somewhat lazy; it stays with probability 2ℓ
.) To
2
2
2
2
2
2
‖𝜇‖
/𝜎
calculate 𝜒 (𝑞||𝑝0 ), first note the 𝜒 distance between 𝑁 (0, 𝜎 𝐼𝑑 ) and 𝑁 (𝜇, 𝜎 𝐼𝑑 ) is ≤ 𝑒
:
𝜒2 (𝑁 (0, 𝜎 2 𝐼𝑑 ), 𝑁 (𝜇, 𝜎 2 𝐼𝑑 )) =
≤
1
Z
𝑑
2
(2𝜋𝜎 2 ) R𝑑
Z
1
(2𝜋𝜎 2 )
𝑑
2
𝑒2(−
𝑒(−
‖𝑥−𝜇‖2
‖𝑥‖2
)+ 2
2𝜎 2
2𝜎
𝑑𝑥 − 1
‖𝑥‖2
+2⟨𝑥,𝜇⟩−2‖𝜇‖2 )/𝜎 2
2
(210)
2
𝑒‖𝜇‖
/𝜎 2
2
𝑑𝑥 = 𝑒‖𝜇‖
/𝜎 2
.
(211)
R𝑑
Then by Lemma 9.3 and Lemma 6.2,
1
𝐼𝑑
𝑤min
𝛽1
Å
Å Å
ãX
ãã
𝑚
ℓ
1
1
2
=𝑂
𝜒 𝑁 𝜇𝑖 , 𝐼𝑑 )||𝑁 (0, 𝐼𝑑
𝑤min 𝑖=1
𝛽1
𝛽1
𝜒2 (𝑝||𝑞 0 ) ≤ 𝑂
ℓ
Å
ã
2
𝑒 𝐷 𝛽1 ℓ
𝑤min
=𝑂
Å
ãã
Å
𝜒2 𝑝e𝛽 ||𝑁 0,
!
Å
=𝑂
ℓ
𝑤min
(212)
(213)
ã
.
(214)
Together with (252) this gives 𝑝 − 𝑞 𝑡 1 ≤ 3𝜀 .
For the term 𝑝𝑡 − 𝑞 𝑡 1 , use Pinsker’s inequality and Lemma 8.1 to get
𝑝𝑡 − 𝑞 𝑡
1
≤
»
2KL(𝑝𝑡 ||𝑞 𝑡 )
(215)
Ä
ä
= 𝑂 𝜂 2 [(𝐷2 + 𝑑)𝑇 𝑡2 + 𝐷2 ] + 𝜂𝑑𝑡𝑇 ≤
Å
ß
for 𝜂 = 𝑂 𝜀 min
1 √
, 1
𝑇 𝑡(𝐷+ 𝑑) 𝑑𝑡𝑇
𝑞 𝑡 1 ≤ 𝜀.
™ã
√
𝜀
3
(216)
𝜀
= 𝑂( 𝑑𝑡𝑇
).
This gives 𝑝 −
Ä ä
Ä ä
1
For the second part, setting 𝜀 = 𝑂 ℓ𝐿
gives that 𝑝𝑙 − 𝑞𝑙𝑡 = 𝑂 𝐿1 . By Lemma 9.1,
Ä
Ä ää
noting Lemma 9.2 gives 𝐶 = 𝑂(1), after collecting 𝑛 = Ω 𝐿2 ln 1𝛿
samples, with probability ≥ 1 − 𝛿,
‘
𝑍
ℓ+1
∈
“1
𝑍
hÄ
1−
‘
“ℓ
𝑍
ℓ+1 /𝑍
𝑍ℓ+1 /𝑍ℓ
ä Ä
1 ℓ
, 1
𝐿
+
−1 ≤
ä
1 ℓ
𝐿
i
1
𝐿.
’
c
Set 𝑍
ℓ+1 = 𝑟 𝑍ℓ .
𝑍ℓ+1
𝑍1 .
39
Then
‘
𝑍
ℓ+1
“ℓ
𝑍
∈ [1 −
1
𝐿, 1
+
1 𝑍ℓ+1
𝐿 ] 𝑍ℓ
and
Now consider general 𝜎. We can transform the problem to the problem where 𝜎 = 1 by the
change of variables 𝑥 ←[ 𝑥𝜎. This changes 𝐷 to 𝐷
𝜎 . Note that running the discretized chain on this
transformed problem with step size 𝜂 corresponds to running the discretized chain on the original
√
problem with step size 𝜂𝜎 2 . This is becauseÄ a step
𝑌
=
𝑌
−𝜂∇𝑔(𝑌
)
𝑑𝑡+
2𝜂𝜉𝑘 with 𝑔(𝑥) = 𝑓 𝜎𝑥
𝑡+1
𝑡
𝑡
ä √
p
corresponds to a step 𝑋𝑡+1 = 𝑋𝑡 − 𝜂𝜎∇𝑔 𝑋𝜎𝑡 + 2𝜂𝜎𝜉𝑘 = 𝑋𝑡 − 𝜂𝜎 2 ∇𝑓 (𝑋𝑡 ) + 2𝜂𝜎 2 𝜉𝑘 .
Now we prove the main theorem, Theorem 3.1.
Proof of Theorem 3.1. Choose 𝛿 =
𝜀
2𝐿
where 𝐿 is the number of temperatures. Use Lemma 9.4 in-
ductively, with probability 1− 2𝜀 each estimate satisfies 𝑍“𝑙 ∈ [ 1𝑒 , 𝑒]. Estimating the final distribution
𝑍1
within 2𝜀 accuracy gives the desired sample.
“
Remark 9.5. One reason that the large powers appear in Lemma 9.4 is that we are going between
conductance and spectral gap multiple times, and each time we lose a square by Cheeger’s inequality.
We care about a spectral gap within sets of the partition, but Theorem A.8 controls the conductance
rather than the spectral gap. It may be possible to tighten the bound by proving a variant of the
theorem that controls the spectral gap directly.
10
Acknowledgements
This work was done in part while the authors were visiting the Simons Institute for the Theory of
Computing. We thank Matus Telgarsky and Maxim Raginsky for illuminating conversations in the
early stages of this work.
References
[Aga+14]
A. Agarwal et al. “Learning Sparsely Used Overcomplete Dictionaries.” In: Conference
on Learning Theory (COLT). 2014.
[AGM12]
S. Arora, R. Ge, and A. Moitra. “Learning Topic Models – Going Beyond SVD”. In:
Proceedings of the 53rd Annual IEEE Symposium on Foundations of Computer Science
(FOCS). 2012.
[Ana+12]
A. Anandkumar et al. “A spectral algorithm for latent dirichlet allocation”. In: Advances in Neural Information Processing Systems. 2012, pp. 917–925.
[AR15]
P. Awasthi and A. Risteski. “On some provably correct cases of variational inference for
topic models”. In: Advances in Neural Information Processing Systems. 2015, pp. 2098–
2106.
[Aro+13]
S. Arora et al. “A practical algorithm for topic modeling with provable guarantees”.
In: International Conference on Machine Learning. 2013, pp. 280–288.
[Aro+15]
S. Arora et al. “Simple, efficient, and neural algorithms for sparse coding”. In: (2015).
[Aro+17]
S. Arora et al. “Provable learning of Noisy-or Networks”. In: Symposium on the Theory
of Computing (STOC). 2017.
40
[Bak+08]
D. Bakry et al. “A simple proof of the Poincaré inequality for a large class of probability
measures including the log-concave case”. In: Electron. Commun. Probab 13 (2008),
pp. 60–66.
[BÉ85]
D. Bakry and M. Émery. “Diffusions hypercontractives”. In: Séminaire de Probabilités
XIX 1983/84. Springer, 1985, pp. 177–206.
[Bel+15]
A. Belloni et al. “Escaping the local minima via simulated annealing: Optimization of
approximately convex functions”. In: Conference on Learning Theory. 2015, pp. 240–
265.
[BEL15]
S. Bubeck, R. Eldan, and J. Lehec. “Sampling from a log-concave distribution with
Projected Langevin Monte Carlo”. In: arXiv preprint arXiv:1507.02564 (2015).
[BGK05]
A. Bovier, V. Gayrard, and M. Klein. “Metastability in reversible diffusion processes II:
Precise asymptotics for small eigenvalues”. In: Journal of the European Mathematical
Society 7.1 (2005), pp. 69–99.
[BGL13]
D. Bakry, I. Gentil, and M. Ledoux. Analysis and geometry of Markov diffusion operators. Vol. 348. Springer Science & Business Media, 2013.
[Bha78]
R. Bhattacharya. “Criteria for recurrence and existence of invariant measures for multidimensional diffusions”. In: The Annals of Probability (1978), pp. 541–553.
[Bov+02]
A. Bovier et al. “Metastability and Low Lying Spectra in Reversible Markov Chains”.
In: Communications in mathematical physics 228.2 (2002), pp. 219–255.
[Bov+04]
A. Bovier et al. “Metastability in reversible diffusion processes I: Sharp asymptotics
for capacities and exit times”. In: Journal of the European Mathematical Society 6.4
(2004), pp. 399–424.
[Dal16]
A. S. Dalalyan. “Theoretical guarantees for approximate sampling from smooth and
log-concave densities”. In: Journal of the Royal Statistical Society: Series B (Statistical
Methodology) (2016).
[Dal17]
A. S. Dalalyan. “Further and stronger analogy between sampling and optimization:
Langevin Monte Carlo and gradient descent”. In: arXiv preprint arXiv:1704.04752
(2017).
[DM16]
A. Durmus and E. Moulines. “High-dimensional Bayesian inference via the Unadjusted
Langevin Algorithm”. In: (2016).
[DS93]
P. Diaconis and L. Saloff-Coste. “Comparison theorems for reversible Markov chains”.
In: The Annals of Applied Probability (1993), pp. 696–730.
[GT14]
S. O. Gharan and L. Trevisan. “Partitioning into expanders”. In: Proceedings of the
twenty-fifth annual ACM-SIAM symposium on Discrete algorithms. Society for Industrial and Applied Mathematics. 2014, pp. 1256–1266.
[HK13]
D. Hsu and S. M. Kakade. “Learning mixtures of spherical gaussians: moment methods
and spectral decompositions”. In: Proceedings of the 4th conference on Innovations in
Theoretical Computer Science. ACM. 2013, pp. 11–20.
[KW13]
D. P. Kingma and M. Welling. “Auto-encoding variational bayes”. In: arXiv preprint
arXiv:1312.6114 (2013).
41
[LM00]
B. Laurent and P. Massart. “Adaptive estimation of a quadratic functional by model
selection”. In: Annals of Statistics (2000), pp. 1302–1338.
[LR16]
Y. Li and A. Risteski. “Algorithms and matching lower bounds for approximatelyconvex optimization”. In: Advances in Neural Information Processing Systems. 2016,
pp. 4745–4753.
[LS93]
L. Lovász and M. Simonovits. “Random walks in a convex body and an improved
volume algorithm”. In: Random structures & algorithms 4.4 (1993), pp. 359–412.
[MR02]
N. Madras and D. Randall. “Markov chain decomposition for convergence rate analysis”. In: Annals of Applied Probability (2002), pp. 581–606.
[RMW14]
D. J. Rezende, S. Mohamed, and D. Wierstra. “Stochastic Backpropagation and Approximate Inference in Deep Generative Models”. In: International Conference on Machine Learning. 2014, pp. 1278–1286.
[RRT17]
M. Raginsky, A. Rakhlin, and M. Telgarsky. “Non-convex learning via Stochastic Gradient Langevin Dynamics: a nonasymptotic analysis”. In: arXiv preprint arXiv:1702.03849
(2017).
[SR11]
D. Sontag and D. Roy. “Complexity of inference in latent dirichlet allocation”. In:
Advances in neural information processing systems. 2011, pp. 1008–1016.
[Vem05]
S. Vempala. “Geometric random walks: a survey”. In: Combinatorial and computational
geometry 52.573-612 (2005), p. 2.
[W+09]
D. B. Woodard, S. C. Schmidler, M. Huber, et al. “Sufficient conditions for torpid
mixing of parallel and simulated tempering”. In: Electronic Journal of Probability 14
(2009), pp. 780–804.
[WSH09]
D. B. Woodard, S. C. Schmidler, and M. Huber. “Conditions for rapid mixing of parallel and simulated tempering on multimodal distributions”. In: The Annals of Applied
Probability (2009), pp. 617–640.
[Zhe03]
Z. Zheng. “On swapping and simulated tempering algorithms”. In: Stochastic Processes
and their Applications 104.1 (2003), pp. 131–154.
A
A.1
Background on Markov chains
Discrete time Markov chains
Definition A.1. A (discrete time) Markov chain is 𝑀 = (Ω, 𝑃 ), where Ω is a measure space and
𝑃 (𝑥, 𝑦) 𝑑𝑦 is a probability measure for each 𝑥. 3 It defines a random process (𝑋𝑡 )𝑡∈N0 as follows.
If 𝑋𝑠 = 𝑥, then
P(𝑋𝑠+1 ∈ 𝐴) = 𝑃 (𝑥, 𝐴) : =
R
Z
𝑝(𝑥, 𝑦) 𝑑𝑦.
(217)
𝐴
A stationary distribution is 𝑝(𝑥) such that if 𝑋0 ∼ 𝑝, then 𝑋𝑡 ∼ 𝑝 for all 𝑡; equivalently,
𝑝(𝑥)𝑃
(𝑥, 𝑦) 𝑑𝑥 = 𝑝(𝑦).
Ω
3
For simplicity of notation, in this appendix we consider chains absolutely continuous with respect to R𝑛 , so we
use the notation 𝑝(𝑥) 𝑑𝑥 rather than 𝑑𝜇(𝑥), and 𝑃 (𝑥, 𝑦) 𝑑𝑦 rather than 𝑃 (𝑥, 𝑑𝑦). The same results and definitions
apply with the modified notation if this is not the case.
42
A chain is reversible if 𝑝(𝑥)𝑃 (𝑥, 𝑦) = 𝑝(𝑦)𝑃 (𝑦, 𝑥).
Definition A.2. For a discrete-time Markov chain 𝑀 = (Ω, 𝑃 ), let 𝑃 operate on functions as
Z
(𝑃 𝑔)(𝑥) = E𝑦∼𝑃 (𝑥,·) 𝑔(𝑦) =
𝑔(𝑥)𝑃 (𝑥, 𝑦) 𝑑𝑦.
(218)
Ω
Suppose 𝑀 = (Ω, 𝑃 ) has unique stationary distribution 𝑝. Let ⟨𝑔, ℎ⟩𝑝 :=
define the Dirichlet form and variance by
R
Ω 𝑔(𝑥)ℎ(𝑥)𝑝(𝑥) 𝑑𝑥
ℰ𝑀 (𝑔, ℎ) = ⟨𝑔, (𝐼 − 𝑃 )ℎ⟩𝑝
Var𝑝 (𝑔) = 𝑔 −
and
(219)
2
Z
𝑔𝑝 𝑑𝑥
Ω
(220)
𝑝
Write ℰ𝑀 (𝑔) for ℰ𝑀 (𝑔, 𝑔). Define the eigenvalues of 𝑀 , 0 = 𝜆1 ≤ 𝜆2 ≤ · · · to be the eigenvalues
of 𝐼 − 𝑃 with respect to the norm ‖·‖𝑝 .
Define the spectral gap by
Gap(𝑀 ) =
inf
𝑔∈𝐿2 (𝑝)
ℰ𝑀 (𝑔)
.
Var𝑝 (𝑔)
(221)
Note that
ℰ𝑀 (𝑔) =
1 x
(𝑔(𝑥) − 𝑔(𝑦))2 𝑝(𝑥)𝑃 (𝑥, 𝑦) 𝑑𝑥 𝑑𝑦
2
(222)
Ω×Ω
and that
Gap(𝑀 ) =
ℰ𝑀 (𝑔
= 𝜆2 (𝐼 − 𝑃 ).
𝑔∈𝐿2 (𝑝),𝑔⊥𝑝 1 ‖𝑔‖2
𝑝
inf
(223)
1
1
Remark A.3. The normalized Laplacian of a graph is defined as ℒ = 𝐼 − 𝐷− 2 𝐴𝐷− 2 , where 𝐴 is
the adjacency matrix and»𝐷 is the diagonal matrix of degrees.
A change of scale by 𝑝(𝑥) turns ℒ into 𝐼 −𝑃 , where 𝑃 has the transition matrix of the random
walk of the graph, so the eigenvalues of ℒ are equal to the eigenvalues of the Markov chain defined
here.
The spectral gap controls mixing for the Markov chain. Define the 𝜒2 distance between 𝑝, 𝑞 by
Z Ç
𝜒2 (𝑝||𝑞) =
Ω
𝑞(𝑥) − 𝑝(𝑥)
𝑝(𝑥)
å2
Z Ç
𝑝(𝑥) 𝑑𝑥 =
Ω
𝑞(𝑥)2
𝑝(𝑥)
å
− 1.
(224)
Let 𝑝0 be any initial distribution and 𝑝𝑡 be the distribution after running the Markov chain for 𝑡
steps. Then
𝜒2 (𝑝||𝑝𝑡 ) ≤ (1 − 𝐺′ )𝑡 𝜒(𝑝||𝑝0 )
where 𝐺′ = min(𝜆2 , 2 − 𝜆max ).
43
(225)
A.2
Restricted and projected Markov chains
Given a Markov chain on Ω, we define two Markov chains associated with a partition of Ω.
Definition A.4. For a Markov chain 𝑀 = (Ω, 𝑃 ), and a set 𝐴 ⊆ Ω, define the restriction of 𝑀
to 𝐴 to be the Markov chain 𝑀 |𝐴 = (𝐴, 𝑃 |𝐴 ), where
𝑃 |𝐴 (𝑥, 𝐵) = 𝑃 (𝑥, 𝐵) + 1𝐵 (𝑥)𝑃 (𝑥, 𝐴𝑐 ).
(In words, 𝑃 (𝑥, 𝑦) proposes a transition, and the transition is rejected if it would leave 𝐴.)
Suppose the unique stationary distribution of 𝑀 is 𝑝. Given a partition 𝒫 = {𝐴𝑗 : 𝑗 ∈ 𝐽},
𝒫
𝒫
define the projected Markov chain with respect to 𝒫 to be 𝑀 = (𝐽, 𝑃 ), where
1
𝑃 (𝑖, 𝑗) =
𝑝(𝐴𝑖 )
𝒫
Z
Z
𝑃 (𝑥, 𝑦)𝑝(𝑥) 𝑑𝑥.
𝐴𝑖
𝐴𝑗
(In words, 𝑃 (𝑖, 𝑗) is the “total probability flow” from 𝐴𝑖 to 𝐴𝑗 .)
We omit the superscript 𝒫 when it is clear.
The following theorem lower-bounds the gap of the original chain in terms of the gap of the
projected chain and the minimum gap of the restrictioned chains.
Theorem A.5 (Gap-Product Theorem[MR02]). Let 𝑀 = (Ω, 𝑃 ) be a Markov chain with stationary
distribution 𝑝.
Let 𝒫 = {𝐴𝑗 : 𝑗 ∈ 𝐽} be a partition of Ω such that 𝑝(𝐴𝑗 ) > 0 for all 𝑗 ∈ 𝐽.
1
𝒫
𝒫
Gap(𝑀 ) min Gap(𝑀 |𝐴𝑗 ) ≤ Gap(𝑀 ) ≤ Gap(𝑀 ).
𝑗∈𝐽
2
A.3
Conductance and clustering
Definition A.6. Let 𝑀 = (Ω, 𝑃 ) be a Markov chain with unique stationary distribution 𝑝. Let
𝑄(𝑥, 𝑦) = 𝑝(𝑥)𝑃 (𝑥, 𝑦)
x
𝑄(𝐴, 𝐵) =
𝑄(𝑥, 𝑦) 𝑑𝑥 𝑑𝑦.
(226)
(227)
𝐴×𝐵
(I.e., 𝑥 is drawn from the stationary distribution and 𝑦 is the next state in the Markov chain.)
Define the (external) conductance of 𝑆, 𝜑𝑀 (𝑆), and the Cheeger constant of 𝑀 , Φ(𝑀 ), by
𝑄(𝑆, 𝑆 𝑐 )
𝑝(𝑆)
Φ(𝑀 ) =
min
𝜑𝑀 (𝑆).
𝜑𝑀 (𝑆) =
𝑆⊂Ω,𝑝(𝑆)≤ 12
(228)
(229)
Definition A.7. Let 𝑀 = (Ω, 𝑃 ) be a Markov chain on a finite state space Ω. We say that 𝑘
disjoint subsets 𝐴1 , . . . , 𝐴𝑘 of Ω are a (𝜑in , 𝜑out )-clustering if for all 1 ≤ 𝑖 ≤ 𝑘,
Φ(𝑀 |𝐴𝑖 ) ≥ 𝜑in
𝜑𝑀 (𝐴𝑖 ) ≤ 𝜑out .
44
(230)
(231)
Theorem A.8 (Spectrally partitioning graphs, [GT14]). Let 𝑀 = (Ω, 𝑃 ) be a reversible Markov
chain with |Ω| = 𝑛 states. Let 0 = 𝜆1 ≤ 𝜆2 ≤ · · · ≤ 𝜆𝑛 be the eigenvalues of the Markov chain.
For any 𝑘 ≥ 2, if 𝜆𝑘 > 0, then √
there exists 1 ≤ ℓ ≤ 𝑘 − 1 and a ℓ-partitioning of Ω into sets
𝑃1 , . . . , 𝑃ℓ that is a (Ω(𝜆𝑘 /𝑘 2 ), 𝑂(ℓ3 𝜆ℓ ))-clustering.
Proof. This is [GT14, Theorem 1.5], except that they use a different notion of the restriction of a
Markov chain 𝑀 |𝐴𝑖 . We reconcile this below.
They consider the Markov chain associated with a graph 𝐺, and consider the Cheeger constant
of the induced graph, in their definition of a (𝜑in , 𝜑out ) clustering: Φ(𝐺[𝐴𝑖 ]) ≥ 𝜑in .
We can recast our definition in graph-theoretic language as follows: construct a weighted graph
𝐺 with weight on edge 𝑥𝑦 given by 𝑝(𝑥)𝑃 (𝑥, 𝑦). Now the restricted chain 𝑀 |𝐴 corresponds to the
graph 𝐺|𝐴 , which is the same as the induced graph 𝐺[𝐴] except that we take all edges leaving a
vertex 𝑥 ∈ 𝐴 and redraw them as self-loops at 𝑥.
On the surface, this seems to cause a problem because if we define the volume of a set 𝑆 ⊆ 𝐴
to be the sum of weights of its vertices, the volume of 𝑆 can be larger in 𝐺|𝐴 than 𝐺[𝐴], yet the
amount of weight leaving 𝑆 does not increase.
However, examining the proof in [GT14], we see that every lower bound of the form 𝜑𝐺[𝐴] (𝑆) is
obtained by first lower-bounding by 𝑤(𝑆,𝐴∖𝑆)
Vol(𝑆) , which is exactly 𝜑𝐺|𝐴 (𝑆). Thus their theorem works
equally well with 𝐺|𝐴 instead of 𝐺[𝐴].
Cheeger’s inequality relates the conductance with the spectral gap.
Theorem A.9 (Cheeger’s inequality). Let 𝑀 = (Ω, 𝑃 ) be a reversible Markov chain on a finite
state space and Φ = Φ(𝑀 ) be its conductance. Then
Φ2
≤ Gap(𝑃 ) ≤ Φ.
2
A.4
Continuous time Markov processes
A continuous time Markov process is instead defined by (𝑃𝑡 )𝑡≥0 , and a more natural object to
consider is the generator.
Definition A.10. A continuous time Markov process is given by 𝑀 = (Ω, (𝑃𝑡 )𝑡≥0 ) where the 𝑃𝑡
define a random proces (𝑋𝑡 )𝑡≥0 by
P(𝑋𝑠+𝑡 ∈ 𝐴) = 𝑃𝑡 (𝑥, 𝐴) :=
Z
𝑃 (𝑥, 𝑦) 𝑑𝑦.
𝐴
Define stationary distributions, reversibility, 𝑃𝑡 𝑓 , and variance as in the discrete case.
Define the generator L by
L 𝑔 = lim
𝑡↘0
𝑃𝑡 𝑔 − 𝑔
.
𝑡
(232)
If 𝑝 is the unique stationary distribution, define
ℰ𝑀 (𝑔, ℎ) = − ⟨𝑔, L ℎ⟩𝑝 .
(233)
The spectral gap is defined as in the discrete case with this definition of ℰ𝑀 . The eigenvalues of 𝑀
are defined as the eigenvalues of −L .4
4
Note that ℒ = 𝐼 − 𝑃 in the discrete case corresponds to −L in the continuous case.
45
Note that in order for (𝑃𝑡 )𝑡≥0 to be a valid Markov process, it must be the case that 𝑃𝑡 𝑃𝑢 𝑔 =
𝑃𝑡+𝑢 𝑔, i.e., the (𝑃𝑡 )𝑡≥0 forms a Markov semigroup.
Definition A.11. A continuous Markov process satisfies a Poincaré inequality with constant 𝐶 if
ℰ𝑀 (𝑔) ≥
1
Var𝑝 (𝑔).
𝐶
(234)
This is another way of saying that Gap(𝑀 ) ≥ 𝐶1 .
For Langevin diffusion with stationary distribution 𝑝,
ℰ𝑀 (𝑔) = ‖∇𝑔‖2𝑝 .
(235)
Since this depends in a natural way on 𝑝, we will also write this as ℰ𝑝 (𝑔). A Poincaré inequality
for Langevin diffusion thus takes the form
ℰ𝑝 (𝑔) =
Z
‖∇𝑔‖2 𝑝 𝑑𝑥 ≥
Ω
1
Var𝑝 (𝑔).
𝐶
(236)
We have the following classical result.
Theorem A.12 ([BGL13]). Let 𝑔 be 𝜌-strongly convex and differentiable. Then 𝑔 satisfies the
Poincaré inequality
ℰ𝑝 (𝑔) ≥ 𝜌 Var𝑝 (𝑔).
2
In particular, this holds for 𝑔(𝑥) = ‖𝑥−𝜇‖
with 𝜌 = 1, giving a Poincaré inequality for the
2
gaussian distribution.
A spectral gap, or equivalently a Poincaré inequality, implies rapid mixing (cf. (225)):
𝑡
‖𝑔 − 𝑃𝑡 𝑔‖2 ≤ 𝑒−𝑡 Gap(𝑀 ) = 𝑒− 𝐶 .
B
(237)
Examples
It might be surprising that sampling a mixture of gaussians require a complicated Markov Chain
such as simulated tempering. However, many simple strategies seem to fail.
Langevin with few restarts One natural strategy to try is simply to run Langevin a polynomial
number of times from randomly chosen locations. While the time to “escape” a mode and enter
a different one could be exponential, we may hope that each of the different runs “explores” the
individual modes, and we somehow stitch the runs together. The difficulty with this is that when
the means of the gaussians are not well-separated, it’s difficult to quantify how far each of the
individual runs will reach and thus how to combine the various runs.
Recovering the means of the gaussians Another natural strategy would be to try to recover
the means of the gaussians in the mixture by performing gradient descent on the log-pdf with a
polynomial number of random restarts. The hope would be that maybe the local minima of the
log-pdf correspond to the means of the gaussians, and with enough restarts, we should be able to
find them.
46
Unfortunately, this strategy without substantial modifications also seems to not work: for
instance, in dimension 𝑑, consider a mixture of 𝑑 + 1 gaussians, 𝑑 of them with means on the
corners of a 𝑑-dimensional simplex with a side-length substantially smaller than the diameter 𝐷 we
are considering, and one in the center of the simplex. In order to discover the mean of the gaussian
in the center, we would have to have a starting point extremely close to the center of the simplex,
which in high dimensions seems difficult.
Additionally, this doesn’t address at all the issue of robustness to perturbations. Though there
are algorithms to optimize “approximately” convex functions, they can typically handle only very
small perturbations. [Bel+15; LR16]
Gaussians with different covariance Our result requires all the gaussians to have the same
variance. This is necessary, as even if the variance of the gaussians only differ by a factor of 2,
there are examples where a simulated tempering chain takes exponential time to converge [W+09].
Intuitively, this is illustrated in Figure 3. The figure on the left shows the distribution in low
temperature – in this case the two modes are separate, and both have a significant mass. The
figure on the right shows the distribution in high temperature. Note that although in this case
the two modes are connected, the volume of the mode with smaller variance is much smaller
(exponentially small in 𝑑). Therefore in high dimensions, even though the modes can be connected
at high temperature, the probability mass associated with a small variance mode is too small to
allow fast mixing.
Figure 3: Mixture of two gaussians with different covariance at different temperature
C
Pertubation tolerance
In the previous sections, we argued that we can sample from distributions of the form 𝑝˜(𝑥) ∝
˜
exp(𝑓 (𝑥)),
where 𝑓˜ is as (3). In this section, the goal we be to argue that we can sample from
distributions of the form 𝑝(𝑥) ∝ exp(𝑓 (𝑥)), where 𝑓 is as (2).
Our main theorem is the following.
Theorem C.1 (Main theorem with perturbations). Suppose 𝑓 (𝑥) and 𝑓˜(𝑥) satisfy (3) and (2).
′
Then, algorithm 2 with parameters
given by Lemma
Ä
ä C.2 produces a sample from a distribution 𝑝
with ‖𝑝 − 𝑝′ ‖1 ≤ 𝜀 in time poly 𝑤min , 𝐷, 𝑑, 1𝜀 , 𝑒Δ , 𝜏 .
47
The theorem will follow immediately from Lemma C.2, which is a straightforward analogue of
9.4. More precisely:
Lemma C.2. Suppose that Algorithm 1 is run on temperatures 0 < 𝛽1 < · · · < 𝛽ℓ ≤ 1, ℓ ≤ 𝐿 with
”1 , . . . , 𝑍
cℓ satisfying
partition function estimates 𝑍
c𝑖
”1
𝑍
𝑍
1
−
≤ 1+
𝑍𝑖 𝑍1
𝐿
Å
ã𝑖−1
(238)
for all 1 ≤ 𝑖 ≤ ℓ and with parameters satisfying
Ç
𝛽1 = 𝑂 min
Ç 2
𝜎
𝐷2
Ñ
𝛽𝑖 − 𝛽𝑖−1 = 𝑂
,Δ
(239)
Ñ
min
Å
𝑇 = Ω 𝐷2 ln
Ö
åå
Ä
𝐷2 𝑑 + ln
1
Å
Ä
1
𝑤min
ää , Δ
(240)
ãã
(241)
𝑤min
𝐷8 𝑑4 + ln
𝑡=Ω
éé
𝜎2
Ä
1
è
ä4
𝑤min
Ç
ln
4
𝜎 8 𝑤min
1
max
𝜖
Ç
𝐷2 𝑑 ln (1/𝑤
min )
𝜎 8 𝑤min
, 𝑒Δ
åå
Ç
max
𝐷2
𝑚16
å
,
𝜎 2 ln(1/𝑤min )4
(242)
Ç
𝜂=𝑂
𝜀𝜎 2
𝑑𝑡𝑇 𝜏 2
å
.
(243)
Ä
Ä
2
ä
ä
Let 𝑞 0 be the distribution 𝑁 0, 𝜎𝛽1 , 1 on R𝑑 × [ℓ]. The distribution 𝑞 𝑡 after running for 𝑡 steps
satisfies 𝑝 − 𝑞 𝑡 1 ≤ 𝜀. Setting 𝜀 = 𝑂
1 − 𝛿 the estimate
Ä ä
1
𝐿
Ä
above and taking 𝑚 = Ω ln
Ñ
“ℓ+1 = 𝑍
“ℓ
𝑍
𝑚
1 X
𝑒(−𝛽ℓ+1 +𝛽ℓ )𝑓𝑖 (𝑥𝑗 )
𝑚 𝑗=1
Ä ää
1
𝛿
samples, with probability
é
(244)
also satisfies (238).
The way we prove this theorem is to prove the tolerance of each of the proof ingredients to
perturbations to 𝑓 .
C.1
Mixing time of the tempering chain
We first show that the mixing time of the tempering chain that uses the continous Langevin
transition 𝑃𝑇 for 𝑝(𝑥) ∝ exp(𝑓 (𝑥)) is comparable to that of 𝑝˜(𝑥) ∝ exp(𝑓˜(𝑥)). Keeping in mind
the statement of Lemma 6.8, the following lemma suffices:
Lemma C.3. Suppose ‖𝑓1 − 𝑓2 ‖∞ ≤
R𝑑 . Then the following hold.
Δ
2
and 𝑝1 ∝ 𝑒−𝑓1 , 𝑝2 ∝ 𝑒−𝑓2 are probability distributions on
1.
ℰ𝑝1 (𝑔)
−Δ ℰ𝑝2 (𝑔)
.
2 ≥𝑒
‖𝑔‖𝑝1
‖𝑔‖2𝑝2
48
(245)
2. Letting L1 , L2 be the generators for Langevin diffusion on 𝑝1 , 𝑝2 ,
𝜆𝑛 (−L1 ) ≥ 𝑒−Δ 𝜆𝑛 (−L2 ).
(246)
3. If a Poincaré inequality holds for 𝑝1 with constant 𝐶, then a Poincaré inequality holds for 𝑝2
with constant 𝐶𝑒Δ .
Note that if we are given probability distributions 𝑝1 , 𝑝2 such that 𝑝1 ∈ [1, 𝑒Δ ]𝑝2 𝑅 for some 𝑅,
then the conditions of the lemma are satisfied.
Proof.
1. The ratio between 𝑝1 and 𝑝2 is at most 𝑒Δ , so
R
R𝑑
R
‖∇𝑔‖2 𝑝1 𝑑𝑥
R𝑑
‖𝑔‖2 𝑝1 𝑑𝑥
≥
𝑒−Δ
𝑒Δ
R
𝑑
RR
R𝑑
‖∇𝑔‖2 𝑝2 𝑑𝑥
‖𝑔‖2 𝑝2 𝑑𝑥
(247)
2. Use the first part along with the variational characterization
𝜆𝑚 (−L ) =
max
𝐿2 (𝑝)
closed subspace 𝑆 ⊆
dim(𝑆 ⊥ ) = 𝑚 − 1
min
𝑔∈𝑆
− ⟨𝑔, L 𝑔⟩
.
‖𝑔‖2𝑝
(248)
3. Use the second part for 𝑚 = 2; a Poincaré inequality is the same as a lower bound on 𝜆2 .
C.2
Mixing time at highest temperature
We show that we can use the same highest temperature corresponding to 𝑓 (𝑥) for 𝑓˜(𝑥) as well, at
the cost of 𝑒Δ in the mixing time. Namely, since ‖𝑓˜ − 𝑓 ‖∞ ≤ Δ, from Lemma 7.3, we immediately
have:
Lemma C.4. If 𝑓 and 𝑓˜ satisfy (2) and (3), there exists a 1/2 strongly-convex function 𝑔, s.t.
‖𝑓 − 𝑔‖∞ ≤ 𝐷2 + Δ.
As a consequence, the proof of Lemma 7.4 implies
Lemma C.5. If 𝑓 and 𝑓˜ satisfy (2) and (3), Langevin diffusion on 𝛽𝑓 (𝑥) satisfies a Poincaré
inequality with constant
C.3
16𝑒2𝛽(𝐷
𝛽
2 +Δ)
.
Discretization
The proof of Lemma 8.3, combined with the fact that ∇𝑓˜ − ∇𝑓
∞
≤ Δ gives
Lemma C.6 (Perturbed reach of continuous chain). Let 𝑃𝑇𝛽 (𝑋) be the Markov kernel corresponding
to evolving Langevin diffusion
𝑑𝑋𝑡
= −𝛽∇𝑓 (𝑋𝑡 ) + 𝑑𝐵𝑡
𝑑𝑡
with 𝑓 and 𝐷 are as defined in 2 for time 𝑇 . Then,
E[‖𝑋𝑡 − 𝑥* ‖2 ] . E[‖𝑋0 − 𝑥* ‖2 ] + (𝛽(𝐷 + 𝜏 )2 + 𝑑)𝑇
49
Proof. The proof proceeds exactly the same as Lemma 8.3, noting that ∇𝑓˜ − ∇𝑓
≤ 𝜏 implies
∞
−⟨𝑋𝑡 − 𝑥* , 𝑋𝑡 − 𝜇𝑖 ⟩ ≤ −‖𝑋𝑡 ‖2 + ‖𝑋𝑡 ‖(‖𝜇𝑖 ‖ + ‖𝑥* ‖ + 𝜏 ) + ‖𝑥* ‖(‖𝜇𝑖 ‖ + 𝜏 )
Furthermore, since ∇2 𝑓˜(𝑥) ⪯ ∇2 𝑓 (𝑥) + 𝜏 𝐼, ∀𝑥 ∈ R𝑑 , from Lemma 8.4, we get
Lemma C.7 (Perturbed Hessian bound).
2
+ 𝜏 𝐼, ∀𝑥 ∈ R𝑑
∇ 𝑓 (𝑥) ⪯
𝜎2
ã
Å
2
As a consequence, the analogue of Lemma 8.5 gives:
Lemma C.8 (Bounding interval drift). In the setting of Lemma 8.5, let 𝑥 ∈ R𝑑 , 𝑖 ∈ [𝐿], and let
𝜂≤
( 𝜎1 +𝜏 )2
.
𝛼
Then,
Å
ã2
ä
𝜂 2 ( 𝜎12 + 𝜏 )3 𝛼 Ä
1
* 2
”
‖𝑥 − 𝑥 ‖2 ) + 𝑇 𝑑 + 𝑑𝑇 𝜂
+𝜏
KL(𝑃𝑇 (𝑥, 𝑖)||𝑃𝑇 (𝑥, 𝑖)) .
2
2𝛼 − 1
𝜎
Putting these together, we get the analogue of Lemma 8.1:
Lemma C.9. Let 𝑝𝑡 , 𝑞 𝑡 : R𝑑 × [𝐿] → R be the distributions after running the simulated tempering
chain for 𝑡 steps, where in 𝑝𝑡 , for any temperature 𝑖 ∈ 𝐿, the Type 1 transitions are taken according
to the (discrete time) Markov kernel 𝑃𝑇 : running Langevin diffusion for time 𝑇 ; in 𝑞 𝑡 , the Type 1
transitions are taken according to running 𝑇𝜂 steps of the discretized Langevin diffusion, using 𝜂 as
1
the discretization granularity, s.t. 𝜂 ≤
. Then,
1
2
𝑡
𝑡
KL(𝑝 ||𝑞 ) . 𝜂
C.4
2
Å
1
+𝜏
𝜎2
ã3 Ä
𝜎2
ä
2
+𝜏
2
(𝐷 + 𝜏 ) + 𝑑 𝑇 𝑡 + 𝜂
2
Å
1
+𝜏
𝜎2
ã3
max E𝑥∼𝑝0 (·,𝑖) ‖𝑥 −
𝑖
𝑥* ‖22
1
+𝜂
+𝜏
𝜎2
Å
ã2
Putting things together
Finally, we prove Theorem C.2
Proof of C.2. The proof is analogous to the one of Lemma 9.4 in combination with the Lemmas
from the previous subsections.
For the analysis of the simulated tempering chain, consider the same partition 𝒫𝑖 we used in
Lemma 9.4. Then, by Lemma C.3,
Gap(𝑀𝑖 |𝐴 ) ≥ Ω 𝑒
1
(ln 𝑤min
)4
−Δ
𝑚16
!!
.
(249)
For the highest temperature, by Lemma C.5, we have
Gap(𝑀1 ) = Ω 𝛽1 𝑒−2𝛽1 (𝐷
) = Ω(min( 1 , 1 )).
Δ 𝐷2
2 +Δ
50
(250)
𝑑𝑡𝑇
Furthermore, by Lemma 9.2, since the condition on 𝛽𝑖 − 𝛽𝑖−1 is satisfied, 𝛿((𝒫𝑖 )ℓ𝑖=1 ) = Ω(1).
Then, same as in Lemma 9.4, the spectral gap of the simulated tempering chain
𝐺 := Gap(𝑀st ) =
Ç
As in Lemma 9.4, since 𝑡 = Ω
𝑤4
𝑒−Δ min
ℓ4
𝑙
,𝑒Δ ))
min
ln( 1𝜖 max( 𝑤
Å
1
ln
Ä
1
ä4 é
𝑤min
𝑚16
.
(251)
å
𝐺
𝑝˜ − 𝑞 𝑡
Ñ
=𝑂
,
𝜀𝑤min
𝜒2 (˜
𝑝||𝑞 0 )
ℓ
ã
(252)
By triangle inequality,
𝜒2 (˜
𝑝||𝑞 0 ) ≤ 𝜒2 (˜
𝑝||𝑝) + 𝜒2 (𝑝||𝑞 0 )
The proof of Lemma 9.4 bounds 𝜒2 (𝑝||𝑞 0 ) = 𝑂
2
𝜒 (˜
𝑝||𝑝) =
Ç
Z
𝑥∈R𝑑
≤
Ä
ℓ
𝑤min
ä
, and
𝑝˜(𝑥) − 𝑝(𝑥)
𝑝(𝑥)
Ç Δ
å2
𝑒 𝑝(𝑥) − 𝑝(𝑥)
𝑝(𝑥)
å2
𝑝(𝑥)𝑑𝑥
𝑝(𝑥)
≤ 𝑒Δ
From this, we get 𝑝˜ − 𝑞 𝑡 1 ≤ 3𝜖 .
For the term 𝑝𝑡 − 𝑞 𝑡 1 , use Pinsker’s inequality and Lemma C.9 to get
𝑝˜𝑡 − 𝑞 𝑡
Å
for
ß
1 √
, 1
𝜂 = 𝑂 𝜀 min √𝑇 𝑡𝜏 3/2 (𝐷+
𝑑+𝜏 ) 𝜏 2 𝑑𝑡𝑇
This gives 𝑝˜ − 𝑞 𝑡 1 ≤ 𝜀.
1
™ã
≤
»
2KL(˜
𝑝𝑡 ||𝑞 𝑡 ) ≤
𝜀
3
(253)
𝜀
= 𝑂( 𝜏 2 𝑑𝑡𝑇
).
The proof of the second part of the Lemma proceeds exactly as C.9.
D
Another lower bound for simulated tempering
Theorem D.1 (Comparison theorem using canonical paths, [DS93]). Let (Ω, 𝑃 ) be a finite Markov
chain with stationary distribution 𝑝.
Suppose each pair 𝑥, 𝑦 ∈ Ω, 𝑥 ̸= 𝑦 is associated with a path 𝛾𝑥,𝑦 . Define the congestion to be
𝛾𝑥,𝑦 ∋(𝑧,𝑤) |𝛾𝑥,𝑦 |𝑝(𝑥)𝑝(𝑦)
ñP
𝜌(𝛾) =
max
𝑝(𝑧)𝑃 (𝑧, 𝑤)
𝑧,𝑤∈Ω,𝑧̸=𝑤
Then
Gap(𝑃 ) ≥
51
1
.
𝜌(𝛾)
ô
.
Definition D.2. Say that partition 𝒫 refines 𝒬, written 𝑃 ⊑ 𝒬, if for every 𝐴 ∈ 𝒫 there exists
𝐵 ∈ 𝒬 such that 𝐴 ⊆ 𝐵.
Define a chain of partitions as {𝒫𝑖 = {𝐴𝑖,𝑗 }}𝐿
𝑖=1 , where each 𝒫𝑖 is a refinement of 𝒫𝑖−1 :
𝒫𝐿 ⊑ · · · ⊑ 𝒫1 .
Theorem D.3. Suppose Assumptions 5.1 hold.
Furthermore, suppose that (𝒫𝑖 )𝐿
𝑖=1 is a chain of partitions. Define 𝛾 for the chain of partitions
as
𝑝𝑖 (𝐴)
𝛾((𝒫𝑖 )𝐿
min
min 1
.
𝑖=1 ) =
1≤𝑖1 ≤𝑖2 ≤𝐿 𝐴∈𝒫𝑖1 𝑝𝑖2 (𝐴)
Then
Gap(𝑀st ) ≥
𝑟2 𝛾𝛿
32𝐿3
min
1≤𝑖≤𝐿,𝐴∈𝒫𝑖
(Gap(𝑀 |𝐴 )).
(254)
Proof. Let 𝑝st be the stationary distribution of 𝑃st . First note that we can easily switch between
𝑝𝑖 and 𝑝st using 𝑝st (𝐴 × {𝑖}) = 𝑟𝑖 𝑝𝑖 (𝐴).
Define the partition 𝒫 on Ω × {0, . . . , 𝑙 − 1} by
𝒫 = {𝐴 × {𝑖} : 𝐴 ∈ 𝒫𝑖 } .
By Theorem A.5,
Gap(𝑀st ) ≥
1
Gap(𝑀 st ) min Gap(𝑀st |𝐵 ).
𝐵∈𝒫
2
(255)
We now lower-bound Gap(𝑀 st ). We will abuse notation by considering the sets 𝐵 ∈ 𝒫 as states
in 𝑀 st , and identify a union of sets in 𝒫 with the corresponding set of states for 𝑀 st .
Consider a tree with nodes 𝐵 ∈ 𝒫, and edges connecting 𝐴 × {𝑖}, 𝐴′ × {𝑖 − 1} if 𝐴 ∈ 𝐴′ .
Designate Ω × {1} as the root. For 𝑋, 𝑌 ∈ 𝒫, define the canonical path 𝛾𝑋,𝑌 to be the unique path
in this tree.
Note that |𝛾𝑋,𝑌 | ≤ 2(𝐿 − 1). Given an edge (𝐴 × {𝑖}, 𝐴′ × {𝑖 − 1}), consider
𝛾𝑋,𝑌 ∋(𝐴×{𝑖},𝐴′ ×{𝑖−1}) |𝛾𝑋,𝑌 |𝑝st (𝑋)𝑝st (𝑌
𝑝st (𝐴 × {𝑖})𝑃st (𝐴 × {𝑖}, 𝐴′ × {𝑖 − 1})
P
)
2(𝐿 − 1)2𝑝st (𝑆)𝑝st (𝑆 𝑐 )
𝑝st (𝐴 × {𝑖})𝑃st (𝐴 × {𝑖}, 𝐴′ × {𝑖 − 1})
≤
(256)
where 𝑆 = 𝐴 × {𝑖, . . . , 𝐿} is the union of all children of 𝐴 × {𝑖} (including itself). This follows
because the paths which go through (𝐴 × {𝑖}, 𝐴′ × {𝑖 − 1}) are exactly those between 𝑋, 𝑌 where
one of 𝑋, 𝑌 is a subset of 𝑆 = 𝐴 × {𝑖, . . . , 𝐿} and the other is not. To upper bound (256), we
𝑝(𝑆)
upper-bound 𝑝(𝐴×{𝑖})
and lower-bound 𝑃 (𝐴 × {𝑖}, 𝐴′ × {𝑖 + 1}).
We upper-bound by definition of 𝛾,
𝑝(𝑆)
=
𝑝(𝐴 × {𝑖})
P𝐿
𝑘=𝑖 𝑟𝑖 𝑝𝑖 (𝐴)
𝑟𝑖 𝑝𝑖 (𝐴)
P
max 𝑟𝑖 𝐿
𝑘=𝑖 𝑝𝑖 (𝐴)
≤
min 𝑟𝑖
𝑝𝑖 (𝐴)
𝐿
≤
.
𝑟𝛾
52
(257)
(258)
(259)
Next we lower bound 𝑃st (𝐴 × {𝑖}, 𝐴′ × {𝑖 − 1}). There is probability
to level 𝑖 − 1, so
′
𝑃st (𝐴 × {𝑖}, 𝐴 × {𝑖 − 1}) ≥
=
≥
≥
®
1
2𝐿
of proposing a switch
´
1
𝑝𝑖−1 (𝑥) 𝑟𝑖−1
𝑟𝑖 𝑝𝑖 (𝑥) min
, 1 𝑑𝑥/(𝑟𝑖 𝑝𝑖 (𝐴))
2𝐿 Ω
𝑝𝑖 (𝑥) 𝑟𝑖
Z
ß
™
𝑟𝑖−1
1
min 𝑝𝑖−1 (𝑥)
, 𝑝𝑖 (𝑥) 𝑑𝑥/𝑝𝑖 (𝐴)
2𝐿 Ω
𝑟𝑖
Z
1 min 𝑟𝑗
min {𝑝𝑖−1 (𝑥), 𝑝𝑖 (𝑥)} 𝑑𝑥/𝑝𝑖 (𝐴)
2𝐿 max 𝑟𝑗 Ω
1
𝑟𝛿.
2𝐿
Z
(260)
(261)
(262)
(263)
Putting (256), (259), and (263) together,
(256) ≤ 2(𝐿 − 1)2
≤
Å
𝐿
𝑟𝛾
ãÅ
2𝐿
𝑟𝛿
ã
8𝐿3
.
𝑟2 𝛾𝛿
(264)
(265)
Using (255) and Theorem D.1,
1
Gap(𝑀 st ) min Gap(𝑀st |𝐵 )
𝐵∈𝒫
2
2
𝑟 𝛾𝛿
≥
min Gap(𝑀st |𝐵 )
16𝐿3 𝐵∈𝒫
𝑟2 𝛾𝛿
Gap(𝑀𝑖 |𝐴 )
≥
min
32𝐿3 1≤𝑖≤𝐿,𝐴∈𝒫𝑖
Gap(𝑀st ) ≥
(266)
(267)
(268)
By taking all the partitions except the first to be the same, we see that this theorem is an
improvement to the bound for simulated tempering in [WSH09, Theorem 3.1], which gives the
bound
™
ß
𝛾 𝐽+3 𝛿 3
Gap(𝑃st ) ≥ 14
min
min (Gap(𝑀𝑖 |𝐴 )), Gap(𝑀1 )
2≤𝑖≤𝐿,𝐴∈𝒫
2 (𝐿 + 1)5 𝐽 3
when 𝑟 = 1, where 𝐽 is the number of sets in the partition. Most notably, their bound is exponential
in 𝐽, while our bound has no dependence on 𝐽.
53
| 8cs.DS
|
Generative Encoder-Decoder Models for Task-Oriented Spoken Dialog
Systems with Chatting Capability
Tiancheng Zhao, Allen Lu, Kyusong Lee and Maxine Eskenazi
Language Technologies Institute
Carnegie Mellon University
Pittsburgh, Pennsylvania, USA
{tianchez,arlu,kyusongl,max+}@cs.cmu.edu
arXiv:1706.08476v1 [cs.CL] 26 Jun 2017
Abstract
Generative encoder-decoder models offer great promise in developing domaingeneral dialog systems. However, they
have mainly been applied to open-domain
conversations.
This paper presents
a practical and novel framework for
building task-oriented dialog systems
based on encoder-decoder models. This
framework enables encoder-decoder models to accomplish slot-value independent
decision-making and interact with external
databases. Moreover, this paper shows the
flexibility of the proposed method by interleaving chatting capability with a slotfilling system for better out-of-domain recovery. The models were trained on both
real-user data from a bus information system and human-human chat data. Results show that the proposed framework
achieves good performance in both offline
evaluation metrics and in task success rate
with human users.
1
Introduction
Task-oriented spoken dialog systems have transformed human-computer interaction by enabling
people interact with computers via spoken language (Raux et al., 2005; Young, 2006; Bohus
and Rudnicky, 2003). The task-oriented SDS is
usually domain-specific. The system creators first
map the user utterances into semantic frames that
contain domain-specific slots and intents using
spoken language understanding (SLU) (De Mori
et al., 2008). Then a set of domain-specific dialog
state variables is tracked to retain the context information over turns (Williams et al., 2013). Lastly,
the dialog policy decides the next move from a
list of dialog acts that covers the expected communicative functions from the system.
Although the above approach has been successfully applied to many practical systems, it has limited ability to generalize to out-of-domain (OOD)
requests and to scale up to new domains. For example, even within in a simple domain, real users
often make requests that are not included in the
semantic specifications. Due to this, proper error handling strategies that guide users back to the
in-domain conversation are crucial to dialog success (Bohus and Rudnicky, 2005). Past error handling strategies were limited to a set of predefined
dialog acts, e.g. request repeat, clarification etc.,
which constrained the system’s capability in keeping users engaged. Moreover, there has been an
increased interest in extending task-oriented systems to multiple topics (Lee et al., 2009; Gašić
et al., 2015b) and multiple skills, e.g. grouping
heterogeneous types of dialogs into a single system (Zhao et al., 2016). Both cases require the
system to be flexible enough to extend to new slots
and actions.
Our goal is to move towards a domain-general
task-oriented SDS framework that is flexible
enough to expand to new domains and skills by
removing domain-specific assumptions on the dialog state and dialog acts (Bordes and Weston,
2016). To achieve this goal, the neural encoderdecoder model(Cho et al., 2014; Sutskever et al.,
2014) is a suitable choice, since it has achieved
promising results in modeling open-domain conversations (Vinyals and Le, 2015; Sordoni et al.,
2015). It encodes the dialog history using deep
neural networks and then generates the next system utterance word-by-word via recurrent neural
networks (RNNs). Therefore, unlike the traditional SDS pipeline, the encoder-decoder model is
theoretically only limited by its input/output vocabulary.
A na`‘ive implementation of an encoderdecoder-based task-oriented system would use
RNNs to encode the raw dialog history and generate the next system utterance using a separate
RNN decoder. However, while this implementation might achieve good performance in an offline
evaluation of a closed dataset, it would certainly
fail when used by humans. There are several reasons for this: 1) real users can mention new entities that do not appear in the training data, such as
a new restaurant name. These entities are, however, essential in delivering the information that
matches users’ needs in a task-oriented system.
2) a task-oriented SDS obtains information from
a knowledge base (KB) that is constantly updated
(“today’s” weather will be different every day), so
simply memorizing KB results that occurred in the
training data would produce false information. Instead, an effective model should learn to query the
KB constantly to get the most up-to-date information. 3) users may give OOD requests (e.g. say,
“how is your day”, to a slot-filling system), which
must be handled gracefully in order to keep the
conversation moving in the intended direction.
This paper proposes an effective encoderdecoder framework for building task-oriented
SDSs. We propose entity indexing to tackle the
challenges of out-of-vocabulary (OOV) entities
and to query the KB. Moreover, we show the extensibility of the proposed model by adding chatting capability to a task-oriented encoder-decoder
SDS for better OOD recovery. This approach
was assessed on the Let’s Go Bus Information
data from the 1st Dialog State Tracking Challenge (Williams et al., 2013), and we report performance on both offline metrics and real human
users. Results show that this model attains good
performance for both of these metrics.
2
matically learn a distributed vector representation
of the dialog state by accumulating the observations at each turn (Williams and Zweig, 2016;
Zhao and Eskenazi, 2016; Dhingra et al., 2016;
Williams et al., 2017). The learned dialog state
is then used by the dialog policy to select the
next action. The second branch of research develops a domain-general action space for dialog
policy. Prior work replaced the domain-specific
dialog acts with domain-independent natural language semantic schema as the action space of dialog managers (Eshghi and Lemon, 2014), e.g. Dynamic Syntax (Kempson et al., 2000). More recently, Wen, et al. (2016) have shown the feasibility of using an RNN as the decoder to generate the
system utterances word by word, and the dialog
policy of the proposed model can be fine tuned using reinforcement learning (Su et al., 2016). Furthermore, to deal with the challenge of developing end-to-end task-oriented dialog models that
are able to interface with external KB, prior work
has unified the special KB query actions via deep
reinforcement learning (Zhao and Eskenazi, 2016)
and soft attention over the database (Dhingra et al.,
2016). The third branch strives to solve both problems at the same time by building an end-to-end
model that maps an observable dialog history directly to the word sequences of the system’s response. By using an encoder-decoder model, it has
been successfully applied to open-domain conversational models (Serban et al., 2015; Li et al.,
2015, 2016; Zhao et al., 2017), as well as to task
oriented systems (Bordes and Weston, 2016; Yang
et al., 2016; Eric and Manning, 2017). In order to
better predict the next correct system action, this
branch has focused on investigating various neural network architectures to improve the machine’s
ability to reason over user input and model longterm dialog context.
Related Work
Past research in developing domain-general dialog systems can be broadly divided into three
branches. The first one focuses on learning
domain-independent dialog state representation
while still using hand-crafted dialog act system actions. Researchers proposed the idea of extracting slot-value independent statistics as the dialog state (Wang et al., 2015; Gašić et al., 2015a),
so that the dialog state representation can be
shared across systems serving different knowledge
sources. Another approach uses RNNs to auto-
This paper is closely related to the third branch,
but differs in the following ways: 1) these models
are slot-value independent by leveraging domaingeneral entity recognizer, which is more extensible to OOV entities, 2) these models emphasize
the interactive nature of dialog and address out-ofdomain handling by interleaving chatting in taskoriented conversations, 3) instead of testing on a
synthetic dataset, this approach focuses on real
world use by testing the system on human users
via spoken interface.
3
Proposed Method
Our proposed framework consists of three steps
as shown in Figure 2: a) entity indexing (EI), b)
slot-value independent encoder-decoder (SiED),
c) system utterance lexicalization (UL). The intuition is to leverage domain-general named entity recognition (NER) (Tjong Kim Sang and
De Meulder, 2003) techniques to extract salient
entities in the raw dialog history and convert the
lexical values of the entities into entity indexes.
The encoder-decoder model is then trained to focus solely on reasoning over the entity indexes in
a dialog history and to make decisions about the
next utterance to produce (including KB query). In
this way, the model can be unaffected by the inclusion of new entities and new KB, while maintaining its domain-general input/output interface for
easy extension to new types of conversation skills.
Lastly, the output from the decoder networks are
lexicalized by replacing the entity indexes and special KB tokens with natural language. The following sections explain each step in detail.
3.1
Entity Indexing and Utterance
Lexicalization
Entity Indexing EI has two parts. First, the EI
utilizes an existing domain-general NER to extract entities from both the user and system utterances. Note that the entity here is assumed to be
a super-set of the slots in the domain. For example, for a flight-booking system, the system may
contain two slots: [from-LOCATION] and [toLOCATION] for the departure and arrival city, respectively. However, EI only extracts every mention of [LOCATION] in the utterances and leaves
the task of distinguishing between departure and
arrival to the encoder-decoder model. Furthermore, this step replaces each KB search result with
its search query (e.g. the weather is cloudy → [kbsearch]-[DATETIME-0]). The second step of EI
involves constructing a indexed entity table. Each
entity is indexed by its order of occurrence in the
conversation. Figure 1 shows an example in which
there are two [LOCATION] mentions.
Properties of Entity Indexing In this section,
several properties of EI and their assumptions are
addressed. First, each entity is indexed uniquely
by its entity type and index. Note that the index is not associated with the entity value, but
rather solely by the order of appearance in the
dialog. Despite the actual words being hidden,
Figure 1: An example of entity indexing and utterance lexicalization.
a human can still easily predict which entity the
system should confirm or search for in the KB
based on logical reasoning. Therefore, that the EI
not only alleviates the OOV problem of deploying
the encoder-decoder model in the real world, but
also forces the encoder-decoder model’s focus on
learning the reasoning process of task-oriented dialogs instead of leveraging too much information
from the language modeling.
Moreover, most slot-filling SDSs, apart from informing the concepts from KBs, usually do not introduce novel entities to users. Instead, systems
mostly corroborate the entities introduced by the
users. With this assumption, every entity mention
in the system utterances can always be found in the
users’ utterances in the dialog history, and therefore can also be found in the indexed entity table.
This property reduces the grounding behavior of
the conventional task-oriented dialog manager into
selecting an entity from the indexed entity table
and confirming it with the user.
Utterance Lexicalization is the reverse of EI.
Since EI is a deterministic process, its effect can
always be reversed by finding the corresponding
entity in the indexed entity table and replacing the
index with its word. For KB search, a simple string
matching algorithm can search for the special [kbsearch] token and take the following generated entities as the argument to the KB. Then the actual
KB results can replace the original KB query. Figure 1 shows an example of utterance lexicalization.
3.2
Encoder-Decoder Models
The encoder-decoder model can then read in the
EI-processed dialog history and predict the system’s next utterance in EI format. Specifically,
a dialog history of k turns is represented by
[(a0 , u0 , c0 ), ...(ak−1 , uk−1 , ck−1 )], in which ai ,
ui and ci are, respectively, the system, user utterance and ASR confidence score at turn i. Each utterance in the dialog history is encoded into fixedsize vectors using Convolutional Neural Networks
Figure 2: The proposed pipeline for task-oriented dialog systems.
(CNNs) proposed in (Kim, 2014). Specifically,
each word in an utterance x is mapped to its word
embedding, so that an utterance is represented as a
matrix R ∈ R|x|×D , in which D is the size of the
word embedding. Then L filters of size 1,2,3 conduct convolutions on R to obtain a feature map, c,
of n-gram features in window size 1,2,3. Then c
is passed through a nonlinear ReLu (Glorot et al.,
2011) layer, followed by a max-pooling layer to
obtain a compact summary of salient n-gram features, i.e. et (x) = maxpool(ReLu(c + b)). Using CNNs to capture word-order information is
crucial, because the encoder-decoder has to be
able to distinguish between fine-grained differences between entities. For example, a simple
bag-of-word embedding approach will fail to distinguish between the two location entities in “leave
from [LOCATION-0] and go to [LOCATION-1]”,
while a CNN encoder can capture the context information of these two entities.
After obtaining utterance embedding, a turnlevel dialog history encoder network similar to
the one proposed in (Zhao and Eskenazi, 2016)
is used. Turn embedding is a simple concatenation of system, user utterance embedding and the
confidence score t = [eu (ai ); eu (ui ); ci ]. Then
an Long Short-Term Memory (LSTM) (Hochreiter and Schmidhuber, 1997) network reads the sequence turn embeddings in the dialog history via
recursive state update si+1 = LSTM(ti+1 , hi ), in
which hi is the output of the LSTM hidden state.
Decoding with/without Attention A vanilla
decoder takes in the last hidden state of the
encoder as its initial state and decodes the
next system utterance word by word as shown
in (Sutskever et al., 2014). This assumes that
the fixed-size hidden state is expressive enough
to encode all important information about the
history of a dialog. However, this assumption may often be violated for a task that has
long-dependency or complex reasoning of the entire source sequence. An attention mechanism
proposed (Bahdanau et al., 2014) in the machine translation community has helped encoderdecoder models improve state-of-art performance
in various tasks (Bahdanau et al., 2014; Xu et al.,
2015). Attention allows the decoder to look over
every hidden state in the encoder and dynamically
decide the importance of each hidden state at each
decoding step, which significantly improves the
model’s ability to handle long-term dependency.
We experiment decoders both with and without
attention. Attention is computed similarly multiplicative attention described in (Luong et al.,
2015). We denote the hidden state of the decoder
at time step j by sj , and the hidden state outputs
of the encoder at turn i by hi . We then predict the
next word by
aji = softmax(hTi Wa sj + ba )
X
cj =
aji hi
(1)
(2)
i
s
sej = tanh(Ws j )
cj
(3)
p(wj |sj , cj ) = softmax(Wo sej )
(4)
The decoder next state is updated by sj+1 =
LSTM(sj , e(wj+1 ), sej ).
3.3
Leveraging Chat Data to Improve OOD
Recovery
Past work has shown that simple supervised learning is usually inadequate for learning robust sequential decision-making policy (Williams and
Young, 2003; Ross et al., 2011). This is because
the model is only exposed to the expert demonstration, but not to examples of how to recover from its
own mistakes or users’ OOD requests. We present
a simple yet effective technique that leverages the
extensibility of the encoder-decoder model in order to obtain a more robust policy in the setting
of supervised learning. Specifically, we artificially
augment a task-oriented dialog dataset with chat
data from an open-domain conversation corpus.
This has been shown to be effective in improving the performance of task-oriented systems (Yu
et al., 2017). Let the original dialog dataset with
N dialogs be D = [d0 ..., dn , ...dN ], where dn is a
multi-turn task-oriented dialog of |dn | turns. Furthermore, we assume we have access to a chat
dataset Dc = [(q0 , r0 ), ...(qm , rm ), ...(qM , rM )],
where qm , rm are common adjacency pairs that
appear in chats, (e.g. q = hello, r = hi, how are
you). Then we can create a new dataset D∗ by repeating the following process a certain number of
times:
1. Randomly sample dialog dn from D
2. Randomly sample turn ti = [ai , ui ] from dn
3. Randomly sample
(qm , rm ) from Dc
an
adjacency
pair
4. Replace the user utterance of ti by qm so that
ti = [ai , qm ]
5. Insert a new turn after ti , i.e. ti+1 = [rm +
ei+1 , ui ]
Figure 3: Illustration of data augmentation. The
turn in the dashed line is inserted in the original
dialog.
In Step 5, ei is an error handling system utterance
after the system answers the user’s OOD request,
qm . In this study, we experimented with a simple
case where ei+1 = ai so that the system should repeat its previous prompt after responding to qm via
rm . Figure 3 shows an example of an augmented
turn. Eventually, we train the model on the union
of the two datasets D+ = D ∪ D∗
Discussion: There are several reasons that the
above data augmentation process is appealing.
First, the model effectively learns an OOD recovery strategy from D∗ , i.e. it first gives chatting
answers to users’ OOD requests and then tries to
pull users back to the main-task conversation. Second, chat data usually has a larger vocabulary and
more diverse natural language expressions, which
can reduce the chance of OOVs and enable the
model to learn more robust word embeddings and
language models.
4
4.1
Experiment Setup
Dataset and Domain
The CMU Let’s Go Bus Information System (Raux et al., 2005) is a task-oriented spoken
dialog system that contains bus information. We
combined the train1a and train1b datasets from
DSTC 1 (Williams et al., 2013), which contain
2608 total dialogs. The average dialog length
is 9.07 turns. The dialogs were randomly splitted into 85/5/10 proportions for train/dev/test data.
The data was noisy since the dialogs were collected from real users via telephone lines. Furthermore, this version of Let’s Go used an inhouse database containing the Port Authority bus
schedule. In the current version, that database was
replaced with the Google Directions API, which
both reduces the human burden of maintaining a
database and opens the possibility of extending
Let’s Go to cities other than Pittsburgh. Connecting to Google Directions API involves a POST call
to their URL, with our given access key as well
as the parameters needed: departure place, arrival
place and departure time, and the travel mode,
which we always set as TRANSIT to obtain relevant bus routes. There are 14 distinct dialog acts
available to the system, and each system utterance
contains one or more dialog acts. Lastly, the system vocabulary size is 1311 and the user vocabulary size is 1232. After the EI process, the sizes
become 214 and 936, respectively.
For chat data, we use a publicly available chat
corpus used in (Yu et al., 2015)1 . In total, there
are 3793 chatting adjacency pairs. We control the
number of data injections to 30% of the number of
turns in the original DTSC dataset, which leads to
a user vocabulary size of 3537 and system vocabulary size of 4047.
4.2
Training Details
For all experiments, the word embedding size was
100. The sizes of the LSTM hidden states for both
the encoder and decoder were 500 with 1 layer.
The attention context size was also 500. We tied
the CNN weights for the encoding system and user
utterances. Each CNN has 3 filter windows, 1, 2,
and 3, with 100 feature maps each. We trained
the model end-to-end using Adam (Kingma and
Ba, 2014), with a learning rate of 1e-3 and a
batch size of 40. To combat overfitting, we apply
dropout (Zaremba et al., 2014) to the LSTM layer
outputs and the CNN outputs after the maxpooling
layer, with a dropout rate of 40%.
5
responses. Then we computed the micro-average
precision, recall, and the F-score.
Slots: This metric measures the model’s performance in generating the correct slot-values. The
slot-values mostly occur in grounding utterances
(e.g. explicit/implicit confirm) and KB queries.
We compute precision, recall, and F-score.
KB Queries: Although the slots metric already
covers the KB queries, here the precision/recall/Fscore of system utterances that contain KB queries
are also explicitly measured, due to their importance. Specifically, this action measures whether
the system is able to generate the special [kbquery] symbol to initiate a KB query, as well as
how accurate the corresponding KB query arguments are.
BLEU (Papineni et al., 2002): compares the ngram precision with length penalty, and has been
a popular score used to evaluate the performance
of natural language generation (Wen et al., 2015)
and open-domain dialog models (Li et al., 2016).
Corpus-level BLEU-4 is reported.
Experiments Results
This approach was assessed both offline and online evaluations. The offline evaluation contains
standard metrics to test open-domain encoderdecoder dialog models (Li et al., 2015; Serban
et al., 2015). System performance was assessed
from three perspectives that are essential for taskoriented systems: dialog acts, slot-values, and KB
query. The online evaluation is composed of objective task success rate, the number of turns, and
subjective satisfaction with human users.
5.1
Offline Evaluation
Dialog Acts (DA): Each system utterance is made
up of one or more dialog acts, e.g. “leaving
at [TIME-0], where do you want to go?” →
[implicit-confirm, request(arrival place)]. To evaluate whether a generated utterance has the same
dialog acts as the ground truth, we trained a multilabel dialog tagger using one-vs-rest Support Vector Machines (SVM) (Tsoumakas and Katakis,
2006), with bag-of-bigram features for each dialog act label. Since the natural language generation module in Let’s Go is handcrafted, the dialog
act tagger achieved 99.4% average label accuracy
on a held-out dataset. We used this dialog act tagger to tag both the ground truth and the generated
1
github.com/echoyuzhou/ticktock text api
Metrics
Vanilla EI
DA
(p/r/f1)
83.5
77.9
80.5
42.0
30.3
35.2
N/A
Slot
(p/r/f1)
KB
(p/r/f1)
BLEU
36.9
79.7
80.1
80.0
60.6
63.6
62.1
48.9
55.3
51.9
54.6
EI
+Attn
80.0
83.1
81.5
63.7
64.7
64.2
55.4
70.8
62.2
59.3
EI+Attn
+Chat
81.8
83.5
82.7
64.6
69.1
66.8
58.2
71.9
64.4
60.5
Table 1: Performance of each model on automatic
measures.
Four systems were compared: the basic
encoder-decoder models without EI (vanilla), the
basic model with EI pre-processing (EI), the
model with attentional decoder (EI+Attn) and the
model trained on the dataset augmented with chatting data (EI+Attn+Chat). The comparison was
carried out on exactly the same held-out test
dataset that contains 261 dialogs. Table 1 shows
the results. It can be seen that all four models achieve similar performance on the dialog act
metrics, even the vanilla model. This confirms
the capacity of encoder-decoders models to learn
the “shape” of a conversation, since they have
achieved impressive results in more challenging
settings, e.g. modeling open-domain conversations. Furthermore, since the DSTC1 data was
collected over several months, there were minor
updates made to the dialog manager. Therefore,
there are inherent ambiguities in the data (the dialog manager may take different actions in the same
situation). We conjecture that ∼80% is near the
upper limit of our data in modeling the system’s
next dialog act given the dialog history.
On the other hand, these proposed methods significantly improved the metrics related to slots
and KB queries. The inclusion of EI alone was
able to improve the F-score of slots by a relative
76%, which confirms that EI is crucial in developing slot-value independent encoder-decoder models for modeling task-oriented dialogs. Likewise,
the inclusion of attention further improved the prediction of slots in system utterances. Adding attention also improved the performance of predicting KB queries, more so than the overall slot accuracy. This is expected, since KB queries are usually issued near the end of a conversation, which
requires global reasoning over the entire dialog
history. The use of attention allows the decoder
to look over the history and make better decisions
rather than simply depending on the context summary in the last hidden layer of the encoder. Because of the good performance achieved by the
models with the attentional decoder, the attention
weights in Equation 1 at every step of the decoding process in two example dialogs from test data
are visualized. For both figures, the vertical axes
show the dialog history flowing from the top to the
bottom. Each row is a turn in the format of (system utterance # user utterance). The top horizontal axis shows the predicted next system utterance.
The darkness of a bar indicates the value of the
attention calculated in Equation 1.
The first example shows attention for grounding the new entity [LOCATION-1] in the previous turn. The attention weights become focus on
the previous turn when predicting [LOCATION1] in the implicit confirm action. The second dialog example shows a more challenging situation,
in which the model is predicting a KB query. We
can see that the attention weights when generating
each input argument of the KB query clearly focus on the specific mention in the dialog history.
The visualization confirms the effectiveness of the
attention mechanism in dealing with long-term de-
pendency at discourse level.
Figure 4: Visualization of attention weights when
generating implicit confirm (top) and KB query
(bottom).
Surprisingly, the model trained on the data augmented with chat achieved slightly better slot accuracy performance, even though the augmented
data is not directly related to task-oriented dialogs. Furthermore, the model trained on chataugmented data achieved better scores for the
KB query metrics. Several reasons may explain
this improvement: 1) since chat data exposes the
model to a significantly larger vocabulary, the resulting model is more robust to words that it had
not seen in the original task-oriented-only training
data, and 2) the augmented dialog turn can be seen
as noise in the dialog history, which adds extra
regularization to the model and enables the model
to learn more robust long-term reasoning mechanisms.
5.2
Human Evaluation
Although the model achieves good performance
in offline evaluation, this may not carray over
to real user dialogs, where the system must simultaneously deal with several challenges, such
as automatic speech recognition (ASR) errors,
OOD requests, etc. Therefore, a real user study
was conducted to evaluate the performance of
the proposed systems in the real world. Due to
the limited number of real users, only two best
performing system were compared, EI+Attn and
EI+Attn+Chat. Users were able to talk to a web
interface to the dialog systems via speech. Google
Chrome Speech API 2 served as the ASR and textto-speech (TTS) modules. Turn-taking was done
via the built-in Chrome voice activity detection
(VAD) plus a finite state machine-based end-ofturn detector (Zhao et al., 2015). Lastly, a hybrid
named entity recognizer (NER) was trained using
Conditional Random Field (CRF) (McCallum and
Li, 2003) and rules to extract 4 types of entities
(location, hour, minute, pm/am) for the EI process.
The experiment setup is as follows: when a user
logs into the website, the system prompts the user
with a goal, which is a randomly chosen combination of departure place, arrival place and time (e.g.
leave from CMU and go to the airport at 10:30
AM). The system also instructs the user to say
goodbye if the he/she thinks the goal is achieved
or wants to give up. The user begins a conversation with one of the two evaluated systems, with a
50/50 chance of choosing either system (not visible to the user). After the user’s session is finished, the system asks the him/her to give two
scores between 1 and 5 for correctness and naturalness of the system respectively. The subjects
in this study consist of undergraduate and graduate students. However, many subjects did not
follow the prompted goal, but rather asked about
bus routes of their own. Therefore, the dialog
was manually labeled for dialog success. A dialog is successful if and only if the systems give at
least one bus schedule that matches with all three
slots expressed by the users. Table 2 shows the
Metrics
EI+Attn
# of Dialog
Slot Precision
KB Precision
Success Rate
Avg Turns
Avg Correctness
Avg Naturalness
75
73.3%
88.6%
73.3%
4.88
3.45 (1.32)
3.46 (1.41)
EI+Attn
+Chat
74
71.8%
93.7%
77.0%
4.91
3.22 (1.40)
3.53 (1.34)
Table 2: Performance of each model on automatic
measures. The standard deviations of subjective
scores are in parentheses.
results. Overall, our systems achieved reasonable
performance in terms of dialog success rate. The
EI+Attn+Chat model achieves slightly higher success and subjective naturalness metrics (although
the difference between EI+Attn+Chat and EI+Attn
2
www.google.com/intl/en/chrome/demos/speech.html
was not statistically significant due to the limited
number of subjects). The precision of grounding the correct slots and predicting the correct KB
query was also manually labelled. EI+Attn model
performs slightly better than the EI+Attn+Chat
model in slot precision, while the latter model performs significantly better in KB query precision.
In addition, EI+Attn+Chat leads to slightly longer
dialogs because sometimes it generates chatting
utterances with users when it cannot understand
users’ utterances.
At last, we investigated the log files and identified the following major types of sources of dialog failure: RNN Decoder Invalid Output: Occasionally, the RNN decoder outputs system utterances as “Okay going to [LOCATION-2]. Did
I get that right?”, in which [LOCATION-2] cannot be found in the indexed entity table. Such invalid output confuses users. This occurred in 149
of the dialogs, where 4.1% of system utterances
contain invalid symbols. Imitation of Suboptimal Dialog Policy: Since our models are only
trained to imitate the suboptimal hand-crafted dialog policy, their limitations show when the original dialog manager cannot handle the situation,
such as failing to understand slots that appeared in
compound utterances. Future plans involves improving the models to perform better than the suboptimal teacher policy.
6
Conclusions
In conclusion, this paper discusses constructing
task-oriented dialog systems using generative encoder decoder models. EI is effective in solving
both the OOV entity and KB query challenges for
encoder-decoder-based task-oriented SDSs. Additionally, the novel data augmentation technique of
interleaving task-oriented dialog corpus with chat
data led to better model performance in both online and offline evaluation. Future work includes
developing more advanced encoder-decoder models that to better deal with long-term dialog history and complex reasoning challenges than current models do. Furthermore, inspired by the success of mixing chatting with slot-filling dialogs,
we will take full advantage of the extensibility of
encoder-decoder models by investigating how to
make systems that are able to interleave various
conversational tasks, e.g. different domains, chatting or task-oriented, which in turn can create a
more versatile conversational agent.
References
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly
learning to align and translate. arXiv preprint
arXiv:1409.0473 .
Dan Bohus and Alexander I Rudnicky. 2003. Ravenclaw: Dialog management using hierarchical task
decomposition and an expectation agenda .
Dan Bohus and Alexander I Rudnicky. 2005. Error handling in the ravenclaw dialog management
framework. In Proceedings of the conference on
Human Language Technology and Empirical Methods in Natural Language Processing. Association
for Computational Linguistics, pages 225–232.
Antoine Bordes and Jason Weston. 2016. Learning end-to-end goal-oriented dialog. arXiv preprint
arXiv:1605.07683 .
Kyunghyun Cho, Bart Van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger
Schwenk, and Yoshua Bengio. 2014. Learning
phrase representations using rnn encoder-decoder
for statistical machine translation. arXiv preprint
arXiv:1406.1078 .
Renato De Mori, Frédéric Bechet, Dilek Hakkani-Tur,
Michael McTear, Giuseppe Riccardi, and Gokhan
Tur. 2008. Spoken language understanding. IEEE
Signal Processing Magazine 25(3).
Bhuwan Dhingra, Lihong Li, Xiujun Li, Jianfeng
Gao, Yun-Nung Chen, Faisal Ahmed, and Li Deng.
2016. End-to-end reinforcement learning of dialogue agents for information access. arXiv preprint
arXiv:1609.00777 .
Mihail Eric and Christopher D Manning. 2017. A
copy-augmented sequence-to-sequence architecture
gives good performance on task-oriented dialogue.
arXiv preprint arXiv:1701.04024 .
Arash Eshghi and Oliver Lemon. 2014. How domaingeneral can we be? learning incremental dialogue
systems without dialogue acts. DialWattSemdial
2014 page 53.
M Gašić, N Mrkšić, Pei-hao Su, David Vandyke,
Tsung-Hsien Wen, and Steve Young. 2015a. Policy
committee for adaptation in multi-domain spoken
dialogue systems. In Automatic Speech Recognition
and Understanding (ASRU), 2015 IEEE Workshop
on. IEEE, pages 806–812.
Milica Gašić, Dongho Kim, Pirros Tsiakoulis, and
Steve Young. 2015b. Distributed dialogue policies for multi-domain statistical dialogue management. In Acoustics, Speech and Signal Processing
(ICASSP), 2015 IEEE International Conference on.
IEEE, pages 5371–5375.
Xavier Glorot, Antoine Bordes, and Yoshua Bengio.
2011. Deep sparse rectifier neural networks. In Aistats. volume 15, page 275.
Sepp Hochreiter and Jürgen Schmidhuber. 1997.
Long short-term memory. Neural computation
9(8):1735–1780.
Ruth Kempson, Wilfried Meyer-Viol, and Dov M Gabbay. 2000. Dynamic syntax: The flow of language
understanding. Wiley-Blackwell.
Yoon Kim. 2014.
Convolutional neural networks for sentence classification. arXiv preprint
arXiv:1408.5882 .
Diederik Kingma and Jimmy Ba. 2014. Adam: A
method for stochastic optimization. arXiv preprint
arXiv:1412.6980 .
Cheongjae Lee, Sangkeun Jung, Seokhwan Kim, and
Gary Geunbae Lee. 2009. Example-based dialog
modeling for practical multi-domain dialog system.
Speech Communication 51(5):466–484.
Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao,
and Bill Dolan. 2015. A diversity-promoting objective function for neural conversation models. arXiv
preprint arXiv:1510.03055 .
Jiwei Li, Will Monroe, Alan Ritter, Michel Galley,
Jianfeng Gao, and Dan Jurafsky. 2016. Deep reinforcement learning for dialogue generation. arXiv
preprint arXiv:1606.01541 .
Minh-Thang Luong, Hieu Pham, and Christopher D
Manning. 2015. Effective approaches to attentionbased neural machine translation. arXiv preprint
arXiv:1508.04025 .
Andrew McCallum and Wei Li. 2003. Early results for
named entity recognition with conditional random
fields, feature induction and web-enhanced lexicons.
In Proceedings of the seventh conference on Natural language learning at HLT-NAACL 2003-Volume
4. Association for Computational Linguistics, pages
188–191.
Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of
the 40th annual meeting on association for computational linguistics. Association for Computational
Linguistics, pages 311–318.
Antoine Raux, Brian Langner, Dan Bohus, Alan W
Black, and Maxine Eskenazi. 2005. Lets go public! taking a spoken dialog system to the real world.
In in Proc. of Interspeech 2005. Citeseer.
Stéphane Ross, Geoffrey J Gordon, and Drew Bagnell.
2011. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS. volume 1, page 6.
Iulian V Serban, Alessandro Sordoni, Yoshua Bengio,
Aaron Courville, and Joelle Pineau. 2015. Building end-to-end dialogue systems using generative hierarchical neural network models. arXiv preprint
arXiv:1507.04808 .
Alessandro Sordoni, Michel Galley, Michael Auli,
Chris Brockett, Yangfeng Ji, Margaret Mitchell,
Jian-Yun Nie, Jianfeng Gao, and Bill Dolan. 2015.
A neural network approach to context-sensitive generation of conversational responses. arXiv preprint
arXiv:1506.06714 .
Pei-Hao Su, Milica Gasic, Nikola Mrksic, Lina RojasBarahona, Stefan Ultes, David Vandyke, TsungHsien Wen, and Steve Young. 2016. Continuously learning neural dialogue management. arXiv
preprint arXiv:1606.02689 .
Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014.
Sequence to sequence learning with neural networks. In Advances in neural information processing systems. pages 3104–3112.
Erik F Tjong Kim Sang and Fien De Meulder.
2003. Introduction to the conll-2003 shared task:
Language-independent named entity recognition. In
Proceedings of the seventh conference on Natural
language learning at HLT-NAACL 2003-Volume 4.
Association for Computational Linguistics, pages
142–147.
and efficient end-to-end dialog control with supervised and reinforcement learning. arXiv preprint
arXiv:1702.03274 .
Jason D Williams and Geoffrey Zweig. 2016. Endto-end lstm-based dialog control optimized with supervised and reinforcement learning. arXiv preprint
arXiv:1606.01269 .
Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho,
Aaron C Courville, Ruslan Salakhutdinov, Richard S
Zemel, and Yoshua Bengio. 2015. Show, attend and
tell: Neural image caption generation with visual attention. In ICML. volume 14, pages 77–81.
Zichao Yang, Phil Blunsom, Chris Dyer, and Wang
Ling. 2016. Reference-aware language models.
arXiv preprint arXiv:1611.01628 .
Steve J Young. 2006. Using pomdps for dialog management. In SLT. pages 8–13.
Zhou Yu, Alan W Black, and Alexander I Rudnicky.
2017. Learning conversational systems that interleave task and non-task content. arXiv preprint
arXiv:1703.00099 .
Grigorios Tsoumakas and Ioannis Katakis. 2006.
Multi-label classification: An overview. International Journal of Data Warehousing and Mining
3(3).
Zhou Yu, Alexandros Papangelis, and Alexander Rudnicky. 2015. Ticktock: A non-goal-oriented multimodal dialog system with engagement awareness.
In Proceedings of the AAAI Spring Symposium.
Oriol Vinyals and Quoc Le. 2015. A neural conversational model. arXiv preprint arXiv:1506.05869 .
Wojciech Zaremba, Ilya Sutskever, and Oriol Vinyals.
2014. Recurrent neural network regularization.
arXiv preprint arXiv:1409.2329 .
Zhuoran Wang, Tsung-Hsien Wen, Pei-Hao Su,
and Yannis Stylianou. 2015. Learning domainindependent dialogue policies via ontology parameterisation. In 16th Annual Meeting of the Special Interest Group on Discourse and Dialogue. page 412.
Tsung-Hsien Wen, Milica Gasic, Nikola Mrksic, PeiHao Su, David Vandyke, and Steve Young. 2015.
Semantically conditioned lstm-based natural language generation for spoken dialogue systems.
arXiv preprint arXiv:1508.01745 .
Tsung-Hsien Wen, David Vandyke, Nikola Mrksic,
Milica Gasic, Lina M Rojas-Barahona, Pei-Hao Su,
Stefan Ultes, and Steve Young. 2016. A networkbased end-to-end trainable task-oriented dialogue
system. arXiv preprint arXiv:1604.04562 .
Jason Williams, Antoine Raux, Deepak Ramachandran, and Alan Black. 2013. The dialog state tracking challenge. In Proceedings of the SIGDIAL 2013
Conference. pages 404–413.
Jason Williams and Steve Young. 2003. Using wizardof-oz simulations to bootstrap reinforcementlearning-based dialog management systems. In
Proceedings of the 4th SIGDIAL Workshop on
Discourse and Dialogue.
Jason D Williams, Kavosh Asadi, and Geoffrey
Zweig. 2017. Hybrid code networks: practical
Tiancheng Zhao, Alan W Black, and Maxine Eskenazi.
2015. An incremental turn-taking model with active system barge-in for spoken dialog systems. In
16th Annual Meeting of the Special Interest Group
on Discourse and Dialogue. page 42.
Tiancheng Zhao and Maxine Eskenazi. 2016. Towards
end-to-end learning for dialog state tracking and
management using deep reinforcement learning. In
17th Annual Meeting of the Special Interest Group
on Discourse and Dialogue.
Tiancheng Zhao, Maxine Eskenazi, and Kyusong Lee.
2016. Dialport: A general framework for aggregating dialog systems. EMNLP 2016 page 32.
Tiancheng Zhao, Ran Zhao, and Maxine Eskenazi.
2017. Learning discourse-level diversity for neural
dialog models using conditional variational autoencoders. arXiv preprint arXiv:1703.10960 .
| 2cs.AI
|
Alternating Back-Propagation for Generator Network
Tian Han † , Yang Lu † , Song-Chun Zhu, Ying Nian Wu
arXiv:1606.08571v4 [stat.ML] 6 Dec 2016
Department of Statistics, University of California, Los Angeles, USA
Abstract
a convolutional neural network (ConvNet or CNN) (LeCun
et al., 1998; Krizhevsky, Sutskever, and Hinton, 2012; Dosovitskiy, Springenberg, and Brox, 2015). It has been shown
recently that the generator network is capable of generating
realistic images (Denton et al., 2015; Radford, Metz, and
Chintala, 2016).
The generator network is a fundamental representation of
knowledge, and it has the following properties: (1) Analysis:
The model disentangles the variations in the observed signals
into independent variations of latent factors. (2) Synthesis:
The model can synthesize new signals by sampling the factors from the known prior distribution and transforming the
factors into the signal. (3) Embedding: The model embeds
the high-dimensional non-Euclidean manifold formed by the
observed signals into the low-dimensional Euclidean space
of the latent factors, so that linear interpolation in the lowdimensional factor space results in non-linear interpolation
in the data space.
This paper proposes an alternating back-propagation algorithm
for learning the generator network model. The model is a nonlinear generalization of factor analysis. In this model, the mapping from the continuous latent factors to the observed signal
is parametrized by a convolutional neural network. The alternating back-propagation algorithm iterates the following two
steps: (1) Inferential back-propagation, which infers the latent
factors by Langevin dynamics or gradient descent. (2) Learning back-propagation, which updates the parameters given
the inferred latent factors by gradient descent. The gradient
computations in both steps are powered by back-propagation,
and they share most of their code in common. We show that
the alternating back-propagation algorithm can learn realistic generator models of natural images, video sequences, and
sounds. Moreover, it can also be used to learn from incomplete
or indirect training data.
1
Introduction
1.2
This paper studies the fundamental problem of learning and
inference in the generator network (Goodfellow et al., 2014),
which is a generative model that has become popular recently.
Specifically, we propose an alternating back-propagation algorithm for learning and inference in this model.
1.1
The factor analysis model can be learned by the Rubin-Thayer
EM algorithm (Rubin and Thayer, 1982; Dempster, Laird,
and Rubin, 1977), where both the E-step and the M-step are
based on multivariate linear regression. Inspired by this algorithm, we propose an alternating back-propagation algorithm
for learning the generator network that iterates the following
two-steps:
(1) Inferential back-propagation: For each training example, infer the continuous latent factors by Langevin dynamics
or gradient descent.
(2) Learning back-propagation: Update the parameters
given the inferred latent factors by gradient descent.
The Langevin dynamics (Neal, 2011) is a stochastic sampling counterpart of gradient descent. The gradient computations in both steps are powered by back-propagation. Because
of the ConvNet structure, the gradient computation in step
(1) is actually a by-product of the gradient computation in
step (2) in terms of coding.
Given the factors, the learning of the ConvNet is a supervised learning problem (Dosovitskiy, Springenberg, and
Brox, 2015) that can be accomplished by the learning backpropagation. With factors unknown, the learning becomes an
unsupervised problem, which can be solved by adding the
inferential back-propagation as an inner loop of the learning
Non-linear factor analysis
The generator network is a non-linear generalization of factor
analysis. Factor analysis is a prototype model in unsupervised learning of distributed representations. There are two
directions one can pursue in order to generalize the factor
analysis model. One direction is to generalize the prior model
or the prior assumption about the latent factors. This led to
methods such as independent component analysis (Hyvärinen, Karhunen, and Oja, 2004), sparse coding (Olshausen
and Field, 1997), non-negative matrix factorization (Lee and
Seung, 2001), matrix factorization and completion for recommender systems (Koren, Bell, and Volinsky, 2009), etc.
The other direction to generalize the factor analysis model
is to generalize the mapping from the continuous latent factors to the observed signal. The generator network is an
example in this direction. It generalizes the linear mapping
in factor analysis to a non-linear mapping that is defined by
†
Alternating back-propagation
Equal contributions.
1
process. We shall show that the alternating back-propagation
algorithm can learn realistic generator models of natural images, video sequences, and sounds.
The alternating back-propagation algorithm follows the
tradition of alternating operations in unsupervised learning,
such as alternating linear regression in the EM algorithm for
factor analysis, alternating least squares algorithm for matrix
factorization (Koren, Bell, and Volinsky, 2009; Kim and Park,
2008), and alternating gradient descent algorithm for sparse
coding (Olshausen and Field, 1997). All these unsupervised
learning algorithms alternate an inference step and a learning
step, as is the case with alternating back-propagation.
1.3
is the generative adversarial network (GAN) (Goodfellow et
al., 2014; Denton et al., 2015; Radford, Metz, and Chintala,
2016), where the assisting network is a discriminator network
that plays an adversarial role against the generator network.
Unlike alternating back-propagation, VAE does not perform explicit explain-away inference, while GAN avoids
inferring the latent factors altogether. In comparison, the alternating back-propagation algorithm is simpler and more
basic, without resorting to an extra network. While it is difficult to compare these methods directly, we illustrate the
strength of alternating back-propagation by learning from
incomplete and indirect data, where we only need to explain
whatever data we are given. This may prove difficult or less
convenient for VAE and GAN.
Meanwhile, alternating back-propagation is complementary to VAE and GAN training. It may use VAE to initialize
the inferential back-propagation, and as a result, may improve the inference in VAE. The inferential back-propagation
may help infer the latent factors of the observed examples for
GAN, thus providing a method to test if GAN can explain
the entire training set.
The generator network is based on a top-down ConvNet.
One can also obtain a probabilistic model based on a bottomup ConvNet that defines descriptive features (Xie et al., 2016;
Lu, Zhu, and Wu, 2016).
Explaining-away inference
The inferential back-propagation solves an inverse problem
by an explaining-away process, where the latent factors compete with each other to explain each training example. The
following are the advantages of the explaining-away inference of the latent factors:
(1) The latent factors may follow sophisticated prior models. For instance, in textured motions (Wang and Zhu, 2003)
or dynamic textures (Doretto et al., 2003), the latent factors
may follow a dynamic model such as vector auto-regression.
By inferring the latent factors that explain the observed examples, we can learn the prior model.
(2) The observed data may be incomplete or indirect. For
instance, the training images may contain occluded objects.
In this case, the latent factors can still be obtained by explaining the incomplete or indirect observations, and the model
can still be learned as before.
1.4
2
2.1
Factor analysis and beyond
Let Y be a D-dimensional observed data vector, such as
an image. Let Z be the d-dimensional vector of continuous
latent factors, Z = (zk , k = 1, ..., d). The traditional factor analysis model is Y = W Z + , where W is D × d
matrix, and is a D-dimensional error vector or the observational noise. We assume that Z ∼ N(0, Id ), where Id stands
for the d-dimensional identity matrix. We also assume that
∼ N(0, σ 2 ID ), i.e., the observational errors are Gaussian
white noises. There are three perspectives to view W . (1)
Basis vectors. Write W = (W1 , ..., Wd ), where each Wk is a
Pd
D-dimensional column vector. Then Y = k=1 zk Wk + ,
i.e., Wk are the basis vectors and zk are the coefficients. (2)
Loading matrix. Write W = (w1 , ..., wD )> , where wj> is
the j-th row of W . Then yj = hwj , Zi + j , where yj and
j are the j-th components of Y and respectively. Each yj
is a loading of the d factors where wj is a vector of loading
weights, indicating which factors are important for determining yj . W is called the loading matrix. (3) Matrix factorization. Suppose we observe Y = (Y1 , ..., Yn ), whose factors
are Z = (Z1 , ..., Zn ), then Y ≈ W Z.
The factor analysis model can be learned by the RubinThayer EM algorithm, which involves alternating regressions
of Z on Y in the E-step and of Y on Z in the M-step, with
both steps powered by the sweep operator (Rubin and Thayer,
1982; Liu, Rubin, and Wu, 1998).
The factor analysis model is the prototype of many subsequent models that generalize the prior model of Z. (1)
Independent component analysis (Hyvärinen, Karhunen, and
Oja, 2004), d = D, = 0, and zk are assumed to follow
independent heavy tailed distributions. (2) Sparse coding
Learning from incomplete or indirect data
We venture to propose that a main advantage of a generative
model is to learn from incomplete or indirect data, which are
not uncommon in practice. The generative model can then
be evaluated based on how well it recovers the unobserved
original data, while still learning a model that can generate
new data. Learning the generator network from incomplete
data can be considered a non-linear generalization of matrix
completion.
We also propose to evaluate the learned generator network
by the reconstruction error on the testing data.
1.5
Factor analysis with ConvNet
Contribution and related work
The main contribution of this paper is to propose the alternating back-propagation algorithm for training the generator
network. Another contribution is to evaluate the generative
models by learning from incomplete or indirect training data.
Existing training methods for the generator network avoid
explain-away inference of latent factors. Two methods have
recently been devised to accomplish this. Both methods involve an assisting network with a separate set of parameters
in addition to the original network that generates the signals. One method is variational auto-encoder (VAE) (Kingma
and Welling, 2014; Rezende, Mohamed, and Wierstra, 2014;
Mnih and Gregor, 2014), where the assisting network is an inferential or recognition network that seeks to approximate the
posterior distribution of the latent factors. The other method
2
(Olshausen and Field, 1997), d > D, and Z is assumed to
be a redundant but sparse vector, i.e., only a small number
of zk are non-zero or significantly different from zero. (3)
Non-negative matrix factorization (Lee and Seung, 2001), it
is assumed that zk ≥ 0. (4) Recommender system (Koren,
Bell, and Volinsky, 2009), Z is a vector of a customer’s desires in different aspects, and wj is a vector of product j’s
desirabilities in these aspects.
2.2
algorithm (Dempster, Laird, and Rubin, 1977), the completedata model is given by
log p(Y, Z; W ) = log [p(Z)p(Y |Z, W )]
1
1
= − 2 kY − f (Z; W )k2 − kZk2 + const. (3)
2σ
2
The observed-data
model is obtained by integrating out Z:
R
p(Y ; W ) = p(Z)p(Y |Z, W )dZ. The posterior distribution of Z is given by p(Z|Y, W ) = p(Y, Z; W )/p(Y ; W ) ∝
p(Z)p(Y |Z, W ) as a function of Z.
For the training data {Y
the complete-data logPi },
n
likelihood is L(W, {Zi }) = i=1 log p(Yi , Zi ; W ), where
we assume σ 2 is given. Learning and inference can be accomplished by maximizing the complete-data log-likelihood,
which can be obtained by the alternating gradient descent
algorithm that iterates the following two steps: (1) Inference
step: update Zi by running l steps of gradient descent. (2)
Learning step: update W by one step of gradient descent.
A more rigorous method is to maximize
Pn the observed-data
log-likelihood,
which
is
L(W
)
=
i=1 log p(Yi ; W ) =
R
Pn
The observed-data logi=1 log p(Yi , Zi ; W )dZi .
likelihood takes into account the uncertainties in inferring
Zi . See Appendix for an in-depth understanding.
The gradient of L(W ) can be calculated according to the
following well-known fact that underlies the EM algorithm:
Z
∂
1
∂
log p(Y ; W ) =
p(Y, Z; W )dZ
∂W
P (Y ; W ) ∂W
∂
log p(Y, Z; W ) .
(4)
= Ep(Z|Y,W )
∂W
ConvNet mapping
In addition to generalizing the prior model of the latent factors Z, we can also generalize the mapping from Z to Y . In
this paper, we consider the generator network model (Goodfellow et al., 2014) that retains the assumptions that d < D,
Z ∼ N(0, Id ), and ∼ N(0, σ 2 ID ) as in traditional factor analysis, but generalizes the linear mapping W Z to a
non-linear mapping f (Z; W ), where f is a ConvNet, and
W collects all the connection weights and bias terms of the
ConvNet. Then the model becomes
Y = f (Z; W ) + ,
Z ∼ N(0, Id ), ∼ N(0, σ 2 ID ), d < D.
(1)
2
The reconstruction error is ||Y − f (Z; W )|| . We may assume more sophisticated models for , such as colored noise
or non-Gaussian texture. If Y is binary, we can emit Y by
a probability map P = 1/[1 + exp(−f (Z; W ))], where the
sigmoid transformation and Bernoulli sampling are carried
out pixel-wise. If Y is multi-level, we may assume multinomial logistic emission model or some ordinal emission
model.
Although f (Z; W ) can be any non-linear mapping, the
ConvNet parameterization of f (Z; W ) makes it particularly
close to the original factor analysis. Specifically, we can write
the top-down ConvNet as follows:
Z (l−1) = fl (Wl Z (l) + bl ),
The expectation with respect to p(Z|Y, W ) can be approximated by drawing samples from p(Z|Y, W ) and then computing the Monte Carlo average.
The Langevin dynamics for sampling Z ∼ p(Z|Y, W )
iterates
(2)
where fl is element-wise non-linearity at layer l, Wl is the
matrix of connection weights, bl is the vector of bias terms
at layer l, and W = (Wl , bl , l = 1, ..., L). Z (0) = f (Z; W ),
and Z (L) = Z. The top-down ConvNet (2) can be considered
a recursion of the original factor analysis model, where the
factors at the layer l − 1 are obtained by the linear superposition of the basis vectors or basis functions that are column
vectors of Wl , with the factors at the layer l serving as the coefficients of the linear superposition. In the case of ConvNet,
the basis functions are shift-invariant versions of one another,
like wavelets. See Appendix for an in-depth understanding
of the model.
3
Zτ +1 = Zτ + sUτ +
s2 1
∂
f
(Z
;
W
)
−
Z
(Y
−
f
(Z
;
W
))
, (5)
τ
τ
τ
2 σ2
∂Z
where τ denotes the time step for the Langevin sampling, s
is the step size, and Uτ denotes a random vector that follows
N(0, Id ). The Langevin dynamics (5) is an explain-away
process, where the latent factors in Z compete to explain
away the current residual Y − f (Zτ ; W ).
To explain Langevin dynamics, its continuous time version for sampling
√ π(x) ∝ exp[−E(x)] is xt+∆t = xt −
0
∆tE (xt )/2 + ∆tUt . The dynamics has π as its stationary distribution, because it can be shown that for any wellbehaved testing function h, if xt ∼ π, then E[h(xt+∆t )] −
E[h(xt )] → 0, as ∆t → 0, so that xt+∆t ∼ π. Alternatively, given xt = x, suppose xt+∆t ∼ K(x, y), then
[π(y)K(y, x)]/[π(x)K(x, y)] → 1 as ∆t → 0.
The stochastic gradient algorithm of (Younes, 1999) can be
used for learning, where in each iteration, for each Zi , only a
single copy of Zi is sampled from p(Zi |Yi , W ) by running a
finite number of steps of Langevin dynamics starting from the
current value of Zi , i.e., the warm start. With {Zi } sampled
Alternating back-propagation
If we observe a training set of data vectors {Yi , i = 1, ..., n},
then each Yi has a corresponding Zi , but all the Yi share
the same ConvNet W . Intuitively, we should infer
Pn{Zi } and
learn W to minimize the reconstruction error i=1 ||Yi −
f (Zi ; W )||2 plus a regularization term that corresponds to
the prior on Z.
More formally, the model can be written as Z ∼ p(Z) and
[Y |Z, W ] ∼ p(Y |Z, W ). Adopting the language of the EM
3
is close to the prior N(0, Id ). If σ 2 is very small, then the
posterior may be multi-modal, but the evolving energy landscape of p(Zi |Yi , W ) may help alleviate the trapping of the
local modes. In practice, we tune the value of σ 2 instead of
estimating it. The Langevin dynamics can be extended to
Hamiltonian Monte Carlo (Neal, 2011) or more sophisticated
versions (Girolami and Calderhead, 2011).
in this manner, we can update the parameter W based on the
gradient L0 (W ), whose Monte Carlo approximation is:
L0 (W ) ≈
n
X
∂
log p(Yi , Zi ; W )
∂W
i=1
n
X
∂ 1
kYi − f (Zi ; W )k2
2
∂W
2σ
i=1
=
−
=
n
X
1
∂
(Yi − f (Zi ; W ))
f (Zi ; W ).(6)
2
σ
∂W
i=1
4
Experiments
The code in our experiments is based on the MatConvNet
package of (Vedaldi and Lenc, 2015).
The training images and sounds are scaled so that the intensities are within the range [−1, 1]. We adopt the structure of
the generator network of (Radford, Metz, and Chintala, 2016;
Dosovitskiy, Springenberg, and Brox, 2015), where the topdown network consists of multiple layers of deconvolution by
linear superposition, ReLU non-linearity, and up-sampling,
with tanh non-linearity at the bottom-layer (Radford, Metz,
and Chintala, 2016) to make the signals fall within [−1, 1].
We also adopt batch normalization (Ioffe and Szegedy, 2015).
We fix σ = .3 for the standard deviation of the noise
vector . We use l = 10 or 30 steps of Langevin dynamics
within each learning iteration, and the Langevin step size
s is set at .1 or .3. We run T = 600 learning iterations,
with learning rate .0001, and momentum .5. The learning
algorithm produces the learned network parameters W and
the inferred latent factors Z for each signal Y in the end. The
synthesized signals are obtained by f (Z; W ), where Z is
sampled from the prior distribution N(0, Id ).
Algorithm 1 describes the details of the learning and sampling algorithm.
Algorithm 1 Alternating back-propagation
Require:
(1) training examples {Yi , i = 1, ..., n}
(2) number of Langevin steps l
(3) number of learning iterations T
Ensure:
(1) learned parameters W
(2) inferred latent factors {Zi , i = 1, ..., n}
1: Let t ← 0, initialize W .
2: Initialize Zi , for i = 1, ..., n.
3: repeat
4:
Inferential back-propagation: For each i, run l steps
of Langevin dynamics to sample Zi ∼ p(Zi |Yi , W )
with warm start, i.e., starting from the current Zi , each
step follows equation (5).
5:
Learning back-propagation: Update W ← W +
γt L0 (W ), where L0 (W ) is computed according to
equation (6), with learning rate γt .
6:
Let t ← t + 1
7: until t = T
4.1
If the Gaussian noise Uτ in the Langevin dynamics (5) is
removed, then the above algorithm becomes the alternating
gradient descent algorithm. It is possible to update both W
and {Zi } simultaneously by joint gradient descent.
Both the inferential back-propagation and the learning back-propagation are guided by the residual Yi −
f (Zi ; W ). The inferential back-propagation is based on
∂f (Z; W )/∂Z, whereas the learning back-propagation is
based on ∂f (Z; W )/∂W . Both gradients can be efficiently
computed by back-propagation. The computations of the
two gradients share most of their steps. Specifically, for
the top-down ConvNet defined by (2), ∂f (Z; W )/∂W and
∂f (Z; W )/∂Z share the same code for the chain rule computation of ∂Z (l−1) /∂Z (l) for l = 1, ..., L. Thus, the code
for ∂f (Z; W )/∂Z is part of the code for ∂f (Z; W )/∂W .
In Algorithm 1, the Langevin dynamics samples from a
gradually changing posterior distribution p(Zi |Yi , W ) because W keeps changing. The updating of both Zi and W collaborate to reduce the reconstruction error kYi − f (Zi ; W )||2 .
The parameter σ 2 plays the role of annealing or tempering
in Langevin sampling. If σ 2 is very large, then the posterior
Qualitative experiments
Figure 1: Modeling texture patterns. For each example, Left:
the 224×224 observed image. Right: the 448×448 generated
image.
Experiment 1. Modeling texture patterns. We learn a
separate model from each texture image. The images are
collected from the Internet, and then resized to 224× 224.
The synthesized images are 448 × 448. Figures 1 shows four
examples.
√
√
The factors Z at the top layer form a d × d image,
√ with
√
each pixel following N(0, 1) independently. The d × d
4
image Z is then transformed to Y by the top-down ConvNet. We use d = 72 in the learning stage for all the texture
experiments. In order to obtain the synthesized image, we
randomly sample a 14 × 14 Z from N(0, I), and then expand
the learned network W to generate the 448 × 448 synthesized
image f (Z; W ).
The training network is as follows. Starting from 7 × 7
image Z, the network has 5 layers of deconvolution with 5×5
kernels (i.e., linear superposition of 5 × 5 basis functions),
with an up-sampling factor of 2 at each layer (i.e., the basis
functions are 2 pixels apart). The number of channels in
the first layer is 512 (i.e., 512 translation invariant basis
functions), and is decreased by a factor 2 at each layer. The
Langevin steps l = 10 with step size s = .1.
Figure 3: Modeling object patterns. Left: the synthesized
images generated by our method. They are generated by
f (Z; W ) with the learned W , where Z = (z1 , z2 ) ∈
[−2, 2]2 , and Z is discretized into 9 × 9 values. Right: the
synthesized images generated using Deep Convolutional Generative Adversarial Net (DCGAN). Z is discretized into 9 × 9
values within [−1, 1]2 .
Figure 2: Modeling sound patterns. Row 1: the waveform
of the training sound (the range is 0-5 seconds). Row 2: the
waveform of the synthesized sound (the range is 0-11 seconds).
Figure 4: Modeling object patterns. Left: each image generated by our method is obtained by first sampling Z ∼
N(0, I100 ) and then generating the image by f (Z; W ) with
the learned W . Middle: interpolation. The images at the four
corners are reconstructed from the inferred Z vectors of four
images randomly selected from the training set. Each image
in the middle is obtained by first interpolating the Z vectors
of the four corner images, and then generating the image
by f (Z; W ). Right: the synthesized images generated by
DCGAN, where Z is a 100 dimension vector sampled from
uniform distribution.
Experiment 2. Modeling sound patterns. A sound signal
can be treated as a one-dimensional texture image (McDermott and Simoncelli, 2011). The sound data are collected
from the Internet. Each training signal is a 5 second clip with
the sampling rate of 11025 Hertz and is represented as a
1 × 60000 vector. We learn a separate model from each sound
signal.
The latent factors Z form a sequence that follows N(0, Id ),
with d = 6. The top-down network consists of 4 layers of
deconvolution with kernels of size 1 × 25, and up-sampling
factor of 10. The number of channels in the first layer is 256,
and decreases by a factor of 2 at each layer. For synthesis,
we start from a longer Gaussian white noise sequence Z with
d = 12 and generate the synthesized sound by expanding
the learned network. Figure 2 shows the waveforms of the
observed sound signal in the first row and the synthesized
sound signal in the second row.
Experiment 3. Modeling object patterns. We model object
patterns using the network structure that is essentially the
same as the network for the texture model, except that we
include a fully connected layer under the latent factors Z,
now a d-dimensional vector. The images are 64 × 64. We use
ReLU with a leaking factor .2 (Maas, Hannun, and Ng, 2013;
Xu et al., 2015). The Langevin steps l = 30 with step size
s = .3.
In the first experiment, we learn a model where Z has two
components, i.e., Z = (z1 , z2 ), and d = 2. The training data
are 11 images of 6 tigers and 5 lions. After training the model,
we generate images using the learned top-down ConvNet for
(z1 , z2 ) ∈ [−2, 2]2 , where we discretize both z1 and z2 into
9 equally spaced values. The left panel of Figure 3 displays
the synthesized images on the 9 × 9 panel.
In the second experiment, we learn a model with d = 100
from 1000 face images randomly selected from the CelebA
dataset (Liu et al., 2015). The left panel of Figure 4 displays
the images generated by the learned model. The middle panel
displays the interpolation results. The images at the four corners are generated by the Z vectors of four images randomly
selected from the training set. The images in the middle are
obtained by first interpolating the Z’s of the four corner images using the sphere interpolation (Dinh, Sohl-Dickstein,
and Bengio, 2016) and then generating the images by the
learned ConvNet.
We also provide qualitative comparison with Deep Convolutional Generative Adversarial Net (DCGAN) (Goodfellow et al., 2014; Radford, Metz, and Chintala, 2016).
The right panel of Figure 3 shows the generated results
5
for the lion-tiger dataset using 2-dimensional Z. The right
panel of Figure 4 displays the generated results trained on
1000 aligned faces from celebA dataset, with d = 100. We
use the code from https://github.com/carpedm20/
DCGAN-tensorflow, with the tuning parameters as in (Radford, Metz, and Chintala, 2016). We run T = 600 iterations
as in our method.
Experiment 4. Modeling dynamic patterns. We model a
textured motion (Wang and Zhu, 2003) or a dynamic texture (Doretto et al., 2003) by a non-linear dynamic system
Yt = f (Zt ; W ) + t , and Zt+1 = AZt + ηt , where we assume the latent factors follow a vector auto-regressive model,
where A is a d×d matrix, and ηt ∼ N(0, Q) is the innovation.
This model is a direct generalization of the linear dynamic
system of (Doretto et al., 2003), where Yt is reduced to Zt
by principal component analysis (PCA) via singular value
decomposition (SVD). We learn the model in two steps. (1)
Treat {Yt } as independent examples and learn W and infer
{Zt } as before. (2) Treat {Zt } as the training data, learn A
and Q as in (Doretto et al., 2003). After that, we can synthesize a new dynamic texture. We start from Z0 ∼ N(0, Id ),
and then generate the sequence according to the learned
model (we discard a burn-in period of 15 frames). Figure
5 shows some experiments, where we set d = 20. The first
row is a segment of the sequence generated by our model,
and the second row is generated by the method of (Doretto et
al., 2003), with the same dimensionality of Z. It is possible
to generalize the auto-regressive model of Zt to recurrent network. We may also treat the video sequences as 3D images,
and learn generator networks with 3D spatial-temporal filters
or basis functions.
puted by summing over all the pixels. For a partially observed
image, we compute it by summing over only the observed
pixels. Then we can continue to use the alternating backpropagation algorithm to infer Z and learn W . With inferred
Z and learned W , the image can be automatically recovered
by f (Z; W ). In the end, we will be able to accomplish the
following tasks: (T1) Recover the occluded pixels of training
images. (T2) Synthesize new images from the learned model.
(T3) Recover the occluded pixels of testing images using the
learned model.
experiment
error
P.7
.0662
P.9
.0771
M20
.0773
M30
.1035
Table 1: Recovery errors in 5 experiments of learning from
occluded images.
Figure 6: Learning from incomplete data. The 10 columns
belong to experiments P.5, P.7, P.9, P.9, P.9, P.9, P.9, M20,
M30, M30 respectively. Row 1: original images, not observed
in learning. Row 2: training images. Row 3: recovered images
during learning.
We want to emphasize that in our experiments, all the
training images are partially occluded. Our experiments are
different from (1) de-noising auto-encoder (Vincent et al.,
2008), where the training images are fully observed, and
noises are added as a matter of regularization, (2) in-painting
or de-noising, where the prior model or regularization has already been learned or given. (2) is about task (T3) mentioned
above, but not about tasks (T1) and (T2).
Learning from incomplete data can be difficult for GAN
and VAE, because the occluded pixels are different for different training images.
We evaluate our method on 10,000 images randomly selected from CelebA dataset. We design 5 experiments, with
two types of occlusions: (1) 3 experiments are about salt and
pepper occlusion, where we randomly place 3 × 3 masks on
the 64 × 64 image domain to cover roughly 50%, 70% and
90% of pixels respectively. These 3 experiments are denoted
P.5, P.7, and P.9 respectively (P for pepper). (2) 2 experiments
are about single region mask occlusion, where we randomly
place a 20×20 or 30×30 mask on the 64×64 image domain.
These 2 experiments are denoted M20 and M30 respectively
(M for mask). We set d = 100. Table 1 displays the recovery
errors of the 5 experiments, where the error is defined as
per pixel difference (relative to the range of the pixel values)
between the original image and the recovered image on the
occluded pixels. We emphasize that the recovery errors are
not training errors, because the intensities of the occluded
Figure 5: Modeling dynamic textures. Row 1: a segment of
the synthesized sequence by our method. Row 2: a sequence
by the method of (Doretto et al., 2003). Rows 3 and 4: two
more sequences by our method.
4.2
P.5
.0571
Quantitative experiments
Experiment 5. Learning from incomplete data. Our method
can learn from images with occluded pixels. This task is
inspired by the fact that most of the images contain occluded
objects. It can be considered a non-linear generalization of
matrix completion in recommender system.
Our method can be adapted to this task with minimal modification. The only modification involves the computation
of kY − f (Z; W )k2 . For a fully observed image, it is com6
experiment
ABP
PCA
pixels are not observed in training. Figure 6 displays recovery
results. In experiment P.9, 90% of pixels are occluded, but
we can still learn the model and recover the original images.
experiment
error
d = 20
.0795
d = 60
.0617
d = 20
.0810
.1038
d = 60
.0617
.0820
d = 100
.0549
.0722
d = 200
.0523
.0621
Table 3: Reconstruction errors on testing images, after learning from training images using our method (ABP) and PCA.
d = 100
.0625
Table 2: Recovery errors in 3 experiments of learning from
compressively sensed images.
Figure 8: Comparison between our method and PCA. Row
1: original testing images. Row 2: reconstructions by PCA
eigenvectors learned from training images. Row 3: reconstructions by the generator learned from training images.
d = 20 for both methods.
Figure 7: Learning from indirect data. Row 1: the original
64 × 64 × 3 images, which are projected onto 1,000 white
noise images. Row 2: the recovered images during learning.
Experiment 6. Learning from indirect data. We can learn
the model from the compressively sensed data (Candès,
Romberg, and Tao, 2006). We generate a set of white noise
images as random projections. We then project the training images on these white noise images. We can learn the
model from the random projections instead of the original images. We only need to replace kY − f (Z; W )k2 by
kSY − Sf (Z; W )k2 , where S is the given white noise sensing matrix, and SY is the observation. We can treat S as a
fully connected layer of known filters below f (Z; W ), so that
we can continue to use alternating back-propagation to infer
Z and learn W , thus recovering the image by f (Z; W ). In
the end, we will be able to (T1) Recover the original images
from their projections during learning. (T2) Synthesize new
images from the learned model. (T3) Recover testing images
from their projections based on the learned model. Our experiments are different from traditional compressed sensing,
which is task (T3), but not tasks (T1) and (T2). Moreover, the
image recovery in our work is based on non-linear dimension
reduction instead of linear sparsity.
We evaluate our method on 1000 face images randomly
selected from CelebA dataset. These images are projected
onto K = 1000 white noise images with each pixel randomly
sampled from N(0, .52 ). After this random projection, each
image of size 64 × 64 × 3 becomes a K-dimensional vector.
We show the recovery errors for different latent dimensions d
in Table 2, where the recovery error is defined as the per pixel
difference (relative to the range of the pixel values) between
the original image and the recovered image. Figure 7 shows
some recovery results.
Experiment 7. Model evaluation by reconstruction error
on testing data. After learning the model from the training
images (now assumed to be fully observed), we can evaluate
the model by the reconstruction error on the testing images.
We randomly select 1000 face images for training and 300
images for testing from CelebA dataset. After learning, we
infer the latent factors Z for each testing image using inferential back-propagation, and then reconstruct the testing image
by f (Z; W ) using the inferred Z and the learned W . In the
inferential back-propagation for inferring Z, we initialize
Z ∼ N(0, Id ), and run 300 Langevin steps with step size .05.
Table 3 shows the reconstruction errors of alternating backpropagation learning (ABP) as compared to PCA learning
for different latent dimensions d. Figure 8 shows some reconstructed testing images. For PCA, we learn the d eigenvectors
from the training images, and then project the testing images
on the learned eigenvectors for reconstruction.
Experiments 5-7 may be used to evaluate generative models in general. Experiments 5 and 6 appear new, and we have
not found comparable methods that can accomplish all three
tasks (T1), (T2), and (T3) simultaneously.
5
Conclusion
This paper proposes an alternating back-propagation algorithm for training the generator network. We recognize that
the generator network is a non-linear generalization of the
factor analysis model, and develop the alternating backpropagation algorithm as the non-linear generalization of
the alternating regression scheme of the Rubin-Thayer EM
algorithm for fitting the factor analysis model. The alternating back-propagation algorithm iterates the inferential backpropagation for inferring the latent factors and the learning
back-propagation for updating the parameters. Both backpropagation steps share most of their computing steps in the
chain rule calculations.
Our learning algorithm is perhaps the most canonical algorithm for training the generator network. It is based on
maximum likelihood, which is theoretically the most accurate estimator. The maximum likelihood learning seeks to
explain and charge the whole dataset uniformly, so that there
is little concern of under-fitting or biased fitting.
As an unsupervised learning algorithm, the alternating
back-propagation algorithm is a natural generalization of
the original back-propagation algorithm for supervised learning. It adds an inferential back-propagation step to the learn7
Y = Wδ Z + bδ + , for an overall bias term that depends on
δ. So the distribution of Y is essentially piecewise Gaussian.
The generator model can be considered an explicit implementation of the local linear embedding (Roweis and Saul,
2000), where Z is the embedding of Y . In local linear embedding, the mapping between Z and Y is implicit. In the
generator model, the mapping from Z to Y is explicit. With
ReLU ConvNet, the mapping is piecewise linear, which is
consistent with local linear embedding, except that the partition of the linear pieces by δ(Z; W ) in the generator model
is learned automatically.
The inferential back-propagation is a Langevin dynamics
on the energy function kY − f (Z; W )k2 /(2σ 2 ) + kZk2 /2.
With f (Z; W ) = Wδ Z, ∂f (Z; W )/∂Z = Wδ . If Z belongs to the piece defined by δ, then the inferential backpropagation seeks to approximate Y by the basis Wδ via
a ridge regression. Because Z keeps changing during the
Langevin dynamics, δ(Z; W ) may also be changing, and the
algorithm searches for the optimal reconfigurable basis Wδ
to approximate Y . We may solve Z by second-order methods
such as iterated ridge regression, which can be computationally more expensive than the simple gradient descent.
ing back-propagation step, with minimal overhead in coding and affordable overhead in computing. The inferential
back-propagation seeks to perform accurate explaining-away
inference of the latent factors. It can be worthwhile for tasks
such as learning from incomplete or indirect data, or learning
models where the latent factors themselves follow sophisticated prior models with unknown parameters. The inferential
back-propagation may also be used to evaluate the generators
learned by other methods on tasks such as reconstructing or
completing testing data.
Our method or its variants can be applied to non-linear
matrix factorization and completion. It can also be applied to
problems where some components or aspects of the factors
are supervised.
Code, images, sounds, and videos
http://www.stat.ucla.edu/~ywu/ABP/main.html
Acknowledgement
We thank Yifei (Jerry) Xu for his help with the experiments
during his 2016 summer visit. We thank Jianwen Xie for
helpful discussions.
The work is supported by NSF DMS 1310391, DARPA
SIMPLEX N66001-15-C-4035, ONR MURI N00014-16-12007, and DARPA ARO W911NF-16-1-0579.
6
6.1
6.2
EM, density mapping, and density shifting
Suppose the training data {Yi , i = 1, ..., n} come from a
data distribution Pdata (Y ). To understand how the alternating back-propagation algorithm or its EM idealization maps
the prior distribution of the latent factors p(Z) to the data
distribution Pdata (Y ) by the learned g(Z; W ), we define
Appendix
ReLU and piecewise factor analysis
The generator network is Y = f (Z; W ) + , Z (l−1) =
fl (Wl Z (l) + bl ), l = 1, ..., L, with Z (0) = f (Z; W ), and
Z (L) = Z. The element-wise non-linearity fl in modern
ConvNet is usually the two-piece linearity, such as rectified
linear unit (ReLU) (Krizhevsky, Sutskever, and Hinton, 2012)
or the leaky ReLU (Maas, Hannun, and Ng, 2013; Xu et al.,
2015). Each ReLU unit corresponds to a binary switch. For
the case of non-leaky ReLU, following the analysis of (Pascanu, Montufar, and Bengio, 2013), we can write Z (l−1) =
δl (Wl Z (l) + bl ), where δl = diag(1(Wl Z (l) + bl > 0)) is a
diagonal matrix, 1() is an element-wise indicator function.
For the case of leaky ReLU, the 0 values on the diagonal are
replaced by a leaking factor (e.g., .2).
δ = (δl , l = 1, ..., L) forms a classification of Z according to the network W . Specifically, the factor space
of Z is divided into a large number of pieces by the hyperplanes Wl Z (l) + bl = 0, and each piece is indexed by
an instantiation of δ. We can write δ = δ(Z; W ) to make
explicit its dependence on Z and W . On the piece indexed
by δ, f (Z; W ) = Wδ Z + bδ . Assuming bl = 0, ∀l, for
simplicity, we have Wδ = δ1 W1 ...δL WL . Thus each piece
defined by δ = δ(Z; W ) corresponds to a linear factor
analysis Y = Wδ Z + , whose basis Wδ is a multiplicative recomposition of the basis functions at multiple layers
(Wl , l = 1, ..., L), and the recomposition is controlled by
the binary switches at multiple layers δ = (δl , l = 1, ..., L).
Hence the top-down ConvNet amounts to a reconfigurable
basis Wδ for representing Y , and the model is a piecewise
linear factor analysis. If we retain the bias term, we will have
Pdata (Z, Y ; W ) = Pdata (Y )p(Z|Y, W )
= Pdata (Z; W )Pdata (Y |Z, W ),
(7)
R
where Pdata (Z; W ) =
p(Z|Y, W )Pdata (Y )dY is obtained by averaging the posteriors p(Z|Y ; W ) over the observed data Y ∼ Pdata . That is, Pdata (Z; W ) can be considered the data prior. The data prior Pdata (Z; W ) is close to
the true prior p(Z) in the sense that
KL(Pdata (Z; W )|p(Z)) ≤ KL(Pdata (Y )|p(Y ; W ))
= KL(Pdata (Z, Y ; W )|p(Z, Y ; W )).
(8)
The right hand side of (8) is minimized at the maximum
likelihood estimate Ŵ , hence the data prior Pdata (Z; Ŵ ) at
Ŵ should be especially close to the true prior p(Z). In other
words, at Ŵ , the posteriors p(Z|Y, Ŵ ) of all the data points
Y ∼ Pdata tend to pave the true prior p(Z).
From Rubin’s multiple imputation point of view (Rubin, 2004) of the EM algorithm, the E-step of EM infers
(m)
Zi
∼ p(Zi |Yi , Wt ) for m = 1, ..., M , where M is the
number of multiple imputations or multiple guesses of Zi .
The multiple guesses account for the uncertainty in inferring Zi from Yi . The M-step of EM maximizes Q(W ) =
Pn P
(m)
M
; W ) to obtain Wt+1 . For each
i=1
m=1 log p(Yi , Zi
data point Yi , Wt+1 seeks to reconstruct Yi by g(Z; W ) from
(m)
the inferred latent factors {Zi , m = 1, ..., M }. In other
(m)
words, the M-step seeks to map {Zi } to Yi . Pooling over
(m)
all i = 1, ..., n, {Zi , ∀i, m} ∼ Pdata (Z; Wt ), hence the
8
If σ 2 is unknown, it can be obtained by averaging the diagonal
elements of Σ. The computation can again be done by the
sweep operator on S, with SZZ being the pivotal matrix.
The E-step is based on the multivariate linear regression of
Z on Y given W . The M-step updates W by the multivariate
linear regression of Y on Z. Both steps can be accomplished
by the sweep operator. We use the notation S and S for the
Gram matrices to highlight the analogy between the two steps.
The EM algorithm can then be considered alternating linear
regression or alternating sweep operation, which serves as a
prototype for alternating back-propagation.
M-step seeks to map Pdata (Z; Wt ) to the data distribution
(m)
Pdata (Y ). Of course the mapping from {Zi } to Yi cannot
(m)
be exact. In fact, g(Z; W ) maps {Zi } to a d-dimensional
patch around the D-dimensional Yi . The local patches for all
{Yi , ∀i} patch up the d-dimensional manifold form by the
D-dimensional observed examples and their interpolations.
The EM algorithm is a process of density shifting, so that
Pdata (Z; W ) shifts towards p(Z), thus g(Z; W ) maps p(Z)
to Pdata (Y ).
6.3
Factor analysis and alternating regression
The alternating back-propagation algorithm is inspired by
Rubin-Thayer EM algorithm for factor analysis, where both
the observed data model p(Y |W ) and the posterior distribution p(Z|Y, W ) are available in closed form. The EM algorithm for factor analysis can be interpreted as alternating
linear regression (Rubin and Thayer, 1982; Liu, Rubin, and
Wu, 1998).
In the factor analysis model Z ∼ N(0, Id ), Y = W Z + ,
∼ N(0, σ 2 ID ). The joint distribution of (Z, Y ) is
Z
0
I
W>
∼N
, d
.
(9)
Y
0
W W W > + σ 2 ID
References
Candès, E. J.; Romberg, J.; and Tao, T. 2006. Robust uncertainty principles: Exact signal reconstruction from highly
incomplete frequency information. IEEE Transactions on
information theory 52(2):489–509.
Dempster, A. P.; Laird, N. M.; and Rubin, D. B. 1977. Maximum likelihood from incomplete data via the em algorithm.
Journal of the Royal Statistical Society: B 1–38.
Denton, E. L.; Chintala, S.; Fergus, R.; et al. 2015. Deep
generative image models using a laplacian pyramid of
adversarial networks. In NIPS, 1486–1494.
Denote
S
S = ZZ
SY Z
SZY
SY Y
E[ZZ > ] E[ZY > ]
=
E[Y Z > ] E[Y Y > ]
Id
W>
=
.
W W W > + σ 2 ID
Dinh, L.; Sohl-Dickstein, J.; and Bengio, S. 2016. Density
estimation using real nvp. CoRR abs/1605.08803.
(10)
Doretto, G.; Chiuso, A.; Wu, Y.; and Soatto, S. 2003. Dynamic textures. IJCV 51(2):91–109.
The posterior distribution p(Z|Y, W ) can be obtained by
linear regression of Z on Y , [Z|Y, W ] ∼ N(βY, V ), where
Dosovitskiy, E.; Springenberg, J. T.; and Brox, T. 2015.
Learning to generate chairs with convolutional neural networks. In CVPR.
β = SZY SY−1Y ,
(11)
V = SZZ − SZY SY−1Y SY Z .
(12)
Girolami, M., and Calderhead, B. 2011. Riemann manifold
langevin and hamiltonian monte carlo methods. Journal
of the Royal Statistical Society: B 73(2):123–214.
Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; WardeFarley, D.; Ozair, S.; Courville, A.; and Bengio, Y. 2014.
Generative adversarial nets. In NIPS, 2672–2680.
The above computation can be carried out by the sweep
operator on S, with SY Y being the pivotal matrix.
Suppose we have observations {Yi , i = 1, ..., n}. In the
E-step, we compute
E[Zi |Yi , W ] = βYi ,
E[Zi Zi> |Yi , W ]
=V +
In the M-step, we compute
SZZ SZY
S=
SY Z SY Y
Pn
>
Pni=1 E[Zi Zi >]/n
=
i=1 Yi E[Zi ] /n
Hyvärinen, A.; Karhunen, J.; and Oja, E. 2004. Independent
component analysis. John Wiley & Sons.
(13)
βYi Yi> β > .
Ioffe, S., and Szegedy, C. 2015. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML.
(14)
Kim, H., and Park, H. 2008. Nonnegative matrix factorization based on alternating nonnegativity constrained least
squares and active set method. SIAM Journal on Matrix
Analysis and Applications 30(2):713–730.
Pn
>
i=1 E[Zi ]Yi /n , (15)
P
n
>
i=1 Yi Yi /n
Kingma, D. P., and Welling, M. 2014. Auto-encoding variational bayes. In ICLR.
E[Zi Zi> ]
where we use E[Zi ] and
to denote the conditional
expectations in (13) and (14). Then we regress Y on Z to
obtain the coefficient vector and residual variance-covariance
matrix
−1
W = SY Z SZZ
(16)
Σ = SY Y − SY Z S−1
ZZ SZY .
(17)
Koren, Y.; Bell, R.; and Volinsky, C. 2009. Matrix factorization techniques for recommender systems. Computer
42(8):30–37.
Krizhevsky, A.; Sutskever, I.; and Hinton, G. E. 2012. Imagenet classification with deep convolutional neural networks. In NIPS, 1097–1105.
9
LeCun, Y.; Bottou, L.; Bengio, Y.; and Haffner, P. 1998.
Gradient-based learning applied to document recognition.
Proceedings of the IEEE 86(11):2278–2324.
Lee, D. D., and Seung, H. S. 2001. Algorithms for nonnegative matrix factorization. In NIPS, 556–562.
Liu, Z.; Luo, P.; Wang, X.; and Tang, X. 2015. Deep learning
face attributes in the wild. In ICCV, 3730–3738.
Liu, C.; Rubin, D. B.; and Wu, Y. N. 1998. Parameter expansion to accelerate em: The px-em algorithm. Biometrika
85(4):755–770.
Lu, Y.; Zhu, S.-C.; and Wu, Y. N. 2016. Learning FRAME
models using CNN filters. In AAAI.
Maas, A. L.; Hannun, A. Y.; and Ng, A. Y. 2013. Rectifier
nonlinearities improve neural network acoustic models. In
ICML.
McDermott, J. H., and Simoncelli, E. P. 2011. Sound texture
perception via statistics of the auditory periphery: evidence
from sound synthesis. Neuron 71(5):926–940.
Mnih, A., and Gregor, K. 2014. Neural variational inference
and learning in belief networks. In ICML.
Neal, R. M. 2011. Mcmc using hamiltonian dynamics. Handbook of Markov Chain Monte Carlo 2.
Olshausen, B. A., and Field, D. J. 1997. Sparse coding with
an overcomplete basis set: A strategy employed by v1?
Vision Research 37(23):3311–3325.
Pascanu, R.; Montufar, G.; and Bengio, Y. 2013. On the
number of response regions of deep feed forward networks
with piece-wise linear activations. arXiv:1312.6098.
Radford, A.; Metz, L.; and Chintala, S. 2016. Unsupervised
representation learning with deep convolutional generative
adversarial networks. In ICLR.
Rezende, D. J.; Mohamed, S.; and Wierstra, D. 2014. Stochastic backpropagation and approximate inference in deep
generative models. In NIPS, 1278–1286.
Roweis, S. T., and Saul, L. K. 2000. Nonlinear dimensionality reduction by locally linear embedding. Science
290(5500):2323–2326.
Rubin, D. B., and Thayer, D. T. 1982. Em algorithms for ml
factor analysis. Psychometrika 47(1):69–76.
Rubin, D. B. 2004. Multiple imputation for nonresponse in
surveys, volume 81. John Wiley & Sons.
Vedaldi, A., and Lenc, K. 2015. Matconvnet – convolutional
neural networks for matlab. In Int. Conf. on Multimedia.
Vincent, P.; Larochelle, H.; Bengio, Y.; and Manzagol, P.A. 2008. Extracting and composing robust features with
denoising autoencoders. In ICML, 1096–1103.
Wang, Y., and Zhu, S.-C. 2003. Modeling textured motion:
Particle, wave and sketch. In ICCV, 213–220.
Xie, J.; Lu, Y.; Zhu, S.-C.; and Wu, Y. N. 2016. A theory of
generative convnet. In ICML.
Xu, B.; Wang, N.; Chen, T.; and Li, M. 2015. Empirical
evaluation of rectified activations in convolutional network.
CoRR abs/1505.00853.
Younes, L. 1999. On the convergence of markovian stochastic algorithms with rapidly decreasing ergodicity rates.
Stochastics: An International Journal of Probability and
Stochastic Processes 65(3-4):177–228.
10
| 9cs.NE
|
Delimited continuations in natural language
Quantification and polarity sensitivity
Chung-chieh Shan
Harvard University
33 Oxford Street
Cambridge, MA 02138 USA
arXiv:cs/0404006v1 [cs.CL] 5 Apr 2004
[email protected]
ABSTRACT
(1)
Making a linguistic theory is like making a programming
language: one typically devises a type system to delineate
the acceptable utterances and a denotational semantics to
explain observations on their behavior. Via this connection,
the programming language concept of delimited continuations can help analyze natural language phenomena such as
quantification and polarity sensitivity. Using a logical metalanguage whose syntax includes control operators and whose
semantics involves evaluation order, these analyses can be
expressed in direct style rather than continuation-passing
style, and these phenomena can be thought of as computational side effects.
The sentence in (2) is ambiguous between at least two readings. On one reading, the speaker must decline to run any
spot that fails to substantiate any claims whatsoever. On
another reading, there exist certain claims (anti-war ones,
say) such that the speaker must decline to run any spot
that fails to substantiate them.
Categories and Subject Descriptors
D.3.3 [Programming languages]: Language constructs
and features—control structures; J.5 [Linguistics]
(2)
Languages, Theory
Keywords
Delimited continuations, control effects, natural language
semantics, quantification, polarity sensitivity
1.
INTRODUCTION
This paper is about computational linguistics, in the sense
of applying insights from computer science to linguistics.
Linguistics strives to scientifically explain empirical observations of natural language. Semantics, in particular, is concerned with phenomena such as the following. In (1) below,
some sentences to the left entail their counterparts to the
right, but others do not.
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
Continuation Workshop 2004 Venice, Italy
Copyright 200X ACM X-XXXXX-XX-X/XX/XX ...$5.00.
We must decline to run any spot that fails to substantiate certain claims.1
Finally, among the four sentences in (3), only (3a) is acceptable. That is, only it can be used in idealized conversation.
The unacceptability of the rest is notated with asterisks.
(3)
General Terms
Every student passed ⊢ Every diligent student passed
No student passed
⊢ No diligent student passed
A student passed
0 A diligent student passed
Most students passed 0 Most diligent students passed
a. No student liked any course.
b. *Every student liked any course.
c. *A student liked any course.
d. *Most students liked any course.
The linguistic entailments and non-entailments in (1) are
facts about English, in that only a speaker of English can
make these judgments. Nevertheless, they presumably have
to do with corresponding logical entailments and non-entailments: both the English speaker who judges that Every student passed entails Every diligent student passed and
the Mandarin speaker who judges that Meige xuesheng dou
jige-le entails Meige yonggong-de xuesheng dou jige-le rely
on knowing that, if every student passed, then every diligent
student passed. Thus the typical linguistic theory specifies
a semantics for natural language by translating declarative
sentences into logical statements with truth conditions. The
linguistic entailments in (1) hold, goes the theory, because
the meanings—truth conditions—of the two sentences are
such that any model that verifies the former also verifies the
latter. Much work in natural language semantics aims in
this way, as depicted in Figure 1, to explain the horizontal by positing the vertical. This approach is reminiscent
of programming language research where an ill-understood
language (perhaps one with a complicating feature like ex1
This sentence is part of a statement made by the cable
television company Comcast after its CNN channel rejected
an anti-war commercial hours before it was scheduled to air
on January 28, 2003.
Every student passed
∀x. student(x) ⇒ passed(x)
⊢
Every diligent student passed
⊢ ∀x. student(x) ∧ diligent(x) ⇒ passed(x)
hsome truth condition on modelsi ⊢ hsome other truth condition on modelsi
Figure 1: The translation/denotation approach to natural language semantics
ceptions) is studied by translation into a simpler language
(without exceptions) that is better understood.
The translation target posited in natural language semantics is often some combination of the λ-calculus and predicate logic. For example, the verb passed might be translated as λx. passed(x). This paper argues by example that
the translation target should be a logical metalanguage with
first-class delimited continuations. The examples are two
natural language phenomena: quantification by words like
every and most in (1), and polarity sensitivity on the part
of words like any in (3).
Quantification was first analyzed explicitly using continuations by Barker (2002). Building on that insight, this
paper makes the following two contributions. First, I analyze natural language in direct style rather than in continuation-passing style. In other words, the logical metalanguage
used here is one that includes control operators for delimited continuations, rather than a pure λ-calculus in which
denotations need to handle continuations explicitly. Natural language is thus endowed with an operational semantics
from computer science that is richer than just βη-reduction.
Second, I propose a new analysis of polarity sensitivity
that improves upon prior theories in explaining why No student liked any course is acceptable but *Any student liked
no course is not. This analysis crucially relies on the notion
of evaluation order from programming languages, thus elucidating the role of control effects in natural language and
supporting the broader claim that linguistic phenomena can
be fruitfully thought of as computational side effects.
The rest of this paper is organized as follows. In §2,
I introduce a simple grammatical formalism. In §3, I describe the linguistic phenomenon of quantification and show
a straw man analysis that deals with some cases but not
others. I then introduce a programming language with delimited continuations and use it to improve the straw man
analysis: quantification in non-subject position is treated
in §4, and inverse scope is covered in §5. In §6, I turn to the
linguistic phenomenon of polarity sensitivity and show how
a computationally motivated notion of evaluation order improves upon previous analyses. In §7, I place these examples
in a broader context and conclude.
2.
A GRAMMATICAL FORMALISM
In this section, I introduce a simple grammatical formalism for use in the rest of the paper. It is a notational variant
of categorial grammar (as introduced by Carpenter (1997;
chapter 4), for instance).
The verb like usually requires an object to its right and a
subject to its left.
(4)
a. Alice liked CS187.
b. *Alice liked.
c. *Alice liked Bob CS187.
Intuitively, like is a function that takes two arguments, and
the sentences (4b–c) are unacceptable due to type mismatch.
We can model this formally by assigning types to the denotations of Alice, CS187, and liked, which we take to be atomic
expressions.
(5)
(6)
(7)
JAliceK = alice : Thing
JCS187K = cs187 : Thing
JlikedK = liked : Thing → Thing → Bool
Here Thing is the type of individual objects, and Bool is the
type of truth values or propositions. Following (justifiable)
standard practice in linguistics, we let liked take its object as
the first argument and its subject as the second argument.
For example, in (4a), the first argument to liked is cs187,
and the second argument is alice.
As (4a) shows, there are two ways to combine expressions.
A function can take its argument either to its right (combining liked with CS187 ) or to its left (combining Alice with
liked CS187 ). We denote these two cases with two infix operators: “′” for forward combination and “8” for backward
combination. (The tick marks depict the direction in which
a function “leans on” an argument.)
(8)
(9)
f ′ x = f (x) : β
x 8 f = f (x) : β
where f : α → β, x : α
where f : α → β, x : α
We can now derive the sentence (4a)—that is, prove it to
have type Bool. The derivation can be written as a tree (10)
or a term (11).
(10)
Alice
liked
CS187
(11) JAliceK 8 (JlikedK ′ JCS187K) = liked cs187 alice : Bool
By convention, the infix operators ′ and 8 associate to the
right, so parentheses such as those in (11) above are optional.
Unfortunately, the system set up so far derives not only
the acceptable sentence (4a) but also the unacceptable sentence (12), with the same meaning.
(12) *Alice CS187 liked.
The reason the system derives (12) is that the direction of
function application is unconstrained: in the derivation below, liked takes its first (object) argument to the left, which
is usually disallowed in English.
(13)
Alice
(14)
CS187
liked
JAliceK 8 JCS187K 8 JlikedK = liked cs187 alice : Bool
To rule out this derivation of (12) in our type system, we
split the function type constructor “→” into two type con-
structors “→”
´ and “→”,
`
one for each direction of application. Using these new type constructors, we change the
denotation of liked to specify that its first argument is to its
right and its second argument is to its left.
(15)
JlikedK = liked : Thing →
´ Thing →
` Bool
We also revise the combination rules (8) and (9) to require
different function type constructors.
(16)
(17)
f ′x:β
x8f :β
where f : α →
´ β, x : α
where f : α →
` β, x : α
The system now rejects (12) while continuing to accept (4a),
as desired.
3.
the grammar can derive the sentence (18a).
(23)
every
(24)
student
liked
CS187
(JeveryK ′ JstudentK) ′ JlikedK ′ JCS187K = (19)
The existential determiner some can be analyzed similarly:
let some denote
(25) JsomeK = λ′r. λ′s. ∃x. r(x) ∧ x 8 s
: (Thing → Bool) →
´ (Thing →
` Bool) →
´ Bool
to derive the sentence Some student liked CS187.
(26)
QUANTIFICATION
some
student
liked
CS187
The linguistic phenomenon of quantification is illustrated
by the following sentences.
(27)
(18) a. Every student liked CS187.
b. Some student liked every course.
c. Alice consulted Bob before most meetings.
To summarize, we treat determiners like every and some
as functions of two arguments: the restrictor and the scope
of a quantifier, both functions from Thing to Bool. Such
higher-order functions are a popular analysis of natural language determiners, and have been known to semanticists
since Montague (1974) as generalized quantifiers. However,
the simplistic account presented above only handles quantificational noun phrases in subject position, as in (18a) but
not (18b) or (18c). For example, in (18b), neither forward
nor backward combination can apply to join the verb liked,
of type Thing →
´ Thing →
` Bool, to its object every course, of
type (Thing →
` Bool) →
´ Bool. Yet, empirically speaking, the
sentence (18b) is not only acceptable but in fact ambiguous
between two available readings. This problem has prompted
a great variety of supplementary proposals in the linguistics
literature (Barwise and Cooper 1981; Hendriks 1993; May
1985; inter alia). The next section presents a solution using
delimited continuations.
As with the previously encountered sentences, the natural
language semanticist wants to translate English into logical
formulas that account for entailment and other properties.
More precisely, the problem is to posit translation rules that
map these sentences thus. For instance, we would like to
map (18a) to a formula like
(19)
∀x. student(x) ⇒ x 8 liked ′ cs187 : Bool,
where the constants
(20) ∀ : (Thing → Bool) → Bool,
⇒ : Bool → Bool → Bool
are drawn from the (higher-order) abstract syntax of predicate logic. To this end, what should the subject noun phrase
every student denote? Unlike with Alice, there is nothing of
type Thing that the quantificational noun phrase every student can denote and still allow the desired translation (19)
to be generated. At the same time, we would like to retain
the denotation that we previously computed for the verb
phrase liked CS187, namely liked ′ cs187. Taking these considerations into account, one way to translate (18a) to (19)
is for the determiner every to denote
(21) JeveryK = λ′r. λ′s. ∀x. r(x) ⇒ x 8 s
: (Thing → Bool) →
´ (Thing →
` Bool) →
´ Bool.
Here the restrictor r and the scope s are λ-bound variables
intended to receive, respectively, the denotations of the noun
student (of type Thing → Bool) and the verb phrase liked
CS187 (of type Thing →
` Bool). (More precisely, r and s are
λ′ -bound variables; the tick mark again signifies the direction
of function application.) In a non-quantificational sentence
like (4a), the verb phrase takes the subject as its argument;
by contrast, in the quantificational sentence (18a), the subject takes the verb phrase as its argument.
Extended with the lexical entry (21) for every, and assuming that student denotes
(22)
JstudentK = student : Thing → Bool,
(JsomeK ′ JstudentK) ′ JlikedK ′ JCS187K
= ∃x. student(x) ∧ x 8 liked ′ cs187 : Bool
4. DELIMITED CONTINUATIONS
First-class continuations represent “the entire (default) future for the computation” (Kelsey, Clinger, Rees et al. 1998).
Refining this concept, Felleisen (1988) introduced delimited
continuations, which encapsulate only a prefix of that future.
This paper uses shift and reset (Danvy and Filinski 1989,
1990), a popular choice of control operators for delimited
continuations.
To review briefly, the shift operator (notated ξ) captures
the current context of computation, removing it and making
it available to the program as a function. For example, when
evaluating the term
(28)
10 × (ξf. 1 + f (2)),
the variable f is bound to the function that multiplies every
number by 10. Thus the above expression evaluates to 21 via
the following sequence of reductions. (The reduced subexpression at each step is underlined.)
(29) [10 × (ξf. 1 + f (2))]
⊲ [1 + (λv. [10 × v])(2)]
⊲ [1 + [10 × 2]] ⊲ [1 + [20]] ⊲ [1 + 20] ⊲ [21] ⊲ 21
Term reductions are performed deterministically in applicative order: call-by-value and left-to-right.
The reset operator (notated with square brackets [ ]) delineates how far shift can reach: shift captures the current
context of computation up to the closest dynamically enclosing reset. Hence “3 ×” below is out of reach.
(30) 3 × [10 × (ξf. 1 + f (2))]
⊲ 3 × [1 + (λv. [10 × v])(2)]
⊲ 3 × [1 + [10 × 2]] ⊲ 3 × [1 + [20]] ⊲ · · · ⊲ 63
Shift and reset come with an operational semantics (illustrated by the reductions above) as well as a denotational
semantics (via the CPS transform). That both kinds of semantics are available is important to linguistics, because
the meanings of natural language expressions (studied in semantics) need to be related to how humans process them
(studied in psycholinguistics).
Quantificational expressions in natural language can be
thought of as phrases that manipulate their context. In a
sentence like Alice liked CS187 (4a), the context of CS187
is the function mapping each thing x to the proposition that
Alice liked x. Compared to the proper noun CS187, what is
special about a quantificational expression like every course
is that it captures its surrounding context when used.
(31) Alice liked [every course].
Thus, loosely speaking, the meaning of the sentence (31) no
longer has the overall shape alice 8 liked ′ · · · once the occurrence of every course is considered, much as the meaning of
the program (28) no longer has the overall shape 10 × · · ·
once the shift expression is evaluated. Let us add shift and
reset to the target language of our translation from English.
We can then translate every course as
(32) Jevery courseK = ξs. ∀x. course(x) ⇒ s(x) : ThingBool
Bool .
The type notation αδγ here indicates an α with a control
effect; the CPS transform maps it to (α → γ) → δ. The
denotation of every course behaves locally as a Thing, but
requires the current context to have the answer type Bool
and maintains that answer type.
To see the new denotation (32) in action, let us derive
the sentence (31). The type of every course is ThingBool
Bool ,
similar to the type Thing of CS187, so the derivation of (31)
is analogous to (10–11).
(33)
Alice
liked
(34)
every course
JAliceK 8 JlikedK ′ Jevery courseK
= [alice 8 liked ′ ξs. ∀x. course(x) ⇒ s(x)]
⊲ [∀x. course(x) ⇒ (λv. [alice 8 liked ′ v])(x)] ⊲ · · ·
⊲ ∀x. course(x) ⇒ alice 8 liked ′ x : Bool
Like the straw man analysis in §3, the denotation in (32)
generalizes to determiners other than every: we can abstract
the noun course out of every course, and deal with some
student similarly.
(35)
(36)
JeveryK = λ′r. ξs. ∀x. r(x) ⇒ s(x),
JsomeK = λ′r. ξs. ∃x. r(x) ∧ s(x)
: (Thing → Bool) →
´ ThingBool
Bool
(We require here that the restrictor r have the type Thing →
Bool, not a type of the form Thing → Boolδγ , so r cannot
incur control effects when applied to x. Any control effect in
the restrictor, such as induced by the quantificational noun
phrase a company in the sentence Every representative of a
company left, must be contained within reset.)
More importantly, unlike the straw man analysis, the new
analysis works uniformly for quantificational expressions in
subject, object, and other positions, such as in (18a–c). Intuitively, this is because shift captures the context of an
expression no matter how deeply it is embedded in the sentence.2 By adding control operators for delimited continuations to our logical metalanguage, we arrive at an analysis
of quantification with greater empirical coverage.
Figure 2 shows a logical metalanguage that formalizes the
basic ideas presented above. It is in this language that
denotations on this page are written and reduced. Refining Danvy and Filinski’s original shift-reset language, we
distinguish between pure and impure expressions. An impure expression may incur control effects when evaluated,
whereas a pure expression only incurs control effects contained within reset (Danvy and Hatcliff 1992, 1994; Nielsen
2001; Thielecke 2003). This distinction is reflected in the
typing judgments: an impure judgment
(37)
Γ ⊢ E : αδγ
not only gives a type α for E itself but also specifies two
answer types γ and δ. By contrast, a pure judgment
(38)
Γ⊢E:α
only gives a type α for E itself. As can be seen in the Lift
rule, pure expressions are polymorphic in the answer type.
As mentioned in §2, the use of directionality in function
types to control word order is not new in linguistics, but the
use of delimited control operators to analyze quantification
is. It turns out that we can tie the potential presence of
control effects in function bodies to directionality. That is,
only directional functions—those whose types are decorated
with tick marks—are potentially impure; all non-directional
functions we need to deal with, including contexts captured
by shift, are pure. Another link between directionality and
control effects is that the →E
´ and →E
` rules for directional
function application are not merely mirror images of each
other: the answer types γ0 through γ3 are chained differently
through the premises. This is due to left-to-right evaluation.
Having made the distinction between pure and impure
expressions, we require in our Shift rule that the body of
a shift expression be pure. This change from Danvy and
Filinski’s original system simplifies the type system and the
CPS transform, but a shift expression ξf. E in their language
may need to be rewritten here to ξf. [E].
The CPS transform for the metalanguage follows from
the typing rules and is standard; it supplies a denotational
semantics. The operational semantics for the metalanguage
specifies a computation relation between complete terms; it
is also standard and shown in Figure 3.
The present analysis is almost, but not quite, the directstyle analogue of Barker’s (2002) CPS analysis. Put in
direct-style terms, Barker’s function bodies are always pure,
whereas function bodies here can harbor control effects. In
2
No matter how deep, that is, up to the closest dynamically
enclosing reset. Control delimiters correspond to islands in
natural language (Barker 2002).
Directions
∆ ::= ′ | 8
∆
α, β, γ, δ ::= Thing | Bool | α → β | α → βγδ
Γ ::= x1 : α1 , . . . , xn : αn
E, F ::= c | x | λx. E | λ∆x. E | F E | F ′ E | E 8 F | [E] | ξf. E
Types
Antecedents
Terms
Constants c : α
∀ : (Thing → Bool) → Bool
∃ : (Thing → Bool) → Bool
⇒ : Bool → Bool → Bool
∧ : Bool → Bool → Bool
···
liked : Thing →
´ (Thing →
` Boolδδ )γγ
student : Thing → Bool
Pure expressions Γ ⊢ E : α
c:α
Γ⊢c:α
Γ, x : α ⊢ E : β
Γ ⊢ λx. E : α → β
Const
Γ, x : α ⊢ x : α
Γ, x : α ⊢ E : βγδ
→I
∆
∆
Γ ⊢ λ x. E : α →
βγδ
Var
∆
→I
Γ ⊢ E : αβα
Γ ⊢ [E] : β
Reset
Γ⊢F :α→β
Γ⊢E:α
Γ ⊢ FE : β
→E
Impure expressions Γ ⊢ E : αδγ
Γ⊢E:α
Γ⊢E:
γ2 γ0
Γ ⊢ F : (α →
´ βγ3
)γ1
αγγ
Γ, f : α → γ ⊢ E : δ
Lift
Γ ⊢ E : αγ1
γ2
γ0
Γ ⊢ F ′ E : βγ3
Γ ⊢ ξf. E : αδγ
→E
´
Γ ⊢ E : αγ0
γ1
Shift
γ2 γ1
Γ ⊢ F : (α →
` βγ3
)γ2
γ0
Γ ⊢ E 8 F : βγ3
→E
`
Figure 2: A logical metalanguage with directionality and delimited control operators
Values
Unknowns
Contexts
V ::= U | λx. E | λ∆x. E
U ::= c | U V | U ′ V | V 8 U
Ch i ::= h i | (Ch i)E | Ch i ′ E | Ch i 8 F
| V (Ch i) | V ′ Ch i | V 8 Ch i
Metacontexts Dh i ::= h i | Ch[Dh i]i
Computations E ⊲ E ′
DhCh(λx. E)V ii
DhCh(λ′x. E) ′ V ii
DhChV 8 (λ8x. E)ii
DhCh[V ]ii
DhChξf. Eii
⊲
⊲
⊲
⊲
⊲
DhChE{x 7→ V }ii
DhChE{x 7→ V }ii
DhChE{x 7→ V }ii
DhChV ii
DhE{f 7→ λx. [Chxi]}i
Figure 3: Reductions for the logical metalanguage
other words, function bodies are allowed to shift, as in the
determiner denotations in (35) and (36). By contrast, Barker
uses choice functions to assign meanings to determiners.
5.
QUANTIFIER SCOPE AMBIGUITY
Of course, natural language phenomena are never as simple as a couple of programming language control operators.
Quantification is no exception, so to speak. For example,
the sentence Some student liked every course (18b) is ambiguous between the following two readings.
(39)
(40)
∃x. student(x) ∧ ∀y. course(y) ⇒ x 8 liked ′ y
∀y. course(y) ⇒ ∃x. student(x) ∧ x 8 liked ′ y
In the surface scope reading (39), some takes scope over every. In the inverse scope reading (40), every takes scope
over some. Given that evaluation takes place from left to
right, the shift for some student is evaluated before the shift
for every course. Our grammar thus predicts the surface
scope reading but not the inverse scope reading. This prediction can be seen in the first few reductions of the (unique)
derivation for (18b):
(41) (JsomeK ′ JstudentK) 8 JlikedK ′ JeveryK ′ JcourseK
′
= (λ r. ξs. ∃x. r(x) ∧ s(x)) ′ student
8 liked ′ (λ′r. ξs. ∀y. r(y) ⇒ s(y)) ′ course
⊲ ξs. ∃x. student(x) ∧ s(x)
8 liked ′ (λ′r. ξs. ∀y. r(y) ⇒ s(y)) ′ course
⊲ ∃x. student(x) ∧ λv. [v
8 liked ′ ((λ′r. ξs. ∀y. r(y) ⇒ s(y)) ′ course)] (x)
Regardless of what evaluation order we specify, as long as
our rules for semantic translation remain deterministic, they
will only generate one reading for the sentence. Hence our
theory fails to predict the ambiguity of the sentence (18b).
To better account for the data, we need to introduce some
sort of nondeterminism into our theory. There are two natural ways to proceed. First, we can allow arbitrary evaluation
order, not just left-to-right. This change would render our
term calculus nonconfluent, a result unwelcome for most programming language researchers but welcome for us in light of
the ambiguous natural language sentence (18b). This route
has been pursued with some success by Barker (2002) and
de Groote (2001). However, there are empirical reasons to
maintain left-to-right evaluation, one of which appears in §6.
A second way to introduce nondeterminism is to maintain left-to-right evaluation but generalize shift and reset to
a hierarchy of control operators (Barker 2000; Danvy and
Filinski 1990; Shan and Barker 2003), leaving it unspecified
at which level on the hierarchy each quantificational phrase
shifts. Following Danvy and Filinski, we extend our logical
metalanguage by superscripting every shift expression and
pair of reset brackets with a nonnegative integer to indicate
a level on the control hierarchy. Level 0 is the highest level
(not the lowest). When a shift expression at level n is evaluated, it captures the current context of computation up to
the closest dynamically enclosing reset at level n or higher
(smaller). For example, whereas the expression
0
(42)
3 × [10 × (ξ 5f. 1 + f (2))]5
evaluates to 63 as in (30), the expression
0
(43)
3 × [10 × (ξ 3f. 1 + f (2))]5
evaluates to 1 + 3 × 10 × 2, or 61. The superscripts can be
thought of “strength levels” for shifts and resets.
Danvy and Filinski (1990) give a denotational semantics
for multiple levels of delimited control using continuations
of higher-order type. We can take advantage of that work
in our quantificational denotations (35–36) by letting them
shift at any level. The ambiguity of (18b) is then predicted
as follows. Suppose that some student shifts at level m and
every course shifts at level n.
(44) Somem student liked everyn course.
If m ≤ n, the surface scope reading (39) results. If m > n,
the inverse scope reading (40) results. In general, a quantifier that shifts at a higher level always scopes over another
that shifts on a lower level, regardless of which one is evaluated first. This way, evaluation order does not determine
scoping possibilities among quantifiers in a sentence unless
two quantifiers happen to shift at the same level.
To summarize the discussion so far, whether we introduce
nondeterministic evaluation order or a hierarchy of delimited
control operators, we can account for the ambiguity of the
sentence (18b), as well as more complicated cases of quantification in English and Mandarin (Shan 2003). For example,
both the nondeterministic evaluation order approach and
the control hierarchy approach predict correctly that the
sentence below, with three quantifiers, is 5-way ambiguous.
(45) Every representative of a company saw most samples.
Despite the fact that there are three quantifiers in this sentence and 3! = 6, this sentence has only 5 readings. Because
a company occurs within the restrictor of every representative of a company, it is incoherent for every to scope over
most and most over a. The reason neither approach generates such a reading can be seen in the denotation of every
in (35): “∀x” is located immediately above “⇒” in the abstract syntax, with no intervening control delimiter, so no
control operator can insert any material (such as most-quantification over samples) in between.
There exist in the computational linguistics literature algorithms for computing the possible quantifier scopings of
a sentence like (45) (Hobbs and Shieber 1987; followed by
Lewin 1990; Moran 1988). Having related quantifier scoping
to control operators, we gain a denotational understanding
of these algorithms that accords with our theoretical intuitions and empirical observations.
An extended logical metalanguage with an infinite hierarchy of control operators is shown in Figure 4. This system is
more complex than the one in Figure 2 in two ways. First,
instead of making a binary distinction between pure and impure expressions, we use a number to measure “how pure”
each expression is. An expression is pure up to level n if it
only incurs control effects at levels above n when evaluated.
Pure expressions are the special case when n = 0. The purity level of an expression is reflected in its typing judgment:
a judgment
(46)
Γ ⊢ E : α!n
states that the expression E is pure up to level n. Here α!n
is a computation type with n levels: as defined in the figure,
it consists of 2n+1 − 1 value types that together specify how
a computation that is pure up to level n affects answer types
between levels 0 and n − 1. In the special case where n = 1,
the computation type α!1 is of the familiar form αδγ .
In the previous system in Figure 2, directional functions
are always impure (that is, pure up to level 1) while nondirectional functions are always pure (that is, pure up to
level 0). In the current system, both kinds of functions declare in their types up to what level their bodies are pure.
For example, the determiners every and some, now allowed
to shift at any level, both have not just the type
(47)
Bool δ!n
(Thing → Bool) →
´ ThingBool γ!n
δ!n ,
γ!n
but also the type
(48)
´ Thing
(Thing → Boolγ0!n
γ1!n ) →
γ0!n
Boolγ2!n
γ1!n
Boolγ2!n
(but see the second technical complication below). As the
argument type Thing → Boolγ0!n
γ1!n above shows, the first argument to these determiners, the restrictor, is non-directional
yet can be impure (that is, pure up to level n + 1).
To traverse the control hierarchy, we add a new Reset rule,
which makes an expression more pure, and a new Lift rule,
which makes
less pure. (Consecutive nested
n
an expression
resets like [E]n+1 can be abbreviated to [E]n without loss
of coherence.)
A second complication in this system, in contrast to Figure 2, is that we can no longer encode logical quantification
using a higher-order constant like ∀ : (Thing → Bool) → Bool,
because such a constant requires its argument—the logical
formula to be quantified—to be a pure function. This requirement is problematic because it is exactly the impurity
of quantified logical formulas that underlies this account of
quantifier scope ambiguity. On one hand, we want to quantify logical formulas that are impure; on the other hand, we
want to rule out expressions like
(49)
∀x. ξf. x,
where the logical variable x “leaks” illicitly into the surrounding context. This issue is precisely the problem of classifying open and closed terms in staged programming (see
Taha and Nielsen 2003 and references therein): the types
Thing and Bool really represent not individuals or truth values but staged programs that compute individuals and truth
values. For this paper, we adopt the simplistic solution of
adjoining to these types a set of free logical variables for
Directions
∆ ::= ′ | 8
∆
Value types
Computation types
Antecedents
Terms
α, β, γ, δ ::= Thing{~
p} | Bool{~
p} | α → β!n | α → β!n
α!0 ::= α,
α!(n + 1) ::= αδ!n
γ!n
Γ ::= x1 : α1 , . . . , xn : αn
E, F ::= c | x | λnx. E | λ∆nx. E | F E | F ′ E | E 8 F | [E]n | ξ nf. E
Constants c : α
p : Thing{p}
∀p : Bool{p, ~
q } → Bool{~
q}
⇒ : Bool{~
p} → Bool{~
q } → Bool{~
p∪~
q}
∃p : Bool{p, ~
q } → Bool{~
q}
∧ : Bool{~
p} → Bool{~
q } → Bool{~
p∪~
q}
γ!n
liked : Thing{~
p} →
´ (Thing{~
q} →
` Bool{~
p∪~
q }δ!m
δ!m )γ!n
student : Thing{~
p} → Bool{~
p}
Expressions Γ ⊢ E : α!n
c:α
Γ⊢c:α
Const
Γ, x : α ⊢ x : α
Var
Γ, x : α ⊢ E : β!n
Γ ⊢ λ(∆)x. E : α (∆)
→ β!n
Γ⊢E:α
Γ⊢E:
αγγ
0
Γ ⊢ [E] : β
(∆)
→I
Γ ⊢ E : αγ0!n
γ1!n
Lift
γ2!n γ0!n
Γ ⊢ F : (α →
´ βγ3!n
)γ1!n
Γ ⊢ E : αβα
γ0!n
βγ2!n
γ1!n
βγ2!n
Γ⊢E:α
Γ ⊢ E : αγ1!n
γ2!n
γ0!n
Γ ⊢ F ′ E : βγ3!n
Γ ⊢ F : (α →
´ β)
Γ⊢E:α
Γ⊢F ′E :β
→E
´
→E
´
β!(n+1)
Γ⊢E:α
Reset
Γ ⊢ F : α → β!n
n+1
Γ ⊢ [E]
Γ⊢E:α
: β!(n + 1)
Γ ⊢ ξ nf. E : αδ!n
γ!n
Shift
γ2!n γ1!n
Γ ⊢ F : (α →
` βγ3!n
)γ2!n
γ0!n
Γ ⊢ E 8 F : βγ3!n
Γ⊢E:α
Γ ⊢ F : (α →
` β)
Γ⊢E8F :β
Reset
→E
Γ, f : α → γ!n ⊢ E : δ!n
Γ ⊢ E : αγ0!n
γ1!n
γ!n
αγ!n
Γ ⊢ F E : β!n
Lift
···
→E
`
→E
`
Figure 4: Extending the logical metalanguage to a hierarchy of control operators
tracking purposes, denoted p, q, . . . . Unfortunately, we also
need to stipulate that these logical variables be freshly αrenamed (“created by gensym”) for each occurrence of a
quantifier in a sentence.
As before, the CPS transform and reductions for this metalanguage are standard; the latter appears in Figure 5.
The present analysis is almost, but not quite, the directstyle analogue of Shan and Barker’s (2003) CPS analysis,
even though both use a control hierarchy. Each level in Shan
and Barker’s hierarchy is intuitively a staged computation
produced at one level higher. More concretely, a computation type with n levels in that system has the shape
δ0
(αγ0
γ1 )δ1
(50)
rather than the shape
(51)
γ0δ0
δ1
αγ1δ2
.
δ3
The issue above of how to encode logical quantification over
impure formulas receives a more satisfactory treatment in
Shan and Barker’s system: no stipulation of α-renaming is
necessary, because there is no analogue of (49) to prohibit.
The relation between that system and staged programming
with effects has yet to be explored.
6.
POLARITY SENSITIVITY
Because the analysis so far focuses on the truth-conditional meaning of quantifiers, it equates the determiners a
and some—both are existential quantifiers with the type and
denotation in (36). Furthermore, sentences like Has anyone
arrived? suggest that the determiner any also means the
same thing as a and some. To the contrary, though, the determiners a, some, and any are not always interchangeable
in their existential usage. The sentences and readings in (52)
show that they take scope differently relative to negation (in
these cases the quantifier no).
(52) a. No student liked some course. (unambiguous ∃¬)
b. No student liked a course.
(ambiguous ¬∃, ∃¬)
c. No student liked any course.
(unambiguous ¬∃)
d. Some student liked no course. (unambiguous ∃¬)
e. A student liked no course.
(ambiguous ¬∃, ∃¬)
f. *Any student liked no course.
(unacceptable)
The determiner any is a negative polarity item: to a first
approximation, it can occur only in downward-entailing contexts, such as under the scope of a monotonically decreasing
quantifier (Ladusaw 1979). A quantifier q, of type (Thing →
Bool) → Bool, is monotonically decreasing just in case
(53)
∀s1 . ∀s2 . ∀x. s2 (x) ⇒ s1 (x) ⇒ q(s1 ) ⇒ q(s2 ).
The quantificational noun phrases no student and no course
are monotonically decreasing since, for instance, if no student liked any course in general, then no student liked any
computer science course in particular.
Whereas any is a negative polarity item, some is a positive
polarity item. Roughly speaking, some is allergic to down-
V ::= U | λx. E | λ∆x. E
Values
Unknowns
Contexts
U ::= c | U V | U ′ V | V 8 U
Ch i ::= h i | (Ch i)E | Ch i ′ E | Ch i 8 F | V (Ch i) | V ′ Ch i | V 8 Ch i
Dn h i ::= h i | Dn+1 hDn+2 h· · · hCh[Dn h i]n ii · · · ii
Metacontexts at level n
Computations E ⊲ E ′
D0 hD1 h· · · hCh(λx. E)V ii · · · ii ⊲ D0 hD1 h· · · hChE{x 7→ V }ii · · · ii
D0 hD1 h· · · hCh(λ′x. E) ′ V ii · · · ii ⊲ D0 hD1 h· · · hChE{x 7→ V }ii · · · ii
D0 hD1 h· · · hChV 8 (λ8x. E)ii · · · ii ⊲ D0 hD1 h· · · hChE{x 7→ V }ii · · · ii
D0 hD1 h· · · hCh[V ]ii · · · ii ⊲ D0 hD1 h· · · hChV ii · · · ii
D0 hD1 h· · · hDn hDn+1 h· · · hChξ nf. Eii · · · iii · · · ii ⊲ D0 hD1 h· · · hDn hE{f 7→ λx. [Dn+1 h· · · hChxii · · · i]n }ii · · · ii
Figure 5: Reductions for the extended logical metalanguage
ward-entailing contexts (especially those with an overtly
negative word like no). These generalizations regarding polarity items cover the data in (52a–e): in principle, goes the
theory, all these sentences are ambiguous between two scopings, but the polarity sensitivity of some and any rule out
one scoping each in (52a), (52c), (52d), and (52f). These
four sentences are thus predicted to be unambiguous, but it
remains unclear why (52f) is downright unacceptable.
In the type-theoretic tradition of linguistics, polarity sensitivity is typically implemented by splitting the answer type
Bool into several types, each a different functor applied to
Bool, that are related by subtyping (Bernardi 2002; Bernardi
and Moot 2001; Fry 1999). For instance, to differentiate the
determiners in (52) from each other in our formalism, we
can add the types BoolPos and BoolNeg alongside Bool, such
that both are supertypes of Bool (but not of each other).
(54)
(57)
α′ ≤ α
′ (∆)
(∆)
α≤α
′
α → β!n ≤ α → β !n
γ ′ !n ≤ γ!n
α ≤α
αδ!n
γ!n
Γ ⊢ E : α!n
≤
δ!n ≤ δ ′ !n
′
α′δγ ′!n
!n
α!n ≤ β!n
Γ ⊢ E : β!n
Sub
Γ ⊢ E : αβα
Γ ⊢ [E] : β
Reset
where β ≤ BoolPos
β!(n+1)
(59)
Γ⊢E:α
γ!n
αγ!n
Γ ⊢ [E] : β!(n + 1)
Reset
where β ≤ BoolPos
Finally, we refine the types of our determiners from (47) to
(60)
Bool δ!n
γ!n
JnoK : (Thing → Bool) →
´ ThingBoolNeg
δ!n ,
γ!n
(61)
(63)
JanyK : (Thing → Bool) →
´ ThingBoolNegδ!n .
γ!n
The chain of answer-type transitions from one quantificational expression to the next acts as a finite-state automaton, shown in Figure 6. The states of the automaton are
the three supertypes of Bool; the ǫ-transitions are the two
subtyping relations in (54); and the non-ǫ transitions are the
determiners in (60–63).
some
any
ǫ
BoolPos
Bool
BoolNeg
no
a
Figure 6: An automaton of answer-type transitions
This three-state machine enforces polarity constraints as
follows. Any valid derivation for a sentence assigns each
of its quantifiers to shift at a certain level in the control
hierarchy. For each level, the quantifiers at that level—in
the order in which they are evaluated—must form
β!n ≤ β ′ !n
We then add a side condition to Reset, requiring that the
produced answer type be Bool or BoolPos, not BoolNeg.
(58)
γ!n
BoolNeg δ!n
γ!n
ǫ
′
(56)
Boolδ!n
γ!n
JaK : (Thing → Bool) →
´ ThingBoolδ!n
,
Bool ≤ BoolNeg
Bool ≤ BoolPos
We also extend the subtyping relation between (value and
computation) types with the usual closure rules, and allow
implicit coercion from a subtype to a supertype.
(55)
(62)
BoolPos δ!n
γ!n
JsomeK : (Thing → Bool) →
´ ThingBoolPosδ!n ,
γ!n
• either a path from BoolPos to Bool in the state machine, in other words a string of determiners matching
the regular expression “some∗ (a | no any∗ )∗ ”;
• or a path from Bool to Bool in the state machine, in
other words a string of determiners matching the regular expression “(a | no any∗ )∗ ”.
Furthermore, the BoolPos-to-Bool levels in the hierarchy
must all be higher than the Bool-to-Bool levels. Every assignment of quantifiers to levels that satisfies these conditions gives a reading for the sentence, in which quantifiers
at higher levels scope wider, and, among quantifiers at the
same level, ones evaluated earlier scope wider.
Consider now the two alternative ways to characterize
scope ambiguity suggested in §5. The first approach is to
allow arbitrary evaluation order (and use a degenerate control hierarchy of one level only). If we take this route,
we can account for all of the acceptability and ambiguity
judgments in (52a–e), but we cannot distinguish the acceptable sentence (52c) from the unacceptable (52f). In other
words, it would be a mystery how the acceptability of a sentence hinges on the linear order in which the quantifiers no
and any appear. This mystery has been noted by Ladusaw (1979; §9.2) and Fry (1999; §8.2) as a defect in current
accounts of polarity sensitivity.
The second approach, using a control hierarchy with multiple levels, fares better by comparison.3 We can stick to
left-to-right evaluation, under which—as desired—an any
must be preceded by a no that scopes over it with no intervening a or some. Indeed, the variations in ambiguity and
acceptability among sentences in (52) are completely captured. For intuition, we can imagine that the hearer of a
sentence must first process the trigger for a downward-entailing context, like no, before it makes sense to process a
negative polarity item, like any.4 Intuition aside, the programming-language notion of evaluation order provides the
syntactic hacker of formal types with a new tool with which
to capture observed regularities in natural language.
7.
LINGUISTIC SIDE EFFECTS
This paper outlines how quantification and polarity sensitivity in natural language can be modeled using delimited
continuations. These two examples support my claim that
the formal theory and computational intuition we have for
continuations can help us construct, understand, and maintain linguistic theories. To be sure, this work is far from the
first time insights from programming languages are applied
to natural language:
• It has long been noted that the intensional logic in
which Montague grammar is couched can be understood computationally (Hobbs and Rosenschein 1978;
Hung and Zucker 1991).
• Dynamic semantics (Groenendijk and Stokhof 1991),
which relates anaphora and discourse in natural languages to nondeterminism and mutable state in programming languages (van Eijck 1998), has been applied to a variety of natural language phenomena, such
as verb-phrase ellipsis (van Eijck and Francez 1995;
Gardent 1991; Hardt 1999).
kind is jumps to labels; one of the second kind is mutable
state. By analogy, I use the term “linguistic side effects”
to refer to aspects of natural language where either it is
unclear what a denotational semantics should look like, or
the “obvious” denotational semantics (such as making each
clause denote whether it is true) turns out to break referential transparency. Besides quantification and polarity
sensitivity, some examples are:
(64) a.
b.
c.
d.
e.
Bob thinks Alice likes CS187.
A man walks. He whistles.
Which star did Alice see?
Alice only saw Venus.
The king of France whistles.
(Intensionality)
(Variable binding)
(Interrogatives)
(Focus)
(Presuppositions)
To study linguistic side effects, I propose to draw an analogy between them and computational side effects. Just
as computer scientists want to express all computational
side effects in a uniform and modular framework and study
how control interacts with mutable state (Felleisen and Hieb
1992), linguists want to investigate properties common to all
linguistic side effects and study how quantification interacts
with variable binding. Furthermore, just as computer scientists want to relate operational notions like evaluation order
and parameter passing to denotational models like continuations and monads, linguists want to relate the dynamics of
information in language processing to the static definition
of a language as a generative device. Whether this analogy
yields a linguistic theory that is empirically adequate is an
open scientific question that I find attractive to pursue.
8. ACKNOWLEDGMENTS
Thanks to Stuart Shieber, Chris Barker, Raffaella Bernardi,
Barbara Grosz, Pauline Jacobson, Aravind Joshi, William
Ladusaw, Fernando Pereira, Avi Pfeffer, Chris Potts, Norman Ramsey, Dylan Thurston, Yoad Winter, and anonymous referees. This work is supported by the United States
National Science Foundation Grants IRI-9712068 and BCS0236592.
9. REFERENCES
However, the link between natural language and continuations has only recently been made explicit, and this paper’s
use of control operators for a direct-style analysis is novel.
The analyses presented here are part of a larger project,
that of relating computational side effects to linguistic side
effects. The term “computational side effect” here covers
all programming language features where either it is unclear
what a denotational semantics should look like, or the “obvious” denotational semantics (such as making each arithmetic expression denote a number) turns out to break referential transparency. A computational side effect of the first
Barker, Chris. 2000. Notes on higher-order continuations.
Manuscript, University of California, San Diego.
3
Although this paper uses Danvy and Filinski’s control hierarchy, polarity sensitivity can be expressed equally well in
Shan and Barker’s system.
4
The syntactic distinction among the types Bool, BoolPos,
and BoolNeg may even be semantically interpretable via the
formulas-as-types correspondence, but the potential for such
a connection has only been briefly explored (Bernardi and
Nilsen 2001) and we do not examine it here. In this connection, Krifka (1995) and others have proposed on pragmatic grounds that determiners like any are negative polarity items because they indicate extreme points on a scale.
Bernardi, Raffaella, and Richard Moot. 2001. Generalized quantifiers in declarative and interrogative sentences.
Journal of Language and Computation 1(3):1–19.
———. 2002. Continuations and the nature of quantification. Natural Language Semantics 10(3):211–242.
Barwise, Jon, and Robin Cooper. 1981. Generalized quantifiers and natural language. Linguistics and Philosophy 4:
159–219.
Bernardi, Raffaella. 2002. Reasoning with polarity in categorial type logic. Ph.D. thesis, Utrecht Institute of Linguistics (OTS), Utrecht University.
Bernardi, Raffaella, and Øystein Nilsen. 2001. Polarity items
in type logical grammar: Connection with DMG. Slides
for talk at Learning Logic and Grammar workshop, Amsterdam.
Carpenter, Bob. 1997. Type-logical semantics. Cambridge:
MIT Press.
Danvy, Olivier, and Andrzej Filinski. 1989. A functional
abstraction of typed contexts. Tech. Rep. 89/12, DIKU,
University of Copenhagen, Denmark. http://www.daimi.
au.dk/∼danvy/Papers/fatc.ps.gz.
———. 1990. Abstracting control. In Proceedings of the
1990 ACM conference on Lisp and functional programming, 151–160. New York: ACM Press.
Danvy, Olivier, and John Hatcliff. 1992. CPS-transformation after strictness analysis. ACM Letters on Programming Languages and Systems 1(3):195–212.
———. 1994. On the transformation between direct and
continuation semantics. In Mathematical foundations
of programming semantics: 9th international conference
(1993), ed. Stephen D. Brookes, Michael G. Main, Austin
Melton, Michael W. Mislove, and David A. Schmidt,
627–648. Lecture Notes in Computer Science 802, Berlin:
Springer-Verlag.
van Eijck, Jan. 1998. Programming with dynamic predicate logic. Report INS-R9810, Centrum voor Wiskunde
en Informatica, Amsterdam. Also as Research Report CT1998-06, Institute for Logic, Language and Computation,
Universiteit van Amsterdam.
van Eijck, Jan, and Nissim Francez. 1995. Verb-phrase
ellipsis in dynamic semantics. In Applied logic: How,
what, and why: Logical approaches to natural language,
ed. László Pólos and Michael Masuch. Dordrecht: Kluwer.
Felleisen, Matthias. 1988. The theory and practice of firstclass prompts. In POPL ’88: Conference record of the annual ACM symposium on principles of programming languages, 180–190. New York: ACM Press.
Felleisen, Matthias, and Robert Hieb. 1992. The revised
report on the syntactic theories of sequential control and
state. Theoretical Computer Science 103(2):235–271.
Fry, John. 1999. Proof nets and negative polarity licensing.
In Semantics and syntax in lexical functional grammar:
The resource logic approach, ed. Mary Dalrymple, chap. 3,
91–116. Cambridge: MIT Press.
Gardent, Claire. 1991. Dynamic semantics and VP-ellipsis.
In Logics in AI: European workshop JELIA ’90, ed. Jan
van Eijck, 251–266. Lecture Notes in Artificial Intelligence
478, Berlin: Springer-Verlag.
Groenendijk, Jeroen, and Martin Stokhof. 1991. Dynamic
predicate logic. Linguistics and Philosophy 14(1):39–100.
de Groote, Philippe. 2001. Type raising, continuations, and
classical logic. In Proceedings of the 13th Amsterdam Colloquium, ed. Robert van Rooy and Martin Stokhof, 97–
101. Institute for Logic, Language and Computation, Universiteit van Amsterdam.
Hardt, Daniel. 1999. Dynamic interpretation of verb phrase
ellipsis. Linguistics and Philosophy 22(2):185–219.
Hendriks, Herman. 1993. Studied flexibility: Categories and
types in syntax and semantics. Ph.D. thesis, Institute
for Logic, Language and Computation, Universiteit van
Amsterdam.
Hobbs, Jerry R., and Stanley J. Rosenschein. 1978. Making computational sense of Montague’s intensional logic.
Artificial Intelligence 9:287–306.
Hobbs, Jerry R., and Stuart M. Shieber. 1987. An algorithm for generating quantifier scopings. Computational
Linguistics 13(1–2):47–63.
Hung, Hing-Kai, and Jeffery I. Zucker. 1991. Semantics of
pointers, referencing and dereferencing with intensional
logic. In LICS ’91: Proceedings of the 6th symposium
on logic in computer science, 127–136. Washington, DC:
IEEE Computer Society Press.
Kelsey, Richard, William Clinger, Jonathan Rees, et al.
1998.
Revised5 report on the algorithmic language
Scheme. Higher-Order and Symbolic Computation 11(1):
7–105. Also as ACM SIGPLAN Notices 33(9):26–76.
Krifka, Manfred. 1995. The semantics and pragmatics of
polarity items. Linguistic Analysis 25:209–257.
Ladusaw, William A. 1979. Polarity sensitivity as inherent scope relations. Ph.D. thesis, Department of Linguistics, University of Massachusetts. Reprinted by New York:
Garland, 1980.
Lewin, Ian. 1990. A quantifier scoping algorithm without
a free variable constraint. In COLING ’90: Proceedings
of the 13th international conference on computational linguistics, vol. 3, 190–194.
May, Robert. 1985. Logical form: Its structure and derivation. Cambridge: MIT Press.
Montague, Richard. 1974. The proper treatment of quantification in ordinary English. In Formal philosophy: Selected
papers of Richard Montague, ed. Richmond Thomason,
247–270. New Haven: Yale University Press.
Moran, Douglas B. 1988. Quantifier scoping in the SRI core
language engine. In Proceedings of the 26th annual meeting of the Association for Computational Linguistics, 33–
40. Somerset, NJ: Association for Computational Linguistics.
Nielsen, Lasse R. 2001. A selective CPS transformation.
In Proceedings of MFPS 2001: 17th conference on the
mathematical foundations of programming semantics, ed.
Stephen Brooks and Michael Mislove. Electronic Notes in
Theoretical Computer Science 45, Amsterdam: Elsevier
Science.
Shan, Chung-chieh. 2003. Quantifier strengths predict
scopal possibilities of Mandarin Chinese wh-indefinites.
Draft manuscript, Harvard University; http://www.eecs.
harvard.edu/∼ ccshan/mandarin/.
Shan, Chung-chieh, and Chris Barker. 2003. Explaining crossover and superiority as left-to-right evaluation.
Draft manuscript, Harvard University and University of
California, San Diego; http://semanticsarchive.net/
Archive/TBjZDQ3Z/.
Taha, Walid, and Michael Florentin Nielsen. 2003. Environment classifiers. In POPL ’03: Conference record of
the annual ACM symposium on principles of programming
languages, 26–37. New York: ACM Press.
Thielecke, Hayo. 2003. From control effects to typed continuation passing. In POPL ’03: Conference record of
the annual ACM symposium on principles of programming
languages, 139–149. New York: ACM Press.
| 6cs.PL
|
1
PsyPhy: A Psychophysics Driven Evaluation
Framework for Visual Recognition
(a) Traditional
Natural
Evaluation
Brandon RichardWebster, Student Member, IEEE,
Samuel E. Anthony, Student Member, IEEE,
and Walter J. Scheirer, Senior Member, IEEE
(b) Visual Psychophysics
arXiv:1611.06448v3 [cs.CV] 14 Sep 2017
Natural
Abstract—By providing substantial amounts of data and standardized evaluation protocols, datasets in computer vision have
helped fuel advances across all areas of visual recognition. But
even in light of breakthrough results on recent benchmarks, it
is still fair to ask if our recognition algorithms are doing as well
as we think they are. The vision sciences at large make use of a
very different evaluation regime known as Visual Psychophysics
to study visual perception. Psychophysics is the quantitative
examination of the relationships between controlled stimuli and
the behavioral responses they elicit in experimental test subjects.
Instead of using summary statistics to gauge performance,
psychophysics directs us to construct item-response curves made
up of individual stimulus responses to find perceptual thresholds,
thus allowing one to identify the exact point at which a subject
can no longer reliably recognize the stimulus class. In this article,
we introduce a comprehensive evaluation framework for visual
recognition models that is underpinned by this methodology. Over
millions of procedurally rendered 3D scenes and 2D images, we
compare the performance of well-known convolutional neural
networks. Our results bring into question recent claims of humanlike performance, and provide a path forward for correcting
newly surfaced algorithmic deficiencies.
Index Terms—Object Recognition, Visual Psychophysics, Neuroscience, Psychology, Evaluation, Deep Learning.
I. I NTRODUCTION
We often attribute “understanding” and other cognitive
predicates by metaphor and analogy to cars, adding
machines, and other artifacts, but nothing is proved by
such attributions.
John Searle
Imagine the following scenario: a marvelous black box
algorithm has appeared that purportedly solves visual object
recognition with human-like ability. As a good scientist, how
might you go about falsifying this claim? By all accounts,
the algorithm achieves superior performance on established
benchmarks in computer vision, but its internal workings
are opaque to the external observer. Such a situation is not
far fetched — it should be familiar to any of us studying
machine learning for visual recognition. But what many of us
in computer vision might not realize is that this setup happens
to be the classic Chinese Room [3] problem proposed by the
philosopher John Searle.
In Searle’s thought experiment, a person who does not speak
Chinese is alone in a locked room and following instructions
from a computer program to generate Chinese characters to
respond to Chinese messages that are slipped under the door.
To the message passer outside of the room, the person inside
B. RichardWebster and W. Scheirer are with the Department of Computer
Science and Engineering, University of Notre Dame, Notre Dame, IN, 46556.
Corresponding Author’s E-mail: [email protected].
S. Anthony is with the Department of Psychology, Harvard University, and
Perceptive Automata, Inc.
p
e
r
t
u
r
b
a
t
i
o
n
Rendered
Gaussian
blur
Linear
Occlusion
Rotation
Fig. 1. In this article, the concept of applying psychophysics [1], [2] on a
recognition model is introduced. In this figure, A and B are two models being
compared. (Top) In traditional dataset-based evaluation, summary statistics
are generated over large sets of data, with little consideration given to
specific conditions that lead to incorrect recognition instances. (Bottom) Psychophysics, a set of experimental concepts and procedures from psychology
and neuroscience, helps us plot the exact relationships between perturbed test
images and resulting model behavior to determine the precise conditions under
which models fail. Instead of comparing summary statistics, we compare itemresponse curves representing performance (y-axis) versus the dimension of the
image being manipulated (x-axis).
understands Chinese. However, this is not the case. The person
inside the room is simply following instructions to complete
the task — there is no real replication of the competency of
knowing the Chinese language. Linking this back to computer
vision, the summary statistics of performance from our algorithms look good on benchmark tests — enough so that we
believe them to be close to human performance in some cases.
But are these algorithms really solving the general problem
of visual object recognition, or are they simply leveraging
“instructions” provided in the form of labeled training data
to solve the dataset?
Datasets in computer vision are intended to be controlled
testbeds for algorithms, where the task and difficulty can be
modulated to facilitate measurable progress in research. A
dataset could be made up of images specifically acquired
for experimentation, or publicly available images crawled
from the web. Under this regime, strong advancements have
been demonstrated for a number of problems, most notably
for object recognition [4]. Deep learning is now a mainstay
in computer vision thanks in part to the 2012 ImageNet
Challenge [5], where AlexNet [6] reduced top-5 object classification error to 16.4% from the previously best reported
result of 25.8%. When algorithms are evaluated on a common
footing, it is possible to track meaningful improvements in
artificial intelligence like this one. However, increases in error
when different datasets are used for training and testing [7],
2
[8] make us wonder if this is the only way to do it.
When it comes to natural intelligence, neuroscientists and
psychologists do not evaluate animals or people in the same
way that computer vision scientists evaluate algorithms — and
for a very good reason. With a collection of images crawled
from the web, there is no straightforward way to determine the
exact condition(s) that caused a subject to fail at recognizing
a stimulus presented during an experiment. A natural image is
the product of the physics at the instant the sensor acquired the
scene; its latent parameters are largely unknown. Instead, for
behavioral experiments meant to discover perceptual thresholds (i.e., the average point at which subjects start to fail), the
vision sciences outside of computer vision use the concepts
and procedures from the discipline of visual psychophysics.
Psychophysics is the quantitative study of the relationships
between controlled stimuli and the behavioral responses they
elicit in a subject [1], [2]. It is a way to probe perceptual
processes through the presentation of incremental and, in
many cases, extremely fine-grained perturbations of visual
stimuli. The properties of each stimulus are varied along one
or more physical dimensions, thus controlling the difficulty
of the task. The result (Fig. 1) is an item-response curve [9],
where performance (e.g., accuracy) on the y-axis is plotted
against the dimension being manipulated (e.g., Gaussian blur)
on the x-axis. Each point on the curve reflects an individual
stimulus, letting us map performance back to causal conditions
in a precise manner. Psychophysics is an indispensable tool
to vision science, and has been deployed to discover the
minimum threshold for stimulation of a retinal photoreceptor
(a single photon) [10], confirm Helmholtz’s assertions on
color absorption in the retina [11], and establish criteria to
diagnose prosopagnosia [12] (the inability to recognize a face).
As in these discoveries from biological vision, we submit
that psychophysics holds much promise for discovering new
aspects of the inner workings of machine learning models.
In this article, we introduce a comprehensive evaluation
framework for visual recognition that is underpinned by the
principles of psychophysics. In this regime, a stimulus can be
an object drawn from purely rendered data or natural scene
data, and a varying physical parameter can control the amount
of transformation in the subsequent set of manipulated images
derived from the original stimulus. A key difference from
traditional benchmarks in computer vision is that instead of
looking at summary statistics (e.g., average accuracy, AUC,
precision, recall) to compare algorithm performance, we compare the resulting item-response curves. For complete control
of the underlying parameter space, we find that procedural
graphics [13]–[16] are a useful way to generate stimuli that
can be manipulated in any way we desire. Because we have
the procedure that rendered each scene, we can find out where
a model is failing at the parametric level. As we will see, by
using this framework to explore artificial vision systems like
psychologists, many interesting new findings can be surfaced
about the strengths and limitations of computer vision models.
To summarize, our main contributions are as follows:
• A general evaluation framework is developed for performing visual psychophysics on computer vision models.
The framework has a strong grounding in well-established
•
•
•
work in psychology and neuroscience for behavioral
experimentation.
An investigation of procedural graphics for large-scale
psychophysics experiments applied to models.
A parallelized implementation of the psychophysics
framework that is deployable as a Python package.
A case study consisting of a battery of experiments
incorporating millions of procedurally rendered images
and 2D images that were perturbed, performed over a
set of well-known Convolutional Neural Network (CNN)
models [6], [17]–[19].
II. R ELATED W ORK
Methods of Evaluation from the Vision Sciences. With respect to work in computer vision directly using psychophysics,
most is related to establishing human baselines for comparison to algorithmic approaches. Riesenhuber and Poggio [20]
described a series of psychophysical comparisons between
humans and the HMAX [21] model of visual cortex using
a limited set of stimuli rendered by computer graphics. Similarly, Eberhardt et al. [22] designed an experiment to measure
human accuracy and reaction time during visual categorization
tasks with natural images, which were then compared to
different layers of CNN models [6], [19]. Geirhos et al.
undertook a similar study for image degradations [23]. With
respect to low-level features, Gerhard et al. [24] introduced
a new psychophysical paradigm for comparing human and
model sensitivity to local image regularities.
Psychophysics can also be used for more than just performance evaluation. Scheirer et al. [25] introduced the notion of
“perceptual annotation” for machine learning, whereby psychophysical measurements are used as weights in a loss function to give a training regime some a priori notion of sample
difficulty. Using accuracy and reaction time measured via the
online psychophysics testing platform TestMyBrain.org [26],
perceptual annotation was shown to enhance face detection
performance. Along these lines, Vondrick et al. [27] devised
a method inspired by psychophysics to estimate useful biases
for recognition in computer vision feature spaces.
Outside of work specifically invoking psychophysics, one
can find other related methods from psychology and neuroscience for behavioral-style model testing. 2D natural images
are the most common type of data in computer vision, and
form a good basis for algorithmic evaluation in this mode.
O’Toole et al. [28], [29] and Philips et al. [30] have designed
controlled datasets of natural images to compare human face
recognition performance against algorithms. With the focus
on algorithmic consistency with human behavior, there is no
explicit model vs. model comparison in these methods.
More control in experimentation can be achieved through
the use of rendered 3D scenes. Cadiue et al. [31], Yamins et
al. [32] and Hong et al. [33] all make use of rendered images
with parametrized variation to compare the representations
of models with those found in the primate brain. Pramod
and Arun [34] describe a set of perceived dissimilarity measurements from humans that is used to study the systematic
differences between human perception and a large number
3
of handcrafted and learned feature representations. Because
of a need for very fine-grained control of object parts and
other latent parameters of scenes, procedural graphics were
introduced by Tenenbaum et al. [13] for the study of oneshot learning using probabilistic generative models. The use
of procedural graphics for generative models was further
developed by Yildirim et al. [14], Kulkarni et al. [15], and
Wu et al. [16]. These studies do not vary the conditions of the
stimuli using the procedures of psychophysics, nor do they use
large-scale renderings on the order of millions of scenes.
Other Manipulations of Stimuli in Visual Recognition
Evaluations. Work coming directly out of computer vision
also addresses stimulus generation for the purpose of isolating
model weaknesses. Hoiem et al. [35] suggest systematically
varying occlusion, size, aspect ratio, visibility of parts, viewpoint, localization error, and background to identify errors in
object detectors. Wilber et al. [36] systematically apply noise,
blur, occlusion, compression, textures and warping effects over
2D scenes to assess face detection performance. Finally, a
whole host of approaches can be found to manipulate the
inputs to CNNs in order to highlight unexpected classification
errors. These include the noise patterns introduced by Szegedy
et al. [37] that are imperceptible to humans, and the fooling
images produced via evolutionary algorithms that were explored by Nguyen et al. [38] and Bendale and Boult [39]. The
level of control in the evaluation procedures varies between
these approaches, but a common starting point based on
model preference for each class is missing (i.e., which object
configuration produces the highest score?). We suggest in this
article that the use of computer graphics helps us address this.
III. M ETHOD : T HE P SY P HY F RAMEWORK
Our procedure for performing psychophysics on a model
largely follows established procedures found in psychology,
with a few key adaptations to accommodate artificial perception. For the purposes of this article, our focus is on
two performance-based forced-choice tasks that yield an interpretable item-response curve. For descriptions of other
procedures in psychophysics, see [1], [2]. First, let us consider
the two-alternative forced choice (2AFC) match-to-sample task
that is common in psychological testing.
In the 2AFC match-to-sample procedure, an observer is
shown a “sample” stimulus, followed by two “alternate”
stimuli where one is a positive (i.e., matching) stimulus and the
other is a negative (i.e., non-matching) stimulus. The observer
is then asked to choose from the alternate stimuli the stimulus
that best matches the sample — the match criterion may or
may not be provided to the observer. The observer repeats the
task at different perturbed stimulus levels in either an adaptive
pattern, which is like gradient descent for humans, or via the
method of constants, which is a predetermined set of perturbed
stimulus levels. Regardless of method, each task has two
presented alternate stimuli (N = 2) and thus two-alternative
forced-choices (M = 2). Analysis of the experiment would
utilize the mean or median accuracy humans achieved at each
stimulus level and mean or median human response time, if
recorded. Models can be tested in precisely the same way
Algorithm 1 Dfm (i, c), the top-1 binary decision of the Softmax layer of a CNN. Used for both preferred view calculation
and MAFC.
Input: f , a single pre-trained network model
Input: i, an input image
Input: c, the expected class
1: V = f (i)
. the Softmax vector
2: c∗ = argmaxj∈[0,|V |) Vj
. find class label
3: ς = Vc∗
4: if c 6= c∗ then
. incorrect class, negate response
5:
ς = −1 ∗ ς
6: end if
7: return ς, the decision score
when the input images are arranged as described above and
accuracy is used as the performance measure.
Second, we can consider a mapping of a more difficult
classification procedure in machine learning to a more general
version of the 2AFC match-to-sample procedure. We call
this mapped classification MAFC match-to-sample. In MAFC,
the probe image in classification is equivalent to the sample
stimulus. In classification, we rarely have only two classes
for a model to choose from. Thus the value of M becomes
the number of labeled training classes (e.g., ImageNet 2012
has 1K learned classes, making M = 1K). Likewise, N —
the number of presented alternate stimuli — changes to the
number of images used in training, as this is the set of images
the model is implicitly matching to (e.g., for ImageNet 2012,
N = ∼1.2M training images).
When testing a model with any psychophysics procedure,
we need a special process for the selection of the stimuli’s
default state, that is, where there is no perturbation. Blanz et
al. [40] show that humans have a natural inclination towards
recognizing a certain view of an object, called a canonical
view. We assume in human trials that an object configuration
close to a canonical view will be chosen, maximizing the
probability that all observers will have no problems performing
at least some part of a match-to-sample task. However, this is
not as simple for any task involving a model because we do not
necessarily know if it follows a similar canonical view. But we
can say that a model’s preferred view is a view that produces
the strongest positive response, as determined by a decision
score. Note that there can be more than one preferred view
(hence our use of the term preferred), because ties are often
observed for the strongest response, especially in quantized
decision score spaces. Choosing a preferred view is crucial to
guaranteeing that when the stimulus is perturbed, the model’s
response will already be at its maximum for that class. Any
perturbation will cause a decline (or possibly no change) in
the strength of the model’s response, not an increase.
PsyPhy Framework for 2AFC and MAFC. Inspired by
software frameworks for subject testing like PsychoPy [41], we
have implemented the 2AFC and MAFC procedures described
above using a Python framework for model testing called
PsyPhy. Here we describe the details of each component of the
framework. The basic steps of (1) stimuli selection, (2) preferred view selection, (3) perturbation, and (4) item-response
4
Algorithm 2 φ2T (Df2 , V, s): an item-response point generation
function supporting 2AFC tasks for any image transformation
function T (s, v)
Input: Df2 , decision function for 2AFC
Input: V , a vector of preferred views for a set of classes
Input: s, the stimulus level
1: h(v) :=random v 0 |v 0 ∈ [V \ {v}]
. pick negative
P
2
2: β =
max(0,
dD
(T
(s,
v),
v,
h(v))e)
f
v∈V
β
3: a = |V |
4: return {s, a}, an x, y coordinate pair (stimulus level,
accuracy over trials) for one item-response point
curve generation apply to any psychophysics procedure, and
the specific 2AFC and MAFC procedures may be viewed
as pluggable modules within the framework. PsyPhy is very
flexible with respect to tasks it can support.
The first step is to select the initial set of stimuli for each
class. For 2D natural images, this is any set of chosen images
I2D for a class c. For a rendered scene, a set of image
specifications I3D is provided to a rendering function R(c, v)
(implemented in this work using Mitsuba [42]) to render a
single object centered in an image. The view v ∈ I3D is
the parameter set {x, y, z, ψ}, where the coordinates x, y,
and z are real numbers in the range (−180.0, 180.0] and ψ,
representing scale, is a real number in the range (0.0, 25.0].
The second step is to find an individual model’s preferred
view for each class. For natural 2D images, the preferred view
function in Eq. 1 is used. The second preferred view function,
Eq. 2, uses R to create rendered images for classification. In
Eq. 2, the search space is almost infinite, thus it does not find
the absolute global maximum, but rather an approximation.
P2D (I2D , c) := argmax Dfm (i, c)
(1)
P3D (I3D , c) := argmax Dfm (R(c, v), c)
(2)
i∈I2D
v∈I3D
A decision function for classification Dfm (i, c) (Alg. 1) normalizes the score output of a model f to a value in the range
[−1.0, 1.0], which gives both a decision and a confidence
associated with that decision. A value in the range [−1.0, 0]
is an incorrect decision and (0, 1.0] is a correct decision. The
parameter i is the input stimulus and c is the expected class.
A natural 2D preferred view (Eq. 1) is a single selected
image i ∈ I2D , where Dfm has the strongest positive response. A 3D preferred view (Eq. 2) is a single selected
set v = {xv , yv , zv , ψv } ∈ I3D , where Dfm has the strongest
positive response. The major difference between Eq. 1 and
Eq. 2 is the use of R in Eq. 2 to render the image prior to
measuring the response from Dfm . Invoking Eq. 1 or Eq. 2 for
each class will create a vector of preferred views V .
After preferred views have been selected for all classes,
whether natural or rendered, the next step is to apply perturbations to them. In this procedure, a set of preferred views
is perturbed at a specific stimulus level (i.e., the amount
of perturbation) using a function T (s, v), where T could
be any image transformation function (e.g., Gaussian blur,
rotation). The parameter v is one preferred view — either in
m
Algorithm 3 φm
T (Df , V, s): an item-response point generation
function supporting MAFC tasks for any image transformation
function T (s, v)
Input: Dfm , decision function for MAFC
Input: V , a vector of preferred views for a set of classes
Input: s,Pthe stimulus level
m
1: β =
v∈V max(0, dDf (T (s, v), c(v))e)
β
2: a = |V |
3: return {s, a}, an x, y coordinate pair (stimulus level,
accuracy over trials) for one item-response point
2D image format or {x, y, z, ψ} for rendered stimuli — and
s is the stimulus level. The function φT (D, V, s) perturbs the
set of preferred views given in V and then makes a decision
on each image using a decision function D. The specific
implementation φ2T (Df2 , V, s) for 2AFC is described in Alg. 2,
m
and φm
T (Df , V, s) for MAFC is described in Alg. 3. Procedure
specific decision functions are required, with Df2 (Alg. 4) used
for 2AFC and Dfm (Alg. 1) used for MAFC. Each individual
image evaluation is a trial. The value returned by φT represents
one point on an item-response curve, which is the computed
accuracy over all trials (one trial per class).
An item-response curve is the set of x, y coordinates that
represent the model behavior for a set of stimuli. Each
x, y value represents a perturbation level and accuracy of
the model’s performance. Note that traditional psychophysics
experiments with live test subjects often apply a psychometric
function to interpolate between the points to generate the
curve. To approximate a psychometric function for better interpretability, we applied rectangular smoothing (i.e., unweighted
sliding-average smoothing) with a window size of 15 while
padding the curve with repeated edge values.
The final step generates item-response curves using the
function CT (φ, D, V, n, bl , bu ). The procedure is simple, and
only requires a repeated execution of φT for each stimulus
level. Its steps are shown in Alg. 5. The procedure will create
a set of stimulus levels starting with a lower bound, bl , and
ending with an upper bound bu . bl is the closest stimulus level
to the preferred view and bu is the stimulus level that is farthest
away. The parameter n is the number of stimulus levels to use.
Typically in visual pyschophysics, log-spaced stepping is used
for finer-grained evaluation near the canonical view; the same
strategy is used for preferred view.
IV. E XPERIMENTS
The first goal of our experiments was to demonstrate PsyPhy
as a large-scale psychophysics evaluation framework. To do
this, we processed millions of procedurally rendered 3D scenes
and 2D images that were perturbed. The second goal was
to demonstrate the utility of procedural graphics for largescale psychophysics experiments. Thus we broke our data
up into two sets: natural scenes and rendered scenes. Our
final goal was to evaluate the strengths and weaknesses of
well-known CNN models. To do this, we looked at model
behavior for 2AFC and MAFC tasks, the behavior of dropout
at test time [43] under perturbations, and comparisons to
5
Algorithm 4 Df2 (i, p, q), best match decision of the final
feature layer of a CNN. Used for 2AFC.
Input: f , a single pre-trained network model
Input: i, an input image
Input: p, the expected positive image
Input: q, the expected negative image
1: Wi = f (i)
. gather activations from final feature layer
2: Wp = f (p)
3: Wq = f (q)
4: ςp = r(Wi , Wp )
. Pearson’s Correlation
5: ςq = r(Wi , Wq )
6: if ςp > ςq then . if incorrect selection, negate response
7:
ς = ςp
8: else
9:
ς = −1 ∗ ςq
10: end if
11: return ς, the decision score
Algorithm 5 CT (φ, D, V, n, bl , bu ): an item-response curve
generation function for any type of decision function
Input: φ, an item-response point generator
Input: f , an input model
Input: V , a vector of preferred views
Input: n, the number of stimulus levels
Input: bl and bu , the lower and upper bound values of the
stimulus levels
1: Let S be n log-spaced stimulus levels from bl to bu
S
2: I =
{φT (D, V, s)}
s∈S
3:
return I, the item-response curve
human behavior. In all of our experiments, we chose to
use five convolutional neural network models that were pretrained on ImageNet 2012 [5]: AlexNet [6], CaffeNet [17],
GoogleNet [18], VGG-16, and VGG-19 [19]. The complete
set of plots and more details on the methods can be found in
the supplemental material for this article1 .
Data Generation. For the natural scene experiments, we
perturbed images from the ImageNet 2012 [5] training dataset,
which consists of ∼1.2M million images and 1K classes.
Using the training set instead of the testing set gives each
model an intentional bias towards “expert” performance. The
following transformations were applied: Gaussian blur, linear occlusion, salt & pepper noise, brightness, contrast, and
sharpness. For each condition, we created 200 perturbed
images starting with the preferred view and log-spaced stepped
towards increasing difficulty. The result was 201 images per
class per network, or 201K images per network, or ∼1M
images per condition. In total, ∼9M images were evaluated.
For the experiments with rendered images, we selected 40
3D objects from the Blend Swap [44] library that corresponded
to classes in ImageNet (see supp. material for a list of the
classes). For each of the 3D objects, we randomly rendered
100K uniformly distributed x, y, z rotations and scales, result1 Supplemental material is accessible at
http://bjrichardwebster.com/papers/psyphy/supp
ing in 4M images. After each preferred view was selected
from that set, the following transformations were applied by
our graphics engine: rotations in the x, y, z dimensions, and
scale. All were applied in the positive and negative direction.
In addition, all of the transformations from the 2D natural
image experiment were repeated using the rendered preferred
views. For each of the 3D transformations, we rendered 200
images starting with the preferred view and log-spaced stepped
towards increasing difficulty. The result was 201 images per
class per network, or ∼8K images per network, or ∼40K
images per transformation. The additional 2D transformations
resulted in a total of ∼362K images, which brought the
rendered image total to ∼683K evaluated images.
2AFC Experiments. The motivation for the 2AFC experiments is twofold: (1) to test decision making at a fundamental
level via activation matching (i.e., to not just look at class
labels), and (2) to test a precise implementation of a wellknown match-to-sample task. Given that this setting is just two
instances of pair-matching, we had initially expected models
to perform relatively well under a variety of perturbations.
The experiments included the use of both natural scenes and
rendered scenes as stimuli.
Model behavior was only very stable for the 3D rotation
(Fig. 2 center and Supp. Fig. 1), and contrast and sharpness (Supp. Figs. 3 & 4) transformations. The rest of the
transformations induced more erratic behavior, with accuracy
declining below 80% (e.g., Fig. 2 left and right). Gaussian blur
was very detrimental to model accuracy, even in this limited
matching setting. This tells us something about the receptive
fields of the convolutional layers of the networks — they are
not large enough to tolerate even modest levels of blur. Also
interesting were results for the 3D-specific scale perturbation
that let us isolate specific failure modes. For example, when
scale decreases to 20% of the original size, the object structure
is still clearly visible, but accuracy drops to ∼60% or lower
for all networks (Supp. Fig. 1). This is an observation that
could not have been made by looking at summary statistics.
What about the differences in behavior across networks?
Are they significant? When examining the item-response
curves with the the 95% confidence interval plotted (Supp.
Figs. 2, 5, 6 & 8) all network behavior consistently demonstrates the same trends for each transformation type across
perturbations. While it is commonly believed that architectural
innovation is advancing deep learning, this finding indicates
that model behavior is more influenced by the training data,
which was the same for all models in these experiments. For
the VGG networks, this suggests that additional layers — beyond a certain point — do not imply better performance under
degrading conditions. Likewise, switching the order of the
pooling and normalization layers in CaffeNet and AlexNet [45]
does not imply better performance under degrading conditions.
MAFC Experiments. The motivation for the MAFC experiments is to evaluate a task that is more closely aligned to
the multi-class classification task the models were originally
trained for. Given that there are 1, 000 choices in this setting
instead of just two, we expected models to perform much
worse under the transformations. And this is exactly what
we observed in the results (Fig. 3 and Supp. Figs. 9, 11,
6
Fig. 2. A selection of item-response curves for the 2AFC task. The natural scene experiment (left) reflects the classification accuracy across all 1K ImageNet
classes and the rendered scene experiments (center and right) reflect the accuracy across 40 classes. Each experiment used five well-known CNNs [6], [17]–
[19]. A perfect curve would be a flat line at the top of the plot. The images at the bottom of each curve show how the perturbations increase from right to
left, starting with no perturbation (i.e., the original image) for all conditions. The 2AFC experiments indicate that the models tolerate different perturbations
in some settings, but not in all cases (e.g., Gaussian blur). The red dot indicates mean human performance for a selected stimulus level. These plots (as well
as the next sets in Figs. 3 & 4) are best viewed in color.
12 & 15). For instance, compare the plots for increasing
brightness (Fig. 3 upper-left) and positive rotation in the
x-axis (Fig. 3 lower-center) to the corresponding plots in
Fig. 2. For both transformation types the networks, which
only experienced moderate dips in performance for the most
extreme perturbations in the 2AFC case, fall to under ∼20%
accuracy at points. A caveat to the MAFC decision function
is that because it is patterned after the classification task in
computer vision, it only uses class labels to make its decisions.
Thus it leaves out the layer-specific activation information
that was used in the 2AFC case. This highlights an important
trade-off that can occur when designing decision functions for
psychophysics evaluations: task fidelity versus task difficulty.
Curiously, there are large asymmetries for some of the
transformations with increasing and decreasing perturbation
levels. See the plots (Fig. 3 upper-left and Supp. Figs. 11
& 12) for brightness, contrast, and sharpness. Contrast is a
particularly intriguing case. As a transformation, contrast is
a non-linear single pixel-level operation applied globally to
an image. In the positive direction, contrast is increased, and
the performance of each network degrades rapidly. In the
negative direction, contrast is decreased, but the performance
of each network remains relatively stable until the objects
have very low contrast. This suggests a contrast sensitivity
problem under the MAFC decision function that is the opposite
of what human patients with contrast-related vision deficits
struggle with. There is a positive aspect to this finding — while
diminished contrast sensitivity may induce night-blindness in
a human driver, CNN-based autonomous driving systems can
be expected to operate more effectively in the dark.
Dropout Experiments. The experiments we have looked
at thus far assume deterministic outputs. What about settings
with stochastic outputs that support uncertainty in decision
making? Gal and Ghahramani [43] introduced dropout at
testing time to implement Bayesian inference in neural networks. What sort of variability does this lead to under various
transformations and perturbation levels, and what does this
tell us about the certainty of the experiments above? The
setup for these experiments is identical to the setup for the
MAFC experiments (including preferred views) except that
during evaluation we applied dropout at test time to the Caffe
version of AlexNet (which was also trained with dropout).
Deploying the pre-trained model for each test, we dropped
out 50% (because the model is large) of the neurons in layers
fe6 and fe7 by uniformly randomly setting their activations
to zero. This is repeated with 5 different random seeds for
each transformation except salt & pepper noise and linear
occlusion, which were not performed due to randomness in
their underlying perturbation functions.
As anticipated, some variability in the base model performance was introduced (Fig. 4 and Supp. Figs. 17-20). But
importantly, most runs still demonstrated a large measure of
consistency (e.g., Fig. 4) across the ranges of perturbations,
indicating higher degrees of model certainty. This is a good
stability property — when a model fails, it does so in an
expected way across different dropout configurations, lending
credibility to the initial characterizations of the behavior in
the earlier experiments. More variability was observed for
the rendered objects versus the natural scenes. This can be
attributed to the use of the 3D objects that were outside of the
training set for all of the models. The maximum difference
observed between points from two runs for any transformation
was 16.5% for sharpness applied to 3D objects (Supp. Fig. 19).
In over half the cases, the maximum difference was over 10%.
Human Comparisons. Using psychophysics, model performance can be directly compared to human performance. To obtain human data points (red dots) for Figs. 2-4, we conducted
a study with 24 participants. Each participant performed the
2AFC task as described above, but to mitigate fatigue, only 5
trials with a fixed psychophysical parameter setting were given
for each of the transformations from Fig. 2. The participants
also performed the MAFC task, but were limited to 3 choices
instead of the full 1000 classes to make the task tractable. For
those experiments, participants performed in 5 trials with a
fixed psychophysical parameter setting for each transformation
in Fig. 3. The original images for each trial were chosen
randomly from the VGG-16 preferred views such that each
class was only used one time for each participant in order to
7
Fig. 3. A selection of item-response curves for the MAFC task. (Top) Natural scenes. (Bottom) Rendered scenes. The plots in the upper-left and bottom-center
are directly comparable to the corresponding plots in Fig. 2. For all perturbations except linear occlusion, human and model behavior is not consistent.
prevent participants from learning while performing the task.
For all trials on both tasks, the sample images were presented
for 50ms and subjects had unlimited time to answer.
Even without generating a full psychometric curve for
the human subjects, it was apparent that only two out of
nine experiments showed any consistency between human and
model performance (Fig. 2 center and Fig. 3 top-center). In
some cases, human and model behavior was wildly divergent
(e.g., salt & pepper noise in Fig. 3). While human performance
was superior to model performance in most cases, there was
one case where humans were slightly worse than the models:
increasing brightness (Fig. 2 left). Brightness adjustment in
image processing is a constant multiplicative change to all the
pixel values, which preserves edges and allows the networks
to recognize the geometry of an object almost until saturation
is reached. Humans were also good at this task, but were still
∼9% worse than VGG-19 for the perturbation level analyzed.
V. D ISCUSSION
In visual psychophysics, we have a convenient and practical
alternative to traditional dataset evaluation. However, the use
of psychophysics testing and datasets are not mutually exclusive. One needs datasets to form a training basis for any datadriven model. Moreover, there is major utility to having a large
amount of such data — this is essential for making machine
learning capture enough intraclass variance to generalize well
to unseen class instances. Data augmentation [6], [46] is
an obvious strategy for leveraging the rendered images that
were problematic for a model during psychophysics testing
to expand the scope of the training set. However, this has
diminishing returns as datasets grow to sizes that exceed
available memory (or even disk space) during training. Using
more limited training data and reinforcement learning that
optimizes over item-response curves to correct for recognition
errors is likely a better path forward.
Recent research has shown that CNNs are able to predict
neural responses in the visual cortex of primates [32]. This,
coupled with excellent benchmark dataset results across multiple recognition domains, suggests that good progress is being
made towards reaching human-like performance. As a strong
counterpoint, our psychophysics experiments show that the
current most popular CNN models sometimes fail to correctly
classify images that humans do not make mistakes on. What
is missing from the models that is causing this behavioral
discrepancy? While certainly capturing some of the operation
of the ventral stream (particularly at the filtering stages of early
vision), today’s CNN architectures lack the intricate structure
and function of biological neural networks [47]. Advances in
connectomics [48], a subfield of neuroscience that is attempting to map circuits in the brain, and functional imaging [49],
which allows for the in vivo study of neuronal responses, are
likely to inform new biologically-inspired architectures in the
near future. With psychophysics as a guide to performance,
candidate models can more easily be accepted or dismissed as
being representative of some form of AI — making it harder
for us to be fooled by the person inside of the Chinese room.
ACKNOWLEDGMENT
The authors thank Lucas Parzianello for helping import
Blendswap models into PsyPhy and Brian Turnquist for providing feedback on an early draft of this work. Funding was
8
Fig. 4. Item-response curves for five different runs of an AlexNet model with
dropout applied at test time [43] for a 3D rotation transformation. The black
line indicates the mean of the five AlexNet curves. The maximum difference
between points on any two curves in this plot is 12.2%.
provided under IARPA contract #D16PC00002, NSF DGE
#1313583, and NSF SBIR Award #IIP-1621689. Hardware
support was generously provided by the NVIDIA Corporation.
R EFERENCES
[1] Z.-L. Lu and B. Dosher, Visual Psychophysics: From Laboratory to
Theory. MIT Press, 2013. 1, 2, 3
[2] F. Kingdom and N. Prins, Psychophysics: a Practical Introduction.
Academic Press, 2016. 1, 2, 3
[3] J. R. Searle, “Minds, brains, and programs,” Behavioral and Brain
Sciences, vol. 3, no. 03, pp. 417–424, 1980. 1
[4] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma,
Z. Huang, A. Karpathy, A. Khosla et al., “ImageNet large scale visual
recognition challenge,” IJCV, vol. 115, no. 3, pp. 211–252, 2015. 1
[5] ImagetNet
Large
Scale
Visual
Recognition
Challenge
2012 (ILSVRC2012), “http://image-net.org/challenges/LSVRC/2012/
index,” Accessed: 2016-10-12. 1, 5
[6] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification
with deep convolutional neural networks,” in NIPS, 2012. 1, 2, 5, 6, 7
[7] A. Torralba and A. A. Efros, “Unbiased look at dataset bias,” in IEEE
CVPR, 2011. 1
[8] T. Tommasi, N. Patricia, B. Caputo, and T. Tuytelaars, “A deeper look at
dataset bias,” ser. LNCS, J. Gall, P. Gehler, and B. Leibe, Eds. Springer,
2015, vol. 9358, pp. 504–516. 1
[9] S. E. Embretson and S. P. Reise, Item response theory for psychologists.
Lawrence Erlbaum Associates, Inc., 2000. 2
[10] S. Hecht, S. Shlaer, and M. H. Pirenne, “Energy, quanta, and vision,”
The Journal of General Physiology, vol. 25, no. 6, pp. 819–840, 1942.
2
[11] J. K. Bowmaker and H. J. Dartnall, “Visual pigments of rods and cones
in a human retina,” The Journal of Physiology, vol. 298, no. 1, pp.
501–511, 1980. 2
[12] B. Duchaine and K. Nakayama, “The Cambridge face memory test:
Results for neurologically intact individuals and an investigation of
its validity using inverted face stimuli and prosopagnosic participants,”
Neuropsychologia, vol. 44, no. 4, pp. 576–585, 2006. 2
[13] J. B. Tenenbaum, C. Kemp, T. L. Griffiths, and N. D. Goodman, “How
to grow a mind: Statistics, structure, and abstraction,” Science, vol. 331,
no. 6022, pp. 1279–1285, 2011. 2, 3
[14] I. Yildirim, T. D. Kulkarni, W. A. Freiwald, and J. B. Tenenbaum,
“Efficient and robust analysis-by-synthesis in vision: A computational
framework, behavioral tests, and modeling neuronal representations,” in
Annual Conference of the Cognitive Science Society (CogSci), 2015. 2,
3
[15] T. D. Kulkarni, P. Kohli, J. B. Tenenbaum, and V. Mansinghka, “Picture:
A probabilistic programming language for scene perception,” in IEEE
CVPR, 2015. 2, 3
[16] J. Wu, C. Zhang, T. Xue, W. T. Freeman, and J. B. Tenenbaum,
“Learning a probabilistic latent space of object shapes via 3d generativeadversarial modeling,” in NIPS, 2016. 2, 3
[17] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick,
S. Guadarrama, and T. Darrell, “Caffe: Convolutional architecture for
fast feature embedding,” arXiv preprint arXiv:1408.5093, 2014. 2, 5, 6
[18] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
V. Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,”
in IEEE CVPR, 2015. 2, 5, 6
[19] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” CoRR, vol. abs/1409.1556, 2014. 2, 5,
6
[20] M. Riesenhuber and T. Poggio, “The individual is nothing, the class everything: Psychophysics and modeling of recognition in object classes,”
MIT, Tech. Rep. AIM-1682, October 2000. 2
[21] ——, “Hierarchical models of object recognition in cortex,” Nature
Neuroscience, vol. 2, no. 11, pp. 1019–1025, 1999. 2
[22] S. Eberhardt, J. Cader, and T. Serre, “How deep is the feature analysis
underlying rapid visual categorization?” in NIPS, 2016. 2
[23] R. Geirhos, D. H. J. Janssen, H. H. Schütt, J. Rauber, M. Bethge,
and F. A. Wichmann, “Comparing deep neural networks against humans: object recognition when the signal gets weaker,” arXiv preprint
1706.06969, 2017. 2
[24] H. E. Gerhard, F. A. Wichmann, and M. Bethge, “How sensitive is the
human visual system to the local statistics of natural images?” PLoS
Computational Biology, vol. 9, no. 1, p. e1002873, 2013. 2
[25] W. J. Scheirer, S. E. Anthony, K. Nakayama, and D. D. Cox, “Perceptual
annotation: Measuring human vision to improve computer vision,” IEEE
T-PAMI, vol. 36, no. 8, August 2014. 2
[26] L. Germine, K. Nakayama, B. C. Duchaine, C. F. Chabris, G. Chatterjee, and J. B. Wilmer, “Is the web as good as the lab? comparable
performance from web and lab in cognitive/perceptual experiments,”
Psychonomic Bulletin & Review, vol. 19, no. 5, pp. 847–857, 2012. 2
[27] C. Vondrick, H. Pirsiavash, A. Oliva, and A. Torralba, “Learning visual
biases from human imagination,” in NIPS, 2015. 2
[28] A. J. O’Toole, P. J. Phillips, F. Jiang, J. Ayyad, N. Penard, and
H. Abdi, “Face recognition algorithms surpass humans matching faces
over changes in illumination,” IEEE T-PAMI, vol. 29, no. 9, pp. 1642–
1646, 2007. 2
[29] A. J. O’Toole, X. An, J. Dunlop, V. Natu, and P. J. Phillips, “Comparing
face recognition algorithms to humans on challenging tasks,” ACM
Transactions on Applied Perception (TAP), vol. 9, no. 4, p. 16, 2012. 2
[30] P. J. Phillips and A. J. O’Toole, “Comparison of human and computer
performance across face recognition experiments,” Image and Vision
Computing, vol. 32, no. 1, pp. 74–85, 2014. 2
[31] C. F. Cadieu, H. Hong, D. Yamins, N. Pinto, N. J. Majaj, and J. J.
DiCarlo, “The neural representation benchmark and its evaluation on
brain and machine,” in ICLR, 2013. 2
[32] D. Yamins, H. Hong, C. F. Cadieu, E. A. Solomon, D. Seibert, and
J. J. DiCarlo, “Performance-optimized hierarchical models predict neural
responses in higher visual cortex,” Proceedings of the National Academy
of Sciences, vol. 111, no. 23, pp. 8619–8624, 2014. 2, 7
[33] H. Hong, D. Yamins, N. J. Majaj, and J. J. DiCarlo, “Explicit information
for category-orthogonal object properties increases along the ventral
stream,” Nature Neuroscience, vol. 19, no. 4, pp. 613–622, 2016. 2
[34] R. T. Pramod and S. P. Arun, “Do computational models differ systematically from human object perception?” in IEEE CVPR, 2016. 2
[35] D. Hoiem, Y. Chodpathumwan, and Q. Dai, “Diagnosing error in object
detectors,” in ECCV, 2012. 3
[36] M. J. Wilber, V. Shmatikov, and S. Belongie, “Can we still avoid
automatic face detection?” in IEEE WACV, 2016. 3
[37] C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow,
and R. Fergus, “Intriguing properties of neural networks,” in ICLR, 2014.
3
[38] A. Nguyen, J. Yosinski, and J. Clune, “Deep neural networks are easily
fooled: High confidence predictions for unrecognizable images,” in IEEE
CVPR, 2015. 3
[39] A. Bendale and T. E. Boult, “Towards open set deep networks,” in IEEE
CVPR, 2016. 3
[40] V. Blanz, M. J. Tarr, and H. H. Bülthoff, “What object attributes
determine canonical views?” Perception, vol. 28, no. 5, pp. 575–599,
1999. 3
[41] J. W. Peirce, “PsychoPy: Psychophysics software in python,” Journal of
Neuroscience Methods, vol. 162, no. 1-2, pp. 8–13, 2007. 3
[42] W. Jakob, “Mitsuba https://pillow.readthedocs.io/en/3.0.0/index.html,”
Accessed: 2016-11-05. 4
[43] Y. Gal and Z. Ghahramani, “Dropout as a bayesian approximation:
Representing model uncertainty in deep learning,” in ICML, 2016. 5, 6,
8
[44] Mitsuba renderer, “http://www.blendswap.com,” 2010. 5
[45] BVLC Caffe, “https://github.com/BVLC/caffe,” Accessed: 2016-11-10.
5
9
[46] K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman, “Return of
the devil in the details: Delving deep into convolutional nets,” in BMVC,
2014. 7
[47] D. D. Cox and T. Dean, “Neural networks and neuroscience-inspired
computer vision,” Current Biology, vol. 24, no. 18, pp. R921–R929,
2014. 7
[48] N. Kasthuri, K. J. Hayworth, D. R. Berger, R. L. Schalek et al.,
“Saturated reconstruction of a volume of neocortex,” Cell, vol. 162,
no. 3, pp. 648–661, 2015. 7
[49] K. Ohki, S. Chung, Y. H. Ch’ng, P. Kara, and R. C. Reid, “Functional
imaging with cellular resolution reveals precise micro-architecture in
visual cortex,” Nature, vol. 433, no. 7026, pp. 597–603, 2005. 7
| 1cs.CV
|
1
Reconstruction Codes for DNA Sequences with
Uniform Tandem-Duplication Errors
arXiv:1801.06022v1 [cs.IT] 18 Jan 2018
Yonatan Yehezkeally, Student Member, IEEE, and Moshe Schwartz, Senior Member, IEEE
Abstract—DNA as a data storage medium has several advantages, including far greater data density compared to electronic
media. We propose that schemes for data storage in the DNA of
living organisms may benefit from studying the reconstruction
problem, which is applicable whenever multiple reads of noisy
data are available. This strategy is uniquely suited to the medium,
which inherently replicates stored data in multiple distinct ways,
caused by mutations. We consider noise introduced solely by
uniform tandem-duplication, and utilize the relation to constantweight integer codes in the Manhattan metric. By bounding
the intersection of the cross-polytope with hyperplanes, we
prove the existence of reconstruction codes with greater capacity
than known error-correcting codes, which we can determine
analytically for any set of parameters.
Index Terms—DNA storage, reconstruction, string-duplication
systems, tandem-duplication errors
I. I NTRODUCTION
DNA is attracting considerable attention in recent years as a
medium for data storage, due to its high density and longevity
[6]. Data storage in DNA may provide integral memory for
synthetic-biology methods, where such is required, and offer
a protected medium for long-period data storage [4], [33].
In particular, storage in the DNA of living organisms is
now becoming feasible [31]; it has varied usages, including
watermarking generically modified organisms [3], [13], [27] or
research material [18], [33], and even afford some concealment
to sensitive information [7]. Naturally, therefore, data integrity
in such media is of great interest.
Several recent works have studied the inherent constraints
of storing and retrieving data from DNA. While desired
sequences (over quaternary alphabet) may be synthesized
(albeit, while suffering from substitution noise), generally data
can only be read by observation of its subsequences, quite
possibly an incomplete observation [19]. Moreover, the nature
of DNA and current technology results in asymmetric errors
which depend upon the dataset [10]. The medium itself also
introduces other types of errors which are atypical in electronic
storage, such as symbol/block-deletion and adjacent transpositions (possibly inverted) [11]. Finally, the purely combinatorial
problem of recovering a sequence from the multiset of all
its subsequences (including their numbers of incidence), was
also studied [1], [32], as well as coding schemes involving
only these multisets (or their profile vectors – describing the
incidence frequency of each subsequence) [30].
A version of this paper was submitted to ISIT’2018.
This work was supported in part by ISF grant no. 130/14.
The authors are with the Department of Electrical and Computer Engineering, Ben-Gurion University of the Negev, Beer Sheva 8410501, Israel (e-mail:
[email protected]; [email protected]).
Other works were concerned with data storage in the
DNA of a living organism. While this affords some level
of protection to the data, and even propagation (through
DNA replication), it is also exposed to specific noise mechanisms due to mutations. Examples of such noise include
symbol insertions, deletion, substitutions (point-mutation), and
duplication (including tandem- and interspersed-duplication).
Therefore, schemes for data storage in live DNA must address
data integrity and error-correction.
In an effort to better understand these typical noise mechanism, their potential to generate the diversity observed in
nature was studied. [12] classified the capacity and/or expressiveness of the systems of sequences over a finite alphabet
generated by four distinct substring duplication rules: endduplication, tandem-duplication, tandem-palindromic- duplication, and interspersed-duplication. [15] fully characterized
the expressiveness of bounded tandem-duplication systems,
proved bounds on their capacity (and, in some cases, even
exact values). [17] later showed that when point-mutations
act together with tandem-duplication as a sequence-generation
process, they may actually increase the capacity of the generated system. [2] looked at the typical duplication distance
of binary sequences; i.e., the number of tandem-duplications
generating a binary sequence from its root. It was proven
that for all but an exponentially small number of sequences
that number is proportional to the sequence length. Further,
when tandem-duplication is combined with point-mutations
(here, only within the duplicated string), it was shown that
the frequency of substitutions governs whether that distance
becomes logarithmic.
The generative properties of interspersed-duplication were
also studied from a probabilistic point of view. [9] showed (under assumption of uniformity) that the frequencies of incidence
for each subsequence converge to the same limit achieved by
an i.i.d. source, thus reinforcing the notion that interspersedduplication is–on its own–capable of generating diversity.
[8] specifically looked at tandem- and end-duplication, and
found exact capacities in the case of duplication length 1
by a generalization of the Pólya urn model that applies to
strings. It also tightly bounded the capacity of complement
tandem-duplication, a process where the duplicated symbol is
complemented (using binary alphabet).
Finally, error-correcting codes for data affected by tandemduplication have been studied in [16], which presented a
construction of optimal-size codes for correcting any number
of errors under uniform tandem-duplication (fixed duplication
length), computing their (and thus, the optimal-) capacity. It
also presented a framework for the construction of optimal
2
codes for the correction of a fixed number of errors. Next, it
studies bounded tandem-duplications, where a characterization
of the capacity of error-correcting codes is made for small
constants. In general, it characterized the cases where the
process of tandem-duplication can be traced back uniquely.
However, classical error-correction coding ignores some
properties of the DNA storage channel; namely, stored information is expected to be replicated, even as it is mutated.
This lends itself quite naturally to the reconstruction problem
[26], which assumes that data is simultaneously transmitted
over several noisy channels, and a decoder must therefore
estimate that data based on several (distinct) noisy versions
of it. Solutions to this problem have been studied in several
contexts. It was solved in [26] for sequence reconstruction over
finite alphabets, where several error models were considered,
such as substitutions, transpositions and deletions. Moreover,
a framework was presented for solving the reconstruction
problem in general cases of interest in coding theory, utilizing
a graph representation of the error model, which was further
developed in [24], [25]. The problem was also studied in the
context of permutation codes with transposition and reversal
errors [20]–[22], and partially solved therein. Later, applications were found in storage technologies [5], [34], [35], since
modern application might preclude the retrieval of a single
data point, in favor of multiple-point requests. However, the
problem hasn’t been addressed yet for data storage in DNA
of living organisms, where it may be most applicable.
In this paper, we study the reconstruction problem over
DNA sequences, with uniform tandem-duplication error. The
paper is organized as follows: In Section II we present notations and definitions. In Section III we demonstrate that
reconstruction codes are error-correcting codes and find their
requisite minimal-distance, as a function of the reconstruction
parameters. In Section IV we then study bounds on the sizes
of such codes by an isometric embedding to constant-weight
codes in the Manhattan metric. Finally, in Section V we
show that by considering reconstruction codes we improve the
capacity of known error-correcting codes, and conclude with
closing remarks in Section VI.
II. P RELIMINARIES
Throughout this paper, though human DNA is composed of
four nucleotide bases, we observe the more general case of
sequences over a finite alphabet; since the alphabet elements
are immaterial to our discussion, we denote it throughout as
Zq . We observe
the set of finite sequences (also: words) over
S
n
∗
it Z∗q = ∞
Z
n=0 q . For any two words u, v ∈ Zq , we denote
n
their concatenation uv. For each word x ∈ Zq , we denote its
length |x| = n. We also take special note of the set of words
with length higher
than or equal to some 0 < k ∈ N, which
we denote Z>k
=
x ∈ Z∗q |x| > k . For ease of notation,
q
we let N stand for the set of non-negative integers throughout
the paper; when an integer is assumed to be strictly positive,
we make special note of that fact.
For 0 < k ∈ N, i ∈ N, we define a tandem-duplication of
duplication-length k by the mappings
(
uvvw x = uvw, |u| = i, |v| = k,
Tk,i (x) =
x
otherwise.
If y = Tk,i (x) and y 6= x (which occurs whenever |x| > i+k),
we say that y is a descendant of x, and denote x =⇒ y.
k
t
Further, given a sequence {xj }j=0 ⊆ Z∗q such that for all
0 6 j < t, xj =⇒ xj+1 , we say that xt is a t-descendant
k
t
of x0 , and denote x0 =⇒ xt . For completeness, we also
0
k
denote x =⇒ x. Finally, if there exists some t ∈ N such that
k
t
∗
x =⇒ y, we also denote x =⇒ y.
k
k
We denote the set of t-descendants of x ∈ Z∗q as
t
Dkt (x) = y ∈ Z∗q x =⇒ y ,
k
for some t ∈SN. We also denote the descendant-cone of x
t
by Dk∗ (x) = ∞
t=0 Dk (x). For ease of notation, we define for
t ∈ N the set of t-ancestors of some given x ∈ Zq>k+1 as
Dk−t (x) = y ∈ Z∗q x ∈ Dkt (y) ,
and similarly Dk−∗ (x).
We say that x ∈ Zq>k is irreducible if it is not the descendant
of any word. We exclude from the definition shorter words,
for which the condition vacuously holds. We denote by Irrk
the set of all irreducible words, and Irrk (n) = Irrk ∩Znq .
It was shown in [17], [23] that for each word x ∈ Z>k
q , a
unique irreducible word exists for which x is a descendant.
We call it the root of x, and denote it by Rk (x). This induces
an equivalence relation by x ∼k y if Rk (x) = Rk (y).
We also follow [17] in defining, for x ∈ Z>k
q , Pref k (x)
as the length-k prefix of x, and Suff k (x) as its suffix; i.e., if
x = uu′ = v ′ v where |u| = |v| = k, then Pref k (x) = u and
Suff k (x) = v. Using this notation, we define an embedding
φk : Zq>k → Zkq × Z∗q by
φk (x) = Pref k (x), Suff |x|−k (x) − Pref |x|−k (x) .
It is seen in [17] that this mapping is indeed injective. Further,
it was shown that, if one also defines ζk,i : Zkq ×Z∗q → Zkq ×Z∗q
by
(
(a, b1 0k b2 ) b = b1 b2 , |b1 | = i,
ζk,i (a, b) =
(a, b)
otherwise,
where 0 < k ∈ N, i ∈ N, then
φk (Tk,i (x)) = ζk,i (φk (x)).
The simplicity of ζk,i in comparison to Tk,i motivates the analysis of tandem-duplications using the φk images of sequence.
If b ∈ Z∗q is composed of the subsequences
b = 0s1 w1 0s2 · · · wm 0sm+1 ;
∗
w1 , . . . , wm ∈ (Zq \ {0})
we define
µ(b) = 0s1 mod k w1 0s2 mod k · · · wm 0sm+1 mod k ,
js
k
j s k
m+1
1
,...,
.
σ(b) =
k
k
3
Lemma 4 For all y, z ∈ Dk∗ (x) it holds that y =⇒ z if and
We may note that wtH (b) = wtH (µ(b)), where wtH is the
Hamming weight, and σ(b) ∈ NwtH (b)+1 = NwtH (µ(b))+1 .
We also observe that b is recoverable from σ(b), µ(b). It was
proven in [17, Cor. 10] that if φk (x) = (a, b) then
only if ψx (z) = ψx (y) + ej , for some unit vector ej ∈ Nm+1 .
φk (Rk (x)) = (a, µ(b)).
φk (y) = (a, 0s1 w1 0s2 · · · wm 0sm+1 ).
Thus, if x, y ∈ Z>k
q , φk (x) = (a1 , b1 ) and φk (y) = (a2 , b2 ),
then x ∼k y if and only if a1 = a2 and µ(b1 ) = µ(b2 ).
Moreover, x ∈ Irrk if and only if σ(b1 ) = (0, 0, . . . , 0).
Finally, for n > k and x, y ∈ Znq we define
dk (x, y) = min t ∈ N Dkt (x) ∩ Dkt (y) 6= ∅ ,
or dk (x, y) = ∞ if {t ∈ N | Dkt (x) ∩ Dkt (y) 6= ∅} = ∅. It was
shown in [17, Lem. 14] that dk (x, y) = ∞ if and only if
x 6∼k y, hence dk (·, ·) is finite on Dkt (x), for any particular
x ∈ Z>k
q . Furthermore, [17, Lem. 19] shows that for any
x ∼k y with |x| = |y| it holds that
dk (x, y) =
1
2
kσ(b1 ) − σ(b2 )k1 ,
k
Proof: Let us denote
Now note that y =⇒ z if and only if there exists i ∈ N
k
such that φk (z) = ζi,k (φk (y)) and φk (z) 6= φk (y), which
occurs if and only if there exists 1 6 j 6 m + 1 such that
φk (z) = (a, 0s1 w1 0s2 · · · wj−1 0sj +k · · · wm 0sm+1 ).
This, in turn, can equivalently be stated as ψx (z) = ψx (y)+ej
where ej is a unit vector.
∗
Corollary 5 For all y, z ∈ Dk∗ (x) it holds that y =⇒ z if
k
and only if ψx (y) ψx (z).
∗
Proof: Using Lemma 4 we observe that y =⇒ z if and
only if y = z (in which case ψx (y) = ψx (z)) or
thus dk (·, ·) defines a metric on Znq for all n > k.
ψx (z) = ψx (y) +
III. R ECONSTRUCTION C ODES
The reconstruction problem in the context of uniform
tandem-duplication errors can be stated as follows: suppose
data is encoded in C ⊆ Znq , and suppose we later are able to
read distinct x0 , x1 , . . . , xN ∈ Dkt (c) for some specific c ∈ C
and t ∈ N; can we uniquely identify c?
As seen in [26], successful reconstruction is ensured if and
only if the code C satisfies the following requirement:
Znq
Definition 1 Take N, t, n > 0. We say that C ⊆
is a
uniform tandem-duplication reconstruction code, which we
abbreviate as an (N, t, k)q -UTR code, if
max |Dkt (c) ∩ Dkt (c′ )| 6 N.
c,c′ ∈C
c6=c′
for some unit vectors ej1 , . . . , ejt ∈ Nm+1 . The claim follows.
Theorem 6 ψx is a poset isomorphism from
to Nm(x)+1 , .
Definition 2 We let ψx : Dk∗ (x) → Nm+1 be defined by
ψx (y) = σ(b′ ), where φk (y) = (a, b′ ).
It was noted in the previous section that ψx is then an
isometry from (Dk∗ (x), dk ) to Nm+1 , 12 k·k1 .
Definition 3 We define on Nm+1 the partial order b′ b′′ if
for every coordinate i = 1, . . . , m + 1 it holds that b′i 6 b′′i (
is the well-known product order).
∗
Dk∗ (x), =⇒
k
Proof: Using Corollary 5 and thefact that Nm+1
, is a
∗
partially ordered set we observe that Dk∗ (x), =⇒ is also a
k
poset, and that ψx is a poset homomorphism. We have already
noted that it is injective (as an isometry).
Finally, note that ψx (x) = 0m+1 . Every element of Nm+1
can be decomposed into
t
X
eju = ψx (x) +
t
X
eju
u=1
u=1
Throughout this section we fix some x ∈ Irrk , and denote
φk (x) = (a, b).
As noted above, for all y ∈ Dk∗ (x), we have φk (y) = (a, b′ ),
with µ(b′ ) = b (hence, in particular, wtH (b′ ) = wtH (b)). We
therefore denote m = m(x) = wtH (b) and make the following
definition:
eju
u=1
In this section we find the size of intersection of descendant
cones, and thereby evaluate the size of (N, t, k)q -UTR codes.
A. The structure of descendant cones
t
X
k
for some unit vectors ej1 , . . . , ejt ∈ Nm+1 . By Lemma 4 we
get a sequence
x =⇒ y1 =⇒ · · · =⇒ yt
k
k
k
Pt
such that ψx (yt ) = u=1 eju , thus ψx is also surjective.
∗
Thus, =⇒ endows Dk∗ (x) with a lattice structure; We
k
denote the join of y, z ∈ Dk∗ (x) as y ∨ z, and their meet
y ∧ z. We have the following immediate results:
Corollary 7 For all y, z ∈ Dk∗ (x),
Dk∗ (y) ∩ Dk∗ (z) = Dk∗ (y ∨ z).
Lemma 8 |Dkt (x)| =
t+m(x)
.
m(x)
Proof: By Lemma 4 we know that
Dkt (x) = {y ∈ Dk∗ (x) | kψx (y)k1 = t}.
4
Since ψx : Dk∗ (x) → Nm(x)+1 is bijective,
|Dkt (x)| equals the
Pm+1
number of distinct integer solutions to j=1 xj = t, where
x1 , . . . , xm+1 > 0 (equivalently, the number of distinct ways
to distribute t identical balls into m(x) + 1 bins).
Corollary 12 For n > |x|, C ⊆ Dk∗ (x) ∩ Znq is an (N, t, k)q UTR code if and only if
B. Reconstruction from uniform tandem-duplication errors
Proof: The claim follows from Lemma 8, Lemma 10.
In other words, the intersection of a uniform-tandemduplication reconstruction code C with the descendant cone
of any irreducible word Dk∗ (x) is simply an error-correcting
code with suitable minimal distance. Note, however, that
this doesn’t hold for C in general: the minimal distance is
dependent on x, and also Dk∗ (x) ∩ Znq depends on |x|.
We start by rephrasing the properties of descendant cones
as seen in the previous section:
Corollary 9 Suppose that x, x′ ∈ Irrk . Then the following
conditions are equivalent:
1) m(x′ ) = m(x).
′
2) For
Dk∗ (x′ ) and y ∈
Dk∗ (x), it holds that
all y ∈
∗
∗
D∗ (y ′ ), =⇒ ∼
= D∗ (y), =⇒ .
k
k
k
k
3) For all t > 0, y ′ ∈ Dk∗ (x′ ) and y ∈ Dk∗ (x), it holds that
|Dkt (y ′ )| = |Dkt (y)|.
4) There exist some t > 0, y ′ ∈ Dk∗ (x′ ) and y ∈ Dk∗ (x)
such that |Dkt (y ′ )| = |Dkt (y)|.
Proof: That (1) =⇒ (2) follows from Theorem 6.
Implications (2) =⇒ (3) =⇒ (4) follow trivially.
To show that (4) =⇒ (1), note by previous parts that
Dk∗ (y ′ ) ∼
= Dk∗ (x′ ) and Dk∗ (y) ∼
= Dk∗ (x). In particular, we have
t
′
t
that |Dk (x )| = |Dk (x)|. Lemma 8 now implies (1).
These observations allow us to classify the intersection size
of descendant-cones.
Lemma 10 For n > k, t > 0 and y1 , y2 ∈ Znq ,
t−dk (y1 ,y2 )
Dkt (y1 ) ∩ Dkt (y2 ) = Dk
(Rk (y1 ))
if y1 ∼k y2 and t > dk (y1 , y2 ), or Dkt (y1 ) ∩ Dkt (y2 ) = ∅
otherwise.
Proof: As already noted, if y1 6∼k y2 or t < dk (y, z) we
have Dkt (y) ∩ Dkt (z) = ∅. Otherwise, by Corollary 7 we have
Dk∗ (y1 ) ∩ Dk∗ (y2 ) = Dk∗ (y1 ∨ y2 ).
Note, then, that if z ∈ Dkt (y1 ) ∩ Dkt (y2 ) then
y1 ∨ y2
t−dk (y1 ,y2 )
=⇒
k
z,
therefore
t−dk (y1 ,y2 )
Dkt (y1 ) ∩ Dkt (y2 ) = Dk
(y1 ∨ y2 ),
and by Corollary 9 we have
t−d (y ,y )
Dk k 1 2 (y1
t−d (y ,y )
∨ y2 ) ∼
= Dk k 1 2 (Rk (y1 ∨ y2 ))
t−dk (y1 ,y2 )
= Dk
(Rk (y1 )).
This allows us, using the following definition, to characterize (N, t, k)q -UTR codes.
Definition 11 Take N, m > 0. We denote for t > 0
t−δ+m
dN,t (m) = min δ ∈ N
6N
m
min dk (c, c′ ) > dN,t (m(x)),
c,c′ ∈C
c6=c′
C. Size of reconstruction codes
In this section we aim to estimate the maximal size of
(N, t, k)q -UTR codes.
Definition 13 For m, r > 0 we denote
Xm+1
m+1
m+1
m
xj = r ,
∆r = (xi )i=1 ∈ N
j=1
which we will call the (m, r)−simplex.
Theorem 14 We take positive integers N, t and n > k. For
C ⊆ Znq and x ∈ Irrk we denote Cx = C ∩ Dk∗ (x) and define
r(x) = n−|x|
k .
If Cx 6= ∅ then r(x) ∈ N and r(x) < nk . Moreover, C is
an (N, t, k)q -UTR code if and only if for all x ∈ Irrk such
m(x)
that Cx 6= ∅, the image ψx (Cx ) ⊆ ∆r(x) satisfies
min 21 kc − c′ k1 c 6= c′ ∈ ψx (Cx ) > dN,t (m(x)).
Proof: If C ∩ Dk∗ (x) 6= ∅ then it follows from the
definitions that for some r ∈ N we have
|x|
+ rk = n;
since |x| > k, necessarily r = r(x) < nk . Furthermore,
C ∩ Dk∗ (x) = C ∩ Dkr (x), hence we’ve seen in the proof
of Theorem 6 that for all y ∈ Dkr (x) we have ψx (y) =
Pr
m(x)
.
u=1 eju ∈ ∆r
To conclude the proof, we recall that for x, x′ ∈ Irrk we
have Dk∗ (x) ∩ Dk∗ (x′ ) = ∅, and that ψx is a bijective isometry
into Nm(x)+1 , 12 k·k1 .
We therefore see that finding error-correcting codes for
uniform tandem-duplication essentially amounts to finding
error-correcting codes in the Manhattan metric over ∆m
r . We
start by notating the maximal size of such codes:
Definition 15 For m, r > 0 and d > 0 we define
o
n
′
1
kc
−
c
k
>
d
.
M (m, r, d) = max |C| C ⊆ ∆m
r , min
1
2
′
c,c ∈C
c6=c′
We now note that if C ⊆ Znq , x, x′ ∈ Irrk (n − rk) (i.e.,
r(x) = r(x′ ) = r) and m(x) = m(x′ ), then Dkn−rk (x) ∼
=
Dkn−rk (x′ ) (through, e.g., ψx−1
′ ◦ ψx ). It is therefore practical
to assume |Cx | = |Cx′ | = M (m, r, dN,t (m)) for all such
x, x′ . This results in the following corollary, which concludes
this section:
5
Corollary 16 If C ⊆ Znq is an (N, t, k)q -UTR code, and for
all x ∈ Irrk it holds that |Cx | = M (m, r, dN,t (m)), then
|C| =
=
⌊n/k⌋−1
X X
r=0
m
· M (m, r, dN,t (m))
X Xn
b ∈ Zn−(r+1)k
q
r=0
since the size of each ball is invariant in r, x, we also denote
B m (d) = B0m (d; 0) for ease of notation.
|{x ∈ Irrk (n − rk) | m(x) = m}|·
⌊n/k⌋−1
b is (0,k−1)q - RLL
wtH (b)=m
m
o
·
· M (m, r, dN,t (m)).
P
Proof: First, trivially, |C| = x∈Irrk |Cx |.
Observe that x ∈ Irrk satisfies Cx 6= ∅, r(x) = r and
m(x) = m, if and only if x ∈ Zqn−rk and in φk (x) = (a, b),
b is (0, k − 1)q -RLL, and wtH (b) = m.
The rest now follows from Theorem 14.
IV. C ODES
ON THE SIMPLEX WITH THE
M ANHATTAN
METRIC
Corollary 16 motivates us to estimate the optimal size of
error-correcting codes in the Manhattan metric over the (m, r)simplex. This section is dedicated to that question; a key
component of which will be the evaluation of the requisite
minimal distance dN,t (m).
A. Sphere size
In this section we evaluate the size of Manhattan-metric
spheres in ∆m
r , then establish the Gilbert-Varshamov bound
in the asymptotic regime.
Lemma 17 [28, Ch.10, Sec.11, Lem.7] For positive integers
k < n it holds that
r
r
k
k
n
n
n
2nH ( n ) 6
2nH ( n )
6
8πk(n − k)
2πk(n − k)
k
where H is the binary entropy function, defined H(p) =
−p log2 p − (1 − p) log2 (1 − p).
Proof: For all 0 < k < n we have n−1 6 k(n−k) 6
hence
1
1
n
1
1+
6 < 1,
6
2πk(n − k)
2π
n−1
π
1
n
>
.
8πk(n − k)
2πn
n2
4 ,
Thus the claim trivially follows from Lemma 17.
Definition 19 For 0 < m ∈ N and r ∈ Z, we denote the
hyperplane
Xm+1
m+1
m+1
x
=
r
,
Am
=
(x
)
∈
Z
i
i i=1
r
i=1
and omit r = 0 in A
=
Am
0 .
Lemma 20 For d, m > 0 denote
min{d,m}
X
m+1 d−1 m−k+d
B m (d) =
.
k
k−1
d
k=1
then
B m (d) 6 |B m (d)| 6 d · B m (d).
Proof: We may count |B m (d)| in the following manner:
1
• Partition by distance from the origin 2 kxk1 = r; then
m
for each 0 6= x ∈ B (d) we have 1 6 r 6 d.
• Further partition by the number k of strictly positive
coordinates xi > 0; then 1 6 k 6 min{r, m}.
• Choose positive coordinates, which may be done in
m+1
distinct ways.
k
Pk
• Find the number of solutions to
j=1 xij = r such that
r−1
each xij > 0. That is k−1
.
• The remaining m + 1 − k coordinates are non-positive,
and their sum is −r. There are m−k+r
such distinct
r
distributions.
Thus,
m
|B (d)| = 1 +
d
X
B m (r)
r=1
min{d,m}
d
m+1 X r−1
m−k+r
k
k−1
r
k=1
r=k
r−1 m−k+r
It now suffices to note that k−1
is a strictly
r
increasing function of r, thus the claim is satisfied.
In the following lemma we make an asymptotic evaluation
of |B m (d)|:
=1+
X
Lemma 21 Take µ ∈ (0, 1) and fix d > 0. Suppose
we’re given a sequence of dimensions (mn )n>0 such that
limn→∞ mnn = µ. Then limn→∞ n1 log2 |B mn (d)| = 0.
Lemma 18 For integers 0 < k < n it holds that
1
n
k
k
− log2 (2πn) 6 log2
.
< nH
nH
n
2
n
k
m
We also denote the ball of radius d > 0 about x ∈ Am
r as
m
m
1
Br (d; x) = y ∈ Ar 2 ky − xk1 6 d ;
Proof: For sufficiently large n we find that
d
X
mn + 1 d − 1 mn − k + d
mn
B (d) =
k
k−1
d
k=1
mn + 1 d−1 mn − 1 + d
6d·
.
2
d
d
Thus using Lemma 18 it is straightforward to show that
lim 1
n→∞ n
log2 B mn (d) = 0.
Recalling that by Lemma 20 we have |B mn (d)| 6 d · B mn (d),
the proof is completed.
We can now quickly state the Gilbert-Varshamov bounds on
M (m, r, d):
6
Lemma 22 For m, r, d > 0 we have
r+m
M (m, r, d) >
|B m (d − 1)| .
m
Next, in order to bound dN,t , we first establish an upper
bound on H(x):
Proof: The proof is an application of Lemma 8 and the
typical GV argument: if if C ⊆ ∆m
r is a maximal code such
1
that
for
all
x
=
6
y
∈
C
we
have
kx
− yk1 > d, then ∆m
r ⊆
2
S
m
x∈C B (d − 1; x).
Thus, we have the following:
Theorem 23 Take µ ∈ (0, 1), ρ > 0 and integer sequences (mn )n>0 , (rn )n>0 such that limn→∞ mnn = µ and
limn→∞ rnn = ρ. Also take a fixed d > 0. Then
!
1
1
.
(1)
lim log2 M (mn , rn , d) = (µ + ρ)H
n→∞ n
1 + µρ
Proof: We use Lemma 22 to find
1
n
log2 |B mn (d − 1)|
lim
t−δ
q Proof: Using Lemma 26 we mayqbound H 1 + m 6
log2 N
t−δ
2 t−δ
we get
m . Reordering the inequality 2
m 6
m
n
=
On the other hand, since M (mn , rn , d) 6 ∆m
rn
Lemma 18 implies the reverse inequality.
(log2 N )2
,
4m
yielding the claim.
OF RECONSTRUCTION CODES
We can now determine the capacity of (N, t, k)q -codes, in
some asymptotic regimes. We define the rate of of a family
of codes (C n ), C n ⊆ Znq , as
1
log2 |B mn (d − 1)| = 0;
n
R(C n ) = lim sup n1 logq |C n |,
rn +mn
mn
,
B. Minimal distance of reconstruction codes
Next, given N, t > 0 and m > 0, we establish bounds on
t−δ+m
dN,t (m) = min δ ∈ N
6N
m
seen in Definition 11.
For ease of notation in what follows, we make the notation,
for 1 6 x ∈ R:
H(x) = xH x1 .
Lemma 24 If N 6 m then dN,t (m) = t.
Proof:
We may verify by substitution that δ = t satisfies
6 N , while δ = t − 1 does not. Using the strict
monotonicity of s 7→ s+m
m , we’re done.
t−δ+m
m
Lemma 25 For N > m > 0 and t > 0 it holds that
log2 N
t−δ
6
.
dN,t (m) 6 min δ ∈ N H 1 +
m
m
Proof: Under the assumption, δ = t − 1 satisfies the
inequality t−δ+m
6 N . Therefore we may restrict the
m
minimum to δ < t, giving 0 < m < (t − δ) + m. Now,
Lemma 18 implies
!
t−δ
t−δ+m
1
log2
6m 1+
H
,
m
m
1 + t−δ
m
which completes the proof.
Theorem 27 Take N > m > 0. Then
(log2 N )2
dN,t (m) 6 max 1, t −
.
4m
V. C APACITY
Using Lemma 21 we know that
n→∞
Proof: The claim can be restated by the substitution
p = x1 as the known inequality H(p)2 6 4p(1 − p) (its proof
follows elementary calculus, and is omitted here).
Finally,
δ > t−
1
n
log2 M (mn , rn , d) >
rn + mn
> lim sup n1 log2
−
mn
lim inf
√
Lemma 26 For x > 1 it holds that H(x) 6 2 x − 1.
and the capacity of (N, t, k)q -codes as the supremum of rates
of families of such codes. Since [16] showed that Irrk (n)
can correct any number of tandem-duplication errors, they are
trivially (N, t, k)q -codes for all N, t. In this section we prove
that reconstruction codes have strictly higher capacity.
First,
we denote for any n, r ∈ N such that n > k and
r < nk , and any N, t ∈ N
MN,t (n, r) =
X
M (m, r, dN,t (m))·
m
n
· b ∈ Zn−(r+1)k
q
b is (0,k−1)q - RLL
wtH (b)=m
o
.
We recall for all n, if rn = arg maxr MN,t (n, r), that by
Corollary 16 we have an (N, t, k)q -code C ⊆ Znq with |C| >
MN,t (n, rn ). On the other hand, for every such C it holds
that |C| 6 nk MN,t (n, rn ). We therefore focus on maximizing
lim sup n1 logq MN,t (n, rn ) by choice of rn .
In what follows, we take γ ∈ (0, 1) and set rn = 1−γ
k n−1
for any n ∈ N for which rn ∈ N; we shall assume that such
n exist (hence, infinitely many exist), and refer only to such
indinces.
For all x ∈ Irrk (n − rn k) = Irrk (k + γn), recall that we
denoted φk (x) = (a, b) with b ∈ Zγn
q in (0, k − 1)q - RLL. We
shall build a reconstruction code in the descendant cones of
only such x, which we denote Cγ .
Lemma 28
There exists a system S ⊆ (0, k − 1)q - RLL and
θ ∈ 12 , 1 such that
1
l→∞ l
cap S = lim
logq S ∩ Zlq = cap (0, k − 1)q - RLL
and for all b ∈ S it holds that wtH (b) > θ|b|.
7
Proof: Let Gq (k − 1) be the strongly connected deterministic digraph generating the system (0, k −1)q - RLL which
adjacency matrix is
q−1
q − 1
.
Tq (k − 1) =
..
q − 1
q−1
0
As is well known for the case of q = 2 (see, e.g., [14], [36]),
its characteristic polynomial is
pq(k−1) (x) = xk − (q − 1)
k−1
X
xj =
j=0
xk+1 − qxk + (q − 1)
,
x−1
hence the Perron eigenvalue λ of Tq (k − 1) is the unique
(k−1)
(x) = xk+1 − qxk + (q − 1) greater than
positive root of p̂q
1 (in fact, λ ∈ (q−1, q), which can readily be confirmed either
using elementary calculus or by information-thoretic methods,
∗
since (Zq \ {0}) ⊆ (0, k − 1)q - RLL ⊆ Z∗q ).
Further, Tq (k − 1) has positive right- and left-eigenvectors
associated with λ, which we denote v̄, w̄ respectively; specifically,
v̄ =
1, λ − (q − 1), . . . , λj−1 − (q − 1)
j−2
X
λi , . . . ,
k−2
X
i=0
λi
!
v
+(q−1)
l · P(e) −
e terminates
at first node
ǫ
k(q − 1)
=⇒ λ > 2 − k(q − λ)
k−2
X
= l(π1 − ǫ)
λ+k(q − λ) = q + (k − 1)(q − λ) > q > 2
k−1
X
1 k
i
k−1
λj
λ =
λ − (q − 1)
vk = λ
− (q − 1)
λ
j=1
i=0
q−1
>0
λ
X
To conclude the proof, note
and we may verify that
=
λk−1 − (q − 1)
λ−1
. (Incidentally, it follows from
and in particular π1 = λ−k(q−λ)
q
π1 ∈ (0, 1) that 1 < k(q − λ) < λ, that is, q − k+1
<λ<
1
q − k .)
Next, recall that for a given ǫ > 0, a (P, ǫ)-stronglytypical path in G is a path γ = (e1 , e2 , . . . , el ) (denoted by
its edges {e1 , e2 , . . . , el } ⊆ Eq (k − 1)) such that for each
e ∈ Eq (k − 1) appears in the path l · τ times, for some τ
satisfying |τ − P(e)| 6 ǫ. If we let Sǫ ⊆ Z∗q be the system
ǫ
induced by (P, k(q−1)
)-strongly-typical paths, then it is well
known that cap Sǫ = cap (0, k−1)q - RLL. Note, for b ∈ Sǫ of
length |b| = l, which is generated by the path γ = (e1 , . . . , el ),
wtH (b) is precisely the number of edges which terminate at
ǫ
)-strongly-typical,
the first node; since γ is (P, k(q−1)
wtH (b) >
,
k
X
λk−1 − λk−i
λ−1
i=2
k
q − 1 X k−i
q−1
λ
+
= λk−1 1 + (k − 1) 1 −
λ−1
λ − 1 i=2
k−2
q−1 X j
q−1
k−1
+
λ
k − (k − 1)
=λ
λ−1
λ − 1 j=0
λk − (q − 1)λk−1
q−1
k−1
+
k − (k − 1)
=λ
λ−1
λ−1
k−1
λ
=
[λ − k(q − λ)]
λ−1
π̂i = λk−1 +
i=0
. . . λk−1 − (q − 1)
w̄ = λk−1 , λk−2 , . . . , λk−j , . . . , 1 .
k
X
i=1
··· 0
..
.
0 1
..
..
. 0
.
0 · · · 0 1
0 ··· ··· 0
1
Note, then, that
, hence every entry of v̄ is indeed positive.
and vj = j+1 λ
v
Denoting qi,j = (Tq (k − 1))i,j · λvji , it follows (see, e.g.,
[29][Sec. 3.5]) that Q = (qi,j )16i,j6k is stochastic, and
represents a transition matrix of a stationary Markov chain P
on Gq (k−1) (a probability measure on its edges set Eq (k−1))
satisfying H(P) = logq λ = cap (0, k − 1)q - RLL. Further,
the stationary distribution ofP
the Markov chain, i.e., a positive
π̄ = (π1 , . . . , πk ) such that kj=1 πj = 1 and π̄ T Q = π̄ T , is
π̂
given by πj = Pk j π̂ , where π̂ is defined by π̂j = wj vj . It
P
i=1 i
holds for all j that πj is the sum of probabilities
P(e) of
edges terminating at the jth node.
=⇒ 2(λ − 1) > λ − k(q − λ) =⇒ π1 >
1
2
Hence we can take any 0 < ǫ < π1 − 21 , and observe that
S = Sǫ , θ = π1 − ǫ satisfy the proposition.
Lemma 28 implies that there exists a subset Sk ⊆ Irrk such
that cap Sk = cap Irrk , and for every x ∈ Sk of length |x| =
k + γn we have m(x) > ⌈θ · γn⌉. For the rest of this section
we only build codes Cγn in the descendant cones of roots in
Sk . Note, then, that if we denote mn = ⌈θ · γn⌉ then the
resulting codes have rate
h
R(Cγn ) > lim sup n1 logq |Irrk (k + γn)|·
i
· M (mn , rn , dN,t (mn ))
= γ cap(Irrk )+
+ lim sup n1 logq M (mn , rn , dN,t (mn )) (2)
8
Theorem 29 As before, we denote rn =
⌈θ · γn⌉. Then
1−γ
k n−1
and mn =
1
(a)
0.8
lim n1 logq M (mn , rn , dNn ,tn (mn )) =
n
θγ
1−γ
·H 1+
log2 q
kθγ
(b)
0.6
R
=
in both of the following two regimes:
1) When Nn = o(n) and tn = t is fixed.
PSfrag replacements
2) When Nn = 2αn and tn = βn for constants α, β > 0
2
satisfying αβ > 4θγ (such that Nn , tn ∈ N for some,
hence infinitely many, indices).
0.4
0.2
0
0
Proof:
1) Note, for sufficiently large n, that Nn < θ · γn 6 mn ,
resulting by Lemma 24 in dNn ,t (mn ) = t. We note that
limn→∞ rnn = 1−γ
and limn→∞ mnn = θγ, hence by
k
Theorem 23 the claim is proven when t is fixed.
2) By Theorem 27:
α2 n
dNn ,tn (mn ) 6 max 1, βn − n
4⌈θ · γn⌉
α2 n
n .
= max 1, β −
4⌈θ · γn⌉
α2
β > 4θγ then for sufficiently large
α2 n
4⌈θ·γn⌉ , hence dNn ,tn (mn ) = 1. Since it
n we have β <
is fixed, we may
now apply the same argument used in the previous part.
If
Before moving on to show that generally R(Cγn ) may be
made to exceed cap(Irrk ) by a careful choice of γ, we look
at the following example.
Example 30 Set√q = k = 2. Then the Perron eigenvalue of
T2 (1) is λ = 1+2 5 , and
√ !
1+ 5
cap(Irr2 ) = log2 (λ) = log2
≈ 0.6942.
2
In addition, any θ which is less than π1 = 21 1 + √15 ≈
0.7236 satisfies Lemma 28.
Alternatively, we may set q = 4 (for the special case of
human DNA) and duplication-length k =
√ 2. Now the Perron
3+ 21
eigenvalue of T4 (1) is given by λ = 2 , hence
√ !
3 + 21
≈ 0.9613.
cap(Irr2 ) = log4 (λ) = log4
2
Further,
qwe may choose any θ which is less than π1 =
1
3
2 1+
7 ≈ 0.8273.
n
R(Cγ ) is shown as a function of γ ∈ (0, 1) for both cases
in Figure 1, under the assumptions of asymptotic regime made
in Theorem 29. The figure demonstrates that the capacity of
reconstruction codes (bounded from below by the maximum
of the curve) is greater than cap(Irrk ).
We now attempt to maximize R(Cγn ) by a proper choice of
γ. Theorem 29 motivates the following definition:
0.2
0.4
γ
0.6
0.8
1
Figure 1. Rate R(Cγn ) in the cases (a) q = k = 2, θ = 0.7236, and (b)
q = 4, k = 2, θ = 0.8273. The value at γ = 1 equals cap(Irrk ), which
was previously the best known construction.
1
2, 1
. We define R : (0, 1) → R
θγ
1−γ
.
R(γ) = γR +
H 1+
log2 q
kθγ
Definition 31 Take R, θ ∈
by
Analysis of R(γ) is simpler using the following change of
variable:
Definition 32 Define x : (0, 1) → (0, ∞) by x(γ) =
1−γ
γ .
We observe that x is a decreasing diffeomorphism, and γ =
1
1+x(γ) .
Lemma 33 One has
x(γ)
x(γ)
R(γ) = γR + θγ 1 +
logq 1 +
kθ
kθ
x(γ)
x(γ)
logq
−
kθ
kθ
Proof:
We observe that for all x > 0, log 1 + x1 =
log x+1
= log(x + 1) − log x; in particular
x
1−γ
1−γ
kθγ
= logq 1 +
− logq
logq 1 +
1−γ
kθγ
kθγ
Hence,
θγ
1−γ
R(γ) =γR +
·H 1+
log2 q
kθγ
1−γ
=γR + θγ logq 1 +
kθγ
kθγ
1−γ
logq 1 +
+
k
1−γ
1−γ
1−γ
logq 1 +
=γR + θγ +
k
kθγ
1−γ
1−γ
logq
−
k
kθγ
1−γ
1−γ
=γR + θγ 1 +
logq 1 +
kθγ
kθγ
1−γ
1−γ
−
logq
kθγ
kθγ
9
We can now show that there always exists a choice of γ for
which we get R(Cγn ) > cap(Irrk ):
Theorem 34 maxγ∈(0,1) R(γ) > R.
Proof: Observe that R(γ) is continuously differentiable
and satisfies limγ→0 R(γ) = 0, limγ→1 R(γ) = R. We find
dx d
θ
x
x
′
R (γ) =R +
·
1+
logq 1 +
dγ dx 1 + x
kθ
kθ
x
x
−
logq
kθ
kθ
x=x(γ)
−θ
x
x
1
logq 1 +
1+
=R − 2
γ (1 + x)2
kθ
kθ
x
x
−
logq
kθ
kθ
θ
x
1
+
·
logq 1 +
(1 + x)
kθ
kθ
x
1
logq
−
kθ
kθ
x=x(γ)
x(γ)
x(γ)
1
(kθ − 1) logq 1 +
+ logq
=R +
k
kθ
kθ
Thus, R′ (γ) = 0 if and only if
kθ−1
x(γ)
x(γ)
q −kR = 1 +
·
kθ
kθ
(3)
This equation has a unique solution x0 = x(γ0 ), since the
RHS is a monotonic increasing function of x, vanishing at
x = 0 and unbounded as x grows. Moreover, 0 < x0 < kθ,
since kθ > 1, hence the RHS is greater than 1 at x = kθ.
Thus R(γ) has a unique local extremum in (0, 1).
It now suffices to show that R(γ) is concave. Indeed,
x
1 dx d
′′
(kθ − 1) logq 1 +
·
R (γ) =
k dγ dx
kθ
x
+ logq
kθ x=x(γ)
kθ − 1
1
−1
<0
+
=
k ln(q)γ 2 kθ + x(γ) x(γ)
Proof: For
fixed
x ∈ [0, ∞) define gx : (0, ∞) → R by
gx (y) = y ln 1 + xy . Then
y
x
x
−x
x
+
−
=
ln
1
+
·
gx′ (y) = ln 1 +
x
2
y
1+ y y
y
y+x
x
x
= − ln 1 −
−
x+y
y+x
x
x
−
> 0.
>− −
x+y
y+x
y
satisfies 1 + x =
Therefore, fx (y) = egx (y) = 1 + xy
y
fx (1) 6 fx (y) = 1 + xy
for all y > 1. In our case kθ > 1
and x0 satisfies Equation 3, hence
1 + x0 x0
x0 + x20
x0 kθ−1 x0
>
=
q −Rk = 1 +
x0 ·
kθ
kθ
1 + kθ kθ
kθ + x0
which we simplify to 0 > x20 + 1 − q −Rk x0 − kθq −Rk .
Thus, the first upper bound is proven. For
√ the second, we
a
,
require only that for a, b > 0 it holds that a + b2 − b 6 2b
which is readily shown by differentiation.
On the other hand, Equation 3 implies that x0 6 kθ.
Therefore
x0
x0 kθ−1 x0
2kθ
·
q −Rk = 1 +
6
kθ
kθ
1 + xkθ0 kθ
⇐⇒ kθq −Rk 6 2kθ − q −Rk x0
which proves the lower bound.
Next, we show that we may tighten the bounds we derived
in the previous lemma.
Lemma 36 Let x0 > 0 be the unique solution to Equation 3,
and denote z0 = xkθ0 . If z 6 z0 6 z then F (z) 6 z0 6 F (z),
where
q −Rk
F (z) =
kθ−1 .
q−Rk
1 + (1+z)kθ−1
kθ−1
It follows that R(γ0 ) > limγ→1 R(γ) = R.
Thus, the main result of this paper is established. In what
remains of this section we show that we can analytically find
γ0 which maximizes R(γ). We begin by establishing bounds
in the following lemma.
Proof: By assumption we have q −Rk = (1 + z0 )
·z0 ,
kθ−1
hence q −Rk 6 (1 + z)
· z0 , implying that z0 > G(z)
q−Rk
where G(z) = (1+z)
kθ−1 . Similarly, z0 6 G(z).
The proposition now trivially follows for F (z) = G(G(z)).
Lemma 35 Let γ0 ∈ (0, 1) be the unique maximum of R(γ),
and denote x0 = x(γ0 ). Then
Finally, we can show that x0 may be found by the following
limiting process:
x0 >
and
x0 6
1
2
kθ
k
(2θ q R )
Theorem 37 The unique solution to Equation 3 is given by
x0 = kθ limn→∞ F n (z1 ), for all z1 ∈ [0, 1].
−1
q
2
(1 − q −Rk ) + kθq 2−Rk − 1 − q −Rk
6
kθq 2
.
4(q Rk − 1)
Proof: As before, we denote the unique solution x0 > 0,
x0
.
and take z0 = kθ
Note that Lemma 36 implies that z0 = F (z0 ). We will prove
that F : [0, 1] → [0, 1] is a contraction; that is, for all z1 , z2 ∈
10
[0, 1] we have |F (z1 ) − F (z2 )| 6 c|z1 − z2 | for some c < 1.
Indeed, recalling kθ > 1 we find
F ′ (z) =
6
2−2Rk (kθ − 1)2
kθ
q−Rk
(1 + z)kθ 1 + (1+z)
kθ−1
(k − 1)2
9
(kθ − 1)2
6
6
< 1,
2R
k
k
(2 )
2
16
where the next to last inequality may be directly verified for
all small k.
Having done so, we utilize Banach’s fixed-point theorem to
deduce that F has a unique fixed point (necessarily z0 ), and for
all z1 ∈ [0, 1], defining zn+1 = F (zn ) we get limn→∞ zn =
z0 .
VI. C ONCLUSION
We have proposed that reconstruction codes can be applied
to data-storage in the DNA of living organisms, due to the
channel’s inherent property of data replication.
We have showed, under the assumption of uniform
tandem-duplication noise, that reconstruction codes are errorcorrecting codes with minimal distance dependent on the
reconstruction parameters. We then proved the existence of
such codes with rates surpassing that of know error-correcting
codes. In particular, Theorem 37 allows us to find the parameters required for code-design in any real application, such that
the resulting code has optimal rate.
We believe that further research should focus on explicit
code constructions. It is also desirable to examine the problem under broader noise models, such as bounded tandemduplication, interspersed-duplication (perhaps inversed), as
well as combinations of multiple error models.
R EFERENCES
[1] J. Acharya, H. Das, O. Milenkovic, A. Orlitsky, and S. Pan, “String
reconstruction from substring compositions,” SIAM J. Discrete Math.,
vol. 29, no. 3, pp. 1340–1371, 2015.
[2] N. Alon, J. Bruck, F. F. Hassanazadeh, and S. Jain, “Duplication distance
to the root for binary sequences,” IEEE Transactions on Information
Theory, vol. PP, no. 99, pp. 1–1, 2017.
[3] M. Arita and Y. Ohashi, “Secret signatures inside genomic DNA,”
Biotechnology Progress, vol. 20, no. 5, pp. 1605–1607, 2004.
[4] F. Balado, “Capacity of DNA data embedding under substitution mutations,” IEEE Trans. on Inform. Theory, vol. 59, no. 2, pp. 928–941,
Feb. 2013.
[5] Y. Cassuto and M. Blaum, “Codes for symbol-pair read channels,” IEEE
Transactions on Information Theory, vol. 57, no. 12, pp. 8011–8020, Dec
2011.
[6] G. M. Church, Y. Gao, and S. Kosuri, “Next-generation digital information storage in DNA,” Science, vol. 337, no. 6102, pp. 1628–1628,
2012.
[7] C. T. Clelland, V. Risca, and C. Bancroft, “Hiding messages in DNA
microdots,” Nature, vol. 399, no. 6736, pp. 533–534, 1999.
[8] O. Elishco, F. Farnoud, M. Schwartz, and J. Bruck, “The capacity
of some Pólya string models,” in Proceedings of the 2016 IEEE
International Symposium on Information Theory (ISIT2016), Barcelona,
Spain, July 2016, pp. 270–274.
[9] F. Farnoud, M. Schwartz, and J. Bruck, “A stochastic model for genomic
interspersed duplication,” in Proceedings of the 2015 IEEE International
Symposium on Information Theory (ISIT2015), Hong Kong, China, Jun.
2015, pp. 904–908.
[10] R. Gabrys, H. M. Kiah, and O. Milenkovic, “Asymmetric Lee distance
codes for DNA-based storage,” IEEE Trans. Information Theory, vol. 63,
no. 8, pp. 4982–4995, Aug 2017.
[11] R. Gabrys, E. Yaakobi, and O. Milenkovic, “Codes in the Damerau distance for DNA storage,” in Proceedings of the 2016 IEEE International
Symposium on Information Theory (ISIT2016), Barcelona, Spain, July
2016, pp. 2644–2648.
[12] F. F. Hassanzadeh, M. Schwartz, and J. Bruck, “The capacity of stringduplication systems,” IEEE Trans. on Inform. Theory, vol. 62, no. 2, pp.
811–824, Feb 2016.
[13] D. Heider and A. Barnekow, “DNA-based watermarks using the DNACrypt algorithm,” BMC Bioinformatics, vol. 8, no. 1, pp. 1–10, 2007.
[14] T. D. Howell, “Statistical properties of selected recording codes,” IBM
Journal of Research and Development, vol. 33, no. 1, pp. 60–73, Jan
1989.
[15] S. Jain, F. F. Hassanzadeh, and J. Bruck, “Capacity and expressiveness
of genomic tandem duplication,” IEEE Transactions on Information
Theory, vol. 63, no. 10, pp. 6129–6138, Oct 2017.
[16] S. Jain, F. F. Hassanzadeh, M. Schwartz, and J. Bruck, “Duplicationcorrecting codes for data storage in the DNA of living organisms,” IEEE
Trans. Information Theory, vol. 63, no. 8, pp. 4996–5010, 2017.
[17] ——, “Noise and uncertainty in string-duplication systems,” in Proceedings of the 2017 IEEE International Symposium on Information Theory
(ISIT2017), Aachen, Germany, June 2017, pp. 3120–3124.
[18] D. C. Jupiter, T. A. Ficht, J. Samuel, Q.-M. Qin, and P. de Figueiredo,
“DNA watermarking of infectious agents: Progress and prospects,” PLoS
Pathog, vol. 6, no. 6, p. e1000950, 2010.
[19] H. M. Kiah, G. J. Puleo, and O. Milenkovic, “Codes for DNA sequence
profiles,” IEEE Trans. on Inform. Theory, vol. 62, no. 6, pp. 3125–3146,
June 2016.
[20] E. Konstantinova, “Reconstruction of permutations distorted by single
reversal errors,” Discrete Appl. Math., vol. 155, pp. 2426–2434, 2007.
[21] ——, “On reconstruction of signed permutations distorted by reversal
errors,” Discrete Math., vol. 308, pp. 974–984, 2008.
[22] E. Konstantinova, V. Levenshtein, and J. Siemons, “Reconstruction of
permutations distorted by single transposition errors,” arXiv preprint
arXiv:math/0702191, 2007.
[23] P. Leupold, C. Martín-Vide, and V. Mitrana, “Uniformly bounded
duplication languages,” Discrete Appl. Math., vol. 146, no. 3, pp. 301–
310, 2005.
[24] V. I. Levenshtein, E. Konstantinova, E. Konstantinov, and S. Molodtsov,
“Reconstruction of a graph from 2-vicinities of its vertices,” Discrete
Appl. Math., vol. 156, pp. 1399–1406, 2008.
[25] V. I. Levenshtein and J. Siemons, “Error graphs and the reconstruction
of elements in graphs,” J. Combin. Theory Ser. A, vol. 116, pp. 795–815,
2009.
[26] V. I. Levenshtein, “Efficient reconstruction of sequences,” IEEE
Trans. on Inform. Theory, vol. 47, no. 1, pp. 2–22, Jan 2001.
[27] M. Liss, D. Daubert, K. Brunner, K. Kliche, U. Hammes, A. Leiherer,
and R. Wagner, “Embedding permanent watermarks in synthetic genes,”
PLoS ONE, vol. 7, no. 8, p. e42465, 2012.
[28] F. J. MacWilliams and N. J. A. Sloane, The Theory of Error-Correcting
Codes. North-Holland, 1978.
[29] B. H. Marcus, R. M. Roth, and P. H. Siegel, “An
introduction
to
coding
for
constrained
systems,”
Oct
2001,
unpublished
Lecture
Notes.
[Online].
Available:
ronny.cswp.cs.technion.ac.il/wp-content/uploads/sites/54/2016/05/chapters1-9.pdf
[30] N. Raviv, M. Schwartz, and E. Yaakobi, “Rank modulation codes
for DNA storage,” in Proceedings of the 2017 IEEE International
Symposium on Information Theory (ISIT2017), Aachen, Germany, June
2017, pp. 3125–3129.
[31] S. L. Shipman, J. Nivala, J. D. Macklis, and G. M. Church, “CRISPRCas encoding of a digital movie into the genomes of a population of
living bacteria,” Nature, vol. 547, p. 345, Jul. 2017.
[32] I. Shomorony, T. A. Courtade, and D. Tse, “Fundamental limits of
genome assembly under an adversarial erasure model,” IEEE Transactions on Molecular, Biological and Multi-Scale Communications, vol. 2,
no. 2, pp. 199–208, Dec. 2016.
[33] P. C. Wong, K. k. Wong, and H. Foote, “Organic data memory using the
DNA approach,” Commun. ACM, vol. 46, no. 1, pp. 95–98, Jan. 2003.
[34] E. Yaakobi, J. Bruck, and P. H. Siegel, “Decoding of cyclic codes
over symbol-pair read channels,” in Proceedings of the 2012 IEEE
International Symposium on Information Theory (ISIT2012), Cambridge,
MA, USA, July 2012, pp. 2891–2895.
[35] E. Yaakobi and J. Bruck, “On the uncertainty of information retrieval in
associative memories,” in Proceedings of the 2012 IEEE International
Symposium on Information Theory (ISIT2012), Cambridge, MA, USA,
Jul. 2012, pp. 106–110.
[36] E. Zehavi and J. K. Wolf, “On runlength codes,” IEEE Trans. on
Inform. Theory, vol. 34, no. 1, pp. 45–54, Jan 1988.
| 7cs.IT
|
CROSSED-PRODUCTS BY LOCALLY COMPACT GROUPS:
INTERMEDIATE SUBFACTORS
arXiv:1611.10121v2 [math.OA] 2 Dec 2016
RÉMI BOUTONNET & ARNAUD BROTHIER
Abstract. We study actions of locally compact groups on von Neumann factors and the associated crossed-product von Neumann algebras. In the setting of totally disconnected groups
we provide sufficient conditions on an action G y Q ensuring that the inclusion Q ⊂ Q ⋊ G is
irreducible and that every intermediate subfactor is of the form Q ⋊ H for a closed subgroup
H < G. This partially generalizes a result of Izumi-Longo-Popa [ILP98] and Choda [Ch78].
We moreover show that one can not hope to use their strategy for non-discrete groups.
1. Introduction
In the theory of von Neumann algebras, the crossed-product construction associates to an action
of a locally compact group G on a von Neumann algebra Q a new von Neumann algebra,
denoted by Q ⋊ G, which encodes the action (to some extent). This construction goes back to
Murray and von Neumann [MvN36] in the case of state preserving actions of countable groups
on abelian von Neumann algebras, and was called the group measure-space construction. Thus
crossed-product algebras appear as one of the most basic examples of von Neumann algebras.
In the case of actions of discrete groups, elementary properties of these crossed-product algebras
are quite well understood. For instance:
(A) If Q is abelian, Q = L∞ (X, µ), and if the corresponding non-singular G-action on (X, µ)
is essentially free then Q′ ∩ (Q ⋊ G) ⊂ Q, so that the crossed-product Q ⋊ G is a factor if
and only if the action is ergodic.
(B) If Q is a factor, then Q′ ∩ (Q ⋊ G) = C if and only if the action is properly outer, meaning
that for all g ∈ G, the corresponding automorphism of Q is not inner.
Moreover, in both settings one can completely describe all the intermediate subalgebras between
Q and Q ⋊ G. Hybrid cases combining aspects of cases (A) and (B) above have been considered
recently in [CS16]
In the case of non-discrete groups the picture is not as nice. The main difference is that there
is no “Fourier decomposition” of elements Q ⋊ G. Namely, not every element of x ∈ Q ⋊ G can
be represented as an L2 -element x ∈ L2 (Q) ⊗ L2 (G).
Nevertheless, in the setting (A) above of actions on abelian algebras there are some satisfying
results. Sauvageot showed in [Sa77, Section 2] that the equivalence between Q′ ∩ (Q ⋊ G) ⊂ Q
and the action G y (X, µ) being essentially free still holds. Moreover, in the case of state
preserving actions of unimodular groups, a powerful tool is available: the so-called crossedsection equivalence relation. It relies on the observation that appropriate II1 corners of Q ⋊ G
can be described by an explicit equivalence relation. We refer to [KPV14] for details and
references.
In this article we will be interested in the setting (B) of actions on factors. An action G y Q of
a locally compact group G on a factor Q is called strictly outer if Q′ ∩ (Q ⋊ G) = C. It is known
in this case that properly outer actions need not be strictly outer. As we will see, assuming
The first author was partially supported by a PEPS grant from INSMI and the second author was partially
supported by NSF Grant DMS-1362138.
1
2
RÉMI BOUTONNET & ARNAUD BROTHIER
that the action is strictly outer allows to deduce more conclusions. For instance we will see
that it implies that the normalizer of Q inside Q ⋊ G is the semi-direct product U (Q) ⋊ G, see
Corollary 3.11. In particular we will derive as in the discrete case that for strictly outer actions
the pairs Q ⊂ Q ⋊ G completely characterize the actions up to cocycle conjugacy. We refer to
[Va05, Theorem 5.1] for examples of strictly outer actions. We will also give a new criterion
providing more examples in Proposition 4.4.
Our main goal is to prove an intermediate subfactor theorem. Namely we will provide examples
of strictly outer actions of non-discrete groups that satisfy the following property.
Definition. We will say that a strictly outer action G y Q of a locally compact group on a
factor Q satisfies the Intermediate Subfactor Property if any subfactor of Q ⋊ G, containing Q
is of the form Q ⋊ H for some closed subgroup H < G.
In the case of outer actions of discrete groups, Choda [Ch78] proved such an intermediate
subfactor property under the extra assumption that the intermediate subfactor is the range
of a normal conditional expectation on Q ⋊ G. Then Izumi-Longo-Popa [ILP98] were able to
show that the existence of such a conditional expectation is automatic for crossed-products by
discrete groups. So the result is known for discrete groups. Unfortunately, as we discuss below,
there is no hope to adapt this strategy for general locally compact groups. In fact, the general
situation is analytically much harder to handle precisely because conditional expectations need
not exist in general.
Our main result relies on a very different approach inspired from II1 -factor techniques. We
will restrict our attention to actions of totally disconnected groups. This will allow us to use
Fourier decomposition arguments. Here is the statement.
Theorem A. Consider an action σ : G y Q of a totally disconnected locally compact group G
on a semi-finite factor Q. Assume that σ is properly outer relative to a compact open subgroup
K0 < G whose action is minimal.
Then σ satisfies the Intermediate Subfactor Property.
We refer to Section 2.5 for the definition of a minimal action and to Definition 4.5 for the notion
of relatively properly outer action. Theorem A applies to Bernoulli shifts G y (⊗G/K Q0 )
where Q0 is an arbitrary II1 -factor, and K is any compact open subgroup of G, see Section
4.4. For instance if G is a closed subgroup of automorphisms of a locally finite tree T , and if G
acts transitively on T , then the Bernoulli shift action G y (⊗T Q0 ) satisfies the Intermediate
Subfactor Property. We refer to Section 4.4 for more examples.
As we mentioned above this result is not a simple generalization of the discrete case, even if Q
is tracial, because in general an intermediate subalgebra Q ⊂ N ⊂ Q ⋊ G does not behave well
from the Hilbert theory perspective. Nevertheless we still manage to use Hilbert techniques to
perform the proof.
Our approach relies on two ingredients. The first one is an averaging argument. This is where
the semi-finiteness assumption on Q will be used. As we will see in Remark 4.13, this technique
also allows to deal with some actions on type III factors Q, but we need to make an assumption
that there is a G-invariant state on Q that has a large centralizer.
Our second ingredient is an extension of the notion of support defined by Eymard [Ey64]. In
the language of quantum groups, the support of an element of Q ⋊ G is the spectrum of the dual
action. In the setting of totally disconnected groups, this notion is particularly well suited,
see for instance the proof of Proposition 4.4. The key fact that we will use in the proof of
Theorem A is that an element whose support is contained in some closed subgroup H < G
actually belongs to the subalgebra Q ⋊ H ⊂ Q ⋊ G. This result is certainly known to experts
CROSSED-PRODUCTS BY LOCALLY COMPACT GROUPS
3
in quantum groups, but as we were not able to find an explicit reference, we will provide a self
contained proof in Section 3.
In view of Theorem A and Remark 4.13 we make the following general conjecture.
Conjecture. Any strictly outer action G y Q on a factor (of any type) satisfies the Intermediate Subfactor Property.
Remark. The conjecture holds in full generality for compact groups. Indeed, if K is a compact
group and K y Q is a strictly outer action on an arbitrary von Neumann algebra, then the pair
Q ⊂ Q ⋊ K can be identified with the basic construction of the inclusion QK ⊂ Q. This result
was showed more generally for integrable (strictly outer) actions of locally compact groups by
Vaes, [Va01, Theorem 5.3]. Moreover, by [ILP98, Theorem 3.15], every intermediate subfactor
of the inclusion QK ⊂ Q is of the form QL form some closed subgroup L < K. Combining
these two facts indeed yells that K y Q satisfies the Intermediate Subfactor Property.
Going back to general (not necessarily compact) groups, can one prove the conjecture even
when Q is tracial and G is an arbitrary locally compact group? As we mentioned, Remark 4.13
allows to produce actions on type III factors. But the condition on large centralizers that we
need is never fulfilled in the case of actions on factors of type III0 . This raises the following
question.
Question 1. Can one provide an explicit example of a strictly outer action (of a non-discrete
group) on a type III0 -factor that satisfies the Intermediate Subfactor Property?
We discuss in Section 5 the (im)possibility to solve the above conjecture by generalizing the
work of Izumi-Longo-Popa. Namely, we completely characterize in the case of arbitrary strictly
outer actions G y Q, existence of normal conditional expectation and operator valued weight
for an inclusion Q ⋊ H ⊂ Q ⋊ G, where H is a closed subgroup of G.
Theorem B. Consider a strictly outer action G y Q on an arbitrary von Neumann algebra
Q and take a closed subgroup H < G. Then we have the following characterizations.
(1) There exists a normal faithful semi-finite operator valued weight T ∈ P(Q ⋊ G, Q ⋊ H)
if and only if the modular functions δG and δH coincide on H.
(2) The inclusion Q ⋊ H ⊂ Q ⋊ G is with expectation if and only if H is open inside G.
For both parts in the above theorem, the only if parts follow easily from modular theory.
Namely both conditions are easily seen to imply the existence of an operator valued weight and
conditional expectation, respectively. Our main contribution is to show that they are actually
necessary.
Note that that for non-strictly outer actions there can exist conditional expectations although
H is not open inside G. For instance, consider a product G = G1 × G2 , with the trivial
action G y C. Assume that G2 is second countable and not discrete. Then G1 is not open
inside G but any faithful normal state on LG2 gives rise to a conditional expectation from
LG ≃ LG1 ⊗ LG2 onto LG1 .
On the other hand we do not know whether the above characterization regarding existence of
normal faithful semi-finite operator valued weights holds for arbitrary actions (not necessarily
strictly outer).
Next, we will show that nevertheless the strategy of Izumi-Longo-Popa and Choda can be
applied to some intermediate subfactors, yielding the following result. We will also mention
applications to crossed-products by Hecke pairs of groups, see Corollary 5.7.
Theorem C. Consider an arbitrary strictly outer action G y Q and a compact open subgroup
K < G. Then any subfactor of Q ⋊ G containing Q ⋊ K is of the form Q ⋊ H for some
intermediate subgroup H; K < H < G.
4
RÉMI BOUTONNET & ARNAUD BROTHIER
Going back to the conjecture, let us finally mention another partial result. It is about intermediate subfactors that are globally invariant under the dual action ∆Q : Q ⋊ G → (Q ⋊ G) ⊗ LG.
Namely if Q is a factor and if N is an intermediate subfactor such that ∆Q (N ) ⊂ N ⊗ LG,
then N is of the form Q ⋊ H for some closed subgroup H < G, see for instance [NT79, Chapter
VII.2].
Question 2. Consider a strictly outer action α of a locally compact quantum group (M, ∆)
on a factor N . Can one show that any von Neumann subalgebra Q ⊂ M ⋉α N containing N
is globally invariant under the dual action? We refer to [Va05] for definitions.
We mention that the Remark above also applies in the quantum setting. Namely, Vaes’ result
[Va01, Theorem 5.3] is valid in the general quantum setting, while the correspondence result of
Izumi-Longo-Popa [ILP98, Theorem 3.15] was generalized by Tomatsu to the quantum setting
in [To08].
Acknowledgement. This work was initiated during the trimester Von Neumann algebras held
at the Hausdorff Institute in Spring 2016. Parts of this work are also due to discussions during
the Conference in the memory of Uffe Haagerup held in Copenhagen in June 2016 and during
the Oberwolfach meeting on C∗ -algebras in August 2016. We warmly thank the organizers of
all these events. The second author was invited at the Mathematical Institute of Bordeaux in
September 2016. He gratefully acknowledges the kind hospitality he received.
Mathematically, we are grateful to Sven Raum for interesting discussions and for pointing out
[Ey64] to us regarding the notion of support. We are also indebted to Amaury Freslon for
explaining to us part of the history on quantum groups related to this work and for providing
references. Theorem B were motivated by discussions of the first author with Hiroshi Ando
and Cyril Houdayer. Many thanks to them. We thank Narutaka Ozawa and Yoshimichi Ueda
for helpful comments on a previous version of this preprint.
Contents
1. Introduction
1
2. Preliminaries
4
3. Support and applications
8
4. Crossed-products by actions of totally disconnected groups
13
5. Existence of conditional expectations and operator valued weights
20
References
28
2. Preliminaries
2.1. General notations. In all the article, the letter G refers to a locally compact1 group; mG
denotes a Haar measure on G, and δG the modular function of G. When we consider L2 -spaces
it is always meant with respect to the measure mG . When integrating functions on G, we will
sometimes use the notation ds instead of dmG (s), s ∈ G. The left regular representation of G
on L2 (G) is denoted by λG .
1We follow the French convention according to which the locally compact assumption also contains the
Hausdorff axiom.
CROSSED-PRODUCTS BY LOCALLY COMPACT GROUPS
5
The letter Q refers to an arbitrary von Neumann algebra on which G acts. The action will be
denoted by G y Q, and called generically σ. By an action, we mean an ultraweakly continuous
homomorphism from G into the automorphism group of Q.
Given a von Neumann algebra M represented on a Hilbert space H, M ′ denotes its commutant,
(M )1 its unit ball in the operator norm, U (M ) its unitary group and Aut(M ) its automorphism
group. We denote by S(M ) and P(M ) the set of normal faithful states on M and the set
of normal faithful semi-finite weights (nfs weights for short), respectively. For any
p weight
Φ ∈ P(Q), consider the left ideal nΦ (M ) = {x ∈ M | Φ(x∗ x) < ∞}, on which x 7→ Φ(x∗ x)
defines a norm k · kΦ . We denote by L2 (M, Φ) the Hilbert space completion of nΦ (M ) and we
write ΛΦ the inclusion map ΛΦ : nΦ (M ) → L2 (M, Φ).
2.2. Group actions and crossed-product von Neumann algebras. Let us give the precise
definition of our main object of study. We refer to [Ta03, Chapter X] and [Ha77a, Ha77b] for
details about the facts below.
Definition 2.1. Fix an action G y Q and represent Q on a Hilbert space H. The crossedproduct von Neumann algebra, denoted by Q ⋊ G, is the von Neumann algebra on L2 (G, H)
generated by the operators {π(x) , x ∈ Q} and {us , s ∈ G} defined by the formulae:
(π(x)ξ)(t) = σt−1 (x)ξ(t)
and
(us ξ)(t) = ξ(s−1 t),
for all ξ ∈ L2 (G, H).
For notational simplicity we will often omit the π and identify π(Q) with Q in the above
definition.
Throughout the article we will always assume that Q is standardly represented on H, with
conjugation operator J and positive cone P. In this case, we abuse with notations and denote
again by σ : g ∈ G → σg ∈ U (H) the canonical implementation of the action G y Q, see [Ha75].
Then for all g ∈ G the operator ρG (g) on L2 (G, H) defined as follows lies in the commutant of
Q ⋊ G:
(2.1)
ρG (g)(f )(s) = δG (g)1/2 σg (f (sg)),
for all s ∈ G, f ∈ L2 (G, H).
Denote by K(G, Q) the ∗-algebra of compactly supported, ∗-ultrastrongly continuous functions,
endowed with product and involution given by the formulae
Z
σs (F1 (ts))F2 (s−1 ) ds
and
F1∗ (t) = δG (t)−1 σt−1 (F1 (t−1 )∗ ),
(F1 · F2 )(t) =
G
for all F1 , F2 ∈ K(G, Q), t ∈ G. The algebra K(G, Q) is also a two sided Q-module with actions
(x · F )(t) = σt−1 (x)F (t),
R
for all F ∈ K(G, Q), x ∈ Q, t ∈ G. The map F 7→ G us F (s) ds defines a Q-bimodular
embedding of K(G, Q) into the crossed-product Q ⋊ G. In this way K(G, Q) is viewed as an
ultraweakly dense ∗-subalgebra of Q ⋊ G.
(F · x)(t) = F (t)x
and
2.3. Crossed-products by subgroups. Given a subgroup H of G, one can restrict any
action G y Q to an action H y Q. In the case where H < G is closed inside G, then the von
Neumann subalgebra of Q ⋊ G generated by Q and by the unitaries uh , h ∈ H, is isomorphic
to Q ⋊ H. This can be seen using induced representations, see [Ta03, Chapter X.4]. With the
same tools one can also compute the commutant of Q ⋊ H inside B(L2 (G, H)).
Theorem 2.2 ([NT79], Theorem VII.1.1). The commutant of Q⋊H inside B(L2 (G, H)) is the
von Neumann algebra generated by the commutant of Q ⋊ G and the subalgebra L∞ (H\G) ⊂
L∞ (G) consisting of left-H-invariant functions on G. Equivalently,
Q ⋊ H = (Q ⋊ G) ∩ L∞ (H\G)′ .
6
RÉMI BOUTONNET & ARNAUD BROTHIER
Remark 2.3. Using the J-map, the above theorem also allows to compute the basic construction hQ ⋊ G, Q ⋊ Hi of the inclusion Q ⋊ H ⊂ Q ⋊ G. In fact this description gives the
isomorphism
hQ ⋊ G, Q ⋊ Hi ≃ (L∞ (G/H) ⊗ Q) ⋊ G,
where G acts diagonally.
2.4. Modular theory and operator valued weights. Given a von Neumann algebra M ,
the modular flow of a weight Φ ∈ P(M ) is denoted by σtΦ ∈ Aut(M ), t ∈ R. The centralizer of
Φ in M is the subalgebra of elements fixed by the flow (σtΦ )t , it is denoted by QΦ . If ψ ∈ P(M )
is another weight, (Dψ : DΦ)t , t ∈ R, denotes the Connes Radon-Nikodym derivate as defined
in [Co73, Section 1.2]. We will need the following simple Lemma.
Lemma 2.4. Consider a von Neumann algebra M , two weights Φ, Ψ ∈ P(M ), and an automorphism θ ∈ Aut(M ). Then (DΨ ◦ θ : DΦ ◦ θ)t = θ −1 ((DΨ : DΦ)t ), for all t ∈ R.
Proof. Denote by (ei,j )1≤i,j≤2 the canonical basis of M2 (C). Put M̃ := M ⊗ M2 (C), and define
ϕ ∈ P(M̃ ) by the formula
2
X
xi,j ⊗ ei,j = Φ(x1,1 ) + Ψ(x2,2 ), for all xi,j ∈ M.
ϕ
i,j=1
Note that the weight on M̃ associated in a similar manner to Φ ◦ θ and Ψ ◦ θ is ϕ ◦ (θ ⊗ id).
By the definition of Connes Radon-Nikodym derivative [Co73, Lemme 1.2.2], we have for all
t ∈ R,
(DΨ : DΦ)t ⊗ e2,1 = σtϕ (1 ⊗ e2,1 )
ϕ◦(θ⊗id)
and (DΨ ◦ θ : DΦ ◦ θ)t = σt
The KMS condition, see [Ta03, Chapter VIII.1], implies
for all t ∈ R. The lemma easily follows.
ϕ◦(θ⊗id)
σt
(1 ⊗ e2,1 ).
= (θ ⊗ id)−1 ◦ σtϕ ◦ (θ ⊗ id)
The notions of modular group and Connes Radon-Nikodym derivative have been defined for
normal faithful conditional expectations in [CD75]. We will need the extended definition for
operator valued weights defined by Haagerup [Ha77d]. Let us fix some notations and recall
known facts about operator valued weights. We refer to [Ha77c, Ha77d] for precise definitions
and proofs.
If N ⊂ M is an inclusion of von Neumann algebras, then we denote by P(M, N ) the set of
nfs operator valued weights from M to N . Given T ∈ P(M, N ), we set nT (M ) := {x ∈
M | T (x∗ x) ∈ N }. One can define the composition Φ ◦ T of a weight T ∈ P(M, N ), with
an nfs weight Φ ∈ P(N ); the resulting weight Φ ◦ T is again normal faithful semi-finite.
More generally, it makes sense to compose operator valued weights: if M1 ⊂ M2 ⊂ M3 are
von Neumann algebras and S ∈ P(M3 , M2 ), T ∈ P(M2 , M1 ) then one can naturally define
T ◦ S ∈ P(M3 , M1 ) in such a way that for all Φ ∈ P(M1 ), we have Φ ◦ (T ◦ S) = (Φ ◦ T ) ◦ S.
If T ∈ P(M, N ) and Φ ∈ P(N ), then for all t ∈ R, the modular automorphism σtΦ◦T associated
to the weight Φ ◦ T leaves the von Neumann subalgebra N c := N ′ ∩ M globally invariant, and
its restriction to N c does not depend on the choice of Φ. This restriction is called the modular
flow of T , denoted by σtT . Moreover, if S ∈ P(M, N ) is another operator valued weight then
for all t ∈ R, the Connes Radon-Nikodym derivative (D(Φ ◦ S) : D(Φ ◦ T ))t is an element
of N c that does not depend on Φ. It is then denoted by (DS : DT )t and called the Connes
Radon-Nikodym derivative at time t.
If P(M, N ) is non-empty then P(N ′ , M ′ ) is non-empty. In particular in this case P(M1 , M ) is
non-empty as well, where M1 denotes the Jones basic construction of the inclusion N ⊂ M , see
b ∈ P(M1 , M ) such that E(e)
b
[Jo83]. Better, if N ⊂ M is with expectation E there exists E
= 1,
CROSSED-PRODUCTS BY LOCALLY COMPACT GROUPS
7
b is called the dual
where e is the Jones projection associated with E, see [Ko85, Lemma 3.1]. E
operator valued weight.
Finally, consider an action G y Q. By [Ha77b, Theorem 3.1], there exists a unique operator
valued weight TQ ∈ P(Q ⋊ G, Q) such that for all F ∈ K(G, Q), g ∈ G and x ∈ Q ⋊ G+ ,
(2.2)
TQ (F ∗ F ) = (F ∗ F )(e) and TQ (ug xu∗g ) = δG (g)ug TQ (x)u∗g .
We call it the Plancherel operator valued weight. For any weight Φ ∈ P(Q), one defines the
dual weight Φ ◦ TQ ∈ P(Q ⋊ G). See [Ha77a] for a different construction.
2.5. Outer actions on von Neumann algebras.
Definition 2.5. We say that a group action G y Q on a von Neumann algebra is
• Properly outer if no non-trivial element of G acts on Q by inner automorphism.
• Strictly outer if the relative commutant Q′ ∩ Q ⋊ G is trivial.
• Minimal if it is faithful and the fixed point subalgebra QG is an irreducible subfactor
of Q: (QG )′ ∩ Q = C.
As observed in [Va01, Va05], a strictly outer action has to be properly outer, but the converse
is not true in general.
Lemma 2.6 ([Va01], Proposition 6.2). An action of a compact group is strictly outer if and
only if it is minimal.
Let us record a basic example for later use.
Example 2.7. Consider a diffuse factor Q0 . For any faithful action of a group H on a finite
set X, the corresponding Bernoulli action H y Q⊗X
is strictly outer, hence minimal.
0
In the setting of totally disconnected groups the following lemma will be useful. Its proof
is very much inspired from that of [Va01, Proposition 6.2], with an other input, the Galois
correspondence theorem [ILP98, Theorem 3.15].
Lemma 2.8. Consider a minimal action K0 y Q of a compact group K0 . For any open
subgroup K < K0 , we have (QK )′ ∩ (Q ⋊ K0 ) = LK ⊂ Q ⋊ K.
Proof. Recall that Q is represented on the Hilbert space H and that Q ⋊ K0 is the subalgebra
of B(L2 (K0 )) ⊗ Q generated by LK0 ⊗ 1 and by the operators π(a) ∈ L∞ (K0 ) ⊗ Q, a ∈ Q,
defined by π(a)(g) = σg−1 (a), g ∈ G. In this picture, we have π(QK0 ) = 1 ⊗ QK0 .
Take x ∈ (QK )′ ∩ M . Since x commutes with QK0 and K0 acts minimally, we deduce that
x ∈ B(L2 (K0 )) ⊗ 1. Note also that (Q ⋊ K0 ) ∩ B(L2 (K0 )) ⊗ 1 = LK0 ⊗ 1, so we can write
x = z ⊗ 1 for some z ∈ LK0 .
For all a ∈ QK and all linear functional µ ∈ Q∗ , we have
z(id ⊗ µ)(π(a)) = (id ⊗ µ)(π(a))z.
Note that the functions (id ⊗ µ)(π(a)) ∈ L∞ (K0 ) are left K-invariant. Let us prove that these
functions generate ℓ∞ (K\K0 ) as a von Neumann algebra as a ∈ QK and µ ∈ Q∗ . To see that
it is sufficient to check that they separate points of K\K0 . Take g, h ∈ K0 such that gh−1 ∈
/ K.
K
Then by [ILP98, Theorem 3.15], there exists a ∈ Q such that σgh−1 (a) 6= a. In particular
π(a)(g) = σg−1 (a) and π(a)(h) = σh−1 (a) are distinct elements of Q. So we can find a linear
functional µ ∈ Q∗ that separates them.
So we arrived at the conclusion that z commutes with ℓ∞ (K\K0 ), and in particular with P (K),
the orthogonal projection onto L2 (K). This means that z leaves L2 (K) invariant : z ∈ LK.
8
RÉMI BOUTONNET & ARNAUD BROTHIER
2.6. Fourier algebra, dual action, multipliers. For a locally compact group G, denote by
A(G) its Fourier algebra as introduced by Eymard [Ey64]. By definition A(G) is the set of
functions on G of the form ξ ∗ η̃, ξ, η ∈ L2 (G), where ∗ denotes the convolution product and η̃
is the function g 7→ η(g −1 ). Note that for all g ∈ G, we have the equality (ξ ∗ η̃)(g) = hλ(g)ξ, ηi.
This set is an algebra under the pointwise multiplication.
The norm of φ ∈ A(G) is defined to be the minimal value of kξkkηk as the functions ξ, η ∈ L2 (G)
satisfy φ = ξ ∗ η̃. With this norm, A(G) is a Banach algebra. The set of compactly supported
continuous functions on G is contained (densely) inside A(G). As a Banach space, A(G) is
isometric to the predual (LG)∗ of LG. The duality pairing is given by the well defined formula
< x, φ >:= hxξ, ηi, for all x ∈ LG, and φ = ξ ∗ η̃ ∈ A(G). We will abuse with notations and
write φ(x) for x ∈ LG to mean < x, φ >. This notation is somewhat consistent with the fact
that φ is a function on G, namely we have: φ(ug ) = φ(g) for all g ∈ G.
Using the product on A(G) we can define multipliers on LG. More precisely, any φ ∈ A(G)
gives rise to a normal completely bounded map mφ : LG → LG, defined by the formula:
< mφ (x), ψ >:=< x, ψφ >
for all x ∈ LG, ψ ∈ A(G).
More generally, for any action G y Q, from an element φ ∈ A(G) one can construct a multiplier
mφ on M := Q ⋊ G in the following way. Consider the unitary operator W on L2 (G × G, H)
such that (W ξ)(g, h) = ξ(g, gh) for all ξ ∈ L2 (G×G, H), g, h ∈ G. Denote by ∆ := Ad(W ∗ ) the
associated automorphism of B(L2 (G × G, H)). We can identify L2 (G × G, H) with L2 (G, H) ⊗
L2 (G) in such a way that
∆(M ⊗ 1) ⊂ M ⊗ LG, and the restriction ∆|L∞ (G)⊗1 = id.
Definition 2.9. With the above notations, the Fourier multiplier associated with an element
φ ∈ A(G) is the normal completely bounded map mφ : Q ⋊ G → Q ⋊ G defined by the formula
mφ (x) = (id ⊗ φ) ◦ ∆(x ⊗ 1),
x ∈ Q ⋊ G.
In practice, the multiplier mφ is characterized by the formula mφ (aug ) = φ(g)aug for all a ∈ Q,
g ∈ G. In this way, one easily checks that in the case where Q = C, the two constructions of
multipliers coincide.
3. Support and applications
In this section we give generalities about the spectrum of the dual action, defined for instance
in [NT79, Chapter IV.1]. We adopt the point of view of Eymard [Ey64], and rather talk about
“support” because we believe it is more transparent for the reader who is familiar with actions
of discrete groups, and not so much with the quantum group language. Our goal is to prove
Theorem 3.7, regarding elements whose support is contained in a subgroup. This is certainly
known to experts but we were not able to find an explicit reference, although it is used in
[NT79, Theorem VII.2.1]. For convenience we tried to keep this section self-contained.
3.1. Definition and first properties. Let us fix an arbitrary action G y Q on a von Neumann algebra.
Definition 3.1. The support of an element x ∈ Q ⋊ G, denoted by supp(x), is the set of
elements g ∈ G such that for all φ ∈ A(G) satisfying φ(g) 6= 0, we have mφ (x) 6= 0.
We also describe the support explicitly in terms of interactions between Q ⋊ G and the copy
of L∞ (G) inside B(L2 (G, H)). If Ω ⊂ G is a measurable subset, we write P (Ω) ∈ L∞ (G) ⊗ 1H
for the orthogonal projection from L2 (G, H) onto L2 (Ω, H).
Proposition 3.2. Take x ∈ Q ⋊ G and g ∈ G. The following are equivalent:
CROSSED-PRODUCTS BY LOCALLY COMPACT GROUPS
9
(i) g ∈ supp(x);
(ii) For any non-empty open set Ω ⊂ G we have P (gΩ)xP (Ω) 6= 0.
Proof. (i) ⇒ (ii). Consider a non-empty open set Ω ⊂ G. Take a non-empty open subset Ω0 ⊂ Ω
and an open neighborhood V of g in G such that V Ω0 ⊂ gΩ. Fix a function φ ∈ A(G) supported
on V such that φ(g) 6= 0. Then for all y ∈ LG, we have mφ (y)P (Ω0 ) = P (gΩ)mφ (y)P (Ω0 ).
Indeed the formula is clear for all y of the form auh , a ∈ Q, h ∈ G, and follows for arbitrary y
by linearity and density.
Note that (Q⋊G)′ L2 (Ω0 , H) spans a dense subset of L2 (G, H). Indeed, for all g ∈ G, we see that
ρG (g)L2 (Ω0 , H) = L2 (Ω0 g, H), where ρG (g) is defined in (2.1). If follows that mφ (x)P (Ω0 ) 6= 0.
Extend arbitrarily φ ∈ A(G) = (LG)∗ to a linear functional φ̃ on B(L2 (G)). With the notations
from Section 2.6, since W commutes with L∞ (G) ⊗ 1, we have
(id ⊗ φ̃) ◦ Ad(W ∗ )(P (gΩ)xP (Ω0 ) ⊗ 1) = P (gΩ)(id ⊗ φ) ◦ Ad(W ∗ )(x ⊗ 1)P (Ω0 )
= P (gΩ)mφ (x)P (Ω0 )
= mφ (x)P (Ω0 ) 6= 0.
We deduce that P (gΩ)xP (Ω0 ) 6= 0, and in particular P (gΩ)xP (Ω) 6= 0.
(ii) ⇒ (i). Fix φ ∈ A(G) such that φ(g) 6= 0. Pick an open neighborhood V of g on which φ
does not vanish. Take ψ ∈ A(G) such that ψ(h) = 1/φ(h) for all h ∈ V , and put φ′ := φψ.
In particular we have φ′ (h) = 1 for all h ∈ V , and it suffices to show that mφ′ (x) 6= 0, since
m φ′ = m ψ ◦ m φ .
Pick a non-empty open set Ω ⊂ G such that gΩ · Ω−1 ⊂ V . We claim that for all y ∈ Q ⋊ G
P (gΩ)mφ′ (y)P (Ω) = P (gΩ)yP (Ω).
By linearity and density, it suffices to check this formula for all y of the form auh , a ∈ Q,
/ V , and
h ∈ G. If h ∈ V , then mφ′ (auh ) = auh for all a ∈ Q, so the formula is obvious. If h ∈
y = auh for some a ∈ Q, then both sides of the formula are equal to 0. Indeed, since mφ′ is a
multiplier, the two terms are scalar multiple of each other, and it suffices to check vanishing of
the right-hand side. Since gΩ · Ω−1 ⊂ V and h ∈
/ V , we deduce that gΩ ∩ hΩ = ∅. This leads to
P (gΩ)auh P (Ω) = P (gΩ)P (hΩ)auh = 0,
as wanted. This proves the claimed equality and hence mφ′ (y) 6= 0.
In the sequel it will sometimes appear that one of the above two descriptions will be better
suited to work with than the other. We will freely switch between these two points of view to
reach the simplest arguments.
Let us record a few properties of the support.
Lemma 3.3. Take x ∈ Q ⋊ G. The following assertions are true.
a) The support of x is a closed subset of G;
b) If x belongs to K(G, Q), then supp(x) coincides with its support as a function on G;
c) If x = ug for some g ∈ G, then supp(x) = {g}.
Proof. a) Consider a net (gi )i of elements in supp(x) that converges to some g ∈ G, and take
φ ∈ A(G) such that φ(g) 6= 0. Since φ is continuous, for i large enough we also have φ(gi ) 6= 0.
Hence mφ (x) 6= 0, as desired.
b) If x is a continuous function in K(G, Q) and ψ ∈ A(G) observe that mψ (x) is the function
in K(G, Q) defined by g 7→ ψ(g)x(g). The statement easily follows.
c) If x = ug , for all φ ∈ A(G) we have mφ (x) = φ(g)x. The result is then obvious.
10
RÉMI BOUTONNET & ARNAUD BROTHIER
Lemma 3.4. Consider x ∈ Q ⋊ G and some open subsets Ω1 , Ω2 ⊂ G satisfying the relation
Ω1 · Ω−1
2 ∩ supp(x) = ∅. Then
P (Ω1 )xP (Ω2 ) = 0.
In particular, for any open set Ω ⊂ G, we have xP (Ω) = P (supp(x)Ω)xP (Ω).
Proof. We first treat the special case where Ω1 · Ω−1
2 ∩ supp(x) = ∅ and Ω2 is compact.
For any (g, h) ∈ Ω1 ×Ω2 , there exists a non-empty open set Vg,h such that P (gh−1 Vg,h )xP (Vg,h ) =
0. Conjugating this equation with ρG (k) for some k ∈ G we can assume that Vg,h is an open
neighborhood of h. Here, recall that ρG is the right action defined in (2.1), which satisfies
ρG (k)P (Vg,h )ρG (k−1 ) = P (Vg,h · k−1 ).
By compactness, for a fixed g ∈ G there exists n > 1, h1 , · · · , hn ∈ Ω2 such that {Vg,hi }ni=1 is a
finite open cover of Ω2 . We can then define an open neighborhood Wg of e ∈ G:
Wg :=
n
\
h−1
i Vg,hi
i=1
We get that P (gWg )xP (Vg,hi ) = 0 for any 1 6 i 6 n, g ∈ Ω1 . Since P (Ω2 ) is smaller than the
supremum of the projections ∨ni=1 P (Vg,hi ), we obtain that P (gWg )xP (Ω2 ) = 0 for any g ∈ Ω1 .
Therefore, P (Ω1 )xP (Ω2 ) = 0 since ∨g∈Ω1 P (gWg ) > P (Ω1 ).
Suppose now that Ω2 is open and non-necessarily relatively compact. For any g ∈ Ω2 , there
exists a compact neighborhood Kg of g such that Ω1 · Kg−1 ∩ supp(x) = ∅. We have that
P (Ω1 )xP (Kg ) = 0 for any g ∈ G, by the proof of above. This proves the desired equality
P (Ω1 )xP (Ω2 ) = 0 because P (Ω2 ) ≤ ∨g∈Ω2 P (Kg ).
The second part of the statement follows from taking Ω2 = Ω, Ω1 = G \ supp(x)Ω.
Before mentioning interesting consequences of this lemma, let us give an essentially equivalent
form involving the multipliers.
Lemma 3.5. Consider x ∈ Q ⋊ G with compact support, and take a function φ ∈ A(G) which
is equal to 1 on a neighborhood of supp(x). Then x = mφ (x).
Proof. Take a non-empty open set Ω ⊂ G such that φ is equal to 1 on supp(x)Ω·Ω−1 . Proceeding
as in the proof of Proposition 3.2, (ii) ⇒ (i), one checks that
P (supp(x)Ω)xP (Ω) = P (supp(x)Ω)mφ (x)P (Ω).
Since supp(mφ (x)) ⊂ supp(x), Lemma 3.4 and the above equality imply that
xP (Ω) = mφ (x)P (Ω).
Moreover, for all g ∈ G, the set Ωg satisfies the same condition as Ω. So the same equality
holds with Ωg in the place of Ω, for all g ∈ G: xP (Ωg) = xP (Ωg). Since Ω is open inside G,
we get that 1 = ∨g P (Ωg). The equality x = mφ (x) follows.
Corollary 3.6. Consider x, y ∈ Q ⋊ G. The following assertions are true.
a)
b)
c)
d)
Adjoint: supp(x∗ ) = supp(x)−1 ;
Sum: supp(x + y) ⊂ supp(x) ∪ supp(y);
Product: supp(xy) ⊂ supp(x) · supp(y).
Vanishing criterion: If supp(x) = ∅, then x = 0.
CROSSED-PRODUCTS BY LOCALLY COMPACT GROUPS
11
Proof. a) This results from the fact that for all φ ∈ A(G), (mφ (x))∗ = mφ∗ (x∗ ), where φ∗ ∈
A(G) is defined by φ∗ (g) = φ(g −1 ), g ∈ G.
b) Consider g in the complementary of supp(x) ∪ supp(y). Then there exists φ1 and φ2 such
that φi (g) 6= 0, i = 1, 2, while mφ1 (x) = 0 and mφ2 (y) = 0. In particular the product φ := φ1 φ2
satisfies φ(g) 6= 0, while mφ (x) = mφ2 ◦ mφ1 (x) = 0 and mφ (y) = mφ1 ◦ mφ2 (y) = 0. In
summary mφ (x + y) = 0 and hence g ∈
/ supp(x + y).
c) Consider an element g ∈ G which is not in supp(x) · supp(y) and take open neighborhoods
of the identity e ∈ G, Ω and V , such that supp(x) · supp(y) ∩ gΩV Ω−1 = ∅. In particular, for all
h ∈ supp(x)−1 gΩ and k ∈ supp(y)Ω we have h−1 k ∈
/ V . Further, the closures X := supp(x∗ )gΩ
and Y := supp(y)Ω do not intersect. Besides, Lemma 3.4 implies that
• yP (Ω) = P (Y )yP (Ω);
• x∗ P (gΩ) = P (X)x∗ P (gΩ).
Altogether, the following equality shows that g ∈
/ supp(xy):
P (gΩ)xyP (Ω) = P (gΩ)xP (X)P (Y )yP (Ω) = P (gΩ)xP (X ∩ Y )yP (Ω) = 0.
d) If supp(x) = ∅, then Lemma 3.4 implies that x = P (G)xP (G) = 0.
Theorem 3.7. Consider an action of a locally compact group G y Q on an arbitrary von
Neumann algebra and take a closed subgroup H < G.
Then an element x ∈ Q ⋊ G belongs to the subalgebra Q ⋊ H if and only if its support is
contained in H.
Proof. First, assume that x ∈ Q ⋊ H. For all g ∈ G \ H we can find a non-empty open set
Ω ⊂ G such that gΩ · Ω−1 ∩ H = ∅. By Lemma 3.4, we deduce that P (gΩ)xP (Ω) = 0, so that
g∈
/ supp(x). Hence supp(x) ⊂ H.
Conversely assume that x ∈ Q ⋊ G is an element with support contained in H. In order to show
that x ∈ Q ⋊ H, we will use Theorem 2.2. This reduces our task to check that x commutes
inside B(L2 (G, H)) with the subalgebra L∞ (H\G) ⊂ L∞ (G) consisting of left H-invariant
functions on G.
Given a left H-invariant open set Ω whose boundary has measure 0, we have equality of the
projections P (Ω) = P (HΩ). By Lemma 3.4, we get xP (Ω) = P (Ω)xP (Ω). Since x∗ also has
support in H we moreover have x∗ P (Ω) = P (Ω)x∗ P (Ω). This shows that x commutes with
P (Ω) ∈ L∞ (H\G).
Claim. The set of functions P (Ω) with Ω as above generates L∞ (H\G).
Denote by q : G → H\G the canonical projection and by µ a quasi-invariant measure on the
coset space H\G. The map f ∈ L∞ (H\G, µ) 7→ f ◦ q ∈ L∞ (H\G) is a normal isomorphism.
With this identification, the indicator function 1U of a set U ⊂ H\G is identified with the
indicator function 1q−1 (U ) = P (q −1 (U )) ∈ L∞ (H\G).
Moreover, for all open subset U ⊂ H\G such that µ(∂U ) = 0, the set Ω := q −1 (U ) ⊂ G is
open, H-invariant and its boundary has Haar measure 0. So we are left to check that the span
of functions 1U for U ⊂ H\G such that µ(∂U ) = 0, is ultraweakly dense in L∞ (H\G). This is
a classical fact about Borel measures on locally compact spaces (see e.g. [Ey64, Proof of (3.33)]
and the references therein).
We deduce from the claim that x commutes with L∞ (H\G). So as wanted, we can use Theorem
2.2 to deduce that x ∈ Q ⋊ H.
12
RÉMI BOUTONNET & ARNAUD BROTHIER
The special case where H is the trivial subgroup of G, the above theorem yields Beurling’s
Theorem, [Ey64, Théorème 4.9]. In particular we stress the following corollary that we will use
several times.
Corollary 3.8. Given an element x in a crossed product von Neumann algebra Q ⋊ G, if the
support of x is a singleton {g}, then there exists a ∈ Q such that x = aug .
Proof. If supp(x) = {g}, then supp(u∗g x) = {e}. In particular u∗g x ∈ Q, by Theorem 3.7.
3.2. Applications. Before moving on to the proof of our main theorems, let us mention a few
classical results that follow easily from the above properties of the support.
Our first application concerns a generalization of the so-called Eymard-Steinespring-Tatsuuma’s
Theorem, [Ta03, Theorem VII.3.9]. We start with analyzing how the support behaves under
the co-product map.
Given an action σ : G y Q, put M = Q ⋊ G and consider the notations W ∈ B(L2 (G × G, H))
and ∆ = Ad(W ∗ ) : M ⊗ 1 → M ⊗ LG introduced in Section 2.6. We will view the algebra
M ⊗ LG as the crossed-product von Neumann algebra associated with the action σ × id :
G × G y Q:
M ⊗ LG ≃ Q ⋊ (G × G).
In this way it makes sense to talk about the support of an element inside M ⊗ LG.
Lemma 3.9. With the above notations, the following fact hold.
a) For all x ∈ M , the support ∆(x ⊗ 1) is equal to {(g, g) , g ∈ supp(x)};
b) For all a ∈ M and b ∈ LG, supp(a ⊗ b) = supp(a) × supp(b).
Proof. a) For all φ, ψ ∈ A(G) the function φ × ψ : (g, h) 7→ φ(g)ψ(h) belongs to the Fourier
algebra A(G × G) and one easily checks the formula
(3.1)
mφ×ψ (∆(x ⊗ 1)) = ∆(mφ·ψ (x) ⊗ 1).
In particular if g 6= h ∈ G, we can find φ, ψ ∈ A(G) such that (φ × ψ)(g, h) = φ(g)ψ(h) 6=
0, but such that φ · ψ = 0 (just take functions with disjoint supports). By (3.1), we get
mφ×ψ (∆(x ⊗ 1)) = 0 and hence (g, h) does not belong to the support of ∆(x ⊗ 1). We thus
deduce that the support of ∆(x ⊗ 1) is contained in the diagonal of G × G. Further, (3.1) easily
implies that if (g, g) belongs to the support of ∆(x ⊗ 1) then g ∈ supp(x).
Conversely, take g ∈ supp(x), and ρ ∈ A(G × G) such that ρ(g, g) 6= 0. We show that
mρ (∆(x ⊗ 1)) 6= 0. Since ρ is continuous, it does not vanish on some open neighborhood U × U
of (g, g). Multiplying ρ by a local inverse if necessary, we may assume that ρ is actually equal
to 1 on U × U . Pick any function φ ∈ A(G) supported on U and such that φ(g) 6= 0. We
deduce from (3.1) that
mφ×φ (mρ (∆(x ⊗ 1))) = m(φ×φ)·ρ (∆(x ⊗ 1)) = mφ×φ (∆(x ⊗ 1)) = ∆(mφ2 (x) ⊗ 1).
Since g ∈ supp(x) and φ2 (g) 6= 0, the above term is non-zero and hence mρ (∆(x ⊗ 1))) 6= 0.
b) Take (g, h) ∈ supp(a⊗ b), and φ, ψ ∈ A(G) such that φ(g), ψ(h) 6= 0. Then (φ× ψ)(g, h) 6= 0,
and we get
mφ (a) ⊗ mψ (b) = mφ×ψ (a ⊗ b) 6= 0.
We deduce that g ∈ supp(a) and h ∈ supp(b).
The converse inclusion can be treated by using a local inverse as in a). Alternatively, we
can use the other description of the support as follows. Take g ∈ supp(a), h ∈ supp(b) and
take a non-empty open set Ω ⊂ G × G. We find non-empty open sets U, V ⊂ G such that
U × V ⊂ Ω. By assumption we know that P (gU )aP (U ) 6= 0 and P (hV )bP (V ) 6= 0, and hence
CROSSED-PRODUCTS BY LOCALLY COMPACT GROUPS
13
P (gU × hV )(a ⊗ b)P (U × V ) 6= 0. It then clearly follows that P ((g, h) · Ω)(a ⊗ b)P (Ω) 6= 0.
This proves that (g, h) ∈ supp(a ⊗ b).
We can now prove the following well known generalization of [Ta03, Theorem VII.3.9]. This
theorem initially deals with the case of group algebras LG, that is, with the case of trivial
actions G y C. The proof given in [Ta03] is rather involved, while an elementary proof relying
(implicitly) on the support already appears in [Ey64, Théorème 3.34]. In the case of general
actions G y Q, the predual (Q ⋊ G)∗ is not identified with an algebra in general, and hence
the notion of character doesn’t apply anymore. Nevertheless we can still provide an easy proof
relying on the notion of support.
Corollary 3.10. Given any action G y Q, put M = Q⋊G and denote by ∆ : M ⊗1 → M ⊗LG
the co-product map defined above. Assume that x ∈ M is an element such that ∆(x ⊗ 1) is of
the form a ⊗ b for some elements a ∈ M and b ∈ LG.
Then there exists g ∈ G and y ∈ Q such that x = yug .
Proof. We may assume that x 6= 0. By Lemma 3.9, the equality ∆(x ⊗ 1) = a ⊗ b implies that
supp(a) × supp(b) is contained in the diagonal of G × G. The only way this can happen is if
supp(∆(x ⊗ 1)) is a singleton {(g, g)}. In this case, supp(x) = {g} and the result follows from
Corollary 3.8.
As an immediate corollary, we deduce the following result.
Corollary 3.11. Given any strictly outer action G y Q, the normalizer of Q inside the
crossed-product M := Q ⋊ G is equal to {aug , g ∈ G, a ∈ U (Q)}.
In particular, two strictly outer actions αi : Gi y Qi , i = 1, 2 of locally compact groups are
cocycle conjugate if and only if the pairs Q1 ⊂ Q1 ⋊ G1 and Q2 ⋊ Q2 ⋊ G2 are isomorphic.
Proof. Take u ∈ NM (Q). Use the notation ∆ : M ⊗ 1 → M ⊗ LG as above. We have that ∆
is the identity map on Q ⊗ 1. Hence one easily checks that (u∗ ⊗ 1)∆(u ⊗ 1) commutes with
Q ⊗ 1. In other words,
(u∗ ⊗ 1)∆(u ⊗ 1) ∈ (Q′ ∩ M ) ⊗ LG = 1 ⊗ LG.
So there exists b ∈ LG such that ∆(u ⊗ 1) = u ⊗ b, and we conclude by Corollary 3.10 that
u = aug for some a ∈ U (Q) and g ∈ G.
The second part of the statement is routine. The only if part is always true, even when the
actions are not strictly outer, see [Ta73, Corollary 3.6]. The if part follows from adapting
[BNP07, Proposition 1.7] to the case of actions of general locally compact groups.
4. Crossed-products by actions of totally disconnected groups
4.1. Notations and tools. There are several advantages in working with totally disconnected
groups. Fix a locally compact group G and an action G y Q on a von Neumann algebra and
put M = Q ⋊ G.
Firstly, given a compact open subgroup K < G, one can define a projection pK := 1K /mG (K).
Note that the net of projections (pK )K increases to 1 as the compact open subgroups K decrease
to {e}. Let us mention two elementary properties of these projections.
Lemma 4.1. Fix a weight Φ ∈ P(Q) and denote by Ψ ∈ P(M ) the corresponding dual
weight.pConsider a compact open subgroup K < G. Then for all a ∈ QK we have kapK kΨ =
kakΦ / mG (K).
Proof. This is an immediate consequence of (2.2).
14
RÉMI BOUTONNET & ARNAUD BROTHIER
Lemma 4.2. Given a compact open subgroup K < G, the map a ∈ QK 7→ apK ∈ pK (Q⋊K)pK
is an onto isomorphism of von Neumann algebras.
Proof. This is a particular case of Proposition 5.6. We give a complete proof of this simpler
case for the convenience of the reader. Since pK commutes with QK , it is clear that the map
is a normal ∗-morphism. By the computations given in Lemma 4.1 above, we see that it is
moreover injective. To check that it is onto, we only need to prove that it has dense image.
For all a ∈ Q and g ∈ K, we have pK (aug )pK = pK apK = E(a)pK , where E : Q → QK is
the conditional expectation. Hence pK (aug )pK belongs to the range of our map, proving the
lemma.
Secondly, for all open subgroup K < G and any action G y Q, there always exists a faithful
normal conditional expectation EK : Q ⋊ G → Q ⋊ K such that EK (ug ) = 1K (g) for all g ∈ G.
In the case where K is compact open, one sees that the multiplier mφ associated with the
function φ := 1K ∈ A(G) gives the desired expectation. In the general case of open subgroups
φ does not necessarily belong to A(G) but it is still positive definite and one can use [Ha77b,
Theorem 3.1.a] to construct the associated multiplier mφ .
An alternative way to construct the expectation EK is by considering modular flows as in
[HR16]. With this point of view, it becomes obvious that EK preserves the Plancherel operator
valued weight: TQ ◦ EK = TQ .
When G is totally disconnected, one checks that the support of an element x ∈ M is described
as follows:
(4.1)
supp(x) = {g ∈ G : EK (u∗g x) 6= 0 for all compact open subgroups K < G}.
Notation. Given a compact open subgroup K of a group G and a set S ⊂ G/K, we denote
by lift(S) a set of representatives of S inside G. This means that for all g ∈ lift(S) we have
gK ∈ S and for all s ∈ S there exists a unique element g ∈ lift(S) such that gK = s.
Lemma 4.3. Consider a compact open subgroup K < G. Given an element x ∈ M with
compact support, the map g 7→ ug EK (u∗g x) is right K-invariant on G and compactly supported.
Moreover,
X
x=
ug EK (u∗g x).
g∈lift(G/K)
Proof. Fix S ⊂ G/K a finite set such that the support of x is contained in lift(S) · K. The
function φ = 1lift(S)K ∈ A(G) is equal to 1 on a neighborhood of the support of x, so by Lemma
P
3.5, we have that x = mφ (x). Moreover, φ can be decomposed as φ = g∈lift(S) 1K (g−1 ·). One
easily checks that the corresponding multiplier satisfies
X
mφ =
ug EK (u∗g ·).
g∈lift(S)
We leave the rest of the proof to the reader.
Although we will not use this fact, let us mention that for a general element x ∈ M , a Kdecomposition as above still makes sense. In this case the sum that appears is infinite but it
converges in the Bures topology associated with the inclusion Q ⋊ K ⊂ Q ⋊ G, with expectation
EK . We refer to [CS16, Section 2] or to the original book [Bu71] for the definition of the Bures
topology.
CROSSED-PRODUCTS BY LOCALLY COMPACT GROUPS
15
4.2. Strictly outer actions. The following proposition combines the ”Fourier coefficient approach” used in the setting of discrete groups with Lemma 2.6 about actions of compact groups.
Proposition 4.4. Consider a properly outer action of a totally disconnected locally compact
group σ : G y Q. The action σ is strictly outer if and only if G admits a compact open
subgroup that acts minimally.
In that case any unitary u ∈ Q ⋊ G normalizing Q is of the form aug for some a ∈ Q, g ∈ G.
Proof. Assume that σ : G y Q is strictly outer. Then one easily checks from the definitions
that σ is properly outer. Moreover, the restriction of σ to any compact subgroup is strictly
outer and thus minimal by [Va01, Proposition 6.2].
Conversely, assume that σ : G y Q is properly outer and admits a compact open subgroup K0
that acts minimally. Note that any subgroup K < K0 acts minimally as well, and hence in a
strictly outer way.
Put M = Q ⋊ G and take a non-zero element x ∈ Q′ ∩ M . By Proposition 3.8 we only need to
show that the support of x is the singleton {e}. The support of x is non-empty by Corollary
3.6.d. Take g ∈ supp(x). For all compact open subgroups K < K0 , put aK := EK (u∗g x) ∈
Q ⋊ K. Since g ∈ supp(x), these elements aK are non-zero.
By minimality, for all K ′ < K < K0 , we have
(4.2)
a∗K ′ aK ∈ Q′ ∩ (Q ⋊ K) = C.
When K = K ′ this relation tells us that aK is a (non-zero) multiple of a unitary element. Then
(4.2) further implies that all these unitaries are proportional to each other. In particular, for
all K < K0 , aK0 is proportional to aK ∈ Q ⋊ K.
We conclude that aK0 is contained in Q ⋊ K, for all K < K0 . Hence its support is equal to
{e}, implying that aK0 ∈ Q. Moreover, aK0 satisfies
aK0 x = σg−1 (x)aK0 ,
for all x ∈ Q.
Since the G-action is properly outer, this gives g = e, as desired. The statement on the
normalizer then follows from Corollary 3.11 (although this could be checked directly by similar
computations on the support).
4.3. Intermediate subfactors. We now turn to the question of determining all intermediate
subfactors Q ⊂ N ⊂ Q ⋊ G. In order to establish our main result Theorem A, we will need to
be able to compute relative commutants of the form (QK )′ ∩ Q ⋊ G for small compact open
subgroups K < G. This forces us to strengthen our assumptions on the action.
Definition 4.5. Given a subgroup K of G we will say that the action G y Q is properly outer
relative to K if the following holds. The only elements g ∈ G for which there exists a non-zero
a ∈ Q such that σg (x)a = ax for all x ∈ QK are the elements of K: g ∈ K.
Note that an action is properly outer if and only if it is properly outer relative to the trivial
subgroup. We will provide examples of relatively properly outer actions in the next section.
Lemma 4.6. Consider an action σ : G y Q of a totally disconnected group G which is properly
outer relative to a compact open subgroup K < G. Put M = Q ⋊ G. We have
(QK )′ ∩ M ⊂ Q ⋊ K.
Proof. Take x ∈ (QK )′ ∩M . We show that supp(x) ⊂ K. Fix g ∈ supp(x), so that EK (u∗g x) 6= 0.
Since the net (pL )L converges strongly to 1 when L ց {e}, we can find a compact open subgroup
L0 < G such that pL0 EK (u∗g x)pL0 6= 0.
16
RÉMI BOUTONNET & ARNAUD BROTHIER
Define L1 = L0 ∩ K ∩ gKg−1 and L := ∩h∈K hL1 h−1 . Since L1 is open, it has finite index
inside K, and hence the intersection defining L is in fact finite. We conclude that L < K is
an open subgroup which is moreover contained in L0 ∩ K ∩ gKg−1 . In particular pL ≥ pL0 , so
that pL EK (u∗g x)pL 6= 0. Since L is normal inside K, the projection pL commutes with uh for
all h ∈ K. By Lemma 4.3, we have
X
X
pL EK (u∗g x)pL =
pL uh EL (u∗gh x)pL =
uh pL EL (u∗gh x)pL .
h∈lift(K/L)
h∈lift(K/L)
The non-vanishing of this sum implies that there exists h ∈ K such that X := pL EL (u∗gh x)pL
is non-zero.
Since L < K ∩ gKg−1 , we have that QK ⊂ QL and σg (QK ) = QgKg
b ∈ QK , the two elements b and σg (b) commute with pL . It follows
−1
⊂ QL . Hence for all
Xb = pL EL (u∗gh xb)pL = pL EL (u∗gh bx)pL = pL σg (b)EL (u∗gh x)pL = σg (b)X.
By Lemma 4.2, pL (Q ⋊ L)pL = QL pL , so there exists a unique (non-zero) a ∈ QL such that
X = apL . By uniqueness, we see that a satisfies ab = σg (b)a for all a ∈ QK . We conclude that
g ∈ K0 , and hence supp(x) ⊂ K. By Theorem 3.7 we have that x ∈ Q ⋊ K.
Proposition 4.7. Consider an action σ : G y Q of a totally disconnected group G which is
properly outer relative to a compact open subgroup K0 < G whose action is minimal. For all
open subgroup K < K0 , we have (QK )′ ∩ M = LK ⊂ Q ⋊ K.
Proof. Fix x ∈ (QK )′ ∩ M . By Lemma 4.6 applied to K0 , we see that x ∈ Q ⋊ K0 . Since K0
acts minimally, the result follows from Lemma 2.8.
In order to prove Theorem A, we will use a convex combination argument. The following lemma
will be needed.
Lemma 4.8 ([HI15], Lemma 4.4). Consider a von Neumann algebra M, a weight Θ ∈ P(M),
and a σ-weakly closed convex subset C of M.
If C is bounded both in the operator norm and the k · kΘ -norm, then ΛΘ (C) is k · kΘ -closed in
L2 (M, Θ).
Proof of Theorem A. Fix an intermediate subfactor Q ⊂ N ⊂ M . Set H := {g ∈ G , ug ∈ N },
so that Q ⋊ H ⊂ N . Take x0 ∈ N . We will show that the support of x0 is contained in H.
This will conclude by Theorem 3.7.
Denote by Tr a normal faithful semi-finite trace on Q, and denote by Ψ ∈ P(M ) the corresponding dual weight.
Step 1. For all compact open subgroups K < K0 , all g ∈ G and all finite trace projection
q ∈ QK , if there exists x ∈ N such that qpK EK (u∗g x)qpK 6= 0, then ug qpK ∈ (N )1 pK .
Fix K < K0 , q ∈ QK , x ∈ N and g ∈ G such that qpK EK (u∗g x)qpK 6= 0. Put x′ := σg (q)xq ∈ N .
We have pK EK (u∗g x′ )pK = qpK EK (u∗g x)qpK 6= 0. By Lemma 4.2, there exists a unique a ∈ QK
such that pK EK (u∗g x′ )pK = apK . Note that a ∈ qQq. The element y := σg (a∗ )x′ ∈ σ(q)N q
satisfies pK EK (u∗g y)pK = a∗ apK 6= 0. Lemma 4.1, implies that
Ψ(pK u∗g ypK ) = Ψ(pK EK (u∗g y)pK ) =
1
kak2Tr 6= 0.
mG (K)
Denote by C ⊂ N the ultraweak closure of the convex hull
C0 := conv({σg (u)yu∗ , u ∈ U (qQK q)}).
CROSSED-PRODUCTS BY LOCALLY COMPACT GROUPS
17
We proceed as in the proof of [HI15, Theorem 4.3 (5) ⇒ (6)] to show that CpK is bounded
in k · kΨ -norm, except that we have slightly weaker assumptions. Since Q is contained in the
centralizer of Ψ, it is clear from the triangle inequality that kXpK kΨ ≤ kypK kΨ for all X ∈ C0 .
Now if X ∈ C is arbitrary, take a net (Xj )j∈J ⊂ C0 that converges ultrastrongly to X. Then
(Xj∗ Xj )j∈J converges ultraweakly to X ∗ X, and since Ψ is lower ultraweakly semi-continuous
(see [Ta03, Theorem VII.1.11.(iii)]) we get that
kXpK k2Ψ = Ψ(pK X ∗ XpK ) ≤ lim inf Ψ(pK Xj∗ Xj pK ) ≤ kypK k2Ψ .
j
Note moreover that kypK kΨ < +∞ since y = yq and qpK ∈ nΨ (M ). Thus the ultraweakly
closed convex set CpK ⊂ M is bounded both in the operator norm and in the k · kΨ -norm and
we can apply Lemma 4.8 to it. In particular we can find z ∈ C, such that zpK is the unique
element of CpK with minimal k · kΨ -norm2.
Let us check that such an element z ∈ N satisfies zpK 6= 0. First, since qpK ∈ nΨ (M ),
Ψ(qpK · pK q) is a bounded normal linear functional on M , which is constant on u∗g C0 . Hence it
is constant on u∗g C. It follows:
Ψ(qpK u∗g zpK q) = Ψ(qpK u∗g ypK q) = Ψ(pK u∗g (σg (q)yq)pK ) = Ψ(pK u∗g ypK ) 6= 0.
So we indeed find that zpK 6= 0.
Note also that CpK is globally invariant under the affine action U (qQK q) y σ(q)M q given
by u · X := σg (u)Xu∗ . Since Ψ centralizes Q, this action is k · kΨ -isometric, so it fixes zpK .
Equivalently, we have u∗g zpK ∈ (qQK q)′ ∩M . Moreover u∗g zpK belongs to qM q. By Proposition
4.7, we have that (qQK q)′ ∩ qM q ⊂ qLK, so u∗g zpK ∈ qLKpK . Since the projection pK is both
central and minimal inside LK, there exists a non-zero scalar λ ∈ C such that
u∗g zpK = λqpK .
(4.3)
We obtain that ug qpK = (z/λ)pK ∈ N pK . Unfortunately we don’t have, a priori, any control
on how small |λ| is, so z/λ could have a very large operator norm. To get around this issue,
we would like to identify polar parts, but we need extra commutations properties. We apply
the convex combination argument a second time.
Arguing as above, one can find an element z1 ∈ C ′ , where C ′ ⊂ C is the ultraweak closure of
conv({σg (u)zu∗ , u ∈ U (qQK q)}) ⊂ N , such that z1∗ zpK is the unique element in (C ′ )∗ zpK with
minimal k · kΨ -norm3. Then z1 enjoys the following properties:
• z1 ∈ N and z1 pK = zpK , because for all u ∈ U (qQK q), σg (u)zu∗ pK = σg (u)zpK u∗ =
zpK . In particular z1 pK 6= 0;
• z1∗ zpK ∈ (qQK q)′ ∩ qM q, by uniqueness of a k · kΨ -norm minimizer inside C ′ zpK . Hence
Proposition 4.7 gives that z1∗ zpK = λ′ qpK for some non-zero scalar λ′ ∈ C.
In particular the above facts give pK z1∗ z1 pK = pK z1∗ zpK = λ′ qpK and λ′ follows positive. Hence
the equality z1∗ z1 pK = λ′ qpK = (λ′ qpK )∗ = pK z1∗ z1 shows that z1∗ z1 commutes with pK . Write
the polar decomposition z1 = u|z1 |, with u ∈ N a partial isometry and |z1 | = (z1∗ z1 )1/2 . Note
that uqpK is a partial isometry since qpK commutes with z1∗ z1 (recall that z1 = z1 q).
√
We have: z1 pK = u|z1 |pK = λ′ uqpK . Combining this with (4.3), we get
√
λug qpK = zpK = z1 pK = λ′ uqpK .
Hence ug qpK and uqpK are proportional partial isometries; they have to coincide,
This proves Step 1.
ug qpK = (uq)pK ∈ (N )1 pK .
2Note however that z itself needs not be unique.
3Note that (C ′ )∗ zp is nothing but the ultraweakly closed convex hull of {uz ∗ zp u∗ , u ∈ U(qQK q)}.
K
K
18
RÉMI BOUTONNET & ARNAUD BROTHIER
Step 2. For all compact open subgroups K < K0 and g ∈ G, if there exists x ∈ N such that
pK EK (u∗g x)pK 6= 0, then ug pK ∈ (N )1 pK .
Fix a subgroup K < K0 , g ∈ G and x ∈ N such that pK EK (u∗g x)pK 6= 0.
Since K acts minimally on Q and QK is with expectation inside Q, [BHV15, Lemma 5.1] implies
that the trace Tr is still semi-finite on QK . Denote by (qi )i∈I an increasing net of projections
in QK with finite trace that converges to 1. Then there exists i0 ∈ I such that for all i ≥ i0 ,
we have that qi pK EK (u∗g x)pK qi 6= 0.
By Step 1, we deduce that for all i ≥ i0 , there exists yi ∈ (N )1 such that
(4.4)
ug qi pK = yi pK .
Denote by y ∈ (N )1 an ultraweak limit of the net (yi )i∈I . Then taking the corresponding limit
in (4.4) gives ug pK = ypK , as desired.
Step 3. The support of x0 is contained in H.
Take g ∈ supp(x0 ). Fix a compact open subgroup K < K0 . Then EK (u∗g x0 ) 6= 0. As
the net of projections (pL )L<G , converges strongly to the identity when L ց {e}, we have
pL EK (u∗g x0 )pL 6= 0 for all small enough compact open subgroups L.
Take a normal open subgroup L < K, so that pL commutes with ug for all g ∈ K. By Lemma
4.3, we can write
X
X
uh pL EL (u∗gh x0 )pL .
pL EK (u∗g x0 )pL =
pL uh EL (u∗h EK (u∗g x0 ))pL =
h∈lift(K/L)
h∈lift(K/L)
For all small enough compact open subgroups L < G that is normal inside K, since the above
sum is non-zero, we can find hL ∈ K such that pL EL (u∗ghL x0 )pL 6= 0. Applying Step 2 to L,
ghL and x0 , we find an element zL ∈ (N )1 such that ughL pL = zL pL .
Note that there exists a net (Li )i of compact open subgroups of G which are all normal in K and
form a neighborhood basis of e ∈ G. This comes from the fact that any open subgroup L < K
has finite index inside K, so that ∩h∈K hLh−1 is an open normal subgroup of K, contained in
L. By compactness of K and (N )1 (for the weak operator topology), there exist subnets of
(hLi )i and (zLi )i that converge to elements h ∈ K and z ∈ (N )1 , respectively. Taking ultraweak
limits, we get that
ugh = lim ughLi pLi = lim zLi pLi = z ∈ N,
i
i
because the net of projections (pL )L converges ultrastrongly to the identity. Hence gh ∈ H,
and we conclude that gK ∩ H 6= ∅.
As K can be arbitrarily small and H is closed, we conclude that g ∈ H. This finishes the proof
of Step 3. Now the theorem follows from Theorem 3.7.
4.4. Examples of actions.
Proposition 4.9. Let G be a totally disconnected group with a compact open subgroup K < G
such that ∩g∈G gKg−1 = {e} (let us call such a subgroup K eventually malnormal). Take a
II1 -factor Q0 , with trace τ0 . The following G-actions satisfy the assumptions of Theorem A, so
they are strictly outer and satisfy the Intermediate Subfactor Property.
(1) The Bernoulli action G y (Q0 , τ0 )⊗G/K obtained by shifting indices;
(2) The free Bernoulli action G y ∗G/K (Q0 , τ0 );
Proof. Let us first check separately the minimality condition on the K-action for each case.
⊗G/K
. Since K is eventually malnormal inside G, it acts faithfully on G/K and
(1) Put Q = Q0
hence on Q. Since K is compact open in G it is commensurated, and hence it acts on G/K
CROSSED-PRODUCTS BY LOCALLY COMPACT GROUPS
19
⊗O
with finite orbits. Let us denote by Oj , j ∈ J these orbits and by Qj := Q0 j . Example 2.7
K
shows that for all j, the fixed point algebra QK
j ⊂ Qj is an irreducible subfactor. Since Q
K
contains ⊗j∈J Qj , it is irreducible inside Q = ⊗j∈J Qj . So K acts minimally.
(2) Put Q = ∗G/K Q0 . Again, since K is not normal inside G, it acts faithfully on Q. Because
of the free product situation, it is clear that the copy of Q0 located at the label K ∈ G/K is
irreducible inside Q. Moreover this algebra is contained in QK , so K acts faithfully.
We now check the relatively properly outer condition simultaneously for both situations. Take
g ∈ G \ K and decompose Q as a product
(K)
(K)
Q ≃ Q0
(gK)
(gK)
⊗ Q0
⊗P
(K)
(or Q ≃ Q0
(gK)
∗ Q0
∗ P ),
where Q0 and Q0
are the copies of Q0 in position K ∈ G/K and gK ∈ G/K and P is the
tensor product (or free product) of all the remaining copies of Q0 .
In both the tensor situation and the free situation, one easily checks that for all nets (un )n ⊂
(K)
(gK)
U (Q0 ) and (vn )n ⊂ U (Q0 ) that converge weakly to 0, and for all a, b ∈ Q, one has
lim τ (un avn b) = 0, where τ is the trace of Q.
n
Assuming that there exists a ∈ Q1 such that σg (x)a = ax for all x ∈ QK
1 , we take for (un )n ⊂
(K)
K
U (Q0 ) ⊂ U (Q ) any net of unitaries that converges weakly to 0, and we set vn = σg (un ).
We get
kak2 = lim τ (u∗n a∗ aun ) = lim τ (un a∗ vn a) = 0.
n
n
Hence a = 0, as desired.
More generally, one can easily check the relative outerness condition in Theorem A when the
action has a large commutant thanks to the following fact.
Lemma 4.10. Consider an action G y Q and a closed subgroup K < G whose action is
minimal. Assume that the centralizer Γ of G in Aut(Q) satisfies: for all x ∈ Q \ C, there exists
γ ∈ Γ such that γ(x) ∈
/ Cx.
Then the action is properly outer relative to K if and only if the only elements of G acting
trivially on QK are the elements of K.
Proof. The only if part is trivial. Conversely assume that K is precisely the set of elements
of G that act trivially on QK . Take g ∈ G such that there exists a non-zero a ∈ Q satisfying
σg (x)a = ax for all x ∈ QK . Since K acts minimally, we can assume that a is a unitary. For
all automorphism γ ∈ Γ, we have
σg (γ(x))γ(a) = γ(a)γ(x),
for all x ∈ QK .
Moreover, using again that γ commutes with the G-action, we see that γ(QK ) = QK , and the
previous equation reads as
σg (x)γ(a) = γ(a)x,
for all x ∈ QK .
In particular, we find that a∗ γ(a) ∈ (QK )′ ∩ Q = C. Hence for all γ ∈ Γ, we have γ(a) ∈ Ca.
By our assumption on Γ, this leads to a ∈ C, and hence g fixes QK pointwise. Thus g ∈ K by
assumption.
Note that the condition on the centralizer Γ in the above lemma is fulfilled as soon as Γ admits
a subgroup Γ0 that preserves a state on Q, and such that QΓ0 = C. Moreover, if Q is a
II1 -factor the trace is invariant under any automorphism, and hence only the second condition
needs to be verified.
We deduce the following result in the spirit of Vaes’ examples [Va05, Theorem 5.1]
20
RÉMI BOUTONNET & ARNAUD BROTHIER
Corollary 4.11. Fix a totally disconnected group and a compact open subgroup K < G. Consider any faithful action G y Q0 on a II1 -factor such that
{g ∈ G | g|QK = id} = K.
Then the diagonal action G y (Q, τ ) := (Q0 , τ0 )⊗N satisfies the assumptions of Theorem A.
Proof. The K-action on Q0 being faithful, [Va05, Theorem 5.1] implies that it is strictly outer,
and hence minimal. Moreover the centralizer Γ of such a diagonal action contains all shift
automorphisms obtained by permuting indices. Hence the condition on Γ appearing in Lemma
4.10 is satisfied (see the comment after Lemma 4.10).
Thus the result follows from Lemma 4.10.
⊗G/K
Remark 4.12. Note that the Bernoulli shift action G y Q0
as in Lemma 4.9 is sometimes
a special case of diagonal action as in Corollary 4.11. For instance this happens if Q0 is the
hyperfinite factor. However this is not clear why this should be the case when Q0 is a prime
factor. Hence even the strict outerness for such actions does not follow from [Va05, Theorem
5.1].
Before moving to the next section, let us briefly explain how to adapt our argument to cover
some actions on type III factors.
Remark 4.13. Let G be a totally disconnected group and let K < G be an eventually malnormal compact open subgroup, see Lemma 4.9. Let Q0 be an arbitrary diffuse factor admitting
a faithful normal state φ0 with large centralizer (meaning that (Qφ0 0 )′ ∩ Q0 = C). Then the
Bernoulli shift G y (Q0 , φ0 )⊗G/K satisfies the Intermediate Subfactor Property.
Let us briefly explain. Denote by (Q, φ) := (Q0 , φ0 )⊗G/K . Then the centralizer Qφ of φ is
irreducible inside Q, and it is invariant under the G-action. One can show that for all compact
open subgroups L < G which is a finite intersection of conjugates of K, the fixed point algebra
QL := (Qφ )L satisfies Q′L ∩ (Q ⋊ G) ⊂ L(L). This is done by following the proof of Lemma 4.6,
and by noting that the action is properly outer relative to L.
Therefore one can use the averaging argument from the proof of Theorem A but instead we
average with elements in U ((Qφ )L ) for small groups L as above. Note moreover that in this
case since φ is a state (and not a weight), one does not need to bother with the projection q
appearing in Step 1 of the proof of Theorem A.
In fact the above remark also applies for free Bernoulli actions with weaker assumptions on
(Q0 , φ0 ), but we will not elaborate on this.
5. Existence of conditional expectations and operator valued weights
In this Section we discuss various results about existence (or non-existence) of conditional
expectations and operator valued weights in connection with Izumi, Longo, and Popa’s paper
[ILP98].
Let us start our discussion by investigating the existence of conditional expectations/operator
valued weights for pairs of the form Q ⋊ H ⊂ Q ⋊ G associated with closed subgroups H < G.
CROSSED-PRODUCTS BY LOCALLY COMPACT GROUPS
21
5.1. Proof of Theorem B. Let us start with two lemmas which rely on the notion of support.
Lemma 5.1. Consider two actions G y Q and G y P and the diagonal action of G on Q⊗P .
If the action G y Q is strictly outer, then
Q′ ∩ ((Q ⊗ P ) ⋊ G) = P.
Proof. Embed G into G × G diagonally. Then M := (Q ⊗ P ) ⋊ G is identified with a subalgebra
of M̃ := (Q ⊗ P ) ⋊ (G × G). Note that
Q′ ∩ M̃ ≃ Q′ ∩ (Q ⋊ G) ⊗ (P ⋊ G) = C ⊗ (P ⋊ G).
In particular any element x ∈ Q′ ∩ M , viewed as an element of M̃ has its support contained
in the diagonal subgroup of G × G (because x ∈ M ) and in {e} × G (because x ∈ Q′ ∩ M̃ ).
So such an element x has its support contained in the trivial group. By Corollary 3.8, we get
x ∈ Q′ ∩ (Q ⊗ P ) = P , as wanted.
Lemma 5.2. Consider an action G y Q on an arbitrary von Neumann algebra and put
M = Q ⋊ G. Take a weight Φ ∈ P(Q) and denote by Ψ ∈ P(M ) the corresponding dual weight.
For any non-zero x ∈ nΨ (M ), the support supp(x) has positive Haar measure inside G.
Proof. We will use the description of the dual weight relying on the Hilbert algebra approach,
[Ha77a]. According to this approach, given the weight Φ, there exists a left Hilbert algebra AΦ
satisfying the following properties.
• The Hilbert completion of AΦ is isomorphic to L2 (G, H) and the left von Neumann
algebra L(AΦ ) is identified with Q ⋊ G;
• The dual weight Ψ corresponds to the canonical weight on L(AΦ ) associated with the
Hilbert algebra AΦ ;
• For any Φ-right bounded vector η ∈ H and any f ∈ K(G), the function f η ∈ K(G, H)
defined by f η : g 7→ f (g)η is a right bounded vector for AΦ . The corresponding operator
πr (f η) is given by
Z
πr (η)σs (ξ(gs)) f (s−1 ) ds, for all ξ ∈ L2 (G, H), g ∈ G,
(5.1)
πr (f η)ξ(g) =
G
where πr (η) denotes the operator on H = L2 (Q, Φ) associated with the right bounded
vector η 4.
Now take a non-zero x ∈ nΨ (M ). By the above facts and [Ta03, Chapter VII.2], there exists
a non-zero left bounded vector ξ ∈ L2 (G, H) such that x = πℓ (ξ) (that is, x is the operator
extending the left multiplication by ξ). We claim that the support of ξ as a function in
L2 (G, H) is contained in the support of x. In fact the equality holds, but we clearly only need
this inclusion to deduce the lemma.
Take g ∈ G in the function support of ξ. Take an open neighborhood Ω ⊂ G of the identity
element e. We have to show that P (gΩ)πℓ (ξ)P (Ω) is non-zero. For all f ∈ K(Ω) and all Φ-right
bounded vector η ∈ H, we have
P (gΩ)πℓ (ξ)P (Ω)(f η) = P (gΩ)πℓ (ξ)(f η) = P (gΩ)πr (f η)ξ.
We now check that for a suitable choice of f and η the above quantity is non-zero.
Since g is in the function support of ξ, there exists η0 ∈ H such that gΩ∩{h ∈ G | hξ(h), η0 i =
6 0}
has positive measure. Since the set {πr (η1 )∗ η2 , η1 , η2 ∈ H, Φ-right bounded} is dense inside
H, we may find two Φ-right bounded vectors η1 , η2 ∈ H such that πr (η1 )∗ η2 is sufficiently
close to η0 so that the set gΩ ∩ {h ∈ G | hξ(h), πr (η1 )∗ η2 i 6= 0} has positive Haar measure. In
4Here we follow [Ha77b] and use the uniqueness of the standard form to identify canonically H with L2 (Q, Φ).
22
RÉMI BOUTONNET & ARNAUD BROTHIER
particular, the function ζ ∈ L2 (G, H) defined by h 7→ πr (η1 )ξ(h) satisfies P (gΩ)ζ 6= 0. Put
C := kP (gΩ)ζk2 > 0.
For f ∈ K(G),
denote by fˇ the function h 7→ f (h−1 ). For all non-negative function f ∈ K(G)
R
such that G f = 1 we have, by (5.1),
Z
2
ˇ
k(πr (fˇη1 )ξ)(h) − ζ(h)k2 dh
kπr (f η1 )ξ − ζk2 =
G
Z Z
f (s)πr (η1 ) σs (ξ(hs)) − ξ(h) dsk2 dh
k
=
ZG ZG
2
f (s)kπr (η1 )kkσs (ξ(hs)) − ξ(h)k ds dh
≤
G
G
Z
= kπr (η1 )k2
f (s)f (t)kσs (ξ(hs)) − ξ(h)kkσt (ξ(ht)) − ξ(h)k ds dt dh
G×G×G
Z
2
≤ kπr (η1 )k
f (s)f (t)kδG (s)−1/2 ρG (s)(ξ) − ξkkδG (t)−1/2 ρG (t)(ξ) − ξk ds dt
G×G
Z
2
f (s)kδG (s)−1/2 ρG (s)(ξ) − ξk ds ,
= kπr (η1 )k
G
where ρG is the representation defined in (2.1). Since ρG is a continuous representation and δG
is a continuous function on G, we get that if f is supported on a small enough neighborhood
of e, then kπr (fˇη1 )ξ − ζk2 < C/2. By definition of C we get:
kP (gΩ)πr (fˇη1 )ξk ≥ kP (gΩ)ζk − kπr (fˇη1 )ξ − ζk ≥ C/2.
So there indeed exists a Φ-right bounded vector η1 and a function f1 = fˇ which is supported
on Ω such that
P (gΩ)xP (Ω)(f1 η1 ) = P (gΩ)πℓ (ξ)(f1 η1 ) 6= 0.
Proof of Theorem B. We prove the two facts separately.
(1) First assume that the modular functions δG and δH coincide on H. Then by [Ha77a,
Theorem 3.2], for any weight Φ ∈ P(Q), the dual weight ΨG ∈ P(Q ⋊ G) and ΨH ∈ P(Q ⋊ H)
satisfy
σtΨG (x) = σtΨH (x), for all x ∈ Q ⋊ H, t ∈ R.
Therefore there exists a nfs operator valued weight T ∈ P(Q ⋊ G, Q ⋊ H) by [Ha77d, Theorem
5.1].
Conversely, assume that P(Q⋊G, Q⋊H) is non-trivial. By [Ha77d, Theorem 5.9], and Remark
2.3 we deduce that there exists an operator valued weight T ∈ P(M̃ , M ), where M = Q⋊G and
M̃ = (L∞ (G/H) ⊗ Q) ⋊ G (because M̃ is isomorphic to the basic construction of Q ⋊ H ⊂ M ).
Our intermediate goal is to deduce that there exists a G-invariant nfs weight on A := L∞ (G/H).
Unfortunately, we don’t know a priori that T is semi-finite on A. To get around this issue we
will exploit the fact that the action is strictly outer and use modular theory.
Let us consider the following operator valued weights in P(M̃ , Q).
• T1 := TG ◦ T , where TG ∈ P(Q ⋊ G, Q) is the Plancherel operator valued weight.
• Tφ := (φ ⊗ id) ◦ T̃G , where T̃G ∈ P(M̃ , A ⊗ Q) is the Plancherel operator valued weight
and φ ⊗ id ∈ P(A ⊗ Q, Q) is the tensor product operator valued weight associated to
some weight φ ∈ P(A) and the identity map on Q (see [Ha77d, Theorem 5.5]).
Fix φ ∈ P(A). By Lemma 5.1, we have that Q′ ∩ M̃ = A, hence the Connes Radon-Nikodym
cocycle (DT1 : DTφ )t in the sense of [Ha77d, Definition 6.2] takes values into A, and it is a
T
1-cocycle for the flow σt φ . By construction, for any weight ψ ∈ P(Q), the weight ψ ◦ Tφ is
CROSSED-PRODUCTS BY LOCALLY COMPACT GROUPS
23
T
simply the dual weight associated to φ ⊗ ψ ∈ P(A ⊗ Q). Hence (σt φ )|A = σtφ = id, since A is
abelian. We conclude that (ut )t := (DT1 : DTφ )t is a one parameter subgroup of unitaries of
A.
By [Co73, Théorème 1.2.4], there exists a nfs weight φ′ ∈ P(A) such that (Dφ′ : Dφ)t = ut for
all t ∈ R.
Claim 1. The weight φ′ is G-invariant. We will denote generically by the letter σ all the G
actions. We fix g ∈ G and show that the Connes Radon-Nikodym derivative (Dφ′ ◦ σg : Dφ′ )t
is equal to 1 for all t ∈ R. We have
(5.2)
(Dφ′ ◦ σg : Dφ′ )t = (Dφ′ ◦ σg : Dφ ◦ σg )t (Dφ ◦ σg : Dφ)t (Dφ : Dφ′ )t .
By Lemma 2.4, we have that (Dφ′ ◦ σg : Dφ ◦ σg )t = σg−1 ((Dφ′ : Dφ)t ) = σg−1 (ut ). Hence (5.2)
becomes
(5.3)
(Dφ′ ◦ σg : Dφ′ )t = σg−1 (ut )(Dφ ◦ σg : Dφ)t ut .
We will show that the right hand side above is equal to 1 by computing all the terms in the
equality
(5.4)
(DT1g : DT1 )t = (DT1g : DTφg )t (DTφg : DTφ )t (DTφ : DT1 )t .
Here we denoted by T1g (resp. Tφg ) the operator valued weight σg−1 ◦ T1 ◦ Ad(ug ) ∈ P(M̃ , Q)
(resp. σg−1 ◦ Tφ ◦ Ad(ug )).
Take a weight ψ ∈ P(Q). By definition of the Connes Radon-Nikodym derivative for operator
valued weights, we have
(DT1g : DTφg )t = (D(ψ ◦ T1g ) : D(ψ ◦ Tφg ))t
= (D(ψ ◦ σg−1 ◦ T1 ) ◦ Ad(ug ) : D(ψ ◦ σg−1 ◦ Tφ ) ◦ Ad(ug ))t
= σg−1 ((D(ψ ◦ σg−1 ◦ T1 ) : D(ψ ◦ σg−1 ◦ Tφ ))t )
= σg−1 ((DT1 : DTφ )t ),
where the third equality follows from Lemma 2.4. Hence (5.4) becomes
(5.5)
(DT1g : DT1 )t = σg−1 (ut )(DTφg : DTφ )t u∗t .
Now [Ha77b, Theorem 3.1] is easily seen to imply that T1g = δG (g)T1 while Tφg = δG (g)Tφ◦σg .
So (DT1g : DT1 )t = δG (g)it and
(DTφg : DTφ )t = δG (g)it (DTφ◦σg : DTφ )t = δG (g)it (Dφ ◦ σg : Dφ)t
Altogether, (5.5) can be rewritten
(5.6)
δG (g)it = σg−1 (ut )δG (g)it (Dφ ◦ σg : Dφ)t u∗t .
And we see that the right hand side of (5.3) is equal to 1, proving Claim 1.
Recall that A = L∞ (G/H) is the subalgebra of right H-invariant functions inside L∞ (G).
Denote by q : G → G/H the quotient map. The formula B ⊂ G/H 7→ φ′ (1q−1 (B) ) defines a
measure ν on the Borel σ-algebra of G/H. This measure is G-invariant and non-zero (because
φ′ is faithful).
Claim 2. The G-invariant measure ν is finite on every compact set of G/H.
Since φ′ is semi-finite (and faithful), there exists a Borel set B ⊂ G/H such that 0 < ν(B) < ∞.
Take a compactly supported non-negative function on G, f ∈ K(G), f 6= 0. Then the function
f ∗ 1B defined as follows is continuous:
Z
f (g)1B (g−1 xH) dmG (g), xH ∈ G/H.
(f ∗ 1B )(xH) :=
G
24
RÉMI BOUTONNET & ARNAUD BROTHIER
By Fubini-Tonelli’s Theorem, we have the key equation
Z
Z
Z
f (g)1B (g−1 xH) dmG (g) dν(xH)
(f ∗ 1B ) dν =
G/H G
G/H
Z
f (g)ν(gB) dmG (g)
=
G
Z
f dmG .
= ν(B)
G
This key equation tells us first that the continuous non-negative function f ∗ 1B is non-zero.
So there exists an open set U ⊂ G/H and α > 0 such that 1U ≤ α(f ∗ 1B ). The key equation
also tells us that f ∗ 1B is ν-integrable, and in particular, ν(U ) < ∞. Since any compact set
K ⊂ G/H can be covered by finitely many translates of U , Claim 2 follows.
By [BHV08, Corollary B.1.7] we deduce from the existence of ν that the modular functions δG
and δH must coincide on H. This proves (1).
(2) As mentioned earlier in the paper, if H is open inside G, then the indicator function 1H is
continuous and positive definite on G. Then the associated multiplier (see [Ha77b, Theorem
3.1.a]) gives the desired conditional expectation from Q ⋊ G onto Q ⋊ H.
Conversely, assume that H is not open inside G. If the modular functions δG and δH do not
coincide on H, then part (1) ensures that there is no nfs operator valued weight from Q ⋊ G
onto Q ⋊ H, and in particular, no conditional expectation.
Assume now that the modular functions do coincide. Fix a nfs weight Φ ∈ P(Q), and denote by
ΨH ∈ P(Q⋊H) and ΨG ∈ P(Q⋊G) the associated dual weights. As we saw in the proof of (1),
[Ha77d, Theorem 5.1] implies that there exists an operator valued weight T ∈ P(Q ⋊ G, Q ⋊ H)
such that ΨH ◦ T = ΨG . As the inclusion Q ⋊ H ⊂ Q ⋊ G is irreducible, it suffices to show
that T is unbounded, by [Ha77d, Theorem 6.6].
Fix an element f ∈ K(H, Q). Note that since H is not open inside G it has measure 0 inside
G5. In particular, since supp(f ) ⊂ H, Lemma 5.2 implies that f ∈
/ nΨG (Q ⋊ G). In contrast,
∗
f ∈ nΨH (Q ⋊ H), and hence the expression ψf := ΨH (f · f ) defines a normal positive linear
functional on Q ⋊ H. We have
T (1)(ψf ) = ψf ◦ T (1) = ΨH ◦ T (f ∗ f ) = ΨG (f ∗ f ) = +∞.
Hence T (1) ∈ Q\
⋊ H + \ Q ⋊ H, and so T is not bounded. The proof is complete.
5.2. When conditional expectations do exist. Applications to Hecke pairs. Before
proving Theorem C, let us mention that the argument of Choda applies beyond the setting of
discrete groups.
Theorem 5.3 ([Ch78], Theorem 3). Consider a strictly outer action G y Q of a locally
compact group and take a von Neumann subalgebra N ⊂ Q ⋊ G that contains Q and that is the
range of a faithful normal conditional expectation EN : Q ⋊ G → N .
Then N is of the form Q ⋊ H for some open subgroup H of G.
Proof. As usual, consider the closed subgroup H of G defined by H = {g ∈ G | ug ∈ N }, so
that Q ⋊ H ⊂ N . Let us show that the converse inclusion also holds.
Since N contains Q, we have u∗g EN (ug ) ∈ Q′ ∩ (Q ⋊ G), for all g ∈ G. As the action is strictly
outer, we deduce that for any g ∈ G, EN (ug ) is a scalar multiple of ug . If this scalar multiple
5Indeed for any positive measure Borel set A inside G, the product A−1 A contains a neighborhood of the
identity.
CROSSED-PRODUCTS BY LOCALLY COMPACT GROUPS
25
is non-zero, this means that ug ∈ N , and hence the scalar in question must be 1. So we obtain
the following computation:
EN (aug ) = 1{g∈H} aug ∈ Q ⋊ H,
for all a ∈ Q and g ∈ G.
Since EN is normal we deduce by linearity and density that EN (Q ⋊ G) ⊂ Q ⋊ H. Thus
N ⊂ Q ⋊ H, and we have equality. The fact that H is open follows from Theorem B.
We now mention a lemma that provides existence of conditional expectations. It follows from
the main technical result of [ILP98].
Lemma 5.4. Consider a compact open subgroup K < G in a locally compact group. Let
σ : G y Q be a strictly outer action and put M := Q ⋊ G, N := Q ⋊ K. Then any intermediate
von Neumann algebra N ⊂ L ⊂ M is the range of a normal faithful conditional expectation
EL : M → L.
Proof. We show that N ⊂ M satisfy the assumptions of [ILP98, Corollary 3.1]. Observe that
the inclusion N ⊂ M is irreducible since the action of G is strictly outer. As mentioned earlier
in the paper, since K is open inside G, there exists a conditional expectation E = EK from
M = Q ⋊ G onto N = Q ⋊ K.
By Remark 2.3, the basic construction M1 of the N ⊂ M is isomorphic to (Q ⊗ ℓ∞ (G/K)) ⋊ G,
where G acts diagonally. The embedding M ⊂ M1 is given by aug 7→ (a ⊗ 1)ug for all a ∈ Q,
g ∈ G. Note that G/K is discrete since K is open. By Lemma 5.1, we have Q′ ∩M1 = ℓ∞ (G/K).
Hence, the relative commutant N c := N ′ ∩ M1 is isomorphic to the von Neumann algebra of
K-bi-invariant maps ℓ∞ (K\G/K).
We represent faithfully M1 on H ⊗ ℓ2 (G/K) ⊗ L2 (G) in the obvious way. In this picture, the
Jones projection eN is the orthogonal projection onto H ⊗ CδK ⊗ L2 (G) where δK is the Dirac
b ∈ P(M1 , M ) of E.
mass of the coset K ∈ G/K. Consider the dual operator valued weight E
b
b
We have that E(eN ) = 1 by [Ko85, Lemma 3.1]. This implies that E(1KgK ) = [K : K ∩gKg−1 ],
where 1KgK is the characteristic function of KgK, g ∈ G. Since K < G is a compact open
subgroup, the index [K : K ∩ gKg−1 ] is finite for any g ∈ G and thus the operator valued
b is semi-finite on N c . Given any weight θ ∈ P(N ), [Ha77d, Theorem 6.6]
weight T := E ◦ E
states that the restriction σ θ◦T |N c of the modular flow associated to θ ◦ T to N c is equal to the
modular flow of the restriction of θ ◦ T to N c . Therefore, σtT (x) = x for any t ∈ R, x ∈ N c
since N c is commutative. Hence, the pair N ⊂ M indeed satisfies the assumptions of [ILP98,
Corollary 3.11], which implies the lemma.
Proof of Theorem C. This follows immediately by combining Theorem 5.3 and Lemma 5.4.
Let us now derive applications to crossed-products by Hecke pairs of groups. By definition, a
Hecke pair is a pair of groups (G, H) such that H is a subgroup of G which is commensurated
(or almost normal) in G, in the sense that H g := H ∩ gHg−1 has finite index in H and gHg−1
for all g ∈ G. We refer to [AD14] and [Pa12, Br16] for more details on facts below.
A typical example of a Hecke pair (G, H) arises when G is a subgroup of the automorphism
group of a locally finite connected graph Γ and H is the subgroup of elements of G that stabilize
a given vertex of Γ. In fact, this example is somewhat generic, see [AD12, Theorem 2.15].
e H),
e which is a new
To any Hecke pair (G, H) can be associated its Schlichting completion (G,
e
e
Hecke pair, for which G is a totally disconnected group and H is a compact open subgroup of
e The precise construction goes as follows: view G as a subgroup of SG/H , the permutation
G.
group of G/H. Endow SG/H with the topology of pointwise converge (where G/H is viewed
e (resp. H)
e to be the closure of G (resp. H) inside SG/H . The
as a discrete space), and define G
26
RÉMI BOUTONNET & ARNAUD BROTHIER
idea of using the Schlichting completion to study operator algebras of Hecke pairs goes back to
Tzanev [Tz01]. The key observation is Proposition 5.6 below.
We will say that an action G y Q is action of the Hecke pair (G, H) if it extends continuously
e Let us define the crossed-product von Neumann
to an action of the Schlichting completion G.
algebra associated with such an action. The construction generalizes that of C∗ -algebras associated with Hecke pairs (which correspond to the trivial action G y C), which originate
in [BC95]. We also refer to [AD14] for a general treatment. Crossed-product C ∗ -algebras for
actions of Hecke pairs were then defined in [Pa12] and the von Neumann algebraic version in
[Br16, Section 4].
Let C[Q; G, H] be the space of continuous functions f : G → Q such that f (hgk) = σh (f (g)) for
any g ∈ G, h, k ∈ H, and such that the induced functions f : G/H → Q is finitely supported.
Note that if f ∈ C[Q; G, H] and g ∈ G, then f (g) is fixed by σ(H ∩ gHg−1 ). We define a
multiplication and an involution ∗ on C[Q; G, H] as follows.
X
f1 f2 (g) =
f1 (s)σs (f2 (s−1 g)), for any f1 , f2 ∈ C[Q; G, H] and g ∈ G,
s∈lift(G/H)
where lift(G/H) is a system of representatives of G/H,
f ∗ (g) = σg (f (g−1 )∗ ), for any f ∈ C[Q; G, H] and g ∈ G.
The space C[Q; G, H] endowed with those operations is a unital ∗-algebra. It contains a copy
of the fixed point von Neumann algebra QH via the map j(a)(g) = χH (g)a, a ∈ QH , g ∈ G,
where χH is the characteristic function of H.
Assume that Q is standardly represented on a Hilbert space H and denote again by σ : G →
U (H) the canonical implementation of the action, see [Ha75]. Let L2 (G/H, H) be the Hilbert
space of L2 -functions from G/H to H, where G/H is viewed as a discrete space (and endowed
with the counting measure). Consider the subspace
K := {ξ ∈ L2 (G/H, H) | ξ(hgH) = σh (ξ(gH)), ∀h ∈ H, gH ∈ G/H}.
A similar proof to that of [Br16, Proposition 5.1] gives us the following.
Lemma 5.5. The map π : C[Q; G, H] → B(K) defined by
X
π(f )ξ(gH) =
f (s)σs (ξ(s−1 gH)), f ∈ C[Q; G, H], ξ ∈ K, gH ∈ G/H,
s∈lift(G/H)
is a bounded representation of the ∗-algebra C[Q; G, H]. We call it the standard representation
of the Hecke algebra.
Denote by vN[Q; G, H] the bicommutant of π(C[Q; G, H]). We call it the crossed-product von
Neumann algebra of Q by (G, H).
The next proposition relates crossed-products by Hecke pairs to the group crossed-product by
the Schlichting completion, it generalizes Lemma 4.2. As mentioned above, this idea goes back
to Tzanev [Tz01].
Proposition 5.6. Consider an action of a Hecke pair (G, H) on a von Neumann algebra Q.
e H)
e the Schlichting completion of (G, H). Then the pairs
Denote by (G,
e e
e e ⊂ p e (Q ⋊ G)p
QH ⊂ vN[Q; G, H] and pHe (Q ⋊ H)p
H
H
H
are isomorphic. Here pHe is the averaging projection associated with the compact open subgroup
e <G
e as defined above.
H
CROSSED-PRODUCTS BY LOCALLY COMPACT GROUPS
27
e H
e are naturally isomorphic. Since the action
Proof. Note that the coset spaces G/H and G/
e the pairs
G y Q extends continuously to an action of G,
e
e H]
e
QH ⊂ vN[Q; G, H] and QH ⊂ vN[Q; G,
e H)
e is equal to itself. Replacing
are isomorphic. Moreover, the Schlichting completion of (G,
e
e
e = G,
(G, H) with (G, H) if necessary, we may assume that H is compact open inside G, and G
e = H.
H
Consider the isometry u : K → L2 (G, H) given by the formula
1
uξ(g) =
σ −1 (ξ(gH)), g ∈ G.
mG (H)1/2 g
Put M := Q ⋊ G. Observe that the map x ∈ pH M pH 7→ u∗ xu ∈ B(K) is a representation that
is equivalent to pH M pH acting on pH JpH JL2 (G, H), where J is the conjugation operator on
the standard representation L2 (G, H) of M . In particular, x ∈ pH M pH 7→ u∗ xu ∈ B(K) is a
faithful representation.
Consider the map φ : C[Q; G, H] → K(G, Q) defined by the formula
1
σ −1 (f (g)), for any f ∈ C[Q; G, H], g ∈ G.
φ(f )(g) =
mG (H) g
Observe that the range of φ is the space of functions F ∈ K(G, Q) such that F (kgl) = σl−1 (F (g))
for any g ∈ G, k, l ∈ H. This range is precisely the corner pH K(G, Q)pH and we have that
(5.7)
u∗ φ(f )u = f for any f ∈ C[Q; G, H].
Therefore, x ∈ pH M pH 7→ u∗ xu ∈ B(K) is an isomorphism of von Neumann algebras onto
vN[Q; G, H]. By Lemma 4.2, we have that pH (Q ⋊ H)pH = QH pH and one easily checks that
u∗ (QH pH )u = QH .
Observe that if H < L < G is an intermediate closed group, then there is a natural identification
of C[Q; L, H] as a subalgebra of C[Q; G, H] of all functions that are supported on L. This
identification extends to the respective crossed-product von Neumann algebras. To see this
with little effort we can use the above proposition as follows. By definition of the Schlichting
completion we have that L̃ is isomorphic to the closure of L inside G̃, where L̃ and G̃ are the
Schlichting completions of L and G with respect to the subgroup H. As mentioned in Section
2.3, that the crossed-product Q ⋊ L̃ is isomorphic to the weak closure of the algebraic crossedproduct Q ⋊alg L̃ inside Q ⋊ G̃. Proposition 5.6 and this later fact imply that there exists an
injective morphism of von Neumann algebras φ : vN[Q; L, H] → vN[Q; G, H] that sends QH to
itself and such that φ(vN[Q; L, H]) is the weak closure of C[Q; L, H] inside vN[Q; G, H].
Theorem C together with Proposition 5.6 implies the following result for crossed-products by
Hecke pairs.
Corollary 5.7. Consider an action of a Hecke pair (G, H) on a von Neumann algebra Q.
If QH ⊂ L ⊂ vN[Q; G, H] is an intermediate von Neumann algebra, then there exists an
intermediate group H < H ′ < G such that L is isomorphic to the crossed-product vN[Q; H ′ , H],
which we identified as a subalgebra of vN[Q; G, H].
Recently, it has been shown that Hecke pairs appear in subfactor theory. Consider a finite index
subfactor N ⊂ M and its symmetric enveloping inclusion T ⊂ S, see [Po94] (or [LR95] in the
type III finite depth setting). Then in some cases, there exists a Hecke pair (G, K) and actions
G y M ⊗ M op , K × K y M ⊗ M op such that T ⊂ S is isomorphic with (M ⊗ M op )K×K ⊂
vN[M ⊗ M op ; G, K], see [Br16, Theorem 5.5]. Hence, our last corollary gives us information
about the lattice of intermediate subfactor in the symmetric enveloping inclusion of N ⊂ M .
28
RÉMI BOUTONNET & ARNAUD BROTHIER
References
[AD12]
[AD14]
[BHV08]
[BNP07]
[BC95]
[BHV15]
[Br16]
[Bu71]
[Ch78]
[CS16]
[CD75]
[Co73]
[Ey64]
[Ha75]
[Ha77a]
[Ha77b]
[Ha77c]
[Ha77d]
[HI15]
[HR16]
[ILP98]
[Jo83]
[Ko85]
[KPV14]
[LR95]
[MvN36]
[NT79]
[Pa12]
[Po94]
[Sa77]
[Ta73]
[Ta03]
[To08]
[Tz01]
C. Anantharaman-Delaroche, Invariant proper metrics on coset spaces. Topology Appl. 160
(2013), 546–552.
C. Anantharaman-Delaroche, Approximation properties for coset spaces and their operator algebras. in Proceedings of the Operator Theory, Timisoara.
B. Bekka, P. de la Harpe & A. Valette, Kazhdan’s property (T). New Mathematical Monographs,
11. Cambridge University Press, Cambridge, 2008.
D. Bisch, R. Nicoara, S. Popa, Continuous families of hyperfinite subfactors with the same standard
invariant. Internat. J. Math. 18 (2007), 255–267.
J.-B. Bost, A. Connes, Hecke algebras, type III factors and phase transitions with spontaneous
symmetry breaking in number theory. Selecta Math. (N.S.) 1 (1995), 411–457.
R. Boutonnet, C. Houdayer, S. Vaes, Strong solidity of free Araki-Woods factors. Preprint 2015,
arXiv:1512.04820.
A. Brothier, On fixed point planar algebras. J. Math. Phys. 57 (2016), 111703.
D. Bures, Abelian subalgebras of von Neumann algebras. Memoirs of the American Mathematical
Society, 110. American Mathematical Society, Providence, R.I., 1971.
H. Choda, A Galois correspondence in a von Neumann algebra. Tohoku Math. Journ. 30 (1978),
491–504.
J. Cameron, R. R. Smith, Intermediate subalgebras and bimodules for crossed products of general
von Neumann algebras. To appear in Internat. J. Math., 2016.
F. Combes, C. Delaroche Groupe modulaire d’une espérance conditionnelle dans une algèbre de von
Neumann. Bull. S.M.F. 103 (1975), 385–426.
A. Connes, Une classification des facteurs de type III. Ann. Sci. École Norm. Sup. 6 (1973), 133–252.
P. Eymard L’algèbre de Fourier d’un groupe localement compact. Bull. S.M.F. 92 (1964), 181–236.
U. Haagerup, The standard form of a von Neumann algebra. Math. Scand. 37 (1975), 271–283.
U. Haagerup, On the dual weight for crossed-product of von Neumann algebras, I. Math. Scand. 43
(1978), 99–118.
U. Haagerup, On the dual weight for crossed-product of von Neumann algebras, II. Math. Scand. 43
(1978), 119–140.
U. Haagerup, Operator valued weights in von Neumann algebras, I. J. Funct. Anal. 32 (1979), 175–
206.
U. Haagerup, Operator valued weights in von Neumann algebras, II. J. Funct. Anal. 33 (1979),
339–361.
C. Houdayer, Y. Isono, Unique prime factorization and bicentralizer problem for a class of type III
factors. Adv. Math. 305 (2017), 402–455.
C. Houdayer, S. Raum, Locally compact groups acting on trees, the type I conjecture and nonamenable von Neumann algebras. Preprint 2016, arXiv:1610.00884.
M. Izumi, R. Longo, S. Popa, A Galois correspondence for compact groups of automorphisms of von
Neumann algebras with a generalization to Kac algebras. J. Funct. Anal. 155 (1998), 25–63.
V.F.R. Jones, Index for subfactors. Invent. Math. 72 (1983), 1–25.
H. Kosaki, Extension of Jones theory on index to arbitrary factors. J. Funct. Anal. 66 (1986), 123–140.
D. Kyed, H. Densing-Pedersen, S. Vaes, L2 -Betti numbers of locally compact groups and their
cross section equivalence relations. Trans. Amer. Math. Soc. 367 (2015), 4917–4956.
R. Longo, K. Rehren, Nets of subfactors. Rev. Math. Phys. 7 (1995), 567–597.
F. Murray & J. von Neumann, Rings of operators. I. Ann. of Math. 37 (1936), 116–229.
Y. Nakagami, M. Takesaki, Duality for crossed products of von Neumann algebras. Lecture Notes
in Mathematics, 731. Springer, Berlin, 1979.
R. Palma, Crossed products by Hecke pairs I: ∗-completions. Mem. Am. Math. Soc. (to be published),
arXiv:1212.5756.
S. Popa, Symmetric enveloping algebras, amenability and AFD properties for subfactors. Math. Res.
Lett. 1 (1994), 409–425.
J.-L. Sauvageot Sur le type du produit croisé dune algèbre de von Neumann par un groupe localement
compact. Bull. Soc. Math. France 105 (1977), 349–368.
M. Takesaki, Duality for crossed products and the structure of von Neumann algebras of type III.
Acta Math. 131 (1973), 249–310.
M. Takesaki, Theory of operator algebras. II. Encyclopaedia of Mathematical Sciences, 125. Operator
Algebras and Non-commutative Geometry, 6. Springer-Verlag, Berlin, 2003. xxii+518 pp.
R. Tomatsu, A Galois correspondence for compact quantum group actions. J. Reine Angew. Math.
633 (2009), 165–182.
K. Tzanev, Hecke C∗ -algebras and amenability. J. Operator Theory 50 (2003), 169–178.
CROSSED-PRODUCTS BY LOCALLY COMPACT GROUPS
[Va01]
[Va05]
29
S. Vaes, The unitary implementation of a locally compact quantum group action. J. of Funct. Anal.
180 (2001), 426–480.
S. Vaes, Strictly outer actions of groups and quantum groups. J. Reine Angew Math. 578 (2005),
147–184.
Rémi Boutonnet, Institut de Mathématiques de Bordeaux, Université de Bordeaux, 351 cours de
la Libération, 33 405 Talence Cedex, France
E-mail address: [email protected]
Arnaud Brothier, Department of Mathematics, University of Rome Tor Vergata, Via della
Ricerca Scientifica 00133 Roma, Italy
E-mail address: [email protected]
https://sites.google.com/site/arnaudbrothier/
| 4math.GR
|
arXiv:1703.08304v1 [math.GR] 24 Mar 2017
DIMENSION QUOTIENTS, FOX SUBGROUPS AND LIMITS OF
FUNCTORS
ROMAN MIKHAILOV AND INDER BIR S. PASSI
Abstract. This paper presents a description of the fourth dimension quotient,
using the theory of limits of functors from the category of free presentations
of a given group to the category of abelian groups. A functorial description
of a quotient of the third Fox subgroup is given and, as a consequence, an
identification (not involving an isolator) of the third Fox subgroup is obtained.
It is shown that the limit over the category of free representations of the third
Fox quotient represents the composite of two derived quadratic functors.
1. Introduction
Given a group G, let Z[G] be its integral group ring and g the augmentation
ideal of Z[G]. The dimension quotients of G are defined to be its subquotients
Dn (G)/γn (G), n ≥ 1, where Dn (G) := G ∩ (1 + gn ) is the nth dimension subgroup
of G and γn (G) is the nth term in the lower central series {γi (G)}i≥1 of G. The
evaluation of dimension quotients is a challenging problem in the theory of group
rings, and has been a subject of investigation since 1935 ([3], [8], [13], [15], [16],
[17], [22], [23]). While these quotients are trivial for free groups ([16], [28]), for
n = 1, 2, 3 in case of all groups, and for odd prime-power groups [22], it was
first shown in 1972 by E. Rips [20] that D4 (G)/γ4 (G) is, in general, non-trivial.
Subsequently, the structure of these fourth dimension quotients has been described
by K. I. Tahara ([24], [25]) and Narain Gupta [8]. Instances of the non-triviality
of dimension quotients in all dimensions n ≥ 4 are now known ([8]; [17], p. 111);
however, their precise structure still remains an open problem.
Another challenging problem concerning normal subgroups determined by twosided ideals in group rings is the so-called Fox subgroup problem ([6], page 557; [1],
Problem 13; [8]). It asks for the identification of the normal subgroup F (n, R) :=
F ∩ (1 + rfn ) for a free group F and its normal subgroup R. A solution to this
problem has been given by I. A. Yunus [27] and Narain Gupta ([8], Chapter III). It
turns out that while F (1, R) = γ2 (R), F (2, R) = [R ∩ γ2 (F ), R ∩ γ2 (F )]γ3 (R), the
identification
p of F (n, R), n ≥ 3, is given as an isolator of a subgroup. For instance,
F (3, R) = G(3, R), where
G(3, R) := γ2 (R ∩ γ3 (F ))[[R ∩ γ2 (F ), R], R ∩ γ2 (F )]γ4 (R).
This identification essentially amounts to the one when the coefficients of the group
ring are in the field of rational numbers, rather than in the ring of integers, and
thus raises the question of the precise determination of the involved torsion.
1
2
ROMAN MIKHAILOV AND INDER BIR S. PASSI
Our aim in this paper is to present an entirely different approach to the above
problems via derived functors and limits of functors over the category of free presentations. Our approach is motivated by the connections between the theory of
limits of functors with homology of groups, derived functors in the sense of DoldPuppe [4], cyclic homology and group rings ([5], [18], [19], [26]). For instance, the
even dimensional integral homology groups turn up as limits [5]:
H2n (G) = lim (R/γ2 (R))⊗n
F ,
←−
where (R/γ2 (R))⊗n is the nth tensor power of the relation module R/γ2 (R)), and
⊗n
(R/γ2 (R))⊗n
, via
F is the group of F -coinvariants, the action of F on (R/γ2 (R))
conjugation in R, being diagonal. Certain derived functors in the sense of DoldPuppe [4] turn out as limits [18]:
L1 SP2 (Gab ) = lim
←−
γ2 (F )
,
γ2 (R)γ3 (F )
L1 SP3 (Gab ) = lim
←−
γ3 (F )
,
[γ2 (R), F ]γ4 (F )
where L1 SP2 and L1 SP3 are the first derived functors of the symmetric square and
cube functor respectively and Gab := G/γ2 (G). The description of derived functors
L1 SPn (Gab ) as limits for n ≥ 4 is given in [19]. An application of the theory of
limits to cyclic homology is given in [26] where it is shown that the cyclic homology
of algebras can be defined as limits over the category of free presentations of certain
simply defined functors. We work in the same direction, but consider the category
of groups. Our approach brings out a fresh context for the study of dimension
subgroups and Fox subgroups.
To describe the main results of the present work, let F be a free group, R a
normal subgroup of F , and G = F/R. Then there is a natural short exact sequence
γ2 (F )
γ2 (G)
R ∩ γ2 (F )
֒→
։
.
γ2 (R)(R ∩ γ4 (F ))
γ2 (R)γ4 (F )
γ4 (G)
Observe that the first two terms can be viewed as functors from the category of
free presentations
R ֒→ F ։ G
of G to the category of abelian groups. The limit functor lim is known to be left
←−
exact, i.e., it sends monomorphisms to monomorphisms, however, it is not right
exact, and therefore short exact sequences of presentations induce long exact
sequences involving higher derived lim i -terms. For instance, the above short
←−
sequence induces the following long exact sequence
lim
←−
R ∩ γ2 (F )
γ2 (G)
γ2 (F )
R ∩ γ2 (F )
֒→ lim
→
→ lim 1
.
←− γ2 (R)γ4 (F )
←− γ2 (R)(R ∩ γ4 (F ))
γ2 (R)(R ∩ γ4 (F ))
γ4 (G)
Our main result on dimension quotients describes the cokernel of the left monomorphism in the above exact sequence. To be precise, we have
Theorem 1. There is a natural short exact sequence
lim
←−
D4 (G)
γ2 (F )
R ∩ γ2 (F )
֒→ lim
։
.
←− γ2 (R)γ4 (F )
γ2 (R)(R ∩ γ4 (F ))
γ4 (G)
DIMENSION QUOTIENTS, FOX SUBGROUPS AND LIMITS OF FUNCTORS
3
Thus we present a description of the fourth dimension quotient purely in functorial
terms (not involving the group ring), which to us, is a very surprising result.
We next give a functorial description of the quotient F (3, R)/G(3, R), together
with a complete identification (not involving an isolator), of the third Fox subgroup
F (3, R).
Theorem 2. Let F be a free group and R a normal subgroup of F .
(a) There is a natural isomorphism
F (3, R)
≃ L1 SP2
G(3, R)
R ∩ γ2 (F )
γ2 (R)(R ∩ γ3 (F ))
.
(b)
F (3, R) = G(3, R)W ,
where W is a subgroup of F , generated by elements
[x, y]m [x, sy ]−1 [y, sx ]
with
xm = rx sx , rx ∈ R ∩ γ3 (F ), sx ∈ γ2 (R)
y m = ry sy , ry ∈ R ∩ γ3 (F ), sy ∈ γ2 (R).
F (3, R)
Finally, we give a description of lim G(3,
R) , where it may be noted that compo←−
sition of derived functors appears.
Theorem 3. There is a natural isomorphism
F (3, R)
lim
≃ L1 SP2 L1 SP2 (Gab ) .
←− G(3, R)
In particular, there a monomorphism
F (3, R)
L1 SP2 L1 SP2 (Gab ) ֒→
.
G(3, R)
The paper is organized as follows. In Section 2 - Preliminaries- we recall the basic
properties and results concerning limits of functors and derived functors needed in
the sequel. Section 3 - Generalized Dimension Quotients - is devoted to proving
several Lemmas on subgroups determined by two-sided ideals in free group rings.
Theorems 1, 2 and 3 are proved in Sections 4, 5 and 6 respectively. We refer the
reader to [21] for more details about limits, and to ([8], [17]) for the identification
of normal subgroups determined by two-sided idelas in group rings.
2. Preliminaries
2.1. Elementary properties of limits. We begin by recalling the basic facts
about limits of functors. Let C, D be arbitrary categories, and let Ab be the category
of abelian groups. Recall ([14], Chapter V) that the limit lim F of a functor
←−
F : C → D is an object of D together with a universal collection of morphisms
{ϕc : lim F → F (c)}c∈C such that F (f )ϕc = ϕc′ for any morphism f : c → c′ .
←−
4
ROMAN MIKHAILOV AND INDER BIR S. PASSI
Universality means that for any object d ∈ D and any collection of morphisms
{ψc : d → F (c)}c∈C such that F (f )ψc = ψc′ for any morphism f : c → c′ there
exists a unique morphism α : d → lim F such that ψc = ϕc α. The limit of a
←−
functor does not always exist; however, if it exists, then it is unique up to unique
isomorphism that commutes with morphisms ϕc .
A category C is said to be strongly connected if for any objects c, c′ ∈ C the homset HomC (c, c′ ) is non-empty. A standard argument implies that, for a strongly
connected category C and a functor F : C → Ab, for any object c ∈ C, there is a
natural embedding
αc : lim F ֒→ F (c).
←−
More precisely, lim F can be identified as follows:
←−
lim F = {x ∈ F (c) | for all c′ ∈ C, and morphisms f1 , f2 : c → c′ ,
←−
F (f1 )(x) = F (f2 )(x)}.
A category with pairwise coproducts is a category C such that for any objects
i2
i1
c2 in C. For a category
c1 ⊔ c2 ←−
c1 , c2 ∈ C there exists the coproduct c1 −→
with pairwise coproducts C, objects c1 , c2 ∈ C, and a functor F : C → Ab, there is
a natural map
T : F (c1 ) ⊕ F (c2 ) → F (c1 ⊔ c2 ).
A functor F is called monoadditive, if T is injective for any pair of objects c1 , c2 ∈ C.
The following lemma is due to S. O. Ivanov. It gives a way to define the limit
of a functor as an equalizer. This lemma will not be used in the proofs of main
statements of the paper.
Lemma 4. Let C be a strongly connected category with pairwise coproducts and
F : C → Ab be a functor. Then there are exact sequences
α
F (i1 )−F (i2 )
c
F (c) −−−−−−−−−→ F (c ⊔ c),
0 −→ lim F −→
←−
T
0 −→ lim F −→ F (c) ⊕ F (c) −→ F (c ⊔ c).
←−
The map lim F → F (c) ⊕ F (c) is given by x 7→ (αc (x), −αc (x)), where
←−
αc : lim F → F (c) is the structure morphism of lim F .
←−
←−
Proof. Since C is strongly connected, the map αc : lim F → F (c) is a monomor←−
phism and its image is equal to the subgroup of C-invariant elements:
lim F ∼
= {x ∈ F (c) | for all c′ , ϕ, ψ : c → c′ F (ϕ)(x) = F (ψ)(x)}.
←−
We identify lim F with this subgroup.
←−
DIMENSION QUOTIENTS, FOX SUBGROUPS AND LIMITS OF FUNCTORS
5
We claim that Ker(T) = {(x, −x) | x ∈ lim F }. Consider two arbitrary arrows
←−
ϕ, ψ : c → c′ and the commutative diagram:
F (c)
F (i1 )−F (i2 )
/ F (c ⊔ c)
(−11 )
T
/ F (c ⊔ c)
F (c) ⊕ F (c)
◆◆◆
◆◆◆
◆◆
F ((ϕ,ψ))
(F (ϕ), F (ψ)) ◆◆◆
&
F (c′ ).
Assume that (x, y) ∈ Ker(T). Then (x, y) ∈ Ker((F (ϕ), F (ψ))), and hence
F (ϕ)(x) = −F (ψ)(y). If we take ϕ = ψ = idc , we get x = −y. It follows that for any
c′ ∈ C and ϕ, ψ : c → c′ we have F (ϕ)(x) = F (ψ)(x). Thus (x, y) = (x, −x), where
x ∈ lim F . Let x ∈ lim F . If we take ϕ = i1 : c → c ⊔ c and ψ = i2 : c → c ⊔ c, we
←−
←−
obtain (x, −x) ∈ Ker(T).
1
Since −1
: F (c) → F (c) ⊕ F (c) is a monomorphism and its image contains the
kernel of T, the kernel of F (i1 ) − F (i2 ) is equal to lim F as well.
←−
Corollary 5. Let C be a strongly connected category with pairwise coproducts and
let F : C → Ab be a functor. Then the functor F is monoadditive if and only if
lim F = 0.
←−
Our main concern is the category E of free presentations of a given group G. The
objects of E are surjective homomorphisms π : F ։ G where F is a free group and
morphisms f : (π1 : F1 ։ G) → (π2 : F2 ։ G) are homomorphisms f : F1 → F2
such that π1 = π2 f . The category E has coproducts given by
(π1 : F1 ։ G) ⊔ (π2 : F2 ։ G) = ((π1 , π2 ) : F1 ∗ F2 ։ G).
The category E is a strongly connected category with pairwise coproducts and
therefore, Lemma 4 applies to functors F : E → Ab. Thus, in particular, the limits
lim F can be identified with corresponding equalizers.
←−
A representation (i.e., a functor) F (F ։ G) : E → Ab will be called a Grepresentation if, for any (F ։ G) ∈ E, the natural inclusion
lim F ֒→ F (F ։ G)
←−
is an isomorphism. That is, a G-representation F (F ։ G) depends only on G.
The limit lim can be defined as the right adjoint to the diagonal functor. It is
←−
left exact, but not right exact. A short exact sequence of representations
F1 ֒→ F2 ։ F3
induces a long exact sequence of derived functors lim i :
←−
1
lim F1 ֒→ lim F2 → lim F3 → lim F1 → lim 1 F2 → lim 1 F3 → . . .
←−
←−
←−
←−
←−
←−
6
ROMAN MIKHAILOV AND INDER BIR S. PASSI
See (Section 2, [21]) for details. In this paper, we will use only one property of higher
limits, namely, their triviality for G-representations. The higher limits lim i F can
←−
be defined as ith cohomology of the category E with coefficients in a representation
F . Since the category E has pairwise products, it is contractible (see, for example,
Lemma 3.5 [21]), hence for any G-representation F , lim i F = 0, i ≥ 1.
←−
Lemma 6. Let F (F ։ G) be a representation, which lives in an exact sequence
F1 (G) → F (F ։ G) → F2 (G),
where Fi (F ։ G) = F (G), i = 1, 2 are G-representations. Then F (F ։ G) is
also a G-representation.
Proof. Let us first consider the case F2 = 0, i.e, the case of a natural surjection
F1 (G) ։ F (F ։ G).
The assertion in this case follows from the following diagram:
lim F1 /
←−
←−
F1 (G)
lim F1
F1 (G)
lim F1
,
←−
lim F /
since the quotient
//
/ F1 (G)
/ F (F ։ G)
//
F (F ։G)
lim F
←−
is zero. Dually, if F1 = 0, the quotient
←−
F2 (G)/F (F ։ G)
is a G-representation, since it is an epimorphic image of a G-representation. The
assertion in this case follows from the following commutative diagram
lim F /
←−
F (F ։ G) /
/ lim F2
←−
/ F2 (G)
/ lim F2
F (F ։G)
←−
//
F2 (G)
F (F ։G) .
Let us now assume that the map F1 (G) → F (F ։ G) is a monomorphism and
F (F ։ G) → F2 (G) is an epimorphism. Since lim 1 F2 = 0, we obtain the
←−
following diagram
lim F1 /
←−
F1 (G) /
and the result follows.
/ lim F
←−
/ F (F ։ G)
/ / lim F2
←−
/ / F2 (G)
DIMENSION QUOTIENTS, FOX SUBGROUPS AND LIMITS OF FUNCTORS
7
2.2. Quadratic functors. We will use the following basic quadratic functors:
tensor square ⊗2 , symmetric square SP2 , exterior square Λ2 and antisymmetric
˜ 2 . Recall that, for an abelian group A,
square ⊗
SP2 (A) = ⊗2 (A)/ha ⊗ b − b ⊗ a, a, b ∈ Ai
Λ2 (A) = ⊗2 (A)/ha ⊗ a, a ∈ Ai
˜ 2 (A) = ⊗2 (A)/ha ⊗ b + b ⊗ a, a, b ∈ Ai.
⊗
For any abelian group A, there is a natural exact sequence
˜ 2 (A) → Λ2 (A) → 0,
0 → A ⊗ Z/2 → ⊗
where the left hand map is given by ā 7→ a ⊗ a, a ∈ A, ā = a + 2A ∈ A ⊗ Z/2.
The derived functors in the sense of Dold-Puppe [4] are defined as follows. For
an abelian group A and an endofunctor F on the category of abelian groups, the
derived functor of F is given as
Li F (A) = πi (F KP∗ ), i ≥ 0,
where P∗ → A is a projective resolution of A, and K is the Dold-Kan transform,
inverse to the Moore normalization functor from simplicial abelian groups to chain
complexes.
The first derived functor L1 SP2 (A) which will be used in many places of this
paper is the natural quotient of Tor(A, A) by diagonal elements (a, a), ma = 0, a ∈
A. The first derived functor L1 Λ2 (A) is the subfunctor of Tor(A, A) generated by
these diagonal elements, i.e., there is a natural short exact sequence
0 → L1 Λ2 (A) → Tor(A, A) → L1 SP2 (A) → 0.
An abstract value of the abelian group L1 SP2 can be easily computed from the
following data:
L1 SP2 (Z/m) = L1 SP2 (Z) = 0, m ≥ 2;
and
L1 SP2 (A ⊕ B) = L1 SP2 (A) ⊕ L1 SP2 (B) ⊕ Tor(A, B),
for all abelian groups A, B.
The derived functors L1 Λ2 and L1 SP2 naturally appear in the homology of
Eilenberg-MacLane spaces. For example, for an abelian A, there are natural short
exact sequences, which do not split (see [2])
0 → Λ3 (A) → H3 (A) → L1 Λ2 (A) → 0
0 → L1 SP2 (A) → H5 K(A, 2) → Tor(A, Z/2) → 0,
where Λ3 is the exterior cube. We refer to the thesis of F. Jean [10] for the structure
of derived functors of higher symmetric powers.
We also need some other natural exact sequences, like the following
(2.1)
2
˜ (A) → L1 Λ2 (A) → 0,
0 → Tor(A, Z/2) → L1 ⊗
and
(2.2)
˜ 2 (A) → SP2 (A) → ⊗2 (A) → ⊗
˜ 2 (A) → 0.
0 → L1 SP2 (A) → Tor(A, A) → L1 ⊗
8
ROMAN MIKHAILOV AND INDER BIR S. PASSI
˜ 2 (A) → SP2 (A) is, in general, non-zero. There sequences
Note that the map L1 ⊗
˜ 2 ։ Λ2 and SP2 → ⊗2 ։ ⊗
˜2
are obtained by deriving the sequences − ⊗ Z/2 → ⊗
respectively.
The following sequence will be used the proofs of our main results several times.
For a free abelian group I and its subgroup J, there is a natural exact sequence
(2.3)
0 → L1 SP2 (I/J) →
Λ2 (I)
→ I ⊗ I/J → SP2 (I/J) → 0,
Λ2 (J)
where the image of an element (x, y), mx = my ∈ J, x, y ∈ I, is mx ∧ y. For the
proof see (Theorem 12, [18]; Section 3, [19]). The proof directly follows from the
result of Köck [12] saying that the Koszul-type complex Λ2 (J) → I ⊗ J → SP2 (I)
represents the element LSP2 (I/J) of the derived category of abelian groups.
3. Generalized dimension quotients
In analogy with the dimension subgroups
Dn (G) = F ∩ (1 + rZ[F ] + fn )/γn (R), n ≥ 1,
when R is a normal subgroup of a free group F with G = F/R, the normal subgroups F ∩ (1 + a + fn ), n ≥ 1, where a is a two-sided ideal of Z[F ] are called
generalized dimension subgroups. We set
D(n, a) := F ∩ (1 + a + fn ).
An example of description of a generalized dimension subgroup and its connection
to a derived functor, which we will use later, is the following. It is shown in [9]
that, there is a natural isomorphism
D(3, fr)
D(3, rf)
=
= L1 SP2 (Gab ).
(3.1)
γ2 (R)γ3 (F )
γ2 (R)γ3 (F )
See [18] and [19] for more examples of such type.
We need the identification of certain generalized dimension subgroups. Recall
that
if R is a normal subgroup of a free group F and Y is a basis of
R, then the two-sided ideal rZ[F ], viewed as a left (resp. right)
Z[F ]-module, is free with basis {y − 1 | y ∈ Y } ([7], Theorem 1,
p. 32).
Lemma 7. If F is a free group of finite rank m ≥ 1, X = {x1 , x2 , . . . , xm } an
ordered basis of F , and R a normal subgroup of F generated by
{xe11 ξ1 , xe22 ξ2 , . . . , xemm ξm , ξm+1 , . . .}
with ξi ∈ γ2 (F ) for i = 1, 2, . . . , and integers ei ≥ 0 satisfying em |em−1 | . . . |e2 |e1 ,
then the generalized dimension subgroup
D(4, frf + f2 r) := F ∩ (1 + frf + f2 r + f4 )
is generated, modulo γ4 (F ), by the commutators
• [[xj , xi ], xi ]ei (j > i),
• [[xj , xi ], xk ]ℓjk (j > i < k), where ℓjk = lcm(ej , ek ).
DIMENSION QUOTIENTS, FOX SUBGROUPS AND LIMITS OF FUNCTORS
9
Proof. It is easy to see that the elements
[[xj , xi ], xi ]ei (j > i), and [[xj , xi ], xk ]ℓjk (j > i < k),
where ℓjk = lcm(ej , ek ) all lie in D(4, frf + f2 r).
Let S = hxei (1 ≤ i ≤ m), γ2 (F )i, and w ∈ D(4, frf + f2 r). Observe that
frf + f2 r + f4 = fsf + f2 s + f4 .
Since D3 (F ) = γ3 (F ) and {[[xj , xi ], xk ] | m ≥ j > i ≤ k ≤ m} is a basis of
γ3 (F )/γ4 (F ), we have, modulo γ4 (F ),
w=
m
Y
wi , wi =
i=1
Y
[[xj , xi ], xk ]aijk ∈ D(4, fsf + f2 s), aijk ∈ Z.
j>i≤k
4
Modulo f , we have
(3.2) Wi := wi − 1 =
X
{aijk [(xj − 1)(xi − 1) − (xi − 1)(xj − 1)](xk − 1)
j>i≤k
− (xk − 1)[(xj − 1)(xi − 1) − (xi − 1)(xj − 1)]}.
Left differentiating (in the sense of free differential calculus [see [8], p.7]) with
respect to xi , we have
X
aiji [(xj − 1)(xi − 1) − 2(xi − 1)(xj − 1)]
(3.3) xi Wi =
j>i
−
X
aijk (xk − 1)(xj − 1) ∈ fs + ei f2 + f3 .
j>i<k
Right differentiating with respect to xi , gives
2aiji (xj − 1) ∈ s + ei f + f2 (j > i).
Hence ej |2aiji , j > i, and Eq(3.3) implies
X
X
aiji (xj − 1)(xi − 1) −
aijk (xk − 1)(xj − 1) ∈ fs + ei f2 + f3 .
(3.4)
j>i
j>i<k
Since the second sum does not involve xi , we conclude that
X
aiji (xj − 1)(xi − 1) ∈ fs + ei f2 + f3
(3.5)
j>i
and
(3.6)
X
aijk (xk − 1)(xj − 1) ∈ fs + ei f2 + f3 .
j>i<k
Eq(3.5) implies that aiji (xi − 1) ∈ s + ei f + f2 , and therefore we have
(3.7)
ei |aiji , j > i.
Similarly Eq(3.6) implies that
(3.8)
ej |aijk , j > i < k.
Eq(3.2) thus reduces to
X
(3.9)
aijk (xj − 1)(xi − 1)(xk − 1) ∈ fsf + f2 s + f4 .
j>i<k
10
ROMAN MIKHAILOV AND INDER BIR S. PASSI
Eq(3.9) yields that
(3.10)
ek |aijk , j > i < k.
Eqs (3.8) and (3.10) imply
(3.11)
ℓij |aijk , j > i < k, ℓ = lcm(ej , ek ).
Hence the element w lies in the subgroup generated by the commutators
• [[xj , xi ], xi ]ei (j > i),
• [[xj , xi ], xk ]ℓjk (j > i < k), where ℓjk = lcm(ej , ek ), as claimed.
Corollary 8. If F is a free group and R a normal subgroup of F such that
F/Rγ2 (F ) is torsion-free, then
D(4, frf + f2 r) = [[R, F ], R]γ4 (F ).
Proof. It is easy to see that [[R, F ], R]γ4 (F ) ⊆ D(4, frf + f2 r). Observe that, for
the reverse inclusion, it suffices to consider the case when F is finitely generated,
and so Lemma 7 applies.
With free group F and its subgroup S as in Lemma 7, we have
Lemma 9.
D(4, sf ∩ f3 ) = [γ2 (F ), S]γ4 (F ).
Proof. Let w ∈ D(4, sf ∩ f3 ). Then, modulo γ4 (F ),
w=
m−1
Y
wi , wi =
i=1
Y
[[xj , xi ], xk ]aijk , aijk ∈ Z,
m≥j>i≤k≤m
3
with wi ∈ D(4, sf ∩ f ). Modulo sf ∩ f3 + f4 , for every i ∈ {1, 2, . . . , m − 1},
(3.12)
Wi := wi − 1 =
X
{aijk [(xj − 1)(xi − 1) − (xi − 1)(xj − 1)](xk − 1)
m≥j>i≤k≤m
− (xk − 1)[(xj − 1)(xi − 1) − (xi − 1)(xj − 1)]}
Since (xj − 1)(xi − 1) − (xi − 1)(xj − 1) ∈ s, it follows that
(3.13)
Wi′ :=
X
aijk (xk − 1)[(xj − 1)(xi − 1) − (xi − 1)(xj − 1)] ∈ sf ∩ f3 + f4 .
m≥j>i≤k≤m
Eq(3.13) implies that for every j ∈ {i + 1, . . . , m} and i ∈ {1, 2, . . . m − 1},
(3.14)
′
xj Wi
=
m
X
aijk (xk − 1)(xi − 1) ∈ s + f3 .
k=i
It follows that
(3.15)
ek |aijk , for m ≥ j > i ≤ k ≤ m.
Thus we see that wi ∈ D(4, sf ∩ f3 ) only if
(3.16)
ek |aijk (m ≥ j > i ≤ k ≤ m),
DIMENSION QUOTIENTS, FOX SUBGROUPS AND LIMITS OF FUNCTORS
11
and so wi and hence w ∈ [γ2 (F ), S]γ4 (F ). Hence D(4, sf ∩ f3 ) ⊆ [γ2 (F ), S]γ4 (F ).
The reverse inclusion is easily seen to hold.
Lemma 10. If F is a free group and R a normal subgroup of F , then, modulo
γ4 (F ),
D(4, fr ∩ f3 ) = D(4, rf ∩ f3 )
= h[f ′ , f ]e | f ′ ∈ γ2 (F ), f ∈ F, f ′e ∈ (R ∩ γ2 (F ))γ3 (F ), f e ∈ Rγ2 (F )i.
Proof. The first equality is an immediate consequence of the canonical
anti-isomorphism of Z[F ] induced by w 7→ w−1 , w ∈ F .
It is easy to check that
[f ′ , f ]e ∈ D(4, rf ∩ f3 )
if f ′ ∈ γ2 (F ), f ∈ F, f ′e ∈ (R ∩ γ2 (F ))γ3 (F ), f e ∈ Rγ2 (F ).
Conversely, let w ∈ D(4, rf ∩ f3 ). To analyze w, we may clearly assume that F
is finitely generated, R, S are as in Lemma 7, and so w ∈ D(4, sf ∩ f3 ). Therefore,
by Lemma 9, we have, modulo γ4 (F ),
w=
m−1
Y
wi
i=1
with
Y
wi =
[[xj , xi ], xk ]ek bijk , bijk ∈ Z.
m≥j<i≤k≤m
On collecting terms, we have,
w≡
m
Y
[fk , xk ]ek , fk ∈ γ2 (F ), 1 ≤ k ≤ m.
k=1
Now w − 1 ∈ rf ∩ f3 + f4 . Therefore we have
m
X
ek (fk − 1)(xk − 1) ∈ rf ∩ f3 + f4 .
k=1
Differentiating with respect to xk , yields
fkek − 1 ∈ r + f3 .
Hence
fkek ∈ (R ∩ γ2 (F ))γ3 (F ), 1 ≤ k ≤ m,
and consequently
w ∈ h[f ′ , f ]e | f ′ ∈ γ2 (F ), f ∈ F, f ′e ∈ (R ∩ γ2 (F ))γ3 (F ), f e ∈ Rγ2 (F )i.
Observe that, for every free group F and its normal subgroup R,
(3.17)
γ2 (R) ∩ γ3 (F ) = [R ∩ γ2 (F ), R].
12
ROMAN MIKHAILOV AND INDER BIR S. PASSI
To see this equality, consider the map between exterior squares
Λ2 (R/(R ∩ F ′ )) → Λ2 (Fab ),
(3.18)
induced by inclusion R/(R ∩ γ2 (F )) ֒→ Fab , where Fab denotes the abelianization
of F . The map (3.18) is a monomorphism, since it is induced by a monomorphism
of free abelian groups. The needed equality (3.17) now follows from the following
identifications:
Λ2 (R/(R ∩ γ2 (F ))) =
γ2 (R)
,
[R ∩ γ2 (F ), R]
Λ2 (Fab ) = γ2 (F )/γ3 (F ).
Thus, in particular,
(γ2 (R) ∩ γ3 (F ))γ4 (F ) = [R ∩ γ2 (F ), R]γ4 (F ).
In view of the known structure ([14], Chapter V, Section 5) of Tor(A, B) for abelian
groups A, B, the preceding Lemma immediately yields the following result.
Lemma 11. There is a natural epimorphism
φ : Tor(γ2 (G)/γ3 (G), Gab ) →
D(4, fr ∩ f3 )
D(4, fr ∩ f3 )
=
.
[R ∩ γ2 (F ), R]γ4 (F )
(γ2 (R) ∩ γ3 (F ))γ4 (F )
4. Proof of theorem 1
First observe that the cokernel of the natural map
D(4, fr)
R ∩ D(4, fr)
֒→
γ2 (R)γ4 (F )
γ2 (R)γ4 (F )
can be naturally identified with the fourth dimension quotient
D4 (G)
D(4, rZ[F ])
=
γ4 (G)
Rγ4 (F )
(see, for example, [8], p.8̇0). There is an obvious short exact sequence
D(4, fr)
D4 (G)
R ∩ D(4, fr)
֒→
→
.
γ2 (R)γ4 (F )
γ2 (R)γ4 (F )
γ4 (G)
The right hand map is, in fact, surjective. For, let w ∈ D(4, rZ[F ]) so that w − 1 ∈
]
rZ[F ] + f4 . The quotient rZ[F
is the relation module R/γ2 (R), and so there is a
fr
natural epimorphism
R/γ2 (R) ։
rZ[F ] + f4
fr + f4
and we can find an element r ∈ R, such that wr − 1 ∈ fr + f4 . It thus follows that
D(4, fr)
4 (G)
every element of the dimension quotient D
γ4 (G) has a preimage in γ2 (R)γ4 (F ) .
DIMENSION QUOTIENTS, FOX SUBGROUPS AND LIMITS OF FUNCTORS
13
Consider the natural diagram with exact rows and columns
(4.1)
R∩D(4, fr)
γ2 (R)γ4 (F )
/
R∩γ2 (F )
γ2 (R)(R∩γ4 (F ))
/
f2
fr+f4
/
f2
(R∩γ2 (F )−1)+fr+f4
//
D(4, fr)
γ2 (R)γ4 (F )
γ2 (F )
γ2 (R)γ4 (F )
/
/
2
f
fr+f4
/
2
f
(γ2 (F )−1)+fr+f4
//
D4 (G)
γ4 (G)
/
γ2 (G)
γ4 (G)
/
Now observe that
f2
f
f
= lim 3 ⊗Z[G]
= 0.
←− fr + f4
←− f
r + f3
f
f
g
This follows from monoadditivity of ff3 ⊗Z[G] r+f
3 = f3 ⊗Z[G] g3 which can be
easily checked. The top two horizontal exact sequences in (4.1) imply the natural
isomorphisms
lim
R ∩ γ2 (F )
R ∩ D(4, fr)
≃ lim
←− γ2 (R)(R ∩ γ4 (F ))
γ2 (R)γ4 (F )
γ2 (F )
D(4, fr)
≃ lim
.
lim
←− γ2 (R)γ4 (F )
←− γ2 (R)γ4 (F )
lim
←−
Therefore, the left hand vertical exact sequence in diagram (4.1) implies that we
have the following the long exact sequence
(4.2) lim
←−
R ∩ γ2 (F )
D4 (G)
γ2 (F )
R ∩ D(4, fr)
֒→ lim
→
→ lim 1
←− γ2 (R)γ4 (F )
←− γ2 (R)γ4 (F )
γ2 (R)(R ∩ γ4 (F ))
γ4 (G)
Next consider the following diagram with exact rows and columns
(4.3)
D(4, fr∩f3 )
(γ2 (R)∩γ3 (F ))γ4 (F )
/
D(4, fr)
γ2 (R)γ4 (F )
/
/
D(3, fr)
γ2 (R)γ3 (F )
γ3 (F )
(γ2 (R)∩γ3 (F ))γ4 (F )
γ2 (F )
γ2 (R)γ4 (F )
γ2 (F )
γ2 (R)γ3 (F )
/
/
f3
fr∩f3 +f4
//
/
/
f2
fr+f4
(observe that, the right hand horizontal map
f2
fr+f3
//
D(4, fr)
γ2 (R)γ4 (F )
general, surjective). The generalized dimension quotient
with the first derived functor of the symmetric square:
D(3, fr)
= L1 SP2 (Gab )
γ2 (R)γ3 (F )
D(3, fr)
γ2 (R)γ3 (F )
D(3, fr)
γ2 (R)γ3 (F ) is
→
is not, in
identified
14
ROMAN MIKHAILOV AND INDER BIR S. PASSI
(see (3.1)). By Lemmas 6 and 11, the dimension quotient
D(4, fr ∩ f3 )
D(4, fr ∩ f3 )
=
(γ2 (R) ∩ γ3 (F ))γ4 (F )
[R ∩ γ2 (F ), R]γ4 (F )
is a G-representation. Therefore, by Lemma 6, using the upper horizontal exact
fr)
sequence in the diagram (4.3), we conclude that γ2D(4,
(R)γ4 (F ) is a G-representation,
and
D(4, fr)
D(4, fr)
lim
=
.
←− γ2 (R)γ4 (F )
γ2 (R)γ4 (F )
Looking at the left hand vertical epimorphism in the diagram (4.1), we conclude
that the natural map
lim
←−
D4 (G)
γ2 (F )
D(4, fr)
= lim
→
←− γ2 (R)γ4 (F )
γ2 (R)γ4 (F )
γ4 (G)
1 R∩D(4, fr)
4 (G)
is an epimorphism, that is, the map D
in (4.2) is zero and
γ4 (G) → lim
←− γ2 (R)γ4 (F )
the asserted short exact sequence follows, and the proof is complete.
5. Proof of theorem 2
(a) Let us set G(2, R) := [R ∩ γ2 (F ), R ∩ γ2 (F )]γ3 (R). Since F (2, R) = G(2, R),
F (3, R) is a subgroup of G(2, R). Observe that, in view of (3.17), we have
γ2 (R ∩ γ2 (F )) ∩ γ3 (R) = [γ2 (R), R ∩ γ2 (F )],
and
γ2 (R ∩ γ3 (F )) ∩ γ3 (R) = [γ2 (R) ∩ γ3 (F ), R ∩ γ3 (F )] =
[[R ∩ γ2 (F ), R], R ∩ γ3 (F )] ⊆ [[R ∩ γ2 (F ), R], R ∩ γ2 (F )].
We have the following natural diagram with exact rows and columns
(5.1)
F (3, R)
G(3, R)
/
γ2 (R∩γ2 (F ))∩(1+rf3 +r3 Z[F ])
γ2 (R∩γ3 (F ))[γ2 (R), R∩γ2 (F )]
γ3 (R)
[R∩γ2 (F ), R, R∩γ2 (F )]γ4 (R)
/
G(2, R)
G(3, R)
/
//
γ2 (R∩γ2 (F ))
γ2 (R∩γ3 (F ))[γ2 (R), R∩γ2 (F )]
r3 Z[F ]
r3 Z[F ]∩rf3
rf2
rf3
rf2
rf3 +r3 Z[F ]
/
/
//
Let us set
I :=
R ∩ γ2 (F )
,
γ2 (R)
J :=
R ∩ γ3 (F )
.
γ2 (R) ∩ γ3 (F )
DIMENSION QUOTIENTS, FOX SUBGROUPS AND LIMITS OF FUNCTORS
15
Clearly J ⊂ I ⊂ R/γ2 (R). Observe that
γ2 (R ∩ γ2 (F ))
Λ2 (I)
=
.
Λ2 (J)
γ2 (R ∩ γ3 (F ))[γ2 (R), R ∩ γ2 (F )]
Invoking the exact sequence (2.3), we have the following sequence
L1 SP2 (I/J) ֒→
Λ2 (I)
→ I ⊗ I/J.
Λ2 (J)
There is a natural isomorphism
rZ[F ]
f2
rf2
=
⊗
rf
f3 + r2
rf3 + r3 Z[F ]
[Observe that, rf3 + r3 = rf3 + r3 Z[F ], and we can omit the last Z[F ] term.] Recall
that D(3, r2 ) = γ2 (R)γ3 (F ) (see [18]). Hence, there are monomorphisms
I ֒→ R/γ2 (R) =
rZ[F ]
,
rf
I/J ֒→
Since both R/γ2 (R) and Coker(I ֒→ R/γ2 (R)) =
that there is a natural monomorphism
f2
.
f3 + r2
R
R∩γ2 (F )
are free abelian, we see
rf2
.
rf + r3
The sequence (2.3) implies the following diagram
I ⊗ I/J ֒→
L1 SP2 (I/J) /
/
3
Λ2 (I)
Λ2 (J)
/ I ⊗ I/J
γ2 (R∩γ2 (F ))
γ2 (R∩γ3 (F ))[γ2 (R), R∩γ2 (F )]
/
rf2
rf3 +r3
We thus obtain the following identification:
γ2 (R ∩ γ2 (F )) ∩ (1 + rf3 + r3 )
= L1 SP2 (I/J).
γ2 (R ∩ γ3 (F ))[γ2 (R), R ∩ γ2 (F )]
There is a simple way to pick representatives of L1 SP2 (I/J) in γ2 (R ∩ γ2 (F )),
which also follows from the sequence (2.3). The subgroup L1 SP2 (I/J) is generated
by elements
[x, y]m , x, y ∈ R ∩ γ2 (F ), xm , y m ∈ (R ∩ γ3 (F ))γ2 (R).
One can easily check that, for such a pair x, y,
[x, y]m − 1 ∈ rf3 + r3 .
We assert that the horizontal arrow (let us call it q) in the diagram (5.1)
F (3, R)
γ2 (R ∩ γ2 (F )) ∩ (1 + rf3 + r3 )
→
G(3, R)
γ2 (R ∩ γ3 (F ))[γ2 (R), R ∩ γ2 (F )]
is an epimorphism. For, let
q:
xm = rx sx , rx ∈ R ∩ γ3 (F ), sx ∈ γ2 (R),
y m = ry sy , ry ∈ R ∩ γ3 (F ), sy ∈ γ2 (R).
16
ROMAN MIKHAILOV AND INDER BIR S. PASSI
Consider the element
w := [x, y]m [x, sy ]−1 [y, sx ].
Clearly, q(w) = [x, y]m , since [x, sy ]−1 [y, sx ] ∈ γ3 (R). Working modulo rf3 , we
have
w − 1 ≡ m(x − 1)(y − 1) − m(y − 1)(x − 1) − ([x, sy ] − 1) + ([y, sx ] − 1)
≡ (x − 1)(y m − 1) − (y − 1)(xm − 1) − ([x, sy ] − 1) + ([y, sx ] − 1)
≡ (x − 1)(sy − 1) − (y − 1)(sx − 1) − ([x, sy ] − 1) + ([y, sx ] − 1)
≡ (sy − 1)(x − 1) − (sx − 1)(y − 1) ≡ 0,
i.e., w ∈ F (3, R). Hence, the map q is an epimorphism.
We next show that the left hand vertical map in the diagram (5.1), namely,
r3 Z[F ]
γ3 (R)
→ 3
[[R ∩ γ2 (F ), R], R ∩ γ2 (F )]γ4 (R)
r Z[F ] ∩ rf3
is a monomorphism, i.e.,
γ3 (R) ∩ (1 + r3 Z[F ] ∩ rf3 ) = [[R ∩ γ2 (F ), R], R ∩ γ2 (F )]γ4 (R).
Clearly,
γ3 (R) ∩ (1 + r3 Z[F ] ∩ rf3 ) = γ3 (R) ∩ (1 + r3 ∩ rf3 ).
Next we will use the following identification (see [11]):
r2 ∩ f3 = r3 + r∆(R ∩ γ2 (F )) + ∆(γ2 (R) ∩ γ3 (F )).
where ∆(H) denotes the augmentation ideal of the group ring Z[H]. We have
r3 ∩ rf3 = r(r2 ∩ f3 )
(recall the quote preceding Lemma 7)
= r(r∆(R ∩ γ2 (F )) + ∆(R ∩ γ2 (F ))r + ∆(γ2 (R) ∩ γ3 (F )) + r3 )
= r(r∆(R ∩ γ2 (F )) + ∆(R ∩ γ2 (F ))r + ∆([R ∩ γ2 (F ), R]) + r3 )
= r∆(R ∩ γ2 (F ))r + r2 ∆(R ∩ γ2 (F )) + r4 ,
Corollary 8 implies that the right hand vertical arrow in (5.1) is a monomorphism.
Hence
R ∩ γ2 (F )
F (3, R)
≃ L1 SP2 (I/J) = L1 SP2
G(3, R)
γ2 (R)(R ∩ γ3 (F ))
and the proof part (a) of Theorem 2 is complete.
Part (b) follows from the part (a) together with the lifting of elements from L1 SP2 (I/J)
described above.
6. Proof of theorem 3
By Theorem 2,
F (3, R)
= lim L1 SP2
lim
←−
←− G(3, R)
R ∩ γ2 (F )
γ2 (R)(R ∩ γ3 (F ))
.
DIMENSION QUOTIENTS, FOX SUBGROUPS AND LIMITS OF FUNCTORS
17
In order to study right hand limit, consider the following diagram with exact rows
and columns
(6.1)
L1 SP2 (Gab )
L1 SP2 (Gab )
R∩γ2 (F )
γ2 (R)(R∩γ3 (F ))
/
/
γ2 (F )
γ2 (R)γ3 (F )
/ / γ2 (G)/γ3 (G)
/ / g2 /g3
r∩f2 +f3
fr+f3
/
/
f2
fr+f3
SP2 (Gab )
SP2 (Gab ).
The middle vertical sequence in (6.1) is the sequence (2.3) (for I = Fab , J =
R/(R ∩ γ2 (F ))):
L1 SP2 (Gab ) /
/
Λ2 (Fab )
Λ2 (R/(R∩γ2 (F )))
/ Fab ⊗ Gab
γ2 (F )
γ2 (R)γ3 (F )
/
/ / SP2 (Gab )
f2
fr+f3
The right hand vertical map γ2 (G)/γ3 (G) → g2 /g3 is a monomorphism, since
D3 (G) = γ3 (G). Let us set
K :=
R ∩ γ2 (F )
rZ[F ] ∩ f2 + f3
.
, L :=
γ2 (R)(R ∩ γ3 (F ))
fr + f3
Then we have the following short exact sequence:
0 → L1 SP2 (Gab ) → K → L → 0.
Observe that, the monomorphism L1 SP2 (Gab ) ֒→ K implies that the induced map
Tor(L1 SP2 (Gab ), L1 SP2 (Gab )) → Tor(K, K)
is also a monomorphism. Therefore, the induced maps
L1 SP2 (L1 SP2 (Gab )) → L1 SP2 (K),
L1 Λ2 (L1 SP2 (Gab )) → L1 Λ2 (K)
are also monomorphisms, since L1 SP2 and L1 Λ2 are subfunctors of the Tor-functor.
Since the functor Tor(−, Z/2) is left-exact, the sequence (2.1) implies that the
induced map
˜ 2 (L1 SP2 (Gab )) → L1 ⊗
˜ 2 (K)
L1 ⊗
18
ROMAN MIKHAILOV AND INDER BIR S. PASSI
is a monomorphism. Now the sequence (2.2) implies that we have the inclusion
Tor(K, K)
L1 SP2 (K)
֒→
.
2
2
2
L1 SP (L1 SP (Gab ))
Tor(L1 SP (Gab ), L1 SP2 (Gab ))
We assert that
(6.2)
lim
←−
Tor(K, K)
= 0.
Tor(L1 SP (Gab ), L1 SP2 (Gab ))
2
The natural consequence of the above vanishing of lim is that we have the isomor←−
phism
L1 SP2 (L1 SP2 (Gab )) = lim L1 SP2 (K)
←−
and Theorem 3 will follow.
It remains to establish the vanishing result (6.2). Observe that we have an exact
sequence
0 → Tor(L1 SP2 (Gab ), L1 SP2 (Gab )) → Tor(L1 SP2 (Gab ), K) → Tor(L1 SP2 (Gab ), L).
Since L ⊂
f2
rf+f3
= Gab ⊗ Fab , there is a natural inclusion
Tor(L1 SP2 (Gab ), L) ֒→ Tor(L1 SP2 (Gab ), Gab ⊗ Fab ).
The short exact sequence
R
0→
⊗ Fab → Fab ⊗ Fab → Gab ⊗ Fab → 0,
R ∩ γ2 (F )
on tensoring with L1 SP2 (Gab ), gives an inclusion
Tor(L1 SP2 (Gab ), Gab ⊗ Fab ) ֒→ Tor(L1 SP2 (Gab ), Gab ⊗ Fab ) ⊗
R
⊗ Fab .
R ∩ γ2 (F )
Since the representation Tor(L1 SP2 (Gab ), Gab ⊗ Fab ) ⊗ R∩γR2 (F ) ⊗ Fab is monoadditive [one can easily check that, for any representation F , the representation
F ⊗ Fab is mono-additive], it follows that
lim Tor(L1 SP2 (Gab ), L) = 0
←−
and consequently
Tor(L1 SP2 (Gab ), L1 SP2 (Gab )) = lim Tor(L1 SP2 (Gab ), K).
←−
A similar argument shows that lim Tor(L, K) = 0 and hence
←−
lim Tor(L1 SP2 (Gab ), K) = lim Tor(K, K).
←−
1
2
←−
2
Since lim Tor(L1 SP (Gab ), L1 SP (Gab )) = 0, we conclude (6.2). ✷
←−
Remark. Theorem 2 implies that there is a natural exact sequence
0 → L1 SP2 (K) →
G(2, R)
rf2
→ 3,
G(3, R)
rf
The quotient
rZ[F ] f2
rf2
⊗ 3
3 ≃
rf
f
rf
DIMENSION QUOTIENTS, FOX SUBGROUPS AND LIMITS OF FUNCTORS
19
is a mono-additive representation. Hence,
lim
←−
G(2, R)
= L1 SP2 (L1 SP2 (Gab )).
G(3, R)
Remark. The middle vertical sequence in (6.1) implies that, if Gab is torsion-free,
γ2 (F )
R∩γ2 (F )
then the group γ2 (R)γ
is torsion-free as well. Since the quotient γ2 (R)(R∩γ
3 (F )
3 (F ))
γ2 (F )
is a subgroup of γ2 (R)γ
, we conclude that, if Gab is torsion-free, then F (3, R) =
3 (F )
G(3, R), by Theorem 2.
Acknowledgement
The research is supported by the Russian Science Foundation grant N 16-1110073. The authors are thankful to Harish-Chandra Research Institute, Allahabad,
for the warm hospitality provided to them during their visit in February 2017.
References
[1] Joan S. Birman: Braids, Links and Mapping Class Groups, Annales of Mathematics Studies
Vol. 82, Princeton Univ. Press, 1974.
[2] L. Breen: On the functorial homology of abelian groups, J. Pure Appl. Alg. 142 (1999),
199–237.
[3] P. M. Cohn: Generalizaion of a theorem of Mangnus, Proc. London Math. Soc., 2 (1952),
297–310.
[4] A. Dold and D. Puppe: Homologie nicht additiver Funktoren Anwendugen, Annales de
l’institut Fourier, tome 11, no . 6 (1961), 201–312.
[5] I. Emmanouil and R. Mikhailov: A limit approach to group homology, J. Algebra 319 (2008),
1450–1461.
[6] R. Fox: Free differential calculus I: Derivation in the free group ring, Ann. Math. 57 (1953),
547–560.
[7] Karl W. Gruenberg: Cohomological Topics in Group Theory, LNM Vol. 143, Springer-Verlag,
1970.
[8] Narain Gupta: Free Group Rings, Contemporary Mathematics, Vol. 66, American Mathematical Society, 1987.
[9] M. Hartl, R. Mikhailov and I. B. S. Passi: Dimension quotients, J. Indian Math. Soc., New
Ser. Spec. Centenary Vol., 63–107 (2007).
[10] F. Jean: Foncteurs derives de lalgebre symetrique: Application au calcul de certains groupes
dhomologie fonctorielle des espaces K(B, n), Doctoral thesis, University of Paris 13, 2002,
available at: http://homepages.abdn.ac.uk/mth192/pages/html/archive/jean.html
[11] R. Karan, D. Kumar and L. Vermani: Some intersection theorems and subgroups determined
by certain ideals in integral group rings II, Algebra Colloq. 9 (2002), 135-142.
[12] B. Köck: Computing the homology of Koszul complexes, Trans. Amer. Math. Soc., 353
(2001), 3115–3147.
[13] G. Losey: On dimension subgroups, Trans. Amer. Math. Soc., 97 (1960), 474–486.
[14] S. Mac Lane: Categories for The Working Mathematicians, Secon Edition, Springer, 1998.
[15] W. Magnus: Beziehngen zwischen Gruppen in einem speziellen Ring, Math. Ann., 111 (1935),
259–280.
[16] W. Magnus: Über Beziehungen zwischen höheren Kommutatoren, J. reine angew. Math.,
177 (1937), 105–115.
[17] Roman Mikhailov and Inder Bir S. Passi: Lower Central and Dimension Series of Groups,
LNM Vol. 1952, Springer 2009.
[18] R. Mikhailov and I. B. S. Passi: Generalized dimension subgroups and derived functors, J.
Pure Appl. Algebra 220 (2016), 2143–2163.
[19] R. Mikhailov and I. B. S. Passi: Free group rings and derived functors, arXiv:1605.08196.
20
ROMAN MIKHAILOV AND INDER BIR S. PASSI
[20] E. Rips: On the fourth integer dimension subgroup, Israel J. Math. 12 (1972), 342–346.
[21] S. Ivanov and R. Mikhailov: A higher limit approach to homology theories, J. Pure and Appl.
Algebra 219 (2015), 1915–1939.
[22] I. B. S. Passi: Dimension subgroups, J. Algebra, 9 (1968), 152–182.
[23] I. B. S. Passi: Group Rings and Their Augmentation Ideals, LNM Vol. 715, Springer-Verlag,
1979,
[24] K.-I. Tahara: On the fourth dimension subgroups and polynomial maps, J. Algebra, 45
(1977), 102–131.
[25] K.-I. Tahara: On the structure of Q3 (G) and the fourth dimension subgroup, Japan J. Math.
(N.S.) 3 (1977), 381–396.
[26] D. Quillen: Cyclic cohomology and algebra extensions, K-Theory 3, (1989), 205–246.
[27] A. I. Yunus: On a problem of Fox, Soviet Math. Dokl., 30 (1984), 346–350.
[28] E. Witt: Treue Darstellung Liescher Ringe, J. reine angew. Math., 177 (1937), 152–160.
DIMENSION QUOTIENTS, FOX SUBGROUPS AND LIMITS OF FUNCTORS
21
Roman Mikhailov
St Petersburg Department of Steklov Mathematical Institute,
Chebyshev Laboratory, St Petersburg State University
14th Line, 29b, Saint Petersburg 199178 Russia
and
School of Mathematics, Tata Institute of Fundamental Research
Mumbai 400005, India
Email: [email protected]
Inder Bir S. Passi
Centre for Advanced Study in Mathematics
Panjab University, Sector 14, Chandigarh 160014 India
and
Indian Institute of Science Education and Research, Mohali (Punjab) 140306 India
Email: [email protected]
| 4math.GR
|
On the Lie bracket approximation approach to distributed
optimization: Extensions and limitations ∗
arXiv:1802.10519v1 [math.OC] 28 Feb 2018
Simon Michalowsky1 , Bahman Gharesifard2 , and Christian Ebenbauer1
1 Institute
for Systems Theory and Automatic Control, University of Stuttgart, Germany
{michalowsky,ce}@ist.uni-stuttgart.de
2 Department of Mathematics and Statistics, Queen’s University, Canada
[email protected]
to relax these assumptions.
In the present work we want to extend these results
and show that the approach can be applied to a large
class of optimization problems under mild assumptions
on the communication network. While in [11], [12], [13]
only optimization problems with linear constraints and
objective functions in the form of a sum of separable functions were considered, in the present work we enhance
the methodology to general convex optimization problems. The main idea of the approach is to use Lie bracket
approximation techniques to find distributed approximations of non-distributed saddle-point dynamics. While in
the previous works only certain Lie bracket approximations were used, we further show here that a whole class
is applicable.
Abstract. We consider the problem of solving a smooth
convex optimization problem with equality and inequality constraints in a distributed fashion. Assuming that
we have a group of agents available capable of communicating over a communication network described by
a time-invariant directed graph, we derive distributed
continuous-time agent dynamics that ensure convergence
to a neighborhood of the optimal solution of the optimization problem. Following the ideas introduced in our
previous work, we combine saddle-point dynamics with
Lie bracket approximation techniques. While the methodology was previously limited to linear constraints and
objective functions given by a sum of strictly convex separable functions, we extend these result here and show that
it applies to a very general class of optimization problems
under mild assumptions on the communication topology.
1 INTRODUCTION
2 PRELIMINARIES
Over the last decades, distributed optimization has been
an active area of research with high practical relevance,
see, e.g., [2, 3, 4] for applications. In these type of problems, the goal is to cooperatively solve an optimization
problem using a group of agents communicating over a
network. While discrete-time algorithms for distributed
optimization constitute the majority in the existing literature, we focus on continuous-time algorithms which have
regained interest in the last decades, [5, 6, 7, 8, 9, 10]. These
algorithms require strong assumptions either on the structure of the optimization problem or on the communication
network. Recently, a novel approach to continuous-time
distributed optimization based on Lie bracket approximations has been proposed in [11], [12] that has the potential
*
2.1 Notation
We denote by Rn the set of n-dimensional real vectors and
further write C p , p ∈ N, for the set of p-times continuously
differentiable real-valued functions. The gradient of a
function f : Rn → R, f ∈ C 1 , with respect to its argument
x ∈ Rn , will be denoted by ∇ x f : Rn → Rn ; we often omit
the subscript, if it is clear from the context. We denote the
(i, j)th entry of a matrix A ∈ Rn×m by aij , and sometimes
denote A by A = [ aij ]. We use ei to denote the vector with
the ith entry equal to 1 and all other entries equal to 0. We
do not specify the dimension of ei but expect it to be clear
from the context. For a vector λ ∈ Rn we let diag(λ) ∈
Rn×n denote the diagonal matrix whose diagonal entries
are the entries of λ. Given two continuously differentiable
vector fields f 1 : Rn → Rn and f 2 : Rn → Rn , the Lie
This article is an extended version of [1] additionally including proofs
for all results, a discussion of the choice of vector fields after Remark 1
and the overview in Table 1.
1
bracket of f 1 and f 2 evaluated at x is defined to be
[ f 1 , f 2 ]( x ) :=
∂ f2
∂f
( x ) f 1 ( x ) − 1 ( x ) f 2 ( x ).
∂x
∂x
set of (4) is non-empty and that the problem has a unique
solution.
Our goal is to design continuous-time optimization algorithms that converge to an arbitrarily small neighborhood of the unique global optimizer of (4) and that can be
implemented in a distributed fashion. More precisely, we
assume that we have a group of n agents available, each capable of interchanging information over a communication
network described by a directed graph G = (V , E ) with
graph Laplacian G = [ gij ], where V = {v1 , v2 , . . . , vn } is
a set of n nodes and E ⊆ V × V is the edge set between
the nodes. In the present setup, each node vi represents
an agent and the edges define the existing communication links between the agents, i.e., if there is an edge from
node i to node j then agent i has access to the information
provided by agent j. We then say that an algorithm is
distributed if each agent only uses its own information as
well as that provided by its out-neighboring agents.
Let L : Rn × Rneq × Rnineq → R, neq = ∑i∈Ieq neq i ,
nineq = ∑i∈Iineq nineq i , denote the Lagrangian associated
to (4), i.e.,
(1)
With a slight abuse of notation we sometimes also write
[ f 1 ( x ), f 2 ( x )]. For a vector x = [ x1 , x2 , . . . , xn ]> ∈ Rn and
a finite set S ⊆ {1, 2, . . . , n}, we denote by xS ∈ R|S| the
ordered stacked vector of all xi with i ∈ S . For example, if
n = 5 and S = {2, 4}, then xS = [ x2 , x4 ]> .
2.2 Basics on graph theory
We recall some basic notions on graph theory, and refer
the reader to [14] or other standard references for more
information. A directed graph (or simply digraph) is an
ordered pair G = (V , E ), where V = {v1 , v2 , . . . , vn } is
the set of nodes and E ⊆ V × V is the set of edges, i.e.
(vi , v j ) ∈ E if there is an edge from node vi to v j . In our
setup the edges encode to which other agents some agent
has access to, i.e. (vi , v j ) ∈ E means that node vi receives
information from node v j . We say that node v j is an outneighbor of node vi if there is an edge from node vi to node
v j . The adjacency matrix A = [aij ] ∈ Rn×n associated to
G is defined as
(
1 if i 6= j and (vi , v j ) ∈ E ,
(2)
aij =
0 otherwise.
L( x, ν, λ) = F ( x ) + ν> a( x ) + λ> c( x ),
where a = [ ai ]i∈Ieq , c = [ci ]i∈Iineq are the stacked vectors of
all ai and ci , respectively, and ν ∈ Rneq , λ ∈ Rnineq are the
associated Lagrange multipliers. In the sequel, we assume
that the state of the ith agent comprises of xi as well as
the dual variables associated to the constraints ai , ci , given
they exist. Without loss of generality we then put the
following assumption on the indexing of the constraints:
We also define the out-degree matrix D = [dij ] associated
to G as
(
if i = j
∑nk=1 aik
dij =
(3)
0
otherwise.
Assumption 1. For any i ∈ Ieq and any j ∈ Iineq , there
exists an x = [ x1 , x2 , . . . , xn ]> ∈ Rn such that
and
Finally, we call G = D − A = [ gij ] ∈ Rn×n the Laplacian of
G . A directed path in G is a sequence of nodes connected
by edges and we write pi1 ir = hvi1 |vi2 | . . . |vir i for a path
from node vi1 to node vir . We say that a path is simple if
the sequence contains no node more than once.
i ∈ Ieq ⊆ {1, 2, . . . , n},
ci ( x ) ≤ 0,
i ∈ Iineq ⊆ {1, 2, . . . , n},
(6)
The following saddle-point dynamics adapted from [15]
is known to converge to a saddle point of the Lagrangian
(see [13] for a proof), thus providing a solution to (4)
F(x)
ai ( x ) = 0,
6= 0
6= 0.
L( x ? , ν, λ) ≤ L( x ? , ν? , λ? ) ≤ L( x, ν? , λ? ).
Consider the following convex optimization problem
s.t.
∂c j
∂x j ( x )
∂ai
∂xi ( x )
In a nutshell, this assumption Newguarantees that the
ith constraints ai , ci are both functions of xi .
It is well-known that if the Lagrangian has Newa saddle
point ( x ? , ν? , λ? ), then x ? is an optimizer of (4). We say
that a point ( x ? , ν? , λ? ) is a saddle point if for all x ∈ Rn ,
nineq
ν ∈ Rneq , λ ∈ R+ , we have
3 PROBLEM SETUP
min
(5)
ẋ = −∇ x L( x, ν, λ)
(4)
= −∇ F ( x ) −
>
>
∂a
∂c
∂x ( x ) ν − ∂x ( x ) λ
(7a)
ν̇ = ∇ν L( x, ν, λ)
where F : Rn → R, ai : Rn → Rneq i , ci : Rn → Rnineq i ,
neq i , nineq i ≥ 1, F ∈ C 2 is strictly convex, the functions
ai , i ∈ Ieq , are affine and the ci ∈ C 2 , i ∈ Iineq , are convex.
We assume further that F, ai , ci are such that the feasible
= a( x )
(7b)
λ̇ = diag(λ)∇λ L( x, ν, λ)
= diag(λ)c( x ).
2
(7c)
notation, for each i = 1, 2, . . . , n, we define the index set
However, (7) is in general not distributed in the aforementioned sense, since the right-hand side of (7) is not
composed only of admissible vector fields, i.e., vector fields
that can be computed locally by the nodes. For example,
if g12 6= 0 and g13 = 0, then the vector field [ e1> x2 , 0, 0]> is
admissible for (7), while [ e1> x3 , 0, 0]> is not.
Recently, a novel approach to distributed optimization
has been proposed that employs Lie bracket approximation techniques to derive distributed approximations
of (7). The idea is to write the right-hand side of (7) by
means of Lie brackets of admissible vector fields. If we
have achieved to rewrite (7) in this form, i.e., we have
ż =
∑
v B B ( z ),
i −1
Ī(i ) := {i } ∪ { j = n +
∑ neq k + `, ` = 1, 2, . . . , neq i }
k =1
i −1
∪ { j = n + neq +
∑ nineq k + `, ` = 1, 2, . . . , nineq i }
(11)
k =1
associating the components of the complete state z to the
ith agent meaning that z j is part of the state of agent i for
all j ∈ Ī(i ). Hence, the state vector of the ith agent is given
by zĪ(i) . Based on this, for all j = 1, 2, . . . , N, we define
I( j) = Ī(i )
(8)
B∈B
(12)
for some i such that j ∈ Ī(i ), i.e., I( j) is the set of all
indices which are associated to the same agent as the jth
index. Note that Ī( j) = I( j), for all j = 1, 2, . . . , n.
[ x > , ν> , λ> ]> ,
where z =
v B ∈ R and B is a set of
Lie brackets of admissible vector fields {φ1 , φ2 , . . . , φ M },
φi : Rn+neq +nineq → Rn+neq +nineq , then we can employ Lie
bracket approximation techniques from [16], [17] to derive
distributed approximations of (8). More precisely, we can
find a family of functions uiσ : R → R parametrized by
σ > 0 such that the trajectories of
4.1 Lie brackets of admissible vector fields
In the following, we first want to discuss which kind of
vector fields can be written in terms of Lie brackets of
admissible vector fields. For i, j = 1, 2, . . . , N, define
M
żσ =
∑ φi (z)uiσ (t)
hi,j (z) = e j f j (zI( j) , zI(i) ),
(9)
(13)
i =1
where f j : R|I( j)| × R|I(i)| → R, f j ∈ C 1 . Observe that hi,j
is admissible if and only if there exist `, k ∈ {1, 2, . . . , n}
such that i ∈ I(`), j ∈ I(k) and gk` 6= 0. In the next
Newresult we consider Lie brackets of admissible vector
fields of the form (13).
uniformly converge to those of (8) as σ increases, given (8)
and (9) are initialized equally. A general algorithm to
compute suitable functions uiσ is presented in [16] and
we will present a modified version thereof tailored to the
problem at hand in [13]. In the present paper we do not
discuss the second step of how to design these functions
uiσ but focus on rewriting the right-hand side of (7) in
terms of Lie brackets of admissible vector fields.
Lemma 1. Consider a graph G of n nodes and let
pi1 ir = hvi1 |vi2 | . . . |vir i denote a simple path in G from
vi1 to vir . Then, for any jk ∈ I(ik ), k = 1, 2, . . . , r, we have
4 MAIN RESULTS
r −2 ∂ f
jk
∂z j
k +1
k =1
= e j1 f jr−1 (zI( jr−1 ) , zI( jr ) ) ∏
Consider the saddle-point dynamics (7) and observe
that the right-hand side is a sum of vector fields of the
form ei ψ( x ) and ei z j ψ( x ), i, j = 1, 2, . . . , n + neq + nineq ,
ψ : Rn → R, where
z = [ x > , ν> , λ> ]> ∈ R N ,
N = n + neq + nineq ,
h
i
h jr ,jr−1 , h jr−1 ,jr−2 , . . . , [h j3 ,j2 , h j2 ,j1 ] . . .
(z)
(14)
(zI( jk ) , zI( jk+1 ) )
and the left-hand side is a Lie bracket of admissible vector
fields.
•
A proof is given in Section 7.1. By the above Lemma,
each non-admissible vector field that takes the same form
as the right-hand side of (14) can be written in terms of a
Lie bracket of admissible vector fields. It is worth mentioning that this does not classify the whole set of vector fields
that can be written as a Lie bracket of admissible vector
fields since we limited ourselves to a single path. We next
discuss a special case that is of particular importance for
the application at hand.
(10)
is the complete state and ei ∈ R N is the ith unit vector. These vector fields might either be admissible or not,
depending on the communication graph as well as the
problem structure. In the following, we wish to discuss
how to write vector fields of this form by means of Lie
brackets of admissible vector fields. For the purpose of
3
Proposition 1. Consider a graph G of n nodes and let pi1 ir =
hvi1 |vi2 | . . . |vir i denote a simple path in G from vi1 to vir . Let
jk ∈ I(ik ), k = 1, 2, . . . , r, be any set of indices and suppose
that
(1)
(2)
k
k
f jk (zI( jk ) , zI( jk+1 ) ) = f j (zI( jk ) ) f j (zI( jk+1 ) )
for k = 2, . . . , r − 1; hence
(1)
k
(2)
k −1
∂z j
k
∂ fj
(zI( jk ) ) = 1
(15)
|
=
(2)
f j (zI( jk+1 ) ) = z jk+1 ,
k
(17)
k = 2, 3, . . . , r − 1,
(18a)
k = 1, 2, . . . , r − 2;
(18b)
As we see from the proof, each bounded vector field
in (17) leads to another unbounded vector field. Hence,
at most half of the vector fields h jk ,jk−1 in (17) can be
bounded. In particular, we can choose the functions f jk ,
k = 1, 2, . . . , r − 1, as follows to guarantee that (16) holds:
(
(1)
f j (x)
k
1
1
cos(α( x −d))
if k is even,
if k is odd, k 6= 1,
sin α( x − d) if k is even, k 6= r − 1,
∫ cos(α(1x−d)) dx if k is odd, k 6= r − 1,
(21a)
(21b)
Example 1. Consider the graph shown in Fig. 1 and assume for the sake of simplicity that I( j) = j for j = 1, . . . , 5
and z ∈ R5 . By (14) we can write non-admissible vector
fields of the form e1 ϕ1 (z2 ) ϕ2 (z3 ) as well as e1 ϕ3 (z4 ) ϕ4 (z5 )
and sums thereof in terms of Lie brackets of admissible
vector fields as long as ϕ1 , ϕ3 admit an analytic expression
of their antiderivatives ∫ ϕ1 (z2 )dz2 , ∫ ϕ3 (z4 )dz4 . In fact,
with
(1)
(1)
(zI( j ) )
k
k
=
cos(α x − d)
where α 6= 0, d ∈ R. However, this choice will lead to
functions f jk that are not globally continuous but only
π
π
); hence we
well-defined in the interval (d − 2α
, d + 2α
need to choose d appropriately and α sufficiently small.
By that choice, all functions f jk with even k are bounded
while all functions f jk with odd k are unbounded.
We next discuss how we can make use of the previous results to rewrite more general vector fields. While
Lemma 1 enables us to write products of functions of variables of nodes which lie on the same path as Lie brackets
of admissible vector fields, we cannot directly use this result to rewrite functions which do not fulfill this property.
To make this clearer, consider the following example:
all k = 1, 2, . . . , r. By (15), f j is also bounded for all
k
k = 1, 2, . . . , r since I( jk ) and I( jk+1 ) are disjunct. However, by (16),
fj
=
(
(2)
f j (x)
k
Proof. Suppose there exists a set of bounded vector fields
such that (17) holds. Then f jk is a bounded function for
(zI( jk ) ) =
∈ C 1.
Remark 1. The same holds true if we do not assume the
structure (15). In fact, the structure is required for all other
variables except of zI( j1 ) , zI( jr ) to cancel out in (17).
•
Lemma 2. Suppose that all assumptions from Proposition 1 are fulfilled. Then there exists no set of bounded
vector fields h jk ,jk−1 ∈ C 1 such that (17) holds.
•
(2)
k −1
∂z j
k
k −1
(2)
hence leading to h jk+1 ,jk (zI( jk ) , zI( jk+1 ) ) = e jk z jk+1 . However, in view of (9) where the φi are given by admissible
vector fields of the form (13), it is often desired that the
admissible vector fields have certain properties such as
boundedness in order to simplify the calculation of the approximating inputs uiσ and improve the transient behavior
of (9). However, as we see next, it is not possible to render
all vector fields bounded.
∂ fj
(2)
, fj
Thus, f j is strictly monotone in z jk which contradicts the
k −1
boundedness assumption, thus concluding the proof.
A proof is given in Section 7.2. In view of (14), the
constraint (16) ensures that all terms depending on zI( jk ) ,
k = 2, 3, . . . , r − 1, cancel out. Equation (17) is of particular
interest since the non-admissible vector fields often take
the form as its right-hand side. According to (16), there
exists a whole class of vector fields h jk ,jk−1 , or equivalently,
functions f jk , such that (17) holds. A particularly simple
choice that has been utilized in the previous works [11],
[12] is to take
k
(20)
k −1
for all z ∈ R N and the left-hand side is a Lie bracket of admissible vector fields.
•
(1)
(zI( jk ) )| ≥ β.
(1)
(1)
(2)
e j1 f j (zI( j1 ) ) f j (zI( jr ) )
1
r −1
f j (zI( jk ) ) = 1,
(2)
k −1
∂z j
k
∂ fj
Note that h jk ,jk−1 ∈ C 1 if and only if f j
(16)
R|I( jk )| ,
for all zI( jk ) ∈
k = 2, 3, . . . , r − 1. Then
h
i
(z)
h jr ,jr−1 , h jr−1 ,jr−2 , . . . , [h j3 ,j2 , h j2 ,j1 ] . . .
is bounded away from
zero, i.e., there exists some constant β > 0 such that for all
zI( jk ) ∈ R|I( jk )| we have
with
f j (zI( jk ) )
(2)
k −1
∂z j
k
∂ fj
(19)
4
h2,1 (z) = e1 ∫ ϕ1 (z2 )dz2 ,
h3,2 (z) = e2 ϕ2 (z3 ),
(22a)
h4,3 (z) = e1 ∫ ϕ3 (z4 )dz4 ,
h5,4 (z) = e4 ϕ4 (z5 ),
(22b)
z2
z3
2
3
and observe that for all z ∈ R N we have
[ψ2 , ψ1 ](z) = e1 ϕ2 (z3 ) ϕ4 (z5 ).
z1 1
(27)
Further, following Proposition 1 and choosing
5
z5
4
z4
h2,1 (z) = e1 z1 ,
h3,2 (z) = e2 ϕ2 (z3 ),
(28a)
h4,3 (z) = e1 ,
h5,4 (z) = e4 ϕ4 (z5 ),
(28b)
we have for all z ∈ R N
Figure 1. The graph considered in Example 1 where the ith agent
has an associated state zi .
ψ1 (z) = [h3,2 , h2,1 ](z),
ψ2 (z) = [h5,4 , h4,3 ](z).
(29)
Using this in (27) we finally managed to rewrite
e1 ϕ2 (z3 ) ϕ4 (z5 ) in terms of admissible vector fields.
•
we have for all z ∈ R5
e1 ϕ1 (z2 ) ϕ2 (z3 ) = [h3,2 , h2,1 ](z)
(23a)
e1 ϕ3 (z4 ) ϕ4 (z5 ) = [h5,4 , h3,2 ](z).
(23b)
Remark 2. Instead of realizing the multiplication by
means of Lie brackets another way is to augment the agent
state by estimates of the respective state of the other agent.
More precisely, in Example 1, we augment the state of
agent 1 by ξ 3 and ξ 5 that are estimates of z3 and z5 , respectively, and let
ż1
w ( z1 , z2 , z4 , ξ 3 , ξ 5 )
+ e2 z 5 + e3 z 5 ,
ξ̇ 3 =
−µξ 3
(30)
−µξ 5
ξ̇ 5
However, we cannot directly use (14) to rewrite a nonadmissible vector field of the form e1 ϕ2 (z3 ) ϕ4 (z5 ).
•
In the next result, we wish to overcome this limitation
and show how to not only write sums of the vector fields
in the form of the right-hand side of (14) in terms of Lie
brackets of admissible vector fields, but also products
thereof.
where w : R5 → R and µ > 0 is sufficiently large, hence
ξ 3 (t) ≈ z3 (t), ξ 5 (t) ≈ z5 (t). The resulting non-admissible
vector fields in the complete augmented system can then
be written in terms of Lie brackets of admissible vector
fields using Proposition 1. However, in the application
at hand this alters the saddle-point dynamics (7) which
necessitates a stability analysis of the augmented system.
Proposition 2. Let ηk : R → R, ηk ∈ C 1 , k = 1, 2, . . . , m,
m ≥ 2, and define
ψ1 (z) = ei η1 (z j1 ) ∫ η0 (zi )dzi ,
ψk (z) = ei zi ηk (z jk ),
(24a)
k = 2, . . . , m − 1,
ψm (z) = ei ηm (z jm ),
(24b)
(24c)
i, jk ∈ N, ψk : R N → R N . Then, for any jk 6= i,
k = 1, . . . , m, we have for all z ∈ R N
h
i
. . . [ψm , ψm−1 ], ψm−2 , . . . , ψ1 (z)
m
= ei η0 (zi ) ∏ ηk (z jk ).
k =1
Hence, under suitable assumptions on the communication graph, this allows us to write vector fields whose
components are sums of products of arbitrary functions
in terms of Lie brackets of admissible vector fields. This
observation gives rise to the next Lemma.
(25)
Lemma 3. Consider a strongly connected graph G of n
nodes and let ϕ : R N → R be an analytic function. Then
any vector field ψ(z) = ei ϕ(z), ei ∈ R N , i = 1, 2, . . . , N,
can be written as a possibly infinite sum of Lie brackets of
admissible vector fields.
•
•
A proof is given in Section 7.3. The vector fields ψk
defined by (24) are in general non-admissible; thus, the
left-hand side of (25) is not a Lie bracket of admissible
vector fields. However, observing that the vector fields
ψk take the same form as the right-hand side of (17), we
can make use of Proposition 1 to write the left-hand side
of (25) as a Lie bracket of admissible vector fields as long
as, for any k ∈ {1, . . . , m}, there exists a path from the ith
node to the node that is associated to state z jk . We illustrate
that by means of Example 1.
Proof. Since ϕ is analytic, by a series expansion it can be
written as a possibly infinite sum of monomials of the
components z j of z, j = 1, 2, . . . , N. Using Proposition 2,
all these monomials can be written in terms of Lie brackets
of vector fields of the form (24). By strong connectivity
of G , all these vector fields can be written in terms of Lie
brackets of admissible vector fields, thus concluding the
proof.
Example 1. [continued] Reconsider Example 1 and suppose we want to rewrite the non-admissible vector field
e1 ϕ2 (z3 ) ϕ4 (z5 ). Following Proposition 2 we let
ψ1 (z) = e1 z1 ϕ2 (z3 ),
ψ2 (z) = e1 ϕ3 (z5 ),
While the result might be more of a theoretical nature
for the application at hand, it nevertheless shows that the
proposed approach in principle applies to a large class of
problems.
(26)
5
Vector field
Admissible if...
(`)
ei
(`)
∂Fi
∂xi
( xi ) ∏ Fj ( x j )
(`)
j∈J F
Rewritable if...
` = 1, 2, . . . , n F
i = 1, 2 . . . , n
gij 6= 0 for all j ∈ J F
∃ pij in G for all j ∈ J F
k ∈ Ieq
i = 1, 2 . . . , n
gik 6= 0 for all k : aki 6= 0
∃ pik in G for all k : aki 6= 0
k ∈ Iineq , ` = 1, 2, . . . , nck
i = 1, 2, . . . , n
gij 6= 0 for all j ∈ Jck
∃ pij in G for all j ∈ Jck
k ∈ Ieq
i = 1, 2 . . . , n
Assumption 2 holds
∃ pki in G for all aki 6= 0
k ∈ Iineq
` = 1, 2, . . . , nck
Assumption 2 holds
∃ pkj in G for all j ∈ Jck
(`)
(`)
j 6 =i
ei āki νk
(`)
ei
∂ck,k
∂xk
(`)
( xk ) ∏ ck,j ( x j )
(`)
j∈Jc
k
j6=k
en+k aki xi
(`)
en+neq +k λk ∏ ck,j ( x j )
(`)
k
j∈Jc
(`)
(`)
(`)
Table 1. An overview of all vector fields appearing in (31).
4.2 Distributed optimization via Lie brackets
functions, i.e.,
nF
F(x) =
In the sequel we apply the results from the last section
to the problem at hand and rewrite the saddle-point dynamics (7) by means of Lie brackets of admissible vector
fields. For the sake of a simpler notation we assume in the
following that neq i = 1 for all i ∈ Ieq and nineq i = 1
for all i ∈ Iineq . We further assume that each agent
has an associated equality and inequality constraint, i.e.,
Ieq = {1, 2, . . . , n}, Iineq = {1, 2, . . . , n}. This can
always be achieved by augmenting the optimization problem (4) by constraints that do not alter the feasible set. We
emphasize that this is not necessary for the methodology
to apply as we will illustrate in the example in Section 5.
We can then write the saddle-point dynamics (7) equivalently as
n
ẋ = − ∑ ei
i =1
ν̇ =
∑
∑
∑ āik νk + ∑
k ∈Ieq
∂ck
∂xk ( x ) λk
`=1
∑ āki xi + b̄k
(32)
(`)
(33)
ck,j ( x j ),
(`)
j∈Jc
k
(`)
(`)
where the Fj
: R → R, j ∈ J F ⊆ {1, 2, . . . , n},
` = 1, 2, . . . , n F , n F > 0, are strictly convex functions
(`)
(`)
and the ck,j : R → R, k ∈ Iineq , j ∈ Jck ⊆ {1, 2, . . . , n},
` = 1, 2, . . . , nci , nck > 0, are convex. Observe that,
(`)
(`)
if n F and nck are infinite and Fj , ck,j are monomials,
(`)
this includes all analytic functions F, ck . If J F
= {`},
(`)
Jck
= {`}, we obtain the particularly important special
case that both the objective function and the constraints
are a sum of separable functions; hence also the case of
linear constraints considered in [11], [12] is covered here.
Under this assumption the vector fields appearing in (31)
are summed up in Table 1. Depending on the communication graph as well as the structure of the constraints
and the objective function, these vector fields can either be
admissible or not. In particular, the vector fields in (31b),
(31c) are admissible if the constraints are compatible with
the communication topology defined by the graph G , i.e.,
if the following assumption holds:
(31a)
k ∈Iineq
(31b)
i =1
e k λ k c k ( x ),
∑ ∏
`=1
(`)
Fj ( x j ),
(`)
j∈J F
nck
ck ( x ) =
n
ek
k ∈Ieq
λ̇ =
∂F
∂xi ( x ) +
∑ ∏
(31c)
k ∈Iineq
Assumption 2. For all i, j = 1, 2, . . . , n with gij = 0 we
where ai ( x ) = āi x + b̄i , āi = [ āi1 , āi2 , . . . , āin ] ∈ R1×n ,
b̄i ∈ R. Motivated by our previous discussions, we assume
in the following that the objective function as well as the
inequality constraints are sums of products of separable
have
∂ai
∂x j ( x )
≡ 0 as well as
∂ci
∂x j ( x )
≡ 0.
•
We point out that all non-admissible vector fields in (31)
can be written in terms of Lie brackets of admissible vector
6
2
1
5
3
4
Figure 2. The communication
graph for the example considered in Section 5.
Vector field
Corresponding path
e5 ν2
p52 = hv5 |v1 |v2 i
e2 x 2 λ 1
p21 = hv2 |v5 |v1 i
e1 λ 3
p13 = hv1 |v2 |v3 i
e4 x 1 λ 4
p41 = hv4 |v5 |v1 i
e1 x 4 λ 4
p14 = hv1 |v2 |v3 |v4 i
e1 x 1 λ 4
p14 = hv1 |v2 |v3 |v4 i
e8 λ 3 x 1
p31 = hv3 |v4 |v5 |v1 i
e9 λ 4 x 4 x 1
p41 = hv4 |v5 |v1 i
e9 λ4 x12
p41 = hv4 |v5 |v1 i
Lie bracket representation
e j z 6 , e5 z j
e j z 7 , e2 z 2 z j
e j λ 3 , e1 z j
e j x 1 , e4 λ 4 z j
e j2 λ4 x4 , [ e j1 z j2 , e1 z j1 ]
e j λ4 , [ e j1 z j2 , e1 x1 z j1 ]
2
e j2 x1 , [ e j1 z j2 , e8 λ2 z j1 ]
e j x 1 , e9 z j x 4 λ 4 ]
2
e j x 1 , e9 λ 4 z j
j ∈ I(1)
j ∈ I(5)
j ∈ I(2)
j ∈ I(5)
j1 ∈ I(2), j2 ∈ I(3)
j1 ∈ I(2), j2 ∈ I(3)
j1 ∈ I(4), j2 ∈ I(5)
j ∈ I(5)
j ∈ I(5)
Table 2. An overview of the non-admissible vector fields in (35) and their Lie bracket representations. Here, the index sets are I(1) = {1, 7}, I(2) = {2, 6}, I(3) = {3, 8}, I(4) = {4, 9},
I(5) = {5}.
fields under appropriate assumptions on the communication graph, see also the last column of Table 1. Specifically,
if the graph is strongly connected, then all non-admissible
vector fields can be rewritten independent of the objective
function as well as the constraints, given that they admit
the structure (32), (33). In most cases, however, much less
restrictive requirements on the communication graph are
sufficient. We do not explicitly discuss how to rewrite
the non-admissible vector fields using Proposition 1 and
Proposition 2 in general but illustrate this by means of an
example in Section 5.
point dynamics (7) are then given by
ẋ = − ∇ F ( x ) − (2e2 − e5 )ν2 − 2( x1 e1 + x2 e2 )λ1
(35a)
− (2x4 − x1 ) e4 λ4 − (2x1 − x4 ) e1 λ4 − ( e1 + e3 )λ3
ν̇2 = 2x2 − x5
(35b)
λ̇ = λ1 e1 ( x12 + x22 − 4) + λ3 e2 ( x1 + x3 − 2)
+ λ4 e3 ( x42 − x4 x1 + x12 − 100),
(35c)
where x ∈ R5 , ν2 ∈ R, and λ = [λ1 , λ3 , λ4 ]> ∈ R3 .
We next rewrite all non-admissible vector fields in (35)
using Proposition 1. We will thereby follow the choice (18)
to make sure that (16) holds. We do not discuss how to
rewrite each non-admissible vector field in detail, but limit
ourselves to the vector field e1 x4 λ4 from (35a). Comparing
5 EXAMPLE
(1)
with (17), we have j1 = 1, jr = 4, and f j (zI( j1 ) ) = 1,
1
In this section we illustrate the previous results by means
of an example. Consider the following optimization problem
(2)
f j (zI( jr ) ) = x4 λ4 . Following Proposition 1, we require
r −1
a path from node 1 to node 4 that is here given by p14 =
hv1 |v2 |v3 |v4 i; thus r = 4. We then obtain
5
min
x ∈ R5
s.t.
F(x) =
e1 x4 λ4 = e j2 λ4 x4 , [ e j1 z j2 , e1 z j1 ] ,
∑ Fi (xi )
i =1
a2 ( x ) = 2x2 − x5 = 0
c1 ( x ) = x12 + x22 − 4 ≤ 0
(36)
where j1 ∈ I(2) = {2, 6}, j2 ∈ I(3) = {3, 8}, and the
right-hand side is a Lie bracket of admissible vector fields.
All other non-admissible vector fields in (35) can be treated
similarly and we sum up the resulting Lie brackets in Table 2. For the simulation we let each j, j1 , j2 be the largest
index in its respective index set. By that choice, we inject
less perturbation in the primal and more in the dual variables New which is also visible in the simulation results
depicted in Fig. 3. As to be seen, the distributed algorithm approximates the trajectories of the non-distributed
saddle-point dynamics (35) and converges to a neighborhood of the optimizer x ? = [0, 2, 2, 4, 4]> . NewWe also
(34)
c3 ( x ) = x1 + x3 − 2 ≤ 0
c4 ( x ) = x42 − x4 x1 + x12 − 100 ≤ 0,
where Fi ( xi ) = ( xi − i )2 , x = [ x1 , x2 , x3 , x4 , x5 ]> ∈ R5 .
We assume that the communication topology is described
by the graph in Fig. 2. Observe that the constraints c3
and c4 are not compatible with the graph topology, hence
Assumption 2 is not fulfilled. The corresponding saddle-
7
6
x(t)
bracket approximation approach to distributed optimization proposed in [11], [12] and discussed which kind of
vector fields can in principle be written in terms of Lie
brackets of admissible vector fields. We did not discuss
the construction of approximating inputs but postpone
this to [13] where we will present a modified version of
the general algorithm from [16] that exploits the structure
of the problem at hand.
x1
x2
x3
x4
x5
4
2
0
-2
0
5
10
15
1
References
ν(t)
0
-1
[1]
S. Michalowsky, B. Gharesifard, and C. Ebenbauer,
“On the Lie bracket approximation approach to distributed optimization: Extensions and limitations,”
in 2018 European Control Conference (ECC), accepted,
2018.
[2]
S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed optimization and statistical learning via the alternating direction method of multipliers,” Foundations and Trends in Machine Learning,
vol. 3, no. 1, pp. 1–122, 2011.
[3]
F. Bullo, J. Cortés, and S. Martínez, Distributed Control
of Robotic Networks, ser. Applied Mathematics Series.
Princeton University Press, 2009.
[4]
J. Zhao and F. Dörfler, “Distributed control and optimization in DC microgrids,” Automatica, vol. 61, no.
Supplement C, pp. 18 – 26, 2015.
[5]
D. Feijer and F. Paganini, “Stability of primal–dual
gradient dynamics and applications to network optimization,” Automatica, vol. 46, no. 12, pp. 1974 – 1981,
2010.
[6]
J. Wang and N. Elia, “A control perspective for centralized and distributed convex optimization,” in
2011 IEEE 50th Conference on Decision and Control
and European Control Conference (CDC-ECC), 2011, pp.
3800–3805.
[7]
H.-B. Dürr, C. Zeng, and C. Ebenbauer, “Saddle point
seeking for convex optimization problems,” IFAC Proceedings Volumes, vol. 46, no. 23, pp. 540–545, 2013.
[8]
B. Gharesifard and J. Cortés, “Distributed
continuous-time convex optimization on weightbalanced digraphs,” IEEE Transactions on Automatic
Control, vol. 59, no. 3, pp. 781–786, 2014.
[9]
S. K. Niederländer and J. Cortés, “Distributed coordination for separable convex optimization with
coupling constraints,” in 2015 54th IEEE Conference
on Decision and Control (CDC), Dec 2015, pp. 694–699.
-2
0
5
10
15
2
λ(t)
1
0
-1
0
5
10
15
time t
Figure 3. Simulation results of the example from Section 5. The
upper plot shows the primal variable x (t), the lower left one
the dual variable ν2 (t) and the lower right one the dual variable λ(t). NewIn each plot, the dotted black lines marked with
squares depict the trajectories of the non-distributed saddlepoint dynamics (35) while the thinner oscillating ones depict the
trajectories of the distributed approximation. The corresponding
thick lines depict the distributed approximation with additional
low-pass filters. The dashed black lines indicate the desired
equilbrium of (35).
included simulation results with additional low-pass filters in the distributed x-, ν- and λ-dynamics. While the
effect on the primal variables is small since we already
reduced the oscillations by our design choice, the dual
variables show significantly less oscillations and better
approximate the non-distributed trajectories. A rigorous
stability analysis of the augmented distributed dynamics
and a performance-oriented design of the filters is up to
future work.
6 CONCLUSIONS AND OUTLOOK
We considered a convex optimization problem and
showed how distributed optimization algorithms can be
designed for a quite general class of problems with little
structural requirements under mild assumptions on the
communication network. We therefore extended the Lie
8
[10] B. Touri and B. Gharesifard, “Saddle-point dynamics for distributed convex optimization on general
directed graphs,” in 2016 IEEE 55th Conference on Decision and Control (CDC), 2016, pp. 862–866.
By the induction hypothesis we then have
h
h jr̄+1 ,jr̄ , h jr̄ ,jr̄−1 , [ . . . , [h j3 ,j2 , h j2 ,j1 ] . . . ]
i
(z)
= [h jr̄+1 ,jr̄ , e j1 f˜jr̄ ,j1 ](z)
[11] C. Ebenbauer, S. Michalowsky, V. Grushkovskaya,
and B. Gharesifard, “Distributed optimization over
directed graphs with the help of Lie brackets,” in Proc.
20th IFAC World Congress, 2017, pp. 15 908–15 913.
N
= e j1 ( ∑
i =1
− e jr̄ (
∂ f˜jr̄ ,j1
∂zi
∑
ei
i ∈I( jr̄ )
∪I( jr̄+1 )
[12] S. Michalowsky, B. Gharesifard, and C. Ebenbauer, “Distributed extremum seeking over directed
graphs,” in 2017 IEEE 56th Conference on Decision and
Control (CDC), 2017, pp. 2095–2101.
= e j1
∂ f˜jr̄ ,j1
∂z jr̄
(z)) e jr̄ f jr̄ (zI( jr̄ ) , zI( jr̄+1 ) )
∂ f jr̄
∂zi
(zI( jr̄ ) , zI( jr̄+1 ) )) e j1 f˜jr̄ ,j1 (z)
(z) f jr̄ (zI( jr̄ ) , zI( jr̄+1 ) ),
(38)
where e ji ∈ Rn+neq +nineq is the ji th unit vector. Here, we
used that pi1 ir̄+1 is a simple path; hence, since I(ik1 ) and
I(ik2 ) are disjunct for any k1 6= k2 , also jk1 6= jk2 for any
k1 , k2 = 1, 2, . . . , r̄ + 1, k1 6= k2 . Using definition (37), we
obtain (14). Finally, since jk ∈ I(ik ) and gik ik+1 6= 0 for
all k = 1, 2, . . . , r − 1, all h jk+1 ,jk are admissible; hence the
left-hand side of (14) is a Lie bracket of admissible vector
fields.
[13] ——, “A Lie bracket approximation approach to distributed optimization over directed graphs,” ArXiv
e-prints arXiv:1711.05486 [math.OC], 2017, https://
arxiv.org/abs/1711.05486.
[14] N. Biggs, Algebraic graph theory. Cambridge university press, 1993.
[15] H.-B. Dürr and C. Ebenbauer, “On a class of smooth
optimization algorithms with applications in control,” IFAC Proceedings Volumes, vol. 45, no. 17, pp.
291 – 298, 2012, 4th IFAC Conference on Nonlinear
Model Predictive Control.
7.2 Proof of Proposition 1
Proof. Observe first that
r −2 ∂ f
jk
∂z j
k +1
k =1
f jr−1 (zI( jr−1 ) , zI( jr ) ) ∏
[16] W. Liu, “An approximation algorithm for nonholonomic systems,” SIAM Journal on Control and Optimization, vol. 35, no. 4, pp. 1328–1365, 1997.
(2)
(zI( jk ) , zI( jk+1 ) )
(1)
= f jr−1 (zI( jr ) ) f jr−1 (zI( jr−1 ) )
r −2
[17] H. J. Sussmann and W. Liu, “Limits of highly oscillatory controls and the approximation of general
paths by admissible trajectories,” in 1991 IEEE 30th
Conference on Decision and Control (CDC), 1991, pp.
437–442.
r − 1 ∂ f (2)
j
× ∏ f j (zI( jk ) ) ∏
k =1
=
(1)
k
(1)
(2)
f j (zI( jr ) ) f j (zI( j1 ) )
r −1
1
k =2
r −1
∏
k =1
k −1
∂z j
k
(zI( jk ) )
(2)
∂ fj
(1)
f j (zI( jk ) ) ∂zk−1 (zI( jk ) ).
j
k
k
Then, using (16) and applying Lemma 1, the result immediately follows.
7 APPENDIX
7.1 Proof of Lemma 1
7.3 Proof of Proposition 2
Proof. We prove this result by induction. First, for r = 1,
(14) is trivially true by definition (13). Suppose now that
the claim holds for all r ≤ r̄, r̄ > 1 and consider r = r̄ + 1.
Define
Proof. We first show by induction that
h
i
. . . [[ψm , ψm−1 ], ψm−2 ], . . . , ψ2 (z) = ei
m
∏ ηk (z jk ).
k =2
f˜jr̄ ,j1 (z) =
(39)
r̄ −2 ∂ f
jk
f jr̄−1 (zI( jr̄−1 ) , zI( jr̄ ) )
∂z j
k +1
k =1
∏
(zI( jk ) , zI( jk+1 ) ).
(37)
For m = 2 it is clear that (39) holds. Suppose now that (39)
holds for all m ≤ m̄, m̄ ≥ 2. For m = m̄ + 1 we then have
9
with a slight abuse of notation
h
i
. . . [[ψm̄+1 , ψm̄ ], ψm̄−1 ], . . . , ψ2 (z)
m̄+1
= [ ei
∏ ηk (z jk ), ei zi η2 (z j2 )]
k =2
∂η2
(z j2 )) ei
= ei ( ei> η2 (z j2 ) + zi e j2 ∂z
j2
m̄+1
+ ei (
∑
k =2
m̄+1
∏ ηk (z jk )
k =2
e jk ∂z k (z jk ) ∏ η` (z j` )) ei zi η2 (z j2 )
∂η
jk
`=2
`6=k
m̄+1
= ei
∏ ηk (z jk ),
(40)
k =2
where we used that i 6= jk for any k = 1, . . . , m in the last
step. This proves (39). Finally, using (39), we obtain
h
i
. . . [[ψm , ψm−1 ], ψm−2 ], . . . , ψ1 (z)
m
= [ ei ∏ ηk (z jk ), ei η1 (z j1 ) ∫ η0 (zi )dzi ]
k =2
m
= ei η0 (zi ) ∏ ηk (z jk ).
(41)
k =1
This finishes the proof.
10
| 3cs.SY
|
Published as a conference paper at ICLR 2018
DYNAMIC N EURAL P ROGRAM E MBEDDINGS FOR P RO GRAM R EPAIR
arXiv:1711.07163v3 [cs.AI] 25 Feb 2018
Ke Wang∗
University of California
Davis, CA 95616, USA
[email protected]
Rishabh Singh
Microsoft Research
Redmond, WA 98052, USA
[email protected]
Zhendong Su
University of California
Davis, CA 95616, USA
[email protected]
A BSTRACT
Neural program embeddings have shown much promise recently for a variety
of program analysis tasks, including program synthesis, program repair, codecompletion, and fault localization. However, most existing program embeddings
are based on syntactic features of programs, such as token sequences or abstract
syntax trees. Unlike images and text, a program has well-defined semantics that
can be difficult to capture by only considering its syntax (i.e. syntactically similar
programs can exhibit vastly different run-time behavior), which makes syntaxbased program embeddings fundamentally limited. We propose a novel semantic
program embedding that is learned from program execution traces. Our key insight is that program states expressed as sequential tuples of live variable values
not only capture program semantics more precisely, but also offer a more natural
fit for Recurrent Neural Networks to model. We evaluate different syntactic and
semantic program embeddings on the task of classifying the types of errors that
students make in their submissions to an introductory programming class and on
the CodeHunt education platform. Our evaluation results show that the semantic
program embeddings significantly outperform the syntactic program embeddings
based on token sequences and abstract syntax trees. In addition, we augment
a search-based program repair system with predictions made from our semantic
embedding and demonstrate significantly improved search efficiency.
1
I NTRODUCTION
Recent breakthroughs in deep learning techniques for computer vision and natural language processing have led to a growing interest in their applications in programming languages and software
engineering. Several well-explored areas include program classification, similarity detection, program repair, and program synthesis. One of the key steps in using neural networks for such tasks is
to design suitable program representations for the networks to exploit. Most existing approaches in
the neural program analysis literature have used syntax-based program representations. Mou et al.
(2016) proposed a convolutional neural network over abstract syntax trees (ASTs) as the program
representation to classify programs based on their functionalities and detecting different sorting routines. DeepFix (Gupta et al., 2017), SynFix (Bhatia & Singh, 2016), and sk p (Pu et al., 2016) are
recent neural program repair techniques for correcting errors in student programs for MOOC assignments, and they all represent programs as sequences of tokens. Even program synthesis techniques
that generate programs as output, such as RobustFill (Devlin et al., 2017), also adopt a token-based
program representation for the output decoder. The only exception is Piech et al. (2015), which
introduces a novel perspective of representing programs using input-output pairs. However, such
representations are too coarse-grained to accurately capture program properties — programs with
the same input-output behavior may have very different syntactic characteristics. Consequently, the
embeddings learned from input-output pairs are not precise enough for many program analysis tasks.
Although these pioneering efforts have made significant contributions to bridge the gap between
deep learning techniques and program analysis tasks, syntax-based program representations are fundamentally limited due to the enormous gap between program syntax (i.e. static expression) and
∗
Work done during an internship at Microsoft Research.
1
Published as a conference paper at ICLR 2018
s t a t i c int[] InsertionSort(int[] A) {
int left = 0;
int right = A.Length;
s t a t i c int[] BubbleSort(int[] A) {
int left = 0;
int right = A.Length - 1;
for (int i = right;i > left;i--) {
for (int j = left;j < i;j++) {
if (A[j] > A[j + 1]) {
int tmp = A[j];
for (int i = left;i < right;i++) {
for (int j = i - 1;j >= left;j--) {
if (A[j] > A[j + 1]) {
int tmp = A[j];
A[j] = A[j + 1];
// instrumentation line
Console.WriteLine(
string.Join(",", A)
);
A[j] = A[j + 1];
// instrumentation line
Console.WriteLine(
string.Join(",", A)
);
A[j + 1] = tmp;
// instrumentation line
Console.WriteLine(
string.Join(",", A)
);
A[j + 1] = tmp;
// instrumentation line
Console.WriteLine(
string.Join(",", A)
);
}}}
}}}
return A;
}
return A;
Bubble
[5,5,1,4,3]
[5,8,1,4,3]
[5,1,1,4,3]
[5,1,8,4,3]
[1,1,8,4,3]
[1,5,8,4,3]
[1,5,4,4,3]
[1,5,4,8,3]
[1,4,4,8,3]
[1,4,5,8,3]
[1,4,5,3,3]
[1,4,5,3,8]
[1,4,3,3,8]
[1,4,3,5,8]
[1,3,3,5,8]
[1,3,4,5,8]
Insertion
[5,5,1,4,3]
[5,8,1,4,3]
[5,1,1,4,3]
[5,1,8,4,3]
[5,1,4,4,3]
[5,1,4,8,3]
[5,1,4,3,3]
[5,1,4,3,8]
[1,1,4,3,8]
[1,5,4,3,8]
[1,4,4,3,8]
[1,4,5,3,8]
[1,4,3,3,8]
[1,4,3,5,8]
[1,3,3,5,8]
[1,3,4,5,8]
}
Figure 1: Bubble sort and insertion sort (code highlighted in shadow box are the only syntactic
differences between the two algorithms). Their execution traces for the input vector A = [8, 5, 1, 4,
3]are displayed on the right, where, for brevity, only values for variable A are shown.
static int max(int[] arr) {
1
Variable Trace
State Trace
{max val : −∞}
{max val : −∞, item : ⊥}
{item : 1}
{max val : −∞, item : 1}
{max val : 1}
{max val : 1, item : 1}
{item : 5}
{max val : 1, item : 5}
{max val : 5}
{max val : 5, item : 5}
{item : 3}
{max val : 5, item : 3}
2
int max_val = int.MinValue;
3
4
foreach(int item in arr)
{
if (item > max_val)
max_val = item;
}
5
6
7
8
9
10
return max_val;
11
12
}
Figure 2: Example for illustrating program
dependency.
Table 1: Variable and state traces obtained by executing function max, given arr = [1, 5, 3].
semantics (i.e. dynamic execution). This gap can be illustrated as follows. First, when a program is
executed at runtime, its statements are almost never interpreted in the order in which the corresponding token sequence is presented to the deep learning models (the only exception being straightline
programs, i.e., ones without any control-flow statements). For example, a conditional statement
only executes one branch each time, but its token sequence is expressed sequentially as multiple
branches. Similarly, when iterating over a looping structure at runtime, it is unclear in which order
any two tokens are executed when considering different loop iterations. Second, program dependency (i.e. data and control) is not exploited in token sequences and ASTs despite its essential role
in defining program semantics. Figure 2 shows an example using a simple max function. On line
8, the assignment statement means variable max val is data-dependent on item. In addition, the
execution of this statement depends on the evaluation of the if condition on line 7, i.e., max val is
also control-dependent on item as well as itself. Third, from a pure program analysis standpoint,
the gap between program syntax and semantics is manifested in that similar program syntax may
lead to vastly different program semantics. For example, consider the two sorting functions shown
in Figure 1. Both functions sort the array via two nested loops, compare the current element to
its successor, and swap them if the order is incorrect. However, the two functions implement different algorithms, namely Bubble Sort and Insertion Sort. Therefore minor syntactic discrepancies
can lead to significant semantic differences. This intrinsic weakness will be inherited by any deep
learning technique that adopts a syntax-based program representation.
2
Published as a conference paper at ICLR 2018
To tackle this aforementioned fundamental challenge, this paper proposes a novel semantic program
embedding that is learned from the program’s runtime behavior, i.e. dynamic program execution
traces. We execute a program on a set of test cases and monitor/record the program states comprising of variable valuations. We introduce three approaches to embed these dynamic executions:
(1) variable trace embedding — consider each variable independently, (2) state trace embedding —
consider sequences of program states, each of which comprises of a set of variable values, and (3)
hybrid embedding — incorporate dependencies into individual variable sequences to avoid redundant variable values in program states.
Our novel program embeddings address the aforementioned issues with the syntactic program representations. The dynamic program execution traces precisely illustrate the program behaves at
runtime, and the values for each variable at each program point precisely models the program semantics. Regarding program dependencies, the dynamic execution traces, expressed as a sequential
list of tuples (each of which represents the value of a variable at a certain program point), provides
an opportunity for Recurrent Neural Network (RNN) to establish the data dependency and control
dependency in the program. By monitoring particular value patterns between interacting variables,
the RNN is able to model their relationship, leading to more precise semantic representations.
Reed & De Freitas (2015) recently proposed using program traces (as a sequence of actions/statements) for training a neural network to learn to execute an algorithm such as addition or sorting.
Their notion of program traces is different from our dynamic execution traces consisting of program
states with variable valuations. Our notion offers the following advantages: (1) a sequence of program states can be viewed as a sequence of input-output pairs of each executed statement, in other
words, sequences of program states provide more robust information than that from sequences of executed statements, and (2) although a sequence of executed statements follows dynamic execution,
it is still represented syntactically, and therefore may not adequately capture program semantics. For
example, consider the two sorting algorithms in Figure 1. According to Reed & De Freitas (2015),
they will have an identical representation w.r.t. statements that modify the variable A, i.e. a repetition
of A[j] = A[j + 1] and A[j + 1] = tmp for eight times. Our representation, on the other hand, can
capture their semantic differences in terms of program states by also only considering the valuation
of the variable A.
We have evaluated our dynamic program embeddings in the context of automated program repair.
In particular, we use the program embeddings to classify the type of mistakes students made to their
programming assignments based on a set of common error patterns (described in the appendix). The
dataset for the experiments consists of the programming submissions made to Module 2 assignment
in Microsoft-DEV204.1X and two additional problems from the Microsoft CodeHunt platform. The
results show that our dynamic embeddings significantly outperform syntax-based program embeddings, including those trained on token sequences and abstract syntax trees. In addition, we show
that our dynamic embeddings can be leveraged to significantly improve the efficiency of a searchbased program corrector S ARF G EN1 (Wang et al., 2017) (the algorithm is presented in the appendix).
More importantly, we believe that our dynamic program embeddings can be useful for many other
program analysis tasks, such as program synthesis, fault localization, and similarity detection.
To summarize, the main contributions of this paper are: (1) we show the fundamental limitation of
representing programs using syntax-level features; (2) we propose dynamic program embeddings
learned from runtime execution traces to overcome key issues with syntactic program representations; (3) we evaluate our dynamic program embeddings for predicting common mistake patterns
students make in program assignments, and results show that the dynamic program embeddings outperform state-of-the-art syntactic program embeddings; and (4) we show how the dynamic program
embeddings can be utilized to improve an existing production program repair system.
2
BACKGROUND : DYNAMIC P ROGRAM A NALYSIS
This section briefly reviews dynamic program analysis (Ball, 1999), an influential program analysis
technique that lays the foundation for constructing our new program embeddings.
Unlike static analysis (Nielson et al., 1999), i.e., the analysis of program source code, dynamic analysis focuses on program executions. An execution is modeled by a set of atomic actions, or events,
1
Currently integrated with Microsoft-DEV204.1X as a feedback generator for production use.
3
Published as a conference paper at ICLR 2018
organized as a trace (or event history). For simplicity, this paper considers sequential executions
only (as opposed to parallel executions) which lead to a single sequence of events, specifically, the
executions of statements in the program. Detailed information about executions is often not readily available, and separate mechanisms are needed to capture the tracing information. An often
adopted approach is to instrument a program’s source code (i.e., by adding additional monitoring
code) to record the execution of statements of interest. In particular, those inserted instrumentation
statements act as a monitoring window through which the values of variables are inspected. This instrumentation process can occur in a fully automated manner, e.g., a common approach is to traverse
a program’s abstract syntax tree and insert “write” statements right after each program statement that
causes a side-effect (i.e., changing the values of some variables).
Consider the two sorting algorithms depicted in Figure 1. If we assume A to be the only
variable of interest and subject to monitoring, we can instrument the two algorithms with
Console.WriteLine(A) after each program location in the code whenever A is modified2 (i.e.
the lines marked by comments). Given the input vector A = [8, 5, 1, 4, 3], the execution traces of
the two sorting routines are shown on the right in Figure 1.
One of the key benefits of dynamic analysis is its ability to easily and precisely identify relevant
parts of the program that affect execution behavior. As shown in the example above, despite the
very similar program syntax of bubble sort and insertion sort, dynamic analysis is able to discover
their distinct program semantics by exposing their execution traces. Since understanding program
semantics is a central issue in program analysis, dynamic analysis has seen remarkable success
over the past several decades and has resulted in many successful program analysis tools such as
debuggers, profilers, monitors, or explanation generators.
3
OVERVIEW OF THE A PPROACH
We now present an overview of our approach. Given a program and the execution traces extracted
for all its variables, we introduce three neural network models to learn dynamic program embeddings. To demonstrate the utility of these embeddings, we apply them to predict common error
patterns (detailed in Section 5) that students make in their submissions to an online introductory
programming course.
Variable Trace Embedding As shown in Table 1, each row denotes a new program point where
a variable gets updated.3 The entire variable trace consists of those variable values at all program
points. As a subsequent step, we split the complete trace into a list of sub-traces (one for each
variable). We use one single RNN to encode each sub-trace independently and then perform max
pooling on the final states of the same RNN to obtain the program embedding. Finally, we add a one
layer softmax regression to make the predictions. The entire workflow is show in Figure 3.
State Trace Embedding Because each variable trace is handled individually in the previous approach, variable dependencies/interactions are not precisely captured. To address this issue, we
propose the state trace embedding. As depicted in Table 1, each program point l introduces a new
program state expressed by the latest variable valuations at l. The entire state trace is a sequence of
program states. To learn the state trace embedding, we first use one RNN to encode each program
state (i.e., a tuple of values) and feed the resulting RNN states as a sequence to another RNN. Note
that we do not assume that the order in which variables values are encoded by the RNN for each
program state but rather maintain a consistent order throughout all program states for a given trace.
Finally, we feed a softmax regression layer with the final state of the second RNN (shown in Figure 4). The benefit of state trace embedding is its ability to capture dependencies among variables
in each program state as well as the relationship among program states.
Dependency Enforcement for Variable Trace Embedding Although state trace embedding can
better capture program dependencies, it also comes with some challenges, the most significant of
which is redundancy. Consider a looping structure in a program. During an iteration, whenever
2
3
On the abstract syntax trees to enable complete automation regardless of the structure of programs.
We ignore the input variable arr since it is read-only (similarly for the state trace later).
4
Published as a conference paper at ICLR 2018
Figure 3: Variable trace for program embedding.
Figure 4: State trace for program embedding.
Figure 5: Dependency enforcement embedding. Dotted lines denoted dependencies.
one variable gets modified, a new program state will be created containing the values of all variables, even of those unmodified by the loop. This issue becomes more severe for loops with larger
numbers of iterations. To tackle this challenge, we propose the third and final approach, dependency enforcement for variable trace embedding (hereinafter referred as dependency enforcement
embedding), that combines the advantages of variable trace embedding (i.e., compact representation
of execution traces) and state trace embedding (i.e., precise capturing of program dependencies). In
dependency enforcement embedding, a program is represented by separate variable traces, with each
variable being handled by a different RNN. In order to enforce program dependencies, the hidden
states from different RNNs will be interleaved in a way that simulates the needed data and control
dependencies. Unlike variable trace embedding, we perform an average pooling on the final states of
all RNNs to obtain the program embedding on which we build the final layer of softmax regression.
Figure 5 describes the workflow.
5
Published as a conference paper at ICLR 2018
4
DYNAMIC P ROGRAM E MBEDDINGS
We now formally define the three program embedding models.
4.1
VARIABLE T RACE M ODEL
Given a program P , and its variable set V (v0 , v1 ,..., vn ∈ V ), a variable trace is a sequence of
values a variable has been assigned during the execution of P .4 Let xt vn denote the value from the
variable trace of vn that is fed to the RNN encoder (Gated Recurrent Unit) at time t as the input,
and ht vn as the resulting RNN’s hidden state. We compute the variable trace embedding for P in
Equation (3) as follows (hT vn denotes the last hidden state of the encoder):
ht v1 = GRU(ht−1 v1 , xt v1 )
...
ht vn = GRU(ht−1 vn , xt vn )
hP = MaxPooling(hT
(1)
v1 , ..., hT vn )
(2)
Evidence = (WhP + b)
(4)
(3)
Y = softmax(Evidence)
(5)
We compute the representation of the program trace by performing max pooling over the last hidden
state representation of each variable trace embedding. The hidden states ht v1 , . . . , ht vn , hP ∈ Rk
where k denotes the size of hidden layers of the RNN encoder. Evidence denotes the output of a
linear model through the program embedding vector hP , and we obtain the predicted error pattern
class Y by using a softmax operation.
4.2
S TATE T RACE M ODEL
The key idea in state trace model is to embed each program state as a numerical vector first and then
feed all program state embeddings as a sequence to another RNN encoder to obtain the program
embedding. Suppose xt vn is the value of variable vn at t-th program state, and ht vn is the resulting
hidden state of the program state encoder. Equation (8) computes the t-th program state embedding.
Equations (9-11) encode the sequence of all program state embeddings (i.e., ht vn , ht+1 vn , . . . ,
ht+m vn ) with another RNN to compute the program embedding.
ht v1 = GRU(ht v0 , xt v1 )
(6)
ht v2 = GRU(ht v1 , xt v2 )
...
ht vn = GRU(ht vn−1 , xt vn )
(7)
(8)
h0t vn = GRU(h0t−1 vn , ht vn )
h0t+1 vn = GRU(h0t vn , ht+1 vn )
...
hP = GRU(h0t+m−1 vn , xt+m vn )
(9)
(10)
(11)
ht v1 , . . . , ht vn ∈ Rk1 ; h0t vn , . . . , hP ∈ Rk2 where k1 and k2 denote, respectively, the sizes of
hidden layers of the first and second RNN encoders.
4.3
D EPENDENCY E NFORCEMENT FOR VARIABLE T RACE E MBEDDING
The motivation behind this model is to combine the advantages of the previous two approaches,
i.e. representing the execution trace compactly while enforcing the dependency relationship among
variables as much as possible. In this model, each variable trace is handled with a different RNN.
A potential issue to be addressed is variable matching/renaming (i.e., α-renaming). In other words
same variables may be named differently in different programs. Processing each variable id with
a single RNN among all programs in the dataset will not only cause memory issues, but more
importantly the loss of precision. Our solution is to (1) execute all programs to collect traces for
all variables, (2) perform dynamic time wrapping (Vintsyuk, 1968) on the variable traces across all
programs to find the top-n most used variables that account for the vast majority of variable usage,
and (3) rename the top-n most used variables consistently across all programs, and rename all other
variables to a same special variable.
4
For presentation simplicity and w.l.o.g., we assume that the program does not take any inputs.
6
Published as a conference paper at ICLR 2018
Given the same set of variables among all programs, the mechanism of dependency enforcement on
the top ones is to fuse the hidden states of multiple RNNs based on how a new value of a variable
is produced. For example, in Figure 2 at line 8, the new value of max val is data-dependent on
item, and control-dependent on both item and itself. So at the time step when the new value of
max val is produced, the latest hidden states of the RNNs encode variable item as well as itself;
they together determine the previous state of the RNN upon which the new value of max val is
produced. If a value is produced without any dependencies, this mechanism will not take effect.
In other words, the RNN will act normally to handle data sequences on its own. In this work we
enforce the data-dependency in assignment statement, declaration statement and method calls; and
control-dependency in control statements such as if , f or and while statements. Equations (11 and
12) expose the inner workflow. hLT vm denotes the latest hidden state of the RNN encoding variable
trace of vm up to the point of time t when xt vn is the input of the RNN encoding variable trace of
vn . denotes element-wise matrix product.
ht−1 vn = hLT v1 hLT vm hLT vn
ht vn = GRU(ht−1 vn , xt vn )
(12)
5
Given vn depends on v1 and vm
hP = AveragePooling(hT v1 , ..., hT
vn )
(11)
(13)
E VALUATION
We train our dynamic program embeddings on the programming submissions
obtained from Assignment 2 from Microsoft-DEV204.1X: “Introduction to
C#” offered on edx and two other problems on Microsoft CodeHunt platform.
• Print Chessboard: Print the chessboard pattern using “X” and “O”
to represent the squares as shown in Figure 6.
• Count Parentheses: Count the depth of nesting parentheses in a
given string.
• Generate Binary Digits: Generate the string of binary digits for a
given integer.
XOXOXOXO
OXOXOXOX
XOXOXOXO
OXOXOXOX
XOXOXOXO
OXOXOXOX
XOXOXOXO
OXOXOXOX
Figure 6: The desired output for the
chessboard exercise.
Regarding the three programming problems, the errors students made in their
submissions can be roughly classified into low-level technical issues (e.g., list
indexing, branching conditions or looping bounds) and high-level conceptual issues (e.g., mishandling corner case, misunderstanding problem requirement or misconceptions on the underlying data
structure of test inputs).5
In order to have sufficient data for training our models to predict the error patterns, we (1) convert
each incorrect program into multiple programs such that each new program will have only one
error, and (2) mutate all the correct programs to generate synthetic incorrect programs such that they
exhibit similar errors that students made in real program submissions. These two steps allow us to
set up a dataset depicted in Table 2. Based on the same set of training data, we evaluate the dynamic
embeddings trained with the three network models and compare them with the syntax-based program
embeddings (on the same error prediction task) on the same testing data. The syntax-based models
include (1) one trained with a RNN that encodes the run-time syntactic traces of programs (Reed &
De Freitas, 2015); (2) another trained with a RNN that encodes token sequences of programs; and
(3) the third trained with a RNN on abstract syntax trees of programs (Socher et al., 2013).
Problem
Print Chessboard
Count Parentheses
Generate Binary Digits
Program Submissions
Correct
Incorrect
2,281
742
505
315
518
371
Synthetic Data
Training Validation Testing
120K
13K
15K
20K
2K
2K
22K
3K
2K
Table 2: Dataset for experimental evaluation.
5
Please refer to the Appendix for a detailed summary of the error patterns for each problem.
7
Published as a conference paper at ICLR 2018
All models are implemented in TensorFlow. All encoders in each of the trace model have two
stacked GRU layers with 200 hidden units in each layer except that the state encoder in the state
trace model has one single layer of 100 hidden units. We adopt random initialization for weight
initialization. Our vocabulary has 5,568 unique tokens (i.e., the values of all variables at each time
step), each of which is embedded into a 100-dimensional vector. All networks are trained using
the Adam optimizer (Kingma & Ba, 2014) with the learning and the decay rates set to their default
values (learning rate = 0.0001, beta1 = 0.9, beta2 = 0.999) and a mini-batch size of 500. For the
variable trace and dependency enforcement models, each trace is padded to have the same length
across each batch; for the state trace model, both the number of variables in each program state as
well as the length of the entire state trace are padded.
During the training of the dependency enforcement model, we have observed that when dependencies become complex, the network suffers from optimization issues, such as diminishing and
exploding gradients. This is likely due to the complex nature of fusing hidden states among RNNs,
echoing the errors back and forth through the network. We resolve this issue by truncating each
trace into multiple sub-sequences and only back-propagate on the last sub-sequence while only feedforwarding on the rest. Regarding the baseline network trained on syntactic traces/token sequences,
we use the same encoder architecture (i.e., two layer GRU of 200 hidden units) processing the same
100-dimension embedding vector for each statement/token. As for the AST model, we learn an embedding (100-dimension) for each type of the syntax node by propagating the leaf (a simple look up)
to the root through the learned production rules. Finally, we use the root embeddings to represent
programs.
Programming
Problem
Print Chessboard
Count Parentheses
Generate Binary Digits
Variable Trace
State Trace
93.9%
92.7%
92.1%
95.3%
93.8%
94.5%
Dependency
Enforcement
99.3%
98.8%
99.2%
Run-Time
Syntactic Trace
26.3%
25.5%
23.8%
Token
AST
16.8%
19.3%
21.2%
16.2%
21.7%
20.9%
Table 3: Comparing dynamic program embeddings with syntax-based program embedding in predicting common error patterns made by students.
As shown in Table 3, our embeddings trained on execution traces significantly outperform those
trained on program syntax (greater than 92% accuracy compared to less than 27% for syntax-based
embeddings). We conjecture this is because of the fact that minor syntactic discrepancies can lead
to major semantic differences as shown in Figure 1. In our dataset, there are a large number of
programs with distinct labels that differ by only a few number of tokens or AST nodes, which
causes difficulty for the syntax models to generalize. Even for the simpler syntax-level errors, they
are buried in large number of other syntactic variations and the size of the training dataset is relatively
small for the syntax-based models to learn precise patterns. In contrast, dynamic embeddings are
able to canonicalize the syntactical variations and pinpoint the underlying semantic differences,
which results in the trace-based models learning the correct error patterns more effectively even
with relatively smaller size of the training data.
In addition, we incorporated our dynamic program embeddings into S ARFGEN (Wang et al., 2017)
— a program repair system — to demonstrate their benefit in producing fixes to correct students
errors in programming assignments. Given a set of potential repair candidates, S ARFGEN uses
an enumerative search-based technique to find minimal changes to an incorrect program. We use
the dynamic embeddings to learn a distribution over the corrections to prioritize the search for the
repair algorithm.6 To establish the baseline, we obtain the set of all corrections from S ARFGEN for
each of the real incorrect program to all three problems and enumerate each subset until we find the
minimum fixes. On the contrary, we also run another experiment where we prioritize each correction
according to the prediction of errors with the dynamic embeddings. It is worth mentioning that one
incorrect program may be caused by multiple errors. Therefore, we only predict the top-1 error each
time and repair the program with the corresponding corrections. If the program is still incorrect, we
repeat this procedure till the program is fixed. The comparison between the two approaches is based
on how long it takes them to repair the programs.
6
Some corrections are merely syntactic discrepancies (i.e., they do not change program semantics such as
modifying a *= 2 to a += a). In order to provide precise fixes, those false positives would need to be eliminated.
8
Published as a conference paper at ICLR 2018
Number of
Fixes
1-2
3-5
6-7
≥8
Enumerative
Search
3.8
44.7
95.9
128.3
Variable Trace
Embeddings
2.5
3.6
4.2
41.6
State Trace
Embeddings
2.8
3.1
3.6
49.5
Dependency Enforcement
Embeddings
3.3
4.1
4.5
38.8
Table 4: Comparing the enumerative search with those guided by dynamic program embeddings in
finding the minimum fixes. Time is measured in seconds.
As shown in Table 4, the more fixes required, the more speedups dynamic program embeddings
yield — more than an order of magnitude speedups when the number of fixes is four or greater.
When the number of fixes is greater than seven, the performance gain drops significantly due to poor
prediction accuracy for programs with too many errors. In other words, our dynamic embeddings
are not viewed by the network as capturing incorrect execution traces, but rather new execution
traces. Therefore, the predictions become unreliable. Note that we ignored incorrect programs
having greater than 10 errors when most experiments run out of memory for the baseline approach.
6
R ELATED W ORK
There has been significant recent interest in learning neural program representations for various
applications, such as program induction and synthesis, program repair, and program completion.
Specifically for neural program repair techniques, none of the existing techniques, such as DeepFix (Gupta et al., 2017), SynFix (Bhatia & Singh, 2016) and sk p (Pu et al., 2016), have considered
dynamic embeddings proposed in this paper. In fact, dynamic embeddings can be naturally extended
to be a new feature dimension for these existing neural program repair techniques.
Piech et al. (2015) is a notable recent effort targeting program representation. Piech et al. explore
the possibility of using input-output pairs to represent a program. Despite their new perspective, the
direct mapping between input and output of programs usually are not precise enough, i.e., the same
input-output pair may correspond to two completely different programs, such as the two sorting
algorithms in Figure 1. As we often observe in our own dataset, programs with the same error
patterns can also result in different input-output pairs. Their approach is clearly ineffective for these
scenarios.
Reed & De Freitas (2015) introduced the novel approach of using execution traces to induce and
execute algorithms, such as addition and sorting, from very few examples. The differences from our
work are (1) they use a sequence of instructions to represent dynamic execution trace as opposed to
using dynamic program states; (2) their goal is to synthesize a neural controller to execute a program
as a sequence of actions rather than learning a semantic program representation; and (3) they deal
with programs in a language with low-level primitives such as function stack push/pop actions rather
than a high-level programming language.
As for learning representations, there are several related efforts in modeling semantics in sentence or
symbolic expressions (Socher et al., 2013; Zaremba et al., 2014; Bowman, 2013). These approaches
are similar to our work in spirit, but target different domains than programs.
7
C ONCLUSION
We have presented a new program embedding that learns program representations from runtime
execution traces. We have used the new embeddings to predict error patterns that students make in
their online programming submissions. Our evaluation shows that the dynamic program embeddings
significantly outperform those learned via program syntax. We also demonstrate, via an additional
application, that our dynamic program embeddings yield more than 10x speedups compared to an
enumerative baseline for search-based program repair. Beyond neural program repair, we believe
that our dynamic program embeddings can be fruitfully utilized for many other neural program
analysis tasks such as program induction and synthesis.
9
Published as a conference paper at ICLR 2018
R EFERENCES
Thoms Ball. The concept of dynamic analysis. In Proceedings of the 7th European Software Engineering Conference Held Jointly with the 7th ACM SIGSOFT International Symposium on Foundations of Software Engineering, pp. 216–234, 1999.
Sahil Bhatia and Rishabh Singh. Automated correction for syntax errors in programming assignments using recurrent neural networks. CoRR, abs/1603.06129, 2016.
Samuel R Bowman. Can recursive neural tensor networks learn logical reasoning? arXiv preprint
arXiv:1312.6192, 2013.
Jacob Devlin, Jonathan Uesato, Surya Bhupatiraju, Rishabh Singh, Abdel rahman Mohamed, and
Pushmeet Kohli. RobustFill: Neural program learning under noisy I/O. In Proceedings of the
34th International Conference on Machine Learning, pp. 990–998, 2017.
Rahul Gupta, Soham Pal, Aditya Kanade, and Shirish K. Shevade. Deepfix: Fixing common c
language errors by deep learning. In Proceedings of the Thirty-First AAAI Conference on Artificial
Intelligence, 2017.
Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization.
abs/1412.6980, 2014. URL http://arxiv.org/abs/1412.6980.
CoRR,
Lili Mou, Ge Li, Lu Zhang, Tao Wang, and Zhi Jin. Convolutional neural networks over tree structures for programming language processing. In Proceedings of the Thirtieth AAAI Conference on
Artificial Intelligence, 2016.
Flemming Nielson, Hanne R. Nielson, and Chris Hankin. Principles of Program Analysis. 1999.
Chris Piech, Jonathan Huang, Andy Nguyen, Mike Phulsuksombati, Mehran Sahami, and Leonidas
Guibas. Learning program embeddings to propagate feedback on student code. In Proceedings
of the 32nd International Conference on Machine Learning, pp. 1093–1102, 2015.
Yewen Pu, Karthik Narasimhan, Armando Solar-Lezama, and Regina Barzilay. Sk p: A neural program corrector for moocs. In Companion Proceedings of the 2016 ACM SIGPLAN International
Conference on Systems, Programming, Languages and Applications: Software for Humanity,
SPLASH Companion 2016, pp. 39–40, 2016.
Scott Reed and Nando De Freitas.
arXiv:1511.06279, 2015.
Neural programmer-interpreters.
arXiv preprint
Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Ng,
and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment
treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pp. 1631–1642, 2013.
Taras K Vintsyuk. Speech discrimination by dynamic programming. Cybernetics, 4(1):52–57, 1968.
Ke Wang, Rishabh Singh, and Zhendong Su. Data-driven feedback generation for introductory
programming exercises. CoRR, abs/1711.07148, 2017. URL http://arxiv.org/abs/
1711.07148.
Wojciech Zaremba, Karol Kurach, and Rob Fergus. Learning to discover efficient mathematical
identities. In Advances in Neural Information Processing Systems, pp. 1278–1286, 2014.
10
Published as a conference paper at ICLR 2018
A PPENDIX
E RROR PATTERNS
Print Chessboard:
• Misprinting “O” to “0” or printing lower case instead of upper case characters.
• Switching across rows are supposed to be the other way around ( i.e. printing OXOXOXOX
for odd number rows and XOXOXOXO for even number rows).
• Printing the first row correctly but failed to make a switch across rows.
• Printing the entire chessboard as “X” or “O” only.
• Printing the chessboard correctly but with extra unnecessary characters.
• Printing the incorrect number of rows.
• Printing the incorrect number of columns.
• Printing the characters correctly but in wrong format (i.e. not correctly seperated with the
spaces to form the rows).
• Others.
Count Parentheses:
• Miss the corner case of empty strings.
• Mistakenly consider the parenthesis to be symbols rather than “(” or “)”.
• Mishandling the string of unmatched parentheses.
• Counting the number of matching parentheses rather then depth.
• Incorrectly assume nested parentheses are always present.
• Miscounting the characters which should have been ignored.
• Others.
Generate Binary Digits:
• Miss the corner case of integer 0.
• Misunderstand the binary digits to be underlying bytes of a string.
• Mistakes in arithmetic calculation regrading shift operations.
• Adding the binary digits rather than concatenating them to a string.
• Miss the one on the most significant bit.
• Others.
11
Published as a conference paper at ICLR 2018
S ARF G EN ’ S A LGORITHM
Algorithm 1: S ARFGEN ’s feedback generation procedure.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
/* P : an incorrect program; Ps : all correct solutions
function FixGeneration(P , Ps )
begin
// Among Ps identify Pcs to be reference programs to fix P
Pcs ← CandidatesIdentification(P , Ps )
// Initialize the minimum number of fixes k to be inifinity
k←∞
// Initialize the minimum set of fixes F(P )
F(P ) ← null
for Pc ∈ Pcs do
// Generates the syntactic discrepencies w.r.t. each Pc
C(P , Pc ) ← DiscrepenciesGeneration(P , Ps )
// Selecting subsets of C(P , Pc ) from size of one itll |C(P , Pc )|
for n ∈ [1, 2, ..., |C(P , Pc )|] do
Csubs (P , Pc ) ← {x | x ⊆ C(P , Pc ) ∧ |x| = n}
// Attemp each subset of C(P , Pc )
for Csub (P , Pc ) ∈ Csubs (P , Pc ) do
P 0 ← PatchApplication(P , Csub (P , Pc ))
// Update k if necessary
if isCorrect(P 0 ) then
if |P 0 | <k then
k ← |P 0 |
F(P ) ← P 0
*/
return F(P )
Algorithm 2: Incorporate pre-trained model to S ARFGEN ’s feedback generation procedure.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
/* P , Ps : same as above; M: learned Model
function FixGeneration(P , Ps , M)
begin
// Among Ps identify Pcs to be reference programs to fix P
Pcs ← CandidatesIdentification(P , Ps )
// Initialize the minimum number of fixes k to be inifinity
k←∞
// Initialize the minimum set of fixes F(P )
F(P ) ← null
for Pc ∈ Pcs do
// Generates the syntactic discrepencies w.r.t. each Pc
C(P , Pc ) ← DiscrepenciesGeneration(P , Ps )
// Executing P to extract the dynamic execution trace
T (P ) ← DynamicTraceExtraction(P )
// Prioritizing subsets of C(P , Pc ) through pre-trained model
Csubs (P , Pc ) ← Prioritization(C(P , Pc ), T (P ), M)
for Csub (P , Pc ) ∈ Csubs (P , Pc ) do
P 0 ← PatchApplication(P , Csub (P , Pc ))
if isCorrect(P 0 ) then
if |P 0 | <k then
k ← |P 0 |
F(P ) ← P 0
return F(P )
12
*/
| 2cs.AI
|
arXiv:1207.5553v1 [math.AC] 23 Jul 2012
REGULARITY 3 IN EDGE IDEALS ASSOCIATED TO
BIPARTITE GRAPHS
OSCAR FERNÁNDEZ-RAMOS AND PHILIPPE GIMENEZ
Abstract. We focus in this paper on edge ideals associated to bipartite graphs and give a combinatorial characterization of those having
regularity 3. When the regularity is strictly bigger than 3, we determine the first step i in the minimal graded free resolution where there
exists a minimal generator of degree > i + 3, show that at this step
the highest degree of a minimal generator is i + 4, and determine the
value of the corresponding graded Betti number βi,i+4 in terms of the
combinatorics of the associated bipartite graph. The results can then
be easily extended to the non-squarefree case through polarization. We
also study a family of ideals of regularity 4 that play an important role
in our main result and whose graded Betti numbers can be completely
described through closed combinatorial formulas.
Introduction
Studying homological invariants of monomial ideals in a polynomial ring
R = K[x1 , . . . , xn ] by looking for combinatorial properties in discrete objects
(graphs, hypergraphs, simplicial complexes, . . . ) associated to them is a well
known technique that has been fruitfully exploited in the last decades (see for
example the surveys [12] and [15] and their references). In fact, it provides
a quite complete dictionary between these two algebraic and combinatorial
classes.
Classical objects used to relate combinatorics with monomial ideals are
Stanley-Reisner ideals, simplicial or cellular resolutions and facet ideals. A
monomial ideal generated by quadrics can be viewed as the facet ideal of a
graph. When the graph is simple, i.e., has no loops, these ideals are called
edge ideals and were first introduced in [18].
The homological invariants of a monomial ideal I that we are interested
in are those encoded in the minimal graded free resolution of the ideal,
namely, the graded Betti numbers and the Castelnuovo-Mumford regularity.
Considering the standard N-grading on the polynomial ring R, the graded
Betti number βi,j is the number of minimal generators of degree j in the
i-th syzygy module of the ideal. If we denote by ui (resp. li ) the maximal
(resp. minimal) degree of a minimal generator in i-th syzygy module then
the fact that the resolution is minimal implies that li ≥ l0 + i, where l0
0
Both authors are partially supported by MTM2010-20279-C02-01 (Ministerio de Ciencia e Innovación, Spain).
1
2
OSCAR FERNÁNDEZ-RAMOS AND PHILIPPE GIMENEZ
is the minimal degree of a generator of I and the regularity of the ideal
is defined as reg(I) := max{ui − i}. An interesting situation is when all
minimal generators of I has same degree l0 and ui = li = l0 + i for all i. In
this case, reg(I) = l0 and we say that the ideal has an l0 -linear resolution.
There is a nice combinatorial characterization of edge ideals having a 2linear resolution, i.e., having regularity 2, in terms of the complement of
the associated graph due to Fröberg ([8]). A new proof of this result has
recently been obtained in [17] where the topology of the lcm-lattice of edge
ideals is studied. Another proof of Fröberg’s characterization can be found
in [6] where, moreover, the least i such that ui > i + 2 is characterized in
a combinatorial way when the edge ideal does not have a linear resolution.
This result was obtained independently in [7] where it is also shown that
ui = i+3 for this value of i where non-linear sygygies first appear. Moreover,
βi,i+3 is determined in terms of the complement of the graph. These results
are recalled in the section 1 (Theorem 1.11) together with all the required
definitions and notations. We will also show in the section that the graded
Betti numbers of an arbitrary edge ideal I satisfy the following property
(Theorem 1.8): for every i ≥ 0 and j ≥ i + 2,
βi,j (I) = βi,j+1 (I) = 0
⇒
βi+1,j+2 (I) = 0 .
It implies in particular that ui+1 ≤ ui + 2 for all i ≥ 0 (Corollary 1.9), a
refinement of [7, Theorem 5.2].
The aim of this paper is to characterize edge ideals associated to bipartite
graphs G having regularity 3 and determine, for those of regularity > 4,
the first step i in the minimal resolution such that ui > i + 3. This will be
done in section 3 where we will also prove that, for this value of i, one has
that ui = i + 4 and show that βi,i+4 is the number of induced subgraphs
of the bipartite complement of G that are isomorphic to cycles of minimal
length. The fundamental role played by these subgraphs is the reason why
we previously devote section 2 to study the graded Betti numbers of the edge
ideal associated to the bipartite complement of an even cycle. We show that
such an edge ideal has regularity 4 and give closed combinatorial formulas
for all its graded Betti numbers. We finally prove in the last section similar
results for monomial ideals generated by quadrics that are not squarefree.
The dependence of the Betti numbers of edge ideals on the characteristic
of the field K, even in the case of edge ideals associated to bipartite graphs,
prevents the possibility of obtaining similar results for higher regularity.
1. Preliminaries
1.1. Graphs and simplicial complexes. We start recalling elementary
concepts on graphs that we will handle along this paper (see [5] and [16]
for the terminology not included here regarding graphs and simplicial complexes, respectively).
REGULARITY 3 IN EDGE IDEALS ASSOCIATED TO BIPARTITE GRAPHS
3
Consider a finite simple graph G and denote by E(G) and V (G) its edge
and vertex sets respectively. We say that a subgraph H of G is induced on
a subset V ′ of V (G) if V (H) = V ′ and E(H) = {{u, v} ∈ E(G) : u, v ∈ V ′ }.
We write H = G[V ′ ] or H < G when H is an induced subgraph on an
unspecified subset of vertices of G. The complement of G is the graph Gc
on the same vertex set as G with E(Gc ) = {{u, v} : u, v ∈ V (G), {u, v} 6∈
E(G)}. Given a vertex u ∈ V (G), we denote by NG (u) the set
S of vertices
of G adjacent to u and, for a subset W ⊂ V (G), NG (W ) := u∈W NG (u).
The degree of u, denoted by deg(u), is the number of elements in NG (u).
A connected graph G with t := |V (G)| ≥ 4 whose vertices are all of degree
two is called a t-cycle and denoted by Ct . We will say that t is the length
of the cycle. An induced subgraph which is also a cycle is called an induced
cycle, and a graph G is said to be chordal if it has no induced cycle. A graph
whose vertices have all degree one has necessarily 2s vertices for some s ≥ 1
and consists of s disconnected edges. We denote it by sK2 .
Definition 1.1. The induced matching number of a graph G is the maximal
s such that sK2 < G. We denote this number by µ(G).
Consider now a simplicial complex ∆. Given a subset W of its vertex set
V (∆), the induced subcomplex of ∆ on W is ∆[W ] := {σ ∈ ∆ : σ ⊂ W }.
Recall that if one has two subcomplexes ∆1 and ∆2 of ∆ such that ∆ =
∆1 ∪ ∆2 , there is a long exact sequence of reduced homologies, called the
Mayer-Vietoris sequence,
(1)
H̃i (∆1 )
⊕
· · · −→ H̃i (∆0 ) −→
−→ H̃i (∆) −→ H̃i−1 (∆0 ) −→ · · ·
H̃i (∆2 )
whenever ∆0 := ∆1 ∩ ∆2 6= ∅.
Definition 1.2. Given a simplicial complex ∆ and u, v 6∈ V (∆), consider
the following two simplicial complexes:
• v ∗ ∆ := ∆ ∪ {{v} ∪ σ : σ ∈ ∆}, the cone on the base ∆ with appex
v;
• Σvu ∆ := ∆ ∪ {{u} ∪ σ : σ ∈ ∆} ∪ {{v} ∪ σ : σ ∈ ∆}, the suspension
of ∆ on the vertices u and v.
The following well-known results on cones and suspensions (see, e.g., [16,
Theorems 8.2 and 25.4]) will be very useful in the sequel.
Proposition 1.3.
(1) H̃i (v ∗ ∆) = 0, ∀ i ≥ 0.
(2) H̃0 (Σvu ∆) = 0 and H̃i (Σvu ∆) ≃ H̃i−1 (∆), ∀ i ≥ 1.
Associated to a graph G, one has its independence complex ∆(G) which is
defined as the simplicial complex on the vertex set V (G) such that F ⊂ V (G)
is a face of ∆(G) if and only if no edge of G is a subset of F . Observe that
if G is a graph and W is an arbitrary subset of V (G), one has that
∆(G)[W ] = ∆(G[W ]) .
4
OSCAR FERNÁNDEZ-RAMOS AND PHILIPPE GIMENEZ
Remark 1.4. A flag complex is a simplicial complex ∆ such that, for any
σ ⊂ V (∆), if every pair of elements in σ is a face of ∆ then σ is also a
face of ∆. In particular, a flag complex containing all pairs of vertices is
necessarily a simplex. Moreover, the independence complex ∆(G) of a graph
G is always, by definition, a flag complex.
Definition 1.5. If ∆ := ∆(G) is the independence complex of a graph G,
for all u ∈ V := V (G), we consider three induced subcomplexes of ∆ that
will be featured in this paper. Note that each of them is the independence
complex of an induced subgraph of G:
• del∆ (u) := {σ ∈ ∆ : u 6∈ σ} = ∆[V \ {u}];
• link∆ (u) := {σ ∈ ∆ : u 6∈ σ and σ∪{u} ∈ ∆} = ∆[V \(NG (u)∪{u})];
• star∆ (u) := {σ ∈ ∆ : σ ∪ {u} ∈ ∆} = ∆[V \ NG (u)].
For any vertex v ∈ V (G), star∆ (v) is a cone with appex v and hence it
is acyclic by Proposition 1.3.1. Since del∆ (v) ∪ star∆ (v) = ∆ and del∆ (v) ∩
star∆ (v) = link∆ (v), we can apply (1) whenever link∆ (v) 6= ∅ and get
(2) · · · −→ H̃i (link∆ (v)) −→ H̃i (del∆ (v)) −→ H̃i (∆) −→ H̃i−1 (link∆ (v)) −→ · · ·
· · · −→ H̃0 (link∆ (v)) −→ H̃0 (del∆ (v)) −→ H̃0 (∆) −→ 0 .
Let’s focus now on bipartite graphs. Recall that a graph G is bipartite if
its vertex set can be splitted into two disjoint sets, V (G) = X ⊔ Y , in such a
way that any edge of G has one vertex in X and the other in Y . When one
deals with bipartite graphs, it is usually convenient to use different symbols
for variables in X and variables in Y . We will denote variables in X by
x1 , . . . , xn and variables in Y by y1 , . . . , ym . The biadjacency matrix of the
bipartite graph G, M (G) = (ai,j ) ∈ Mn×m ({0, 1}), is defined by ai,j = 1 if
{xi , yj } ∈ E(G), 0 otherwise. The bipartite complement of a bipartite graph
G is the bipartite graph Gbc on the same vertex set as G, V (Gbc ) = X ⊔ Y ,
with E(Gbc ) = {{x, y} : x ∈ X, y ∈ Y, {x, y} 6∈ E(G)}. One has that
M (Gbc ) = 1n×m − M (G) where 1n×m is the n × m matrix whose entries are
all 1. Note that the bipartition V (G) = X ⊔ Y may not be unique if the
graph G is not connected and the notions of biadjacency matrix or bipartite
complement depend on the bipartition. That’s the reason why in section 3
we will restrict ourselves to connected bipartite graphs.
The next lemma will be useful to handle the homology of the independence
complex of a bipartite graph G. The last three items are rules that one can
apply for a reduction to a simpler case by removing vertices of G when the
biadjacency matrix M of G satisfies some properties.
Lemma 1.6. Let G be a bipartite graph with bipartition V (G) = {x1 , . . . , xn }⊔
{y1 , . . . , ym }, biadjacency matrix M = (ai,j ) ∈ Mn×m ({0, 1}), and independence complex ∆.
(1) If M has a row or a column whose entries are all 0, then H̃i (∆) = 0
for all i ≥ 0.
REGULARITY 3 IN EDGE IDEALS ASSOCIATED TO BIPARTITE GRAPHS
5
(2) If there exist r and c such that ar,c = 1 and the rest of entries on the
row r and the column c are zeros then, for all i > 0,
H̃i (∆) ≃ H̃i−1 (∆[V (G) \ {xr , yc }]) .
(3) If M has more than one row (resp. column) and if the entries on
the row r (resp. column c) are all 1 then, for all i ≥ 0,
H̃i (∆) ≃ H̃i (∆[V (G) \ {xr }]) (resp. H̃i (∆) ≃ H̃i (∆[V (G) \ {yc }])) .
(4) If M has a row r (resp. column c) with a unique zero entry, say
ar,c = 0, and if there is another zero entry on the column c (resp.
row r) then for all i ≥ 0,
H̃i (∆) ≃ H̃i (∆[V (G) \ {xr }]) (resp. H̃i (∆) ≃ H̃i (∆[V (G) \ {yc }])) .
(5) If M has two rows r and r ′ (resp. two columns c and c′ ) such that
{j : ar,j = 0} ⊂ {j : ar′ ,j = 0} (resp. {i : ai,c = 0} ⊂ {j : ai,c′ = 0})
then for all i ≥ 0,
H̃i (∆) ≃ H̃i (∆[V (G) \ {xr }]) (resp. H̃i (∆) ≃ H̃i (∆[V (G) \ {yc }])) .
Proof. 1: The vertex z of G corresponding to the row or column of M with
zero entries is isolated in G and hence ∆ is a cone with appex z, so it is
acyclic by Proposition 1.3.1.
2: By 1, one has that H̃i (del∆ (yc )) = H̃i (del∆ (xr )) = 0. Since ∆ =
del∆ (xr ) ∪ del∆ (yc ) and del∆ (xr ) ∩ del∆ (yc ) = deldel∆ (xr ) (yc ), the result
follows from the Mayer-Vietoris sequence (2).
3 and 4 are particular cases of 5 that follows by applying again the Mayer
Vietoris sequence (2), observing that link∆ (xr ) is acyclic by 1.
Example 1.7. If Σm is the independence complex of mK2 , then
1 if i = m − 1,
dimK (H̃i (Σm )) =
0 otherwise .
This can be shown by induction on m ≥ 1 as follows. Since Σ1 consists of
two disjoint vertices, H̃0 (Σ1 ) ≃ K and H̃i (Σ1 ) = 0 for all i > 0. If m > 1,
H̃0 (Σm ) = 0 because Σm is connected and for i > 0, applying Lemma 1.6.2,
one gets that H̃i (Σm ) ≃ H̃i−1 (Σm−1 ) and the result follows.
1.2. Some properties of the graded Betti numbers of an edge ideal.
Given a minimal graded free resolution of a homogeneous ideal I in R =
K[x1 , · · · , xn ],
M
M
ϕ0
ϕ1 M
R(−j)β0,j −→ I −→ 0 ,
0 −→
R(−j)βp,j −→ . . . −→
R(−j)β1,j −→
j
j
j
the regularity of I (in the sense of Castelnuovo-Mumford) is defined in terms
of the graded Betti numbers βi,j as reg(I) := max{j − i : βi,j 6= 0}. The
graded Betti numbers of I are usually arranged in a table called the Betti
diagram of I where βi,j is placed in the i-th column and (j − i)-th row of
6
OSCAR FERNÁNDEZ-RAMOS AND PHILIPPE GIMENEZ
the table. Note that the index of the last nonzero row in the Betti diagram
of I is its regularity.
If I is a monomial ideal, we can provide the polynomial ring R with
the usual Nn -multigrading and I has a minimal multigraded free resolution.
We can then define its multigraded Betti numbers, βi,m , as the number of
minimal generators of degree m ∈ Nn in the i-th syzygy module.
There is a one-to-one correspondence between squarefree monomial ideals
generated in degree 2 and simple graphs: associated to a simple graph G, one
has the edge ideal I(G) generated by the monomials of the form xi xj with
{xi , xj } ∈ E(G). The edge ideal I(G) is indeed the Stanley-Reisner ideal
of the independence complex ∆(G) of G, I(G) = I∆(G) . The multigraded
Betti numbers of an edge ideal I(G) can be expressed in terms of the reduced
homology of ∆(G) using Hochster’s Formula ([11, Theorem (5.1)]) that we
recall now. For any m ∈ Nn and i ≥ 0, one has that βi,m (I(G)) = 0 if the
mn
1
monomial xm := xm
1 · · · xn is not squarefree, and
(3)
βi,m (I(G)) = dimK H̃|m|−i−2(∆(G)[W ])
otherwise, where W := {j ∈ [n] : mj = 1}. The graded Betti numbers of
I(G) are then given by the following formula:
X
(4)
βi,j (I(G)) =
dimK (H̃j−i−2 (∆(G)[W ])) .
W ⊂V (G),|W |=j
Hochster’s Formula is a powerful tool when one wants to get information
on the Betti numbers of edge ideals. It can be used for example to prove
the following property on the graded Betti numbers of an edge ideal I(G):
Theorem 1.8. For any i ≥ 0 and any j ≥ i+2, if βi,j (I(G)) = βi,j+1 (I(G)) =
0 then βi+1,j+2 (I(G)) = 0.
Proof. Denote by ∆ := ∆(G) the independence complex of G and assume
that βi+1,j+2 (I(G)) 6= 0. By (3), there exists W ⊂ V (G) with |W | = j + 2
such that dimK H̃j−i−1 (∆[W ]) > 0. As ∆[W ] = ∆(G[W ]) is a flag complex,
there exist u, v ∈ W such that {u, v} 6∈ ∆(G)[W ] by Remark 1.4. Consider
then the following decomposition of ∆[W ],
∆[W ] = ∆[W \ {u}] ∪ ∆[W \ {v}] ,
with ∆[W \ {u}] ∩ ∆[W \ {v}] = ∆[W \ {u, v}] which is not empty since
|W | = j + 2 > 2. Invoking Hochster’s Formula (3) again, one has that
H̃j−i−1 (∆[W \ {u}]) = H̃j−i−1 (∆[W \ {v}]) = 0 because βi,j+1 (I) = 0 and
H̃j−i−2 (∆[W \ {u, v}]) = 0 because βi,j (I) = 0. Using the Mayer-Vietoris
sequence (1), one gets that H̃j−i−1 (∆[W ]) = 0, a contradiction.
Note that Theorem 1.8 can easily be extended to monomial ideals generated in degree two that may not be squarefree through polarization (see
section 4). The following direct consequence answers a question by Aldo
REGULARITY 3 IN EDGE IDEALS ASSOCIATED TO BIPARTITE GRAPHS
7
Conca who asked if our [7, Theorem 5.2] could be improved in this direction. In [1], Avramov, Conca and Iyengar proved bounds for the syzygies of
Koszul algebras and that question arose in this context.
Corollary 1.9. Let I be a monomial ideal generated in degree two and
denote by ui the maximal degree of a minimal generator in its i-th syzygy
module. Then, for all i ≥ 0, ui+1 ≤ ui + 2.
When an edge ideal I(G) has a linear resolution, all the nonzero entries
in its Betti diagram are located on the first row. Fröberg proved that an
edge ideal I(G) has a linear resolution if and only if the graph Gc is chordal.
We can rephrase this nice combinatorial characterization as follows:
Theorem 1.10 ([8]). An edge ideal I(G) has regularity 2 if and only if Gc
does not have induced cycles.
In [6], the authors go one step further and show that if reg(I(G)) > 2,
the non-linear syzygies appear for the first time at the (t − 3)-th step of the
resolution where t is the minimal length of an induced cycle in Gc . This
result is contained in the following stronger statement:
Theorem 1.11 ([7]). If I(G) is an edge ideal with reg(I(G)) > 2, let t ≥ 4
be the minimal length of an induced cycle in Gc . Then:
• βi,j (I(G)) = 0 for all i < t − 3 and j > i + 2;
• βt−3,t (I(G)) = |{induced t-cycles in Gc }|;
• βt−3,j (I(G)) = 0 for all j > t;
• for any m ∈ Nn such that |m| = t, one has that βt−3,m (I(G)) = 1
if m ∈ {0, 1}n and Gc [W ] ≃ Ct where W := {xi : mi = 1}.
Otherwise, βt−3,m (I(G)) = 0.
Observe that in the previous theorem, induced cycles in Gc play an important role. That’s why we previously focused on a particular family of edge
ideals, those associated to complements of cycles, and gave in [7, Proposition
3.1] closed combinatorial formulas for all its graded Betti numbers.
Following the same philosophy, we will focus now on graphs that are
the bipartite complement of an even cycle since induced even cycles in the
bipartite complement of an arbitrary graph will play a fundamental role
later in our main Theorem 3.2. Again, for this family of graphs, we describe
all the graded Betti numbers of the associated edge ideals in Theorem 2.1.
2. Bipartite complement of a cycle of even length
The following result is a direct consequence of Propositions 2.7, 2.9 and
2.11, that we will prove in this section.
Theorem 2.1. The edge ideal associated to the bipartite complement of an
even cycle of length t := 2s ≥ 6 has regularity 4 and its Betti diagram is
8
OSCAR FERNÁNDEZ-RAMOS AND PHILIPPE GIMENEZ
0
1 ...
2 β0,2 β1,3 . . .
3
β1,4 . . .
4
s−3
...
βs−3,s−1
...
...
t−5
t−4
βt−5,t−2
1
where the nonzero entries are located in the shadowed area. Moreover, βj−2,j
for 2 ≤ j ≤ s − 1, and βj−3,j for 4 ≤ j ≤ t − 2, are given respectively by the
following closed combinatorial formulas:
j−1 X
k
X
s k−1 s−k−1 s−k−c
βj−2,j =
,
c c−1
c−1
j−k
c=1
k=1
⌊j/2⌋
βj−3,j
X j − m − a − 1t − j − m
X m − 1 t − j − 1 j−2m
.
=
t
m
m−1
m−1
a
a=0
m=2
bc be the bipartite complement of an even cycle C
Let G := C2s
2s with at
least 6 vertices, i.e., s ≥ 3. The vertices and edges of the even cycle C2s
will be V := {x1 , . . . , xs } ⊔ {y1 , . . . , ys } and {{x1 , y1 }, {y1 , x2 }, . . . , {ys , x1 }}
respectively along this section. We will sometimes refer to the two subsets
in the bipartition of V as X and Y . The biadjacency matrix M of G has
exactly two zero entries on each row and column:
M =
0
1
0
0
1
..
.
..
.
1
0
..
.
0
..
.
1
..
.
..
.
..
.
..
.
1
1
..
.
..
.
.
...
0
1
0
0
1
0
...
..
.
..
...
...
0
.
In order to use Hochster’s Formula to determine the graded Betti numbers of
I(G), we need to compute the reduced simplicial homologies H̃i (∆(G[W ]))
for all subsets W of V . The case W = V is solved in Proposition 2.3. Its
proof requires the following lemma.
Lemma 2.2. For every v ∈ V , del∆(G) (v) is acyclic.
Proof. Without loss of generality, let’s choose v = x1 . As observed in Definition 1.5, del∆(G) (x1 ) is the independence complex of G[V \ {x1 }] whose
biadjacency matrix N is obtained by removing the first row of M . Observe
that the first and last columns of N satisfy the condition in Lemma 1.6.4
and hence can be removed. Again, the first and last rows of this new matrix
satisfy the same condition and we remove them. Recursively, when s is odd
(respectively even), we reduce the computation of the homology to the case
of the independence complex of a graph whose biadjacency matrix is a 2 × 3
(respectively 3 × 2) matrix whose central column (respectively row) has its
two entries equal to zero and del∆(G) (v) is acyclic by Lemma 1.6.1.
REGULARITY 3 IN EDGE IDEALS ASSOCIATED TO BIPARTITE GRAPHS
Proposition 2.3. dimK (H̃i (∆(G))) =
1
0
9
if i = 2,
otherwise .
Proof. Since ∆(G) is connected, dimK (H̃0 (∆(G))) = 0. For i > 0, using the
Mayer-Vietoris sequence (2), the previous lemma implies that H̃i (∆(G)) ≃
H̃i−1 (link∆(G) (v)) for every vertex v of G and i > 0. The biadjacency matrix
(or its transpose) of G[V \ (NG (v) ∪ {v})] is a 2 × (s − 1) matrix with exactly
two zero entries, one in each row and located on two different columns.
Applying Lemma 1.6.3 (if s > 3) as many times as necessary, one gets that
H̃i (∆(G)) ≃ H̃i−1 (Σ2 ) and the result follows from Example 1.7.
Let W be now a proper subset of V , W ( V = X ⊔ Y . Set WX := W ∩ X,
WY := W ∩ Y , and denote by kW the number of connected components of
the graph C2s [W ] that are not isolated vertices. Note that if kW 6= 0, then
WX 6= ∅ and WY 6= ∅.
Proposition 2.4.
(1) If WX = ∅ or WY = ∅ then ∆(G[W ]) is acyclic.
(2) If WX 6= ∅, WY 6= ∅ and kW = 0 then
1 if i = 0,
dimK (H̃i (∆(G[W ]))) =
0 otherwise .
(3) If kW > 0 then
dimK (H̃i (∆(G[W ]))) =
kW − 1
0
if i = 1,
otherwise .
Remark 2.5. Observe that Gc = C2s ∪ KX ∪ KY , where KA denotes the
complete graph on a set of vertices A ⊂ V . Thus, Gc [W ] = C2s [W ] ∪ KWX ∪
KWY . Since KWX and KWY are connected, if one of them is empty or if
they are connected to each other in Gc [W ], i.e., if kW 6= 0, then Gc [W ] is
connected. Otherwise, KWX and KWY are its connected components. Thus,
the condition in Proposition 2.4.2 is satisfied if and only if Gc [W ] is not
connected. When WX 6= ∅ and WY 6= ∅, denote by M [W ] the biadjacency
matrix of G[W ]. It is easy to check that if WX 6= ∅ and WY 6= ∅, then
kW = 0 ⇔ NC2s (WX ) ∩ WY = ∅ ⇔ M [W ] has no zero entries.
Proof of Proposition 2.4. If WX = ∅ or WY = ∅, ∆(G[W ]) is a simplex and
1 follows. Assume now that WX 6= ∅, WY 6= ∅ and kW = 0. Then, M [W ]
has no zero entries by Remark 2.5 and, by Lemma 1.6.3, for all i ≥ 0,
H̃i (∆(G[W ])) ≃ H̃i (∆(K2 )) = H̃i (Σ1 ). So 2 follows from Example 1.7.
Assume now that M [W ] has at least one zero entry. First observe that the
number of zero entries in any row and column of M [W ] is at most two (and
that for at least two of the columns or rows, it is one). By Lemma 1.6.3,
the dimension of the reduced homologies will not change if we remove from
M [W ] any row and any column with no zero entry. In other words, since
a row or a column of M [W ] with no zero entry corresponds to an isolated
vertex in C2s [W ], if W ′ is the subset of W formed by all the elements in
W that are not isolated vertices in C2s [W ], one has that H̃i (∆(G[W ])) ≃
10
OSCAR FERNÁNDEZ-RAMOS AND PHILIPPE GIMENEZ
H̃i (∆(G[W ′ ])). Moreover, kW = kW ′ . Using now Lemma 1.6.4, one can
remove from M [W ′ ] any row (resp. column) with exactly one zero entry
and such that in the column (resp. row) where this zero entry is located,
there is another zero entry. Such a row or column of M [W ′ ] corresponds
to a vertex of C2s [W ′ ] of degree one whose (unique) neighbor is of degree
two. Removing such a vertex does not change the number of connected
components of C2s [W ′ ] and it creates in C2s [W ′ ] a vertex of the same kind,
until we reach a vertex of degree one whose neighbor also has degree one.
Thus, H̃i (∆(G[W ])) ≃ H̃i (∆((kW K2 )bc )) for all i ≥ 0. The result is now a
direct consequence of the following technical lemma.
m − 1 if i = 1,
Lemma 2.6. For all m ≥ 1, dimK (H̃i (∆((mK2 )bc ))) =
0
otherwise.
Proof. The biadjacency matrix of the graph (mK2 )bc is an m × m matrix
whose entries are all 1 except the ones on the principal diagonal that are
zero. Denote by Θm its independence complex, Θm := ∆((mK2 )bc ). Since
Θm is connected, dimK (H̃0 (Θm )) = 0 for all m ≥ 1.
In order to determine the homology for i ≥ 1, consider the family of
subcomplexes of Θm , F = {Θm [X], Θm [Y ], {x1 , y1 }, . . . , {xm , ym }} whose
elements we index by x, y, z1 , . . . , zm . Recall that the nerve of F , N (F ), is
the simplicial complex on the vertex set VF := {x, y, z1 , . . . , zm } whose faces
are the subsets of VF such that the intersection of the corresponding elements
in F is non empty. The simplicial complex NS
(F ) has 2m facets, {x, zi } and
{y, zi } for all i = 1, . . . , m. Since Θm = i∈VF Fi , applying the Nerve
Theorem (see, for example, [2, Theorem 10.6]), one gets that H̃i (Θm ) ≃
H̃i (N (F )) for all i ≥ 0. On the other hand, N (F ) = Σyx h{z1 }, . . . , {zm }i
and hence, by Proposition 1.3.2, H̃i (N (F )) ≃ H̃i−1 (h{z1 }, . . . , {zm }i) for all
i ≥ 1 and the result follows.
As a straightforward consequence, one gets that the last row of the Betti
bc ) with a nonzero entry is the one indexed by 4, i.e., regI(C bc ) =
diagram of I(C2s
2s
bc )) = 1 is the only nonzero entry on this row.
4, and that β2s−4,2s (I(C2s
bc )) = 0 if j > i + 4;
• βi,j (I(C2s
1 if i = 2s − 4 ,
bc )) =
• βi,i+4 (I(C2s
0 otherwise .
Proposition 2.7.
Proof. Putting together Propositions 2.3 and 2.4, one has that for every
subset W of V , dimK (H̃i (∆(G[W ]))) = 0 for all i ∈
/ {0, 1, 2}. Moreover,
dimK (H̃2 (∆(G[W ]))) 6= 0 if and only if W = V and dimK (H̃2 (∆(G))) = 1.
The result then follows from Hochster’s Formula (4).
bc ), one
In order to complete the description of the Betti diagram of I(C2s
has to determine the graded Betti numbers on the first two rows, i.e., βi,j
for i + 2 ≤ j ≤ i + 3.
REGULARITY 3 IN EDGE IDEALS ASSOCIATED TO BIPARTITE GRAPHS
11
We start with the first row. Using Hochster’s Formula (4) and Proposition 2.4.2, one needs to determine all the proper subsets W of V such that
WX 6= ∅, WY 6= ∅ and Gc [W ] is not connected. Indeed, βi,i+2 (I(G)) is the
number of induced subgraphs Gc [W ] on i+2 vertices that are non connected.
Let’s denote by CX the cycle on the vertex set X whose edges are {x1 , x2 },
{x2 , x3 }, . . . , {xs , x1 }. Note that the edges of CX correspond to the pairs
{xi , xj } of elements in X such that NC2s (xi ) ∩ NC2s (xj ) 6= ∅.
Lemma 2.8. Assume that Gc [W ] is not connected.
(1) There exists x ∈ WX such that NC2s (x) 6⊂ NC2s (WX \{x}).
(2) |NC2s (WX )| = |WX | + |comp(CX [WX ])| where comp(CX [WX ]) is the
set of connected components of CX [WX ].
Proof. If NC2s (x) ⊂ NC2s (WX \{x}) for some x ∈ WX , then NCX (x) ⊂ WX .
Thus, if NC2s (x) ⊂ NC2s (WX \{x}) for all x ∈ WX , then WX = X and one
can not have WY 6= ∅ and NC2s (WX ) ∩ WY = ∅. This implies that Gc [W ] is
connected by Remark 2.5 and 1 follows.
We will prove 2 by induction on r := |WX |. If WX = {x} then |NC2s (x)| =
2, |{x}| = 1, |comp((Cx )[{x}])| = 1 and the statement holds. Consider now
W such that |WX | = r > 1 and assume that the statement holds for subsets
X ′ such that |WX ′ | = r − 1. By 1, we know that there exists x0 ∈ WX such
that NC2s (x0 ) 6⊂ NC2s (WX \{x0 }), and one has two possibilities:
• If x0 is connected in CX to some x ∈ WX \{x0 }, i.e., if NC2s (x0 ) ∩
NC2s (WX \{x0 }) 6= ∅, then |NC2s (WX )| = |NC2s (WX \{x0 })| + 1. In
this case, |comp((CX )[WX ])| = |comp(CX [WX \{x0 }])|;
• Otherwise, |NC2s (WX )| = |NC2s (WX \{x0 })|+2 and |comp(CX [WX ])| =
|comp(CX [WX \{x0 }])| + 1.
In both cases, applying our inductive hypothesis, one gets that |NC2s (WX )| =
|WX \{x0 })| + |comp(CX [WX ])| + 1 = |WX | + |comp(CX [WX ])|.
The nonzero entries on the first row of the Betti diagram are given by the
following result.
bc )) = 0.
Proposition 2.9.
(1) For all j ≥ s, βj−2,j (I(C2s
(2) For j = 2, . . . , s − 1,
j−1 X
k
X
s k−1 s−k−1 s−k−c
bc
βj−2,j (I(C2s )) =
.
c c−1
c−1
j−k
k=1 c=1
Proof. Consider a proper subset W of V with |W | = j ≥ 2. As already obbc )) in Hochster’s
served, H̃0 (∆(G[W ])) will contribute (by 1) to βj−2,j (I(C2s
c
Formula if and only if G [W ] is not connected.
By Remark 2.5, if Gc [W ] is not connected then |WX | > 0, |WY | > 0 and
|NC2s (WX )| + |WY | ≤ |Y | = s. Thus, |WY | ≤ s − |NC2s (WX )| < s − |WX | by
6 0. It implies that
Lemma 2.8.2 since WX 6= ∅ and hence |comp(CX [WX ])| =
12
OSCAR FERNÁNDEZ-RAMOS AND PHILIPPE GIMENEZ
0 < |WX | < |W | = |WX | + |WY | < s. Thus if |W | ≥ s, Gc [W ] is connected
and 1 follows.
Now for j with 2 ≤ j ≤ s − 1, we have to count how many subsets W
of V with |W | = j satisfy that Gc [W ] is not connected. For each choice of
WX with k elements (1 ≤ k ≤ j − 1 in order to have WX 6= ∅ and WY 6= ∅),
we must choose j − k elements from Y \NC2s (WX ) for WY , so there are
s−|NC2s (WX )|
(CX [WX ])|
= s−k−|comp
possible choices by Lemma 2.8.2. If
j−k
j−k
we fix the number of connected components of
X [WX ] and denote it by c,
C
s−k−1
possible subsets WX
according to [7, Lemma 3.3], there are sc k−1
c−1
c−1
with |WX | = k and |comp(CX [WX ])| = c, and the result follows.
Corollary 2.10. The first and the last nonzero entries on the first row of
bc
bc
bc ) coincide, i.e., β
the Betti diagram of I(C2s
s−3,s−1 (I(C2s )) = β0,2 (I(C2s )).
Proof. For j = s − 1 one has that s−k−c
6= 0 if and only if c = 1. In
j−k
k−1
s−k−1
s−k−c
bc )) =
this case c−1 = c−1 = j−k = 1, and hence βs−3,s−1 (I(C2s
Ps−2
bc
bc
k=1 s = s(s − 2) = |E(C2s )| = β0,2 (I(C2s )).
bc ) will be complete once we
The description of the Betti diagram of I(C2s
give the graded Betti numbers located on the second row. This is our next
result.
bc )) = 0.
Proposition 2.11.
(1) For all j ≥ 2s − 1, βj−3,j (I(C2s
(2) For j = 4, . . . , 2s − 2,
⌊j/2⌋
bc
βj−3,j (I(C2s
)) =
X
m=2
(m−1)
j−2m
X
a=0
2s j − m − a − 1 2s − j − 1 2s − j − m
.
m
m−1
m−1
a
bc ))
Proof. By Proposition 2.4, H̃1 (∆(G[W ])) will contribute to βj−3,j (I(C2s
in Hochster’s Formula (4) if and only if W is a proper subset of V with
|W | = j ≥ 4 such that C2s [W ] has at least 2 connected components that
are not isolated vertices. More precisely, denoting by w(j, m) the number of
proper subsets W of V with |W | = j and such that C2s [W ] has m connected
components that are not isolated vertices, then
⌊j⌋
(5)
βj−3,j (I(G)) =
2
X
(m − 1)w(j, m) .
m=2
In particular, since for any subset W of V with 2s − 1 elements, one has
that C2s [W ] is connected, 1 follows.
Now for j ≤ 2s − 2, denote by W (j, m, a) the set of proper subsets W
of V with |W | = j and such that C2s [W ] has a isolated vertices and m
connected components that are not isolated vertices. Then, w(j, m) =
Pj−2m
a=0 w(j, m, a) where w(j, m, a) = |W (j, m, a)|, and we are reduced to
compute w(j, m, a) for all possible j, m, a.
REGULARITY 3 IN EDGE IDEALS ASSOCIATED TO BIPARTITE GRAPHS
13
As in the proof of [7, Lemma 3.3], observe that a subset W of V can be
represented as a vector of length 2s whose ℓ-th entry is 1 if the ℓ-th element
in V belongs to W , 0 otherwise. Using this correspondence, the number of
nonzero entries in this vector is the number of vertices in C2s [W ] and the
number of blocks of nonzero entries is related to the number of connected
components of C2s [W ]. In order to avoid distinguishing cases as when the
vector starts/ends with 1/0, we will allow to modify the starting vertex and
focus only on vectors whose first entry is 1 and last entry is 0. Denote by
B(2s, j, k) the set of vectors of length 2s, with entries in {0, 1}, whose first
entry is 1 and last entry is 0, and whose j nonzero entries are located in k
different blocks. Let H(j, m, a) be the subset of B(2s, j, m + a) formed by
vectors with m blocks of 1’s of length strictly bigger than 1 and a blocks of
1’s of length 1 and whose first block of nonzero entries has length strictly
bigger than 1. To each element w in H(j, m, a) corresponds 2s elements in
W (j, m, a) (one for each choice of a vertex of C2s as the vertex corresponding
to the first entry of w), and to an element W in W (j, m, a) corresponds m
distinct elements in H(j, m, a) (one for each connected component of C2s [W ]
that we choose as the one that gives the first block of nonzero entries in the
vector). Thus, w(j, m, a) = 2s
m |H(j, m, a)|.
Finally, in order to determine |H(j, m, a)|, note that each element in
H(j, m, a) comes from a vector h in B(2s − m − a, j − m − a, m) by adding
1 in each block of 1’s of h (there are m), and by inserting a times a 1 between two zero entries of h. As already observed
in the proof of [7, Lemma
j−m−a−1 2s−j−1
3.3], |B(2s − m − a, j − m − a, m)| =
m−1
m−1 . Moreover, for any
element in B(2s − m − a, j − m − a, m), each block of zero entries of length
ℓ will give ℓ − 1 places where one can add a 1 between two zero entries, and
since and element in B(2s − m − a, j − m − a, m) has 2s − j zero entries
located in m different
blocks, each element in B(2s − m − a, j − m − a, m)
will provide 2s−j−m
elements in H(j, m, a). Putting all together, one gets
a
2s 2s−j−m j−m−a−1 2s−j−1
and we are done.
that w(j, m, a) = m
a
m−1
m−1
Corollary 2.12. The first and the last nonzero entries on the second row of
bc ) coincide, i.e., β
bc
bc
the Betti diagram of I(C2s
2s−5,2s−2 (I(C2s )) = β1,4 (I(C2s )).
Proof. For j = 2s−2, 2s−j−1
6= 0 if and only if m = 2, and then 2s−j−m
6=
m−1
a
2s 2s−5
bc
0 if and only if a = 0, and hence β2s−5,2s−2 (I(C2s )) = 2 1 = s(2s − 5).
bc )) = 2s 1 2s−5 2s−6 and we are done.
On the other hand, β1,4 (I(C2s
2 1
1
0
Remark 2.13. Recall from [14] that the induced matching number of a graph
G and the Castelnuovo-Mumford regularity of the edge ideal I(G) satisfy
that µ(G)+1 ≤ reg(I(G)). In the case of the bipartite complement of an even
cycle, one can easily determine the induced matching number. Since C6bc =
bc because C 6< C , and hence
3K2 , µ(C6bc ) = 3. Now if s > 3, 3K2 6< C2s
6
2s
bc ) < 3. As C bc [{x , x , y , y }] is formed by the two non-connected
µ(C2s
1 3 1 2
2s
bc ) = 2 if s > 3. Hence, the matching
edges {x1 , y2 } and {y1 , x3 }, µ(C2s
14
OSCAR FERNÁNDEZ-RAMOS AND PHILIPPE GIMENEZ
bc and the Castelnuovo-Mumford regularity of I(C bc ) ⊂ R =
number of C2s
2s
K[x1 , . . . , xs , y1 , . . . , ys ] are related as follows:
bc ) + 1
if s = 3 ,
µ(C2s
bc
reg(I(C2s
)) =
bc ) + 2
µ(C2s
if s ≥ 4 .
This is not the only difference between the cases s = 3 and s ≥ 4. Indeed,
bc ) is not even
R/I(C6bc ) is a complete intersection while, for s ≥ 4, R/I(C2s
Cohen-Macaulay (if it was then it would be Gorenstein which is impossible
since its Betti diagram is not symmetric).
3. Regularity 3 in bipartite edge ideals
In this section we focus on edge ideals associated to bipartite graphs,
which we call bipartite edge ideals. We only consider connected graphs because the Betti numbers of the edge ideal associated to a disconnected graph
can be computed from the Betti numbers of the edge ideals associated to its
connected components; see [13, Lemma 2.1].
Bipartite edge ideals having regularity 2 can be characterized using Theorem 1.10. They are shown to be the edge ideals associated to Ferrer’s graphs
in [3, Theorem 4.2].
Our aim here is to prove our main results, Theorems 3.1 and 3.2. The first
one, analogous to Fröberg’s classical Theorem 1.10, provides a combinatorial
characterization of bipartite edge ideals having regularity 3. The second
one, analogous to Theorem 1.11, gives some extra information when the
bipartite edge ideal I(G) has regularity > 3: we determine the first step
i in the minimal graded free resolution of I(G) where there are syzygies
contributing to a graded Betti number located outside the first two rows of
the Betti diagram. We also show that these syzygies are then concentrated
in degree i + 4 and compute the corresponding graded Betti number βi,i+4 .
Theorem 3.1. Let G be a connected bipartite graph. The edge ideal I(G)
has regularity 3 if and only if Gc has at least one induced cycle (of length
≥ 4) and Gbc does not have any induced cycle of length ≥ 6.
Theorem 3.2. Let G be a connected bipartite graph and set r := |V (G)|.
Assume that reg(I(G)) > 3 and let t = 2s ≥ 6 be the minimal length of an
induced cycle in Gbc . Then:
• βi,j (I(G)) = 0 for all i < t − 4 and j > i + 3;
• βt−4,t (I(G)) = |{induced t-cycles in Gbc }|;
• βt−4,j (I(G)) = 0 for all j > t;
• for any m ∈ Nr such that |m| = t, one has that βt−4,m (I(G)) = 1 if
m ∈ {0, 1}r and Gbc [W ] ≃ Ct where W := {vi ∈ V (G) : mi = 1}.
Otherwise, βt−4,m (I(G)) = 0.
Before we prove these results, let’s recall a construction and some results
from [4] that will be useful. Given a simplicial complex Γ on the vertex set
X = {x1 , . . . , xn } whose facets are denoted by F1 , . . . , Fm , consider m new
REGULARITY 3 IN EDGE IDEALS ASSOCIATED TO BIPARTITE GRAPHS
15
vertices, Y := {y1 , . . . , ym }, and define a new simplicial complex, ∆(Γ), on
the vertex set X ⊔ Y by
(6)
∆(Γ) := ∆′ ∪ ∆X ,
where ∆X denotes the (n − 1)-simplex on the vertex set X, and ∆′ is the
simplicial complex given by ∆′ = {σ ∪ τ : σ ∈ Γ, τ ⊂ {yj : σ ⊂ Fj }}. Then,
[4, Theorem 4.7] states that
(7)
H̃i+1 (∆(Γ)) ≃ H̃i (Γ) , ∀ i ≥ 0 .
Let G be a connected bipartite graph on the vertex set V (G) = X ⊔Y with
X = {x1 , . . . , xn }, Y = {y1 , . . . , ym }, set R := K[x1 , . . . , xn , y1 , . . . , ym ], and
denote as before WX := W ∩ X, WY := W ∩ Y for any subset W of V (G).
One has that the set
(8)
ΓG := {σ ⊂ NGbc (y) : y ∈ Y }
is a simplicial complex on X \ {x ∈ X that are isolated vertices of Gbc }.
Definition 3.3. We say that a subset W ⊂ V (G) is relevant if |W | ≥ 3
and, for all u, v ∈ W , NG[W ] (u) 6⊂ NG[W ] (v).
Remark 3.4.
• If W is not relevant, then there exist u, v ∈ W such
that NG[W ] (u) ⊂ NG[W ] (v) and Hi (∆(G[W ])) ≃ Hi (∆(G[W \ {v}]))
for all i ≥ 0 by Lemma 1.6.5.
• If W is relevant, then |WX | > 1 and |WY | > 1.
Lemma 3.5. If W ⊂ V (G) is relevant, then ∆(ΓG[W ] ) = ∆(G)[W ].
Proof. Denote by Γ := ΓG[W ] the simplicial complex associated to the graph
G[W ] as in (8), let F(Γ) be its set of facets, and set ∆ := ∆(Γ) as defined
in (6). Since W is relevant, G[W ] has no isolated vertex and hence WX is
the vertex set of Γ. Moreover, F(Γ) = {NGbc [W ] (y) : y ∈ WY }. This implies
that ∆ = ∆′ ∪ ∆WX where ∆′ = {σ ∪ τ : σ ∈ ΓG[W ] , τ ⊂ {y ∈ WY : σ ⊂
NGbc [W ] (y)}}. Consider σ ⊂ W . If σ ⊂ WX then σ ∈ ∆WX ⊂ ∆ and also
σ ∈ ∆(G)[W ]. Otherwise, one has that
σ ∈ ∆ \ ∆W X
⇔ σX ∈ ΓG[W ], σY 6= ∅ and σX ⊂ NGbc [W ] (y) , ∀y ∈ σY
⇔ σY 6= ∅ and σX ⊂ NGbc [W ](y) , ∀y ∈ σY
⇔ σY 6= ∅ and {x, y} 6∈ E(G[W ]) , ∀x ∈ σX , ∀y ∈ σY
⇔ σ⊂
6 WX and σ ∈ ∆(G)[W ]
⇔ σ ∈ ∆(G)[W ] \ ∆WX .
Thus, σ ∈ ∆ ⇔ σ ∈ ∆(G)[W ].
Lemma 3.6. Let I := IΓG[W ] be the Stanley-Reisner ideal associated to
the simplicial complex ΓG[W ] and let {m1 , . . . , ms } be its monomial minimal
generating set. One has that:
16
OSCAR FERNÁNDEZ-RAMOS AND PHILIPPE GIMENEZ
• if W is relevant then deg(mi ) ≥ 2, ∀i ∈ [s];
• max{deg(mi ) : i ∈ [s]} ≤ µ(G[W ]).
Proof. If W is relevant and x ∈ W then NG[W ] (x) ( WY . Thus, x ∈
NGbc [W ] (y) for some y ∈ WY , and hence {x} ∈ ΓG[W ]. Therefore, non-faces
must have dimension strictly greater than 1. Since minimal generators of I
correspond to minimal non-faces of ΓG[W ] , the first claim follows.
If g = xi1 · · · xid is a minimal generator of I, then {xi1 , . . . , xid } 6⊂ NGbc [W ] (y)
for all y ∈ WY and xgi 6∈ I for all k ∈ [d]. Hence, for every l ∈ [d],
k
Fl := {xik : k 6= l} ⊂ NGbc [W ] (y) for some element y in WY that we denote
by y(l). Then, xil 6∈ NGbc [W ] (y(l)), or equivalently, xil ∈ NG[W ] (y(l)) and
xik 6∈ NG[W ] (y(l)) if k 6= l. So {{xil , y(l)} : l ∈ [d]} is a set consisting of d
disconnected edges of G[W ]. This implies that µ(G[W ]) ≥ d.
Proof of Theorems 3.1 and 3.2. We will first prove the equivalence in Theorem 3.1 and show that the extra information contained in Theorem 3.2 then
follows quite easily. First assume that reg(I(G)) = 3. By Theorem 1.10, Gc
contains an induced cycle of length l ≥ 4. Moreover, if there exists a subset
W of V (G) such that Gbc [W ] ≃ Cl for some (even) l ≥ 6 then, since on one
hand βi,j (I(G)) ≥ βi,j (I(G[W ])) for all i, j by Hochster’s Formula (4), and
on the other βl−4,l (I(Clbc )) = 1 by Theorem 2.1, one gets that reg(I(G)) ≥ 4,
a contradiction.
Conversely, assume that reg(I(G)) 6= 3. If reg(I(G)) = 2 then there is no
induced cycle in Gc by Theorem 1.10 and the result holds. If reg(I(G)) > 3
then, by Theorem 1.8, there exists i such that βi,i+4 (I) 6= 0. Denote by i4
the smallest integer with this property. By [14, Lemma 2.2], i4 ≥ 2 and if
i4 = 2, then β2,6 (I) 6= 0 is the number of induced subgraphs of G isomorphic
to 3K2 . We only have to notice that (3K2 )bc ≃ C6 to obtain that if i4 = 2,
Gbc contains an induced cycle of length 6. On the other hand, all the items in
Theorem 3.2 follow in this case from Theorem 1.8 and [9, Theorem 2.1] which
states that, for any monomial in R, xm , if one collects at each step of the
minimal multigraded free resolution of I(G), the minimal generators whose
multidegree divides xm , one gets a minimal multigraded free resolution of
I(G)m , the edge ideal whose minimal generators divide xm .
If i4 ≥ 3, Hochster’s Formula (3) tells us that there exists W ⊂ V (G)
such that
(9)
|W | = i4 + 4 and dimK (H̃2 (∆(G)[W ])) > 0 .
As in the case i4 = 2, we will be done using Theorem 1.8 and [9, Theorem 2.1]
if we show that the subsets W ⊂ V (G) satisfying (9) are the ones such that
(10)
G[W ] ≃ (Ci4 +4 )bc .
If W satisfies (10), then it satisfies (9) by Proposition 2.3. Now take W
satisfying (9) and consider the simplicial complex Γ := ΓG[W ]. Note that,
REGULARITY 3 IN EDGE IDEALS ASSOCIATED TO BIPARTITE GRAPHS
17
using Remark 3.4, W has to be a relevant subset of vertices by minimality
of i4 . Applying Lemma 3.5 and (7), one has that
dimK (H̃1 (Γ)) = dimK (H̃2 (∆(G)[W ])) > 0 .
Moreover, dimK (H̃1 (Γ[X ′ ])) = 0 for all X ′ ( WX since ∆(ΓX ′ ) ≃ ∆(G)[W ′ ]
where W ′ = X ′ ⊔ WY and if dimK (H̃1 (Γ[X ′ ])) = dimK (H̃2 (∆(G)[W ′ ])) > 0,
we will reach a contradiction with the minimality of the size of W .
As i4 > 2, we have β2,6 (I(G)) = 0 and hence, by [14, Lemma 2.2], µ(G) =
2. Thus, by Lemma 3.6, IΓ is generated in degree 2, i.e., it is an edge ideal,
and hence we can write Γ = ∆(G∗ ) for some simple graph G∗ on the vertex
set WX . Thus, dimK (H̃1 (∆(G∗ ))) > 0 and dimK (H̃1 (∆(G∗ )[X ′ ])) = 0 for
all X ′ ( WX . Applying Theorem 1.11, we have that Cl < (G∗ )c for some
l ≥ 4 but Cl 6< (G∗ )c [X ′ ] for all X ′ ( WX , so necessarily, (G∗ )c = Cl
and l = |WX |. Therefore, Γ = ∆(Clc ) = Cl and we have |NGbc [W ] (y)| = 2,
for all y ∈ WY . Together with the fact that NGbc [W ] (u) 6⊂ NGbc [W ] (v) for
all u, v ∈ W such that u 6= v (so |NGbc [W ] (u)| =
6 1, u ∈ WX ) and that
P
P
v∈WY degGbc [W ] (v) (so NGbc [W ] (u)| ≤ 2, u ∈ WX ),
u∈WX degGbc [W ] (u) =
this implies that |NGbc [W ] (y)| = 2 for all y ∈ W . Moreover, Gbc [W ] is
connected because Γ is, and hence Gbc [W ] ≃ C|W | .
Remark 3.7. One can find in [14] several examples of edge ideals whose
regularity is 3 or 4 depending on the characteristic of the field K. This
shows that in Theorem 3.1 the bipartite hypothesis can not be removed
since the information provided there only depends on the combinatorics of
the graph G. That is why we restricted ourselves to bipartite edge ideals in
this work. Now observe that even for bipartite edge ideals, it is hopeless to
try an extrapolation of our results for higher values of the regularity as an
example in [4] shows.
4. The non-squarefree case
Let I be an ideal in R := K[x1 , . . . , xn ] generated by monomials of degree
two which is not squarefree. Assume, without loss of generality, that I is
minimally generated by {m1 , . . . , ms } where m1 = x21 , . . . , ml = x2l and
ml+1 , . . . , ms are squarefree for some l ∈ [s]. We define
• Isqf := (ml+1 , . . . , ms ) ⊂ R, and
• Ipol := (x1 y1 , . . . , xl yl , ml+1 , . . . , ms ) ⊂ R∗ := K[x1 , . . . , xn , y1 , . . . , yl ].
The ideal Ipol , called the polarization of I, has the following useful property:
if we provide R and R∗ with a Nn -multigrading such that deg(xi ) = ei for
all i ∈ [n] and deg(yj ) = ej for all j ∈ [l] then, by [10, Corollary 1.6.3],
(11)
βi,m (I) = βi,m (Ipol ), ∀i ≥ 0, ∀m ∈ Nn .
Both ideals Isqf and Ipol are edge ideals, the first one on the vertex set
{x1 , . . . , xn } and the second on {x1 , . . . , xn , y1 , . . . , yl }. We will call G the
non simple graph associated to I and denote, as in the squarefree case,
18
OSCAR FERNÁNDEZ-RAMOS AND PHILIPPE GIMENEZ
I = I(G). Denote by Gsqf and Gpol the simple graphs associated to Isqf
and Ipol , respectively. Observe that Gsqf and Gpol are obtained by removing
loops in G and substituting whiskers for loops in G, respectively.
Definition 4.1. We say that two edges e1 , e2 ∈ E(G) are totally disjoint
provided {u, v} 6∈ E(G) if u ∈ e1 and v ∈ e2 .
Assume that the simple graph Gsqf is connected and bipartite. In this
case, we say that the non simple graph G is bipartite and define the bipartite
complement of G as the bipartite complement of the simple graph Gsqf , i.e.,
Gbc := (Gsqf )bc . We also define the complement of G as the complement of
the simple graph Gsqf , i.e., Gc := (Gsqf )c .
We can complete the characterization of ideals associated to bipartite
graphs having regularity 3 with the non-squarefree case as follows:
Proposition 4.2. Let I ⊂ R be a non-squarefree monomial ideal generated
in degree two and assume that the non simple graph G associated to I is
bipartite. Then, I has regularity 3 if and only if
• G either has two totally disjoint edges or Cl < Gc for some l ≥ 5,
• G does not have three edges that are pairwise totally disjoint, and
• Gbc has no induced cycle of length ≥ 8.
Proof. By (11), reg(I) = 3 if and only if reg(Ipol ) = 3 and, using Theorem 3.1, this occurs if and only if (Gpol )c has an induced cycle of length 4
and (Gpol )bc has no induced cycle of length ≥ 6. Rewriting these properties
of the graph Gpol in terms of the graph G, the result follows.
When reg(I) > 3, the claims in Theorem 3.2 remain valid if G does
not contain three edges that are pairwise totally disjoint since l-cycles in
(Gsqf )c and in (Gsqf )bc coincide with the l-cycles in (Gpol )c and (Gpol )bc
respectively, provided l > 6. However, if G has three edges that are pairwise
totally disjoint, then:
• βi,j = 0 if i ≤ 1 and j > i + 3;
• β2,6 is the number of induced subgraphs of G isomorphic to three
pairwise totally disjoint edges;
• β2,j = 0 for all j > 6;
• considering the Nn -multigrading on R, for all m ∈ Nn such that
|m| = 6, one has that β2,m = 1 if G[{xi : mi = 1}] consists of three
totally disjoint edges. Otherwise, β2,m = 0.
Example 4.3. The ideal I = (x21 , x1 x5 , x2 x5 , x2 x7 , x3 x5 , x3 x6 , x3 x7 , x4 x6 )
satisfies that β2,6 = 1. The bipartite graph Gbc does not have any induced
6-cycle but there are three pairwise disjoint edges in G.
References
[1] L. Avramov, A. Conca and S. Iyengar. Free resolutions over commutative Koszul
algebras. Math. Res. Lett. 17:197–210, 2010.
REGULARITY 3 IN EDGE IDEALS ASSOCIATED TO BIPARTITE GRAPHS
19
[2] A. Björner. Topological methods. In: Handbook of combinatorics, Vol. 1, 2 :1819–
1872. Elsevier, Amsterdam, 1995.
[3] A. Corso and U. Nagel. Monomial and toric ideals associated to Ferrers graphs. Trans.
Amer. Math. Soc. 361:1371–1395, 2009.
[4] K. Dalili and M. Kummini. Dependence of Betti numbers on characteristic.
arXiv:1009.4243, Sep. 2010.
[5] R. Diestel. Graph theory. Graduate Texts in Mathematics, Vol. 173. Springer, Heidelberg, Fourth edition, 2010.
[6] D. Eisenbud, M. Green, K. Hulek, and S. Popescu. Restricting linear syzygies: algebra
and geometry. Compos. Math. 141:1460–1478, 2005.
[7] O. Fernández-Ramos and P. Gimenez. First nonlinear syzygies of ideals associated to
graphs. Comm. Algebra 37:1921–1933, 2009.
[8] R. Fröberg. On Stanley-Reisner rings. In: Topics in algebra, Part 2 (Warsaw 1988),
Banach Center Publ. Vol. 26:57–70, 1990.
[9] V. Gasharov, T. Hibi, and I. Peeva. Resolutions of a-stable ideals. J. Algebra 254:375–
394, 2002.
[10] J. Herzog and T. Hibi. Monomial ideals. Graduate Texts in Mathematics, Vol. 260.
Springer-Verlag, London, 2011.
[11] M. Hochster. Cohen-Macaulay rings, combinatorics, and simplicial complexes. In:
Ring theory, II (Proc. Second Conf., Univ. Oklahoma, Norman, Okla., 1975). Lecture
Notes in Pure and Appl. Math. Vol. 26: 171–223. Dekker, New York, 1977.
[12] H.-T. Ha and A. Van Tuyl. Resolutions of square-free monomial ideals via facet ideals:
a survey. In: Algebra, geometry and their interactions, Contemp. Math. Vol. 448:91–
117. Amer. Math. Soc., Providence, RI, 2007.
[13] S. Jacques and M. Katzman. The Betti numbers of forests. math.AC/0501226, Jan.
2005.
[14] M. Katzman. Characteristic-independence of Betti numbers of graph ideals. J. Combin. Theory Ser. A 113:435–454, 2006.
[15] S. Morey and R. H. Villarreal. Edge ideals: algebraic and combinatorial properties.
arXiv:1012.5329, Dec. 2010.
[16] J. R. Munkres. Elements of algebraic topology. Addison-Wesley Publishing Company,
Menlo Park, CA, 1984.
[17] E. Nevo. Regularity of edge ideals of C4 -free graphs via the topology of the lcmlattice.. J. Combin. Theory Ser. A 118:491–501, 2011.
[18] R. H. Villarreal. Cohen-Macaulay graphs. Manuscripta Math. 66:277–293, 1990.
[19] R. H. Villarreal. Monomial algebras. Monographs and Textbooks in Pure and Applied
Mathematics, Vol. 238. Marcel Dekker Inc., New York, 2001.
University of Valladolid
E-mail address: [email protected]
University of Valladolid
E-mail address: [email protected]
| 0math.AC
|
Geometrically Exact Finite Element Formulations for Curved Slender Beams:
Kirchhoff-Love Theory vs. Simo-Reissner Theory
Christoph Meiera,∗, Wolfgang A. Wallb , Alexander Poppb
a Mechanosynthesis
b Institute
Group, Massachussets Institute of Technology, 77 Massachusetts Avenue, Cambridge, 02139, MA, USA
for Computational Mechanics, Technical University of Munich, Boltzmannstrasse 15, D–85748 Garching b. München, Germany
arXiv:1609.00119v1 [cs.CE] 1 Sep 2016
Abstract
The present work focuses on geometrically exact finite elements for highly slender beams. It aims at the proposal of
novel formulations of Kirchhoff-Love type, a detailed review of existing formulations of Kirchhoff-Love and SimoReissner type as well as a careful evaluation and comparison of the proposed and existing formulations. Two different
rotation interpolation schemes with strong or weak Kirchhoff constraint enforcement, respectively, as well as two different choices of nodal triad parametrizations in terms of rotation or tangent vectors are proposed. The combination of
these schemes leads to four novel finite element variants, all of them based on a C 1 -continuous Hermite interpolation
of the beam centerline. Essential requirements such as representability of general 3D, large-deformation, dynamic
problems involving slender beams with arbitrary initial curvatures and anisotropic cross-section shapes, preservation
of objectivity and path-independence, consistent convergence orders, avoidance of locking effects as well as conservation of energy and momentum by the employed spatial discretization schemes, but also a range of practically
relevant secondary aspects will be investigated analytically and verified numerically for the different formulations. It
will be shown that the geometrically exact Kirchhoff-Love beam elements proposed in this work are the first ones of
this type that fulfill all these essential requirements. On the contrary, Simo-Reissner type formulations fulfilling these
requirements can be found in the literature very well. However, it will be argued that the shear-free Kirchhoff-Love
formulations can provide considerable numerical advantages such as lower spatial discretization error level, improved
performance of time integration schemes as well as linear and nonlinear solvers or smooth geometry representation
as compared to shear-deformable Simo-Reissner formulations when applied to highly slender beams. Concretely,
several representative numerical test cases confirm that the proposed Kirchhoff-Love formulations exhibit a lower
discretization error level as well as a considerably improved nonlinear solver performance in the range of high beam
slenderness ratios as compared to two representative Simo-Reissner element formulations from the literature.
Keywords: Geometrically exact beams, Simo-Reissner, Kirchhoff-Love, Finite elements, Objectivity, C 1 -continuity
1. Introduction
Highly slender fiber- or rod-like components represent essential constituents of mechanical systems in countless fields
of application and scientific disciplines such as mechanical engineering, biomedical engineering, material science
and bio- or molecular physics. Examples are high-tensile industrial webbings, fiber-reinforced composite materials,
fibrous materials with tailored porosity, synthetic polymer materials or also cellulose fibers determining the characteristics of paper [49, 146, 91, 90, 56, 116]. On entirely different time and length scales, such slender components
are relevant when analyzing the supercoiling process of DNA strands, the characteristics of carbon nanotubes or the
Brownian dynamics within the cytoskeleton of biological cells, a biopolymer network of highly slender filaments that
crucially influences biologically relevant processes such as cell division and cell migration [45, 108, 148, 153]. Often,
these slender components can be modeled as 1D Cosserat continua based on a geometrically nonlinear beam theory.
∗ Corresponding
author
Email address: [email protected] (Christoph Meier)
September 2, 2016
In all mentioned cases, mechanical contact interaction crucially influences the overall system behavior. While in the
authors’ recent works [106] and [107] a novel beam-to-beam contact formulation has been proposed that faces the
challenges arising from high beam slenderness ratios and complex 3D contact configurations, the present contribution
focuses on the development of geometrically exact beam element formulations suitable for an accurate, efficient and
robust modeling of highly slender beams. While the vast majority of geometrically exact beam element formulations
available in the literature is based on the Simo-Reissner theory of thick rods, the current work proposes geometrically
exact finite elements based on the Kirchhoff-Love theory of thin rods that are taylored for high beam slenderness ratios.
Basically, two essential motivations for applying a beam theory instead of a 3D continuum mechanics theory for the
modeling of slender bodies can be identified: In the early days of beam theories, it was the accessibility of analytic
solutions as for example ”Euler’s Elastica”, even for large deformation problems, that motivated the development and
application of one-dimensional theories. Nowadays, it is the knowledge that the modeling of highly slender bodies
via beam theories yields considerably more efficient, but also more well-posed, numerical formulations as it would
be the case for 3D continuum theories. So-called induced beam theories can be regarded as reduced 1D continuum
theories consistently derived from the 3D theory of continuum mechanics under consideration of a basic kinematic
constraint that reflects the deformation states expected for slender bodies in a reasonable manner. Such 1D beam
theories typically allow to describe the motion and deformation of slender bodies in 3D space on the basis of proper
kinematic, kinetic and constitutive resultant quantities. In the case of induced beam theories, these resultant quantities
can for example be derived via integration of 3D stress measures over the beam cross-section. The 3D stress measures
typically result from the constrained 3D displacement field as well as standard 3D strain measures and constitutive
relations. In this context, the cross-section of a beam represents the collection of all material points sharing the same
beam length coordinate in the stress-free configuration. On the contrary, so-called intrinsic beam theories directly
postulate the 1D resultant quantities. These theories are internally consistent in the sense that the resultant quantities
as well as the 1D relations connecting these quantities still fulfill essential mechanical principles such as equilibrium
of forces and moments, conservation of energy or rather existence of work conjugated stress-strain pairs, observer invariance or path-independence of conservative problems. Nevertheless, intrinsic beam theories are decoupled from the
3D continuum mechanics theory. Typically, the postulated constitutive constants relating stress and strain measures
are determined experimentally, while the constitutive constants of induced beam theories follow directly from the
corresponding 3D constitutive laws. Such postulated constitutive laws based on experimentally determined constants
are favorable for applications where no 3D continuum foundation exists: Considering the low number of discrete
molecules distributed over the thickness of macromolecules as occurring for example in biopolymer networks, DNA
strands or carbon nanotubes - to come back to the applications mentioned above - no 3D continuum theory can be
applied in a reasonable manner. Nevertheless, it is well-established that these slender components can be described in
good approximation by 1D continuum theories and associated experimentally determined constitutive constants [126].
A compromise between induced and intrinsic theories are so-called semi-induced beam theories, where only the constitutive law is postulated. The remaining kinetic and kinematic relations are consistently derived from the 3D theory.
Based on the Bernoulli hypothesis of undeformed cross-sections and the work of Euler, the “Kirchhoff beam” [86],
proposed by Kirchhoff zu Heidelberg in 1859, was the first formulation allowing for arbitrary initial curvatures and
large three-dimensional deformations including the states of bending and torsion. In 1944, this theory was enhanced
by Love [101] to also account for small axial tension. A comprehensive historic overview of these early developments
is given in the work of Dill [47]. It was Reissner in 1972 for the two-dimensional case [114] and in 1981 for the general three-dimensional case [115], who completed the theory by two additional deformation measures representing
the shear deformation of the beam. While the 3D problem statement of Reissner was still based on some additional
approximations, Simo [132] extended the work of Reissner to yield a formulation that is truly consistent in the sense
of a semi-induced beam theory. Thus, starting from a basic kinematic assumption, all kinetic and kinematic quantities
and relations are consistently derived from the 3D continuum theory, while the constitutive law has been postulated.
Originally, this theory has been denoted as geometrically exact beam theory. Nowadays, it is also referred to as
Simo-Reissner beam theory. According to the definition of Simo [132], also in this work, a beam theory is denoted
as geometrically exact, if “the relationships between the configuration and the strain measures are consistent with the
virtual work principle and the equilibrium equations at a deformed state regardless of the magnitude of displacements,
rotations and strains” ([43], p. 1126). For that reason, also the notation “finite-strain beams” has been applied in the
2
original work [132]. However, as later argued by several authors (see e.g. [43]) and in accordance with derivations in
the literature (see e.g. [3, 2, 101], but also Section 3.5 of this work), a consistency of the geometrically exact beam
theory and the 3D theory of continuum mechanics in the sense of a (fully) induced beam theory can only be assumed
as long as small strains are considered. The fulfillment of the basic kinematic assumption of rigid cross-sections underlying the geometrically exact beam theory requires pointwise six (translational and rotational) degrees of freedom
in order to uniquely describe the (centroid) position and orientation of the cross-sections. Consequently, this beam
theory can be identified as 1D Cosserat continuum [38], derived from a 3D Boltzmann continuum with pointwise three
(translational) degrees of freedom. While there exists a variety of beam theories that also consider in-plane as well as
out-of-plane cross-section distortion, this contribution focuses on geometrically exact beam formulations on the basis
of the rigid cross-section assumption as applied by Simo and Reissner. Furthermore, throughout this work, the notion
Simo-Reissner theory will be preferred since the notion geometrically exact beam theory, when following the definition presented above, also applies to consistently derived shear-free formulations on the basis of the Kirchhoff-Love
theory. In the remainder of this work, the notion “shear-free” represents the opposite of ”shear-deformable” and thus
is equivalent to “vanishing shear strains”, but, of course, not to ”vanishing shear stresses”. Unfortunately, no absolute
consensus concerning naming of different beam models can be found in the literature. For that reason, the following nomenclature, trying to be consistent with the most important representatives in the literature, will be applied here:
Name
Simo-Reissner
(Anisotropic) Kirchhoff-Love
Straight Kirchhoff-Love
Isotropic Kirchhoff-Love
Nonlinear Euler-Bernoulli
Torsion-Free Simo-Reissner
Torsion-Free Kirchhoff-Love
Timoshenko
Euler-Bernoulli
Inextensible ...
Theory
Nonlinear
Nonlinear
Nonlinear
κ0 , 0
+
+
-
I2 , I3
+
+
+
Tension
+
+
+
Shear
+
-
Torsion
+
+
+
Bending
+
+
+
Nonlinear
-
-
+
-
+
+
Nonlinear
Nonlinear
Linear
Linear
...
...
+
+
...
+
+
+
+
-
+
+
-
+
+
...
+
+
+
+
...
Table 1: Nomenclature applied within the current contribution for different (geometrically exact) beam theories / models.
Geometrically nonlinear beam models capturing the modes of axial tension, torsion and bending and being appropriate
for initially straight beams with isotropic cross-section shapes (identical moments of inertia of area I2 = I3 ) are denoted
as nonlinear Euler-Bernoulli beams. The extension to anisotropic cross-section shapes (different moments of inertia of
area I2 , I3 ) and arbitrary initial curvatures κ0 , 0 is referred to as Kirchhoff-Love theory. In order to simplify the subsequent comparison of shear-free beam formulations from the literature the more refined nomenclature of i) isotropic,
ii) straight and iii) anisotropic Kirchhoff-Love formulations is introduced for theories capturing i) neither anisotropic
cross-section shapes nor initial curvatures (=nonlinear Euler-Bernoulli), ii) anisotropic cross-section shapes but no
initial curvatures or iii) both, anisotropic cross-section shapes and initial curvatures. The supplementation by shear deformation modes is covered by the Simo-Reissner theory. Euler-Bernoulli or Kirchhoff-Love formulations that neglect
the mode of axial tension are denoted as inextensible Euler-Bernoulli or Kirchhoff-Love variants. Recently, sheardeformable and shear-free beam formulations have been proposed that also neglect the mode of torsion [121, 105].
These formulations are capable of accurately modeling ropes and cables (providing a mechanically consistent bending
stabilization), quasi-continua such as chains but also general 1D continua with full bending and torsional stiffness if
some restrictions concerning external loads (no torsional moment loads) and initial geometry (I2 = I3 ,κ0 = 0) are fulfilled. Eventually, the restriction of shear-free or shear-deformable theories to the geometrically linear regime yields
the linear Euler-Bernoulli and Timoshenko beam models [144]. An overview of these beam models is given in Table 1.
By identifying the configuration space underlying the geometrically exact beam theory as nonlinear, differentiable
manifold with Lie group structure and by pointing out important algorithmic consequences resulting from the nonadditivity and non-commutativity of the associated group elements, the original work by Simo [132] and the subse3
quent work by Simo and Vu-Quoc [134] laid the foundation for abundant research work on this topic in the following
years. The static beam theory [132, 134] has been extended to dynamics by Cardona and Geradin [34, 35] and by Simo
and Vu-Quoc [135]. The contributions of Kondoh et al. [88], Dvorkin et al. [50] as well as Iura and Atluri [78] can be
regarded as further pioneering works in this field. These contributions mark the starting point for the development of
a large variety of geometrically exact beam element formulations [44, 71, 74, 137, 110, 81, 79, 124] which basically
differ in the type of rotation interpolation (e.g. interpolation of incremental, iterative or total rotational vectors), the
choice of nodal rotation parametrization (via rotation vectors, quaternions etc.), the type of iterative rotation updates
(multiplicative or additive), or the time integration scheme applied to the rotational degrees of freedom (e.g. based on
additive or multiplicative rotation increments). Also extensions of the geometrically exact beam theory to arbitrary
cross-section shapes with shear centers differing from the cross-section centroid can be found [65]. An overview of
the most important developments at that time is exemplarily given in the text books of Crisfield [40] as well as Geradin
and Cardona [58]. An break in this development is given by the works of Crisfield and Jelenić [43, 80], who have
shown that none of the rotation field discretizations of the formulations existent at that time could preserve both of
the important properties objectivity and path-independence (see also [76] for a discussion of this topic). Furthermore,
in [43] and [80], a new, objective and path-independent orthogonal interpolation scheme was proposed that directly
acts on the rotation manifold and not on any of its rotation vector parametrizations as done in the works before.
This formulation was the starting point for the development of many alternative rotation interpolation strategies for
geometrically exact beams that also preserve these properties. Among others, orthogonal interpolations of relative
rotation vectors (see e.g. [60, 122]) or quaternions (see e.g. [59, 117, 158]), non-orthogonal interpolation strategies
in combination with modified beam models (see e.g. [120, 22, 51]) and non-orthogonal interpolation strategies with
subsequent orthogonalization (see e.g. [117]) can be identified. As reported in the original work [43, 80], the rotation
interpolation scheme proposed by Crisfield and Jelenić can exactly represent the state of constant curvature. Thus, it
can be interpreted as geodesic, i.e. shortest, connection between two configurations on the rotation manifold. Consequently, these geodesic rotation interpolation schemes represent the counterpart to linear interpolations of translational
quantities. The works [24, 25] as well as the recent contributions [122, 145, 139, 138] can be identified as further
geometrically exact beam element formulations based on geodesic interpolations of the rotational (and translational)
primary variable fields. A 2D extension of these so-called helicoidal interpolations to higher-order elements is given
in [48]. A formulation with smooth centerline representation based on an isogeometric collocation scheme is proposed
in [149]. Besides these purely displacement-based elements, also interpolation schemes directly acting at the strain
level combined with a subsequent derivation of the position and rotation field via integration (see e.g. [156, 157, 145])
as well as mixed formulations [125] have been proposed. Furthermore, a variety of contributions considering time
integration of rotational variables can be found in this context [136, 26, 61, 14, 21, 75, 98, 119, 83, 46].
While all the formulations presented above have been based on the finite element method (FEM), also discrete
representatives of Simo-Reissner beam formulations based on finite difference schemes can be found in the literature [93, 92, 100, 82]. These are often denoted as discrete elastic rods and based on the concept of discrete differential
geometry (DDG). In the context of finite element formulations for geometrically nonlinear beam problems, again, a
variety of alternatives to the geometrically exact formulations considered in the last two paragraphs can be found.
The maybe most prominent representatives of these alternatives are the corotational method [39, 40, 42, 53, 57] as
well as Absolute Nodal Coordinate (ANC) [129, 128] and solid beam element [11, 55] formulations. The corotational
technique was initially introduced by Wempner [152] as well as Belytschko et al. [17, 18] and shows strong similarities to the ”natural approach” of Argyris et al. [4]. The basic idea is to split the overall non-linear deformation
into a contribution stemming from large rotations and a part stemming from local deformations expressed in a local,
”corotated” frame. Often, the local deformation can be modeled on the basis of first- (or second-) order theories such
that the entire degree of nonlinearity is represented by the rotation of the local frame. The basic idea of ANC beam
element formulations is to employ standard shape functions as known from solid finite element formulations in order
to interpolate the 3D displacement field within the beam. Instead of introducing a kinematic constraint and deriving a
resultant 1D model, different polynomial degrees are typically applied for the interpolation in beam length direction
and in transverse directions. Numerical comparisons as performed e.g. by Romero [117, 118] and Bauchau et al. [15]
advocate geometrically exact beams in general, and orthogonal triad interpolation schemes (see e.g. [43]) in particular, with regard to computational accuracy and efficiency. Especially the arguments I-III stated below in terms of
possible benefits of shear-free as compared to shear-deformable beam element formulations in the range of high beam
4
slenderness ratios are pronounced even stronger when comparing geometrically exact Kirchhoff-Love beam element
formulations with ANC beam element formulations, with the latter additionally exhibiting highly stiff in-plane deformation modes when dealing with very slender beams. On the other hand, a clear distinction between corotational and
geometrically exact formulations will not always be possible. For example, the interpolation scheme of the geometrically exact formulation proposed by Crisfield and Jelenić [43, 80] is also based on the definition of an element-local,
corotated reference triad and can consequently be interpreted as corotational formulation at the same time. All in all,
it can be stated that geometrically exact finite element formulations have become well-established in the meantime
and can arguably be regarded as state-of-the-art methods for the computational treatment of geometrically nonlinear
beam problems. An overview of the different FEM discretiation techniques is given in Figure 1.
Figure 1: Overview of different discretization techniques for nonlinear problems of slender continua based on the finite element method.
In the context of the geometrically nonlinear Kirchhoff-Love beam theory, several discrete realizations based on finite
difference schemes have recently been proposed [62, 19, 20, 95, 1]. In contrast to the Simo-Reissner theory, also several works based on an analytic treatment of Kirchhoff-Love beam problems exist in the modern literature [130, 94].
Interestingly, most approaches of these two categories can be found in the field of bio- or molecular physics. Although, the theoretical basis of shear-free Kirchhoff-Love beam formulations has a much longer tradition than the
Simo-Reissner theory of shear-deformable beams, there are only a few geometrically nonlinear shear-free finite element representations, which have not reached the excellent properties of geometrically exact Simo-Reissner formulations so far. In his recent works [5] and [6], Armero and Valverde gave a historic overview of existing Kirchhoff finite
elements and pointed out their drawbacks. Accordingly, the first Kirchhoff type element formulations have applied
different interpolations (C 0 -continuous Lagrange polynomials for the axial displacements and C 1 -continuous Hermite
polynomials for the transversal displacements) to the different displacement-components, which led to a loss of objectivity [5]. In later works, the objectivity could be preserved by employing e.g. trigonometric shape functions, but
the corresponding formulations were limited to the investigation of plane circular arches (see e.g. [8] or [9]). A different approach was the development of framework or corotational shear-free beams [41, 66, 67, 68, 96], a category of
formulations, which naturally preserves the objectivity of the continuous problem. As pointed out in [5], these types
of Kirchhoff type formulations often exhibit a comparatively poor accuracy; a fact, which can directly be traced back
to the lack of an exact representation of the kinematic quantities. A further critical issue relevant in the context of thin
Kirchhoff beams is membrane locking, a locking phenomenon given distinction to by Stolarski and Belytschko [141].
In general, membrane locking denotes the inability of curved structural elements, e.g. beams or shells, to represent the
inextensibility constraint of vanishing membrane / axial strains. For thin Kirchhoff beams, [54] was one of the first
contributions in which this effect was investigated by relating the beam slenderness ratio to the condition number of the
stiffness matrix, but without explicitly using the term locking. Diverse methods have been proposed in the literature in
order to avoid membrane locking of Kirchhoff rods. Amongst others, these are the approaches of reduced / selective
5
integration (see e.g. [109, 111, 112]), assumed strains based on the Hu-Washizu functional (see e.g. [36, 85, 97]),
assumed stresses based on the Hellinger-Reissner functional (see e.g. [33, 109]) or penalty relaxation / stabilization
techniques in combination with membrane correction factors (see e.g. [102, 143]). Most of these works are limited to
2D beam problems. A historic overview concerning the development of Kirchhoff beam elements in general and the
key issues of objectivity and membrane locking in particular is given in the recent works of Armero and Valverde [5, 6].
The shear-free beam elements presented by now are typically based on additional kinematic assumptions, thus not
being consistent with the concept of geometrically exact beams. Besides these contributions considered so far, also
finite element and finite difference discretization approaches for Kirchhoff beams can be found in the literature which
are in fact geometrically exact, but which rely on global interpolation strategies [153, 20, 52]. These are typically
based on a rotation or curvature interpolation strategy and a subsequent integration of the rotation field along the
entire beam length in order to yield an explicit beam centerline representation. Unfortunately, these global approaches
typically rely on a serial finite element evaluation and yield dense system matrices and not the very desirable sparse
system matrices with small bandwidths as typical for standard FEM approaches. Consequently, these approaches
are not suitable for high-performance computing and will not further be considered in the following. The number of
existing geometrically exact (locally supported) finite element formulations on the basis of the Kirchhoff-Love theory
is very limited. For example, the recent contribution of Sansour [123] proposes an energy-momentum method for
2D, initially straight geometrically exact elements based on the nonlinear Euler-Bernoulli theory. On the other hand,
Armero and Valverde developed plane and three-dimensional locking-free Kirchhoff-Love beam element formulations accounting for arbitrarily curved initial geometries as well as anisotropic cross-section shapes that guarantee the
fundamental properties of objectivity and geometrical exactness [5, 6]. However, these beam elements only cover the
geometrically linear case of infinitesimal deformations, but not the general large-deformation case.
Only a few geometrically exact, 3D, large-deformation, shear-free beam elements can be found in the literature. Arguably the first formulations of this kind have been proposed by Boyer [28] and Weiss [150, 151]. In his recent
work [27], Boyer extended the original formulation [28] for the modeling of undersea cables. Also the essential requirement of objectivity and path-independence ar fulfilled by the approaches of Boyer and Weiss. However, these
geometrically nonlinear Euler-Bernoulli formulations only treat the special case of beams with circular cross-sections
and a straight initial configuration, i.e. the case of beams with rotationally symmetric reference geometry in the following denoted as isotropic case (see first column of Table 2). This limitation simplifies the theory considerably, and
already the modeling of simple piecewise straight frames is difficult since no variables are available that determine the
cross-section orientation required for kinematic constraints at beam-to-beam joints. The more recent contributions of
Zhang and Zhao [154, 155] allow for anisotropic cross-sections but still focus on initially straight beams (denoted as
straight case in the second column of Table 2). However, later it will become clear that the first-order twist angle interpolation underlying these formulations might in general not allow for optimal spatial convergence rates considering
the employed third-order centerline interpolation. In the very recent works of Greco et al. [63, 64], first steps towards
geometrically nonlinear isogeometric Kirchhoff-Love beam elements accounting for initial curvature and anisotropic
cross-sections have been made. However, the formulation has only been applied to geometrically linear examples.
Bauer et al. [16] adapted the ideas of Greco and facilitated an application to geometrically nonlinear examples. These
formulations are denoted as anisotropic case in Table 2. Unfortunately, important properties such as objectivity and
path-independence have not been considered in the works of Greco and Bauer. Also the question of membrane locking has not yet been consistently treated within the geometrically nonlinear realm by the textitisotropic, straight and
anisotropic Kirchhoff-Love formulations considered in this paragraph. Nevertheless, they can be considered as the
most elaborate geometrically exact Kirchhoff-Love beam element formulations from the literature, since being the
only ones accounting for the numerical treatment of general 3D, large-deformation problems. Consequently, these
formulations will be in the focus of detailed evaluations and comparisons performed throughout this work.
All in all, it can be concluded that none of the existing geometrically exact shear-free beam elements of Kirchhoff-Love
type is comparable to the existing shear-deformable formulations of Simo-Reissner type in terms of generality and
fulfillment of essential properties so far. Also detailed comparisons of shear-free and shear-deformable geometrically
exact beam elements in the range of high slenderness ratios are still missing. This backlog is the motivation for the development of novel geometrically exact Kirchhoff-Love beam elements fulfilling the following essential requirements:
6
1) Geometrical exactness: As already mentioned above, the proposed beam element formulations have to be geometrically exact in the sense that the derived deformation measures are consistent with the virtual work principle and
the equilibrium equations at any deformed state independent of the magnitude of displacements, rotations and strains.
2) Representability of general geometries and external loads: The most general case of 3D, large-deformation
problems of thin beams with a) anisotropic cross-section shapes as well as b) arbitrary initial geometries and curvatures have to be represented. c) The proposed rotation interpolation schemes have to be capable of representing such
general scenarios without exhibiting any singularities for practically relevant configurations.
3) Fulfillment of essential mechanical principles: The essential mechanical principles of objectivity, i.e. observer
or frame invariance, as well as path-independence have to be preserved by the employed discretization schemes.
4) Fulfillment of general requirements on spatial finite element discretization: For all finite element formulations presented in this work, accuracy will be verified in terms of convergence towards proper reference solutions.
Furthermore, the following criteria have to be fulfilled: a) Optimal spatial convergence rates measured in appropriate
error norms have to be achieved. b) Furthermore, effective methods for the avoidance of membrane locking are required such that no remaining locking-related deterioration of the spatial convergence behavior will be observed for
the resulting finite elements, even in the range of very high slenderness ratios. c) Further properties such as conservation of linear momentum, conservation of angular momentum and conservation of energy for arbitrarily rough spatial
discretizations are very desirable for the proposed spatial interpolation schemes. d) It can also be beneficial for beam
element formulations to fulfill simple patch tests, e.g. to exactly represent the states of constant axial tension, torsion
and bending. e) It may be preferable that the finite element solution is independent from the chosen node numbering.
5) Fulfillment of general requirements on temporal discretization: The main focus of this work lies on the development and evaluation of spatial finite element discretizations for geometrically nonlinear beam problems. Nevertheless, it is at least required that a) dynamics can be represented in general, and that b) energy-stable time integration
can be achieved. c) Often, energy-momentum conserving time integration schemes may be favorable.
6) Simple realization of essential boundary conditions and joints: Choices of nodal primary variables, especially
with respect to the rotation parametrization, are demanded that enable the formulation of practically relevant Dirichlet
boundary conditions but also of node-wise joints between several beams without the need of additional constraint
equations. This is not standard for most of the existing geometrically exact shear-free beam element formulations.
7) Avoidance of Lagrange multipliers and saddle point systems: Also existing geometrically exact beam elements
of Simo-Reissner type can be subjected to the Kirchhoff constraint via additional Lagrange multiplier fields. Such a
procedure typically results in saddle point systems and the need of a special class of linear solvers or requires global
condensation strategies. The shear-free beam elements considered here should neither rely on Lagrange multipliers
nor should they yield in saddle point systems. All required calculations have to be feasible in an element-local manner.
8) Suitability for high-performance computing: In the literature, several finite difference and finite element discretizations of Kirchhoff-Love beams have been proposed that rely on global strategies for rotation field construction. Even though, these schemes show otherwise desirable properties, they typically suffer from two elementary
drawbacks: Mostly, these schemes result in dense discrete system matrices and depend on a successive, i.e. serial,
evaluation of the individual finite elements within a discretization. These two properties make such formulations virtually impossible for high-performance computing. From the finite elements considered here, it is required to result
in sparse system matrices with small bandwidths and to be suitable for parallel computing element evaluation routines.
In the authors’ recent contribution [104], the first 3D large-deformation geometrically exact Kirchhoff-Love beam
element formulation that fulfills the essential properties of objectivity and path-independence and that is capable of
representing arbitrary initial curvatures and anisotropic cross-section shapes has been proposed. In the subsequent
work [105], also the important question of membrane locking has successfully been addressed. The current con7
Type
Source
Formulation
1) Geometrically Exact
2a) Initial Curvature
2b) Anisotropic Cross-Sections
2c) Singularity-Free
3) Objective & Path-Independent
4a) Optimal Convergence Order
4b) Locking Avoided
4c) Conservation Properties (FE)
4d) Patch Test Passed
4e) Symmetric Discretization
5a) Suitable for Dynamics
5b) Energy-Stable (FD)
5c) Conservation Properties (FD)
6) Realization of DBCs & Joints
7) Lagrange Multipliers Avoided
8) High-Performance Computing
Kirchhoff-Love
Isotropic
+
+
+
+
o
+
+
o
o
+
+
Literature
Authors
Straight
Anisotropic [104, 105]
Current
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
?
o
+
+
+
+
o
+
+
+
?
?
?
+
?
?
+
+
+
+
+
+
+
+
+
+
Simo-R.
Literature
[43, 80]
+
+
+
+
+
+
+
+
o
+
+
+
+
+
Table 2: Fulfillment of essential requirements on geometrically exact beam elements: Comparison of existing and proposed formulations.
tribution extends these methodologies by providing considerable improvements in terms of accuracy and practical
applicability, a generalization to dynamic problems as well as theoretical and numerical comparisons to existing formulations. In Table 2, the fulfillment of the essential requirements stated above is verified for the existing categories
of 3D, large-deformation geometrically exact Kirchhoff-Love beam elements identified so far, i.e. for the isotropic,
straight and anisotropic formulations from the literature, the author’s recent contributions [104, 105] as well as the
formulations proposed in the current work. Of course, not all of the stated requirements have explicitly been investigated in the mentioned references. However, the evaluation results presented in Table 2 have been derived on the
best of the authors’ knowledge. Explanations and Derivations required for the evaluation of the individual categories
can be found in Sections 5-10 in form of paragraphs beinning with the expression ”Review: ...”. In general, the
symbols ”+”, ”o” and ”-” in Table 2 mean that a requirement is fully, partly or not fulfilled. The question mark ”?”
indicates that the required information could not definitely be extracted from the literature. According to Table 2, the
formulations newly proposed in the current work can indeed close most of the gaps left by existing geometrically exact Kirchhoff-Love formulations. However, several existing Simo-Reissner beam element formulations already fulfill
most of the stated requirements, a fact that might indicate the increased complexity of consistently incorporating the
additional constraint of vanishing shear strains in Kirchhoff-Love formulations. In order to underline this statement,
also the well-known Simo-Reissner type formulation proposed by Crisfield and Jelenić [43, 80], the first one of this
type preserving objectivity and path-independence, is represented as one possible example in Table 2 (right column).
Given the excellent properties of existing geometrically exact shear-deformable finite element formulations, the question arises which benefits can be gained from applying Kirchhoff-Love instead of Simo-Reissner beam element formulations. It is quite obvious, that the range of low beam slenderness ratios, where shear-deformation is not negligible,
requires the application of beam element formulations taking this fact into account. Also the underlying continuum
theory, based on an unconstrained 1D Cosserat continuum with pointwise six degrees of freedom, and the resulting
discrete problem statement seem to be easier to be formulated for the Simo-Reissner case. However, in the range
of high beam slenderness ratios, thus, exactly in scenarios prevalent in many of the practically relevant applications
mentioned in the beginning of this contribution, shear-free beam element formulations of Kirchhoff-Love type may
exhibit considerable advantages. In the following, some of these potential benefits shall be outlined:
I) Lower discretization error level: The most essential difference between the proposed Kirchhoff-Love and exist8
ing Simo-Reissner beam element formulations lies in the neglect of shear deformation in the beam theory underlying
the former category. This property is independent from the chosen discretization strategy. Consequently, if otherwise
comparable interpolation strategies are applied, it can be assumed that the Kirchhoff type formulations require less
degrees of freedom in order to yield the same polynomial approximation quality, and eventually the same discretization error level, since no additional primary variables are required in order to represent shear deformation.
II) Numerical benefits in the range of high slenderness ratios: In the range of very high slenderness ratios, the
influence of the shear modes on the overall beam deformation is not only negligible, it can also be very beneficial to
abstain from these high stiffness contributions from a numerical point of view: Mechanical problems of slender beams
typically lead to stiff differential equations and ill-conditioned numerical problems deteriorating the performance of
time integration schemes, nonlinear solvers and linear solvers. The avoidance of the stiff shear mode contributions
can considerably improve the situation. Concretely, detailed numerical investigations on several numerical test cases
involving highly slender beams will be considered. These test cases will reveal a considerably improved performance
of nonlinear solution schemes when Kirchhoff type instead of Reissner type discretizations are employed. Similar
trends will be predicted - at least theoretically - for the behavior of linear solvers and time integration schemes.
III) Smooth geometry representation: The proposed Kirchhoff-Love beam elements will be based on C 1 -continuous
centerline interpolations. These interpolations will eventually result in smooth beam-to-beam contact kinematics, a
property that is highly desirable in order to yield efficient and robust contact algorithms (see e.g. [106, 107]).
IV) Derivation of reduced models: In [105], a special reduced model, denoted as torsion-free beam theory, has
been derived from the general Kirchhoff-Love theory. This reduced model has been shown to only be valid for special
problem classes concerning beam geometry and external loads, which are, however, present in many fields of application. The finite elements resulting from such a reduced model typically feature a simplified numerical implementation
and an increased computational efficiency. Although this torsion-free theory does not lie in the focus of this work, it
has to be emphasized1 that the proposed Kirchhoff-Love formulations lie the foundation for such reduced models.
These possible benefits will be further detailed in Section 8 and verified numerically in Section 11. After having
motivated the intention of the current work, its main scientific contributions shall briefly be highlighted:
• Two novel rotation interpolation strategies for geometrically exact Kirchhoff-Love beam element formulations
are proposed: i) The first represents an orthonormal interpolation scheme that fulfills the Kirchhoff constraint in
a strong manner. It can be regarded as generalization of the scheme proposed in [104] that allows for an exact
conservation of energy and momentum and for alternative sets of nodal degrees of freedom that simplify the
prescription of essential boundary conditions. ii) The space-continuous theory of the second variant is based
on a weak enforcement of the Kirchhoff constraint. The discrete realization of the Kirchhoff constraint relies
on a properly chosen collocation strategy. In combination with the employed smooth centerline interpolation,
also this strategy can completely avoid any additional Lagrange multipliers. For each of these two element
formulations, two different sets of nodal rotation parametrizations are proposed. One is based on nodal rotation
vectors and one on nodal tangent vectors. While these different choices are shown to yield identical FEM solutions, they differ in the resulting performance of nonlinear solvers and the effort required to prescribe essential
boundary conditions and joints. The four finite element formulations resulting from a combination of the two
interpolation schemes and the two choices of nodal primary variables are subject to detailed comparisons with
respect to resulting discretization error levels and the performance of nonlinear solution schemes.
• The resulting finite element formulations are combined with a finite difference time integration scheme for
large rotations recently proposed in the literature by Brüls and Cardona [31]. This implicit scheme allows for
energy-stable, second-order accurate time integration on the basis of optimized numerical dissipation and can
be identified as a Lie-group extension of the well-known generalized-α scheme. Up to the best of the author’s
knowledge, the current work represents the first application of a Lie group time integration scheme based on
optimized numerical dissipation to geometrically exact Kirchhoff-Love beam elements, and one of the first applications of such a scheme to geometrically nonlinear beam element formulations at all.
9
• Furthermore, the current work intends to review and evaluate existing geometrically exact beam element formulations and to compare these with the formulations newly proposed in this work. Concretely, the fulfillment
of the essential requirements according to Table 2 will be studied analytically for the three identified categories
of isotropic, straight and anisotropic formulations, which can be identified as the most general and elaborated
3D, large-deformation geometrically exact Kirchhoff-Love formulations previously proposed in the literature,
for one exemplary representative of Simo-Reissner formulations from the literature as well as for the KirchhoffLove formulations presented in the current and former works of the authors. A further original contribution
of this work is given by the detailed and systematic numerical comparison performed between the proposed
geometrically exact Kirchhoff-Love and two representative Simo-Reissner beam element formulations from the
literature. Specifically, resulting spatial convergence rates, discretization error levels as well as the performance
of nonlinear solution schemes are compared for different beam slenderness ratios.
Eventually, the organization of the remainder of this contribution shall briefly be given. One distinctive property of
geometrically exact beam formulations is the presence of large rotations within the associated configuration space. In
order to provide the theoretical basis for subsequent derivations, in the following Section 2, the SO(3) group of large
rotations as well as possible parametrizations will be introduced. In Section 3, the most general type of geometrically
exact beam formulations considered in this work, the Simo-Reissner theory of thick (shear-deformable) beams, will be
presented. Subsequently, in Section 4, the general theory will be restricted to the Kirchhoff-Love theory of thin (shearfree) beams. There, the different methodologies of imposing the Kirchhoff constraint of vanishing shear strains in a
strong or weak sense will be investigated. Afterwards, the space-time-continuous beam problem will be discretized.
In Section 5, a recently proposed extension of the generalized-α method from vector spaces to Lie groups [7, 31, 32],
will be applied to the beam formulations proposed in this work. Afterwards, in Section 6, spatial discretization
is performed, which represents the core topic in the development of geometrically exact Kirchhoff beam elements
considered in this work. In Sections 6.1 and 6.2, specific finite element interpolations employed to the translational
and rotational primary variable fields considered in this work will be proposed. In Section 6.3, the already stated
essential requirements on the employed spatial discretizations and the resulting finite element formulations will be
further detailed. Also detailed explanations and derivations for the evaluation results of Table 2 will be presented in
Sections 5 and 6. In the subsequent Sections 7-10, different finite element realizations resulting from the proposed
interpolation schemes and rotation parametrizations will be presented and the fulfillment of the basic requirements
stated above will be confirmed analytically. Concretely, Section 7 presents the basics of the geometrically exact SimoReissner type formulation proposed in [43, 80], which will serve as reference for several numerical comparisions
performed in Section 11. In the subsequent Section 8, the benefits of applying Kirchhoff-Love instead of SimoReissner element formulations in the range of high slenderness ratios will be further quantified. Afterwards, the two
major developments of this work, a Kirchhoff-Love element formulation based on strong constraint enforcement as
well as a Kirchhoff-Love element formulation based on weak constraint enforcement, will be proposed in Sections 9
and 10. For each formulation, two different variants of nodal rotation parametrization, one based on nodal rotation
vectors and one based nodal tangent vectors will be derived. Eventually, in Section 11, all the proposed concepts and
the resulting finite element formulations will be verified by means of proper numerical test cases. The reader who is
primarily interested in the practical implementation of the newly proposed geometrically exact Kirchhoff-Love beam
element formulations is referred to Sections 5, 9, 10 as well as to the appendices of this work.
2. The Rotation Group SO(3)
The category of beam theories considered in this contribution assumes the beam cross-sections to be rigid. Consequently, the cross-section kinematics are point-wise uniquely defined by six degrees of freedom, three translational
ones representing the position vector of the cross-section centroid and three rotational ones describing the crosssection orientation. Thereto, an orthonormal triad consisting of the base vectors g1 , g2 , g3 ∈ <3 is attached on the
beam cross-sections. Furthermore, a right-handed inertial Cartesian frame E1 , E2 , E3 ∈ <3 associated with the material configuration and a corresponding right-handed inertial Cartesian frame e1 , e2 , e3 ∈ <3 of the spatial configuration
are introduced. Nevertheless, for simplicity, it is assumed that both frames coincide, thus ei = Ei for i = 1, 2, 3. Now,
10
the rotation from the global frame Ei onto the local frame gi is described via the orthogonal transformation Λ ∈ SO(3):
gi = ΛEi
with
Λ = g j ⊗ E j = (g1 , g2 , g3 )E j
for
i, j = 1, 2, 3.
(1)
Here and in the following, the summation convention over repeated indices holds. Throughout this work, the index
near a matrix (for example the index (.)Ei in equation (1)) denotes the basis in which the associated tensor is represented. In the context of geometrically exact beam theories, the two-point tensor Λ acts as push-forward operator (see
e.g. [103]) between material and spatial objects. From a rather mathematical point of view, the rotation tensor Λ can
be identified as an element of the Special Orthogonal group SO(3) of orthogonal transformations according to
SO(3) := {Λ ∈ <3×3 |ΛTΛ = I3 , det(Λ) = 1},
(2)
under the action of non-commutative multiplication SO(3)×SO(3) → SO(3), Λ3 = Λ1 Λ2 , Λ2 Λ1 with inverse element
Λ−1 =ΛT and identity element I3 . Here, det(.), (.)T and (.)−1 are the determinant, transpose and inverse. I3 is a 3×3
identity matrix. The nonlinear manifold SO(3) can be classified as Lie group with tangent space
δΛ ∈ T Λ SO(3) := {S(a)Λ|S(a) ∈ so(3)}.
(3)
Here, so(3) denotes the set of skew symmetric tensors with S(a)b = a× b for a, b ∈ <3 . The isomorphism between
so(3) and <3 enables a unique expression of S(a) ∈ so(3) by the vector a ∈ <3 denoted as axial vector. By inserting
the special choice Λ = I3 into (3) it can easily be verified that so(3) represents the tangent space to SO(3) at the identity:
so(3) := T I SO(3) = {S(a)|S(a) = −S(a)T ∀ a ∈ <3 }.
(4)
The Lie group SO(3) and its Lie algebra so(3) are related by the exponential map exp (.): so(3) → SO(3) according to:
exp(S(a)) := I3 +S(a)+
S(a)2 S(a)3
+
+... .
2!
3!
(5)
Up to now, the rotation group SO(3) has been introduced without stating any specific parametrization of the rotation
tensor Λ. In the following sections, two possible parametrizations, which will be useful in the development of beam
element formulations according to the Simo-Reissner and Kirchhoff-Love theory, will be presented.
2.1. SO(3) parametrization via rotation vectors
There exist various parametrizations of the rotation tensor such as rotation (pseudo-) vectors, Euler angles or Rodrigues parameters that are based on a minimal set of three parameters. Also four-parametric representations of the
rotation tensor such as quaternions have proven to be very useful for practical purposes. Within this work, two different parameterizations will be employed: The one presented in this section is based on rotation vectors ψ ∈ <3 .
In Section 2.2, an alternative parametrization especially suited for Kirchhoff formulations is presented. The rotation
vector parametrization can explicitly be given via the well-known Rodrigues formula:
h
i
Λ(ψ) = exp(S(ψ)) = I + sin ψS(eψ ) + (1 − cos ψ)S(eψ )S(eψ ) .
(6)
Here, ψ = ||ψ|| represents the scalar rotation angle and eψ = ψ/||ψ|| the axis of rotation. Throughout this work, ||(.)||
denotes the Euclidean norm in <3 . As indicated by the notation exp(S(ψ)), equation (6) represents a closed-form representation of the exponential map initially introduced in (5) of the last section. The rotation vector of a given rotation
tensor can for example be extracted by employing Spurrier’s algorithm [140]. The infinitesimal spatial quantity δθ,
denoted as multiplicative rotation vector variation or spin vector, allows to express the variation of Λ:
δΛ :=
d
d
=0
exp (S(δθ))Λ = S(δθ)Λ
or δgi = δθ×gi .
(7)
Alternatively, the variation δΛ can be expressed by the material spin vector δΘ = ΛT δθ via right translation:
δΛ :=
d
d
=0
Λ exp (δΘ) = ΛS(δΘ).
11
(8)
Expressing the spin vectors δΘ and δθ in the associated frames Ei and gi , respectively, yields:
δΘ = δΘi Ei ,
δθ = δΘi gi .
(9)
Thus, the components of the spatial spin vector expressed in the local basis gi are identical to the components of
the material spin vector expressed in the global basis Ei . This relation also holds for all the other pairs of spatial /
material quantities considered in this work (see e.g. [132]). Based on the defining equation (8) and the representation
of skew-symmetric tensors by means of the Levi-Civita-Symbol i jk , the components of S(δΘ) can be determined:
S(δΘ) = ΛTδΛ = (Ei ⊗gi )·(δg j ⊗E j ) = gTiδg j Ei ⊗E j with S i j (δΘ) = −i jk δΘk .
|{z}
(10)
=:S i j (δΘ)
For later use, the components δΘi shall be expressed by gi and δgi . From the previous equation (10), it follows:
δΘ1 = gT3δg2 = −gT2δg3 ,
δΘ2 = gT1δg3 = −gT3δg1 ,
δΘ3 = gT2δg1 = −gT1δg2 .
(11)
A relation between infinitesimal additive and multiplicative increments is given by the tangent operator T defined as
!
1
1
ψ/2
1
I − 2 S(ψ)S(ψ)T − S(ψ).
(12)
δψ =: Tδθ, T = 2 S(ψ)S(ψ)T +
tan (ψ/2)
2
ψ
ψ
The inversion of the previous equation (12), expressing multiplicative by means of additive increments, is given by:
!
1
sin (ψ)
sin (ψ) 1−cos (ψ)
δθ = T−1 δψ, T−1 = 2 1−
S(ψ)S(ψ)T +
I+
S(ψ).
(13)
ψ
ψ
ψ
ψ2
For details on the derivation of the transformations T and T−1 , the interested reader is e.g. referred to [135, 34, 74,
40]. While the rotation vector parametrization presented so far represents a well-known tool in the formulation of
geometrically exact beam elements of Simo-Reissner type, in the following section, an alternative parametrization of
large rotations will be proposed which offers some advantages in the description of Kirchhoff type beam elements.
2.2. SO(3) parametrization via ”smallest rotation” triads
The alternative parametrization considered in this section consists of four degrees of freedom (t, ϕ) with t ∈ <3 and
ϕ ∈ <. In the context of Kirchhoff beam elements presented later in this work, t will be the non-unit tangent vector
aligned to the space curve representing the beam centerline. Due to the Kirchhoff constraint of vanishing shear strains,
the first base vector g1 of the cross-section triad Λ = (g1 , g2 , g3 )Ei , can be expressed by this tangent vector:
g1 =
t
.
||t||
(14)
Based on the tangent vector t and an arbitrary given triad ḡi , one can determine a triad g Mi , in the following denoted
as interMediate or Medium triad (index M), that results when the triad ḡi is rotated onto the tangent vector t via the
“Smallest Rotation” (SR) (see e.g. [40, 104]. The resulting base vectors can be represented by the expressions
g M1 = g1 =
ḡT g1
ḡT g1
t
, g M2 = ḡ2 − 2 T (g1 + ḡ1 ) , g M3 = ḡ3 − 3 T (g1 + ḡ1 )
||t||
1+ ḡ1 g1
1+ ḡ1 g1
→
Λ M =: sr(Λ̄, g1 ).
(15)
In order to shorten notation, the abbreviation sr(.) has been introduced for the SR mapping of (15). Practical choices
for the triad Λ̄ will be presented in Section 6.2 (see also [105] for a discussion of possible singularities of the SR
mapping). Subsequently to the definition of an intermediate triad Λ M according to (15), the cross-section triad gi can
be defined based on a relative rotation of the intermediate triad Λ M with respect to the tangent t by an angle of ϕ:
Λ = exp(S(ϕg1 ))Λ M
⇔
g1 = g M1 , g2 =g M2 cos ϕ + g M3 sin ϕ, g3 = g M3 cos ϕ − g M2 sin ϕ.
(16)
Equations (15) and (16) uniquely define a triad Λ parametrized by the four degrees of freedom (t, ϕ). Evidently, one
of these four degrees of freedom, namely the norm ||t|| of the tangent vector, will not influence the triad orientation.
12
However, as it will turn out in the next sections, the non-unit tangent vector t is a quantity that directly results from
the beam centerline description. Thus, the only additional degree of freedom introduced in order to describe the triad
orientation is the relative angle ϕ and consequently, the proposed type of triad parametrization is not redundant. For
later use, the spatial spin vector δθ shall be expressed by means of additive increments (δt, δϕ) of the four parameters
(t, ϕ). Therefore, it is split into a component δθk parallel to g1 and a component δθ⊥ perpendicular to g1 as follows:
δθ = δθk + δθ⊥ = δΘ1 g1 + δθ⊥ .
(17)
Throughout this work, the indices (.)k and (.)⊥ of a vector will denote the components of the vector which are parallel
or perpendicular to the vector g1 , respectively. Taking advantage of (7), δθ⊥ can be easily derived (see also [105]):
!
t
t × δt g1 × δt
t
×δ
=
=
with δt ∈ <3 .
(18)
δθ⊥ = g1 × δg1 =
||t||
||t||
||t||
||t||2
In a next step, (11) and (16) are exploited in order to formulate the tangential component δΘ1 of the spin vector:
δΘ1 = gT3δg2 = gTM3 δg M2 + δϕ =: δΘ M1 + δϕ
(19)
By variation of the basis vector δg M2 defined in (15), the tangential component δΘ M1 can be determined to:
δΘ M1 =
ḡT1 S(g1 ) δt
(g1 × ḡ1 )T
δg
=
−
=: TΘM1t δt
1
1+gT1 ḡ1
1+gT1 ḡ1 ||t||
→
δΘ1 = TΘM1t δt+δϕ.
(20)
Inserting equations (18)-(20) into the split relation (17) yields the following expression for the spatial spin vector:
g1 ⊗ ḡT1
1
t
δθ = g1 δϕ + I −
(21)
S(g1 ) δt with g1 = , t := ||t||.
T
t
t
1 + g1 ḡ1
|
{z
}
=:Tθt
So far, the four degrees of freedom (t, ϕ) have been applied in order to uniquely describe a tangent vector t, defined by
its orientation and its length, as well as the orientation of a triad Λ = (g1 , g2 , g3 ) aligned parallel to this tangent vector.
Next, the non-unit tangent vector t and the triad Λ shall be described by the alternative set of four degrees of freedom
(ψ, t). Here, ψ represents the rotation vector associated with the triad Λ via the Rodrigues formula (6) and t := ||t|| is
the norm of the tangent vector. The following transformations hold between the two sets (t, ϕ) and (ψ, t):
Λ = exp(S(ψ)), g1 = ΛE1 , t = tg1 , exp(S(ϕg1 )) = Λsr(Λ̄, g1 )−1 .
(22)
Based on (22), the set (t, ϕ) can be calculated from (ψ, t) and the other way round. Next, also a transformation rule
between the variations (δt, δϕ) and the variations (δθ, δt) associated with the set (ψ, t) shall be derived:
δt = δ||t|| =
tT
δt = gT1 δt.
t
Combining equations (21) and (23) yields a transformation rule between the two sets (δθ, δt) and (δt, δϕ):
!
!
!
!
!
!
(g1 + ḡ1 )T
δθ
Tθt g1
δt
δt
−tS(g1 ) g1
δθ
⇔
=
with Tϕθ =
.
=
T
δt
g
0
δϕ
δϕ
Tϕθ
0
δt
1 + gT1 ḡ1
| 1{z }
|
{z
}
(23)
(24)
=:T M
=:T−1
M
The inverse transformation in (24) has been derived in a similar manner. The mappings T M and T−1
M transform between
multiplicative rotation increments and additive increments of the chosen parametrization. Thus, they represent the
analogon to the transformations T and T−1 in case of a rotation vector parametrization (see Section 2.1). Since the
Kirchhoff constraint of vanishing shear deformation solely influences the component δθ⊥ of the spin vector, it will in
13
the following sections often be useful to express only this component by additive increments δt, while the tangential
spin vector component δΘ1 instead of the additive increment δϕ is regarded as independent primary variable:
!
!
!
!
!
!
1
δθ
δt
δt
−tS(g1 ) g1
δθ
S(g ) g1
= t T1
,
=
,
(25)
δt
δΘ1
δΘ1
gT1
0
δt
g1
0
|
{z
}
|
{z
}
=:T̃
=:T̃−1
The transformation (25) basically represents a reformulation of (17) and (18). Again, T̃ and T̃−1 represent the corresponding mappings. Since these mappings solely transform the component δθ⊥ , they are independent from the actual
definition of the triad Λ M . Consequently, the index M has been omitted for the transformation matrices T̃ and T̃−1 .
3. Simo-Reissner Beam Theory
In this section, the fundamentals of the geometrically exact Simo-Reissner beam theory based on the work of Reissner [114, 115] as well as Simo and Vu-Quoc [132, 134] will be briefly summarized. The results of this section will
provide an essential basis for the subsequent derivation of Kirchhoff type beam formulations.
Figure 2: Kinematic quantities defining the initial and deformed configuration of the geometrically exact beam.
3.1. Basic kinematic assumptions
Throughout this work, prismatic beams with anisotropic cross-section shape are considered. In the initial (unstressed)
configuration, the beam centerline, which is defined as the line connecting the cross-section centroids, is described by
the space curve s → r0 (s) ∈ <3 . Here and in the following, the index 0 of a quantity refers to the unstressed, initial
configuration. Furthermore, s ∈ [0, l] =: Ωl ⊂ < is an arc-length parametrization of the curve and l ∈ < the beam
length in the initial configuration. The description of the initial configuration is completed by a field of right-handed
orthonormal triads s → g01 (s), g02 (s), g03 (s) ∈ <3 , also denoted as material triads in the following. These are attached
to the beam cross-sections, which are assumed to be undeformable according to the Bernoulli hypothesis. In this
context, g01 (s) = r00 (s) represents the unit tangential vector to the initial centerline and the base vectors g02 (s) and
d
g03 (s) coincide with the principal axes of inertia of the cross-section at s. Throughout this work, the prime (.)0 = ds
(.)
denotes the derivative with respect to the arc-length parameter s. The rotation from the global frame Ei onto the initial
local frame g0i (s) is described via the orthogonal transformation s → Λ0 (s) ∈ SO(3) as introduced in (1) leading
to the following definition s → C0 := (r0 (s), Λ0 (ψ0 (s))) ∈ <3 × SO(3) of the initial configuration. Correspondingly,
the deformed configuration of the beam at time t ∈ < is given by s, t → C := (r(s, t), Λ(ψ(s, t))) ∈ <3 × SO(3),
where the orthogonal transformation Λ(ψ(s, t)) maps from the global frame Ei onto the current local frame gi (s)
and the base vector g1 (s, t) := g2 (s, t) × g3 (s, t) is in general not tangential to the deformed centerline anymore due
to shear deformation. According to Section 2, Λ(s, t) can be represented by three rotation parameters (e.g. by a
rotation vector ψ(s, t)), leading to pointwise six, three translational and three rotational, degrees of freedom. The basic
kinematic assumption of the geometrically exact Simo-Reissner theory considered so far can easily be summarized
by the following constrained position vector field associated with the initial and deformed configuration
X(s) = r0 (s) + s2 g02 (s) + s3 g03 (s),
x(s, t) = r(s, t) + s2 g2 (s, t) + s3 g3 (s, t),
14
s := (s, s2 , s3 )T =: (s1 , s2 , s3 )T , (26)
where (s2 , s3 ) represent convective coordinates describing the position of an arbitrary material point within the (rigid)
cross-section. In order to simplify the notation for subsequent derivations, the convective coordinate vector given by
s = (s1 , s2 , s3 )T as well as the redundant name s1 := s for the arc-length s has been introduced. The kinematic quantities
defining the initial and deformed configuration as introduced so far are illustrated in Figure 2. In order to simplify
notation, the time argument t will be dropped in the following. Later in this section, the arc-length derivative of the
base vectors gi (s) will be required. Similar to equation (7), this derivative can be formulated as
Λ0 (s) = S(k(s))Λ(s)
or
Λ0 (s) = Λ(s)S(K(s)),
(27)
where k(s) and K(s) := ΛT (s)k(s) are referred to as spatial and material curvature vector, which can be derived as:
S(k(s)) = Λ0 (s)ΛT (s),
S(K(s)) = ΛT (s)Λ0 (s)
K = Ki Ei , k = Ki gi .
(28)
In a similar manner, the spatial and material angular velocity vectors w and W are defined according to:
S(w(s)) = Λ̇(s)ΛT (s),
S(W(s)) = ΛT (s)Λ̇(s)
W = Wi Ei , w = Wi gi .
(29)
For completeness, also the spin vectors, which have been introduced in Section 2, are repeated here:
S(δθ(s)) = δΛ(s)ΛT (s),
S(δΘ(s)) = ΛT (s)δΛ(s)
δΘ = δΘi Ei , δθ = δΘi gi .
(30)
˙ = d (.) denotes the derivative with respect to time t. By applying the Young theorem
Throughout this work, the dot (.)
dt
0
0
(Λ˙ ) = (Λ̇) and making use of S(a)S(b)−S(b)S(a) = S(S(a)b), the following compatibility relations can be shown:
k̇ = w0 + w × k,
δk = δθ0 + δθ × k,
K̇ = W − W × K,
δK = δΘ − δΘ × K,
0
0
δw = δθ̇ + δθ × w,
δW = δΘ̇ − δΘ × W.
(31)
In the following sections, stress resultants mechanical equilibrium and proper constitutive relations will be presented.
3.2. Stress resultants, mechanical equilibrium and objective deformation measures
With f̃ and m̃ denoting distributed external forces and moments per unit length and fρ and mρ representing the force
and moment contributions due to inertia effects, the strong form of equilibrium reads (see e.g. [3, 115, 132]):
f 0 + f̃ + fρ = 0,
m0 + r0 × f + m̃ + mρ = 0.
(32)
In (32), f and m are the force and moment stress resultants acting on the beam cross-section area A. A material form
of the 1D equilibrium equations can be derived by inserting the material stress resultants F:=ΛTf and M:=ΛTm into
the balance equations (32). Following the principle of virtual work, the admissible variations, i.e. infinitesimal small
and arbitrary (additive or multiplicative) changes of the current configuration s → δC := (δr(s), δθ(s)) ∈ <3 ×<3 that
are compatible with the employed boundary conditions are introduced in a next step. Here, δr(s) ∈ <3 represents the
vector of (additive) virtual displacements and δθ(s) ∈ <3 the vector of (multiplicative) virtual rotations, also denoted
as spin vector. By multiplication of (32) with δr and δθ and integration by parts, the spatial weak form is derived:
G=
Zl
0
h
i
h
i
0
0 T
T
T
T
T
0T
=˙ 0.
δθ
m
+
(δr
−
δθ
×
r
)
f
−δθ
(
m̃
+
m
)
−
δr
(
f̃
+
f
)
ds
−
δr
f
−
δθ
m
ρ
ρ
σ
σ
| {z } |
Γσ
Γσ
{z
}
δo ωT m
δo
(33)
γT f
Here, fσ and mσ denote external forces and moments at the Neumann boundary Γσ of the considered beam. Based on
the principle of virtual work, the following objective spatial deformation measures can be identified:
γ = r0 −g1 , S(ω) = Λ0 ΛT−ΛΛT0Λ00 ΛT = S(k)−S(ΛΛT0 k0 ) → ω = k−ΛΛT0 k0 .
15
(34)
If the material counterpart of equation (32) is chosen as starting point (see e.g [43]), the material deformation measures
Γ and Ω being work-conjugated to the material stress resultants F and M can be determined in an analogous manner:
Γ = ΛTr0 −E1 , S(Ω) = ΛTΛ0 −ΛT0Λ00 = S(K)−S(K0 ) → Ω = K− K0 .
(35)
The “objective variation” δo of an arbitrary vector a ∈ <3 appearing in (33) is defined as δo a := δa−δθ ×a (see e.g.
[132]). It is easy to verify that all of these deformation measures vanish for the stress-free initial configuration, i.e.
when the relations Λ = Λ0 , r0 = r00 and g1 = g01 = r00 = Λ0 E1 are valid, and that the following variations can be derived:
δo γ = δr0 − δθ × r0 ,
δΓ = ΛT δr0 + r0 × δθ ,
δo ω = δθ0 ,
δΩ = ΛT δθ0 .
(36)
The components of Γ represent axial tension and shear, the components of Ω represent torsion as well as bending.
3.3. Constitutive relations
Finally, constitutive relations between the stress resultants M and F and the deformation measures Ω and Γ are
required. The simplest constitutive law of this type is given by the length-specific hyperelastic stored energy function:
1
1
Π̃int (Ω, Γ) = ΩTC M Ω+ ΓT CF Γ,
2
2
M=
∂Π̃int
= C M ·Ω,
∂Ω
F=
∂Π̃int
= CF ·Γ.
∂Γ
Here, the material constitutive tensors C M and CF have the following diagonal structure:
C M = diag GIT , EI2 , EI3 Ei and CF = diag EA, G Ā2 , G Ā3 Ei .
(37)
(38)
Here, E and G are the Young’s modulus and the shear modulus, A, Ā2 and Ā3 are the cross-section and the two reduced
cross-sections, I2 and I3 are the two principal moments of inertia and IT is the torsional moment of inertia. In a similar
manner to (37), the length-specific kinetic energy Π̃kin of the beam can be formulated:
1
1
Π̃kin (w, ṙ) = wTcρ w+ ρAṙTṙ,
2
2
Cρ = diag ρ(I2 +I3 ), ρI2 , ρI3 Ei .
|{z}
cρ = ΛCρ ΛT ,
(39)
=:IP
Here, ρ is the mass density, Cρ the material inertia tensor and w the spatial angular velocity vector, which has already
been introduced in (29). From (39), the length-specific linear momentum l̃ and angular momentum h̃ can be derived:
l̃ :=
∂Π̃kin
= ρAṙ,
∂ṙ
h̃ :=
∂Π̃kin
= cρ w = ΛCρ ΛTw = ΛCρ W.
∂w
(40)
Similar to the length-specific external forces f̃ and m̃, also the length-specific energies Π̃int , Π̃kin and momenta l̃, h̃ have
˜
been furnished with the (.)-symbol.
The total counterparts fext , mext , Πint , Πkin , l and h are obtained by integration:
Zl
h i
fext := f̃ds+ fσ ,
Γσ
Zl
h
i
mext := r × f̃ + m̃ ds+ r × fσ +mσ ,
Γσ
0
Zl
Πint := Π̃int ds,
0
0
Zl
Πkin := Π̃kin ds,
0
Zl
l := l̃ds,
0
Zl
h := (h̃+r× l̃)ds.
(41)
0
These definitions will be required subsequently. Based on (40), the inertia forces fρ and moments mρ yield:
h
i
−fρ = ˙l̃ = ρAr̈, −mρ = h̃˙ = Λ S(W)Cρ W+Cρ A = S(w)cρ w+cρ a, a := ẇ, A := Ẇ.
(42)
Similar to w = ΛW, also the angular accelerations are related via the push-forward operator:
a := ẇ =
d
(ΛW) = ΛS(W)W + ΛẆ = ΛẆ =: ΛA, since S(W)W = 0.
dt
16
(43)
When the beam problem has to be discretized in time (see Section 5 for details), the vectors w and a can either be
directly employed in a time integration scheme [135, 136, 80] or, alternatively, they can be expressed via the (additive)
rate of the primary variable ψ [34]. Similar to (12), the following relations between w and a can be formulated:
w = T−1 ψ̇, a = Ṫ−1 ψ̇ + T−1 ψ̈.
(44)
Finally, the problem setup presented in Sections 3.1-3.3 has to be completed by boundary and initial conditions in
order to end up with a well-defined initial boundary value problem:
r = ru , ψ = ψu on Γu ,
f = fσ , m = mσ on Γσ ,
Γσ ∩ Γu = ∅, Γσ ∪ Γu = {0, L}
(45)
r = r0 , ṙ = v0 , Λ = Λ0 , w = w0 at t = 0.
Based on a trial space (r(s, t), Λ(ψ(s, t))) ∈ U of functions with square-integrable first derivatives satisfying (45) and an
weighting space (δr(s), δθ(s)) ∈ V of functions with square-integrable first derivatives satisfying δr = 0, δθ = 0 on Γu ,
the weak form (33) is equivalent to the strong form (32) supplemented by the boundary conditions (45). The following two Sections 3.4 and 3.5 are intended to supplement the geometrically exact beam theory sdsd so far. However,
the provided information is not necessarily required for the derivation of the space-continuous problem statement of
the Simo-Reissner and Kirchhoff-Love beam theory and the subsequent discretization procedures. Thus, the reader
may alternatively skip these content and proceed with Section 4. Concretely, in Section 3.4, the alternative procedure
of deriving the weak and strong form of the balance equations presented in Section 3.2 on the basis of a variational
problem statement and properly defined Lagrangian is undertaken. Section 3.5 is meant to confirm the constitutive
relations presented above based on a derivation from 3D continuum mechanics. There, the consistency of the 1D and
3D continuum model in the sense of a fully induced beam theory will be verified for the case of locally small strains.
Remark: Two possible time integration schemes can be derived from the variants of either employing W
and A directly or expressing them via additive rates given by (44):
1)
(ṙ, r̈, w, a)n+1 = f((r, Λ)n+1 , (r, Λ)n , (ṙ, w)n , (r̈, a)n ),
2)
(ṙ, r̈, ψ̇, ψ̈)n+1 = f((r, ψ)n+1 , (r, ψ)n , (ṙ, ψ̇)n , (r̈, ψ̈)n ).
(46)
Here, the indices (.)n and (.)n+1 refer to two successive time steps of the time-discrete problem and f(.)
represents a typical finite difference time integration scheme (e.g. a Newmark scheme). The first variant
(see Section 5) can be considered as being more flexible since it does not require any specific rotation
parametrization. It can directly be applied to Reissner type beam formulations as well as to Kirchhoff
type beam formulations with strong or weak Kirchhoff constraint enforcement without the need for further
adaptions. Due to this flexibility and the very simple and compact time integrator resulting from this
procedure, this will be the method of choice employed throughout this work.
3.4. Interlude: Variational problem formulation of Simo-Reissner beam theory
For the case that no external forces act on the beam, i.e. fσ = f̃ = mσ = m̃ = 0, the strong and weak forms (32) and (33)
shall equivalently be formulated on the basis of the well-known Hamilton principle:
δ
ZT
L dt = 0
with
[δr = δθ = 0]t=T
t=0 .
(47)
t=0
The Lagrangian L occurring in (47) is defined as the difference of kinetic and potential energy:
L=
Zl
(Π̃kin − Π̃int ) ds.
s=0
17
(48)
Based on the kinetic energy (39) and the hyper-elastic energy (37), the variation (47) reads:
δ
ZT
L dt =
ZT Zl
δo wTcρ w + δo ṙTρAṙ − δo ωTcm ω − δo γTc f γ ds dt
t=0 s=0
t=0
=
ZT Zl
δθ̇T h̃ + δṙT l̃ − δθ0Tm − δr0 −δθ×r0 Tf ds dt
(49)
t=0 s=0
ZT Zl
=−
−δθTmρ − δrTfρ + δθ0Tm + δr0 −δθ×r0 Tf ds dt = 0.
t=0 s=0
From the first to second line, the objective variations δo w = δθ̇, δo ω = δθ0 and δo γ = [δr0 −δθ×r0 ] (see (31) and (36))
have been inserted. Furthermore, the additional relation δo ṙT ṙ = (δṙ − δθ × ṙ)T ṙ = δṙT ṙ has been applied. From the
second to the third line, partial integration of the inertia terms together with the boundary conditions in (47) have been
applied. Since no external forces and moments are considered, i.e. fσ = mσ = 0, partial integration of (49) yields:
ZT
ZT Zl
δθT m0 + r0 × f + mρ + δrT f 0 + fρ ds dt = 0.
δ L dt =
|
{z
}
|{z}
t=0
t=0 s=0
=0
˙
(50)
=0
˙
The arbitrariness of δθ(s, t) and δr(s, t) directly yields the strong form (32). Inverting the last step from (49) to (50),
by partial integration of the weighted strong form (32), this time only along the beam length l, yields the weak form
of the balance equations (33). The terms occurring in this weak form for an unloaded beam can already be identified
in the third line of (49). While conservative external forces could also be included into the Lagrangian (48), for a
consideration of non-conservative external forces and 3D external moments, which are known to be non-conservative
(see e.g. [134]), the starting point has to be the strong form of the mechanical balance equations (32).
3.5. Interlude: Relation between 1D and 3D constitutive laws
The aim of the following considerations is to derive the constitutive laws (37) and (38) in a consistent manner from
the 3D continuum theory. Thereto, the deformation gradient F of the 3D position field subject to the kinematic
constraints (26) of the geometrically exact beam theory shall be derived. Subsequently, also the Cauchy-Green deformation tensor E is required. These two objects can be formulated based on the following definitions:
E :=
1 T
F F − I3
2
with F :=
∂x
∂x ∂s
=
= gi ⊗ Gi .
∂X
∂s ∂X
(51)
The (non-orthonormal) covariant basis vectors gi = ∂x(s)/∂si and Gi = ∂X(s)/∂si have to be determined from (26):
g1 = r0 + k × (s2 g2 + s3 g3 ) , g2 = g2 , g3 = g3 ,
G1 = g01 + k0 × (s2 g02 + s3 g03 ) , G2 = g02 , G3 = g03 .
(52)
Also the contravariant base vectors Gi can be determined from the second line of (52) via the definition GiT G j = δi j :
G1 =
1
K01 s3
K01 s2
g01 , G2 = g02 +
g01 , G3 = g03 −
g01 , C := 1−K03 s2 +K02 s3 .
C
C
C
(53)
Making use of the relation r0 = g1 +γ (34) and inserting the first line of (52) as well as (53) into (51) yields:
F=
1
K01 s3
K01 s2
g1 +γ+k×(s2 g2 + s3 g3 ) ⊗g01 +g2 ⊗ g02 +
g01 +g3 ⊗ g03 −
g01 .
C
C
C
(54)
Finally, by inserting (54) into (51), the individual components of the Cauchy-Green deformation tensor can be determined. However, in order to gain further insight into the underlying structure of the deformation gradient, the very
18
ellegant procedure suggested by Geradin and Cardona [58] (for initially straight beams) as well as Linn et al. [100]
(for initially curved beams) is employed by slightly reformulating the expression (54) on the basis of the relative
rotation tensor Λ̃ := ΛΛT0 . By applying the auxiliary relation 1/C = 1+(1−C)/C to the pre-factor of g1 and solving all
the products in (54), the deformation gradient can eventually be reformulated according to:
F = Λ̃ (I3 + H⊗g01 )
with
H := Hi g0i , I3 = g0i ⊗g0i , Λ̃ = gi ⊗g0i .
(55)
The components of the vector H identified in (55), denoted as material strain vector in [58] and [100], read:
1
[Γ1 +(K2 −K02 )s3 −(K3 −K03 )s2 ] ,
C
1
1
H2 = [Γ2 −(K1 −K01 )s3 ] , H3 = [Γ3 −(K1 −K01 )s2 ] .
C
C
H1 =
(56)
Based on the deformation gradient (55), the Cauchy-Green deformation tensor can be derived according to:
E=
H TH
1
H⊗ g01 + g01 ⊗ H +
g01 ⊗ g01 .
2C
2C 2
(57)
The result (57) has been consistently derived from the basic kinematic assumptions (26) without any additional approximations. However, in order to finally end up with the simple constitutive laws of the geometrically exact beam
theory based on the quadratic form (37), the following well-known assumption of small local strains is made:
Γi 1
and R · Ki 1, R · K0i 1 for i = 1, 2, 3.
(58)
The assumptions (58) state that small local axial and shear strains are considered and that the radii of initial and
deformed centerline curvature have to be small as compared to the cross-section radius R. In the following, a firstorder approximation in these small quantities is considered by setting C ≈ 1 and neglecting the last, quadratic term
in (57). Based on these small-strain assumptions, the approximated Cauchy-Green deformation tensor reads
E ≈ Ē =
1
H⊗ g01 + g01 ⊗ H =: Ēi j g0i ⊗ g0 j .
2
(59)
The components Ēi j of the approximated Cauchy-Green deformation tensor Ē resulting from this procedure yield:
Ē11 = Γ1 + (K2 − K02 )s3 − (K3 − K03 )s2 ,
1
Ē12 = Ē21 = (Γ2 − (K1 − K01 )s3 ) ,
2
1
Ē13 = Ē31 = (Γ3 + (K1 − K01 )s2 ) ,
2
Ē22 = Ē33 = Ē23 = Ē32 = 0.
(60)
By applying a Saint-Venant-Kirchhoff material with constitutive tensor C = Ci jkl g0i ⊗ g0 j ⊗ g0k ⊗ g0l , the 2. PiolaKirchhoff stress tensor S̄ = S̄ i j g0i ⊗ g0 j can be formulated based on the approximated Cauchy-Green tensor Ē:
S̄ 11 = Ē(Γ1 + (K2 − K02 )s3 − (K3 − K03 )S 2 ),
S̄ 22 = S̄ 33 = −νE11 ,
S̄ 12 = S̄ 21 = G (Γ2 − (K1 − K01 )s3 ) ,
S̄
13
S̄
23
= S̄
31
= G (Γ3 + (K1 − K01 )s2 ) ,
= S̄
32
= 0.
(61)
Here, the scaled Young’s modulus Ē and the shear modulus G have been introduced in (61) according to
Ē :=
(1 − ν)E
,
(1 + ν)(1 − 2ν)
19
G :=
E
,
2(1 + ν)
(62)
where E is Young’s modulus and ν is Poisson’s ratio. From (61), it gets obvious that the standard relations known in the
geometrically exact beam theory in terms of vanishing in-plane stress components S̄ 22 = S̄ 33 = 0 and of a constitutive
parameter E in front of the normal stress S̄ 11 , only holds for the special case ν = 0. This is a consequence of the
kinematic assumption of rigid cross-sections, which requires the existence of in-plane reaction forces in general.
In order to resolve these two putative contradictions for general cases ν , 0, the constraint of rigid cross-sections
can be weakened by allowing for a uniform lateral contraction of the cross-section with in-plane strain components
Ē22 = Ē33 = −νE11 and for a proper in-plane warping field (see [100] or [150] for further details). Alternatively, the
approximation ν = 0 can be employed in the first two lines of (61). In praxis, this slight inconsistency is often taken
into account, which is not unusual in the field of structural theories (see e.g. [3] or [87]). In a last step, the material
force and moment stress resultants are determined by integration of the stress vector t1 = S · g01 acting on a beam
cross-section with material normal vector g01 according to the definition proposed in the original work [132]:
Z
Z
F̃ := t1 dA =
E(Γ1 +(K2 −K02 )s3 −(K3 −K03 )s2 )g01
A
A
+ G (Γ2 −(K1 −K01 )s3 ) g02 +G(Γ3 +(K1 −K01 )s2 )g03 dA
(63)
EA 0
0 Γ1
· Γ2 = C̃F Γ̃.
= 0 GA
0
0 GA g0i Γ3 g0i
Z
Z
M̃ := xA × t1 dA = (s2 g02 + s3 g03 )× E(Γ1 +(K2 −K02 )s3 −(K3 −K03 )s2 )g01
A
A
+G (Γ2 −(K1 −K01 )s3 ) g02 + G (Γ3 +(K1 −K01 )s2 ) g03 dA
(64)
GIP
0
0 K1 − K01
EI2
= 0
· K2 − K02 = C̃ M Ω̃.
0
0 EI3 g
K3 − K03 g
0i
0i
R
R
R
R
Here, the defintions of the moments of inertia of area I2 := A s23 dA and I3 := A s22 dA as well as A s3 dA = A s2 dA =
R
s s dA = 0 have been applied. As expected, (63) and (64) yield a constitutive law that is identical to the one posA 2 3
tulated in (37). Comparable derivations based on similar small-strain assumptions can e.g. be found in the original
works of Kirchhoff [86] and Love [101] in the context of shear-free beam formulations as well as in the current contributions [58, 84, 100] in the context of geometrically exact Simo-Reissner type formulations. It has to be mentioned
that the presented derivation yields alternative material objects F̃, C̃F and Γ̃ as well as M̃, C̃ M and Ω̃ that are pulledback to the curved, initial reference configuration and not to the straight reference configuration as it was the case for
the material objects considered so far (see also the remark in [104], page 452). However, the components of the alternative material objects, e.g. F̃ = Fi g0i , when expressed with respect to the ”curved”, local basis g0i are identical to the
components of the original material objects, e.g. F = Fi Ei , when expressed with respect to the ”straight”, global basis
Ei . Via push-forward, also the 1. Piola-Kirchhoff stress tensor P and the Cauchy stress tensor σ can be determined.
P = FS,
σ=
1
FSFT .
detF
(65)
Starting with the 1. Piola-Kirchhoff stress tensor or with the Cauchy stress tensor, the spatial stress resultants, e.g.
f = Fi gi , can be derived similarly to (63) and (64). Since again only first-order terms of the small strains (58) are
relevant, it is sufficient to approximate the deformation gradient, required for the push-forward, according to:
F ≈ Λ̃ = gi ⊗ g0i ,
detF ≈ 1.
(66)
Consequently, as already postulated in the sections before, the relevant pull-back / push-forward operator is given by
a rotation tensor Λ̃. If for the derivations above a 3D continuum formulation with material strain and stress measures
based on a straight reference configuration had been applied, the resulting 1D material objects would be based on the
global basis Ei and the total rotation tensor Λ could be identified as the relevant pull-back / push-forward operator.
20
4. Kirchhoff-Love Beam Theory
The configuration space of Reissner type beams is described by pointwise six degrees of freedom, namely the three
translational components of r(s) and three rotational degrees of freedom ψ(s) parametrizing Λ(s). In this section, the
assumption of vanishing shear strains is made, which can be assumed as a sensible approximation for highly slender
beams (see e.g. [101]). Thus, the cross-sections spanned by g2 and g3 have to remain perpendicular to t(s) := r0 (s):
g2 (s) · t(s) ≡ 0
and g3 (s) · t(s) ≡ 0
or g1 (s) ≡
t(s)
.
||t(s)||
(67)
Principally, this so-called Kirchhoff constraint of vanishing shear deformations can be enforced in a strong or in a
weak manner. If the same parametrization (r(s), ψ(s)) as in the Reissner case is chosen, additional fields of Lagrange
multipliers λ2 (s) and λ3 (s) are necessary, in order to integrate (67) into a constrained variational problem in a weak
sense (see Section 4.4). In the following Sections 4.1-4.3, the concept of developing a parametrization with 4 degrees
of freedom (r(s), ϕ(s)) that fulfills the Kirchhoff constraint in a strong manner as already derived in [104] will be
briefly repeated and extended to the dynamic case. As already introduced in Section 2.2, the scalar-valued quantity
ϕ(s) will describe the relative rotation between the material frame gi (s) and an intermediate frame g Mi (s) with respect
to the tangent vector t(s) according to (16). By means of (15), one example for a suitable intermediate frame g Mi (s) has
already been given, the ”Smallest Rotation” intermediate frame. Nevertheless, the following derivations are presented
in a rather general manner, which allows to insert arbitrary alternative intermediate frame definitions g Mi (s).
4.1. Kinematics
Within the following three sections, the Kirchhoff constraint (67) is strongly enforced (see also (16)) according to
g1 (s) =g M1 (s) =
t(s)
||t(s)||
with
t(s) = r0 (s)
and
Λ = exp(S(ϕg1 ))Λ M
with Λ M = (g M1 , g M2 , g M3 )Ei ,
(68)
where the centerline-aligned intermediate triad base vectors g Mi (s) = g Mi (r0 (s)) are completely defined by the centerline field r(s) but not further specified for now (a possible example is given by (15)). Now, the Kirchhoff constraint is
incorporated by expressing the current configuration s, t → C := (r(s, t), Λ(r(s, t), ϕ(s, t))) ∈ <3 × SO(3) via the new
set of primary variables (r(s, t), ϕ(s, t)). Inserting (68) in the definition of the curvature vectors (28) yields (see [104]):
k = kk + k⊥ = K1 g1 + S(g1 )g01 = (K M1 + ϕ0 )g1 +
S(r0 )r00
||r0 ||2
| {z }
with
K M1 := gTM3 g0M2 .
(69)
=:κ
Here, K M1 represents the torsion of the (arbitrary) intermediate triad field and κ is the Frenet-Serret curvature of the
beam centerline r(s). In components, the spatial as well as the material curvature vector read:
T
T
k = K M1 +ϕ0 , gT2 κ, gT3 κ
and K = K M1 +ϕ0 , gT2 κ, gT3 κ .
(70)
gi
Ei
The intermediate torsion K M1 is the only term in (70) that depends on the specific choice of the intermediate triad.
Besides the curvature vectors, also the spin vector δθ has to be adapted to the Kirchhoff constraint (see also (25)):
δθ = δθk + δθ⊥ = δΘ1 g1 + S(g1 )δg1 = (δΘ M1 + δϕ)g1 +
S(r0 )δr0
.
||r0 ||2
(71)
In analogy to Reissner beam formulations, the first component δΘ1 of the spin vector, representing a multiplicative
increment, will directly be employed in the weak form and not further expressed via additive increments according to
δΘ1 = δΘ M1 +δϕ. Consequently, the admissible variations are s → δC := (δr(s),δθ(δr(s), δΘ1 (s), r(s, t))) ∈ <3 ×<3
with the new set of variational primary variables (δr(s), δΘ1 (s)) defining the Kirchhoff case. Again, compatibility
conditions similar to (31) can be stated. Due to the Kirchhoff constraint, these conditions are only required for the
tangential vector components. Left-multiplication of the relations in the second line of (31) with ET1 = (ΛT g1 )T yields:
K̇1 = W10 + ġT1(g1 ×g01 ), δK1 = δΘ01 +δgT1(g1 ×g01 ), δW1 = δΘ̇1 +δgT1(g1 × ġ1 ).
21
(72)
4.2. Deformation measures and stress resultants
Having defined kinematics that are compatible with the Kirchhoff constraint according to (67), the deformation measures, constitutive relations and stress resultants presented in Section 3 can be adapted to the shear-free case. Inserting
the constrained curvature vectors from (70) into the deformation measures Ω and ω according to (34) and (35) yields
T
T
Ω = K M1 +ϕ0 −K M01 −ϕ00 , gT2 κ−gT02 κ0 , gT3 κ−gT03 κ0 , ω = K M1 +ϕ0 −K M01 −ϕ00 , gT2 κ−gT02 κ0 , gT3 κ−gT03 κ0 . (73)
Ei
gi
The objective variation of (73) is still given by δΩ = Λ δθ , δo ω = δθ , but now the spin vector is constrained according
to (71). By construction, the shear components of Γ and γ in (34) and (35), vanish due to the Kirchhoff constraint (67):
T
0
0
Γ = E1 , γ = g1 → δΓ = δE1 , δo γ = δg1
with := ||r0 || − 1, δ = δr0T g1 .
(74)
Here, the abbreviation has been introduced for the remaining component representing the axial tension. Based on
the force split f = fk +f⊥ = F1 g1 +f⊥ , also the constitutive relations simply as consequence of the Kirchhoff constraint :
1
∂Π̃int
∂Π̃int
1
= C M Ω = ΛTm, F1 =
= EA.
Π̃int (Ω, ) = ΩTC M Ω+ EA 2 , M =
2
2
∂Ω
∂
(75)
The inertia forces fρ as well as the inertia moments mρ are identical to (42) and (43). Like in the Reissner case,
the spatial or material angular velocities w and W as well as the spatial or material angular accelerations a and A
can either be directly used in the employed time integration scheme or they can be expressed via the (additive) rate
of the primary variables (r(s), ϕ(s)). For the latter approach, transformation matrices depending on the definition
of the employed intermediate triad (see e.g. T−1
M in (24) in the case of the SR intermediate triad) as well as their
time-derivatives are required in order to formulate relations similar to (44).
w = Tθt ṫ + ϕ̇g1 ,
a = Ṫθt ṫ + Tθt ẗ + ϕ̇ġ1 + ϕ̈g1 .
(76)
In the Kirchhoff case, a third variant can be advantageous: Similar to the curvature vector k (see (69)) and the spin
vector δθ (see (71) or (25)), also the angular velocity w can be split into the following two components
w = wk +w⊥ = W1 g1 +S(g1 )ġ1 = W1 g1 +
S(r0 )ṙ0
,
||r0 ||2
a = Ẇ1 g1 +W1 ġ1 +S(g1 )g̈1 .
(77)
While w⊥ is determined by r(s), W1 is not specified any further, i.e. it is not expressed via r(s) and ϕ(s) as done in (76).
Remark: Three possible time integration schemes result from the different variants given above:
1)
(ṙ, r̈, w, a)n+1 = f((r, Λ)n+1 , (r, Λ)n , (ṙ, w)n , (r̈, a)n ),
2)
(ṙ, r̈, ϕ̇, ϕ̈)n+1 = f((r, ϕ)n+1 , (r, ϕ)n , (ṙ, ϕ̇)n , (r̈, ϕ̈)n ),
(78)
3) (ṙ, r̈, W1 , Ẇ1 )n+1 = f((r, Λ)n+1 , (r, Λ)n , (ṙ, W1 )n , (r̈, Ẇ1 )n ).
sd
Remark: From the tangent transformations (69), (71) and (77), the following similarities become obvious:
1
δt !
1
t0 !
1
ṫ !
S(g
)
g
S(g
)
g
S(g
)
g
sdδθ = ||t||
, k = ||t||
, w = ||t||
. (79)
1
1
1
1
1
1
δΘ1
K1
W1
4.3. Strong and weak form
In this section, the spatial representation of mechanical equilibrium will be considered. In the following, the notation
will be simplified by summarizing external forces and moments as well as inertia forces and moments according to
f̃ρ := f̃ +fρ and m̃ρ := m̃+mρ . The shear forces f⊥ , which provide no work contribution in the Kirchhoff case, can be
eliminated from the strong form of mechanical equilibrium yielding the following set of four differential equations:
gT1 m0 + m̃ρ = 0,
" 0
(80)
#0
r
0
0
(F1 g1 ) +
×
m
+
m̃
+
f̃
=
0.
ρ
ρ
||r0 ||2
22
The set (80) is sufficient in order to solve for the four primary variables r and ϕ as soon as the stress resultants fk and
m are expressed by kinematic and constitutive relations from Section 4.2. Multiplying (80) with the admissible translational and rotational variations δr and δΘ1 and integrating by parts gives the equivalent weak form of equilibrium:
Zl h
i
h
i
0T
G = δθ
m + δr0Tg1 F1 −δrT f̃ρ − δθTm̃ρ ds − δrTfσ +δθTmσ =˙ 0,
|{z}
Γσ
| {z }
0
δo
ωTm
δθ = δΘ1 g1 +
r0 ×δr0
.
||r0 ||2
(81)
δF1
In (81), the constrained spatial spin vector according to (71) has been identified and already substituted with the
symbol δθ. As indicated by the curly brackets in (81), the pre-factors of the stress resultants m and F1 are represented by the objective variations δo ω and δ, underlining the geometrical exactness of the proposed Kirchhoff beam
formulation. Finally, the problem setup has to be completed by proper boundary and initial conditions:
r = ru , g1 = g1u , ϕ = ϕu on Γu , f = fσ , m = mσ on Γσ , Γσ ∩ Γu = ∅, Γσ ∪ Γu = {0, l}
r = r0 , ṙ = v0 , ϕ = ϕ0 , w = w0 at t = 0.
(82)
Here g1u prescribes the orientation of the tangent vector and ϕu the orientation of the cross-section with respect to
a rotation around the tangent vector. How these conditions can be modeled in practice is shown in Appendix B.
By introducing the trial space (r, Λ(r, ϕ)) ∈ U satisfying (82) and the weighting space (δr, δθ(δr, δΘ1 , r)) ∈ V, with
δr = 0, δΘ1 = 0 on Γu , the beam problem is fully defined. It should be emphasized that only the concrete analytic
expressions for g M2 (r(s)) and g M3 (r(s)) in (68) depend on the specific choice of the intermediate triad definition.
Remark: In [104], a special choice of intermediate triad field in (68) has been proposed, the Frenet-Serret
triad of the beam centerline. It has been argued in this work, that this choice is not favorable for numerical
purposes as consequence of singularities occurring on straight beam segments. However, the Frenet-Serret
frame is very beneficial for the analytical treatment of Kirchhoff-Love beam problems and has e.g. been
applied in order to derive analytical reference solutions for the numerical examples of Section 11.
Remark: As shown in [105], the Kirchhoff-Love beam theory presented so far provides an ideal basis
for the derivation of reduced beam models, i.e. beam models that are valid under certain restrictions concerning initial geometry and external loads and will eventually yield considerably simplified finite element
formulations. On this basis, a special torsion-free beam element formulation could be derived that completely abstains from any rotational degrees of freedom, a distinct property that drastically simplifies many
standard procedures in the numerical treatment of geometrically exact beam element formulations.
4.4. Weak enforcement of Kirchhoff constraint
In the last sections, the set of primary variable fields (r(s), ϕ(s)) has been chosen in a way such that the Kirchhoff
constraint (67) of vanishing shear strains is strongly fulfilled by construction. However, more flexibility in the subsequent discretization process (see e.g. Section 6) could be gained by formulating a Reissner type beam problem, which
allows for two independent interpolations for the centerline field r(s) as well as the triad field Λ(s), and by weakly
enforcing the Kirchhoff constraint of vanishing shear strains by means of additional constraint equations:
Γ j (s) ≡ gTj (s)r0 (s) ≡ 0
for
j = 2, 3.
(83)
In order to integrate these constraints into a variational framework, an additional Lagrange multiplier potential
ΠλΓ23 =
Zl
(λΓ2 (s)Γ2 (s) + λΓ3 (s)Γ3 (s))ds
(84)
0
is required. The Lagrange multiplier fields λΓ2 (s) and λΓ3 (s) in (84) can be interpreted as the shear force components
F2 (s) and F3 (s), i.e. reaction forces which enforce the constraint of vanishing shear strains. Variation of the Lagrange
multiplier potential (84) leads to the contribution of the Kirchhoff constraint to the weak form:
δΠλΓ23 =
Zl
Zl
(δλΓ2 (s)Γ2 (s)+δλΓ3 (s)Γ3 (s))ds+
0
(λΓ2 (s)δΓ2 (s)+λΓ3 (s)δΓ3 (s))ds.
0
23
(85)
The first term in (85) represents the weak statement of the Kirchhoff constraint (83) while the second term can be
interpreted as the work contribution of the shear reaction forces. Similar to the displacement primary fields, a proper
trial space (λΓ2 , λΓ3 ) ∈ UλΓ23 and a proper weighting space (δλΓ2 , δλΓ3 ) ∈ VλΓ23 have to be introduced which uniquely
define the resulting mixed beam formulation. The discrete realization of (85) will be presented in Section 10.
5. Temporal Discretization of Primary Fields
In this and the next section, the space-time-continuous beam problems presented so far will be discretized. While
spatial discretization as discussed in Section 6 will be based on the finite element method (FEM), a recently proposed
extension of the well-known generalized-α method [37] from vector spaces to Lie groups [7, 31, 32], which is directly
applicable to the beam element formulations proposed in this work, will be presented in this section. This time integrator represents an implicit, one-step finite difference scheme that inherits most of the desirable properties of the standard
generalized-α variant. In the context of finite element methods for solid mechanics, it is often more convenient to perform time discretization on the semi-discrete problem setting resulting from spatial discretization. Here, just the opposite succession, i.e. the initial time discretization is followed by a subsequent spatial discretization, is chosen. This
second variant is often applied in the development of geometrically exact beam finite element formulations and will
lead to simpler discrete expressions. In the following, the considered total time interval t ∈ [0, T ] is subdivided into
equidistant subintervals [tn , tn+1 ] with constant time step size ∆t, where n ∈ N0 is the time step index. Consequently,
the solution for the primary variable fields describing the current configuration C(s, t) := (r(s, t), Λ(s, t)) is computed
at a series of discrete points in time with associated configurations C(s, tn ) := (r(s, tn ), Λ(s, tn )) =: (rn (s), Λn (s)). In
order to simplify notation required for subsequent derivations, the weak form G (see e.g. (33) or (81)) is split into the
contributions Gint of internal forces, Gkin of kinetic forces and Gext of external forces:
G =Gint +Gkin −Gext .
(86)
In a next step, the basics of the Lie group generalized-α method originally proposed by [7, 31, 32] will be presented.
This method will be applied for time discretization of the Reissner and Kirchhoff type beam element formulations
presented in subsequent sections, whose configuration space C(s, t) = (r(s, t), Λ(s, t)) ∈ <3 × SO(3) is defined by the
position field r(s, t) and the rotation field Λ(s, t). It is emphasized that the following procedure is independent from
the rotation parametrization of Λ(s) employed to these different beam element formulations. In order to express the
(material) angular velocities and accelerations Wn+1 and An+1 at the end of a time interval [tn , tn+1 ] in terms of known
quantities at time tn and the unknown rotation field Λn+1 (s), the vectors θ̃n+1 and Θ̃n+1 , representing the spatial and
material multiplicative rotation increment between the time steps tn and tn+1 , are introduced:
exp(S(Θ̃n+1 )) = ΛTn Λn+1 ,
Θ̃n+1 = ΛTn θ̃n+1 = ΛTn+1 θ̃n+1 .
(87)
Besides the distinctions that can be made for vector space time integrators (e.g. implicit or explicit scheme, onestep or multi-step scheme, employed methodology in order to guarantee stability and/or conservation properties) two
further classifications can be made for time integration schemes applied to rotational variables: First, depending on the
type of spatial rotation interpolation, the succession of spatial and temporal discretization will in most cases influence
the resulting discrete solution. Secondly, it can be distinguished between approaches that apply a time integration
scheme directly to the vectors (Θ̃, W, A) and approaches that express angular velocities and accelerations by means of
(additive) rates of the primary variables, i.e. (ψ, ψ̇, ψ̈) (see also (44)). The axial vectors Θ̃, W and A are associated with
elements of the Lie algebra so(3). Consequently, time integration schemes of the former type are commonly denoted
as Lie group time integration schemes. In this context, it can be further distinguished between Lie group schemes that
are based on the material vectors (Θ̃, W, A) and schemes that are based on their spatial counterparts (θ̃, w, a). Only
the former variant will be considered in this work. Arguably, one of the first Lie group time integration schemes, at
least in the context of geometrically exact beam formulations, has been proposed by Simo and Vu-Quoc [135] and
represents the Lie group extension of the classical Newmark scheme. On the contrary, the scheme of [7, 31, 32],
which will be presented in the following, is the Lie group extension of the standard generalized-α method. Also this
scheme is based on the four parameters β, γ, αm and α f and simplifies to the variant of Simo and Vu-Quoc [135] for
24
the special choice αm = α f = 0. A distinctive feature of the Lie group generalized-α scheme lies in the fact that all
terms of the weak form are evaluated at the end point tn+1 of the considered time interval:
Gn+1 =G(rn+1 , ṙn+1 , r̈n+1 , Λn+1 , Wn+1 , An+1 , f̃n+1 , m̃n+1 , fσ,n+1 , mσ,n+1 )=0.
˙
(88)
In a next step, the update formulas for translational quantities are given by a standard Newmark scheme
ũn+1 := rn+1 − rn ,
ũn+1 = ∆tṙn + +∆t2 [(0.5 − β)r̈mod,n + βr̈mod,n+1 ],
(89)
ṙn+1 = ṙn + ∆t[(1 − γ)r̈mod,n + γr̈mod,n+1 ].
This update scheme is slightly changed in form of a multiplicative configuration update for the rotations:
exp(S(Θ̃n+1 )) = ΛTn Λn+1 ,
Θ̃n+1 = ∆tWn + +∆t2 [(0.5 − β)Amod,n + βAmod,n+1 ],
(90)
Wn+1 = Wn + ∆t[(1 − γ)Amod,n + γAmod,n+1 ].
The modified acceleration vectors r̈mod occuring in (89) are related to the physical acceleration vectors r̈ according to:
(1 − αm )r̈mod,n+1 + αm r̈mod,n = (1 − α f )r̈n+1 + α f r̈n
with
r̈mod,0 = r̈0 .
(91)
In a similar manner, the modified as well as the real / physical angular accelerations are related according to:
(1 − αm )Amod,n+1 + αm Amod,n = (1 − α f )An+1 + α f An
with Amod,0 = A0 .
For later use, it is favorable to express the vectors ṙn+1 and r̈n+1 in terms of the primary unknown rn+1 :
!
!
γ
γ
γ
ṙn + ∆t 1 −
r̈mod,n ,
ũn+1 + 1 −
ṙn+1 =
β∆t
β
2β
1 − αm
1 − αm
r̈n+1 =
ũn+1 −
ṙn
β∆t(1 − α f )
β∆t2 (1 − α f )
"
#
αf
(1 − αm )(0.5 − β)
αm
+ −
r̈mod,n −
+
r̈n .
β(1 − α f )
1 − αf
1 − αf
A similar relation can be formulated for the material vectors of angular velocity Wn+1 and acceleration An+1 :
!
!
γ
γ
γ
Θ̃n+1 + 1 −
Wn + ∆t 1 −
Amod,n ,
Wn+1 =
β∆t
β
2β
1 − αm
1 − αm
An+1 =
Θ̃n+1 −
Wn
β∆t(1 − α f )
β∆t2 (1 − α f )
"
#
αf
(1 − αm )(0.5 − β)
αm
+ −
+
Amod,n −
An .
β(1 − α f )
1 − αf
1 − αf
(92)
(93)
(94)
In [31] and [32], it has been proven that the integration scheme given by equations (88)-(94) yields the same favorable properties as the standard generalized-α method, which are second-order accuracy, unconditional stability
(within the linear regime), controllable damping of the high-frequency modes and minimized damping of the lowfrequency modes. Remarkably, the parameter choice leading to this optimal behavior is identical to that of the standard generalized-α scheme. Furthermore, it has been shown that this scheme can consistently treat non-constant mass
matrix contributions, such as the term ΛCρ A occurring in geometrically exact Reissner and Kirchhoff type beam
formulations. It has to be stated that the extended generalized-α scheme cannot guarantee for exact conservation of
energy, linear and angular momentum. In the field of Lie group time integration schemes, a large variety of methods aiming to guarantee these conservation properties has been proposed [136, 26, 61, 14, 21, 75, 98, 119, 83, 46].
25
However, the perhaps most essential advantage of the extended generalized-α scheme as compared to these alternatives lies again in its simplicity and flexibility. Independent of the beam theory (Reissner or Kirchhoff type), the
employed spatial interpolation schemes as well as the chosen set of nodal primary variables (e.g. in terms of rotation
parametrization), this time integration scheme can directly be applied without the need for any further adaptions.
Review: As mentioned in Section 5, a large number of scientific contributions considers the development
of energy-momentum schemes for Simo-Reissner type formulations. However, this does not hold true for
Kirchhoff-Love beam element formulations. For example, the formulations of anisotropic type (see requirements 5a)-5c) in Table 2) haven’t considered dynamics so far. In [151], a energy-momentum scheme
is proposed, but only for temporal discretization of the beam centerline. The torsional problem associated
with the rotational degree of freedom is only considered in a static manner, i.e. arbitrary dynamic problems containing also rigid body rotations with respect to the beam axis are not accessible by this approach.
The temporal discretization applied in the present work encompasses both, the translational and the rotational fields. As verified numerically in Section 11, the numerical dissipation provided by the extended
generalized-α scheme will enable energy-stable time integration for highly nonlinear problems. Concretely,
in the well-known ellbow-cantilever example of Section 11.7, a considerably improved energy stability can
be observed as compared to [80], where a Lie-group extension of the standard Newmark scheme [135] has
been employed. Nevertheless, a supplementation of the proposed Kirchhoff-Love beam elements by energymomentum schemes for temporal discretization seems to be a interesting direction of future research.
6. Spatial Discretization Methods for Primary Fields
Spatial discretization is exclusively considered in the context of finite element methods within this work. It represents
the core topic in the development of geometrically exact Kirchhoff beam elements considered in this work.
6.1. Discretization of beam centerline
First, the spatial discretization of the beam centerline r(s) will be conducted. Thereto, an elementwise parameter space
ξ ∈ [−1; 1]. is introduced with the element Jacobian J(ξ) := ||r0h,ξ (ξ)|| mapping between infinitesimal increments in
the parameter space and the arc-length space according to ds = J(ξ)dξ. In the following two sections, two different
interpolation schemes based on Lagrange polynomials or Hermite polynomials respectively will be presented.
6.1.1. Discretization of beam centerline based on Lagrange polynomials
The highest derivative of the primary variable r(s) occurring in the weak form (33) of the Simo-Reissner beam theory
is the first derivative r0 (s) of the centerline curve. Consequently, a C 0 -continuous interpolation of the beam centerline
r(s) is sufficient in this case. Thus, the standard choice of trial functions is given by Lagrange polynomials, yielding
rh (ξ) =
nr
X
Li (ξ)d̂i =: Ld̂ and
r0h (ξ) =
nr
X
Li (ξ)d̂i0 =: Ld̂0 .
(95)
i=1
i=1
In (95), the vectors d̂i , d̂i0 ∈ <3 represent the nodal positions, whereas Li (ξ) are standard Lagrange polynomials satisfying the interpolation property Li (ξ j ) = δi j at the element node coordinates ξ j as well as proper completeness conditions.
Here, δi j represents the Kronecker delta symbol. The matrix L and the vector d̂ represent element-wise assemblies of
the shape functions Li and the position vectors d̂i for i = 1, ..., nr , which are given by L := (L1 (ξ)I3 , ..., Lnr (ξ)I3 ) as well
as d̂ := (d̂1T , ..., d̂nr T )T . Following a Bubnov-Galerkin approach, the interpolation of the trial functions δr(s) reads:
δrh (ξ) =
nr
X
Li (ξ)δd̂i =: Lδd̂.
i=1
26
(96)
6.1.2. Discretization of beam centerline based on Hermite polynomials
The highest derivative of the primary variable r(s) occurring in the weak form (81) of the Kirchhoff beam is given by
the second derivative r00 (s). Consequently, for the interpolation of the centerline r(s) shape functions are required that
fulfill C 1 -continuity at the element boundaries. Besides this requirement, a C 1 -continuous centerline representation
can be very beneficial for problem class requiring a smooth geometry representation such as beam-to-beam contact
problems (see e.g. [106, 107]). In order to guarantee for C 1 -continuity, Hermite shape functions are employed:
2
2
2
2
X
X
cX i
cX i
Ht (ξ)t̂i =: Hd̂, r0h (ξ) = Hdi (ξ)d̂i0 +
Ht (ξ)t̂i0 =: Hd̂0 .
rh (ξ) = Hdi (ξ)d̂i +
2
2
i=1
i=1
i=1
i=1
(97)
In (97), the vectors d̂i , d̂i0 ∈ <3 and t̂i , t̂i0 ∈ <3 represent nodal position and nodal tangent vectors at the two boundary
nodes of the resulting finite elements. Again, the matrix H and the vector d̂ represent proper element-wise assemblies
of the shape functions Hti and Hdi as well as the nodal position and tangent vectors d̂i and t̂i for i = 1, 2. The explicit
expressions are H := (Hd1 (ξ)I3 , 0.5cHt1 (ξ)I3 , Hd2 (ξ)I3 , 0.5cHt2 (ξ)I3 ) as well as d̂ := (d̂1T , t̂1T , d̂2T , t̂2T )T . In [104], it has
already been shown that the Hermite shape functions in (97) fulfill the interpolation property for the nodal positon and
tangent vectors as well as proper completeness conditions for third-order polynomials and that the optimal choice for
the constant c = lele,h is given by the element length lele,h of the discretized initial geometry, which can be determined
in an iterative manner. On the basis of (97), the first and second arc-length derivative can be determined as
r0h (ξ) =
1
rh,ξ (ξ),
J(ξ)
J,ξ (ξ)
1
rh,ξξ (ξ) −
rh,ξ (ξ),
2
J(ξ)
J(ξ)3
r00h (ξ) =
(98)
where the Jacobi factor J(ξ) and its parameter derivative J,ξ (ξ) appearing in (98) are given by the following relations:
J(ξ) := ||r0h,ξ (ξ)||
and
rT0h,ξ r0h,ξξ
d
.
J,ξ (ξ) = ||r0h,ξ || =
dξ
||r0h,ξ ||
(99)
Finally, also the variations δr(s) have to be discretized by properly chosen test functions. Following a BubnovGalerkin approach, also the interpolation of δr(s) is based on Hermite polynomials and given by the expression:
δrh (ξ) =
2
X
i=1
2
Hdi (ξ)δd̂i +
cX i
H (ξ)δt̂i =: Hδd̂.
2 i=1 t
(100)
In an analogous manner, the procedure presented here can also be extended to Hermite polynomials of higher order.
6.2. Discretization of rotation field
In Section 2.1, two parametrizations of rotation tensors have been investigated: a parametrization via rotation vectors
ψ and a parametrization via the SR mapping on the basis of the set (t, ϕ). In the following two sections, these two
variants will be employed in order to parametrize the rotation tensors Λi at the element nodes i = 1, ..., nΛ . In the
subsequent Sections 6.2.3 and 6.2.4, also two possible approaches for the interpolation of these nodal triads in the
elements interior, one based on rotation vectors and one based on the SR mapping, will be presented.
6.2.1. Parametrization of nodal triads via rotation vectors
According to Section 2.1, rotation vectors ψ̂1n , ..., ψ̂nnΛ can be employed as primary variables describing the nodal triads
Λin = Λin (ψ̂in ), i = 1, 2, ..., nΛ , at time step tn . The update from iteration k to k + 1 of a nonlinear solution scheme might
either be based on additive rotation increments ∆ψ̂i,k+1
or on multiplicative rotation increments ∆θ̂ni,k+1 given by:
n
i,k
i,k+1
)
Λi,k+1
(ψ̂i,k+1
) = Λi,k
n (ψ̂n +∆ψ̂n
n
n
i,k
) = exp(S[∆θ̂ni,k+1 ])Λi,k
or Λi,k+1
(ψ̂i,k+1
n
n (ψ̂n ).
n
(101)
Only for rotations with a magnitude smaller than 180◦ a unique rotation vector can be extracted from a given triad
(by applying e.g. Spurrier’s algorithm, see [140]). Within this work, rotation vectors are always extracted in a manner
such that ψ ∈] − π; π]. Within this range, the transformation matrix T between additive and multiplicative rotation
vector increments (see (12)) and its inverse, which do not exist at ψ = 2π, are always well-defined.
27
6.2.2. Parametrization of nodal triads via smallest rotation mapping
Alternatively, the nodal triads can be defined via a relative rotation of nodal intermediate triads ΛiMϕ̂ by a relative angle
ϕ̂i with respect to the tangent (see (16)). The nodal intermediate triads are defined by the smallest rotation mapping
of the nodal intermediate triad Λ̄iMϕ̂ ,n := ΛiMϕ̂ ,n−1 of the last time step onto the basis vector gi1,n of the current step tn :
Λin = exp(S[ϕ̂in gi1,n ])ΛiMϕ̂ ,n ,
ΛiMϕ̂ ,n = sr(Λ̄iMϕ̂ ,n , gi1,n )
with Λ̄iMϕ̂ ,n := ΛiMϕ̂ ,n−1 .
(102)
This variant will be used for Kirchhoff type beam element formulations in combination with the Hermite centerline
interpolation (97), where the first base vector gi1 is defined via the tangent vector to the beam centerline, i.e. gi1 =
t(ξi )/||t(ξi )||. All in all, the nodal triad is defined by the nodal relative angle ϕ̂i and the tangent vector t(ξi ) at the
node i, i.e. Λi = Λi (ϕi , t(ξi )). However, it has to be emphasized that in contrary to ϕ̂i , the vector t(ξi ) does not
necessarily have to be a nodal primary variable. If the considered node i coincides with one of the two element
boundary nodes employed in the Hermite interpolation (97), this tangent vector indeed represents a nodal primary
variable, i.e. t(ξi ) = t̂i for i = 1, 2, otherwise the tangent vector simply represents the interpolated centerline derivative
at this position, i.e. t(ξi ) = r0 (ξi ). Based on (additive) increments ∆ϕ̂i,k+1
of the nodal relative angles, the configuration
n
update from iteration k to iteration k + 1 of the Newton-Raphson scheme at time step tn reads:
i,k+1
i,k+1
i,k+1
i,k+1
i
i,k
i,k+1 i,k+1
Λi,k+1
(ϕ̂i,k+1
n
n , g1,n ) = exp(S[ϕ̂n +∆ϕ̂n , g1,n ])Λ Mϕ̂ ,n , Λ Mϕ̂ ,n = sr(Λ̄ Mϕ̂ ,n , g1,n ).
(103)
0 k+1 i
, d̂2,k+1
(ξ )/||r0nk+1 (ξi )|| are fully defined by the centerline r0nk+1 = r0 (d̂1,k+1
, t̂n1,k+1 , t̂2,k+1
The base vectors gi,k+1
)
n
n
n
1,n = rn
i,k+1
i,k
i,k+1
i,k+1
i,k
i,k+1
based on the additive updates d̂n = d̂n +∆d̂n as well as t̂n = t̂n +∆t̂n .
Remark: Within this work, intermediate triads Λ M based on the SR mapping are used for two different
purposes: Firstly, they are used for the definition of nodal material triads Λi based on nodal relative angles
ϕ̂i and associated nodal intermediate triads ΛiMϕ̂ (SR mapping ”in time” from Λ̄iMϕ̂ ,n = ΛiMϕ̂ ,n−1 to ΛiMϕ̂ ,n ,
see (102)). Secondly, they are used for the definition of an interpolated material triad field Λ(ξ) based on
a relative angle field ϕ(ξ) and an associated intermediate triad field Λ Mϕ ,n (ξ) (SR mapping ”in space” from
a reference triad Λr,n = Λn (ξr ) to Λ Mϕ ,n (ξ), see (113)). In order to distinguish these two applications, the
additional index ϕ̂ or ϕ of Λ M refers to the associated relative angle.
Review: In Appendix Appendix B it is shown that a rotation vector-based parametrization of the nodal
triads according to the last section will considerably simplify the modeling of complex Dirichlet boundary
conditions and joints as compared to the tangent vector-based variant considered in this section. Only the
straight formulations from the literature (see requirement 6) of Table 2) employ a rotation vector-based
triad parametrization, which supports a simple prescription of such conditions. In Section 11, it will turn
out that the tangent vector-based parametrization of nodal triads on the other hand will lead to a better
nonlinear solver performance. Thus, in practice one might combine the advantage of these to approaches
by employing a rotation vector-based parametrization only at nodes where complex Dirichlet or constraint
conditions have to be modeled and the tangent vector-based variant to all the remaining nodes.
6.2.3. Triad interpolation based on local rotation vectors
In this section, a triad interpolation is presentented that has originally been proposed by Shoemake [131] and for the
first time employed to geomerically exact beam element formulations by Crisfield and Jelenić [43, 80]. On each of
the considered nΛ nodes, a triad Λi , with i = 1, 2, ..., nΛ , is defined by primary degrees of freedom either according to
Section 6.2.1 or to Section 6.2.2. The interpolation strategy presented in this section is independent from the specific
choice of nodal primary variables. First, a reference triad Λr based on the triads at nodes I and J is defined:
Λr = ΛI exp(S(ΦI J /2))
with
exp(S(ΦI J )) = ΛITΛ J .
(104)
The nodes I and J are chosen as the two middle triads for elements with an even number nΛ of nodes and as the one
middle triad (i.e. I = J) for elements with an odd number nΛ of nodes (see also equation (6.2) in [43], which is based
28
on a sightly different node numbering). Based on the definition Λr , the interpolated triad field is defined as follows:
Λh (ξ) = Λr exp(S(Φlh (ξ))),
Φlh (ξ) =
nΛ
X
Li (ξ)Φil ,
exp(S(Φil )) = ΛTrΛi .
(105)
i=1
Again, Li (ξ) represent the standard Lagrange polynomials of order nΛ − 1 and Φil the (material) rotation vectors
associated with the relative rotation between the triad Λi at node i and the reference triad Λr . The interpolation (105)
represents an orthonormal interpolation scheme. Thus, the interpolated triad field is still an element of the rotation
group, i.e. Λh (ξ) ∈ SO(3) ∀ ξ ∈ [−1; 1]. Furthermore, the interpolation scheme (105) preserves the objectivity of the
space-continuous deformation measures (see [43]). The curvature vector (see (28)) resulting from (105) reads
Kh (ξ) = T−T (Φlh (ξ))Φ0lh (ξ),
(106)
and can exactly represent the state of constant curvature Kh = const . Thus, the two-noded variant of (105) can be
identified as a geodesic interpolation scheme, since it connects two points on the nonlinear manifold SO(3) via the
”shortest distance”. Consequently, the two-noded variant of this interpolation represents the SO(3)-counterpart to the
linear interpolation (95) of quantities in <3 . In contrast to the interpolations (95) and (97) of the beam centerline, the
rotation interpolation (105) is nonlinear in the nodal degrees of freedom. Thus, if e.g. the field of rotation vectors
ψh (ξ) with nodal values ψh (ξi ) = ψ̂i is employed for triad parametrization, the rotation vector interpolation resulting
from (105) can be written in an abstract manner in the form ψh (ξ) = nl(ψ̂1 , ..., ψ̂nΛ , ξ). While an explicit interpolation
rule for the rotation vectors ψh (ξ) is not needed for practical purposes - the triad field is already given by (105) - the
discrete version of the spin vector field δθh (ξ) and the field of (multiplicative) rotation vector increments ∆θh (ξ) will
be required in the next sections for the spatially discretized weak form of the balance equations and its linearization.
In [43], these discretized fields have consistently been derived from the triad interpolation (105), leading to:
∆θh (ξ) =
nΛ
X
Ĩi (ξ)∆θ̂i =: Ĩ∆θ̂θ ,
δθh (ξ) =
i=1
nΛ
X
Ĩi (ξ)δθ̂i =: Ĩδθ̂θ .
(107)
i=1
The arc-length derivatives of the interpolations (107) follow in a straightforward manner and are given by:
∆θh0 (ξ) =
nΛ
X
1 i
Ĩ,ξ (ξ)∆θ̂i ,
J(ξ)
i=1
δθh0 (ξ) =
nΛ
X
1 i
Ĩ,ξ (ξ)δθ̂i .
J(ξ)
i=1
(108)
The generalized shape function matrices Ĩi (ξ) ∈ <3 × <3 as well as their derivatives Ĩi,ξ (ξ) have been derived in the
original work [43] (see also Appendix A). Again, assembly matrices and vectors Ĩ := (Ĩ1 , ..., ĨnΛ ) as well as ∆θ̂θ :=
(∆θ̂1T , ..., ∆θ̂nΛ T )T and δθ̂θ := (δθ̂1T , ..., δθ̂nΛ T )T have been introduced. These shape functions depend on the rotational
primary variables in a nonlinear manner, e.g. Ĩi (ξ) = nl(ψ̂1 , ..., ψ̂n , ξ) if nodal rotation vectors according to Section 6.2.1
are employed. Consequently, they have to be re-calculated for every new configuration and this dependency on the
rotational primary variables would have to be considered within a consistent linearization procedure in case the spin
vector interpolation δθh (ξ) given in (107) is used in the weak form according to a Bubnov-Galerkin procedure. In order
to avoid this additional linearization, it can be sensible to follow a Petrov-Galerkin approach based on an interpolation
of δθh (ξ) via Lagrange polynomials. This strategy is also applied within this work and leads to:
δθh (ξ) =
nΛ
X
Li (ξ)δθ̂i =: Lδθ̂θ ,
δθh0 (ξ) =
i=1
nΛ
X
1 i
L,ξ (ξ)δθ̂i =: L0 δθ̂θ .
J(ξ)
i=1
(109)
Nevertheless, the interpolation ∆θh (ξ) is still based on (107) in order to end up with a consistent linearization. As
emphasized in [80], the generalized shape functions fulfill the following interpolation and completeness properties
Ĩi (ξ j ) = δi j I3 ,
nΛ
X
Ĩi (ξ) ≡ I3 ,
i=1
nΛ
X
Ĩi,ξ (ξ) ≡ 0,
(110)
i=1
i.e. these shape functions can exactly represent constant rotation vector increment fields. Since these properties
are also fulfilled by the Lagrange polynomials, both the Bubnov-Galerkin interpolation (107) as well as the PetrovGalerkin interpolation (109) of the spin vector field δθh (ξ) can exactly represent a constant distribution δθh (ξ) = const.
in an element. This property is important with respect to conservation of angular momentum (see e.g. Sections 6.3.5).
29
6.2.4. Triad interpolation based on ”Smallest Rotation” mapping
In this section, again, a triad interpolation with nΛ nodes is considered. On each of these nodes, a triad Λi , with
i = 1, 2, ..., nΛ , is defined by primary degrees of freedom either according to Section 6.2.1 or to Section 6.2.2. Similar
to the last section, the interpolation strategy presented in the following is independent from the specific choice of
nodal primary variables. Concretely, a novel interpolation scheme is proposed that defines an orthonormal triad
field Λh (ξ) ∈ SO(3) ∀ ξ ∈ [−1; 1] based on a given tangent vector field t(ξ) = r0 (ξ) and nodal triads Λi = (gi1 , gi2 , gi3 )
with i = 1, 2, ..., nΛ . In the following, this tangent vector field is defined by a Hermite interpolation of the beam
centerline according to (97) based on two nodes at the element boundary with six degrees of freedom d̂ j , t̂ j and
j = 1, 2, respectively. It has to be emphasized that the number of nodes of the triad interpolation can in general differ
from the number of nodes of the Hermite centerline interpolation, i.e. nΛ , 2 (see Figure 3(b)).
(a) Node numbering applied to Simo-Reissner elements.
(b) Node numbering applied to Kirchhoff-Love elements.
Figure 3: Element-local node numbering of translational and rotational primary fields.
The nodal triads are oriented tangential to the beam centerline curve. Thus, the first base vectors yield:
g1 (ξ) =
r0 (ξi )
r0 (ξ)
→ gi1 = 0 i .
0
||r (ξ)||
||r (ξ )||
(111)
Similarly to Section 6.2.3, one nodal triad ΛI initially has to be chosen as reference triad of the interpolation scheme:
Λr = Λ I
with
I ∈ {1, 2, ..., nΛ }.
(112)
Based on the reference triad Λr and the nodal triads Λi the interpolation procedure is defined according to:
Λh (ξ) = exp(S[ϕh (ξ)g1 (ξ)]) Λ Mϕ (ξ),
nΛ
X
ϕh (ξ) =
Li (ξ)ϕi ,
Λ Mϕ (ξ) = sr(Λr , g1 (ξ)),
exp(S[ϕi gi1 ]) = Λi ΛTMϕ (ξi ).
(113)
i=1
The general curvature vector Kh (ξ) for interpolations that fulfill the Kirchhoff constraint in a strong manner is given
by (70). The total torsion resulting from (113) can be derived in a straight-forward manner to:
K1 = K Mϕ 1 + ϕ0h ,
K Mϕ 1 = gTMϕ 3 g0Mϕ 2 = −
κTg1I
1+gT1 g1I
.
Remark: Ihe nodal relative angles ϕi = ϕh (ξi ) in (113) are different from the nodal primary variables ϕ̂i
in (102) of Section 6.2.2 for parametrization of nodal triads. For both quantities, the symbol ϕ has been
chosen since in both cases the relative angle between the material triad Λ(ξi ) and an intermediate triad
Λ Mϕ (ξi ) in case of ϕi (stemming from a SR mapping ”in space”) or an intermediate triad ΛiMϕ̂ in case of ϕ̂i
(stemming from a SR mapping ”in time”) is measured. The difference becomes clear by realizing that the
intermediate triad Λ Mϕ (ξi ) resulting from the smallest rotation of Λr onto g1 (ξi ) (see (113)) will in general
differ from the intermediate triad ΛiMϕ̂ resulting from the smallest rotation of Λ̄iMϕ̂ onto g1 (ξi ) (see (102)).
30
(114)
Review: In [104], it has been argued that the SR mapping exhibits a singularity for ḡT1 g1 = −1 in (15).
For the proposed finite element formulations this can only occur for rotation increments per time step
that are larger than 180◦ or large element deformations exhibiting relative rotations between the element
center and boundary nodes that are larger than 180◦ (see [104]). Consequently, these singularities are not
practically relevant for reasonable spatial and temporal discretizations. In [104], it has also been argued that
these singularities occurring at relative rotations of 180◦ represent the optimum that can be achieved for
mappings on the tangent vector. The approach employed by the anisotropic formulations from the literature
is sightly different (see requirement 2c) in Table 2). There, only one spatially fixed reference triad at the
beam endpoint is used for the initial SR mapping in space, and only one temporarily fixed reference triad at
every centerline position s for the SR mapping in time. Consequently, for these formulations, singularities
could occur for practically relevant configurations in case the relative rotation between the cross-sections at
the beam end point and at an arbitrary centerline position s exceeds 180◦ or in case the total cross-section
rotation from the initial to the current configuration exceeds 180◦ at any centerline position s.
Review: In the literature, the smallest rotation mapping defined in (15) of Section 2.2 is often denoted
as ”rotation without twist”. Thus, sometimes it is mistakenly assumed that an intermediate triad field as
employed in (113) would exhibit a vanishing torsion (see e.g. [63], where the slightly different interpolation
scheme (128) has been employed). However, according to (114), the torsion of the intermediate triad
field (113) constructed via the SR mapping does not vanish for general curved 3D configurations. It can
easily be shown that the torsion vanishes in the limit of fine discretizations lele := s(e),2 − s(e),1 → 0. In this
limit, the intermediate triad field becomes identical to an elementwise Bishop frame [23] and the relation
lim K Mϕ 1 (ξ) = − lim
I
ξ →ξ
lele →0
κTg1I
1+gT1 g1I
= − lim
I
ξ →ξ
(g1 (ξ) × g01 (ξ))Tg1 (ξ I )
1+gT1 (ξ)g1 (ξ I )
= 0.
(115)
holds true. However, as verified numerically in Section 11.5, a neglecting of K Mϕ 1 in the range of finite
element lengths will in general lead to a decline in the spatial convergence rate. On the contrary, the arclength derivatives (and variations) derived in the alternative anisotropic Kirchhoff-Love formulation [16]
(see requirement 4a) in Table 2), contain the required terms stemming from the SR mapping.
Again, the discrete version of the spin vector field δθh (ξ) has to be determined. Following a Petrov-Galerkin approach,
the spin vector of (71) can be discretized as follows:
δθh (ξ) = δΘ1h (ξ)g1 +
S(r0 )δr0
||r0 ||2
with δΘ1h (ξ) =
nΛ
X
Θ1 ,
Li (ξ)δΘ̂i1 =: Lk δΘ̂
(116)
i=1
where the interpolation of r0 as well as δr0 follows (97) and (100). The matrix Lk := (L1 , ..., LnΛ ) as well as the vector
Θ1 := (δΘ̂11 , ..., δΘ̂n1Λ )T represent assemblies of the Lagrange shape functions Li and the nodal twist components δΘ̂i1 .
δΘ̂
Alternatively, it can be discretized in a Bubnov-Galerkin manner based on (113). In this case, it reads:
δθh (ξ) = δΘ1h (ξ)g1 +
nΛ
X
S(r0 )δr0
,
δΘ
(ξ)
=
δΘ
(ξ)+δϕ
(ξ),
δϕ
(ξ)
=
Li (ξ)δϕi .
1h
M
1
h
h
ϕ
||r0 ||2
i=1
(117)
Θ1 and δd̂, i.e. via assemblies of
The spin vector δθh (ξ) in (117) shall be completely expressed via nodal variations δΘ̂
nodal primary variables. Thereto, δΘ Mϕ 1 (ξ) and δϕi are expressed as
δΘ Mϕ 1 (ξ) = δΘ1I +
δϕ
i
gT1 (g1I ×δg1I )−g1I T (g1 ×δg1 )
1+gT1 g1I
,
gi T (gI ×δg1I )−g1I T (gi1 ×δgi1 )
= δΘi1 −δΘiMϕ 1 = δΘi1 −δΘ1I − 1 1
,
1+gi1T g1I
31
(118)
which directly follows from interpolation (113). Inserting the relations (118) into the spin vector (117) yields:
nΛ
X
giT(gI ×δg1I )−g1IT(gi1 ×δgi1 ) gT1(g1I ×δg1I )−g1IT(g1 ×δg1 )
Θ 1 − Li 1 1
.
δΘ1h (ξ) = Lk δΘ̂
+
I
1+gT1 g1I
1+giT
1 g1
i=1
|
{z
}
(119)
=:δΘ1,di f f (ξ)
In (119), the notion δΘ1,di f f (ξ) has been introduced for the term distinguishing the Petrov-Galerkin (116) and the
Bubnov-Galerkin variant (117). By making use of the abbreviations xb = x(ξb ), xa = x(ξa ) and the relations
gT1a (g1b ×δg1b ) = gT1a (t̃b ×δr0b ) = (g1a × t̃b )T H0 (ξb )δd̂,
t̃ :=
g1
r0
=
,
||r0 || ||r0 ||2
(120)
the Bubnov-Galerkin interpolation of the spin vector given in equation (117) can finally be formulated as:
T
T
T
Θ1vθkΘ +δd̂ vθ⊥ +δd̂ vθkd vθkΘ = LTk ⊗gT1 , vθ⊥ = −H0T S(t̃),
δθhT (ξ) = δΘ̂
nΛ
X i
H0T (ξ)(g1I × t̃) − H0T (ξI )(g1 × t̃I )
,
vθkd =
L v1i − v1 ⊗gT1 , v1 =
1+g1T g1I
i=1
v1i = v1 (ξi ) =
H0T (ξi )(g1I × t̃i ) − H0T (ξI )(gi1 × t̃I )
1+gi1T g1I
(121)
.
Comparing equations (116) and (121) leads to the conclusion that the difference between the Bubnov-Galerkin and the
Petrov-Galerkin variant is expressed by one additional term involving vθkd . The arc-length derivative of (121) reads:
T
T
T
Θ1 v0θkΘ+δd̂ v0θ⊥+δd̂ v0θkd ,
δθh0T (ξ) = δΘ̂
00T
0T
T
0
0
0T
0
T
v0θkΘ = L0T
k ⊗g1 + Lk ⊗g1 , vθ⊥ = −H S(t̃)− H S(t̃ ), t̃ =
v0θkd =
nΛ
X
nΛ
X
Li v1i − v1 ⊗g0T
Li0 v1i − v01 ⊗gT1 +
1 ,
(122)
i=1
i=1
v01 =
r00
2(r0T r00 )r0
−
0
2
||r ||
||r0 ||4
H0T (ξ)(g1I × t̃0 )+ H00T (ξ)(g1I × t̃)− H0T (ξI )(g01 × t̃I )
1+gT1g1I
−
I
(g0T
1 g1 )v1
1+g1T g1I
.
It shall be investigated if the variants (116) and (117) can represent a constant distribution δθhT (ξ) = w0 = const. and
δθh0T (ξ) = 0, as it is the case for their counterparts (107) and (109). Thereto, the nodal variations are chosen as
δd̂ j = w0 × d̂ j ,
δt̂ j = w0 × t̂ j ,
δΘ̂i1 = gT1 (ξi )w0 for j = 1, 2; i = 1, ..., nΛ .
(123)
Inserting δd̂ j and δt̂ j according to (123) into (100) gives the desired result δrh (ξ) = w0 ×rh (ξ). With this result and by
inserting (123) into the spin vector interpolations (116) and (117), which both fulfill the interpolation property for the
tangential spin vector components δΘ̂i1 , it can be shown that w0 is at least represented correctly at the element nodes:
δθh (ξi ) = (gT1 (ξi )w0 )g1 (ξi ) +
r0 (ξi )×(w0 ×r0 (ξi ))
= w0 for i = 1, ..., nΛ .
||r0 (ξi )||2
(124)
Here, the Grassmann identity for cross-products a×(b×c) = (aT c)b−(aT b)c with a, b, c ∈ <3 as well as g1 = r0 /||r0 || have
been employed. Next, it will be investigated for the variants (116) and (117), if the choice (123) leads to a constant
spin vector field along the beam element. Inserting (123) into (116) and using the Grassmann identity yields:
nΛ
X i
r0 (ξ)×(w0 ×r0 (ξ))
i
T
δθh (ξ) =
L (ξ)g1 (ξ )w0 g1 (ξ) +
||r0 (ξ)||2
i=1
"
!#
nΛ
X
(125)
= I3 − g1 (ξ) ⊗ gT1 (ξ) −
Li (ξ)gT1 (ξi ) w0 , w0 .
i=1
|
{z
,0
32
}
Thus, the Petrov-Galerkin variant (116) cannot represent a constant spin vector field. Since (116) and (117) solely
differ in δΘ1,di f f (ξ) (see (119)), only this term has to be further investigated. Inserting (123) in δΘ1,di f f (ξ) yields:
nΛ
nΛ
X
X
δΘ1,di f f (ξ) = − Li (ξ)(gT1(ξi )w0 )+ Li (ξ)(gT1(ξ I )w0 )+(gT1 (ξ)w0 )−(gT1(ξ I )w0 )
i=1
|i=1{z }
(126)
=1
= gT1 (ξ) −
nΛ
X
!
Li (ξ)gT1 (ξi ) w0 .
i=1
Thus, adding the term δΘ1,di f f (ξ)g1 (ξ) to the Petrov-Galerkin variant (125) yields the desired result δθh (ξ) = w0 = const.
for the Bubnov-Galerkin spin vector interpolation in case the nodal variations are given by (123). Alternatively, this
result can be obtained by considering that (117) represents the consistent variation of the objective triad interpolation (113) (see also Section 6.3.2). Since this interpolation is objective, the variation of the discrete internal energy
has to vanish for infinitesimal rigid body rotations. For an arbitrary stress resultant m, this is only possible if δθh0 (ξ) ≡ 0
and consequently δθh (ξ) ≡ const. can be displayed exactly (see also the weak form (81)). Recapitulatory, the interpolation (117) can represent arbitrary constant spin vector distributions w0 , while for the interpolation (116) this is
only possible for 2D problems or in the special case w0 = 0. This result will be important in order to investigate the
conservation properties of the resulting finite element formulations (see e.g. Sections 6.3.5 and 9.4). Finally, the field
of (multiplicative) rotation vector increments ∆θh (ξ) and the derivative ∆θh0 (ξ), required for a consistent linearization
of the discretized weak form, follow from equations (121) and (122) by simply replacing the variations δ(.) by increments ∆(.). The spin vector considered above is expressed via multiplicative nodal increments δΘ̂i1 as nodal primary
variables. In case a rotation parametrization of the nodal triads via nodal rotation vectors according to Section 6.2.1
is employed, the nodal vector of multiplicative iterative rotation increments given by ∆Θ̂i = (∆Θ̂i1 , ∆Θ̂i2 , ∆Θ̂i3 )Tgi can
be directly used for triad update as shown in (101). However, if a rotation parametrization of the nodal triads via the
SR mapping and nodal relative angles ϕ̂i according to Section 6.2.2 is employed, the rotation vector increments shall
be expressed by means of additive increments ∆ϕ̂i of the nodal primary variables ϕ̂i as shown in (103). A relation
between ∆Θ̂i1 and ∆ϕ̂i can easily be derived on the basis of equations (19) and (20):
∆Θ̂i1 = ∆ΘiMϕ̂ 1 + ∆ϕ̂i = −
i
i
ḡiT
1 S(g1 ) ∆t
+ ∆ϕ̂i ,
iT i ||ti ||
1 + g1 ḡ1
∆ti = ∆r0 (ξi ) = H0 (ξi )∆d̂.
(127)
At the element boundary nodes, the last term in (127) can be simplified: ∆t j = ∆r0 (ξ j ) = H0 (ξ j )∆d̂ = ∆t̂ j if j = 1 or j = 2.
Review: The triad interpolation scheme presented in this section is very similar to the approach proposed
in the authors’ earlier contributions [104] (see Section 3.5.2). There, an intermediate triad field has been
constructed in a manner similar to (113), but with the choice Λr = Λ1Mϕ̂ . While the most essential properties
of these two approaches are comparable, there are some slight advantages of the procedure presented
here: Choosing a material triad Λi as reference triad Λr makes the interpolation scheme independent
from the choice of the nodal primary variables (according to Section 6.2.1 or according to Section 6.2.2).
Furthermore, locating the reference triad at the element middle node makes the element formulation
symmetric and extends the maximal orientation difference of the material triads at the element boundary
nodes that can be represented from 180◦ to 360◦ . The latter property results from the maximal orientation
difference of 180◦ allowed for two tangent vectors in order to yield a unique SR mapping (see also
Section 2.2). Besides the authors’ earlier work [104], also the straight and anisotropic Kirchhoff-Love
formulations from the literature exhibit the mentioned lack of symmetry (see requirement 4e) in Table 2).
33
Review: In [104], an alternative triad interpolation (see Section 3.5.1 of [104]) has been investigated,
which defines an intermediate triad field Λ Mϕ (ξ) directly via the smallest rotation from the intermediate
triad field Λ̄ Mϕ (ξ) of the last time step onto the current tangent vector field g1 (ξ) according to:
Λh (ξ) = exp(S[ϕh (ξ)g1 (ξ)]) Λ Mϕ (ξ),
nΛ
X
ϕh (ξ) =
Li (ξ)ϕ̂i ,
Λ Mϕ (ξ) = sr(Λ̄ Mϕ (ξ), g1 (ξ)),
Λ̄ Mϕ ,n+1 (ξ) = Λ Mϕ ,n (ξ).
(128)
i=1
At first glance, this interpolation seems to be more straightforward than (113) since no nodal triads are
required for constructing the intermediate triad field. However, as shown in [104], an interpolation of this
kind is neither objective nor path-independent. This non-objective and path-dependent interpolation has
e.g. been applied by the anisotropic formulations from the literature (see requirement 3) of Table 2).
6.3. Requirements on spatial discretization methods
In this section, essential requirements on the spatial discretizations of translational and rotational fields will be stated.
Subsequently, different beam elements will be presented and the fulfillment of these requirements investigated.
6.3.1. Differentiability of discrete fields
The first requirement for spatial discretization methods concerns differentiability. On the one hand, this requirement is
related to the weak form of the balance equations: The highest arc-length derivative occurring in the weak form of the
Simo-Reissner beam theory is of order one, leading to the requirement of at least C 0 -continuous discrete centerline
and triad fields. Such a continuity at the element boundaries is provided by the Lagrange centerline interpolation (95)
as well as by the two discussed approaches of triad interpolation according to Sections 6.2.3 and 6.2.4. The second
arc-length derivative of the beam centerline in the weak form of the balance equations is a distinctive property of the
Kirchhoff-Love beam theory and requires the interpolation of the centerline to be at least C 1 -continuous as guaranteed
by (97). On the other hand, certain applications such as beam-to-beam contact formulations (see e.g. [106, 107])
benefit considerably from a smooth geometry representation and the existence of a well-defined tangent vector field
along the entire beam centerline, which can conveniently be furnished by the Hermite interpolation (97).
6.3.2. Objectivity and path-independence
The properties of objectivity and path-independence play a central role in the development of (geometrically exact)
beam finite element formulations. The importance of these properties can be traced back to the nonlinear nature of the
configuration space (resulting from the occurrence of large rotations) which complicates the interpolation of rotational
quantities. Furthermore, it can be explained by the historic background that none of the early geometrically exact
beam formulations fulfilled both of these properties (see [43]). As already explained in [43], the path-independence
of the employed discretizations can directly be concluded from the fact that none of these interpolation schemes is
based on any history values of interpolated quantities. Only the nodal primary variables depend on history values.
However, the corresponding nodal displacements can always arise in a way such that the finite element solution is
independent from the actual load path in case the considered physical problem is path-independent. In other words,
the arising nodal displacements yield the path-independent solution to the discrete optimization problem (based on
a proberly defined Lagrangian) which is associated with the path-independent physical problem. In the numerical
investigations performed at the end of this chapter, this property will be verified. However, throughout this section,
the fundamental property of objectivity, i.e. the invariance of the applied deformation measures under rigid body
motions, will be investigated. Thereto, a rigid body translation rR and a rigid body rotation ΛR are superimposed onto
the beam centerline curve r(ξ) and the triad field Λ(ξ). A rigid body motion is characterized by constant fields rR and
ΛR along the beam, thus r0R ≡ 0 and Λ0R ≡ 0. In the following, the subscript (.)∗ denotes quantities that result from the
superimposed rigid body motion, thus r∗ (ξ) = ΛR (r(ξ)+rR ) and Λ∗ (ξ) = ΛR Λ(ξ). A formulation is denoted as being
objective if such a rigid body motion does not affect the material deformation measures Ω and Γ. It is straightforward
to show that the space-continuous versions of the deformation measures Ω and Γ are objective, i.e. Ω∗ = Ω and
Γ∗ = Γ (see e.g. [43]). The question of interest is if this objectivity is preserved by the employed spatial discretization
34
schemes. In [43], it is shown that the fulfillment of the following requirement guarantees for the objectivity of a
geometrically exact beam formulation based on the Simo-Reissner theory as introduced in Section 3:
r∗h (ξ)=Λ
˙ R (rh (ξ)+rR ) ,
Λ∗h (ξ)=Λ
˙ R Λh (ξ) → Ω∗h = Ωh ,
Γ∗h = Γh .
(129)
The following investigations will exclusively be applied to discretized quantities. In order to shorten notation, the
subscript (.)h will be omitted throughout this section. First, it shall be shown that the validity of (129) is also sufficient
for the invariance of the deformation measures (73) and (74) of the Kirchhoff beam theory. If (129) is valid, it follows:
r∗0 (ξ) = ΛR r0 (ξ),
r∗00 (ξ) = ΛR r00 (ξ),
g∗i (ξ) = ΛR gi (ξ),
0
g∗0
i (ξ) = ΛR gi (ξ).
(130)
Based on (130), the axial tension, the Frenet-Serret curvature vector as well as the total torsion read:
0.5
||r∗0 || = r0T ΛTR ΛR r0 = ||r0 ||
→
∗ = ||r∗0 || − 1 = ||r0 || − 1 = ,
r∗0 × r∗00 ΛR (r0 × r00 )
∗
T T
T
=
= ΛR κ → g∗T
i κ = gi ΛR ΛR κ = gi κ,
||r∗0 ||2
||r0 ||2
∗
∗
0T T
0T
0
K1∗ = K M1
+ ϕ∗0 = g∗0T
2 g3 = g2 ΛR ΛR g3 = g2 g3 = K M1 + ϕ = K1 .
κ∗ =
(131)
The identities Ω∗ = Ω and Γ∗ = Γ for the original and the rotated deformation measures (73) and (74) are a direct consequence of (131). Thus, also for Kirchhoff beam elements, the requirements (129) are sufficient in order to ensure
objectivity. In the following, the validity of (129) will be investigated for the interpolations in Sections 6.1 and 6.2.
6.3.2.1. Objectivity of centerline interpolations. Due to the linear dependence of the centerline interpolations (95)
or (97) on the nodal vectors, the proof of the first part of (129) is independent from the intermediate triad field:
r∗ (ξ) = ΛR (r(ξ)+rR ) for d̂∗i = ΛR d̂i +rR and t̂∗i = ΛR d̂i .
(132)
6.3.2.2. Objectivity of triad interpolation based on local rotation vectors. The fulfillment of (129), i.e. the objectivity of interpolation (105) has been shown in the original work [43]. The interested reader is referred to this reference.
6.3.2.3. Objectivity of triad interpolation based on ”Smallest Rotation” mapping. Based on the relations (130)-(132)
and the strong fulfillment of the Kirchhoff constraint, the base vector g∗1 (ξ) resulting from the rigid body rotation yields:
g∗1 (ξ) =
ΛR r0 (ξ)
r∗0 (ξ)
= 0
= ΛR g1 (ξ).
∗0
||r (ξ)||
||r (ξ)||
(133)
In a next step, the nodal primary variables are chosen such that the nodal triads are also rigidly rotated:
Λi∗ = ΛR Λi .
(134)
Using (133), (134) and (15), the following relation between the vectors g Mϕ 2 (ξ) and g Mϕ 3 (ξ) of the intermediate triad
Λ Mϕ (ξ) (see (113)) and their counterparts g∗Mϕ 2 (ξ) and g∗Mϕ 3 (ξ) resulting from the rigid body motion can be derived:
g∗Mϕ i (ξ) = g∗ri −
=
∗
g∗T
ri g1 (ξ)
g∗1 (ξ)+g∗r1
∗T ∗
1+gr1 g1 (ξ)
gT ΛT ΛR g1 (ξ)
(ΛR g1 (ξ)+ΛR gr1 ) = ΛR g Mϕ i (ξ),
ΛR gri − ri T R T
1+gr1 ΛR ΛR g1 (ξ)
(135)
i = 2, 3.
From (133) and (135), it can be concluded that the intermediate triad field is rigidly rotated:
Λ∗Mϕ (ξ) = ΛR Λ Mϕ (ξ).
35
(136)
If the transformation property Λexp(S[a])ΛT =exp(ΛS[a]) ∀ Λ ∈ SO(3), a ∈ <3 together with (133), (134) and (136) is
considered, the following result can be derived from the fourth equation of the triad interpolation scheme (113):
i∗
i
i
exp(S[ϕi∗ gi∗
˙ i∗ Λ∗T
Mϕ (ξ ) =
1 ]) = exp(S[ϕ ΛR g1 ])=Λ
ΛR Λi ΛTMϕ (ξi )ΛTR = ΛR exp(S[ϕi gi1 ])ΛTR = exp(S[ϕi ΛR gi1 ]) → ϕi∗ = ϕi .
(137)
Thus, the interpolation ϕ(ξ) in (113) is unchanged by the rigid body motion, i.e. ϕ∗ (ξ) = ϕ(ξ). Together with equations (133) and (136), the desired result for Λ∗ (ξ) as already stated in (129) can be derived:
Λ∗ (ξ) = exp(S[ϕ∗ (ξ)g∗1 (ξ)]) Λ∗Mϕ (ξ) = exp(S[ϕ(ξ)ΛR g1 (ξ)])ΛR Λ Mϕ (ξ)
=ΛR exp(S[ϕ(ξ)g1 (ξ)])ΛTR ΛR Λ Mϕ (ξ) = ΛR exp(S[ϕ(ξ)g1 (ξ)])Λ Mϕ (ξ) = ΛR Λ(ξ).
(138)
In the reformulations made in (138), again use has been made of the transformation property from above and (113).
According to (138), the triad interpolation proposed in Section 6.2.4 fulfills the requirement of objectivity. As mentioned above and derived in the third line of (131), the fulfillment of (129) guarantees for objective deformation
measures, provided that these are consistently derived from the triad interpolation. In order to verify this latter restriction, the two individual contributions K M1 and ϕ0 appearing in the third line of (131) shall be subject to a closer
investigation. Besides the relation ϕ∗ (ξ) = ϕ(ξ) → ϕ∗0 (ξ) = ϕ0 (ξ), which has already been deduced above, also the
∗
torsion K M1
of the intermediate system can be calculated for the configuration resulting from the rigid body motion:
∗
KM
=−
ϕ1
κ∗Tg1I∗
=−
I∗
1+g∗T
1 g1
κT ΛTR ΛR g1I
1+gT1 ΛTR ΛR g1I
=−
κTg1I
1+gT1 g1I
= K Mϕ 1 .
(139)
∗
As expected, K Mϕ 1 is not affected by the rigid body motion and the torsion K1∗ = K M1
+ϕ∗0 = K M1 +ϕ0 = K1 remains
unchanged. This underlines the objectivity of the interpolation (113) and the consistency of the torsion measure.
6.3.3. Avoidance of locking effects
It is well-known that purely displacement-based finite elements are prone to locking. Locking effects particularly relevant for geometrically exact beam formulations are shear locking as well as membrane locking. While shear locking
can - by definition - only appear in beam formulations of Simo-Reissner type, membrane locking has already been
observed for geometrically linear Kirchhoff beams (see e.g. [5] or [6]). In general, membrane locking refers to the
inability of elements to exactly reproduce inextensibility, viz. a vanishing axial strain ≡ 0, for curved structures
such as shells or beams. This behavior can be traced back to a coupling between the kinematic quantities describing
the axial tension mode and the curved geometry. While the focus of the subsequent investigations lies on membrane
locking, at least some remarks concerning shear locking will be made at the end of this section.
6.3.3.1. Characterization of locking. One possible definition of locking is the deterioration of the spatial convergence
rate in dependence of a certain key parameter. Subsequent investigations will reveal that the element slenderness ratio
ζele := lele /R plays the role of such a key parameter associated with the membrane locking effect of slender beams.
One question of interest is how the liability of a finite element formulation to locking can be assessed in a quantitative
manner. From a mathematical point of view, this question can be answered by investigating the stability of the finite
element formulation. For example, for mixed finite element formulations, a stability criterion is given by the wellknown Ladyshenskaya-Babuska-Brezzi (LBB) condition, also denoted as inf-sup condition (see [29, 30]). Since a
direct and general analysis of such conditions can often be intricate, also numerical inf-sup tests have been suggested
in the literature (see e.g. [77, 12]). From a mechanical point of view, locking is typically explained by the occurrence
of parasitic stresses, viz. the occurrence of modes in the discrete solution that are not part of the analytical solution.
Consequently, the question if a formulation is prone to locking or not can also be answered by investigating proper
representative test cases for parasitic stresses. Besides these mathematical and mechanical interpretations of locking,
a third, namely a numerical perspective (see [89]) can be helpful. From a numerical point of view, locking can be seen
as the consequence of an over-constrained system of equations. As introduced in [69], the so-called constraint ratio
36
allows for some, at least heuristic, evaluation of the locking behavior of a finite element formulation. The constraint
ratio is defined as the ratio of the total number of equilibrium equations neq to the total number of constraints neq,c :
r :=
neq
.
neq,c
(140)
In order to analyze the locking behavior, the constraint ratio r of the continuous problem and the constraint ratio rh of
the discretized problem evaluated for an infinite number of elements have to be compared. The underlying proposition
is that elements with rh < r (especially rh < 1, which means that more constraints than degrees of freedom are present)
have a tendency to lock, whereas values rh > r of the constraint ratio indicate that not enough constraint equations
are available in order to reproduce the constraint in an accurate manner. Following this hypothesis, the case rh = r
has to be regarded as the optimal constraint ratio. Throughout this contribution, the relevant locking phenomena will
be analyzed based on a mechanical as well as a numerical perspective. The corresponding concepts are applicable
in a straightforward manner. In future work, the stability of the most promising element formulations also has to be
investigated in a mathematically rigorous manner, either based on direct analysis or on numerical stability tests.
6.3.3.2. Membrane locking. In [105], the effect of membrane locking in the context of geometrically exact Kirchhoff
beam elements based on a Hermite centerline interpolation according to Section 6.1.2 has been investigated in detail.
Here, only the main results shall be recaptulated. In [105], it has been found that the state of (exactly) vanishing axial
strains can only be represented for straight configurations but not for abritariy curved configurations. Furthermore,
the amount of parasitic axial strain energy occuring in states of pure pending has been shown to increase quadratically
with the beam element slenderness ratio leading to a progressively (over-) stiff system answer. The constraint ratios
of the space-continuous and discrete 2D Kirchhoff-Love beam problem have been determined to r = 2 > rh = 1,
indicating that the considered element formulation is prone to membrane locking. Different solution strategies exist
in order to get rid of these locking effects. In [105], the approach of Assumed Natural Strains (ANS, see e.g. [70]),
Reduced Integration (RI, see e.g. [109]) and an alternative procedure proposed in [105] and denoted as Minimally
Constrained Strains (MCS), have been compared. The proposed MCS scheme can be characterized as an assumed
strain/B-bar approach (see e.g. [133]), and it can be derived in a variationally consistent manner. Concretely, the
contribution G of the axial tension to the weak form of the mechanical equilibirum equations is replaced by G¯ :
G = δr0Tg1 EA ||r0 ||−1
|{z}
| {z }
¯ (ξ) =
G¯ = δ¯ EA¯ ,
δ
nCP
X
→
i
i
L (ξ)(ξ ),
i=1
δ¯ (ξ) =
nCP
X
k=1
i
i
L (ξ)δ(ξ ),
1
Π̃int, = EA¯ 2 .
2
(141)
In (141), ξi denotes the parameter coordinate and nCP the number of Collocation Points (CPs) where the ”original“
strains are evaluated. Moreover, Li (ξ) are Lagrange shape functions of polynomial order nCP −1. Linearizing (141) at
the undeformed, straight configuration C0 , yields the B-bar structure typical for geometrically linear finite elements:
nCP
X
T
T T
L(G )|C0 := δd̂ H0Tt0 K tT0 H0 ∆d̂, L(G¯ )|C0 = δd̂ B̄ K B̄∆d̂, B̄(ξ) = Li(ξ)B(ξi ).
|{z} |{z}
BT
i=1
B
In the last equation, the additional abbreviation K := EA has been introduced in order to shorten the notation.
Remark: Actually, the residual vector stated in (161) consists of functional expressions that cannot
be integrated exactly by Gauss integration. However, in numerical simulations, the deviations in the
results for a number of Gauss points nG > 4 turned out to be very small as compared to the discretization
error. Therefore, the notion ”exact integration“ will be used whenever four or more Gauss points are applied.
37
(142)
Remark: In the last section, it has been shown that the strain field (ξ) is objective and path-independent,
meaning that the strain field does not change as a consequence of a rigid body motion and that its value
at a certain configuration is independent from the deformation path leading to that configuration. Since
the assumed strain field ¯ (ξ) represents a pure re-interpolation of the original strain field (ξi ) evaluated at
fixed collocation points ξi , the former will also fulfill objectivity and path-independence.
In [105], the number and location of CPs has been chosen according to nCP = 1 and ξ1 = −1, ξ2 = 1, ξ3 = 0. In order to
motivate this choice, the constraint ratio resulting from the MCS method and different sets of CPs as well as for the
alternative methods RI and ANS will be presented in the following. In case of a strain re-interpolation such as given
by (141), it is sufficient that the constraint of vanishing axial strains is fulfilled at the collocation points in order to
end up with a vanishing axial strain energy. Since the Hermite interpolation (97) provides a C 1 -continuous centerline
curve, the first derivative r0 and consequently also the axial tension is C 0 -continuous at the element boundaries.
Thus, exactly one constraint equation ||r0 (ξi )||=1
˙ results from each (interior and element boundary) CP.
Remark: The chosen CPs are motivated by strain-continuity: If the axial strains were not C 0 -continuous at
the element boundaries, each element boundary node would provide two constraint equations, one for the
previous element (e) and one for the subsequent element (e+1), i.e. ||r0 (ξ(e) = 1)||=1
˙ and ||r0 (ξ(e+1) = −1)||=1.
˙
In the following table, the constraint ratios resulting from the three methods ANS, RI and MCS are compared for
different choices concerning number and locations of collocation or Gauss points. As shown in [105], the ANS
approach has to be based on CPs with vanishing axial strain values for a pure bending state, which requires four CPs.
neq,c
rh /r
MCS-3b
2nele +1
1
MCS-2b
nele +1
2
MCS-3i
3nele
2/3
MCS-2i
2nele
1
ANS-4i
4nele
0.5
RI-3i
3nele
2/3
RI-2i
2nele
1
Table 3: Quantitative comparison of different ”anti-locking” methods.
The index i in Table 3 indicates that all collocation or Gauss points lie in the elements interior while variants that are
marked by an index b also employ the element boundary nodes ξ = ±1. According to the statements made above, the
variants with rh /r = 1 represent the constraints associated with the axial tension in an optimal manner, whereas variants
with rh /r < 1 have a tendency to lock. Consequently, at first glance, the variants MCS-3b, MCS-2i and RI-2i seem
to be equally suitable. However, as derived in [105], the variants MCS-2i and RI-2i will lead to an underconstrained
system of equations, allowing for zero-energy modes and yielding a singular system of equations and a rank deficient stiffness matrix in the straight configuration, while the variant MCS-3b exactly provides the minimal number of
constraint equations required in order to avoid such zero-energy modes in the straight configuration. For that reason,
this method has been denoted as method of Minimally Constrained Strains. Similarly, an optimal constraint ratio of
rh = r = 4 and the avoidance of zero-energy modes can be verified for this choice CPs in the general 3D case.
Remark: Of course, locking would be avoided if the beam elements could exactly represent the internal
energy associated with pure bending. According to [105], the MCS method fulfills this requirement only for
straight configurations while arbitrary curved configurations will yield a slightly over-constrained system
of equations. Thus, the state of constant curvature and vanishing axial tension cannot be displayed exactly.
However, it has been shown numerically that membrane locking is avoided. This result is reasonable, since
the variant MCS-3i still fulfills the optimal constraint ratio r/rh = 1 for arbitrarily curved configurations.
Remark: As an alternative to the variant MCS-3i, also a reduced Gauss-Lobatto integration scheme with
integration points ξ1 = −1, ξ2 = 1 and ξ3 = 0 could be applied to the axial tension term of the weak form,
yielding the same number of constraint equations neq,c = 2nele +1 as the MCS approach. However, within
this contribution, the MCS method will be preferred due to its arguably more consistent variational basis
and the uniform integration scheme resulting for the individual work contributions of the weak form.
All in all, it can be concluded that the choice of the element boundary nodes and the element midpoint as CPs of the
MCS method leads to the minimal possible number of constraint equations and to an optimal constraint ratio. Con38
sequently, a successful avoidance of locking effects could be confirmed for this approach. A similar effectiveness can
be expected for a reduced Gauss-Lobatto integration of the axial tension terms with integration points ξ1 = −1, ξ2 = 1
and ξ3 = 0. On the contrary, the minimal number of three integration points possible for a reduced Gauss-Legendre
integration scheme leads to an increased constraint ratio of rh = 3 > r and consequently to a suboptimal locking behavior. While the MCS and RI methodologies aim at a reduction of the number of constraint equations, the functional
principle of the ANS approach is different: There, parameter coordinates have to be determined where the constraint
is already correctly fulfilled by the original element formulation (when applying the latter to a representative test
case). These parameter coordinates are typically chosen as CPs for the ANS approach. Such a procedure can avoid
locking effects in a manner that is independent of the number of constraint equations. The drawback of the ANS
method is that the positions of these points may change for general deformation states in the geometrically nonlinear
regime, which might considerably deteriorate the effectiveness of this approach. All these theoretical considerations
recommend the proposed MCS approach as method of choice in order to avoid locking effects for the Kirchhoff beam
elements considered in this contribution. The numerical results presented in [105] have confirmed this prediction.
Review: In all the three categories isotropic, straight and anisotropic large-deformation geometrically exact
Kirchhoff-Love finite elements available in the literature (see requirement 4b) in Table 2), the consequences
of membrane locking have been observed. However, a rigorous treatment seems to be missing in these
works. For example, in [151], oscillations of the membrane forces have been observed and cured by means
of a special force averaging procedure. However, this procedure seems to be rather a post-processing step
than an invasion in the actual finite element formulation that would also improve the final displacement
solution. In [154], the axial tension term of the weak form is replaced by an averaged element-wise constant
approximation of (ξ) in order to be able of exactly representing the solution of a straight beam under axial
load. However, the influence of this procedure on membrane locking effects in curved configurations, i.e.
configurations that are actually relevant for membrane locking, has not been investigated. Finally, [64]
proposes a mixed finite element formulation in combination with a multi patch B-spline approach in
order to treat the locking effects. Interestingly, the third-order variant of these B-spline patches could be
identified as Hermite interpolation comparable to (97), which turned out to be favorable for the avoidance
of locking as compared to single B-spline patches due to the lower inter-element continuity enforced by
the former approach. However, the geometrical compatibility equations of the multi patch approach as
well as the considered numerical examples seem only to cover the geometrically linear regime. In contrast,
the numerical examples in [105] and Section 11 of this work confirm that this gap could be closed by
the proposed MCS method, which successfully cures membrane locking in 3D large-deformation problems.
6.3.3.3. Shear locking. The phenomenon of shear locking does not lie in the focus of the current work. Nevertheless,
the cause for such locking effects shall at least be briefly compared with the situation already discussed for membrane
locking. Shear locking denotes the inability of a finite element to exactly represent the state of vanishing shear strains.
This situation can again be illustrated by means of a pure bending example. It is assumed that the beam centerline is
discretized either by Lagrange polynomials according to (95) or by Hermite polynomials according to (97) and that
the triad interpolation is given by (105). It has already been stated that the triad interpolation (105) can exactly represent constant curvatures. On the other hand, it has been shown that the Hermite centerline representation (and also its
counterpart based on Lagrange polynomials) cannot exactly display the state of constant curvature. This means, the
state r0T(ξ)g2 (ξ) ≡ r0T(ξ)g3 (ξ) ≡ 0 cannot be displayed exactly in combination with a constant curvature. In other words,
the interpolation spaces applied to the translation (i.e. to the beam centerline) and to the rotation field (i.e. to the triad
field) do not optimally match in the sense that a state of constant curvature and vanishing shear deformation cannot
be represented exactly. Similar to membrane locking, the ratio of the shear stiffness to the bending stiffness increases
quadratically with the beam element slenderness ratio. Thus, again the element slenderness ratio represents the key
parameter for this locking effect. Further 2D investigations on membrane and shear locking effects in geometrically
linear and nonlinear shear-deformable beam element formulations can for example be found in [73].
39
Review: At the end of this section, also a view considerations shall be made concerning element patch tests
(requirement 4d) in Table 2). Such a patch test for geometrically exact beams could for example require the
exact representability of a state of constant 3D curvature (bending and torsion) and constant axial tension
(as well as constant shear-deformation in case of Simo-Reissner formulations). The beam centerline curve
associated with such a configuration is given by a helix with constant slope as investigated in the numerical
example of Section 11.3. In the following, it will be distinguished, if only the strain energy associated with
such a state can exactly be represented or if additionally also the beam geometry in terms of centerline and
triad field of such a state can exactly be represented. In Sections 9 and 11.3, it will be verified theoretically
and numerically that the SK element formulations based on strong enforcement of the Kirchhoff constraint
and proposed in Section 9 can in general neither represent the strain energy nor the beam geometry
associated with such a state. Due to the identical measure κ of centerline curvature and the comparable
centerline discretizations employed by the isotropic, straight and anisotropic Kirchhoff-Love formulations
from the literature, it can directly be concluded that also none of these formulations can fulfill this patch
test, neither in the energetic nor in the geometrical sense (see again Table 2). On the other hand, it will be
verified theoretically and numerically that the WK element formulations based on weak enforcement of the
Kirchhoff constraint (see Section 10) as well as the Simo-Reissner formulation proposed by Crisfield and
Jelenić [43, 80] (denoted as CJ element in Section 7) can represent the exact energy state of the patch test
due to the employed geodesic triad interpolation (and an appropriate collocation scheme such as MCS or
reduced integration for coupling the centerline and the triad field), but not the exact geometry.
There exist a few very recent 3D Simo-Reissner beam element formulations that can exactly represent such
a patch case in the energetic but also in the geometrical sense. Unfortunately, an explicite analytic representation of the discretization underlying these formulations in form of a helicoidal interpolation applied to
the triad and to the centerline field is only possible for first-order interpolations [145, 139, 138]. As emphasized in [99], such a helicoidal scheme can be identified as one geodesic interpolations on the semidirect
product manifold SE(3) := <3 o SO(3), denoted as Special Euclidean group, while e.g. the formulation
of Crisfield and Jelenić [43, 80] consists of two individual geodesic interpolations, a linear one with constant slope on the Euclidean vector space <3 as well as a spherical interpolation with constant curvature
on the Special Orthogonal group SO(3), together composing a interpolation scheme on the direct product
manifold <3 × SO(3). An extension of these helicoidal interpolations to higher-order schemes is e.g. given
by [156, 157]. There, a closed-form representation of the interpolation has not been possible anymore and
the triad and centerline field have been generated out of the strain field in an implicit manner via numerical
integration. In contrast to Simo-Reissner formulations, Kirchhoff-Love element formulations should typically be of higher-order in order to guarantee for C 1 -continuity and reasonable convergence orders (see also
Section 6.3.4). It is not clear if the increased numerical effort required for implicitly defining higher-order
helicoidal fields via numerical integration could overcompensate possible advantages in terms of higher approximation quality when developing comparable Kirchhoff-Love element formulations that exactly fulfill
the patch test in the energetic and geometrical sense. At least, it seems to be a promising direction of future
research to compare formulations of this kind with the Kirchhoff-Love elements proposed here.
6.3.4. Optimal convergence orders
In order to compare the convergence behavior of different finite element formulations, a well-defined error measure is
required. Thereto, the following relative L2 -error ||e||2rel will be considered in the numerical examples of Section 11:
s
||e||2rel
=
1
umax
1
l
Z
l
||rh − rre f ||2 ds.
(143)
0
In (143), rh denotes the numerical solution of the beam centerline for a certain discretization. For all examples without
analytic solution, the standard choice for the reference solution rre f is a numerical solution via the WK-TAN element
(see Section 10.2) employing a spatial discretization that is by a factor of four finer than the finest discretization shown
in the convergence plot. The normalization with the element length l makes the error independent of the length of
the considered beam. The second normalization leads to a more convenient relative error measure, which relates the
40
L2 -error to the maximal displacement umax of the load case. For some examples, also the relative energy error
||e||e,rel =
Πint,h − Πint,re f
,
Πint,re f
(144)
will be considered. Here, Πint,h and Πint,re f represent the stored energy functions (see e.g. 75) associated with a certain
discretization and with the reference solution. Before the convergence plots of selected numerical examples will
be considered, the optimal convergence rates in the norms (143) and (144) expected for the different beam element
formulations shall be briefly discussed. The convergence in the energy error (which is minimized by the finite element
method) is dominated by the highest derivative m of the primary variable fields r(ξ) and Λ(ξ) occurring in the energy
(see e.g. 75) and consequently in the weak form. Since the employed hyperelastic stored energy functions represent
quadratic forms in the derivatives of these primary variable fields, the convergence rate in the energy error yields:
||e||e = O(h2(k−m+1) ).
(145)
Here, h := lele is the element length and k the polynomial degree completely represented by the trial functions. In
[142], the convergence rate of the L2 -error for the Ritz solution to a variational problem of order m is shown to be:
||e||2 = O(hk+1 + h2(k−m+1) ).
(146)
The second term in (146) represents the dependence of the L2 -error convergence rate on the energy error convergence,
reflecting the variational basis of the finite element method. The first term represents the pure polynomial approximation of the trial functions with respect to the considered primary variable field, e.g. r(ξ). In most cases, e.g. when
displacement-based solid elements are considered, the first exponent is smaller than the second one and dominates the
overall discretization error. For this reason, only the first term is considered by many authors. However, in the following, it will be shown that especially for Kirchhoff type beam elements also the second term of (146) is important.
Thereto, the expected convergence rates for Reissner and Kirchhoff type beam elements shall briefly be discussed.
6.3.4.1. Element formulations of Simo-Reissner type. The highest derivative of primary variable fields occurring in
the weak form associated with the Simo-Reissner beam problem is m = 1. Thus, for the third-order (k = 3) Reissner
beam elements considered in this work, a convergence rate of six is expected in the energy error (144), while the
L2 -error is dominated by the first term of (146) leading to a corresponding optimal convergence rate of four.
6.3.4.2. Element formulations of Kirchhoff-Love type. The subsequently proposed Kirchhoff beam elements will lead
to the values k = 3 for the polynomial degree of the triad functions and m = 2 for the highest arc-length derivative in the
weak form. Consequently, the convergence of the energy error is of order four. Furthermore, the exponents of both
terms in (146) take on a value of four, also leading to an expected convergence rate of four for the L2 -error. Thus,
also the second, energy-related term has to be considered for Kirchhoff problems of this kind. For that reason, at least
polynomials of order three should be chosen as trial functions for Kirchhoff beam elements: Reducing the polynomial
degree from k = 3 to k = 2 would lead to an undesirable decline in the L2 -convergence rate from four to two. Thus,
the third-order Kirchhoff beam elements proposed in this work can be regarded as approximations of the lowest order
that is reasonable from a numerical point of view. However, the fact that both exponents of (146) take on the same
value for k = 3 also means that in some cases, usually for examples involving complex deformation states and strain
distributions, the second, energy-related term might determine the overall error-level. Thus, in order to fully exploit
the approximation power of the employed discretization, it can often be sensible to apply trial functions of increased
polynomial degree k > 3. In this case, the first term of (146) will determine the overall error level for sufficiently
fine discretizations (since the second term converges faster) and consequently the discretization error is exclusively
limited by the approximation power of the applied polynomial order. The extension of the proposed Kirchhoff beam
elements to Hermite interpolations of order k > 3 is possible in a straightforward manner and will be treated in future
research work. In the numerical example of Section 11.2, a first proof of concept will be given for such an extension.
41
Remark: Based on these considerations the question arises, which trial function orders kS R and kKL of
formulations based on the Simo-Reissner and the Kirchhoff-Love theory have to be chosen in numerical
examples in order to perform a reasonable comparison of their convergence behavior. The answer to this
question depends on the primary interest, which might either lie in the L2 - or in the energy convergence.
Here, the third-order Kirchhoff elements are compared to third-order Reissner elements, leading to equal
L2 - but to different energy error rates. Alternatively, one could compare the third-order Kirchhoff elements
with second-order Reissner elements, leading to equal energy error rates but to different L2 -error rates.
Review: Later in Section 9, a quadratic interpolation will be applied to the relative angle field ϕh (ξ) of the
interpolation scheme (113). Since the orientation of the material triad field is determined by the relative
angle ϕh (ξ) as well as the tangent vector field r0h (ξ), with the latter being a polynomial of order two, this
second-order interpolation is sufficient for triad field discretization. Even a higher polynomial degree for
ϕh (ξ) could not further improve the exact polynomial representation of rotational strains, which is of first
order. On the other hand, if ϕh (ξ) was only interpolated linearly, only the constant term of rotational strains
could be exactly represented and the second term in (146) would dominate the L2 -convergence leading to a
decline in the expected optimal order to two. In Meier et al. [104], these considerations have been confirmed
numerically. The Kirchhoff-Love formulations of straight type (see requirement 4a) in Table 2) also apply
only a first-order interpolation for the relative angle ϕh (ξ) and a comparable intermediate triad definition.
Thus, also for these formulations a decline in the optimal convergence order from for to two is expected.
6.3.5. Conservation properties
Since the finite element solution converges towards the corresponding analytic solution in the limit of fine spatial
discretizations, elementary properties of the analytic solution such as conservation of linear momentum (or rather
equilibrium of forces in statics), conservation of angular momentum (or rather equilibrium of moments in statics) as
well as conservation of energy (or rather balance of external and internal work for non-conservative problems) will
also be fulfilled by the numerical solution for lele → 0. However, often it is desirable to provide such properties
already for arbitrarily rough spatial discretizations. The question, if these properties of the space-continuous problem
are inherited by the spatially discretized problem, will later be investigated for the different beam element formulations
proposed in subsequent sections. Thereto, use will be made of the fact that the discretized weak form of the balance
equations is fulfilled for arbitrary values of the nodal primary variable variations. Choosing the nodal primary variable
variations such that the associated virtual motion represents a rigid body translation given by
(δrh (ξ), δθh (ξ)) ≡ (u0 , 0) ∈ Vh
with
u00 = 0,
(147)
allows to investigate the conservation of linear momentum. The special choice (147) leads to δr0h (ξ) ≡ δθh0 (ξ) ≡ 0 and
to vanishing contributions of the internal forces and moments in the discretized weak forms (33) and (81) associated
with the Simo-Reissner and the Kirchhoff-Love beam theory. Inserting (147) into these weak forms yields
l̇ = fext
Zl
with l :=
l̃ds, fext =
0
Zl
h i
f̃ds + fσ ,
(148)
Γσ
0
and to exact conservation of linear momentum l = const. for the unloaded system, viz. if fext = 0. Since reaction forces
at Dirichlet supports are also included in fext , (148) is equivalent to the equilibrium of forces in the static case, i.e. if
l = 0. Similarly, a choice of the nodal primary variable variations representing a rigid body rotation given by
(δrh (ξ), δθh (ξ)) ≡ (w0 × rh (ξ), w0 ) ∈ Vh
with w00 = 0,
(149)
allows to investigate the angular momentum. Relation (149) leads to δr0h (ξ) ≡ w0 ×r0h (ξ) as well as δθh0 (ξ) ≡ 0 and again
to vanishing contributions of the internal forces and moments in the discrete versions of the weak forms (33) and (81)
associated with the Simo-Reissner and the Kirchhoff-Love beam theory. Inserting (149) into these weak forms yields:
Zl
Zl
h
i
ḣ = mext , h := (h̃+r× l̃)ds, mext =
r × f̃ + m̃ ds+ r × fσ +mσ ,
Γσ
0
0
42
(150)
and to exact conservation of angular momentum h = const. for the unloaded system, viz. if mext = 0. Since possible
reaction moments at Dirichlet supports are also included in mext , relation (150) is equivalent to the equilibrium of
moments in the static case, i.e. if h = 0. Finally, a choice of the nodal primary variable variations according to
(δrh (ξ), δθh (ξ)) ≡ (ṙh (ξ), wh (ξ)) ∈ Vh ,
(151)
allows to investigate the mechanical power balance. Inserting (151) into the discrete versions of the weak forms (33)
˙ =
and (81) associated with the Simo-Reissner and the Kirchhoff-Love theory, and making use of the relations Π̃
int
˙
w0T m+(ṙ0 −w×r0 )T f and Π̃kin = wTcρ a (see Section 3.4) as well as wT S(w)cρ w = 0, yields the following relation
Π̇kin + Π̇int = Pext
with
Pext
#
"
#
Zl "
=
ṙT f̃ + wTm̃ ds + ṙTfσ + wTmσ ,
0
(152)
Γσ
and consequently to exact energy conservation Πkin + Πint = const. for the unloaded system, viz. if Pext = 0.
So far, it has been shown that exact conservation of linear momentum, angular momentum and energy (see equations (148), (150) and (153)) can be guaranteed for the spatially discretized (and time-continuous) problem, provided
the special choices (147), (149) and (151) for the translational and rotational variation fields are contained in the discrete weighting space Vh of the considered finite element formulation. In the following sections, the question if (147),
(149) and (151) can indeed be represented by the discrete weighting functions, will be investigated for the proposed
beam element formulations. Of course, also the time integration scheme influences the conservation properties of the
fully discrete system considerably. However, the investigation of this factor does not lie within the scope of this work.
In general, it can be stated that the three conservation properties considered above will be fulfilled by a finite element
formulation if i) the formulation is objective and path-independent and ii) the test functions are consistently derived in
a Bubnov-Galerkin manner. The first pre-requisite ensures that a unique potential exists in form of a discrete storedenergy function that is path-independent and invariant under rigid body motions. The second pre-requisite ensures
that the corresponding contribution to the weak form represents exact increments of this potential, which fulfills the
third requirement (153) per definition, but also the first and second requirement (148) and (150), since infinitesimal
rigid body translations and rotations according to (147) and (149) will lead to vanishing increments of this objective
potential. However, the opposite conclusion does obviously not hold true: Also Petrov-Galerkin formulations can
at least fulfill the first and second requirement (148) and (150) if the chosen test functions can represent rigid body
translations and rotations (147) and (149) (see e.g. CJ element of Section 7 or WK element of Section 10).
Review: Following the argumentation from above, it can be expected that the objective and pathindependent Bubnov-Galerkin FEM discretizations employed in the formulations of the straight type (see
requirement 4) of Table 2) will in general fulfill the above conservation properties as long as the variation
and discretization process required to derive the test functions is conducted in a consistent manner and no
additional approximations (e.g. small tension assumptions according to ||r|| ≈ 1 as discussed in [155]) are
applied that could spoil this consistency. The situation is slightly different for the anisotropic formulations:
The non-objective and path-dependent character of the employed discrete deformation measures might in
general lead to non-vanishing energy increments for infinitesimal rigid body translations and rotations according to (147) and (149). The exact representation of these non-vanishing energy increments resulting
from consistently derived Bubnov-Galerkin test functions may in this case exactly be the reason that the
conservation properties (148) and (150) will be spoiled. Eventually, also the existing Kirchhoff-Love formulations of the straight type in Table 2 shall be discussed. There, a standard polynomial interpolation
based on Lagrange shape functions is applied to the tangential spin vector component δΘ1 similar to the
Petrov-Galerkin variant (116). Consequently, as shown in equation (125), also these formulations cannot
represent constant spin vector distributions and exact conservation of angular momentum (150).
43
It can be shown that these straight formulations indeed represent Petrov-Galerkin formulations, which in
turn can also not guarantee for the conservation of energy (153). To verify this statement, the following
considerations are made: If the discrete spin vector field employed in [28] is consistent with the underlying
triad interpolation in a Bubnov-Galerkin sense is not straight-forward to be answered, since no explicit triad
interpolation scheme is given (and required) by the considered isotropic formulations. However, at least
the arc-length derivative of the discrete triad field is defined by the employed torsion interpolation, which
is based on Lagrange polynomials in [28] or [151]. The consistent Bubnov-Galerkin interpolation can then
be identified based on the second compatibility condition in (72) an reads:
δΘ01h (ξ) = δK1h (ξ)−δg1h (ξ)T(g1h (ξ)×g01h (ξ)),
(153)
Since K1h (ξ) has been interpolated via Lagrange polynomials, r0h (ξ) has been interpolated via Hermite polynomials and the relation g01h (ξ) = r0h (ξ)/||r0h (ξ)|| holds, a consistent Bubnov-Galerkin interpolation of the
tangential spin vector component δΘ1h (ξ) fulfilling (153) cannot be based on a pure polynomial interpolation as done in the mentioned references. Consequently, these interpolations are of Petrov-Galerkin type.
7. Simo-Reissner Beam Element
Throughout this work, the Reissner type beam element formulation proposed by Crisfield and Jelenić [43, 80], in
the following referred to as CJ element, will serve as reference formulation for numerical comparisons. In the next
section, the main constituents required to derive the element residual vector will be presented. In the subsequent
Sections 7.2 and 7.3, this element formulation will be investigated with respect to possible locking effects and the
fulfillment of mechanical conservation properties as introduced in Sections 6.3.3 and 6.3.5.
7.1. Element residual vector
In this section, the element residual vector of the CJ element will be derived from the general, space-continuous
Simo-Reissner beam problem. First, all trial and weighting functions are replaced by their discrete counterparts taken
from the finite-dimensional trial subspace (rh , Λh ) ∈ Uh ⊂ U and the weighting subspace (δrh , δθh ) ∈ Vh ⊂ V. In the
following, nn -noded finite elements with the vectors of nodal primary variables x̂CJ := (d̂1T, ψ̂1T, ..., d̂nn T, ψ̂nn T )T and
δx̂CJ := (δd̂1T, δθ̂1T, ..., δd̂nn T, δθ̂nn T )T are considered. The centerline interpolation is based on Lagrange polynomials
of order nn −1 according to Section (6.1.1), i.e. nr = nn . Furthermore, the rotation field interpolation follows equation (105) of Section 6.2.3 based on nΛ = nn nodes, in combination with a Petrov-Galerkin approach for the spin vector
discretization given by (109). In contrast to the original works [43, 80], here the modified generalized-α scheme of
Section 5 is employed for time integration, defining the velocities and accelerations required for the inertia forces fρ
and moments mρ as stated in (42). Inserting these interpolation schemes as presented in the previous sections into the
weak form of the balance equations (33) yields the element residual vector contributions rCJ,d̂ and rCJ,θ̂ according to
rCJ,d̂
Z1
h
i
=
L0T f − LT f̃ρ J(ξ)dξ − LT fσ ,
Γσ
−1
rCJ,θ̂ =
Z1
(154)
h
i
L0T m − LT S(t)f − LT m̃ρ J(ξ)dξ − LT mσ = 0.
Γσ
−1
The subscripts (.)d̂ and (.)θ̂ distinguish the residual vector contributions associated with the variations δd̂i and δθ̂i .
Within this work, a linearization ∆rCJ = kCJ ∆x̂CJ based on multiplicative rotation increments ∆θ̂i according to
∆x̂CJ := (∆d̂1T, ∆θ̂1T, ..., ∆d̂nn T, ∆θ̂nn T )T as given in [80] is employed. In dynamics, the element residual vector rCJ
and stiffness matrix kCJ slightly differ from the original work [80] due to the applied time integration scheme.
44
7.2. Avoidance of locking effects
In [43] and [80], the authors proposed a reduced Gauss integration scheme in order to avoid shear locking and membrane locking in the range of high beam slenderness ratios. Thereto, nn − 1 integration points have been employed
for the integration of the internal force contribution of a nn -noded element. The effectiveness of this procedure will
be verified in subsequent numerical examples and shall be briefly motivated by the following considerations: The
3D Simo-Reissner beam problem is based on neq = 6 differential equations (32) describing the beam problem and
pointwise neq,c = 3 constraint equations in order to represent the state of vanishing axial strains r0T g1 ≡ 1 and vanishing shear strains r0T g2 ≡ r0T g3 ≡ 1 prevalent in a pure bending problem. Consequently, the constraint ratio of the
space-continuous problem yields r = neq /neq,c = 2. As consequence of the reduced integration, the discrete number of
constraint equations takes on a value of neq,c = 3 · (nn −1) · nele . Given the total number of equations neq = 6(nn −1) · nele
after application of proper Dirichlet conditions, the constraint ratio of the discrete problem results in:
rh = lim
nele →∞
3 · (nn −1) · nele
= 2 = r.
6(nn −1) · nele
(155)
Relation (155) yields the optimal constraint ratio for this element formulation. Consequently, no locking effects are
expected. The investigations made so far can be further refined by realizing that the CJ element can exactly represent
the internal energy associated with a 3D pure bending state. In order to understand this statement, the internal energy
is split into contributions stemming from torsion and bending and into contributions stemming from axial tension and
shear deformation, i.e. Πint,h = Πint,Ω,h +Πint,Γ,h . For a pure bending state, the energy contribution Πint,Γ,h has to vanish,
and thus the total internal energy of a pure bending state is given by Πint,Ω,h , which in turn is uniquely defined by the
curvature vector field Kh (ξ) = const. In order to represent the desired (constant) distribution of the curvature vector
field Kh (ξ), which is possible for the employed triad interpolation (105), only (nn − 1) · nele of the (nn − 1) · nele + 1
nodal rotation vectors have to arise properly, while the one remaining nodal rotation vector describes rotational rigid
body modes of the beam. Since the curvature vector field is defined via the arc-length derivative of the rotation field,
this one remaining nodal rotation vector can also be interpreted as integration constant resulting from an integration
of the curvature field. Additionally, the nodal position vectors have to arise in a way such that Πint,Γ,h = 0. While
the requirement Γh (ξ) ≡ 0 cannot be fulfilled exactly for the employed triad trial function spaces, the reduced Gauss
integration scheme applied for the CJ element yields a finite number of 3·(nn −1)·nele constraint equations in order to
satisfy Πint,Γ,h = 0. Thus, similar to the rotation field, only (nn−1) · nele of the (nn−1) · nele+1 nodal position vectors have
to arise properly in order to fulfill these constraint equations, while the one remaining nodal position vector describes
translational rigid body modes of the beam. All in all, the (nn−1)·nele+1 nodal position and rotation vectors can always
arise in a way such that a 3D pure bending case can be represented, which consists of a constant curvature vector field
Kh (ξ) =const., a vanishing (reduced integrated) energy contribution Πint,Γ,h of axial and shear strains as well as six
superposed rigid body modes. Consequently, the torsion and bending modes represented by Ωh (ξ) as well as the axial
tension and shear values at the Gauss points represented by Γh (ξGP ) are non-competing and no locking effects are
expected. These considerations can easily be extended to arbitrary curvature fields Kh (ξ) that are representable by the
employed triad interpolation and arbitrary fields Γh (ξ) for which the term ΓTh Γh as occurring in the energy integral can
be integrated exactly by the reduced Gauss integration scheme. In Sections 11.2 and 11.3, the expected result that the
discrete hyperelastic energies associated with pure bending states in 2D and in 3D can be represented exactly by this
beam element formulation will be verified by means of corresponding numerical test cases.
7.3. Conservation properties
In the following, it will be investigated if the Simo-Reissner beam element formulation proposed by Crisfield and
Jelenić and repeated in the section above can represent the variational fields (147), (149) and (151) required in order
to guarantee conservation of linear momentum, conservation of angular momentum as well as conservation of energy.
The representation of a rigid body translation (147) is trivial and given by the nodal primary variable variations
δd̂i = u0 ,
δθ̂i = 0 for i = 1, ..., nn .
(156)
Similarly, a rigid body rotation (149) can be displayed by the nodal primary variable variations
δd̂i = w0 × d̂i ,
δθ̂i = w0 for i = 1, ..., nn .
45
(157)
It follows from (156) and (157), that conservation of linear and angular momentum can be guaranteed. This statement
holds for both spin-vector discretizations (107) and (109) and for the discretized centerline variation (96) since all of
these variants fulfill proper completeness conditions and can exactly represent a constant vector field δθ = w0 . If the
nodal velocities and angular velocities of the time-continuous problem are chosen as primary variable variations
δd̂i = ḋi ,
δθ̂i = wi for i = 1, ..., nn ,
(158)
only the Bubnov-Galerkin variant (107) leads (per definition) to an exact representation of the rates of the spatially
discretized hyperelastic and kinetic energy and consequently to exact energy conservation for the spatially discretized,
time-continuous problem. On the contrary, the Petrov-Galerkin variant (109), which has been employed in (154), is
not variationally consistent with the triad interpolation (105) occurring in the discrete internal and kinetic energies.
Consequently, the weak form (154) does not represent exact energy rates of the spatially discretized problem.
8. Motivation for ”shear-free” beam theories
Geometrically exact Simo-Reissner beam elements unify high computational efficiency and accuracy. In fields of
application where thick beams are involved and the effect of shear deformation is important, they are favorable as
compared to the Kirchhoff type counterparts. However, with increasing beam slenderness ratio ζ = l/R, the shear
contribution to the overall beam deformation decreases. Furthermore, it is exactly the avoidance of the high stiffness
contributions of the shear modes which makes the Kirchhoff-Love theory of thin beams not only applicable, but also
favorable in the range of high beam slenderness ratios. In this brief section, possible benefits of applying Kirchhoff
type beam elements in the range of high slenderness ratios will be illustrated and, at least approximately, quantified.
In Section 11, most of these effects will also be investigated and verified by means of numerical examples.
1: Improved stability properties of time integration scheme: The dynamic equations of motion of highly slender
beams typically result in very stiff Partial Differential Equations (PDEs). With increasing beam slenderness ratio, the
ratio between high eigenfrequencies (associated with shear modes), intermediate eigenfrequencies (associated with
axial tension and twisting modes) and low eigenfrequencies (associated with bending modes) increases considerably.
As a consequence, the stability requirement of explicit time integration schemes leads to very small critical time step
sizes as compared to the large oscillation periods of the bending modes. On the contrary, implicit time integration
schemes can provide unconditional stability in the linear regime of small deformations. However, in the large deformation regime, also their performance is considerably deteriorated by such high-frequency contributions. Despite the
stability aspect, high-frequent modes are strongly affected by the time discretization error and should be avoided as
long as no high-frequency analysis is required by a specific application. In order to illustrate the relevant frequency
spectrum, in the following, the proportionalities of the eigenfrequencies resulting from pure bending ωb , pure torsion
ωt , pure axial tension ωa and pure shearing ω s are given for the linearized beam problem:
s
s
s
s
s
s
s
EI
E
GIT
E
E
GA
E
1
∼
, ωt ∼
∼
, ωa ∼
, ωs ∼
∼ζ
.
(159)
ωb ∼
4
2
2
2
2
ζ ρl
ρI
ρAl
ρIP l
ρl
ρl
ρl2
According to the relations (159), the ratio of axial and torsional eigenfrequencies to bending eigenfrequencies increases linearly with increasing slenderness ratio. The ratio of shear eigenfrequencies to bending eigenfrequencies
increases quadratically with increasing slenderness ratio. Thus, from a theoretical point of view, the avoidance of
shear modes could already improve the numerical behavior considerably. Since the numerical examples of Section 11
mainly focus on static analysis and only present a brief outlook on possible dynamic investigations, a numerical
verification of these theoretical considerations does not lie within the scope of this work. However, numerical investigations of this question available in the literature confirm the predicted trend: Lang and Arnold [92] investigated
the geometrically nonlinear oscillations of a slender beam, which has been modeled by means of the geometrically
exact Simo-Reissner theory and discretized via finite differences (see also [93]). In order to measure the influence of
high-frequency modes on time integration stability, the maximal possible time step sizes of the applied explicit time
integration scheme have been determined for three different mechanical beam models: The full shear-deformable and
extensible Simo-Reissner beam formulation, a beam formulation subject to the Kirchhoff constraint of vanishing shear
46
deformation, and a beam formulation subject to the Kirchhoff constraint and an additional inextensibility constraint
enforcing vanishing axial tension. For the roughest discretization considered in this numerical experiment and an
investigated slenderness ratio of ζ = 100, a rather moderate slenderness ratio as compared to many applications as
mentioned in Section 1, the time step size could be increased by a factor of ≈ 100 when abstaining from the shear
mode and by a further factor of ≈ 5 when additionally abstaining from the axial tension mode. These results indicate
the potential of the Kirchhoff type beam formulations. Furthermore, they suggest that the first step towards an extensible Kirchhoff beam formulation might already represent the essential one with respect to numerical savings.
2: Improved performance of (iterative) linear solvers: According to the previous argumentation, a high ratio
of the highest to the lowest dynamical eigenfrequencies, measured by the dynamic spectral radius, deteriorates the
performance of time integration schemes. In a similar manner, the performance of iterative linear solvers (see e.g.
[113]) decreases with increasing ratio of the highest to the lowest eigenvalue of the tangent stiffness matrix, a measure
for the condition number of this matrix. Furthermore, even if direct linear solvers are applied, very high condition
numbers might considerably limit the achievable numerical accuracy. Especially in dynamics, where such round-off
errors tend to accumulate, these effects are undesirable. In the following, the influence of the different deformation
modes on the condition number is investigated. For simplicity, the physical units of the considered beam problem
are chosen such that the element length lies in the range of lele ≈ 1. Since the element length, or better the element
Jacobian, typically enters the element formulation with different exponents occurring in different stiffness matrix
entries, an element length of lele ≈ 1 seems to be a reasonable choice with respect to conditioning. In this case, the
resulting contributions to the element stiffness matrix of a Simo-Reissner beam element formulation that has been
linearized with respect to the straight configuration, typically obeys the following proportionalities:
kb ∼ EI ∼ ER4 , kt ∼GIT ∼ ER4 , ka ∼ EA ∼ ER2 , k s ∼GA ∼ ER2 .
(160)
Again, kb , kt , ka and k s denote stiffness contributions from bending, torsion, axial tension and shear modes. As long
as lele ≈ 1 holds and the discretization is kept fixed, the cross-section radius R decreases linearly with increasing slenderness ratio ζ. According to (160), it is expected that the ratio of high stiffness contributions (from shear and axial
tension modes) to low stiffness contributions (from torsional and bending modes), and also the condition number,
increases quadratically with the beam slenderness ratio ζ. Furthermore, it is expected that a pure neglect of shear
modes is not sufficient in order to improve conditioning. Thus, the supplementation of the proposed Kirchhoff-Love
formulations by an additional inextensibility constraint seems to be beneficial in order to get also rid of the axial stiffness contributions and to improve also the performance of linear solvers. However, the formulation of an additional
inextensibility constraint does not lie in the focus of this contribution (see also the remark at the end of this section).
3: Improved performance of nonlinear solvers: The relation between the performance of a nonlinear solver, e.g. of
a Newton-Raphson scheme, and the conditioning of the considered problem, e.g. measured via the condition number
of the tangent stiffness matrix, is not that clear as in the case of linear solvers. Nevertheless, typically, it is at least
expected that the performance of tangent-based nonlinear solvers also deteriorates for ill-conditioned problems showing slope differences of the target function by several orders of magnitude when stepping in different directions (e.g.
in directions that activate shear and axial tension modes or in directions that activate bending and torsional modes).
Interestingly, all numerical examples investigated in this work will confirm the trend that the nonlinear solver performance of the considered Reissner type beam elements deteriorates drastically with increasing slenderness ratio while
the total number of Newton iterations required by the Kirchhoff type formulations remains almost unchanged.
4: Reduced system size: Kirchhoff type beam element formulations do not require any degrees of freedom for representing the mode of shear deformation. It can be expected that, at least as long as no convergence deteriorating
phenomena such as locking occur, the same polynomial approximation and the same discretization error level can be
guaranteed with fewer degrees of freedom. This prediction will be confirmed by the numerical examples in Section 11.
5: Smooth geometry representation: The proposed Kirchhoff-Love beam elements will be based on C 1 -continuous
centerline interpolations. These interpolations will eventually result in smooth beam-to-beam contact kinematics, a
property that is highly desirable in order to yield efficient and robust contact algorithms (see e.g. [106, 107]).
47
6: Abstaining from algorithmic treatment of large rotations: It has already been mentioned earlier in this work
that the proposed Kirchhoff-Love beam formulations provide an ideal basis for the derivation of reduced beam models
which are valid under certain restrictions concerning external loads and initial geometry. For example in [105], a
torsion-free beam element formulation could be derived (and extended to dynamic problems in [106]) that is based on
a pure centerline representation and can consequently avoid the treatment of large rotations and associated degrees of
freedom. Thus, many steps within a nonlinear finite element algorithm that are typically complicated by the presence
of large rotations (e.g. spatial interpolation, time discretization, non-symmetric tangent stiffness matrix, non-constant
and non-symmetric mass matrix, incremental and iterative configuration updates) are comparable to those of standard
solid finite elements for this torsion-free formulation. For further details, the reader is referred to [105, 106].
These sources of potential benefits were the motivation for the development of shear-free element formulations based
on the Kirchhoff-Love theory of thin beams. Different realizations of such element formulations, e.g. based on a weak
or on a strong enforcement of the Kirchhoff constraint, will be presented in the next sections. The influence of the
aforementioned aspects on the resulting numerical behavior will be verified in Section 11 via proper test cases.
Remark: As discussed above, the performance of iterative linear solvers could be improved by supplementing the proposed Kirchhoff type element formulations by an additional inextensibility constraint. Unfortunately, in contrast to the Kirchhoff constraint, there is no straightforward way to enforce the inextensibility
constraint directly through a special choice of the primary variables or by a collocation approach that would
allow for Lagrange multiplier elimination on element level, as long as the interpolation property r(xi ) = d̂i
with i = 1, 2 has to be fulfilled for the position vector field r(s) at the element boundary nodes. This statement can easily be illustrated by considering a straight beam element of arbitrary order. In order to avoid
zero-energy modes, inextensibility means in such a case that the solution for the nodal position vectors at
the two boundary nodes cannot arise independently, but have to fulfill a constraint (e.g. ||dˆ1 − dˆ2 ||=l
˙ ele ).
9. Kirchhoff-Love Beam Element Based on Strong Constraint Enforcement
In this section, a finite element formulation based on a strong enforcement of the Kirchhoff constraint is presented. In
Section 9.1, a variant based on nodal triads parametrized via tangent vectors according to Section 6.2.2 is investigated.
In Section 9.2, the transition to a rotation vector-based parametrization as in Section 6.2.1 is conducted. Similar to the
CJ element presented in the previous section, also for this element formulation, the avoidance of possible membrane
locking effects as well as the fulfillment of mechanical conservation properties will be verified in Sections 9.3 and 9.4.
In Appendix B the suitability of the tangent-based and rotation vector-based nodal triad parametrizations for the
modeling of practically relevant Dirichlet boundary conditions and joints is evaluated.
9.1. Residual vector of tangent-based parametrization
Similar to the Simo-Reissner case, the trial and weighting functions are replaced by their discrete counterparts taken
from the finite-dimensional trial subspace (rh , ϕh ) ∈ Uh ⊂ U and the weighting subspace (δrh , δΘ1h ) ∈ Vh ⊂ V.
In the following, 3-noded elements with the nodal primary variables x̂T AN := (d̂1T, t̂1T, ϕ̂1 , d̂2T, t̂2T, ϕ̂2 , ϕ̂3 )T as well as
the set of nodal primary variable variations δx̂T AN := (δd̂1T, δt̂1T, δΘ̂11 , δd̂2T, δt̂2T, δΘ̂21 , δΘ̂31 )T are considered (see also
Figure 3(b)). The centerline interpolation is based on Hermite polynomials according to Section 6.1.2 and completely
defined by the two element boundary nodes. The rotation field interpolation follows equation (113). Concretely,
a quadratic rotation interpolation based on three nodes, thus also involving the element center node, is considered.
Since the orientation of the material triad field is determined by the relative angle ϕh (ξ) as well as the tangent vector
field r0h (ξ), with the latter being a polynomial of order two, this second-order interpolation is sufficient for triad field
discretization. In Meier et al. [104], it has been confirmed that a higher interpolation order nΛ > 3 will not further
improve the approximation quality while a lower interpolation order nΛ < 3 will lead to a decline in the convergence
rate. The time integration of Section 5 is employed, thus leading to the inertia forces fρ and moments mρ given in (42).
Inserting these discretizations into equation (81) and taking advantage of the spin vector interpolation given by (117)
48
eventually yields the element residual vector of the Bubnov-Galerkin variant of this element formulation:
Z1h
h
i
i
v0θ⊥+ v0θkd m+ v F1 − HT f̃ρ − vθ⊥+ vθkd m̃ρ J(ξ)dξ
rS K−T AN+CS ,d̂ =
−1
"
#
h
i
T
− H fσ + vθ⊥+ vθkd mσ ,
(161)
Γσ
"
#
Z1
H0T t
.
rS K−T AN+CS ,Θ̂1 = v0θkΘ m − vθkΘ m̃ρ J(ξ)dξ − vθkΘ mσ with v =
||t||
Γσ
−1
The Bubnov-Galerkin formulation (161) can be transformed into a Petrov-Galerkin variant based on the spin interpolation scheme (116) by simply omitting the terms vθkd , which yields:
"
#
Z1
T
T
0
rS K−T AN,d̂ = vθ⊥m+ v F1 − H f̃ρ − vθ⊥m̃ρ J(ξ)dξ− H fσ + vθ⊥ mσ ,
Γσ
−1
(162)
Z1
rS K−T AN,Θ̂1 =
"
0T
#
v0θkΘ m − vθkΘ m̃ρ J(ξ)dξ − vθkΘ mσ
−1
with v =
Γσ
H t
||t||
.
In order to avoid membrane locking in the range of high element slenderness ratios, the following re-interpolation of
the axial tension and its variation δ based on the MCS procedure (141) (see also [105]) is applied:
3
3
X
X
F̄1 = EA¯ , ¯ (ξ) = Li (ξ)(ξi ), δ¯ (ξ) = Li (ξ)δ(ξi ),
i=1
i=1
3
X
r
(ξi ) = ||r0 || − 1 (ξi ) , δ(ξi ) = δr0T 0
, v̄ = Li (ξ)v (ξi ).
||r || (ξi )
i=1
0
(163)
!
Thus, a locking-free finite element formulation can be obtained by simply replacing the axial force F1 by F̄1 and the
discrete axial tension variation operator v by v̄ in the discrete weak form (161). The discrete expression for the
internal energy associated with the modified axial tension ¯ eventually reads:
Π̃int,¯ =
1
EA¯ 2 .
2
(164)
The element formulation based on the degrees of freedom x̂T AN and δx̂T AN , the residual (162) and (163) will be denoted as SK-TAN element, which stands for ”Strong Kirchhoff constraint enforcement combined with nodal triad
parametrization via nodal TANgents”. Correspondingly, the combination of the degrees of freedom x̂T AN and δx̂T AN ,
the residual (161) and (163) will be denoted as SK-TAN+CS element, referring to the Consistent Spin vector interpolation underlying the Bubnov-Galerkin variant. It has to be emphasized that the replacement of the original axial tension
terms with the corresponding MCS terms according to (163) is standard for all Kirchhoff type beam element formulations considered in this work. Only in examples where for comparison reasons also variants without MCS method
are considered, an additional abbreviation ...-MCS, e.g. SK-TAN-MCS, is employed. For the SK-TAN element, a linearization ∆rS K−T AN = kS K−T AN ∆x̂T AN based on the increment vector ∆x̂T AN := (∆d̂1T, ∆t̂1T, ∆ϕ̂1 , ∆d̂2T, ∆t̂2T, ∆ϕ̂2 , ∆ϕ̂3 )T
will be employed (see Appendix C). In contrary to the multiplicative rotation variations δΘ̂i1 occurring in δrS K−T AN ,
the quantities ∆ϕ̂i represent additive rotation increments of the nodal relative angles ϕ̂i .
9.2. Residual vector of rotation vector-based parametrization
In some scenarios, e.g. applications where complex rotational Dirichlet or coupling conditions should be prescribed
at the element boundary, it can be beneficial to employ the alternative parametrization of the triads Λ1 and Λ2 at the
49
element boundary nodes via rotation vectors according to Section 6.2.1. In such a case, an alternative set of nodal
primary variables given by x̂ROT := (d̂1T, ψ̂1T, tˆ1 , d̂2T, ψ̂2T, tˆ2 , ϕ̂3 )T and δx̂ROT := (δd̂1T, δθ̂1T, δtˆ1 , δd̂2T, δθ̂2T, δtˆ2 , δΘ̂31 )T can
be employed. Here, ψ̂1 and ψ̂2 represent the rotation vectors associated with these boundary triads, δθ̂1 and δθ̂2 are the
corresponding spin vectors and tˆ1 and tˆ2 represent the magnitudes of the nodal tangents, i.e. tˆ1 := ||t1 || and tˆ2 := ||t2 ||, as
introduced in Section 2.2. In this case, the nodal tangents of the Hermite interpolation (97) are no primary variables
anymore, but have to be expressed by ψ̂1 and tˆ1 as well as ψ̂2 and tˆ2 (see also the transformation rule (22)):
t̂i → ti = tˆi exp(S(ψ̂i ))E1
for
i = 1, 2.
(165)
The transformation between the variations (δθ̂i , δtˆi ) as well as (δt̂i , δΘ̂i1 ) for i = 1, 2 is given by the transformation
matrices T̃ and T̃−1 according to (22) and leads to the following relation:
I3
T̃1
,
I3
(166)
δx̂T AN = T̃x̂ δx̂ROT with T̃x̂ =
2
T̃
1
where all blank entries are zero. The required transformation matrices T̃1 as well as T̃2 follow from equation (22) as:
!
!
−tˆ1 S(g11 ) g11
−tˆ2 S(g21 ) g21
1
2
T̃ =
, T̃ =
with gi1 = exp(S(ψ̂i ))E1 .
(167)
g1T
0
g2T
0
1
1
In order to simplify the transformation between the sets of degrees of freedom considered in Sections 9.1 and 9.2, the
residual rT AN,d̂ and rT AN,Θ̂1 according to (161) are slightly reordered and the sought-after residual rROT is introduced:
rT AN,d̂ := (rTT AN,d̂1 , rTT AN,t̂1 , rTT AN,d̂2 , rTT AN,t̂2 )T ,
rT AN,Θ̂1 := (rT AN,Θ̂1 , rT AN,Θ̂2 , rT AN,Θ̂3 )T ,
1
1
1
rT AN := (rTT AN,d̂1 , rTT AN,t̂1 , rT AN,Θ̂1 , rTT AN,d̂2 , rTT AN,t̂2 , rT AN,Θ̂2 , rT AN,Θ̂3 )T ,
1
1
1
rROT := (rTROT,d̂1 , rTROT,θ̂1 , rROT,tˆ1 , rTROT,d̂2 , rT
(168)
T
ˆ2 , rROT,tˆ2 , rROT,Θ̂3 ) .
ROT,θ
1
Inserting relation (166) into the virtual work contribution resulting from one beam element yields:
T
δx̂TT AN rT AN = δx̂TROT T̃x̂ rT AN =˙ δx̂TROT rROT
→
T
rROT = T̃x̂ rT AN .
(169)
According to (167) and (25), the matrix T̃ and its inverse T̃−1 are well-defined as long as tˆ1 , 0 and tˆ2 , 0. The
physical interpretation of tˆi = 0 is that an arc-segment on the beam centerline at the position of the node i with initial
length ds would be compressed to a length of zero. Since such a scenario is impossible from a physical point of view,
these requirements are assumed to be fulfilled. Consequently, the transformation from the residual vector rT AN to the
residual vector rROT is based on a non-singular matrix T̃x̂ . The same statement holds for the transformation of the
global residual vectors RT AN and RROT via the matrix T̃X which represents an assembly of the element matrices T̃x̂ .
Based on these considerations, the following relation can be established:
T
RROT = T̃X RT AN , T̃X ∈ <nX×<nX , rank(T̃X ) = nX
→
RT AN = 0 ⇔ RROT = 0.
(170)
As long as a unique solution of RT AN = 0 exists, the solution of RROT = 0 will lead to the same mechanical equilibrium configuration. In other words, the pure re-parametrization performed from Section 9.1 to Section 9.2 will
not change the results of the discretized beam problem. Nevertheless, the transformation matrix T̃X depends on the
primary degrees of freedom in a nonlinear manner and has to be considered within a consistent linearization procedure.
Throughout this work, the element formulation based on the degrees of freedom x̂ROT and δx̂ROT , the residual (162)
transformed via (169) together with the MCS approach (163) will be denoted as SK-ROT element, which stands for
50
”Strong Kirchhoff constraint enforcement combined with nodal triad parametrization via nodal ROTation vectors”.
Correspondingly, the combination of the degrees of freedom x̂ROT and δx̂ROT , the residual (162) transformed via (169)
together with the MCS approach (163) will be denoted as SK-ROT+CS element. Since the SK-ROT and the SKTAN element yield the same finite element solution (expressed via different nodal primary variables), the following
theoretical investigations concerning locking behavior and conservation properties will only be performed for the SKTAN element. For the SK-ROT element, a linearization ∆rS K−ROT = kS K−ROT ∆x̂ROT based on the increment vector
∆x̂ROT := (∆d̂1T, ∆θ̂1T, ∆tˆ1 , ∆d̂2T, ∆θ̂2T, ∆tˆ2 , ∆ϕ̂3 )T will be employed (see Appendix E). Here, ∆θ̂1 and ∆θ̂2 represent
multiplicative increments while ∆ϕ̂3 can be identified as an additive increment.
9.3. Avoidance of locking effects
In order to investigate the locking behavior of the SK-TAN element, the investigations already made in Section 6.3.3
only have to be extended from 2D to the general 3D case. In 3D, the Kirchhoff beam problem is described by
neq = 4 differential equations (80) and constrained by neq,c = 1 constraint equation in case a pure bending state shall be
represented. Thus, the constraint ratio of the space-continuous problem yields r = neq /neq,c = 4. Due to the employed
MCS method, the discrete number of constraint equations takes on a value of neq,c = 2nele+1. Given the total number of
equations neq = 8nele +1 after application of proper Dirichlet boundary conditions, the discrete constraint ratio yields:
rh = lim
nele →∞
8nele +1
= 4 = r.
2nele +1
(171)
Relation (171) yields the optimal constraint ratio for this element formulation. Consequently, no locking effects are
expected. Furthermore, it has been shown in Section 6.3.3 that the requirement of representing a straight beam configuration with arbitrary distribution of ¯h (ξ) yields a number of independent equations that equals the number of degrees
of freedom. Consequently, such a state can be represented exactly and no zero-energy modes associated with this state
have to be expected. The extension of this statement to 3D is straightforward and will not be further investigated here.
Moreover, it has also been stated in Section 6.3.3 that the discrete hyperelastic energy Πint,h associated with a pure
2D bending case cannot be displayed exactly by the SK-TAN element. Of course, this statement still holds in 3D. In
Sections 11.2 and 11.3, the expected result that the discrete hyperelastic energies associated with pure bending states
in 2D and in 3D cannot be represented in an exact manner by this beam element formulation will be verified by means
of corresponding numerical test cases. There, it will also be shown that this property leads to a slightly increased
discretization error level as compared to the subsequently derived WK-TAN element. However, this observation is
independent from the element slenderness ratio and cannot be attributed to membrane locking.
9.4. Conservation properties
Also for the proposed SK-TAN beam element with strong enforcement of the Kirchhoff constraint, it shall be investigated if the variational fields (147), (149) and (151) required for conservation of linear momentum, conservation of
angular momentum and conservation of energy can be represented by the corresponding discrete weighting subspace
Vh . The representation of a rigid body translation (147) is given by the nodal primary variable variations
δd̂ j = u0 ,
δt̂ j = 0,
δΘ̂i1 = 0 for j = 1, 2; and i = 1, 2, 3.
(172)
This result can be verified by inserting the choices for δd̂ j and δt̂ j made in (172) into the Hermite interpolation (100)
and making use of the completeness conditions underlying the Hermite polynomials (see [104]), which yields δrh (ξ) =
u0 = const. as well as δr0h (ξ) = δg01h (ξ) = 0. Inserting these relations together with δΘ̂i1 = 0 into either (116) or (117)
results in the required vanishing of the discrete spin vector field. Thus, both the Petrov-Galerkin as well as the BubnovGalerkin variant for the spin vector interpolation lead to an exact conservation of the linear momentum. Next, a rigid
body rotation (149) has to be displayed by the following choice of nodal primary variable variations
δd̂ j = w0 × d̂ j ,
δt̂ j = w0 × t̂ j ,
δΘ̂i1 = gT1 (ξi )w0 for j = 1, 2; and i = 1, 2, 3.
(173)
Inserting δd̂ j and δt̂ j according to (157) into (100) gives the desired result δrh (ξ) = w0 × rh (ξ). In Section 6.2.4
(see (123)-(126)), it has been shown that based on the nodal values (173) the Bubnov-Galerkin interpolation (117)
can exactly represent such a ”virtual” rigid body rotation, while the Petrov-Galerkin variant (116) cannot. Thus, only
51
the Bubnov-Galerkin interpolation can guarantee for exact conservation of angular momentum. This result will be
confirmed by subsequent numerical examples. Finally, the conservation of energy has to be investigated. If the nodal
velocities and angular velocities of the time-continuous problem are chosen as nodal primary variable variations,
δd̂ j = ḋ j ,
δt̂ j = ṫ j ,
δΘ̂i1 = gT1 (ξi )wi for j = 1, 2; and i = 1, 2, 3,
(174)
again, only the Bubnov-Galerkin variant (117) leads (per definition) to an exact representation of the rates of the
discrete internal and kinetic energy and to exact energy conservation for the spatially discretized, time-continuous
problem. In contrast, the Petrov-Galerkin variant (116) is not variationally consistent with the triad interpolation (113)
underlying the discrete energies and cannot guarantee for energy conservation of the time-continuous problem.
10. Kirchhoff-Love Beam Element Based on Weak Constraint Enforcement
As alternative to the formulation presented in the last section, a beam element will now be presented that is based
on the weak fulfillment of the Kirchhoff constraint. Thus, the basis for the intended element formulation is provided
by the Simo-Reissner beam theory. In a first step, a finite element formulation of Simo-Reissner type with a C 1 continuous centerline representation will be derived in Section 10.1. Afterwards, the Kirchhoff constraint of vanishing
shear strains will be enforced in order to end up with a finite element formulation of Kirchhoff type. Following
the derivations in Section 4.4, the weak statement of the Kirchhoff constraint (85) can be realized by introducing
spatial interpolations for the Lagrange multipliers and their variations, i.e. by choosing a proper discrete trial space
(λΓ,2,h , λΓ,3,h ) ∈ UλΓ23 ,h ∈ UλΓ23 and a proper discrete weighting space (δλΓ,2,h , δλΓ,3,h ) ∈ VλΓ23 ,h ∈ VλΓ23 . The resulting
nonlinear system of discrete equilibrium equations will contain discrete Lagrange multipliers as additional nodal
primary variables and exhibit a saddle point type structure. In order to avoid the additional effort of solving a large
system of equations with saddle point structure, a slightly different approach is chosen here. In the next section, a
modified Reissner type beam element formulation will be presented, which is based on a smooth Hermite centerline
interpolation and a MCS type strain re-interpolation not only of the axial strain h (ξ) but also of the shear strains
Γ2,h (ξ) and Γ3,h (ξ). Applying the constraint of vanishing shear strains in a consistent manner directly on the reinterpolated strain fields Γ̄2,h (ξ) and Γ̄3,h (ξ) yields a collocation point type approach of constraint enforcement which
does not require additional Lagrange multipliers (see Section 10.2). Throughout this contribution, this variant will be
preferred since it does neither yield additional Lagrange multiplier degrees of freedom nor a saddle point type system
of equations. Also for this element formulation, the two variants concerning nodal rotation parametrization according
to Sections 6.2.1 and 6.2.2 will be presented in the following Sections 10.2 and 10.3.
10.1. Basic formulation: Hermitian Simo-Reissner element
The Reissner type beam element formulated in this section represents an intermediate step in the derivation of a corresponding Kirchhoff type beam element formulation in the next section. The discrete beam centerline representation
is given by the Hermite interpolation (97) based on the position and tangent vectors d̂i and t̂i at the two element
boundary nodes. Furthermore, the rotation interpolation is given by a three-noded representation of (105) with nodal
triads Λ1 , Λ2 and Λ3 . Again, a finite element formulation will be considered on the basis of a strain re-interpolation
similar to the MCS method (163). While in the Kirchhoff case (163), only the axial strain has been treated, now the
entire deformation measure Γ will be re-interpolated in order to avoid membrane as well as shear locking:
Γ̄(ξ) =
3
X
i=1
Li (ξ)Γ(ξi ),
δΓ̄(ξ) =
3
X
Li (ξ)δΓ(ξi )
ξ1 = −1, ξ2 = 1, ξ3 = 0.
(175)
i=1
On the basis of (175), also the hyper-elastic stored energy function Π̃int (Ω, Γ) given in (37) has to be replaced by
Π̃int (Ω, Γ̄). Now, one can introduce the following set of degrees of freedom x̂HS R := (d̂1T, t̂1T, ψ̂1T, d̂2T, t̂2T, ψ̂2T , ψ̂3T )T as
well as the associated variation vector δx̂HS R := (δd̂1T, δt̂1T, δθ̂1T, δd̂2T, δt̂2T, δθ̂2T, δθ̂3T )T . Based on the weak form (33),
52
(175), the definitions (36) and (35), (97) and (105), the element residual vector can be derived as follows:
rHS R,d̂ =
Z1
h
vΓ1 F̄ − LT f̃ρ J(ξ)dξ − LT fσ
i
Γσ
−1
rHS R,θ̂ =
Z1
h
(176)
i
L0T m + vΓ2 F̄ − LT m̃ρ J(ξ)dξ − LT mσ = 0,
Γσ
−1
3
X
F̄ = CF Γ̄, vΓ1 = Li (ξ) H0TΛ
(ξi )
3
X
, vΓ2 = Li (ξ) LTS(t)Λ
i=1
(ξi )
.
i=1
This element formulation could e.g. be applied to problems of thick beams with higher continuity requirements
(e.g. beam contact). However, here, the formulation (176) solely represents an intermediate step in the derivation of
Kirchhoff beam elements with weak enforcement of the Kirchhoff constraint as performed in the next two sections.
10.2. Residual vector of tangent-based parametrization
Due to (175), the general weak constraint enforcement of Section (4.4) can be simplified:
Γ̄2 (ξ) ≡ Γ̄3 (ξ) ≡ 0 → r0T(ξi )g2 (ξi ) =˙ r0T(ξi )g3 (ξi ) =˙ 0
for
i = 1, 2, 3.
(177)
According to (177), the Kirchhoff constraint is exactly fulfilled at the three collocation points. In the following, a
parametrization is chosen that directly fulfills these constraints without the need for additional Lagrange multipliers.
Thereto, the same set of nodal degrees of freedom x̂T AN = (d̂1T, t̂1T, ϕ̂1 , d̂2T, t̂2T, ϕ̂2 , ϕ̂3 )T as well as the set of nodal
primary variable variations δx̂T AN = (δd̂1T, δt̂1T, δΘ̂11 , δd̂2T, δt̂2T, δΘ̂21 , δΘ̂31 )T as in Section 9.1 are chosen. In case of
a Bubnov-Galerkin approach, the following discrete spin vector field results from the triad interpolation (105) in
combination with the Kirchhoff constraint according to (177) (see also (107)):
3
X
δθh (ξ) = Ĩi (ξ)δθi
with δθi = δΘ̂i1 g1 (ξi ) +
i=1
!
S[r0 (ξi )]δr0 (ξi )
.
||r0 (ξi )||2
(178)
Since the Kirchhoff constraint is exactly fulfilled at the three element nodes, the constrained variant (71) of the nodal
spin vectors δθi has been combined with (107). Based on the alternative Petrov-Galerkin approach (109), the Kirchhoff
constraint given by the relations (177) yields the following expression for the spin vector field:
3
X
δθh (ξ) = Li (ξ)δθi
with δθ =
i
δΘ̂i1 g1 (ξi )
i=1
!
S[r0 (ξi )]δr0 (ξi )
+
.
||r0 (ξi )||2
(179)
Similar to the CJ element, only the latter version (179) will be employed throughout this contribution. The final finite
element residual vector resulting from the discretized counterparts of these fields reads:
"
# "
#
Z1
T
T
0
rWK−T AN,d̂ = v̄θ⊥ m + v̄ F̄1 − H f̃ρ − v̄θ⊥ m̃ρ J(ξ)dξ − H fσ − v̄θ⊥ mσ ,
Γσ
−1
Γσ
"
#
Z1
rWK−T AN,Θ̂1 = v̄0θkΘ m − v̄θkΘ m̃ρ J(ξ)dξ − v̄θkΘ mσ ,
Γσ
−1
v̄θ⊥ =−
3
X
Li (ξ)vθ⊥ (ξi ),
i=1
3
X
v̄θkΘ =
i=1
Li (ξ)vθkΘ (ξi )
3
X
v̄ =
Li (ξ)v (ξi ),
i=1
3 Li (ξ)
X
,ξ
with v̄0...=
v... (ξ i ).
J(ξ)
i=1
53
(180)
In the following, the formulation based on the degrees of freedom x̂T AN and δx̂T AN and on the residual (180) will be
denoted as WK-TAN element, which stands for ”Weak Kirchhoff constraint enforcement combined with nodal triad
parametrization via nodal TANgents”. For the WK-TAN element, a linearization ∆rWK−T AN = kWK−T AN ∆x̂T AN based
on the increment vector ∆x̂T AN := (∆d̂1T, ∆t̂1T, ∆ϕ̂1 , ∆d̂2T, ∆t̂2T, ∆ϕ̂2 , ∆ϕ̂3 )T is employed (see Appendix D).
Remark: Actually, a collocation type approach has been applied in order to enforce the Kirchhoff constraint. Nevertheless, the notion ”weak constraint enforcement” is kept throughout this work since such a
procedure still represents the basis of the space-continuous formulation. Moreover, the difference to the
formulation based on ”strong constraint enforcement” of Section 9 shall be emphasized by this naming.
10.3. Residual vector of rotation vector-based parametrization
Also for the element formulation of Section 10.2 based on weak Kirchhoff constraint enforcement, a coordinate
transformation from x̂T AN and δx̂T AN to the alternative primary variables x̂ROT := (d̂1T, ψ̂1T, tˆ1 , d̂2T, ψ̂2T, tˆ2 , ϕ̂3 )T and
δx̂ROT := (δd̂1T, δθ̂1T, δtˆ1 , δd̂2T, δθ̂2T, δtˆ1 , δΘ̂31 )T can be performed. The transformation rule for the element residual
T
vector rROT = T̃x̂ rT AN is identical to (169) in Section 9.2. Throughout this work, the element formulation based
on the degrees of freedom x̂ROT and δx̂ROT and on the element residual vector (180) transformed via (169) will be
denoted as WK-ROT element, which stands for ”Weak Kirchhoff constraint enforcement combined with nodal triad
parametrization via nodal ROTation vectors”. For the WK-ROT element, a linearization ∆rWK−ROT = kWK−ROT ∆x̂ROT
based on the increment vector ∆x̂ROT := (∆d̂1T, ∆θ̂1T, ∆tˆ1 , ∆d̂2T, ∆θ̂2T, ∆tˆ1 , ∆ϕ̂3 )T will be employed (see Appendix E).
10.4. Avoidance of locking effects
In the investigation of the locking behavior of the proposed WK-TAN element, many results already derived in Section 7.2 and 9.3 can be re-used. Since the numbers neq and neq,c for the space-continuous as well as for the discrete
problem are identical to the SK-TAN element, it can readily be concluded that also the WK-TAN element formulation
shows an optimal constraint ratio of r = rh = 4 and that no membrane locking effects are expected for this element.
Similar to Section 7.2, it shall be shown that also the WK-TAN/-ROT elements can exactly represent the internal
energy associated with a 3D pure bending state. This time, the internal energy is split into contributions stemming
from torsion and bending and into contributions stemming from axial tension, i.e. Πint,h = Πint,Ω,h +Πint,,h . For a pure
bending state, the energy contribution Πint,,h has to vanish, and thus the total internal energy of a pure bending state
is given by Πint,Ω,h , which is uniquely defined by the curvature vector field Kh (ξ) = const. In order to represent the
desired (constant) distribution of the curvature vector field Kh (ξ), which is possible for the employed triad interpolation (105), only 2nele of the 2nele +1 nodal triads have to arise properly, while the one remaining nodal triad describes
rotational rigid body modes of the beam. Although the nodal triads are not necessarily parametrized by nodal rotation
vectors, still three conditions result from each of these 2nele nodal triads, thus resulting in a total of neq,Ω = 6nele conditions. Additionally, the axial strains at the collocation points have to vanish in order to yield a vanishing contribution
Πint,,h = 0. This requirement results in neq, = 2nele +1 additional conditions that have to be fulfilled at the collocation
points. If again six further conditions are considered in order to superpose arbitrary rigid body modes (representing
the minimally required number of Dirichlet boundary conditions in static problems), the total number of neq = 8nele +7
equations equals the total number of nuk = 7(nele+1)+nele unknowns contained in the global vector X for the considered
WK-TAN/ROT elements. Thus, in case a unique FEM solution is existent, a 3D pure bending case can be represented
exactly. Similar to the CJ element, the torsion and bending modes represented by Ωh (ξ) as well as the axial tension
values at the collocation points represented by (ξCP ) are non-competing and no locking effects are expected. Again,
these considerations can easily be extended to arbitrary curvature fields Kh (ξ) that are representable by the employed
triad interpolation and arbitrary second-order polynomials ¯ (ξ) according to (163). In Sections 11.2 and 11.3, the
expected result that the discrete hyperelastic energies associated with pure bending states in 2D and in 3D can exactly
be represented by the WK-TAN/ROT elements will be verified by means of corresponding numerical test cases.
10.5. Conservation properties
Since the WK-TAN beam element proposed above basically combines the triad interpolation and the spin vector
interpolation δθh (ξ) of the CJ element (see Section 7) with the centerline interpolation and its variation δr0h (ξ) applied
54
to the SK-TAN element of Section 9, the corresponding conservation properties can directly be concluded from the
investigations of Sections 7.3 and 9.4. Consequently, this element will exactly fulfill conservation of linear and angular
momentum. Conservation of energy can only be guaranteed for the spatially discretized, time-continuous problem in
case the Petrov-Galerkin spin vector interpolation (179) is replaced by its Bubnov-Galerkin counterpart (178).
11. Numerical Examples
In this section, the previously proposed beam element formulations will be investigated numerically by means of
proper test cases. All simulations results presented in the following rely on a software implementation of the proposed finite element formulations and numerical algorithms within the in-house finite element research code BACI
(cf. Wall and Gee [147]), developed jointly at the Institute for Computational Mechanics at the Technical University
of Munich. While most of the numerical examples (see Sections 11.1-11.6) are considered as quasi-static problems,
eventually, in Section 11.7, also a dynamic test case is investigated. In a first step, these numerical examples aim to
verify the principle applicability and accuracy of the proposed general and reduced Kirchhoff-Love beam elements in
the range of different beam slenderness ratios. This verification crucially relies on detailed comparisons with analytic
reference solutions, benchmark tests known form the literature as well as numerical reference solutions generated by
means of well-established geometrically exact beam element formulations of Simo-Reissner type. More specifically,
also the essential requirements formulated in Section 6.3 such as objectivity and path-independence, avoidance of
locking effects, consistent spatial convergence behavior as well as the fulfillment of conservation properties will be
verified for the different beam element formulations presented in Sections 7-10. Finally, based on the arguments given
in Section 8, the focus will also lie on detailed comparisons of Reissner and Kirchhoff type beam element formulations
for example with respect to the resulting discretization error level or the performance of the Newton-Raphson scheme.
Since the Kirchhoff type beam element formulations based on a tangent-based triad parametrization and the formulations based on a rotation vector-based triad parametrization (compare e.g. Sections 9.1 and 9.2 or Sections 10.2
and 10.3) have been shown to yield identical FEM solutions, only the former category will be investigated with respect to spatial discretization errors. Furthermore, for all examples without analytic solution, the standard choice for
the reference solution rre f (see also (143) of Section 6.3.4) is a numerical solution via the WK-TAN element (see
Section 10.2) employing a spatial discretization that is by a factor of four finer than the finest discretization shown
in the corresponding convergence plot. In order to achieve a good comparability among the different geometries and
load cases, a standard set of geometrical and constitutive parameters has been applied in all simulations unless stated
otherwise. This standard set consists of a beam with initial length l = 1000 and square cross-section with side length
R. These parameters lead to a cross-section area of A = R2 and to moments of inertia of area of I2 = I3 =: I = R4 /12
and IT = R4 /6. Different beam slenderness ratios ζ := l/R are generated by varying the value R = 0.1, 1.0, 10, 100 of the
cross-section side length. The standard choice for the constitutive parameters is E = 1.0 and G = 0.5, thus leading to
EI =GIP = R4 /12. For all numerical examples considered in the following sections, a Newton-Raphson scheme based
on consistent linearization has been applied in order to solve the set of nonlinear equations resulting from the temporally and spatially discretized weak form of the balance equations. As convergence criteria, the Euclidean norms
k+1
of the displacement increment vector ∆Xk+1
n+1 and of the residual vector R(Xn+1 ) are checked. For convergence, these
k+1
k+1
norms have to fall below prescribed tolerances δR and δX , i.e. ||R(Xn+1 )|| < δR and ||∆Xn+1
|| < δX . Typical convergence
−8
tolerances chosen for the subsequent examples are in the range of δX = 10 as well as δR = 10−7 , 10−9 , 10−11 , 10−13 for
the slenderness ratios ζ = 10, 100, 1000, 10000. For the (quasi-)static problems presented in the following sections, the
external loads are applied on the basis of an incremental procedure, where N shall denote the number of load steps and
∆t the load step size. As long as nothing is stated to the contrary, the following simple procedure is applied in order to
adapt the load step size during the static simulation in an efficient manner: Initially, a comparatively small load step
size ∆t0 is chosen, e.g. according to N0 = 1. If the Newton-Raphson scheme has not converged within a prescribed
number of niter,max iterations, the step size is halved and the load step is repeated. This procedure is repeated until
convergence can be achieved. Then, after four converging load steps on the low step size level, the step size is doubled again. Also this procedure of successively doubling the step size after four converging load steps at the current
step size level is repeated until the original step size ∆t0 is reached again. This procedure will not only drastically
increase the overall computational efficiency, it also allows for comparatively objective and fair comparisons of the
55
performance of the Newton-Raphson scheme for different element formulations. In subsequent numerical examples,
such comparisons will be made on the basis of the accumulated number of Newton iterations
niter,tot :=
N
X
niter,n ,
(181)
n=1
required to solve the entire problem. Here, niter,n is the number of iterations required for load step tn . In ”load step
adaption scheme” above, non-converging steps are considered in the total number of iterations with niter,n = niter,max .
11.1. Example 1: Verification of objectivity
The objectivity of the Kirchhoff beam element formulations proposed in Sections 9 and 10 has already been proven
theoretically. In order to verify these results numerically, the following test case will be investigated (see Figure 4(a)):
At the clamped end of an initially curved beam with slenderness ratio ζ = 10, whose stress-free centerline configuration equals a quarter circle, a Dirichlet rotation with respect to the global x-axis is imposed. For the presented
quasi-static example, a total rotation angle of 20π, increasing linearly over 100 load steps, is prescribed. In order
to investigate objectivity, the normalized internal (hyperelastic) energy is plotted over the total number of rotations
(see Figure 4(b)) for the WK-TAN and SK-TAN element as well as for the Bubnov-Galerkin variant of the SK-TAN
element formulation with Consistent Spin (SK-TAN+CS) vector interpolation according to (161). For comparison
reasons, also the non-objective SR Kirchhoff beam element formulation investigated in Meier et al. [104] will be considered (see also the last remark at the end of Section 6.2.4). For clearness, the internal energy Πint is normalized by
the factor Πint,r = 0.5EIπ2 /(4l), which is equal to the amount of mechanical work that is required to bend the initially
stress-free quarter circle into a straight beam by means of a discrete, external end-moment. Of course, the internal
energy should vanish for a beam that is merely rotated out of its stress-free initial configuration.
(a) Problem setup.
(b) Internal energy due to imposed rigid body rotation.
Figure 4: Objectivity test: Rigid body rotation of an initially stress-free quarter circle.
From Figure 4(b), it becomes obvious, however, that the internal energy of the SR formulation increases over the
number of rotations, which is a clear indication for the already theoretically predicted non-objectivity. Within 10
rotations, the normalized energy reaches a value of almost Πint,r /4, which results in a clearly visible deformation of
the initial quarter circle. On the contrary, the internal energy of all the other investigated element formulations results
in a value that is zero up to machine precision. Finite element formulations based on such interpolation schemes as
the investigated SR element might show reasonable results for static test cases (see e.g. [104]). However, especially in
dynamic problems involving considerable rigid body motions, non-physical results as well as a drastic deterioration of
the conservation properties investigated in Section 6.3.5 can follow from the application of such element formulations.
56
11.2. Example 2: Pure bending in 2D
The examples shown in this section exclusively focus on 2D geometries and load cases. The section is sub-divided
into two subsections: In Section 11.2.1, two load cases, a pure pending case as well as a combined moment-and-force
load case, yielding geometrically nonlinear, but still moderate centerline deformations, are considered. This section
aims at the investigation of membrane locking effects and at the comparison of different anti-locking tools, especially
of the MCS method proposed in Section 6.3.3. In Section 11.2.2, again a pure bending and a combined moment-andforce load case will be considered. However, due to higher load factors, the resulting degree of deformation is further
increased as compared to the examples of Section 11.2.1. This higher degree of deformation reveals clear differences
in the approximation quality of the WK and SK Kirchhoff beam element variants. Besides the comparison of these
two variants, also a first proof of concept for the development of higher-order Hermitian Kirchhoff elements is given.
11.2.1. Comparison of different anti-locking methods
An initially straight beam is clamped at one end. Two different load cases will be analyzed: The first load case M
is identical to the example analyzed in Section 6.3.3 and solely consists of a discrete end-moment M = (0, 0, M)T
applied in one load step. The moment M = EIπ/(2l) exactly bends the beam into a quarter-circle shaped arc. In the
second load case M+F, the end-moment and an additional tip force F = (0, F, 0)T in global y-direction are applied in
one load step. The initial and deformed geometries for these load cases are illustrated in Figure 5.
(a) Straight beam bent by end-moment.
(b) Straight beam bent by end-moment and -force.
Figure 5: Initial and deformed configuration of an initially straight beam for two load cases.
While the standard parameters and the slenderness ratios ζ = 10, 100, 1000, 10000 have been chosen for the load
case M, only the highest and therefore most critical slenderness ratio ζ = 10000 in combination with an external
force of F = 1.0 · 10−10 ≈ 10M/l has been investigated for the load case M+F. In a first step, the SK-TAN element
formulation according to (162) is applied in combination with a full Gauss-Legendre integration (”SK-TAN-FI“)
with nG = 4, a reduced Gauss-Legendre integration (”SK-TAN-RI“) with nG = 3, a classical assumed natural strain
approach (”SK-TAN-ANS“) as well as the MCS method according to (163) based on nG = 4 integration points (”SKTAN-MCS“) (see Section 6.3.3 and [105] for further details on these variants). In Figure 6(a), the relative L2 -error
of the load case M and different slenderness ratios is plotted with respect to an analytic reference solution. For
spatial discretization, the variant SK-TAN-FI based on meshes with 1, 2, 4, 8, 16, 32 and 64 elements has been applied.
Accordingly, the convergence is slowed down dramatically with increasing slenderness ratio. If the beam is e.g.
discretized by one finite element (lele = 1000), the relative error increases almost by two orders of magnitude when
enhancing the slenderness ratio from ζ = 10 to ζ = 10000. However, Figure 6(a) also reveals that this effect decreases
with decreasing element sizes and almost completely disappears for discretizations with more than 32 elements.
57
10
0
10
10 -2
10 -2
-4
2
rel
10
10 -6
10 -8
10
-10
||e||
||e||
2
rel
10
SK-TAN-FI,
SK-TAN-FI,
SK-TAN-FI,
SK-TAN-FI,
4
10
th
10 -6
10
-10
10
-12
10 2
10 3
element length
10 4
10
2
rel
10
10 -12
10 1
||e||
2
rel
||e||
-4
-10
10 2
10 3
element length
10 4
0
10 -2
10 -6
10
10 1
(b) Reference: Analytic, different slenderness ratios.
0
10 -8
ζ=10
ζ=100
ζ=1000
ζ=10000
4 th order
10 -2
10
SK-TAN-MCS,
SK-TAN-MCS,
SK-TAN-MCS,
SK-TAN-MCS,
order
(a) Reference: Analytic, different slenderness ratios.
10
-4
10 -8
ζ=10
ζ=100
ζ=1000
ζ=10000
-12
10 1
0
10 -6
ζ=10000, SK-TAN-MCS
ζ=10000, SK-TAN-ANS
ζ=10000, SK-TAN-RI
ζ=10000, SK-TAN-FI
ζ=10000, WK-TAN-MCS
4 th order
10 -8
ζ=10000, SK-TAN-MCS
ζ=10000, SK-TAN-ANS
ζ=10000, SK-TAN-RI
ζ=10000, WK-TAN-MCS
4 th order
10 2
10 3
element length
-4
10
10 4
(c) Reference: Analytic, ζ = 10000.
-10
10 -12
10 1
10 2
10 3
element length
10 4
(d) Reference: WK-TAN-MCS, ζ = 10000.
Figure 6: Straight beam subject to the load cases M and M+F: Relative L2 -error.
The reason for this behavior lies in the fact that the element slenderness ratio ζele = lele /R is the key-parameter for
the observed locking effect and that the latter also decreases with decreasing element sizes. However, for typical
engineering applications with relative error bounds in the range of 1%, the effect is by no means negligible. For
sufficiently fine discretizations the expected convergence order of four is reached. In Figure 6(b), the relative L2 error is plotted for the same slenderness ratios as before, but for the variant SK-TAN-MCS, which is supplemented
by the MCS method according to (163). As expected, the locking effect completely disappears for all investigated
slenderness ratios. However, as shown in Figure 6(c), for the load case M and the highest investigated slenderness ratio
of ζ = 10000, the same effect can alternatively be achieved by applying a simple reduced integration procedure (variant
SK-TAN-RI) or a classical ANS approach (variant SK-TAN-ANS). On the contrary to load case M, Figure 6(d)
reveals that no distinctive improvement of the locking behavior can be obtained by these alternative methods for
the load case M+F: Both the ANS approach as well as the reduced integration scheme can only slightly alleviate
the locking effect in the range of rather coarse discretizations as compared to the variant SK-TAN-FI. The MCS
approach, however, completely eliminates the error offset due to membrane locking also for this load case. The
explanation for this observation is obvious and in agreement with the statements of Section 6.3.3: Similar to the
working principle of the MCS method, the reduced integration scheme can alleviate locking by reducing the number
of constraint equations. Yet, as shown in Section 6.3.3, the MCS method leads to a lower number of constraint
equations as compared to the simple reduced integration scheme, which makes the latter method less effective. While
58
the reduced integration scheme seems to be sufficient for the load case M, a special case yielding symmetric curvature
distributions within the elements, the more general deformed configurations resulting from the load case F+M already
demonstrate the limits of this simple method. On the other hand, the working principle of the standard ANS method
does not aim at a reduction of the number of constraint equations, but rather at an evaluation of the critical axial
tension term at selected collocation points with vanishing parasitic strains. In the geometrically nonlinear regime
of large deformations, the parameter space positions of these optimal collocation points are deformation-dependent.
Obviously, the load case M+F already leads to a change in these positions up to an extent that almost completely
destroys the working principle and impact of the ANS method. Summing up, it is to say that based on these two
examples the proposed MCS method seems to be superior to standard methods such as reduced integration or ANS
in terms of locking avoidance when combined with the considered geometrically exact Kirchhoff beam elements. In
example 8 of Section 11.6, a further comparison of these different anti-locking methodologies on the basis of a very
general problem setting involving 3D deformation states and an initially curved geometry will be presented, which
will confirm this result. For completeness, in Figures 6(c) and 6(d), also the L2 -error of the WK-TAN-MCS element
based on a weak enforcement of the Kirchhoff constraint according to (180) has been plotted. For the load cases M
and M+F, this formulation yields a comparable convergence behavior and discretization error level as the SK-TANMCS element. Next, it will be shown that this behavior will change with increasing deformation. In the following,
the MCS method will be employed per default and the abbreviation ...-...-MCS omitted.
11.2.2. Comparison of different element formulations
In the convergence plots investigated for the two load cases M and M+F, no noteworthy differences between the SK
and WK element could be observed. In order to investigate the difference in the two general approaches of enforcing the Kirchhoff constraint in a strong or in a weak manner further in detail and to perform first comparisons with
geometrically exact beam element formulations of Simo-Reissner type, two additional load cases will be considered
(see Figure 7): The first load case considered in this section, in the following denoted as M̃, simply increases the
magnitude of the external moment by a factor of eight as compared to the previous load case M, i.e. M̃ = 8M, thus
leading to a deformed geometry that is represented by a double circle (see Figure 7(a)). Since the contribution of Simo
and Vu-Quoc [134], this load case has been established as a standard test case for geometrically exact beam element
formulations. Finally, in a fourth load case, denoted as M̃+F̃, the end-moment M̃ = 8M is supplemented by a tip force
F = (0, F̃, 0)T in global y−direction, whose magnitude is this time exactly chosen as F̃ = 10 M̃/l = 0.08M. The initial
geometry as well as the final configuration of this last load case is illustrated are Figure 7(b).
(a) Straight beam bent by end-moment.
(b) Straight beam bent by end-moment and -force.
Figure 7: Initial and deformed configuration of an initially straight beam for two load cases.
In Figure 8(a), the L2 -error resulting from the load case M̃ has been plotted for the WK-TAN and the SK-TAN element
as well as for the Reissner type CJ beam element formulation proposed by Crisfield and Jelinic [43, 80] and presented
in Section 7. Here, discretizations with 8, 16, 32, 64, 128 and 256 elements have been employed. These discretizations
59
10
10
0
10 -2
10 -4
10 -4
-6
10 -8
10
-10
||e||
10
e,rel
10 -2
2
rel
||e||
0
10
-6
10 -8
Crisfield & Jelenic
WK-TAN
SK-TAN
SK-TAN, L/8
10
-10
4 th order
10 -12
10 1
10 2
10 3
degrees of freedom
10 4
(a) Reference: Analytic, L2 -error.
10 -12
10 1
SK-TAN
SK-TAN, L/8
4 th order
10 2
10 3
degrees of freedom
10 4
(b) Reference: Analytic, energy error.
Figure 8: Straight beam and load case M̃: convergence of L2 - and energy error.
are comparable to the load case M since similarly to that example, also here, the roughest discretization is based
on one finite element per 90◦ -arc segment of the analytic solution. In order to enable a reasonable comparison of
different element formulations, here and in the following, the discretization error will be plotted over the total number
of degrees of freedom resulting from the respective finite element discretization. Since no shear deformation is
present for this example, the Reissner and Kirchhoff type elements converge towards the same analytic solution.
All element formulations exhibit the expected optimal convergence order of four, indicated by the black dashed
line. The WK-TAN element shows the expected result that Kirchhoff element formulations can represent the same
discretization error level with less degrees of freedom as compared to the Reissner type element formulation (see
Section 8). Furthermore, for this example, it can even be shown that the lines representing the discretization error
of the WK-TAN element and of the CJ element formulation would be almost identical if the discretization error
was plotted solely over the degrees of freedom associated with the centerline interpolation. Thus, the observable
difference in Figure 8(a) is a pure result of the additional rotational degrees of freedom required for Reissner type
element formulations in order to represent shear deformation. Such a behavior is expected for this pure bending
example since the two considered element formulations can exactly represent the internal energy associated with a
pure bending state (see Sections 7.2 and 10.4). Consequently, the discretization error contribution stemming from
the second term in (146) vanishes, the finite element problem degenerates to a pure problem of polynomial curve
approximation represented by the first term in (146) and, thus, the discretization error plotted over the number of
centerline DoFs yields similar results for the Lagrange centerline interpolation of the Reissner type element and the
Hermite centerline interpolation of the Kirchhoff type element. The situation is completely different for the SKTAN element, which cannot exactly represent the internal energy associated with a pure bending state. A closer
investigation would confirm the expected result that the SK-TAN element exhibits a remaining error in the lengthspecific hyperelastic stored energy (75), which is more or less constant along the beam length. Based on this finding,
it can easily be answered why the discretization error level of the SK-TAN element applied to the load case M̃ is
considerably increased as compared to the first load case M (while an identical level of the length-specific L2 -error
has been observed for the WK-TAN and the CJ element): By the FEM, solely the distribution of the second centerline
derivative r00 is optimized in order to yield a minimal energy error within the beam domain Ωl , while the centerline
field r itself is only constrained at the clamped end of the beam. Thus, with increasing distance from the clamped end,
the discretization error in the centerline field r, resulting from a two-fold integration of the (more or less) constant
error in the second derivative r00 along an increasing arc-length segment, also increases. Consequently, by assuming
comparable errors in the length-specific energy for comparable discretizations (i.e. the same number of finite elements
representing the same angle segment of the analytic solution), a higher length-specific discretization error is expected
for the load case M̃ as compared to the load case M. Figure 8(b) confirms the expected result that the energy error
of the SK-TAN element does not vanish for this example and exhibits a convergence order of four. Furthermore, it is
60
10
0
10
Crisfield & Jelenic
SK-TAN
4 th order
10 -2
2
rel
-6
10 -8
10
6 th order
10 -4
||e||
2
rel
||e||
10
Crisfield & Jelenic
SK-TAN
4 th order
10 -2
6 th order
10 -4
0
10
-6
10 -8
-10
10
10 -12
10 1
10 2
10 3
degrees of freedom
10 4
(a) Reference: Analytic, third-order elements.
-10
10 -12
10 1
10 2
10 3
degrees of freedom
10 4
(b) Reference: Analytic, fifth-order elements.
Figure 9: Straight beam and load case M̃: convergence of third- and fifth-order elements.
shown that the length-specific energy error ”averaged” along the entire beam length is identical to the length-specific
energy error ”averaged” only along the first eighth of the beam (representing a quarter circle). As consequence of
the error accumulation described above, the length-specific L2 -error is lower (and similar to load case M) if it is only
”averaged” along the first eighth of the beam (see Figure 8(a)). From a rather mathematical point of view, the increased
discretization error level of the SK-TAN element can be explained by a high level of the second, energy-related term
in (146) that dominates the overall discretization error. This behavior, in turn, is a pure consequence of the fact that
the two exponents k+1 and 2(k−m+1) in (146) are identical for trial functions of polynomial degree k = 3. However,
for polynomial degrees k > 3, the second term in (146) is expected to converge with a higher rate and consequently,
for sufficiently fine discretizations, the first term reflecting the pure polynomial approximation power will determine
the overall discretization error level. In this range, a lower discretization error per DoF can be expected for the
Kirchhoff type beam element formulations independently of the beam length, the complexity of the deformation
state or the type of boundary conditions. For a first proof of principle, in Figure 9(b), the L2 -error resulting from
the CJ element with fifth-order Lagrange interpolation as well as from the SK-TAN element based on a fifth-order
Hermite interpolation are depicted: While for very rough discretizations, the energy-related error contribution with
higher convergence rate still seems to dominated the overall discretization error of the SK-TAN element, the expected
optimal gap between the Reissner discretization error (higher level) and the Kirchhoff discretization error (lower
level) can be observed. For comparison reasons, in Figure 9(a), the results of the corresponding third-order variants
of Figure 8(a) are repeated. Since this work focuses on the development of third-order Kirchhoff beam elements, no
further details on the construction of higher-order Hermite polynomials (either by introducing additional nodes or by
considering higher-order derivatives) will be given at this point. However, it is expected that a comparable behavior
as illustrated in Figure 9(b), can also be achieved for the other test cases considered in the following, if fifth-order SK
and WK elements are employed. A detailed investigation of general geometrically exact beam element formulations
of Kirchhoff-Love type with polynomial degree k > 3 will be considered in future research work. Eventually, also
the convergence behavior of the fourth load case shall be investigated. In Figures 10(a) and 10(b) the L2 -error of the
CJ, the WK-TAN and the SK-TAN element is plotted for the two beam slenderness ratios ζ = 100 and ζ = 10000.
Since no closed-form analytic solution has been available for this example, a numerical reference solution based on
the element formulation of Crisfield and Jelenić has been employed. As a consequence of shear deformation induced
by the tip force F̃, the result derived from the Reissner and Kirchhoff type beam element formulations will differ
in the limit of very fine discretizations h → 0. This ”model error” of the shear-free Kirchhoff elements becomes
visible in form of a kink in the convergence diagram and a certain cutoff error level that remains constant even for
arbitrarily fine discretizations. As expected, the model difference between the Simo-Reissner and the Kirchhoff-Love
beam theory decreases with increasing beam slenderness ratio, a property that is reflected by a lower cutoff error level
for the higher slenderness ratio ζ = 10000. For the lower slenderness ratio ζ = 100, the relative error distinguishing the
61
10
10
0
10 -2
10 -4
10 -4
-6
||e||
10
2
rel
10 -2
2
rel
||e||
0
10 -8
10
-10
10
-6
10 -8
Crisfield & Jelenic
WK-TAN
SK-TAN
10
Crisfield & Jelenic
WK-TAN
SK-TAN
-10
4 th order
10 -12
10 1
4 th order
10 2
10 3
degrees of freedom
10 -12
10 1
10 4
(a) Reference: Crisfield & Jelenić, ζ = 100.
10 2
10 3
degrees of freedom
10 4
(b) Reference: Crisfield & Jelenić, ζ = 10000.
Figure 10: Load case M̃+F̃: L2 -error for different element formulations.
0
10
0
10 -2
10 -4
10
10
e,rel
10 -2
-6
||e||
||e||
e,rel
10
10 -8
10
-10
10 -12
10 1
-4
10 -6
10 -8
WK-TAN
SK-TAN
4 th order
10
10 2
10 3
degrees of freedom
10 4
(a) Reference: WK-TAN, ζ = 100.
-10
10 -12
10 1
WK-TAN
SK-TAN
4 th order
10 2
10 3
degrees of freedom
10 4
(b) Reference: WK-TAN, ζ = 10000.
Figure 11: Load case M̃+F̃: Energy-error for different element formulations.
Kirchhoff from the Reissner solution lies below 10−3 , which can be assumed as reasonable approximation for many
engineering applications. For the high slenderness ratio ζ = 10000, the relative error between these two models is
smaller than 10−7 . For the investigated cases of ζ = 100 and ζ = 10000, the cutoff error scales almost quadratically
with the slenderness ratio, which would be the expected result for the solution of the geometrically linear theory. This
result is remarkable for this highly nonlinear example. From Figures 10(a) and 10(b), it can again be observed that all
element formulations exhibit the expected convergence rate of four, that the error level of the SK-TAN element lies
slightly above and that the error level of the WK-TAN element lies below the error level of the CJ element. Despite
the fact that the Simo-Reissner formulations yield the more general solutions, which also contain the effects of shear
deformation, the Kirchhoff type WK-TAN element formulation will be considered as numerical reference solution for
all remaining examples throughout this work. Such a procedure seems to be sensible, since within this contribution,
the convergence behavior of the Kirchhoff type elements and not of the Reissner type elements shall be studied.
Nevertheless, the model error of the Kirchhoff type beam elements is still observable, but this time in form of a kink
and a remaining cutoff error level in the convergence plots of the Reissner type formulation. While the WK-TAN
element was able to exactly represent the internal energy of the load case M̃, now, for the load case M̃+F̃, the energy
62
2000
2000
Crisfield & Jelenic
WK-TAN
SK-TAN
WK-ROT
SK-ROT
1500
Newton iterations
Newton iterations
1500
Crisfield & Jelenic
WK-TAN
SK-TAN
WK-ROT
SK-ROT
1000
500
0
1000
500
0
0
20
40
60
80
100
number of elements
120
140
0
(a) Moderate slenderness ratio: ζ = 100.
20
40
60
80
100
number of elements
120
140
(b) High slenderness ratio: ζ = 10000.
Figure 12: Load case M̃+F̃: Total number of Newton iterations.
convergence of the formulations WK-TAN and SK-TAN can be compared (see Figures 11(a) and 11(b) for the two
slenderness ratios ζ = 100 and ζ = 10000). Both element formulations exhibit the expected convergence order of four,
and similar to the load case M̃, the WK-TAN element yields a better approximation of the internal energy than the
SK-TAN element. This is observable in form of a lower energy error level in Figures 11(a) and 11(b) and the reason
for the lower L2 -error level visible in Figures 10(a) and 10(b). A possible explanation for the better performance
of the WK-TAN element may be found by considering the interaction of the employed translational and rotational
interpolation schemes: As shown in Section 10.4, the number of unknowns equals the number of equations required
for the WK-TAN element to (energetically) represent a pure bending state. This means that an exact representation
of the internal energy associated with states of constant axial tension, bending curvature and torsion is possible. This
property does not hold for the SK-TAN element (see e.g. Section 9.3), where the corresponding system of equations
that has to be fulfilled for representing a pure bending state is slightly over-constrained. While the difference between
the WK and the SK elements in the L2 -error level is expected to vanish with higher-order trial functions, such an
effect can in general not be predicted for the energy error. Eventually, it has to be mentioned that for both the L2 -error
as well as the energy error plots of the load case M̃+F̃, no difference between the slenderness ratios ζ = 100 and
ζ = 10000 is evident, which again underlines the successful avoidance of membrane locking and the effectiveness of
the MCS method. Finally, also the performance of the Newton-Raphson scheme shall be investigated and compared
between Kirchhoff and Reissner type element formulations (see Figure 12). Since the computationally expensive steps
of solving a nonlinear system of equations and evaluating the tangent stiffness matrix have to be conducted in every
Newton iteration, a reduction in the total number of Newton iterations niter,tot as defined in (181) would considerably
increase the overall efficiency of the numerical algorithm. In Figures 12(a) and 12(b), the total number of Newton
iterations niter,tot of the load case M̃+F̃ in combination with slenderness ratios of ζ = 100 and ζ = 10000 has been plotted
for the element formulations CJ, WK-TAN, SK-TAN, WK-ROT and SK-ROT and different spatial discretizations.
While the final FEM solutions have been shown to be independent from the choice of nodal rotation parametrization,
the number of Newton iterations required for the SK/WK-ROT and SK/WK-TAN variants might differ considerably.
Therefore, also the Newton performance of these variants has been investigated. For solving the highly nonlinear
beam problem, the load step adaption scheme mentioned above based on an initial number of N0 = 2 load steps has
been employed. By comparing Figures 12(a) and 12(b), one realizes that the Newton performance of the Kirchhoff
type element formulations is rarely influenced by the considered slenderness ratio, while the number of Newton
iterations required by the Reissner element increases drastically with increasing slenderness ratio. Furthermore, it
seems that the SK/WK-TAN variants require fewer Newton iterations than the SK/WK-ROT variants. These trends
will be confirmed, and even more pronounced, in more general 3D examples presented in subsequent sections. In
summary, the following conclusions can be drawn from the 2D examples considered in this section: 1) The proposed
Kirchhoff elements yield accurate results with acceptable model errors for slenderness ratios of ζ ≥ 100 and a model
63
error that decreases quadratically with increasing beam slenderness ratio. 2) The expected convergence orders of four
in the L2 - as well as in the energy error could be confirmed for all investigated Kirchhoff elements. 3) In combination
with the MCS method, none of the considered element formulations exhibited an influence of the element slenderness
ratio on the resulting discretization error. This result confirms the effectiveness of the MCS method in the avoidance of
membrane locking. 4) The discretization error level of the WK-TAN element lies below the error level of the Reissner
type element and also below the error level of the SK-TAN element. The increased error level of the SK-TAN element
has been shown to vanish with higher polynomial degree k > 3 of the trial functions. 5) While the total number of
Newton iterations required by the Reissner type element formulations considerably increases with increasing beam
slenderness ratio, the number of iterations remains more or less constant for the Kirchhoff type formulations. These
conclusions drawn from these 2D tests will be confirmed by the 3Dl examples investigated in the following sections.
11.3. Example 3: Pure bending in 3D
In this section, the 3D extension of the pure bending examples (load cases M and M̃) of the last section will be
considered. Again, the focus lies on an initially straight, clamped beam of standard length l = 1000 investigated for
the two slenderness ratios ζ = 100 and ζ = 10000. However, this time, the beam is loaded by a 3D end-moment
M1 := (M, 0, M)T , with M = 10 for ζ = 100 and M = 10−7 for ζ = 10000, which contains an additional moment component in beam length direction inducing torsion. The initial and deformed configuration are illustrated in Figure 13(a).
10
0
10 -2
||e||
2
rel
10
-4
10 -6
10 -8
10
-10
10 -12
10 1
Cri. & Jel., ζ=10
2
Cri. & Jel., ζ=10 4
WK-TAN, ζ=10 2
WK-TAN, ζ=10
SK-TAN, ζ=10
SK-TAN, ζ=10
4 th order
4
2
4
10 2
10 3
degrees of freedom
10 4
(b) reative L2 -error, reference: analytic.
(a) Initial and deformed geometry
Figure 13: Straight beam loaded with a 3-dimensional discrete end-moment.
As argued in [105], the analytic solution of this example is given by the following space curve representation:
1
√2 (sin β + β)
s
EI
r(s) = R0 (1 − cos β) with R0 =
and β = √
.
(182)
2M
√1 (β − sin β)
2R0
2
This solution represents a helix whose mid-axis points into the (1, 0, 1)T -direction, viz. in the direction of the applied
external moment. The special parameter choice of this example leads to a radius R0 of the enveloping cylinder that is
identical to the slope of the helix. In Figure 13(b), the relative L2 -error resulting from the two investigated slenderness
ratios is plotted for the element formulations CJ, WK-TAN and SK-TAN as well as spatial discretizations based on
8, 16, 32, 64, 128 and 256 elements. Again, all element formulations exhibit the expected convergence order of four
and the discretization error level of the WK-TAN element is lower, whereas the discretization error of the SK-TAN
element is slightly higher than for the CJ element. Furthermore, no visible difference can be observed between the
discretization error levels associated with the two different slenderness ratios. Due to the choice GIT = EI2 = EI3 ,
it can easily be verified that this example results in an analytic solution exhibiting vanishing axial tension and shear
64
15000
Crisfield & Jelenic
Simo & Vu Quoc
WK-TAN
SK-TAN
WK-ROT
SK-ROT
10000
Newton iterations
Newton iterations
15000
5000
0
10000
Crisfield & Jelenic
Simo & Vu Quoc
WK-TAN
SK-TAN
WK-ROT
SK-ROT
5000
0
0
20
40
60
80
100
120
140
0
number of elements
20
40
60
80
100
120
140
number of elements
(a) Moderate slenderness ratio: ζ = 100.
(b) High slenderness ratio: ζ = 10000.
Figure 14: Load case ”straight to helix“: Total number of Newton iterations.
deformation as well as a constant spatial and material curvature vector along the entire beam pointing into the direction of the external moment vector, i.e. k = K = M/(EI)(1, 0, 1)T = const. Thus, already the roughest discretizations
of the CJ and WK-TAN elements can exactly represent the hyperelastic stored energy function for this pure bending
case, which can be interpreted as a simple 3D patch test for geometrically exact beams. Finally, also the number
of Newton iterations shall be investigated (see Figure 14). In order to enable more general conclusions, this time, a
second Reissner type beam element formulation, which is based on a completely different triad interpolation scheme,
has additionally been included in the comparison. Concretely, this element represents an ”interpretation” formulated
by Crisfield [40] (see Chapter 17.2) of the original variant proposed by Simo and Vu-Quoc [134], in the following denoted as SV element. This time, the load step adaption scheme presented above based on an initial number of N0 = 10
load steps has been employed. Similarly to the 2D case considered in the last section, the Newton performance of
the Reissner type element formulations drastically deteriorates with increasing slenderness ratio whereas the performance of the Kirchhoff type elements remains unchanged (or is even slightly improved in case of the WK/SK-TAN
variants). Concretely, for the slenderness ratio ζ = 10000, all investigated discretizations of the WK/SK-TAN elements
exhibit a remarkably constant number of niter,tot = 140±4 iterations, while the total number of iterations required by
the WK/SK-ROT elements increases from niter,tot ≈ 500 to niter,tot ≈ 2000 with increasing number of elements and the
total number of iterations required by the Reissner type formulations is almost by two orders of magnitude higher
than for the WK/SK-TAN elements and lies constantly above a value of niter,tot > 12000. Seemingly, the considerable difference between the WK/SK-TAN elements and the Reissner type element formulations can be attributed to
two different effects: Firstly, the parametrization of nodal triads via tangent vectors seems to be more good-natured
than the parametrization based on nodal rotation vectors. This effect already gets visible as difference between the
WK/SK-TAN and the WK/SK-ROT variants and seems to be more or less independent from the beam slenderness
ratio. Secondly, the high stiffness contributions resulting from the shear mode seem to considerably deteriorate the
Newton convergence in the range of high slenderness ratios: This effect becomes obvious as difference between the
WK/SK-ROT elements and the Reissner type elements. The linearizations of all of these four elements are based on
multiplicative updates of nodal rotation vectors. This observation is emphasized by the two elements types WK-ROT
and CJ, which additionally exhibit the same triad interpolation. These two elements only differ in the centerline interpolation (based on Lagrange or Hermite polynomials), which is not expected to influence the Newton convergence
in such a drastic manner, and the fact that the WK-ROT element additionally enforces the constraint of vanishing
shear strains. Consequently, the avoidance of shear modes seems to be the main reason for the considerably improved
performance of the Kirchhoff type element formulations. Finally, the observation that the total number of Newton
iterations required by the WK/SK-ROT variants increases with increasing number of elements is only of secondary
practical interest since the discretizations relevant for practical applications are located in the range of small element
numbers (on the left of Figure 14(b)). The observations made so far, will be confirmed by the subsequent 3D examples.
65
Remark: For some of the discretizations investigated in Figure 14, solutions of the SK-TAN element
could already be found in one load step. However, since in these cases no convergence could be achieved
for simulations based on two or three load steps, the solution of the problem by means of one load
step can rather be regarded as a ”lucky shot” than as a representative convergence behavior. In order
to avoid a biased comparison resulting from such effects, the initial number of load steps has been
increased to N0 = 10. By this means and the mentioned load step adaption scheme, an evaluation and
comparison process is intended that is as fair and objective as possible. Nevertheless, this example
shows that an absolute statement concerning the robustness of the nonlinear solution scheme based
on a single example / discretization and a deliberately chosen ”good-natured” step size, as sometimes
done in the literature, is questionable. Here, the degree of arbitrariness is intended to be minimized by
employing an automated scheme for determining the optimal load step size, by comparing the results
of different test cases, different discretizations, different element types (here Reissner and Kirchhoff
type beam elements) as well as different representatives for each element type. In order to avoid biased
results as consequence of incorrect linearizations, the results of the Reissner type elements derived on
the basis of an analytic representation of the consistent tangent stiffness matrix have been verified by
simulations on the basis of a consistent tangent stiffness matrix derived via an automatic differentiation tool.
11.4. Example 4: Verification of path-independence
In Section 6.3.2, the fundamental property of objectivity has already been verified for the proposed Kirchhoff beam
elements. In this section, it will be shown that these element formulations are also path-independent, i.e. for beam
problems whose analytic solution is independent from a specific loading path, these beam element formulations also
yield a discrete solution that is independent from a specific loading path. As numerical test case for path independence, an initially straight clamped beam with initial length l = 1000 and slenderness ratio ζ = 100 (thus R = 10) is
considered that is loaded by an end-moment M = (0, 0, M)T , with the moment M = 4EIπ/l ≈ 10.47 being defined such
that it exactly bends the beam into a ”double-circle“, and an additional end-force F = (0, 0, F)T , with F = 0.01 ≈ M/l.
Again, for comparison reasons, also the case of an increased slenderness ratio ζ = 10000 with correspondingly adapted
loads M = 4EIπ/l ≈ 1.047 · 10−7 and F = 10−10 ≈ M/l will be investigated. The problem setup as well as the deformed
configuration for this example are shown in Figure 15. In the following, two different possibilities how to apply these
tip loads are investigated: In a first load case, the moment and the force are applied simultaneously (load case ”sim“),
while, in a second load case, the moment and the force are applied successively (load case ”suc“). In the latter case,
the external moment is increased linearly from zero to M within the pseudo-time interval t ∈ [0; 0.5], whereas the
external force is increased linearly from zero to F within the pseudo-time interval t ∈ [0.5; 1.0].
Figure 15: Problem setup: Initially straight beam bent by a discrete end-moment and -force.
The deformed shapes of both load cases have been plotted for the pseudo-time steps t = 0.25, t = 0.5, t = 0.75 and
t = 1.0 in Figure 16. Apparently, the two load cases lead to different deformation paths, but to an identical final
configuration. It contradicts intuition that this final deformed configuration lies completely in the half space with
z ≤ 0, although the tip force points into the positive z-direction. For the case of small forces F, this observation has
been verified by deriving an analytical solution based on a linearization of the equilibrium equations with respect
to the double-circle configuration resulting from the end-moment M (see also [105]). Furthermore, this observation
66
(a) Step t=0.25 (sim).
(b) Step t=0.50 (sim).
(c) Step t=0.75 (sim).
(d) Step t=1.0 (sim).
(e) Step t=0.25 (suc).
(f) Step t=0.50 (suc).
(g) Step t=0.75 (suc).
(h) Step t=1.0 (suc).
Figure 16: Deformed configurations for simultaneous (sim) and successive (suc) loading of an straight beam with a moment M and a force F.
is in agreement with the results obtained in [72] and [13], where a similar example based on a slightly modified
parameter choice has been analyzed. In order to investigate possible path dependence effects also in a quantitative
manner, the relative L2 -error has been calculated between the solution rh,suc of the load case ”suc“ for a certain
discretization and the solution rh,sim of the load case ”sim“ for the same centerline discretization. Thus, basically
the relative L2 -error definition of equation (143) has been applied, with rh = rh,suc and rre f = rh,sim . The results
obtained for the two different slenderness ratios and the investigated element formulations CJ, SK-TAN and WKTAN are illustrated in Figure 17. Accordingly, for all investigated element types, discretizations and slenderness
ratios, this error vanishes up to machine precision, which verifies the path independence of these formulations. For
completeness, Figure 18 represents the discretization error resulting from the CJ, SK-TAN and WK-TAN elements for
the considered slenderness ratios of ζ = 100 and ζ = 10000. There, the observations already made in earlier examples
10
10
0
10
Crisfield & Jelenic
WK-TAN
SK-TAN
-2
10 -4
-6
10
2
rel
10 -8
-6
10 -8
10
-10
10
-12
10 -12
10 -14
10 -14
10
-16
10 -18
10
10
0
20
40
60
80
100
number of elements
Crisfield & Jelenic
WK-TAN
SK-TAN
-2
10 -4
||e||
||e||
2
rel
10
10
0
120
-10
-16
10 -18
140
(a) Moderate slenderness ratio: ζ = 100.
0
20
40
60
80
100
number of elements
(b) High slenderness ratio: ζ = 10000.
Figure 17: L2 -error between the load cases ”Simultaneously” and ”Successively”.
67
120
140
10
0
10 -2
||e||
2
rel
10 -4
10
-6
10 -8
10
-10
Crisfield & Jelenic
WK-TAN
SK-TAN
Isotropic
4 th order
10 -12
10 1
10 2
10 3
degrees of freedom
10 4
(a) Reference: WK-TAN, ζ = 100.
(b) Reference: WK-TAN, ζ = 10000.
Figure 18: Path independence: L2 -error for different element formulations.
with respect to convergence rate, discretization error level and cutoff error between Kirchhoff and Reissner type
element formulations are confirmed. For comparison reasons, also the L2 -error resulting from the reduced isotropic
beam element formulation proposed in [105] has been plotted. The result is very similar to the SK-TAN element.
The only reason why the discretization error level is slightly lower for the isotropic than for the SK-TAN element
lies in the twist interpolation, which only requires two DoFs for the isotropic element but three DoFs for the SKTAN element. In Section 6.2.4, the importance of a consistent torsion (114) of the intermediate triad field has been
emphasized. Now, in Figure 18(b), the discretization error for a variant (SK-TAN no K M1 ) has been plotted, where
exactly this torsion term has been neglected. Surprisingly, the resulting discretization error level is identical to the
”correct” SK-TAN element formulation. How can this contradiction be explained? In order to answer this question, it
has to be realized that the actual triad orientation is not important for isotropic examples in order to yield a consistent
centerline convergence. It can easily be verified that only the mechanical torsion has to be represented correctly, which
is also the functional principle of the isotropic beam element. If the torsion of the intermediate triad field is neglected,
the total torsion is solely represented by the derivative of the relative angle field ϕ(ξ). Consequently, the relative
angle arises in a way such that the total torsion is represented exactly, which in turn results in an inconsistent triad
orientation. However, since for isotropic beams, only the torsion, but not the triad orientation, enters the weak form,
the final result for the beam centerline is correct. Later in Section 11.5, it will be shown that the situation changes
for anisotropic beams, i.e. beams with initial curvature or with anisotropic cross-section shapes. There, the neglect
of the intermediate triad torsion will indeed lead to an inconsistent centerline solution resulting in a decreased spatial
convergence rate. Furthermore, this investigation explains why certain Kirchhoff element formulations available in the
literature, which accidentally neglect this torsion term, nevertheless produce correct results and consistent convergence
rates for the centerline solution as long as isotropic beam problems are considered. Finally, in Figure 19, the total
number of Newton iterations required by the different finite element formulations is plotted for the load case ”sim”
and the two investigated slenderness ratios. Here, the load step adaption scheme based on N0 = 10 has been employed.
The obtained results are very similar to the last section: The Newton performance of the Reissner type element
formulations drastically deteriorates with increasing slenderness ratio whereas the performance of the Kirchhoff type
elements remains unchanged. For the slenderness ratio ζ = 10000, all investigated discretizations of the WK/SK-TAN
elements exhibit a remarkably constant number of niter,tot = 107 ± 1 iterations, while the total number of iterations
required by the WK/SK-ROT elements increases from niter,tot ≈ 800 to niter,tot ≈ 2500 with increasing number of
elements. Again, the total number of iterations required by the Reissner type beam element formulations is almost by
two orders of magnitude higher than for the WK/SK-TAN elements and lies constantly above a value of niter,tot > 6000
for the CJ element formulation and niter,tot > 7000 for the SV element formulation.
68
8000
8000
Crisfield & Jelenic
Simo & Vu Quoc
WK-TAN
SK-TAN
WK-ROT
SK-ROT
Newton iterations
6000
5000
7000
6000
Newton iterations
7000
4000
3000
2000
1000
Crisfield & Jelenic
Simo & Vu Quoc
WK-TAN
SK-TAN
WK-ROT
SK-ROT
5000
4000
3000
2000
1000
0
0
0
20
40
60
80
100
120
140
0
number of elements
20
40
60
80
100
120
140
number of elements
(a) Moderate slenderness ratio: ζ = 100.
(b) High slenderness ratio: ζ = 10000.
Figure 19: Load case ”Simultaneously“: Total number of Newton iterations.
11.5. Example 5: Arc-segment with out-of-plane load
Besides the objectivity test in Section 11.1, all examples investigated in the previous sections were based on isotropic
geometries, i.e. straight beams with quasi-circular cross-sections EI2 = EI3 . Now, an initially curved beam will be
considered. The initial geometry is represented by a 45◦ -degree circular arc-segment with curvature radius r0 = 100
that lies completely in the global x-y-plane and that is clamped at one end. The section constitutive parameters of
the beam result from a quadratic cross-section shape with side length R = 1 and a Young’s modulus of E = 107 as
well as a shear modulus of G = 0.5 · 107 . This initial geometry is loaded by an out-of-plane force f = (0, 0, fz )T in
global z-direction with magnitude fz = 600. This example has initially been proposed by Bathe and Bolourchi [11]
and can meanwhile be considered as standard benchmark test for geometrically exact beam element formulations
that has been investigated by many authors (see e.g. [134, 34, 71, 50, 39, 80, 44, 127, 51, 117, 118, 10]). While the
original definition of the slenderness ratio yields a value of ζ = l/R = 100π/4 for this example, a slightly modified
definition of the slenderness ratio according to ζ̃ = r0 /R = 100 is employed in the following. For comparison reasons,
also a second variant of this example with increased slenderness ratio ζ̃ = r0 /R = 10000, i.e. R = 0.01, and adapted
Figure 20: Arc-segment with out-of-plane force: Initial (red) and final (green) configuration.
69
Formulation
Crisfield & Jelenić
Crisfield & Jelenić
Simo & Vu-Quoc
WK-TAN
WK-TAN
SK-TAN
# Elements
32
8
8
32
8
8
u x (l)
47.15044
47.15044
47.14634
47.15215
47.15178
47.15201
uy (l)
15.68480
15.68480
15.69146
15.68535
15.68510
15.68557
uz (l)
53.47486
53.47486
53.47362
53.47176
53.47225
53.47216
Table 4: Case ζ̃ = 100: tip displacement and relative error for different formulations.
Formulation
Crisfield & Jelenić
Crisfield & Jelenić
Simo & Vu-Quoc.
WK-TAN
WK-TAN
SK-TAN
# Elements
32
8
8
32
8
8
u x (l)
47.15129
47.15129
47.14719
47.15129
47.15093
47.15115
uy (l)
15.68508
15.68508
15.69174
15.68508
15.68482
15.68530
uz (l)
53.46860
53.46860
53.46736
53.46860
53.46908
53.46900
Table 5: Case ζ̃ = 10000: tip displacement and relative error for different formulations.
force fz = 6 · 10−6 will be investigated. The initial and deformed geometry are illustrated in Figure 20. In Tables 4
and 5, the tip displacements resulting from the two slenderness ratios and different discretizations with WK-TAN and
SK-TAN Kirchhoff type elements as well as with Reissner type elements of Crisfield and Jelenić and Simo and VuQuoc are plotted. Due to rough spatial discretizations, and in some cases also due to additional model simplifications,
the corresponding values derived in the literature for the case ζ̃ = 100 show a comparatively large variation. On the
contrary, the deviation in the results displayed in Table 4 is smaller than 0.1% for all investigated formulations. The
fact that these results have been derived by representatives of different beam theories, i.e. of the Simo-Reissner and
of the Kirchhoff-Love theory, indicates their correctness. While the Reissner and Kirchhoff values resulting from
a discretization with 32 elements coincide up to the fourth significant digit for the case ζ̃ = 100, the corresponding
values are identical in all seven significant digits displayed for the case of the high slenderness ratio ζ̃ = 10000. The
observations described above are also confirmed by the convergence plots in Figure 21. All formulations yield the
expected convergence orders, and, similar to the last example, the SK-TAN element exhibits an identical discretization
10
0
10 -2
||e||
2
rel
10
-4
10 -6
10 -8
10
-10
Crisfield & Jelenic
WK-TAN
SK-TAN
4 th order
10 -12
10 1
10 2
10 3
degrees of freedom
10 4
(a) Reference: WK-TAN, ζ̃ = 100.
(b) Reference: WK-TAN, ζ̃ = 10000.
Figure 21: Arc-segment with out-of-plane force: L2 -error for different formulations.
70
500
500
Crisfield & Jelenic
Simo & Vu Quoc
WK-TAN
SK-TAN
WK-ROT
SK-ROT
300
400
Newton iterations
Newton iterations
400
200
100
300
Crisfield & Jelenic
Simo & Vu Quoc
WK-TAN
SK-TAN
WK-ROT
SK-ROT
200
100
0
0
0
5
10
15
0
number of elements
5
10
15
number of elements
(a) Number of Newton iterations for ζ̃ = 100.
(b) Number of Newton iterations for ζ̃ = 10000.
(c) Number of load steps for ζ̃ = 100.
(d) Number of load steps for ζ̃ = 10000.
Figure 22: Arc-segment with out-of-plane force: Number of Newton iterations and load steps.
error level as the WK-TAN element, since again no multiple centerline loops are involved for this example. Furthermore, similar to the example of Section 11.3, also a variant (SK-TAN no K M1 ) has been investigated where the torsion
of the intermediate triad field has been omitted. While the omission of this term did not influence the convergence
order observed in Section 11.3, this inconsistency yields a decline in the convergence rate from four to two for the
anisotropic example considered here. This underlines the importance of consistently considering this term (see also
Section 11.3 for further explanation). Also for this example, the performance of the Newton-Raphson scheme will
be evaluated. However, in order to enable a comparison with the values available in the literature, this time, no load
step adaption scheme is employed, but the following alternative procedure in order to determine the maximal constant
load step size ∆t =const.: Starting with a scheme based on one load step N0 = 1, the number of load steps is increased
by one, i.e. Nnew = Nold +1, in the range N = 1, ..., 10 and increased by increments of 10, i.e. Nnew = Nold +10, in the
range N > 10, until Newton convergence is achieved for all load steps. In order to avoid ”lucky shots” (see the remark
at the end of this section and also the remark in Section 11.3), a maximal load step size and an associated minimal
number of load steps Nmin is accepted only if also the next incrementation step of the load step size according to the
procedure described above leads to Newton convergence for all load steps. In Figure 22, the total number of Newton
iterations as well as the minimal number of load steps Nmin resulting from the maximal constant load step size are
plotted for the two different slenderness ratios. The results are similar to the observations made in previous sections,
however, with a smaller difference between the SK/WK-TAN elements and the SK/WK-ROT elements. Concretely,
71
Reference
[44]
[134]
[34]
[127]
[127]
[11]
[39]
[80]
CJ
CJ
SV
SV
SK-TAN
WK-TAN
SK-ROT
WK-ROT
Elements
8 first-order
8 first-order
8 first-order
8 first-order
8 first-order
8 first-order
8 first-order
8 first-order
8 first-order
8 third-order
8 first-order
8 third-order
8 third-order
8 third-order
8 third-order
8 third-order
Nmin
6
3
6
3
2
60
3
1
7
7
7
7
1
1
1
2
niter,tot
27
47
30
11
16
4
59
58
52
58
8
8
24
30
Remark
Number of iterations has not been reported.
Application of a standard Newton scheme.
Application of an accelerated Newton scheme.
Number of iterations has not been reported.
Non-objective variant proposed in [81].
Objective variant proposed in [80].
Objective variant proposed in [80].
Interpretation of the formulation of [134].
Interpretation of the formulation of [134].
Kirchhoff type beam element formulation.
Kirchhoff type beam element formulation.
Kirchhoff type beam element formulation.
Kirchhoff type beam element formulation.
Table 6: Case ζ̃ = 100: Number of load steps and Newton iterations from the literature (top) and from this work (bottom).
the beam problem with slenderness ratio ζ̃ = 100 is solved in 1 load step and a total of 8 iterations for the SK/WKTAN discretizations, in 1-3 load steps and a total of 10-40 iterations for the SK/WK-ROT discretizations and in 7 load
steps and a total of 57-58 iterations for the discretizations based on Reissner type elements. For the slenderness ratio
ζ̃ = 10000, the problem is again solved in 1 load step and a total of 8 iterations for the SK/WK-TAN discretizations,
in 1-4 load steps an a total of 10-50 iterations for the SK/WK-ROT discretizations and in 30-60 load steps and a total
of 350-450 iterations for the Reissner discretizations. In Table 6, the corresponding values reported in the literature
for the slenderness ratio ζ̃ = 100 are summarized. As already mentioned earlier, a direct comparison of these results is
difficult since it is not clear which procedure has been applied by the different authors in order to determine the minimal number of Newton iterations (e.g. if it was required that also ”subsequent refinement steps” have to be convergent
or if, on the contrary, also singular occurrences of convergence for special, good-natured loading paths were accepted).
Nevertheless, the numbers summarized in Table 6 should at least give a first impression on the behavior of the NewtonRaphon scheme resulting from different finite element formulations. Accordingly, only a few formulations can solve
the problem in less than 20 iterations. Furthermore, for the case ζ̃ = 100, only the Reissner type formulation proposed
in [81] and investigated in [80] yields a lower number of Newton iterations than the SK/WK-TAN elements. However,
as shown in [80], this beam element formulation is non-objective and path-dependent. Moreover, for all examples investigated so far, the real advantage of the Kirchhoff type formulations occurred especially for the high slenderness
ratio ζ̃ = 10000, which has not been investigated in the literature. For the range of moderate and high slenderness
ratios, it can be concluded that the proposed Kirchhoff beam elements can be considered as very robust and efficient
formulations as compared to many (Reissner type) alternatives from the literature.
Remark: Maybe the reader is wondering why the SV element formulation, which is an interpretation of the
element formulation proposed in [134], based on an identical discretization with eight first-order elements
as investigated in [134], required more Newton iterations than reported in that reference. Actually, also in
the numerical tests performed here, the nonlinear problem resulting from a discretization with eight firstorder SV elements could be solved in three load steps. However, since a subsequent simulation based on
four load steps was not convergent, the procedure for the avoidance of ”lucky shots” as explained above has
been applied, thus leading to a total of 7 load steps and 52 Newton iterations.
72
11.6. Example 6: Helix loaded with axial force
In this example, the generality of the initial geometry shall be further increased: A helix with linearly increasing slope,
clamped at one of its ends, is loaded with a end-force F = (0, 0, F)T (see Figure 23 for illustration).
Figure 23: Helix with varying slope loaded with discrete force: Problem setup.
The space curve representing initial geometry of the helix can be described via the following analytic representation:
sin β
l
r0 (β) = R0 cos β−1 , R0 = q
(183)
q ! ≈ 34.36.
6 2
2
27π2
4
4 2
β
6 3π
+1+
ln
+
1+
81π2
4
8
3π
3π
The radius R0 of the enveloping cylinder of the helix is chosen such that the helix exactly consists of 4.5 loops, i.e.
β ∈ [0; 9π], along the standard length of l = 1000. Also this example is investigated for two different slenderness ratios
ζ = 100 and ζ = 10000 with associated axial forces F = 2 · 10−1 as well as F = 2 · 10−9 . The ratio of these forces is
chosen identical to the ratio of the bending stiffnesses of the cases ζ1 = 100 and ζ2 = 10000 leading to the comparable
values uz,max,1 ≈ 267 and uz,max,2 ≈ 266 for the maximal tip-displacement in z-direction. In Figure 24, the resulting
relative L2 -error of the CJ, SK-TAN and WK-TAN element is plotted for discretizations with 16, 32, 64, 128, 256 and
512 elements. All element formulations show the expected convergence rate of four, the discretization error level
of the WK-TAN element is slightly lower than the discretization error level of the SK-TAN element. However, both
discretization error levels lie below the error level of the Reissner type CJ element. In the authors’ former contribution,
also the different anti-locking approaches already investigated for the 2D case in Section 11.2.1 have been compared
for this 3D helix example. The derived results and the drawn conclusion are similar as for this 2D example. For further
details on this comparison, the interested reader is referred to this reference. Finally, the helix problem shall be even
10
10
0
10 -2
10 -4
10
-6
||e||
10
2
rel
10 -2
2
rel
||e||
0
10 -8
10
-10
-4
10 -6
10 -8
Crisfield & Jelenic
WK-TAN
SK-TAN
10
-10
10
-12
4 th order
10
4 th order
-12
10 1
Crisfield & Jelenic
WK-TAN
SK-TAN
10 2
10 3
degrees of freedom
10 4
(a) Reference: WK-TAN, ζ = 100.
10 1
10 2
10 3
degrees of freedom
(b) Reference: WK-TAN, ζ = 10000.
Figure 24: Helix loaded with axial force: L2 -error for different formulations.
73
10 4
Figure 25: Twisted helix with axial force: Initial and final shape for ζ = 100 and ζ = 1000 (visualization based on ζ = 1000).
more generalized by not only accounting for 3D initial curvatures but also for anisotropic cross-section shapes with
initial twist as illustrated in Figure 25. Again, the two slenderness ratios ζ = 100 and ζ = 10000 have been investigated.
The square cross-section of the last example is extended to a rectangular cross-section shape with dimensions b = 10
and h = 5 and an assumed torsional moment of inertia IT ≈ 3.2875·10−2 for the case ζ = 100. The case ζ = 10000 is
defined by b = 0.1 and h = 0.05 and an assumed torsional moment of inertia IT ≈ 3.2875·10−6 . The external forces have
been chosen as F = 5·10−2 for ζ = 100 and F = 5 · 10−10 for ζ = 10000 and the initial twist as one twist rotation per helix
loop. The resulting L2 -error plotted in Figure 26 again shows a consistent convergence behavior similar to Figure 24.
Additionally, in Figure 25, also the Bubnov-Galerkin variant (SK-TAN+CS) of the SK-TAN element with consistent
spin vector interpolation has been plotted. Accordingly, no visible difference compared to the Petrov-Galerkin (SKTAN) variant can be observed. In a last step, also the balances of forces and moments are investigated for this most
general example of Figure 25. In Table 7, the reaction forces and moments at the clamped end of the helix at s = 0 and
the force and moment contributions (with respect to the point s = 0) resulting from the external load applied at s = l
are plotted for discretizations with eight elements. It can easily be verified that the balance of forces and moments
is exactly fulfilled by the variants CJ, WK-TAN and SK-TAN+CS, while the Petrov-Galerkin variant SK-TAN only
fulfills the balance of forces but not the balances of moments. This confirms the prediction made in Section 9.4.
74
10
10 -2
10 -2
-4
2
rel
10
10 -6
10
10
-8
-10
||e||
2
rel
10
||e||
0
10 0
-4
10 -6
10 -8
Crisfield & Jelenic
WK-TAN
SK-TAN
SK-TAN+CS
10
-10
4 th order
4 th order
10 -12
10 1
Crisfield & Jelenic
WK-TAN
SK-TAN
SK-TAN+CS
10 2
10 3
degrees of freedom
10 4
(a) Reference: WK-TAN, ζ = 100.
10 -12
10 1
10 2
10 3
degrees of freedom
10 4
(b) Reference: WK-TAN, ζ = 10000.
Figure 26: Twisted helix loaded with axial force: L2 -error for different formulations.
F1 (0)
F1 (l)
F2 (0)
F2 (l)
F3 (0)
F3 (l)
M1 (0)
M1 (l)
M2 (0)
M2 (l)
M3 (0)
M3 (l)
Cri. & Jel.
0.0
0.0
0.0
0.0
5.00000000E-6
5.00000000E-6
-1.64350142E-4
-1.64350142E-4
-4.41005618E-5
-4.41005617E-5
0.0
0.0
WK-TAN
0.0
0.0
0.0
0.0
5.00000000E-6
5.00000000E-6
-1.54617971E-4
-1.54617971E-4
-8.55776851E-6
-8.55776851E-6
0.0
0.0
SK-TAN
0.0
0.0
0.0
0.0
5.00000000E-6
5.00000000E-6
-1.51929992E-4
-1.52158798E-4
-1.13519642E-5
-2.30787384E-5
0.0
0.0
SK-TAN+CS
0.0
0.0
0.0
0.0
5.00000000E-6
5.00000000E-6
-1.65509346E-4
-1.65509346E-4
-5.43051999E-6
-5.43051999E-6
0.0
0.0
Table 7: Slenderness ζ = 10000: Reaction forces and moments for different formulations.
11.7. Example 7: Free oscillations of an elbow cantilever
The final example represents a dynamic test case. The example has initially been investigated in [135] and subsequently been considered in several contributions in the field of geometrically nonlinear beam element formulations
(see e.g. [80] and [32]). A right-angle elbow cantilever beam consists of two straight beam segments of length l = 10
being rigidly connected at one of their ends. In the initial configuration, the first segment points into global y-direction
and the second segment into global x-direction. The first segment of the cantilever is clamped at the position s = 0.
In the original work, the cross-section of the beam is described by the section parameters EA = GA2 = GA3 = 106 ,
GIT = EI2 = EI3 = 103 as well as ρA = 1 and ρIP /2 = ρI2 = ρI3 = 10. It can easily be verified that these section parameters
can for example be represented by a quadratic cross-section with radius R ≈ 0.1, thus resulting in a slenderness ratio
of ζ ≈ 100 for each of the two segments, for which the rotational inertia contributions are additionally scaled by a
factor of 104 . This artificial scaling has been applied in order to emphasize (and properly verify) the rotational inertia
contributions, which would otherwise be comparatively small for the chosen slenderness ratio. This cantilever beam
is loaded by a discrete force F = (0, 0, Fz )T pointing in global z-direction and acting at the rigid corner of the elbow.
The magnitude of the force is linearly increased from Fz = 0 at t = 0 to Fz = 50 at t = 1 and then linearly decreased
to Fz = 0 at t = 2. In the remaining simulation time until T = 50, the cantilever executes geometrically nonlinear free
oscillations in 3D space with amplitudes being in the range of the segment lengths. The initial and deformed configurations at different time steps are illustrated in Figure 27. For time integration, the modified generalized-α scheme
75
(a) Step 0.
(b) Step 5.
(c) Step 10.
(d) Step 15.
(e) Step 20.
(f) Step 25.
(g) Step 30.
(h) Step 35.
(i) Step 40.
(j) Step 45.
Figure 27: Free oscillations of an elbow cantilever: Initial and deformed configurations.
of Section 5 with ρ∞ = 0.95 has been employed. As spatial discretizations, the CJ element, the WK-ROT element
as well as the Petrov-Galerkin variant SK-ROT and the Bubnov-Galerkin variant SK-ROT+CS with consistent spin
vector interpolation according to (161) have been applied. This example yields a further complexity that has not been
present in the previous test cases: The modeling of a rigid beam connection occurring at the corner of the elbow. As
already mentioned in earlier sections, the modeling of such kinks in the beam centerline is easier to be realized by
the SK/WK-ROT variants. Nevertheless, the resulting solutions are identical as for the SK/WK-TAN variants. In Figure 28, the internal, kinetic and total system energy is plotted for different time step sizes and spatial discretizations.
For discretizations with 8 WK-ROT and SK-ROT elements per elbow segment and a time step size of ∆t = 0.01 as
illustrated in Figures 28(a) and 28(b), no visible oscillations in the total system energy and no visible differences in
the energy contributions of the Reissner and Kirchhoff type elements can be observed. In Figures 28(c)- 28(f), the
energy contributions resulting from a larger time step size of ∆t = 0.25 and a rougher spatial discretization based on
one beam element per elbow segment has been plotted. Accordingly, also for these rough discretizations, the overall
system energy is preserved very well. The fact that the total system energy approximation resulting from the (energetically consistent) Bubnov-Galerkin discretization SK-ROT+CS is rarely improved as compared to the Petrov-Galerkin
variants CJ, WK-ROT and SK-ROT indicates that for the chosen spatial and temporal discretizations, the influence of
the temporal discretization error might dominate the error in the total system energy. As compared to reference [80],
where an identical time step size of ∆t = 0.25 and a comparable spatial discretization consisting of one second-order
CJ element per elbow segment has been applied, the oscillations of the total system energy could be considerably
decreased and the instability observed there at the end of the considered time interval could be completely avoided.
This improvement can be attributed to the applied Lie group extension of the generalized-α scheme of Section 5,
whereas in reference [80] the Lie group extension of a Newmark time integration scheme as proposed in [134] has
been considered. Based on (simplified) Reissner type beam element formulation (see [98]), similar observations have
already been made in reference [32], where the Lie group extension of the generalized-α scheme has been proposed.
76
(a) WK-ROT, 16 elements, ∆t = 0.01.
(b) SK-ROT, 16 elements, ∆t = 0.01.
(c) Cri. & Jel., 2 elements, ∆t = 0.25.
(d) WK-ROT, 2 elements, ∆t = 0.25.
(e) SK-ROT, 2 elements, ∆t = 0.25.
(f) SK-ROT+CS, 2 elements, ∆t = 0.25.
Figure 28: Free oscillations of an elbow cantilever: Energy conservation.
77
12. Conclusion
The current work focused on geometrically exact finite elements for highly slender beams. It aimed at the proposal of
novel formulations of Kirchhoff-Love type, a detailed review of existing formulations of Kirchhoff-Love and SimoReissner type and a careful evaluation and comparison of the proposed and existing formulations. In the authors’
recent contribution [104], the first 3D large-deformation geometrically exact Kirchhoff-Love beam element formulation that fulfills the essential properties of objectivity and path-independence and that is capable of representing
arbitrary initial curvatures and anisotropic cross-section shapes has been proposed. The developed beam element formulation incorporates the modes of axial tension, torsion and anisotropic bending. In the subsequent work [105], also
the important question of membrane locking has successfully been addressed. The current contribution has extended
these methodologies by providing considerable improvements in terms of accuracy and practical applicability as well
as a generalization to dynamic problems. Thereto, two alternative interpolation schemes have been proposed: i) The
first scheme is based on a strong enforcement of the Kirchhoff constraint (SK) enabled by a novel orthonormal rotation
interpolation scheme. ii) The space-continuous theory of the second scheme is based on a weak enforcement of the
Kirchhoff constraint (WK). The discrete realization of the Kirchhoff constraint relies on a properly chosen collocation
strategy which can entirely abstain from the use of additional Lagrange multipliers. While this second formulation
allows for arbitrary rotation interpolations, the investigated numerical realization has employed the well-known orthonormal, geodesic triad interpolation scheme proposed in [43]. Furthermore, for each of these two interpolation
schemes, two different sets of nodal rotation parametrizations are proposed, one based on nodal rotation vectors
(ROT) and one on nodal tangent vectors (TAN). While these different choices have been shown to yield identical
FEM solutions, they differ in the resulting performance of nonlinear solvers and in the effort required for prescribing
essential boundary and transition conditions such as rigid joints. The four finite element formulations resulting from
a combination of the two interpolation schemes and the two choices of nodal primary variables have been denoted as
SK-TAN, SK-ROT, WK-TAN and WK-ROT elements, respectively. Taking advantage of a third-order Hermite interpolation, both element formulations provide a C 1 -continuous centerline representation. In order to avoid membrane
locking effects in the regime of high beam slenderness ratios, the concept of Minimally Constrained Strains (MCS)
recently proposed in [105] has been employed for re-interpolation of the axial tension field. Eventually, the proposed
beam elements are supplemented by an implicit, second-order accurate time integration scheme recently proposed in
the literature for time discretization of large rotations. This integration scheme can be identified as Lie group extension
of the well-known generalized-α method with comparable properties. The generality and flexibility of this scheme
allow for a straightforward combination with the different element formulations considered in this work.
The review of existing geometrically exact Kirchhoff-Love beam elements revealed that there are only a few approaches of this kind that are suitable for general 3D, large-deformation problems. These approaches have been categorized in isotropic, straight and anisotropic Kirchhoff-Love formulations. A detailed evaluation of these formulations
led to the result that typically only a few of the essential requirements summarized in Table 2 are fulfillled. On the contrary, for the finite elements proposed in this work, the fulfillment of these essential properties such as objectivity and
path-independence, consistent spatial convergence behavior, the avoidance of locking in the high slenderness regime
or the conservation of energy and momentum by the spatial discretization scheme have been shown theoretically and
verified by means of representative numerical examples. In the context of conservation properties, especially, the influence of applying either Bubnov-Galerkin or Petrov-Galerkin discretizations has been in the focus. Concerning locking
behavior, the recently proposed MCS method has been compared with alternative methods known from the literature
such as Assumed Natural Strains (ANS) or Reduced Integration (RI) schemes (see also [105]). In contrast to these
alternative methods, the MCS method could effectively avoid any evidence of membrane locking for all investigated
load cases and slenderness ratios. In contrast to previously existing Kirchhoff-Love formulations, Simo-Reissner formulations fulfilling the requirements of Table 2 can be found in the literature very well. However, it has been argued
that the shear-free Kirchhoff-Love formulations can provide considerable numerical advantages such as lower spatial
discretization error level, improved performance of time integration schemes as well as linear and nonlinear solvers
or smooth geometry representation as compared to shear-deformable Simo-Reissner formulations when applied to
highly slender beams. On the basis of several numerical examples, detailed and systematic numerical comparisons of
the resulting discretization error levels and performance of the nonlinear solver have been performed between the four
variants of the proposed geometrically exact Kirchhoff-Love beam elements and two well-established geometrically
78
exact Simo-Reissner beam element formulations known from the literature. Most of the examples have been investigated for the two different slenderness ratios ζ = 100 and ζ = 10000. For the low slenderness ratio ζ = 100, the general
model difference between the shear-free Kirchhoff-Love theory and the Simo-Reissner theory of shear-deformable
beams, measured in form of the relative L2 -error remaining in the limit of arbitrarily fine spatial discretizations, typically lay below 0.1%. Also the quadratic decrease of this model difference with increasing slenderness ratio could
be confirmed numerically. In all investigated examples, the proposed WK elements have shown a lower discretization
error level than the investigated Simo-Reissner beam element formulation. These results confirm the theoretical prediction that Kirchhoff type formulations can achieve the same discretization error level as Reissner type formulations
with less degrees of freedom, since no shear deformation has to be represented. Compared to the excellent results of
the WK elements, the SK elements showed an increased discretization error level, in some examples even higher as
for the Simo-Reissner elements. Based on the underlying convergence theory, this phenomenon could be attributed
to the polynomial degree k = 3 of the employed trial functions and predicted to vanish for higher-order elements with
k > 3. This prediction has been confirmed by means of a first numerical test case employing Hermite polynomials
of order k = 5 which resulted in the expected optimal discretization error level lying below the error level of the
Simo-Reissner reference formulation. While most of the investigated examples have been conducted in a quasi-static
manner, also one well-known dynamic benchmark test from the literature has been conducted. There, the accuracy
of the inertia contributions as well as the energy stability of the employed time integration scheme could be confirmed.
Besides the resulting discretization error level, also the total number of Newton-Raphson iterations required to solve
the considered test cases by means of the different element formulations and for different slenderness ratios has been
analyzed in a systematic manner. For all investigated examples and slenderness ratios, the proposed SK-TAN and
WK-TAN elements required less Newton iterations to solve the problem as compared to the two well-established
Simo-Reissner formulations chosen as reference. In the small slenderness range ζ = 100, the results of the four proposed Kirchhoff-Love variants and the two investigated Simo-Reissner formulations lay at least in the same order of
magnitude. While the behavior of the Kirchhoff-Love formulations remained more or less unchanged, the number of
Newton iterations required by the two different Simo-Reissner formulations increased considerably with increasing
slenderness ratio. In the investigated examples with slenderness ratio ζ = 10000, this number was up to two orders of
magnitude(!) higher for the Simo-Reissner elements as compared to the proposed SK-TAN and WK-TAN elements.
Also the number of iterations required by the SK-ROT and WK-ROT elements, which are based on nodal rotation vectors for triad parametrization, has been independent from the considered slenderness ratio, higher as for the SK-TAN
and WK-TAN elements but still considerably lower as for the Reissner type elements. Recapitulatory, out of the four
proposed Kirchhoff variants, the WK-TAN element, based on a weak enforcement of the Kirchhoff constraint and a
triad parametrization via nodal tangent vectors, can be recommended in terms of a low discretization error level and
an excellent performance of the Newton-Raphson scheme. Of course, further factors could be considered in a comprehensive comparison. For example, the SK-ROT and WK-ROT elements based on nodal rotation vectors simplify
the prescription of Dirichlet conditions. The flexibility of the proposed beam element variants allows to combine the
advantages of the two different rotation parametrizations by choosing e.g. the WK-TAN element as basic formulation that provides an excellent Newton Raphson performance and by replacing the nodal tangents by nodal rotation
vectors at nodes where complex boundary or coupling conditions have to be prescribed. This can be realized by a
simple transformation applied to the residual and stiffness contributions of the relevant node. The abstaining from
stiff shear mode contributions underlying the proposed Kirchhoff-Love element formulations may not only yield an
improved Newton-Raphson performance. Also the highest eigenfrequency band of slender beams, which is associated
with the shear modes, can be avoided by this means. The theoretical considerations made in this work give hope for
considerably improved stability properties of numerical time integration schemes when combined with the developed
shear-free elements. A future numerical investigation of this topic seems to provide a considerable scientific potential.
79
Appendix A. Definition of rotational shape function matrices
In this appendix, the shape functions Ĩi (ξ) required for the multiplicative rotation increments
∆θ(ξ) =
nΛ
X
Ĩi (ξ)∆θ̂i ,
∆θ0 (ξ) =
i=1
nΛ
X
Ĩi0 (ξ)∆θ̂i .
(A.1)
i=1
associated with the triad interpolation 105 and originally derived in [80] shall be presented:
Ĩi (ξ) = Li (ξ)Λr T−1 (Φlh (ξ))T(Φil )ΛTr
nΛ
X
I T
j
j
L
(ξ)T(Φ
)
+ δiI Λr I3 − T−1 (Φlh (ξ))
v Λr
l
j=1
nΛ
X
J T
j
j
−1
iJ
L (ξ)T(Φl )
+ δ Λr I3 − T (Φlh (ξ))
v Λr .
j=1
In (A.2), no summation over double indices is applied. The vectors vI and v J are defined as
!
!
!
!
1
1
ΦI J
1
1
ΦI J
I
IJ
J
IJ
v = I3 + I J tan
S(Φ ) , v = I3 − I J tan
S(Φ ) ,
2
Φ
4
2
Φ
4
(A.2)
(A.3)
with the common abbreviation ΦI J = ||ΦI J ||. Moreover, the arc-length derivative Ĩi0 (ξ) reads:
i
T
Ĩi (ξ) = Li0 (ξ)Λr T−1 (Φlh (ξ))T(Φil )ΛTr + Li (ξ)Λr T−1
,s (Φlh (ξ))T(Φl )Λr
nΛ
nΛ
X
X
I T
j
j
j
j0
iI
−1
−1
L
(ξ)T(Φ
)
L
(ξ)T(Φ
)
− δ Λr T,s (Φlh (ξ))
+T
(Φ
(ξ))
v Λr
lh
l
l
j=1
j=1
nΛ
nΛ
X
X
J T
j
j
j
j0
−1
− δiJΛr T−1
L
(ξ)T(Φ
)
L
(ξ)T(Φ
)
+T
(Φ
(ξ))
lh
,s (Φlh (ξ))
v Λr .
l
l
j=1
j=1
(A.4)
Finally, the required arc-length derivative T−1
,s (Φlh (ξ)) is given by (see also [79, 40]):
1 − cos Φlh
Φlh sin Φlh − 2(1 − cos Φlh )
S(Φlh ) +
S(Φ0lh )
4
Φlh
Φ2lh
!
sin Φlh
1−
S(Φlh )S(Φ0lh ) + S(Φ0lh )S(Φlh )
Φlh
T 0
T−1
,s (Φlh (ξ)) = Φlh Φlh
+
1
Φ2lh
+ ΦTlh Φ0lh
(A.5)
3 sin Φlh − Φlh (2 + cos Φlh )
S(Φlh )S(Φlh ).
Φ5lh
Here, the abbreviations Φlh = Φlh (ξ) as well as Φlh = ||Φlh (ξ)|| have been applied. As mentioned in the original
0
work [80], the limit T−1
,s (Φlh (ξ)) → 0.5S(Φlh (ξ)) can be derived for small angles Φlh (ξ) → 0.
Appendix B. Modeling of Dirichlet boundary conditions and joints
For many applications, the formulation of proper Dirichlet boundary conditions or joints between the nodes of different
beam elements are of a high practical relevance. This appendix represents a brief summary, where the possibility of
formulating some basic constraint conditions will be investigated for the SK-ROT and the SK-TAN element.
80
Appendix B.1. SK-ROT element
Since the SK-ROT element simplifies the formulation of Dirichlet boundary conditions and kinematic constraints in
many practically relevant cases, it will be considered first.
i) Dirichlet boundary conditions: A simple support at element node a can be realized via
d̂a = d̂au = d̂a0 → ∆d̂a = 0.
(B.1)
If a clamped end should be modeled, also the cross-section orientation has to be fixed, i.e.
Λa = Λau = Λa0 ,
and
ψ̂a = ψ̂a0 → ∆θ̂a = 0.
(B.2)
Thus, the modeling of Dirichlet boundary conditions for the employed translational and rotational degrees of freedom
is similar to standard finite elements that are purely based on translational degrees of freedom. This procedure can
also be extended to inhomogeneous conditions. However, the determination of ∆θ̂a requires special care in this case:
d̂a = d̂au (t) → ∆d̂an+1 = d̂au,n+1 − d̂au,n ,
(B.3)
a
Λa = Λau (t) → exp(S(∆θ̂n+1
)) = Λau,n+1 ΛaT
u,n .
The multiplicative procedure of the second line simplifies to the additive procedure according to the first line if the
prescribed rotation is additive, which only holds for 2D rotations.
ii) Connections: A simple (moment-free) joint between the two nodes a and b of two connected elements reads:
d̂b = d̂a ,
δd̂b = δd̂a ,
∆d̂b = ∆d̂a .
(B.4)
Thus, the degrees of freedom d̂b can be eliminated from the global system of equations in a standard manner by
simply assembling the corresponding lines and columns of the global residual vector and stiffness matrix properly. A
rigid joint between two elements prescribed at the nodes a and b additionally requires to suppress any relative rotation
between the associated nodal triads. It is assumed that these nodal triads differ by some fixed relative rotation Λ0 :
Λa = Λb Λ0
exp(S(∆θ̂a )) = exp(S(∆θ̂b ))Λ0
or
→
Λ0 = ΛbT Λa .
(B.5)
From (B.5), the following relations between the associated rotation increments can be derived:
δΛa = δΛb Λ0 → S(δθ̂a )Λa = S(δθ̂b )Λb Λ0 → δθ̂b = δθ̂a → ∆θ̂b = ∆θ̂a .
(B.6)
Consequently, also the rotational degrees of freedom ψ̂ can be eliminated in a standard manner by simply assembling
the corresponding lines and columns of the global residual vector and of the global stiffness matrix properly.
b
Remark: It is emphasized that a rigid joint according to (B.5) is formulated via right-translation of the
rotation tensor Λ0 . This is mandatory since a rigid joint represents a fixed orientation difference between
material quantities, i.e. a fixed relative rotation with respect to material axes. A left-translation via
Λa = Λ0 Λb
→
Λ0 = Λa ΛbT
→
δθ̂b = ΛT0 δθ̂a , δθ̂a ,
(B.7)
i.e. a fixed relative rotation with respect to spatial axes, has a different physical meaning.
Remark: If additive increments ∆ψ̂a and ∆ψ̂b of the rotation vectors ψ̂a and ψ̂b instead of the multiplicative
increments ∆θ̂a and ∆θ̂b were applied in the linearization process, equation (B.6) has to be replaced by:
∆θ̂b = ∆θ̂a
→
∆ψ̂b = T(ψ̂b )T−1 (ψ̂a )∆ψ̂a , ∆ψ̂a .
(B.8)
In this case, a direct elimination of the degrees of freedom ψ̂b via a proper assembly of the global stiffness
matrix is not possible. Instead, the corresponding columns have to be scaled with the matrix T(ψ̂b )T−1 (ψ̂a ).
Remark: Physically reasonable boundary conditions can be completely defined by the cross-section orientation and centroid position. For all considered types of boundary conditions, the degrees of freedom tˆa and
tˆb , which are a measure for the nodal axial force, are part of the FEM solution and must not be prescribed.
81
Appendix B.2. SK-TAN element
The treatment of the translational degrees of freedom required for the subsequently considered boundary conditions
is identical to the last section and will therefore be omitted here.
i) Dirichlet boundary conditions: In order to model a clamped end with the SK-TAN element, the simplest case of a
tangent vector that is parallel to a global base vector, e.g. t̂a k e1 , is considered. Then, (B.1) has to be supplemented by
t̂aT e2 = t̂aT e3 = 0 → ∆tˆ2a = ∆tˆ3a = 0
and ϕ̂a = ϕ̂a0 → ∆ϕ̂a = 0.
(B.9)
Here, the representation t̂a = tˆia ei of the tangent in the global frame ei has been exploited. In order to prescribe boundary
conditions with arbitrary triad orientation, the tangent has to be expressed in the basis of the prescribed triad:
Λa = Λau = Λa0 , t̂a = T̂ ia gai → T̂ 2a = T̂ 3a = 0 → ∆T̂ 2a = ∆T̂ 3a = 0, ϕ̂a = ϕ̂a0 → ∆ϕ̂a = 0.
(B.10)
Consequently, in this case, the equations of the linearized residual vector that are associated with the degrees of
freedom t̂a have to be transformed by the rotation tensor Λa0 and the Dirichlet conditions have to be formulated in this
rotated coordinate system. Again, the first component T̂ 1 of the tangent vector, when expressed in the material frame,
represents its magnitude and must not be prescribed. If the Dirichlet conditions are time-dependent, the prescribed
evolution of the relative angle has to be adapted, since now the intermediate frame ΛaMϕ̂ might change in time:
a
a
a
exp(S[ϕ̂an+1 ga1,n+1 ]) = Λan+1 ΛaT
Mϕ̂ ,n+1 with Λ Mϕ̂ ,n+1 = sr(Λ Mϕ̂ ,n , g1,n+1 ).
(B.11)
Thus, the required value ϕ̂an+1 has to be determined based on the prescribed current triad Λan+1 and the intermediate
triad ΛaMϕ̂ ,n of the last step (see also Section 6.2.2). The remaining conditions remain unchanged as compared to (B.10).
ii) Connections: Based on (B.5), (B.6) and (25), the following relations between (t̂a , ϕ̂a ) and (t̂b , ϕ̂b ) can be stated:
!
1
a
a
a
a
b
b
b
b
b b
b
b
δt̂ = −t S(g1 ) |{z}
δθ +g1 δt = −t S(g1 ) a S(g1 )δt̂ + g1 δΘ̂1 + gb1 δtˆb ,
t
=δθ
˙ a
!
(B.12)
b
bT
b
bT 1
a
a
a
a
δΘ̂1 = g1 |{z}
δθ = g1 a S(g1 )δt̂ + g1 δΘ̂1 .
t
a
=δθ
˙
Combining these two relations eventually yields the following total transformation matrix:
!
! δt̂a
1
b
b
b
a
b
b
a
b
δt̂
−t S(g1 ) ta S(g1 ) −t S(g1 )g1 g1
δΘ̂a1 .
=
bT 1
a
bT a
b
g1 ta S(g1 )
g g
0 b
δΘ̂1
|
{z 1 1
} δtˆ
(B.13)
=:TRC
A similar relation can also be formulated for the iterative increments. Since the multiplicative rotation increment
components ∆Θ1 = TΘM1t ∆t + ∆ϕ (see (20)) have to be expressed by additive increments ∆t and ∆ϕ for the chosen
linearization scheme, an additional transformation is required as compared to (B.13):
a
!
!
∆t̂
I3
0 0
b
I3
0
∆t̂
= T̃RC1 TRC T̃RC2 ∆ϕ̂a1 , T̃RC1 :=
, T̃RC2 := TΘM1t 1 0 .
(B.14)
−TΘM1t 1
∆ϕ̂b
ˆb
T
0
0 1
∆t
Equations (B.13) and (B.14) allow to transform the corresponding lines and columns of the global residual vector and
of the global tangent stiffness matrix properly and to eliminate the degrees of freedom (t̂b , ϕ̂b ) from the global system
of equations. Again, the magnitude of the tangent vector tˆb is not influenced by the rigid joint and enters the system of
equations as new degree of freedom. While in the last section, the motion of the rigid joint was completely determined
by the set (d̂a , ψ̂a , tˆa , tˆb ), in this section the alternative set (d̂a , t̂a , ϕ̂a , tˆb ) is employed.
82
All in all, it can be concluded that the realization of clamped ends with arbitrary orientation or of rigid joints between beams is simpler for the SK-ROT formulation based on nodal rotation vectors. While for these elements such
conditions can directly be formulated in the global coordinate system, the tangent vector-based SK-TAN formulation
requires an additional transformation of the corresponding lines and columns of the global residual vector and stiffness matrix. In Section 11, some properties of the tangent vector-based variant will become apparent which make this
type of formulation favorable for many applications. If certain element nodes require Dirichlet conditions of the type
considered here, it is still possible to apply a hybrid approach, and to replace the nodal tangents by nodal rotation
vectors as primary variables only at the specific nodes where such conditions are required. All the results derived so
far apply in a similar manner to the WK-TAN and WK-ROT elements that will be derived in the next section.
Appendix C. Linearization of SK-TAN element
Before deriving the linearization of the SK-TAN element, some former definitions are repeated:
t := r0 ,
g1 :=
r0
,
||r0 ||
t̃ :=
r0
,
||r0 ||2
g01 =
r00 (r0T r00 )r0
−
,
||r0 ||
||r0 ||3
t̃0 =
r00
2(r0T r00 )r0
−
.
0
2
||r ||
||r0 ||4
(C.1)
These quantities will be required for later derivations. Linearization of (C.1) yields:
1
1
I3 −g1 ⊗gT1 H0 ∆d̂,
∆t̃ = 0 2 I3 −2g1 ⊗gT1 H0 ∆d̂,
0
||r ||
||r ||
0T 00
(r
r
)
1
0
∆g01 = − 0 3 I3 −g1 ⊗gT1 H0 ∆d̂ − 0 g01 ⊗gT1 +g1 ⊗g0T
1 H ∆d̂
||r ||
||r ||
1
+ 0 I3 −g1 ⊗gT1 H00 ∆d̂,
||r ||
2
2(r0T r00 )
0
∆t̃0 = −
I3 −2g1 ⊗gT1 H0 ∆d̂ − 0 2 g01 ⊗gT1 +g1 ⊗g0T
1 H ∆d̂
0
4
||r ||
||r ||
1
+ 0 2 I3 −2g1 ⊗gT1 H00 ∆d̂.
||r ||
∆g1 =
(C.2)
In the following, the linearization of the SK-TAN element will be derived. For completeness, the underlying residual
vector (162) with inserted strain re-interpolation (163) is repeated here:
"
#
Z1
T
T
0
rd̂ = vθ⊥m+ v̄ F̄1 − H f̃ρ − vθ⊥m̃ρ J(ξ)dξ− H fσ + vθ⊥ mσ ,
Γσ
−1
(C.3)
"
#
Z1
rΘ̂1 = v0θkΘ m − vθkΘ m̃ρ J(ξ)dξ − vθkΘ mσ .
Γσ
−1
The linearization of the element residual vector (C.3) obeys the following general form:
Z1
∆rd̂ = ∆v0θ⊥m+ v0θ⊥∆m+∆v̄ F̄1 + v̄ ∆F̄1 J(ξ)dξ
−1
−
Z1
"
#
HT ∆fρ +∆vθ⊥m̃ρ + vθ⊥∆mρ J(ξ)dξ− ∆vθ⊥mσ ,
(C.4)
Γσ
−1
"
#
Z1
∆rΘ̂1 = ∆v0θkΘ m+ v0θkΘ ∆m − ∆vθkΘ m̃ρ − vθkΘ ∆mρ J(ξ)dξ − ∆vθkΘ mσ .
Γσ
−1
83
In order to identify the element stiffness matrix kS K−T AN , (C.4) has to be brought in the form
∆rS K−T AN =: kS K−T AN ∆x̂T AN .
(C.5)
The vector ∆x̂T AN has already been defined in Section 9.1. The moment-related terms yield:
∆v0θ⊥m = H00T S(m)∆t̃+ H0T S(m)∆t̃0 ,
∆m = −S(m)∆θ+cm ∆θ0 ,
∆vθ⊥m̃ρ = H0T S(m̃ρ )∆t̃,
∆vθ⊥mσ = H0T S(mσ )∆t̃,
∆vθkΘ m̃ρ = (LTk ⊗ m̃Tρ )∆g1 ,
∆vθkΘ mσ = (LTk ⊗mTσ )∆g1 ,
(C.6)
T
T
T
0
∆v0θkΘ m = (L0T
k ⊗m )∆g1 +(Lk ⊗m )∆g1 .
Here, many of the relations already derived in Section 6.2.4 could be re-used. The field of multiplicative rotation
vector increments ∆θ follows directly from equation (121):
Θ1+(vTθ⊥ + vTθkd )∆d̂,
∆θ = vTθkΘ ∆Θ̂
vθkΘ = LTk ⊗gT1 , vθ⊥ = −H0T S(t̃),
nΛ
X i
H0T (ξ)(g1I × t̃) − H0T (ξI )(g1 × t̃I )
,
vθkd =
L v1i − v1 ⊗gT1 , v1 =
1+g1T g1I
i=1
v1i = v1 (ξi ) =
H0T (ξi )(g1I × t̃i ) − H0T (ξI )(gi1 × t̃I )
1+gi1T g1I
(C.7)
.
In a similar manner, the associated arc-length derivative ∆θ0 follows from equation (122):
0T
Θ1+(v0T
∆θ0 = v0T
θ⊥+ vθkd )∆d̂,
θkΘ ∆Θ̂
00T
0T
T
0
0
0T
0
T
v0θkΘ = L0T
k ⊗g1 + Lk ⊗g1 , vθ⊥ = −H S(t̃)− H S(t̃ ), t̃ =
v0θkd =
nΛ
X
r00
2(r0T r00 )r0
−
,
||r0 ||2
||r0 ||4
nΛ
X
Li0 v1i − v01 ⊗gT1 +
Li v1i − v1 ⊗g0T
1 ,
(C.8)
i=1
0T
v01 =
H
i=1
00T
I
0
I
(ξ)(g1 × t̃ )+ H (ξ)(g1I × t̃)− H0T (ξI )(g01 × t̃I ) (g0T
1 g1 )v1
−
.
1+gT1g1I
1+g1T g1I
The remaining linearizations required in equation (C.6) have already been derived in (C.2). In contrast to the spin
vector field δθ, the increment field ∆θ has to be expressed via additive relative angle increments ∆ϕ̂i . The required
relation is given by (127) and repeated here:
Θ1 = (Θ̂11 , Θ̂21 , Θ̂31 )T ,
∆Θ̂
∆Θ̂i1 = −
i
0 i
ḡiT
1 S(g1 ) H (ξ )∆d̂
+ ∆ϕ̂i .
i
i ||
||t
1 + giT
ḡ
1 1
The linearization of the element residual terms associated with axial tension results in:
3
X
H0T
∆F̄1 = EA∆¯ = EAv̄T ∆d̂, ∆v̄ = Li (ξ)∆v (ξi ), ∆v = 0 (I3 +g1 ⊗gT1 )H0 ∆d̂.
||r ||
i=1
(C.9)
(C.10)
Based on equation (93), the linearization of the inertia forces can be written as:
−HT∆fρ = ρAcr̈1 HT H∆d̂,
cr̈1 =
1−αm
.
(1−αm )β∆t2
(C.11)
The time integration factor cr̈1 of the modified generalized-α scheme according to Section 5 slightly differs from the
corresponding factor of the standard generalized-α scheme. The linearization of the inertia moments yields:
−∆mρ = S(mρ )∆θ + [cW {S(W)Cρ − S(Cρ W)} + cA Cρ ]∆Θ̃n+1 ,
1−αm
γ
cA =
, cW =
, ∆Θ̃n+1 = ΛTn ∆θ̃n+1 = ΛTn T(θ̃n+1 )∆θ.
2
β∆t
(1−αm )β∆t
84
(C.12)
For clarity, the indices n + 1 and n of the current and previous time step have explicitly been noted for some of the
quantities occurring in (C.12). All the other quantities are evaluated at tn+1 . As already introduced in Section 5, the
fields Θ̃n+1 and θ̃n+1 are the material and spatial multiplicative rotation increments relating the current configuration
and the converged configuration of the previous time step tn . The two vectors are related by the transformation
Θ̃n+1 = ΛTn+1 θ̃n+1 = ΛTn θ̃n+1 → ∆Θ̃n+1 = ΛTn ∆θ̃n+1 .
(C.13)
The second step in (C.13) is valid since θ̃n+1 is an eigenvector with eigenvalue one of the rotation tensor Λn+1 ΛTn
between the configurations n and n+1, thus Λn+1 ΛTn θ̃n+1 = θ̃n+1 . Furthermore, ∆Θ̃n+1 and ∆θ̃n+1 represent the fields
of additive increments of Θ̃n+1 and θ̃n+1 between two successive Newton iterations, whereas ∆θ as given by (C.7)
represents the field of multiplicative rotation increments between two successive Newton iterations.
Appendix D. Linearization of WK-TAN element
The residual vector of the WK-TAN element is given in equation (180) and repeated here:
"
#
Z1
rd̂ = v̄0θ⊥m+ v̄ F̄1 − HT f̃ρ − v̄θ⊥m̃ρ J(ξ)dξ− HT fσ + v̄θ⊥ mσ ,
Γσ
−1
(D.1)
"
#
Z1
rΘ̂1 = v̄0θkΘ m − v̄θkΘ m̃ρ J(ξ)dξ − v̄θkΘ mσ .
Γσ
−1
The linearization of the element residual vector (D.1) obeys the following general form:
Z1
∆rd̂ = ∆v̄0θ⊥m+ v̄0θ⊥∆m+∆v̄ F̄1 + v̄ ∆F̄1 J(ξ)dξ
−1
−
Z1
"
#
H ∆fρ +∆v̄θ⊥m̃ρ + v̄θ⊥∆mρ J(ξ)dξ− ∆v̄θ⊥mσ ,
T
(D.2)
Γσ
−1
"
#
Z1
0
0
∆rΘ̂1 = ∆v̄θkΘ m+ v̄θkΘ ∆m − ∆v̄θkΘ m̃ρ − v̄θkΘ ∆mρ J(ξ)dξ − ∆v̄θkΘ mσ .
Γσ
−1
In order to identify the element stiffness matrix kWK−T AN , (D.2) has to be brought in the form
∆rWK−T AN =: kWK−T AN ∆x̂T AN .
(D.3)
The linearization of the vectors v̄... and v̄0... originally defined in (180) follows to:
∆v̄θ⊥ =−
3
X
Li (ξ)∆vθ⊥ (ξi ),
3
X
∆v̄ = Li (ξ)∆v (ξi ),
i=1
i=1
3
X
∆v̄θkΘ = Li (ξ)∆vθkΘ (ξi )
i=1
3
X
Li , ξ(ξ)
∆v... (ξi ).
with ∆v̄0...=
J(ξ)
i=1
(D.4)
The linearization of the vectors v... and v0... has already been stated in the last section. Also the linearization of the
moment stress resultant has the same form as in the last section:
∆m = −S(m)∆θ+cm ∆θ0 .
85
(D.5)
However, the fields ∆θ and ∆θ0 originally defined in Section 6.2.3 are this time given by
∆θ =
3
X
i=1
3
X
1 i
Ĩ (ξ)∆θ(ξ ), ∆θ =
Ĩ,ξ (ξ)∆θ(ξi ).
J(ξ)
i=1
i
0
i
(D.6)
Due to the Kirchhoff constraint, the nodal increments ∆θ(ξi ) can be expressed according to:
∆θ(ξi ) = ∆Θ̂i1 g1 (ξi )+ vTθ⊥ (ξi )∆d̂,
∆Θ̂i1 = −
i
0 i
ḡiT
1 S(g1 ) H (ξ )∆d̂
+ ∆ϕ̂i .
i
||ti ||
1 + giT
1 ḡ1
(D.7)
The linearization of the inertia forces is identical to the corresponding results of the last section:
−HT∆fρ = ρAcr̈1 HT H∆d̂,
cr̈1 =
1−αm
.
(1−αm )β∆t2
(D.8)
This statement also holds for the linearization of the inertia moments, which reads:
−∆mρ = S(mρ )∆θ + [cW {S(W)Cρ − S(Cρ W)} + cA Cρ ]∆Θ̃n+1 ,
1−αm
γ
, ∆Θ̃n+1 = ΛTn ∆θ̃n+1 = ΛTn T(θ̃n+1 )∆θ.
cA =
, cW =
β∆t
(1−αm )β∆t2
(D.9)
However, for the WK-TAN element, the rotation increment field ∆θ is given by equation (D.6).
Appendix E. Linearization of SK-ROT and WK-ROT elements
The nodal primary variable variations of the SK/WK-TAN and the SK/WK-ROT elements read:
δx̂T AN := (δd̂1T, δt̂1T, δΘ̂11 , δd̂2T, δt̂2T, δΘ̂21 , δΘ̂31 )T ,
(E.1)
δx̂ROT := (δd̂1T, δθ̂1T, δtˆ1 , δd̂2T, δθ̂2T, δtˆ1 , δΘ̂31 )T .
In a similar manner, the set of iterative nodal primary variable increments have been defined as:
∆x̂T AN := (∆d̂1T, ∆t̂1T, ∆ϕ̂1 , ∆d̂2T, ∆t̂2T, ∆ϕ̂2 , ∆ϕ̂3 )T ,
(E.2)
∆x̂ROT := (∆d̂1T, ∆θ̂1T, ∆tˆ1 , ∆d̂2T, ∆θ̂2T, ∆tˆ1 , ∆ϕ̂3 )T .
The transformations between these primary variable variations and increments is given by:
δx̂T AN = T̃x̂ δx̂ROT
and ∆x̂T AN = T Mx̂ ∆x̂ROT .
(E.3)
The transformation matrices T̃x̂ , originally defined in (166), and T Mx̂ have the following form:
I3
T̃x̂ =
T̃1
I3
T̃2
1
and
I3
T Mx̂ =
T1M
I3
T2M
1
.
(E.4)
These two different matrices are required, since the primary variable variations of the SK/WK-TAN elements are
based on the multiplicative quantities δΘ̂i1 , whereas the corresponding iterative primary variable increments are based
on the additive quantities ∆ϕ̂i . The matrices T̃i and TiM ((25) and (24)) are evaluated at the element boundary nodes:
T̃i := T̃(ξi )
and
TiM := T M (ξi )
86
for
i = 1, 2.
(E.5)
In Section 9.2, it has already been shown that the following residual transformation is valid:
T
rROT = T̃x̂ rT AN .
(E.6)
In a similar manner, also the linearized element residual vector can be transformed:
T
T
T
∆rROT = ∆T̃x̂ rT AN + T̃x̂∆rT AN =: H̃x̂ (rT AN )+ T̃x̂ kT AN T Mx̂ ∆x̂ROT .
|{z} |
{z
}
=kT AN ∆x̂T AN
(E.7)
:=kROT
Here, the matrix H̃x̂ (rT AN ) has been introduced, in order to represent the linearization of T̃x̂ :
T
H̃x̂ (rT AN )∆x̂ROT := ∆T̃x̂ rT AN
0
with H̃x̂ (rT AN ) =
1
H̃
0
2
H̃
0
.
(E.8)
i
After calculating the derivative of T̃x̂ and re-ordering the result, the submatrices H̃ can be stated:
S(rT AN,t̂i )S(gi1 ) − rT AN,Θ̂i1 S(gi1 ) S(gi1 )rT AN,t̂i
i
, i = 1, 2.
H̃ =
−rTT AN,t̂i S(gi1 )
0
(E.9)
From (E.7), the following transformation rule for the the element stiffness matrix can be stated:
∆rROT = kROT ∆x̂ROT
T
kROT = H̃x̂ (rT AN )+ T̃x̂ kT AN T Mx̂ .
with
(E.10)
In order to apply this transformation, the components of the element stiffness matrices kT AN and kROT have to be
arranged in the same order as the components of the element residual vectors:
rT AN := (rTT AN,d̂1 , rTT AN,t̂1 , rT AN,Θ̂1 , rTT AN,d̂2 , rTT AN,t̂2 , rT AN,Θ̂2 , rT AN,Θ̂3 )T ,
1
1
1
rROT := (rTROT,d̂1 , rTROT,θ̂1 , rROT,tˆ1 , rTROT,d̂2 , rT
T
ˆ2 , rROT,tˆ2 , rROT,Θ̂3 ) .
ROT,θ
(E.11)
1
References
[1] L. X. Ademir. Static Kirchhoff rods under the action of external forces: Integration via Runge-Kutta method. Journal of Computational
Methods in Physics, 2014:650365, 2014.
[2] S. S. Antman. Kirchhoff’s problem for nonlinearly elastic rods. Quarterly of Applied Mathematics, 32(3):221–240, 1974.
[3] S. S. Antman. Nonlinear problems of elasticity. Springer, 1995.
[4] J. H. Argyris, H. Balmer, J. St. Doltsinis, P. C. Dunne, M. Haase, M. Kleiber, G. A. Malejannakis, H.-P. Mlejnek, M. Müller, and D. W.
Scharpf. Finite element method–the natural approach. Computer Methods in Applied Mechanics and Engineering, 17-18(Part 1):1–106,
January 1979.
[5] F. Armero and J. Valverde. Invariant Hermitian finite elements for thin Kirchhoff rods. I: The linear plane case. Computer Methods in
Applied Mechanics and Engineering, 213-216:427–457, 2012.
[6] F. Armero and J. Valverde. Invariant Hermitian finite elements for thin Kirchhoff rods. II: The linear three-dimensional case. Computer
Methods in Applied Mechanics and Engineering, 213-216:458–485, 2012.
[7] M. Arnold and O. Brüls. Convergence of the generalized-α scheme for constrained mechanical systems. Multibody System Dynamics,
18(2):185–202, 2007.
[8] D. G. Ashwell and A. B. Sabir. Limitations of certain curved finite elements when applied to arches. International Journal of Mechanical
Sciences, 13(2):133–139, 1971.
[9] D. G. Ashwell, A. B. Sabir, and T. M. Roberts. Further studies in the application of curved finite elements to circular arches. International
Journal of Mechanical Sciences, 13(6):507–517, 1971.
[10] A. Avello, J. G. de Jaln, and E. Bayo. Dynamics of flexible multibody systems using cartesian co-ordinates and large displacement theory.
International Journal for Numerical Methods in Engineering, 32(8):1543–1563, 1991.
[11] K.-J. Bathe and S. Bolourchi. Large displacement analysis of three-dimensional beam structures. International Journal for Numerical
Methods in Engineering, 14(7):961–986, 1979.
[12] K.-J. Bathe, A. Iosilevich, and D. Chapelle. An inf-sup test for shell finite elements. Computers & Structures, 75(5):439–456, 2000.
87
[13] J.-M. Battini and C. Pacoste. Co-rotational beam elements with warping effects in instability problems. Computer Methods in Applied
Mechanics and Engineering, 191(17–18):1755–1789, 2002.
[14] O. A. Bauchau and C. L. Bottasso. On the design of energy preserving and decaying schemes for flexible, nonlinear multi-body systems.
Computer Methods in Applied Mechanics and Engineering, 169(12):61–79, 1999.
[15] O. A. Bauchau, S. Han, A. Mikkola, and M. K. Matikainen. Comparison of the absolute nodal coordinate and geometrically exact formulations for beams. Multibody System Dynamics, 32(1):67–85, 2014.
[16] A. M. Bauer, M. Breitenberger, B. Philipp, R. Wüchner, and K.-U. Bletzinger. Nonlinear isogeometric spatial Bernoulli beam. Computer
Methods in Applied Mechanics and Engineering, 303:101–127, 2016.
[17] T. Belytschko and B. J. Hseih. Nonlinear transient finite element analysis with convected coordinates. International Journal for Numerical
Methods in Engineering, 7:255–271, 1973.
[18] T. Belytschko and W. G. Lawrence. Applications of higher order corotational stretch theories to nonlinear finite element analysis. Computers
& Structures, 10(1):175–182, 1979.
[19] M. Bergou, M. Wardetzky, S. Robinson, B. Audoly, and E. Grinspun. Discrete elastic rods. ACM Transactions on Graphics, 27(3):63:1–
63:12, August 2008.
[20] F. Bertails, B. Audoly, M.-P. Cani, B. Querleux, F. Leroy, and J.-L. Lévêque. Super-helices for predicting the dynamics of natural hair. ACM
Transactions on Graphics, 25(3):1180–1187, July 2006.
[21] P. Betsch and P. Steinmann. Constrained integration of rigid body dynamics. Computer Methods in Applied Mechanics and Engineering,
191(35):467–488, 2001.
[22] P. Betsch and P. Steinmann. Frame-indifferent beam finite elements based upon the geometrically exact beam theory. International Journal
for Numerical Methods in Engineering, 54(12):1775–1788, 2002.
[23] R. L. Bishop. There is more than one way to frame a curve. The American Mathematical Monthly, 82(3):246–251, 1975.
[24] M. Borri and C. Bottasso. An intrinsic beam model based on a helicoidal approximation - Part I: Formulation. International Journal for
Numerical Methods in Engineering, 37(13):2267–2289, 1994.
[25] M. Borri and C. Bottasso. An intrinsic beam model based on a helicoidal approximation - Part II: Linearization and finite element implementation. International Journal for Numerical Methods in Engineering, 37(13):2291–2309, 1994.
[26] C. L. Bottasso and M. Borri. Integrating finite rotations. Computer Methods in Applied Mechanics and Engineering, 164(34):307–331,
1998.
[27] F. Boyer, G. De Nayer, A. Leroyer, and M. Visonneau. Geometrically exact Kirchhoff beam theory: Application to cable dynamics. Journal
of Computational and Nonlinear Dynamics, 6:1–14, 2011.
[28] F. Boyer and D. Primault. Finite element of slender beams in finite transformations: A geometrically exact approach. International Journal
for Numerical Methods in Engineering, 59(5):669–702, 2004.
[29] F. Brezzi. On the existence, uniqueness and approximation of saddle-point problems arising from lagrangian multipliers. ESAIM: Mathematical Modelling and Numerical Analysis - Modélisation Mathématique et Analyse Numérique, 8(R2):129–151, 1974.
[30] F. Brezzi and M. Fortin. Mixed and hybrid finite elements. Springer, 1991.
[31] O. Brüls and A. Cardona. On the use of Lie group time integrators in multibody dynamics. Journal of Computational and Nonlinear
Dynamics, 5:031002, 2010.
[32] O. Brüls, A. Cardona, and M. Arnold. Lie group generalized-α time integration of constrained flexible multibody systems. Mechanism and
Machine Theory, 48:121–137, 2012.
[33] M. Cannarozzi and L. Molari. A mixed stress model for linear elastodynamics of arbitrarily curved beams. International Journal for
Numerical Methods in Engineering, 74(1):116–137, 2008.
[34] A. Cardona and M. Géradin. A beam finite element non-linear theory with finite rotations. International Journal for Numerical Methods in
Engineering, 26(11):2403–2438, 1988.
[35] A. Cardona, M. Géradin, and D. B. Doan. Rigid and flexible joint modelling in multibody dynamics using finite elements. Computer
Methods in Applied Mechanics and Engineering, 89(1):395–418, 1991.
[36] J. K. Choit and J. K. Lim. General curved beam elements based on the assumed strain fields. Computers & Structures, 55(3):379–386, 1995.
[37] J. Chung and G. M. Hulbert. A time integration algorithm for structural dynamics with improved numerical dissipation: The generalized-α
method. Journal of Applied Mechanics, 60:371–375, 1993.
[38] E. Cosserat and F. Cosserat. Théorie des Corps Déformables. Traité de Physique. Paris, second ed. edition, 1909.
[39] M. A. Crisfield. A consistent co-rotational formulation for non-linear, three-dimensional, beam-elements. Computer Methods in Applied
Mechanics and Engineering, 81(2):131–150, 1990.
[40] M. A. Crisfield. Non-linear finite element analysis of solids and structures, Volume 2: Advanced topics. Wiley & Sons, 1997.
[41] M. A. Crisfield. Non-linear finite element analysis of solids and structures, Volume 1: Essentials. Wiley & Sons, 2003.
[42] M. A. Crisfield, U. Galvanetto, and G. Jelenić. Dynamics of 3-D co-rotational beams. Computational Mechanics, 20(6):507–519, 1997.
[43] M. A. Crisfield and G. Jelenić. Objectivity of strain measures in the geometrically exact three-dimensional beam theory and its finite-element
implementation. Proceedings of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences, 455(1983):1125–
1147, 1999.
[44] L. A. Crivelli and C. A. Felippa. A three-dimensional non-linear Timoshenko beam based on the core-congruential formulation. International Journal for Numerical Methods in Engineering, 36(21):3647–3673, 1993.
[45] C. J. Cyron and W. A. Wall. Numerical method for the simulation of the Brownian dynamics of rod-like microstructures with threedimensional nonlinear beam elements. International Journal for Numerical Methods in Engineering, 90(8):955–987, 2012.
[46] F. Demoures, F. Gay-Balmaz, S. Leyendecker, S. Ober-Blöbaum, T. S. Ratiu, and Y. Weinand. Discrete variational Lie group formulation
of geometrically exact beam dynamics. Numerische Mathematik, 130(1):73–123, 2015.
[47] E. H. Dill. Kirchhoff’s theory of rods. Archive for History of Exact Sciences, 44(1):1–23, 1992.
[48] E. P. Dukić, G. Jelenić, and M. Gaćeša. Configuration-dependent interpolation in higher-order 2D beam finite elements. Finite Elements in
Analysis and Design, 78:47–61, 2014.
88
[49] D. Durville. Simulation of the mechanical behaviour of woven fabrics at the scale of fibers. International Journal of Material Forming,
3(2):1241–1251, 2010.
[50] E. N. Dvorkin, E. Oñate, and J. Oliver. On a non-linear formulation for curved Timoshenko beam elements considering large displacement/rotation increments. International Journal for Numerical Methods in Engineering, 26(7):1597–1613, 1988.
[51] S. R. Eugster, C. Hesch, P. Betsch, and Ch. Glocker. Director-based beam finite elements relying on the geometrically exact beam theory
formulated in skew coordinates. International Journal for Numerical Methods in Engineering, 97(2):111–129, 2014.
[52] W. W Fan and W. D. Zhu. An accurate singularity-free formulation of a three-dimensional curved EulerBernoulli beam for flexible multibody
dynamic analysis. Journal of Vibration and Acoustics, 138(5):051001, 2016.
[53] C. A. Felippa and B. Haugen. A unified formulation of small-strain corotational finite elements: I. Theory. Computer Methods in Applied
Mechanics and Engineering, 194(21–24):2285–2335, 2005.
[54] I. Fried. Shape functions and the accuracy of arch finite elements. American Institute of Aeronautics and Astronautics Journal, 11:287–291,
1973.
[55] J. Frischkorn and S. Reese. A solid-beam finite element and non-linear constitutive modelling. Computer Methods in Applied Mechanics
and Engineering, 265:195–212, 2013.
[56] B. Gadot, O. R. Martinez, S. R. du Roscoat, D. Bouvard, D. Rodney, and L. Orgéas. Entangled single-wire NiTi material: A porous metal
with tunable superelastic and shape memory properties. Acta Materialia, 96:311–323, 2015.
[57] M. Géradin and A. Cardona. Kinematics and dynamics of rigid and flexible mechanisms using finite elements and quaternion algebra.
Computational Mechanics, 4(2):115–135, 1989.
[58] M. Géradin and A. Cardona. Flexible multibody dynamics - a finite element approach. Wiley, 2001.
[59] S. Ghosh and D. Roy. Consistent quaternion interpolation for objective finite element approximation of geometrically exact beam. Computer
Methods in Applied Mechanics and Engineering, 198(3–4):555–571, 2008.
[60] S. Ghosh and D. Roy. A frame-invariant scheme for the geometrically exact beam using rotation vector parametrization. Computational
Mechanics, 44(1):103–118, 2009.
[61] O. Gonzalez. Time integration and discrete Hamiltonian systems. Journal of Nonlinear Science, 6(5):449–467, 1996.
[62] S. Goyal, N. C. Perkins, and C. L. Lee. Nonlinear dynamics and loop formation in Kirchhoff rods with implications to the mechanics of
DNA and cables. Journal of Computational Physics, 209(1):371–389, 2005.
[63] L. Greco and M. Cuomo. B-Spline interpolation of Kirchhoff-Love space rods. Computer Methods in Applied Mechanics and Engineering,
256:251–269, 2013.
[64] L. Greco and M. Cuomo. An isogeometric implicit mixed finite element for Kirchhoff space rods. Computer Methods in Applied Mechanics
and Engineering, 298:325–349, 2016.
[65] F. Gruttmann, R. Sauer, and W. Wagner. A geometrical nonlinear eccentric 3D-beam element with arbitrary cross-sections. Computer
Methods in Applied Mechanics and Engineering, 160(34):383–400, 1998.
[66] K. M. Hsiao, J. Y. Lin, and W. Y. Lin. A consistent co-rotational finite element formulation for geometrically nonlinear dynamic analysis of
3-D beams. Computer Methods in Applied Mechanics and Engineering, 169(1–2):1–18, 1999.
[67] K. M. Hsiao and R. T. Yang. A co-rotational formulation for nonlinear dynamic analysis of curved Euler beam. Computers & Structures,
54(6):1091–1097, 1995.
[68] K. M. Hsiao, R. T. Yang, and A. C. Lee. A consistent finite element formulation for non-linear dynamic analysis of planar beam. International Journal for Numerical Methods in Engineering, 37(1):75–89, 1994.
[69] T. J. R. Hughes. The finite element method - linear static and dynamic finite element analysis. Dover, 2000.
[70] T. J. R. Hughes, R. L. Taylor, and W. Kanoknukulchai. A simple and efficient finite element for plate bending. International Journal for
Numerical Methods in Engineering, 11(10):1529–1543, 1977.
[71] A. Ibrahimbegović. On finite element implementation of geometrically nonlinear Reissner’s beam theory: Three-dimensional curved beam
elements. Computer Methods in Applied Mechanics and Engineering, 122(1–2):11–26, 1995.
[72] A. Ibrahimbegović. On the choice of finite rotation parameters. Computer Methods in Applied Mechanics and Engineering, 149(1–4):49–71,
1997.
[73] A. Ibrahimbegović and F. Frey. Finite element analysis of linear and non-linear planar deformations of elastic initially curved beams.
International Journal for Numerical Methods in Engineering, 36(19):3239–3258, 1993.
[74] A. Ibrahimbegović, F. Frey, and I. Kozar. Computational aspects of vector-like parametrization of three-dimensional finite rotations. International Journal for Numerical Methods in Engineering, 38(21):3653–3673, 1995.
[75] A. Ibrahimbegović and S. Mamouri. Energy conserving/decaying implicit time-stepping scheme for nonlinear dynamics of threedimensional beams undergoing finite rotations. Computer Methods in Applied Mechanics and Engineering, 191(3738):4241–4258, 2002.
[76] A. Ibrahimbegović and R. L. Taylor. On the role of frame-invariance in structural mechanics models at finite rotations. Computer Methods
in Applied Mechanics and Engineering, 191(45):5159–5176, 2002.
[77] A. Iosilevich, K.-J. Bathe, and F. Brezzi. On evaluating the infsup condition for plate bending elements. International Journal for Numerical
Methods in Engineering, 40(19):3639–3663, 1997.
[78] M. Iura and S. N. Atluri. Dynamic analysis of finitely stretched and rotated three-dimensional space-curved beams. Computers and
Structures, 29:875–889, 1988.
[79] G. Jelenić and M. A. Crisfield. Interpolation of rotational variables in nonlinear dynamics of 3D beams. International Journal for Numerical
Methods in Engineering, 43(7):1193–1222, 1998.
[80] G. Jelenić and M. A. Crisfield. Geometrically exact 3D beam theory: Implementation of a strain-invariant finite element for statics and
dynamics. Computer Methods in Applied Mechanics and Engineering, 171(1–2):141–171, 1999.
[81] G. Jelenić and M. Saje. A kinematically exact space finite strain beam model - finite element formulation by generalized virtual work
principle. Computer Methods in Applied Mechanics and Engineering, 120(12):131–161, 1995.
[82] P. Jung, S. Leyendecker, J. Linn, and M. Ortiz. A discrete mechanics approach to the Cosserat rod theory - Part 1: Static equilibria.
International Journal for Numerical Methods in Engineering, 85(1):31–60, 2011.
89
[83] C. Kane, J. E. Marsden, M. Ortiz, and M. West. Variational integrators and the Newmark algorithm for conservative and dissipative
mechanical systems. International Journal for Numerical Methods in Engineering, 49(10):1295–1325, 2000.
[84] R. K. Kapania and J. Li. A formulation and implementation of geometrically exact curved beam elements incorporating finite strains and
finite rotations. Computational Mechanics, 30(5):444–459, April 2003.
[85] D. Karamanlidis and R. Jasti. Curved mixed beam elements for the analysis of thin-walled free-form arches. Ingenieur-Archiv, 57(5):361–
367, 1987.
[86] G. Kirchhoff. Ueber das Gleichgewicht und die Bewegung eines unendlich dünnen elastischen Stabes. Journal für die reine und angewandte
Mathematik, 56:285–313, 1859.
[87] W. T. Koiter. On the nonlinear theory of thin elastic shells. In Proceedings of the Koninklijke Nederlandse Akademie van Wetenschappen,
volume 69, 1966.
[88] K. Kondoh, K. Tanaka, and S. N. Atluri. An explicit expression for the tangent-stiffness of a finitely deformed 3-D beam and its use in the
analysis of space frames. Computers & Structures, 24(2):253–271, 1986.
[89] F. Koschnick. Geometrische Locking-Effekte bei Finiten Elementen und ein allgemeines Konzept zu ihrer Vermeidung. PhD thesis, Lehrstuhl
für Statik, Technische Universität München, 2004.
[90] A. Kulachenko, T. Denoyelle, S. Galland, and S. B. Lindström. Elastic properties of cellulose nanopaper. Cellulose, 19(3):793–807, 2012.
[91] A. Kulachenko and T. Uesaka. Direct simulations of fiber network deformation and failure. Mechanics of Materials, 51:1–14, 2012.
[92] H. Lang and M. Arnold. Numerical aspects in the dynamic simulation of geometrically exact rods. Applied Numerical Mathematics,
62(10):1411–1427, 2012.
[93] H. Lang, J. Linn, and M. Arnold. Multi-body dynamics simulation of geometrically exact Cosserat rods. Multibody System Dynamics,
25(3):285–312, 2010.
[94] J. Langer and D. A. Singer. Lagrangian aspects of the Kirchhoff elastic rod. SIAM Review, 38(4):605–618, December 1996.
[95] A. Lazarus, J. J. T. Miller, and P. M. Reis. Continuation of equilibria and stability of slender elastic rods using an asymptotic numerical
method. Journal of the Mechanics and Physics of Solids, 61(8):1712–1736, 2013.
[96] T.-N. Le, J.-M. Battini, and M. Hjiaj. A consistent 3D corotational beam element for nonlinear dynamic analysis of flexible structures.
Computer Methods in Applied Mechanics and Engineering, 269:538–565, 2014.
[97] P. G. Lee and H. C. Sin. Locking-free straight beam element based on curvature. Communications in Numerical Methods in Engineering,
9(12):1005–1011, 1993.
[98] E. V. Lens and A. Cardona. A nonlinear beam element formulation in the framework of an energy preserving time integration scheme for
constrained multibody systems dynamics. Computers & Structures, 86(12):47–63, 2008.
[99] J. Linn. Discrete kinematics of Cosserat rods based on the difference geometry of framed curves. In The 4th Joint International Conference
on Multibody System Dynamics, Montréal, Canada, 2016.
[100] J. Linn, H. Lang, and A. Tuganov. Geometrically exact Cosserat rods with Kelvin-Voigt type viscous damping. Mechanical Sciences,
4(1):79–96, 2013.
[101] A. E. H. Love. A treatise on the mathematical theory of elasticity. Dover, 1944.
[102] M. Lyly, R. Stenberg, and T. Vihinen. A stable bilinear element for the Reissner-Mindlin plate model. Computer Methods in Applied
Mechanics and Engineering, 110(3–4):343–357, 1993.
[103] J. E. Marsden and T. J. R. Hughes. Mathematical foundations of elasticity. Dover, 1994.
[104] C. Meier, A. Popp, and W. A. Wall. An objective 3D large deformation finite element formulation for geometrically exact curved Kirchhoff
rods. Computer Methods in Applied Mechanics and Engineering, 278:445–478, 2014.
[105] C. Meier, A. Popp, and W. A. Wall. A locking-free finite element formulation and reduced models for geometrically exact Kirchhoff rods.
Computer Methods in Applied Mechanics and Engineering, 290:314–341, 2015.
[106] C. Meier, A. Popp, and W. A. Wall. A finite element approach for the line-to-line contact interaction of thin beams with arbitrary orientation.
Computer Methods in Applied Mechanics and Engineering, accepted for publication.
[107] C. Meier, A. Popp, and W. A. Wall. A unified approach for beam-to-beam contact. Computer Methods in Applied Mechanics and Engineering, submitted for publication.
[108] K. W. Müller, C. Meier, and W. A. Wall. Resolution of sub-element length scales in Brownian dynamics simulations of biopolymer networks
with geometrically exact beam finite elements. Journal of Computational Physics, 303:185–202, 2015.
[109] A. K. Noor and J. M. Peters. Mixed models and reduced/selective integration displacement models for nonlinear analysis of curved beams.
International Journal for Numerical Methods in Engineering, 17(4):615–631, 1981.
[110] E. Petrov and M. Géradin. Finite element theory for curved and twisted beams based on exact solutions for three-dimensional solids Part 1:
Beam concept and geometrically exact nonlinear formulation. Computer Methods in Applied Mechanics and Engineering, 165(1–4):43–92,
1998.
[111] G. Prathap. The curved beam/deep arch/finite ring element revisited. International Journal for Numerical Methods in Engineering,
21(3):389–407, 1985.
[112] G. Prathap and B. P. Naganarayana. Analysis of locking and stress oscillations in a general curved beam element. International Journal for
Numerical Methods in Engineering, 30(1):177–200, 1990.
[113] A. Quarteroni, R. Sacco, and S. Fausto. Numerical mathematics. Springer, 2000.
[114] E. Reissner. On one-dimensional finite-strain beam theory: The plane problem. Zeitschrift für Angewandte Mathematik und Physik (ZAMP),
23(5):795–804, 1972.
[115] E. Reissner. On finite deformations of space-curved beams. Zeitschrift für Angewandte Mathematik und Physik (ZAMP), 32(6):734–744,
1981.
[116] D. Rodney, B. Gadot, O. R. Martinez, S. R. Roscoat, and L. Orgéas. Reversible dilatancy in entangled single-wire materials. Nature
Materials, 15:72–77, 2016.
[117] I. Romero. The interpolation of rotations and its application to finite element models of geometrically exact rods. Computational Mechanics,
34:121–133, 2004.
90
[118] I. Romero. A comparison of finite elements for nonlinear beams: The absolute nodal coordinate and geometrically exact formulations.
Multibody System Dynamics, 20(1):51–68, 2008.
[119] I. Romero. Formulation and performance of variational integrators for rotating bodies. Computational Mechanics, 42(6):825–836, 2008.
[120] I. Romero and F. Armero. An objective finite element approximation of the kinematics of geometrically exact rods and its use in the formulation of an energymomentum conserving scheme in dynamics. International Journal for Numerical Methods in Engineering, 54(12):1683–
1716, 2002.
[121] I. Romero, M. Urrecha, and C. J. Cyron. A torsion-free non-linear beam model. International Journal of Non-Linear Mechanics, 58:1–10,
2014.
[122] O. Sander. Geodesic finite elements for Cosserat rods. International Journal for Numerical Methods in Engineering, 82(13):1645–1670,
2010.
[123] C. Sansour, T. L. Nguyen, and M. Hjiaj. An energy-momentum method for in-plane geometrically exact Euler-Bernoulli beam dynamics.
International Journal for Numerical Methods in Engineering, 102(2):99–134, 2015.
[124] C. Sansour and W. Wagner. Multiplicative updating of the rotation tensor in the finite element analysis of rods and shells - a path independent
approach. Computational Mechanics, 31(1):153–162, May 2003.
[125] H. A. F. A. Santos, P. M. Pimenta, and J. P. M. Almeida. A hybrid-mixed finite element formulation for the geometrically exact analysis of
three-dimensional framed structures. Computational Mechanics, 48(5):591–613, 2011.
[126] M. G. Schmidt, A. E. Ismail, and R. A. Sauer. A continuum mechanical surrogate model for atomic beam structures. International Journal
for Multiscale Computational Engineering, 13(5):413–442, 2015.
[127] M. Schulz and F. C. Filippou. Non-linear spatial Timoshenko beam element with curvature interpolation. International Journal for Numerical Methods in Engineering, 50(4):761–785, 2001.
[128] A. A. Shabana, H. A. Hussien, and J. L. Escalona. Application of the absolute nodal coordinate formulation to large rotation and large
deformation problems. Journal of Mechanical Design, 120(2):188–195, 1998.
[129] A. A. Shabana and R. Y. Yakoub. Three dimensional absolute nodal coordinate formulation for beam elements: Theory. Journal of
Mechanical Design, 123(4):606–613, December 2001.
[130] Y. Shi and J. E. Hearst. The Kirchhoff elastic rod, the nonlinear Schrödinger equation, and DNA supercoiling. The Journal of Chemical
Physics, 101(6):5186–5200, 1994.
[131] K. Shoemake. Animating rotation with quaternion curves. ACM SIGGRAPH Computer Graphics, 19(3):245–254, July 1985.
[132] J. C. Simo. A finite strain beam formulation. The three-dimensional dynamic problem. Part I. Computer Methods in Applied Mechanics and
Engineering, 49:55–70, 1985.
[133] J. C. Simo and T. J. R. Hughes. On the variational foundations of assumed strain methods. Journal of Applied Mechanics, 53:51–54, 1986.
[134] J. C. Simo and L. Vu-Quoc. A three-dimensional finite strain rod model. Part II: Computational aspects. Computer Methods in Applied
Mechanics and Engineering, 58:79–116, 1986.
[135] J. C. Simo and L. Vu-Quoc. On the dynamics in space of rods undergoing large motions - a geometrically exact approach. Computer
Methods in Applied Mechanics and Engineering, 66(2):125–161, 1988.
[136] J. C. Simo and K. K. Wong. Unconditionally stable algorithms for rigid body dynamics that exactly preserve energy and momentum.
International Journal for Numerical Methods in Engineering, 31(1):19–52, 1991.
[137] W. M. Smolenski. Statically and kinematically exact nonlinear theory of rods and its numerical verification. Computer Methods in Applied
Mechanics and Engineering, 178(1–2):89–113, 1998.
[138] V. Sonneville, A. Cardona, and O. Brüls. Geometric interpretation of a non-linear beam finite element on the Lie group SE(3). Archive of
Mechanical Engineering, 61:305–329, 2014.
[139] V. Sonneville, A. Cardona, and O. Brüls. Geometrically exact beam finite element formulated on the special Euclidean group. Computer
Methods in Applied Mechanics and Engineering, 268:451–474, 2014.
[140] R. A. Spurrier. Comment on ”singularity-free extraction of a quaternion from a direction-cosine matrix”. Journal of Spacecraft and Rockets,
15:255–255, 1978.
[141] H. Stolarski and T. Belytschko. Membrane locking and reduced integration for curved elements. Journal of Applied Mechanics, 49:172–176,
1982.
[142] G. Strang and G. Fix. An analysis of the finite elment method. Wellesley-Cambrigde Press, 2008.
[143] A. Tessler and L. Spiridigliozzi. Curved beam elements with penalty relaxation. International Journal for Numerical Methods in Engineering, 23(12):2245–2262, 1986.
[144] S. P. Timoshenko. On the correction for shear of the differential equation for transverse vibrations of prismatic bars. Philosophical Magazine
Series 6, 41(245):744–746, 1921.
[145] P. Češarek, M. Saje, and D. Zupan. Kinematically exact curved and twisted strain-based beam. International Journal of Solids and Structures,
49(13):1802–1817, 2012.
[146] T. D. Vu, D. Durville, and P. Davies. Finite element simulation of the mechanical behavior of synthetic braided ropes and validation on a
tensile test. International Journal of Solids and Structures, 58:106–116, 2015.
[147] W. A. Wall and M. W. Gee. BACI: A multiphysics simulation environment. Technical report, Technical University of Munich, 2016.
[148] Q. Wang and C. M. Wang. The constitutive relation and small scale parameter of nonlocal continuum mechanics for modelling carbon
nanotubes. Nanotechnology, 18(7):075702, 2007.
[149] O. Weeger, S.-K. Yeung, and M. L. Dunn. Isogeometric collocation methods for Cosserat rods and rod structures. Computer Methods in
Applied Mechanics and Engineering, page http://dx.doi.org/10.1016/j.cma.2016.05.009, 2016.
[150] H. Weiss. Dynamics of geometrically nonlinear rods: I. Mechanical models and equations of motion. Nonlinear Dynamics, 30(4):357–381,
2002.
[151] H. Weiss. Dynamics of geometrically nonlinear rods: II. Numerical methods and computational examples. Nonlinear Dynamics, 30(4):383–
415, 2002.
[152] G. Wempner. Finite elements, finite rotations and small strains of flexible shells. International Journal of Solids and Structures, 5(2):117–
91
153, 1969.
[153] Y. Yang, I. Tobias, and W. K. Olson. Finite element analysis of DNA supercoiling. The Journal of Chemical Physics, 98(2):1673–1686,
1993.
[154] Z. Zhang, Z. Qi, Z. Wu, and H. Fang. A spatial Euler-Bernoulli beam element for rigid-flexible coupling dynamic analysis of flexible
structures. Shock and Vibration, 2015:208127, 2015.
[155] Z. Zhao and G. Ren. A quaternion-based formulation of Euler–Bernoulli beam without singularity. Nonlinear Dynamics, 67(3):1825–1835,
2012.
[156] D. Zupan and M. Saje. Finite-element formulation of geometrically exact three-dimensional beam theories based on interpolation of strain
measures. Computer Methods in Applied Mechanics and Engineering, 192(49–50):5209–5248, 2003.
[157] D. Zupan and M. Saje. The linearized three-dimensional beam theory of naturally curved and twisted beams: The strain vectors formulation.
Computer Methods in Applied Mechanics and Engineering, 195(33–36):4557–4578, 2006.
[158] E. Zupan, M. Saje, and D. Zupan. On a virtual work consistent three-dimensional Reissner–Simo beam formulation using the quaternion
algebra. Acta Mechanica, 224(8):1709–1729, 2013.
92
| 5cs.CE
|
arXiv:1802.06259v1 [cs.CV] 17 Feb 2018
Exact and Consistent Interpretation for Piecewise Linear
Neural Networks: A Closed Form Solution
Lingyang Chu
Xia Hu
Juhua Hu
Simon Fraser University
Burnaby, Canada
[email protected]
Simon Fraser University
Burnaby, Canada
[email protected]
Simon Fraser University
Burnaby, Canada
[email protected]
Lanjun Wang
Jian Pei
Huawei Technology Co. Ltd
Beijing, China
[email protected]
JD.com and Simon Fraser University
Beijing/Burnaby, China/Canada
[email protected]
ABSTRACT
Strong intelligent machines powered by deep neural networks are
increasingly deployed as black boxes to make decisions in risksensitive domains, such as finance and medical. To reduce potential
risk and build trust with users, it is critical to interpret how such
machines make their decisions. Existing works interpret a pretrained neural network by analyzing hidden neurons, mimicking
pre-trained models or approximating local predictions. However,
these methods do not provide a guarantee on the exactness and
consistency of their interpretation. In this paper, we propose an
elegant closed form solution named OpenBox to compute exact and
consistent interpretations for the family of Piecewise Linear Neural
Networks (PLNN). The major idea is to first transform a PLNN into
a mathematically equivalent set of linear classifiers, then interpret
each linear classifier by the features that dominate its prediction.
We further apply OpenBox to demonstrate the effectiveness of nonnegative and sparse constraints on improving the interpretability
of PLNNs. The extensive experiments on both synthetic and real
world data sets clearly demonstrate the exactness and consistency
of our interpretation.
KEYWORDS
Deep neural network, exact and consistent interpretation, closed
form.
ACM Reference Format:
Lingyang Chu, Xia Hu, Juhua Hu, Lanjun Wang, and Jian Pei . 2018. Exact
and Consistent Interpretation for Piecewise Linear Neural Networks: A
Closed Form Solution. In Proceedings of ACM (arXiv version). ACM, New
York, NY, USA, 9 pages. https://doi.org/10.1145/nnnnnnn.nnnnnnn
1
INTRODUCTION
More and more machine learning systems are making significant
decisions routinely in important domains, such as medical practice,
autonomous driving, criminal justice, and military decision making [15]. As the impact of machine-made decisions increases, the
demand on clear interpretations of machine learning systems is
growing ever stronger against the blind deployments of decision
machines [17]. Accurately and reliably interpreting a machine learning model is the key to many significant tasks, such as identifying
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
arXiv version, 2017, Canada
© 2018 Copyright held by the owner/author(s).
ACM ISBN 978-x-xxxx-xxxx-x/YY/MM.
https://doi.org/10.1145/nnnnnnn.nnnnnnn
failure models [1], building trust with human users [36], discovering
new knowledge [35], and avoiding unfairness issues [46].
The interpretation problem of machine learning models has been
studied for decades. Conventional models, such as Logistic Regression and Support Vector Machine, have all been well interpreted
from both practical and theoretical perspectives [4]. Powerful nonnegative and sparse constraints are also developed to enhance the
interpretability of conventional models by sparse feature selection [21, 27]. However, due to the complex network structure of a
deep neural network, the interpretation problem of modern deep
models is yet a challenging field that awaits further exploration.
As to be reviewed in Section 2, the existing studies interpret
a deep neural network in three major ways. The hidden neuron
analysis methods [9, 29, 32, 45] analyze and visualize the features
learned by the hidden neurons of a neural network; the model mimicking methods [2, 3, 7, 20] build a transparent model to imitate
the classification function of a deep neural network; the local explanation methods [11, 38, 40, 42] study the predictions on local
perturbations of an input instance, so as to provide decision features
for interpretation. All these methods gain useful insights into the
mechanism of deep models. However, there is no guarantee that
what they compute as an interpretation is truthfully the exact behavior of a deep neural network. As demonstrated by Ghorbani [13],
most existing interpretation methods are inconsistent and fragile,
because two perceptively indistinguishable instances with the same
prediction result can be easily manipulated to have dramatically
different interpretations.
Can we compute an exact and consistent interpretation for a pretrained deep neural network? In this paper, we provide an affirmative answer, as well as an elegant closed form solution for the
family of piecewise linear neural networks. Here, a piecewise linear neural network (PLNN) [18] is a neural network that adopts
a piecewise linear activation function, such as MaxOut [16] and the
family of ReLU [14, 19, 31]. The wide applications [26] and great
practical successes [25] of PLNNs call for exact and consistent interpretations on the overall behaviour of this type of neural networks.
We make the following technical contributions.
First, we prove that a PLNN is mathematically equivalent to a set
of local linear classifiers, each of which being a linear classifier that
classifies a group of instances within a convex polytope in the input
space. Second, we propose a method named OpenBox to provide
an exact interpretation of a PLNN by computing its equivalent set
of local linear classifiers in closed form. Third, we interpret the
classification result of each instance by the decision features of
its local linear classifier. Since all instances in the same convex
polytope share the same local linear classifier, our interpretations
are consistent per convex polytope. Fourth, we also apply OpenBox
arXiv version, 2017, Canada
to study the effect of non-negative and sparse constraints on the
interpretability of PLNNs. We find that a PLNN trained with these
constraints selects meaningful features that dramatically improve
the interpretability. Last, we conduct extensive experiments on both
synthetic and real-world data sets to verify the effectiveness of our
method.
The rest of this paper is organized as follows. We review the
related works in Section 2. We formulate the problem in Section 3
and present OpenBox in Section 4. We report the experimental
results in Section 5, and conclude the paper in Section 6.
2
RELATED WORKS
How to interpret the overall mechanism of deep neural networks
is an emergent and challenging problem.
2.1
Hidden Neuron Analysis Methods
The hidden neuron analysis methods [9, 29, 32, 45] interpret a
pre-trained deep neural network by visualizing, revert-mapping or
labeling the features that are learned by the hidden neurons.
Yosinski et al. [45] visualized the live activations of the hidden
neurons of a ConvNet, and proposed a regularized optimization
to produce a qualitatively better visualization. Erhan et al. [10]
proposed an activation maximization method and a unit sampling method to visualize the features learned by hidden neurons.
Cao et al. [5] visualized a neural network’s attention on its target
objects by a feedback loop that infers the activation status of the
hidden neurons. Li et al. [28] visualized the compositionality of
clauses by analyzing the outputs of hidden neurons in a neural
model for Natural Language Processing.
To understand the features learned by the hidden neurons, Mahendran et al. [29] proposed a general framework that revert-maps
the features learned from an image to reconstruct the image. Dosovitskiy et al. [9] performed the same task as Mahendran et al. [29]
did by training an up-convolutional neural network.
Zhou et al. [47] interpreted a CNN by labeling each hidden neuron with a best aligned human-understandable semantic concept.
However, it is hard to get a golden dataset with accurate and complete labels of all human semantic concepts.
The hidden neuron analysis methods provide useful qualitative
insights into the properties of each hidden neuron. However, qualitatively analyzing every neuron does not provide much actionable
and quantitative interpretation about the overall mechanism of the
entire neural network [12].
2.2
Model Mimicking Methods
By imitating the classification function of a neural network, the
model mimicking methods [2, 3, 7, 20] build a transparent model
that is easy to interpret and achieves a high classification accuracy.
Ba et al. [2] proposed a model compression method to train a
shallow mimic network using the training instances labeled by
one or more deep neural networks. Hinton et al. [20] proposed a
distillation method that distills the knowledge of a large neural
network by training a relatively smaller network to mimic the
prediction probabilities of the original large network. To improve
the interpretability of distilled knowledge, Frosst and Hinton [12]
extended the distillation method [20] by training a soft decision
tree to mimic the prediction probabilities of a deep neural network.
Che et al. [7] proposed a mimic learning method to learn interpretable phenotype features. Wu et al. [43] proposed a tree regularization method that uses a binary decision tree to mimic and
regularize the classification function of a deep time-series model.
Lingyang Chu, Xia Hu, Juhua Hu, Lanjun Wang, and Jian Pei
The mimic models built by model mimicking methods are much
simpler to interpret than deep neural networks. However, due to the
reduced model complexity of a mimic model, there is no guarantee
that a deep neural network with a large VC-dimension [18, 24, 41]
can be successfully imitated by a simpler shallow model. Thus, there
is always a gap between the interpretation of a mimic model and
the actual overall mechanism of the target deep neural network.
2.3
Local Interpretation Methods
The local interpretation methods [11, 38, 40, 42] compute and visualize the important features for an input instance by analyzing the
predictions of its local perturbations.
Simonyan et al. [39] generated a class-representative image and
a class-saliency map for each class of images by computing the
gradient of the class score with respect to an input image. Ribeiro et
al. [36] proposed LIME to interpret the predictions of any classifier
by learning an interpretable model in the local region around the
input instance.
Zhou et al. [48] proposed CAM to identify discriminative image
regions for each class of images using the global average pooling
in CNNs. Selvaraju et al. [37] generalized CAM [48] by Grad-CAM,
which identifies important regions of an image by flowing classspecific gradients into the final convolutional layer of a CNN.
Koh et al. [23] used influence functions to trace a model’s prediction and identify the training instances that are the most responsible
for the prediction.
The local interpretation methods generate an insightful individual interpretation for each input instance. However, the interpretations for perspectively indistinguishable instances may not be
consistent [13], and can be purposefully manipulated by a simple
transformation of the input instance without affecting the prediction result [22].
3
PROBLEM DEFINITION
For a PLNN N that contains L layers of neurons, we write the l-th
layer of N as Ll . Hence, L1 is the input layer, LL is the output
layer, and the other layers Ll , l ∈ {2, . . . , L−1} are hidden layers.
A neuron in a hidden layer is called a hidden neuron. Let nl
represent the number of neurons in Ll , the total number of hidden
Í
neurons in N is computed by N = lL−1
n .
=2 l
(l )
(l −1)
Denote by ui the i-th neuron in Ll , by bi
its bias, by
(l )
(l )
ai its output, and by zi the total weighted sum of its inputs.
For all the nl neurons in Ll , we write their biases as a vector
(l −1)
(l −1)
b(l −1) = [b1 , . . . , bnl ]⊤ , their outputs as a vector a(l ) =
(l )
(l )
(l )
(l )
[a1 , . . . , anl ]⊤ , and their inputs as a vector z(l ) = [z1 , . . . , znl ]⊤ .
Neurons in successive layers are connected by weighted edges.
(l )
Denote by Wi j the weight of the edge between the i-th neuron
in Ll +1 and the j-th neuron in Ll , that is, W (l ) is an nl +1 -by-nl
matrix. For l ∈ {1, . . . , L − 1}, we compute z(l +1) by
z(l +1) = W (l ) a(l ) + b(l )
(1)
Denote by f : R → R the piecewise linear activation function
(l )
(l )
for each neuron in the hidden layers of N . We have ai = f (zi )
for all l ∈ {2, . . . , L − 1}. We extend f to apply to vectors in an
(l )
(l )
element-wise fashion, such that f (z(l ) ) = [f (z1 ), . . . , f (znl )]⊤ .
(l
)
Then, we compute a for all l ∈ {2, . . . , L − 1} by
a(l ) = f (z(l ) )
(2)
Exact and Consistent Interpretation for Piecewise Linear
Neural Networks: A Closed Form Solution
Table 1: Frequently used notations.
arXiv version, 2017, Canada
4.1
The Configuration of a PLNN
(l )
Notation
(l )
ui
nl
N
(l )
zi
(l )
ci
Ch
Ph
Fh (·)
Qh
Description
The i-th neuron in layer Ll .
The number of neurons in layer Ll .
The total number of hidden neurons in N.
The input of the i-th neuron in layer Ll .
The configuration of the i-th neuron in layer Ll .
The h-th configuration of the PLNN N.
The h-th convex polytope determined by Ch .
The h-th linear classifier that is determined by Ch .
The set of linear inequalities that define Ph .
For a hidden neuron ui , the piecewise linear activation function
(l )
f (zi )
is in the following form.
(l )
(l )
r 1 zi + t 1 , if zi ∈ I 1
(l )
(l )
r 2 zi + t 2 , if zi ∈ I 2
(l )
f (zi ) =
..
.
r z(l ) + t , if z(l ) ∈ I
k i
k
k
i
(3)
(l )
An input instance of N is denoted by x ∈ X, where X ⊆ Rd is
a d-dimensional input space. x is also called an instance for short.
Denote by xi the i-th dimension of x. The input layer L1 contains
(1)
n 1 = d neurons, where ai = xi for all i ∈ {1, . . . , d}. That is,
(1)
a = x.
The output of N is a(L) ∈ Y, where Y ⊆ Rn L is an n L dimensional output space. The output layer LL adopts the softmax
function to compute the output by a(L) = softmax(z(L) ).
A PLNN works as a classification function F : X → Y that
maps an input x ∈ X to an output a(L) ∈ Y. It is widely known
that F (·) is a piecewise linear function [30, 34]. However, due to the
complex network of a PLNN, the overall behaviour of F (·) is hard
to understand. Thus, a PLNN is usually regarded as a black box.
How to interpret the overall behavior of a PLNN in a humanunderstandable manner is an interesting problem that has attracted
much attention in recent years.
Following a principled approach of interpreting a machine learning model [4], we regard an interpretation of a PLNN N as the
decision features that define the decision boundary of N . We call a
model interpretable if it explicitly provides its interpretation (i.e.,
decision features) in closed form.
Definition 3.1. Given a fixed PLNN N with constant structure
and parameters, our task is to interpret the overall behaviour of N
by computing an interpretable model M that satisfies the following
requirements.
• Exactness: M is mathematically equivalent to N such that
the interpretations provided by M truthfully describe the
exact behaviour of N .
• Consistency: M provides similar interpretations for classification of similar instances.
Table 1 summarizes a list of frequently used notations.
4
THE OPENBOX METHOD
In this section, we describe the OpenBox method, which produces
an exact and consistent interpretation of a PLNN by computing an
interpretation model M in a piecewise linear closed form.
We first define the configuration of a PLNN N , which specifies
the activation status of each hidden neuron in N . Then, we illustrate
how to interpret the classification result of a fixed instance. Last, we
illustrate how to interpret the overall behavior of N by computing
an interpretation model M that is mathematically equivalent to N .
where k ≥ 1 is a constant integer, f (zi ) consists of k linear functions, {r 1 , . . . , r k } are constant slopes, {t 1 , . . . , tk } are constant
intercepts, and {I 1 , . . . , Ik } is a collection of constant real intervals that partition R.
Given a fixed PLNN N , an instance x ∈ X determines the value
(l )
(l )
of zi , and further determines a linear function in f (zi ) to apply.
(l )
According to which linear function in f (zi ) is applied, we encode
the activation status of each hidden neuron by k states, each of
which uniquely corresponds to one of the k linear functions of
(l )
(l )
(l )
(l )
f (zi ). Denote by ci ∈ {1, . . . , k} the state of ui , we have zi ∈
(l )
(l )
Iq if and only if ci = q (q ∈ {1, . . . , k }). Since the inputs zi ’s
are different from neuron to neuron, the states of different hidden
neurons may differ from each other.
(l )
(l )
Denote by a vector c(l ) = [c1 , . . . , cnl ] the states of all hidden
neurons in Ll . The configuration of N is an N -dimensional vector,
denoted by C = [c(2) , . . . , c(L−1) ], which specifies the states of all
hidden neurons in N .
The configuration C of a fixed PLNN is uniquely determined by
the instance x. We write the function that maps an instance x ∈ X
to a configuration C ∈ {1, . . . , k} N as conf : X → {1, . . . , k} N .
(l )
(l )
(l )
For a neuron ui , denote by variables ri and ti the slope and
intercept, respectively, of the linear function that corresponds to
(l ) (l )
(l )
(l )
the state ci . ri and ti are uniquely determined by ci , such that
(l )
(l )
(l )
ri = rq and ti = tq , if and only if ci = q (q ∈ {1, . . . , k }).
For all hidden neurons in Ll , we write the variables of slopes and
(l )
(l )
(l )
(l )
intercepts as r(l ) = [r1 , . . . , rnl ]⊤ and t(l ) = [t1 , . . . , tnl ]⊤ , respectively. Then, we rewrite the activation function for all neurons
in a hidden layer Ll as
f (z(l ) ) = r(l ) ◦ z(l ) + t(l )
(4)
where r(l ) ◦ z(l ) is the Hadamard product between r(l ) and z(l ) .
Next, we interpret the classification result of a fixed instance.
4.2
Exact Interpretation for the Classification
Result of a Fixed Instance
Given a fixed PLNN N , we interpret the classification result of a
fixed instance x ∈ X by deriving the closed form of F (x) as follows.
Following Equations 2 and 4, we have, for all l ∈ {2, . . . , L − 1}
a(l ) = f (z(l ) ) = r(l ) ◦ z(l ) + t(l )
By plugging a(l ) into Equation 1, we rewrite z(l +1) as
z(l +1) = W (l ) (r(l ) ◦ z(l ) + t(l ) ) + b(l ) = W̃ (l ) z(l ) + b̃(l )
(5)
arXiv version, 2017, Canada
Lingyang Chu, Xia Hu, Juhua Hu, Lanjun Wang, and Jian Pei
where b̃(l ) = W (l ) t(l ) + b(l ) , and W̃ (l ) = W (l ) ◦ r(l ) is an extended
version of Hadamard product, such that the entry at the i-th row
(l )
(l ) (l )
and j-th column of W̃ (l ) is W̃i j = Wi j rj .
By iteratively plugging Equation 5 into itself, we can write z(l +1)
for all l ∈ {2, . . . , L − 1} as
z(l +1) =
l −2
Ö
W̃ (l −h) z(2) +
l l −h−1
Õ
Ö
Algorithm 1: OpenBox(N, D train )
Input: N B a fixed PLNN, D train ⊂ X the set of training
instances used to train N .
Output: M B a set of active LLCs
1:
2:
3:
4:
5:
6:
7:
8:
W̃ (l −q) b̃(h)
h=2 q=0
h=0
By plugging z(2) = W (1) a(1) + b(1) and a(1) = x into the above
equation, we rewrite z(l +1) , for all l ∈ {2, . . . , L − 1}, as
z(l +1) =
l −2
Ö
W̃ (l −h)W (1) x +
h=0
l −2
Ö
h=0
W̃ (l −h) b(1) +
l l −h−1
Õ
Ö
W̃ (l −q) b̃(h)
h=2 q=0
= Ŵ (1:l ) x + b̂(1:l )
(6)
Î −2 (l −h) (1)
where Ŵ (1:l ) = lh=0
W̃
W is the coefficient matrix of x,
and b̂(1:l ) is the sum of the remaining terms. The superscript (1 : l)
indicates that Ŵ (1:l ) x + b̂(1:l ) is equivalent to PLNN’s forward propagation from layer L1 to layer Ll .
Since the output of N on an input x ∈ X is F (x) = a(L) =
softmax(z(L) ), the closed form of F (x) is
F (x) = softmax(Ŵ (1:L−1) x + b̂(1:L−1) )
(7)
For a fixed PLNN N and a fixed instance x, Ŵ (1:L−1) and b̂(1:L−1)
are constant parameters uniquely determined by the fixed configuration C = conf (x). Therefore, for a fixed input instance x, F (x) is
a linear classifier whose decision boundary is explicitly defined
by Ŵ (1:L−1) x + b̂(1:L−1) .
Inspired by the interpretation method widely used by conventional linear classifiers, such as Logistic Regression and linear
SVM [4], we interpret the prediction on a fixed instance x by the
decision features of F (x). Specifically, the entries of the i-th row of
Ŵ (1:L−1) are the decision features for the i-th class of instances.
Equation 7 provides a straightforward way to interpret the classification result of a fixed instance. However, individually interpreting the classification result of every single instance is far from
the understanding of the overall behavior of a PLNN N . Next, we
describe how to interpret the overall behavior of N by computing
an interpretation model M that is mathematically equivalent to N .
4.3
Initialization: M = ∅, C = ∅.
for each x ∈ D train do
Compute the configuration by Ch ← conf (x).
if Ch < C then
C ← C ∪ Ch and M ← M ∪ (Fh (x), Ph ).
end if
end for
return M.
Exact Interpretation of a PLNN
A fixed PLNN N with N hidden neurons has at most k N configurations. We represent the h-th configuration by Ch ∈ C, where
C ⊆ {1, . . . , k } N is the set of all configurations of N .
Recall that each instance x ∈ X uniquely determines a configuration conf (x) ∈ C. Since the volume of C, denoted by |C|, is at
most k N , but the number of instances in X can be arbitrarily large,
it is clear that at least one configuration in C should be shared by
more than one instances in X.
Denote by Ph = {x ∈ X | conf (x) = Ch } the set of instances
that have the same configuration Ch . We prove in Theorem 4.1 that
for any configuration Ch ∈ C, Ph is a convex polytope in X.
Theorem 4.1. Given a fixed PLNN N with N hidden neurons,
∀Ch ∈ C, Ph = {x ∈ X | conf(x) = Ch } is a convex polytope in X.
Proof. We prove by showing that conf (x) = Ch is equivalent
to a finite set of linear inequalities with respect to x.
When l = 2, we have z(2) = W (1) x + b(1) . For l ∈ {3, . . . , L − 1}, it
follows Equation 6 that z(l ) = Ŵ (1:l −1) x + b̂(1:l −1) , which is a linear
function of x, because Ŵ (1:l ) and b̂(1:l ) are constant parameters
when Ch is fixed. In summary, given a fixed Ch , z(l ) is a linear
function of x for all l ∈ {2, . . . , L − 1}.
We show that Ph is a convex polytope by showing that conf (x) =
Ch is equivalent to a set of 2N linear inequalities with respect to
(l )
(l )
x. Recall that zi ∈ Iq if and only if ci = q (q ∈ {1, . . . , k }).
Denote by ψ : {1, . . . , k } → {I 1 , . . . , Ik } the bijective function that
(l )
maps a configuration ci to a real interval in {I 1 , . . . , Ik }, such
(l )
(l )
that ψ (ci ) = Iq if and only if ci = q (q ∈ {1, . . . , k}). Then,
conf (x) = Ch is equivalent to a set of constraints, denoted by
(l )
(l )
(l )
Qh = {zi ∈ ψ (ci ) | i ∈ {1, . . . , nl }, l ∈ {2, . . . , L − 1}}. Since zi
(l )
is a linear function of x and ψ (ci ) is a real interval, each constraint
(l )
(l )
zi ∈ ψ (ci ) in Qh is equivalent to two linear inequalities with
respect to x. Therefore, conf (x) = Ch is equivalent to a set of 2N
linear inequalities, which means Ph is a convex polytope.
□
According to Theorem 4.1, all instances sharing the same configuration Ch form a unique convex polytope Ph that is explicitly
defined by 2N linear inequalities in Qh . Since Ch also determines
the linear classifier for a fixed instance in Equation 7, all instances
in the same convex polytope Ph share the same linear classifier
determined by Ch .
Denote by Fh (·) the linear classifier that is shared by all instances
in Ph , we can interpret N as a set of local linear classifiers
(LLCs), each LLC being a linear classifier Fh (·) that applies to all
instances in a convex polytope Ph . Denote by a tuple (Fh (·), Ph ) the
h-th LLC, a fixed PLNN N is equivalent to a set of LLCs, denoted by
M = {(Fh (·), Ph ) | Ch ∈ C}. We use M as our final interpretation
model for N .
For a fixed PLNN N , if the states of the N hidden neurons are
independent, the PLNN N has k N configurations, which means M
contains k N LLCs. However, due to the hierarchical structure of a
PLNN, the states of a hidden neuron in Ll strongly correlate with
the states of the neurons in the former layers Lq (q < l). Therefore,
the volume of C is much less than k N , and the number of local linear
classifiers in M is much less than k N . We discuss this phenomenon
later in Table 3 and Section 5.4.
In practice, we do not need to compute the entire set of LLCs
in M all at once. Instead, we can first compute an active subset
of M, that is, the set of LLCs that are actually used to classify the
available set of instances. Then, we can update M whenever a new
LLC is used to classify a newly coming instance.
Exact and Consistent Interpretation for Piecewise Linear
Neural Networks: A Closed Form Solution
(l )
are the coefficients of x in zi .
We also discover that some linear inequalities in Qh are redundant whose hyperplanes do not intersect with Ph . To simplify our
interpretation on the polytope boundaries, we remove such redundant inequalities by Caron’s method [6] and focus on studying the
PBFs of the non-redundant ones.
The advantages of OpenBox are three-fold as follows. First, our
interpretation is exact, because the set of LLCs in M are mathematically equivalent to the classification function F (·) of N . Second,
our interpretation is group-wise consistent. It is due to the reason
that all instances in the same convex polytope are classified by exactly the same LLC, and thus the interpretations are consistent with
respect to a given convex polytope. Last, our interpretation is easy
to compute, since OpenBox computes M by a one-time forward
propagation through N for each instance in D train .
Data Sets
SYN
FMNIST-1
FMNIST-2
1 https://github.com/marcotcr/lime
PLNN-NS
✓
×
LR
×
×
LR-F
×
✓
LR-NS
✓
×
LR-NSF
✓
✓
# Neurons
(n 1, n 2, . . . , n L )
(2, 4, 16, 2, 2)
(784, 8, 2, 2)
(784, 8, 2, 2)
PLNN
|C| kN
266 222
78
210
23
210
PLNN-NS
|C| kN
41
222
3
210
18
210
Table 4: Detailed description of data sets.
Training Data
# Positive
# Negative
6,961
13,039
4,000
4,000
4,000
4,000
Data Sets
SYN
FMNIST-1
FMNIST-2
1.5
x2
EXPERIMENTS
In this section, we evaluate the performance of OpenBox, and compare it with the state-of-the-art method LIME [36]. In particular, we
address the following questions: (1) What are the LLCs look like?
(2) Are the interpretations produced by LIME and OpenBox exact
and consistent? (3) Are the decision features of LLCs easy to understand, and can we improve the interpretability of these features by
non-negative and sparse constraints? (4) How to interpret the PBFs
of LLCs? (5) How effective are the interpretations of OpenBox in
hacking and debugging a PLNN model?
Table 2 shows the details of the six models we used. For both
PLNN and PLNN-NS, we use the same network structure described
in Table 3, and adopt the widely used activation function: ReLU [14].
We apply the non-negative and sparse constraints proposed by
Chorowski et al. [8] to train PLNN-NS. Since our goal is to comprehensively study the interpretation effectiveness of OpenBox
rather than achieving state-of-the-art classification performance,
we use relatively simple network structures for PLNN and PLNNNS, which are still powerful enough to achieve significantly better
classification performance than Logistic Regression (LR). The decision features of LR, LR-F, LR-NS and LR-NSF are used as baselines
to compare with the decision features of LLCs.
The Python code of LIME is published by its authors1 . The other
methods and models are implemented in Matlab. PLNN and PLNNNS are trained using the DeepLearnToolBox [33]. All experiments
are conducted on a PC with a Core-i7-3370 CPU (3.40 GHz), 16GB
main memory, and a 5,400 rpm hard drive running Windows 7 OS.
We use the following data sets. Detailed information of the data
sets is shown in Table 4.
Synthetic (SYN) Data Set. As shown in Figure 1(a), this data set
contains 20,000 instances uniformly sampled from a quadrangle in
PLNN
×
×
Table 3: The network structures (n 1 , n 2 , . . . , n L ) and the number of configurations |C| of PLNN and PLNN-NS. The neurons in successive layers are initialized to be fully connected.
k = 2 is the number of linear functions of ReLU, N is the
number of hidden neurons.
Testing Data
# Positive
# Negative
N/A
N/A
3,000
3,000
3,000
3,000
1.5
1
1
0.5
0.5
0
-0.5
0
-0.5
-1
-1
-1.5
-1.5
-1.5
-1
-0.5
0
0.5
1
1.5
-1.5
-1
-0.5
x1
0
0.5
1
1.5
x1
(a) training data of SYN
(b) prediction results of PLNN
1.5
x2
5
Models
NS
Flip
x2
(l )
By Equation 6, zi is a linear function with respect to x. The PBFs
Table 2: The models to interpret. LR is Logistic Regression.
NS means non-negative and sparse constraints. Flip means
the model is trained on the instances with flipped labels.
1.5
1
1
0.5
0.5
x2
Algorithm 1 summarizes the OpenBox method, which computes
M as the active set of LLCs that are actually used to classify the
set of training instances, denoted by D train .
Now, we are ready to introduce how to interpret the classification
result of an instance x ∈ Ph , h ∈ {1, . . . , |C|}. First, we interpret
the classification result of x using the decision features of Fh (x)
(Section 4.2). Second, we interpret why x is contained in Ph using
the polytope boundary features (PBFs), which are the decision
features of the polytope boundaries. More specifically, a polytope
(l )
(l )
boundary of Ph is defined by a linear inequality zi ∈ ψ (ci ) in Qh .
arXiv version, 2017, Canada
0
-0.5
0
-0.5
-1
-1
-1.5
-1.5
-1.5
-1
-0.5
0
0.5
1
1.5
-1.5
x1
(c) convex polytopes
-1
-0.5
0
0.5
1
1.5
x1
(d) LLCs
Figure 1: The LLCs of the PLNN trained on SYN.
2-dimensional Euclidean space. The red and blue points are positive
and negative instances, respectively. We use all instances in SYN as
training data to visualize the LLCs of a PLNN.
FMNIST-1 and FMNIST-2 Data Sets. Each of these data sets
contains two classes of images in the Fashion MNIST data set [44].
FMNIST-1 consists of the images of Ankle Boot and Bag. FMNIST-2
consists of the images of Coat and Pullover. All images in FMNIST1 and FMNIST-2 are 28-by-28 grayscale images. We represent an
image by cascading the 784 pixel values into a 784-dimensional
feature vector. The Fashion MNIST data set is available online2 .
2 https://github.com/zalandoresearch/fashion-mnist
arXiv version, 2017, Canada
5.2
Are the Interpretations Exact and
Consistent?
Exact and consistent interpretations are naturally favored by human
minds. In this subsection, we systematically study the exactness
and consistency of the interpretations of LIME and OpenBox on
FMNIST-1 and FMNIST-2. Since LIME is too slow to process all
instances in 24 hours, for each of FMNIST-1 and FMNIST-2, we
uniformly sample 600 instances from the testing set, and conduct
the following experiments on the sampled instances.
We first analyze the exactness of interpretation by comparing
the predictions computed by the local interpretable model of LIME,
the LLCs of OpenBox and PLNN, respectively. The prediction of an
instance is the probability of classifying it as a positive instance.
In Figure 2, since LIME does not guarantee zero approximation
error on the local predictions of PLNN, the predictions of LIME are
not exactly the same as PLNN on FMNIST-1, and are dramatically
different from PLNN on FMNIST-2. The difference of predictions is
more significant on FMNIST-2, because the images in FMNIST-2 are
more difficult to distinguish, which makes the decision boundary
of PLNN more complicated and harder to approximate. We can also
see that the predictions of LIME exceed [0, 1]. This is because the
output of the interpretable model of LIME is not a probability at all.
As a result, it is arguable that the interpretations computed by LIME
may not truthfully describe the exact behavior of PLNN. In contrast,
since the set of LLCs computed by OpenBox is mathematically
equivalent to F (·) of PLNN, the predictions of OpenBox are exactly
the same as PLNN on all instances. Therefore, the decision features
of LLCs exactly describe the overall behavior of PLNN.
Next, we study the interpretation consistency of LIME and
OpenBox by analyzing the similarity between the interpretations
of similar instances.
In general, a consistent interpretation method should provide
similar interpretations for similar instances. For an instance x, denote by x ′ the nearest neighbor of x by Euclidean distance, by
γ , γ ′ ∈ Rd the decision features for the classification of x and x ′ ,
respectively. We measure the consistency of interpretation by the
cosine similarity between γ and γ ′ , where a larger cosine similarity
indicates a better interpretation consistency.
PLNN,OpenBox
LIME
Prediction
LIME
Prediction
1
0.5
0
0
200
400
PLNN,OpenBox
1
0.5
0
600
0
200
400
Index of Instance
Index of Instance
(a) FMNIST-1
(b) FMNIST-2
600
Figure 2: The predictions of LIME, OpenBox and PLNN. We
sort the results by PLNN’s predictions in descending order.
LIME
OpenBox
LIME
1
0.5
0
OpenBox
1
CS
What Are the LLCs Look Like?
CS
5.1
We demonstrate our claim in Theorem 4.1 by visualizing the LLCs
of the PLNN trained on SYN.
Figures 1(a)-(b) show the training instances of SYN and the prediction results of PLNN, respectively. Since all instances are used
for training, the prediction accuracy is 99.9%.
In Figure 1(c), we plot all instances with the same configuration in
the same colour. Clearly, all instances with the same configuration
are contained in the same convex polytope. This demonstrates our
claim in Theorem 4.1.
Figure 1(d) shows the LLCs whose convex polytopes cover the
decision boundary of PLNN and contain both positive and negative
instances. As it is shown, the solid lines show the decision boundaries of the LLCs, which capture the difference between positive
and negative instances, and form the overall decision boundary
of PLNN. A convex polytope that does not cover the boundary of
PLNN contains a single class of instances. The LLCs of these convex
polytopes capture the common features of the corresponding class
of instances. As to be analyzed in the following subsections, the
set of LLCs produce exactly the same prediction as PLNN, and also
capture meaningful decision features that are easy to understand.
Lingyang Chu, Xia Hu, Juhua Hu, Lanjun Wang, and Jian Pei
0
200
400
600
0.5
0
0
200
400
Index of Instance
Index of Instance
(a) FMNIST-1
(b) FMNIST-2
600
Figure 3: The cosine similarity (CS) between the decision features of each instance and its nearest neighbour. The results
of LIME and OpenBox are separately sorted by cosine similarity in descending order.
As shown in Figure 3, the cosine similarity of OpenBox is equal
to 1 on about 50% of the instances, because OpenBox consistently
gives the same interpretation for all instances in the same convex
polytope. Since the nearest neighbours x and x ′ may not belong to
the same convex polytope, the cosine similarity of OpenBox is not
always equal to 1 on all instances. In constrast, since LIME computes
individual interpretation based on the unique local perturbations
of every single instance, the cosine similarity of LIME is significantly lower than OpenBox on all instances. This demonstrates the
superior interpretation consistency of OpenBox.
In summary, the interpretations of OpenBox are exact, and are
much more consistent than the interpretations of LIME.
5.3
Decision Features of LLCs and the Effect of
Non-negative and Sparse Constraints
Besides exactness and consistency, a good interpretation should
also have a strong semantical meaning, such that the “thoughts”
of an intelligent machine can be easily understood by a human
brain. In this subsection, we first show the meaning of the decision
features of LLCs, then study the effect of the non-negative and
sparse constraints in improving the interpretability of the decision
features. The decision features of PLNN and PLNN-NS are computed
by OpenBox. The decision features of LR, LR-F, LR-NS and LR-NSF
are used as baselines. Table 5 shows the accuracy of all models.
Figure 4 shows the decision features of all models on FMNIST-1.
Interestingly, the decision features of PLNN are as easy to understand as the decision features of LR and LR-F. All these features
clearly highlight meaningful image parts, such as the ankle and heel
of Ankle Boot, and the upper left corner of Bag. A closer look at the
the average images suggests that these decision features describe
the difference between Ankle Boot and Bag.
The decision features of PLNN capture more detailed difference
between Ankle Boot and Bag than the decision features of LR and
LR-F. This is because the LLCs of PLNN only capture the difference
between a subset of instances within a convex polytope, however,
LR and LR-F capture the overall difference between all instances
Exact and Consistent Interpretation for Piecewise Linear
Neural Networks: A Closed Form Solution
arXiv version, 2017, Canada
1
Table 5: The training and testing accuracy of all models.
0.5
Data Set
Accuracy
LR
LR-F
PLNN
LR-NS
LR-NSF
PLNN-NS
0
-0.5
-1
(a) Avg. Image
(b) LR
(c) LR-NS
(d) PLNN
(e) PLNN-NS
1
0.5
0
FMNIST-1
Train
Test
0.998
0.997
0.998
0.997
1.000
0.999
0.772
0.776
0.989
0.989
1.000
0.999
FMNIST-2
Train
Test
0.847
0.839
0.847
0.839
0.907
0.868
0.711
0.698
0.782
0.791
0.894
0.867
-0.5
-1
(f) Avg. Image
(g) LR-F
(h) LR-NSF
(i) PLNN
(j) PLNN-NS
Figure 4: The decision features of all models on FMNIST-1.
(a)-(e) and (f)-(j) show the average image and the decision
features of all models for Ankle Boot and Bag, respectively.
For PLNN and PLNN-NS, we show the decision features of
the LLC whose convex polytope contains the most instances.
1
0.5
0
-0.5
easier to understand than the decision features of PLNN. In particular, as shown in Figure 5, the decision features of PLNN-NS clearly
highlight the collar and breast of Coat, and the shoulder of Pullover,
which are much easier to understand than the cluttered features of
PLNN. These results demonstrate the effectiveness of non-negative
and sparse constraints in selecting meaningful features. Moreover,
the decision features of PLNN-NS capture more details than LR-NS
and LR-NSF, thus PLNN-NS achieves a comparable accuracy with
PLNN, and significantly outperforms the accuracy of LR-NS and
LR-NSF on FMNIST-2.
In summary, the decision features of LLCs are easy to understand,
and the non-negative and sparse constraints are highly effective in
improving the interpretability of the decision features of LLCs.
-1
(a) Avg. Image
(b) LR
(c) LR-NS
(d) PLNN
5.4
(e) PLNN-NS
1
0.5
0
-0.5
-1
(f) Avg. Image
(g) LR-F
(h) LR-NSF
(i) PLNN
(j) PLNN-NS
Figure 5: The decision features of all models on FMNIST-2.
(a)-(e) and (f)-(j) show the average image and the decision
features of all models for Coat and Pullover, respectively. For
PLNN and PLNN-NS, we show the decision features of the
LLC whose convex polytope contains the most instances.
of Ankle Boot and Bag. The accuracies of PLNN, LR and LR-F are
comparable because the instances of Ankle Boot and Bag are easy to
distinguish. However, as to be shown in Figure 5, when the instances
are hard to distinguish, PLNN captures much more detailed features
than LR and LR-F, and achieves a significantly better accuracy.
Figure 5 shows the decision features of all models on FMNIST-2.
As it is shown, LR and LR-F capture decision features with a strong
semantical meaning, such as the collar and breast of Coat, and the
shoulder of Pullover. However, these features are too general to
accurately distinguish between Coat and Pullover. Therefore, LR
and LR-F do not achieve a high accuracy. Interestingly, the decision
features of PLNN capture much more details than LR and LR-F,
which leads to the superior accuracy of PLNN.
The superior accuracy of PLNN comes at the cost of cluttered
decision features that may be hard to understand. Fortunately, applying non-negative and sparse constraints on PLNN effectively
improves the interpretability of the decision features without affecting the classification accuracy.
As shown in Figures 4 and 5, the decision features of PLNN-NS
highlight similar image parts as LR-NS and LR-NSF, and are much
Are PBFs of LLCs Easy to Understand?
The polytope boundary features (PBFs) of polytope boundaries (PBs) interpret why an instance is contained in the convex
polytope of a LLC. In this subsection, we systematically study the
semantical meaning of PBFs. Limited by space, we only use the
PLNN-NS models trained on FMNIST-1 and FMINST-2 as the target model to interpret. The LLCs of PLNN-NS are computed by
OpenBox.
(l )
(l )
Recall that a PB is defined by a linear inequality zi ∈ ψ (ci ),
(l )
where the PBFs are the coefficients of x in zi . Since the activation
(l )
zi
(l )
∈ ψ (ci )
(l )
(l )
(l )
zi
(l )
function is ReLU,
is either
> 0 or zi ≤ 0. Since
the values of PBFs are non-negative for PLNN-NS, for a convex
(l )
polytope Ph , if zi > 0, then the images in Ph strongly correlate
with the PBFs of zi ; if zi ≤ 0, then the images in Ph are not
(l )
strongly correlated with the PBFs of zi .
The above analysis of PBs and PBFs is demonstrated by the results in Tables 6 and 7, and Figure 6. Take the first convex polytope
(2)
(3)
in Table 6 as an example, the PBs are z11 > 0 and z2 > 0, whose
PBFs in Figures 6(b)-(c) show the features of Ankle Boot and Bag, respectively. Therefore, the convex polytope contains images of both
Ankle Boot and Bag. A careful study of the other results suggests
that the PBFs of the convex polytopes are easy to understand and
accurately describe the images in each convex polytope.
We can also see that the PBFs in Figure 6 look similar to the decision features of PLNN-NS in Figures 4 and 5. This shows the strong
correlation between the features learned by different neurons of
PLNN-NS, which is probably caused by the hierarchy network structure. Due to the strong correlation between neurons, the number
of configurations in C is much less than k N , as shown in Table 3.
Surprisingly, as shown in Table 7, the top-1 convex polytope on
FMNIST-2 contains more than 98% of the training instances. On
these instances, the training accuracy of LLC is much higher than
the training accuracies of LR-NS and LR-NSF. This means that the
arXiv version, 2017, Canada
Lingyang Chu, Xia Hu, Juhua Hu, Lanjun Wang, and Jian Pei
Table 6: The PBs of the top-3 convex polytopes (CP) containing the most instances in FMNIST-1. “/” indicates a redundant linear inequality. Accuracy is the training accuracy of
LLC on each CP.
1
0.5
0
-0.5
-1
(2)
(2)
(a) z6
(3)
(b) z11
(3)
(c) z2
CP
1
2
3
(d) z4
1
0.5
(2)
z6
/
≤0
/
(2)
z11
>0
>0
≤0
(3)
z2
>0
/
/
(3)
z4
/
≤0
>0
#Ankle Boot
3,991
9
0
#Bag
3,997
0
3
Accuracy
0.999
1.000
1.000
0
-0.5
Table 7: The PBs of the top-3 convex polytopes (CP) containing the most instances in FMNIST-2. Accuracy is the training
accuracy of LLC on each CP.
-1
(e)
(2)
z4
(f)
(2)
z5
(g)
(2)
z8
(h)
(3)
z2
CP
1
2
3
Figure 6: (a)-(d) show the PBFs of the PLNN-NS on FMNIST-1.
(e)-(h) show the PBFs of the PLNN-NS on FMNIST-2.
600
450
300
150
0
LIME
Avg. CPP
OpenBox
0
50
100
150
200
1
0.75
0.5
0.25
0
5
10
15
# Hacked Features
(a) FMNIST-1
(b) FMNIST-2
LIME
0
0
OpenBox
50
100
150
200
600
450
300
150
0
LIME
0
5
20
OpenBox
10
15
# Hacked Features
# Hacked Features
(c) FMNIST-1
(d) FMNIST-2
20
Figure 7: The hacking performance of LIME and OpenBox. (a)(b) show the average (Avg.) CPP. (c)-(d) show the NLCI.
training instances in the top-1 convex polytope are much easier to
be linearly separated than all training instances in FMNIST-2. From
this perspective, the behavior of PLNN-NS is like a “divide and
conquer” strategy, which set aside a small proportion of instances
that hinder the classification accuracy such that the majority of the
instances can be better separated by a LLC. As shown by the top-2
and top-3 convex polytopes in Table 7, the set aside instances are
grouped in their own convex polytopes, where the corresponding
LLCs also achieve a very high accuracy. Table 6 shows similar phenomenon on FMNIST-1. However, since the instances in FMNIST-1
are easy to be linearly separated, the training accuracy of PLNN-NS
marginally outperforms LR-NS and LR-NSF.
5.5
(2)
z5
>0
≤0
≤0
(2)
z8
>0
>0
≤0
(3)
z2
>0
>0
>0
#Coat
3,932
32
18
#Pullover
3,942
10
0
Accuracy
0.894
0.905
0.944
OpenBox
# Hacked Features
NLCI
NLCI
Avg. CPP
LIME
1
0.75
0.5
0.25
0
(2)
z4
>0
>0
>0
Can We Hack a Model Using OpenBox?
Knowing what an intelligent machine “thinks” provides us the
privilege to “hack” it. Here, to hack a target model is to significantly change its prediction on an instance x ∈ X by modifying
as few features of x as possible. In general, the biggest change of
prediction is achieved by modifying the most important decision
features. A more precise interpretation on the target model reveals
the important decision features more accurately, thus requires to
modify less features to achieve a bigger change of prediction. Following this idea, we apply LIME and OpenBox to hack PLNN-NS,
and compare the quality of their interpretations by comparing the
change of PLNN-NS’s prediction when modifying the same number
of decision features.
For an instance x ∈ X, denote by γ ∈ Rd the decision features
for the classification of x. We hack PLNN-NS by setting the values
of a few top-weighted decision features in x to zero, such that the
prediction of PLNN-NS on x changes significantly. The change
of prediction is evaluated by two measures as follows. First, the
change of prediction probability (CPP) is the absolute change
of the probability of classifying x as a positive instance. Second,
the number of label-changed instance (NLCI) is the number of
instances whose predicted label changes after being hacked. Again,
due to the inefficiency of LIME, we use the sampled data sets in
Section 5.2 for evaluation.
In Figure 7, the average CPP and NLCI of OpenBox are always
higher than LIME on both data sets. This demonstrates that the
interpretations computed by OpenBox are more effective than LIME
when they are applied to hack the target model.
Interestingly, the advantage of OpenBox is more significant on
FMNIST-1 than on FMNIST-2. This is because, as shown in Figure 2(a), the prediction probabilities of most instances in FMNIST-1
are either 1.0 or 0.0, which provides little gradient information for
LIME to accurately approximate the classification function of the
PLNN-NS. In this case, the decision features computed by LIME
cannot describe the exact behavior of the target model.
In summary, since OpenBox produces the exact and consistent
interpretations for a target model, it achieves an advanced hacking
performance over LIME.
5.6
Can We Debug a Model Using OpenBox?
Intelligent machines are not perfect and predictions fail occasionally.
When such failure occurs, we can apply OpenBox to interpret why
an instance is mis-classified.
Figure 8 shows some images that are mis-classified by PLNN-NS
with a high probability. In Figures 8(a)-(c), the original image is a
Coat, however, since the scattered mosaic pattern on the cloth hits
more features of Pullover than Coat, the original image is classified
as a Pullover with a high probability. In Figures 8(d)-(f), the original
image is a Pullover, however, it is mis-classified as a Coat because
the white collar and breast hit the typical features of Coat, and
the dark shoulder and sleeves miss the most significant features of
Pullover. Similarly, the Ankle Boot in Figure 8(g) highlights more
features on the upper left corner, thus it is mis-classified as a Bag.
Exact and Consistent Interpretation for Piecewise Linear
Neural Networks: A Closed Form Solution
(a) CO
(b) CO: 0.04 (c) PU: 0.96
(d) PU
(e) CO: 1.00 (f) PU: 0.00
(g) AB
(h) AB: 0.16 (i) BG: 0.84
(j) BG
(k) AB: 1.00 (l) BG: 0.00
Figure 8: The mis-classified images of (a) Coat (CO), (d)
Pullover (PU), (g) Ankle Boot (AB), and (j) Bag (BG). (a), (d), (g)
and (j) show the original images. For the rest subfigures, the
caption shows the prediction probability of the corresponding class; the image shows the decision features supporting
the prediction of the corresponding class.
The Bag in Figure 8(j) is mis-classified as an Ankle Boot because it
hits the features of ankle and heel of Ankle Boot, however, misses
the typical features of Bag on the upper left corner.
In conclusion, as demonstrated by the mis-classified examples
in Figure 8, OpenBox accurately interprets the mis-classifications,
which is potentially useful in debugging abnormal behaviors of the
interpreted model.
6
CONCLUSIONS AND FUTURE WORK
In this paper, we tackle the challenging problem of interpreting
PLNNs. By studying the states of hidden neurons and the configuration of a PLNN, we prove that a PLNN is mathematically equivalent
to a set of LLCs, which can be efficiently computed by the proposed
OpenBox method. Extensive experiments show that the decision
features and the polytope boundary features of LLCs provide exact
and consistent interpretations on the overall behavior of a PLNN.
Such interpretations are highly effective in hacking and debugging
PLNN models. As future work, we will extend our work to interpret more general neural networks that adopt smooth activation
functions, such as sigmoid and tanh.
REFERENCES
[1] Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. 2016. Analyzing the behavior
of visual question answering models. arXiv:1606.07356 (2016).
[2] Jimmy Ba and Rich Caruana. 2014. Do deep nets really need to be deep?. In NIPS.
2654–2662.
[3] Osbert Bastani, Carolyn Kim, and Hamsa Bastani. 2017. Interpreting Blackbox
Models via Model Extraction. arXiv:1705.08504 (2017).
[4] C Bishop. 2007. Pattern Recognition and Machine Learning (Information Science
and Statistics). Springer, New York (2007).
[5] C. Cao, X. Liu, Y Yang, Y. Yu, J. Wang, Z. Wang, Y. Huang, L. Wang, C. Huang,
et al. 2015. Look and think twice: Capturing top-down visual attention with
feedback convolutional neural networks. In ICCV. 2956–2964.
[6] RJ Caron, JF McDonald, and CM Ponic. 1989. A degenerate extreme point strategy
for the classification of linear constraints as redundant or necessary. JOTA 62, 2
(1989), 225–237.
[7] Z. Che, S. Purushotham, R. Khemani, and Y. Liu. 2015. Distilling knowledge from
deep networks with applications to healthcare domain. arXiv:1512.03542 (2015).
[8] Jan Chorowski and Jacek M Zurada. 2015. Learning understandable neural
networks with nonnegative weight constraints. TNNLS 26, 1 (2015), 62–69.
[9] Alexey Dosovitskiy and Thomas Brox. 2016. Inverting visual representations
with convolutional networks. In CVPR. 4829–4837.
[10] D. Erhan, Yoshua Bengio, A. Courville, and P. Vincent. 2009. Visualizing higherlayer features of a deep network. University of Montreal 1341 (2009), 3.
[11] Ruth Fong and Andrea Vedaldi. 2017. Interpretable Explanations of Black Boxes
by Meaningful Perturbation. arXiv:1704.03296 (2017).
arXiv version, 2017, Canada
[12] Nicholas Frosst and Geoffrey Hinton. 2017. Distilling a Neural Network Into a
Soft Decision Tree. arXiv:1711.09784 (2017).
[13] Amirata Ghorbani, Abubakar Abid, and James Zou. 2017. Interpretation of Neural
Networks is Fragile. arXiv:1710.10547 (2017).
[14] Xavier Glorot, Antoine Bordes, and Yoshua Bengio. 2011. Deep sparse rectifier
neural networks. In ICAIS. 315–323.
[15] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. 2016. Deep Learning. MIT
Press. http://www.deeplearningbook.org.
[16] Ian J Goodfellow, David Warde-Farley, Mehdi Mirza, Aaron Courville, and Yoshua
Bengio. 2013. Maxout networks. arXiv:1302.4389 (2013).
[17] B. Goodman and S. Flaxman. 2016. European Union regulations on algorithmic
decision-making and a" right to explanation". arXiv:1606.08813 (2016).
[18] Nick Harvey, Chris Liaw, and Abbas Mehrabian. 2017. Nearly-tight VC-dimension
bounds for piecewise linear neural networks. arXiv:1703.02930 (2017).
[19] K. He, X. Zhang, S. Ren, and J. Sun. 2015. Delving deep into rectifiers: Surpassing
human-level performance on imagenet classification. In ICCV. 1026–1034.
[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in
a neural network. arXiv:1503.02531 (2015).
[21] Patrik O Hoyer. 2002. Non-negative sparse coding. In WNNSP. 557–565.
[22] Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Maximilian Alber, Kristof T
Schütt, Sven Dähne, Dumitru Erhan, and Been Kim. 2017. The (Un) reliability of
saliency methods. arXiv:1711.00867 (2017).
[23] Pang Wei Koh and Percy Liang. 2017. Understanding black-box predictions via
influence functions. arXiv:1703.04730 (2017).
[24] Pascal Koiran and Eduardo D Sontag. 1996. Neural networks with quadratic VC
dimension. In NIPS. 197–203.
[25] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. 2012. Imagenet classification with deep convolutional neural networks. In NIPS. 1097–1105.
[26] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. 2015. Deep learning. nature
521, 7553 (2015), 436.
[27] Honglak Lee, Alexis Battle, Rajat Raina, and Andrew Y Ng. 2007. Efficient sparse
coding algorithms. In NIPS. 801–808.
[28] Jiwei Li, Xinlei Chen, Eduard Hovy, and Dan Jurafsky. 2015. Visualizing and
understanding neural models in NLP. arXiv:1506.01066 (2015).
[29] Aravindh Mahendran and Andrea Vedaldi. 2015. Understanding deep image
representations by inverting them. In CVPR. 5188–5196.
[30] Guido F Montufar, Razvan Pascanu, Kyunghyun Cho, and Yoshua Bengio. 2014.
On the number of linear regions of deep neural networks. In NIPS. 2924–2932.
[31] Vinod Nair and Geoffrey E Hinton. 2010. Rectified linear units improve restricted
boltzmann machines. In ICML. 807–814.
[32] Jiquan Ngiam, Aditya Khosla, Mingyu Kim, Juhan Nam, Honglak Lee, and Andrew Y Ng. 2011. Multimodal deep learning. In ICML. 689–696.
[33] R. B. Palm. 2012. Prediction as a candidate for learning deep hierarchical models
of data. (2012).
[34] Razvan Pascanu, Guido Montufar, and Yoshua Bengio. 2013. On the number of response regions of deep feed forward networks with piece-wise linear activations.
arXiv:1312.6098 (2013).
[35] Nadeem N Rather, Chintan O Patel, and Sharib A Khan. 2017. Using Deep
Learning Towards Biomedical Knowledge Discovery. IJMSC 3, 2 (2017), 1.
[36] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. Why should i
trust you?: Explaining the predictions of any classifier. In KDD. ACM, 1135–1144.
[37] R. R Selvaraju, A. Das, R. Vedantam, M. Cogswell, D. Parikh, and D. Batra. 2016.
Grad-cam: Why did you say that? visual explanations from deep networks via
gradient-based localization. arXiv:1610.02391 (2016).
[38] A. Shrikumar, P. Greenside, and A. Kundaje. 2017. Learning important features
through propagating activation differences. arXiv:1704.02685 (2017).
[39] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. 2013. Deep inside
convolutional networks: Visualising image classification models and saliency
maps. arXiv:1312.6034 (2013).
[40] D. Smilkov, N. Thorat, B. Kim, F. Viégas, and M. Wattenberg. 2017. SmoothGrad:
removing noise by adding noise. arXiv:1706.03825 (2017).
[41] Eduardo D Sontag. 1998. VC dimension of neural networks. NATO ASI Series F
Computer and Systems Sciences 168 (1998), 69–96.
[42] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic Attribution
for Deep Networks. arXiv:1703.01365 (2017).
[43] M. Wu, M. C Hughes, S. Parbhoo, M. Zazzi, V. Roth, and F. Doshi-Velez. 2018.
Beyond Sparsity: Tree Regularization of Deep Models for Interpretability. AAAI
(2018).
[44] Han Xiao, Kashif Rasul, and Roland Vollgraf. 2017. Fashion-MNIST: a Novel
Image Dataset for Benchmarking Machine Learning Algorithms. (2017).
arXiv:cs.LG/cs.LG/1708.07747
[45] J. Yosinski, J. Clune, A. Nguyen, T. Fuchs, and H. Lipson. 2015. Understanding
neural networks through deep visualization. arXiv:1506.06579 (2015).
[46] Rich Zemel, Yu Wu, Kevin Swersky, Toni Pitassi, and Cynthia Dwork. 2013.
Learning fair representations. In ICML. 325–333.
[47] Bolei Zhou, David Bau, Aude Oliva, and Antonio Torralba. 2017. Interpreting
Deep Visual Representations via Network Dissection. arXiv:1711.05611 (2017).
[48] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba.
2016. Learning deep features for discriminative localization. In CVPR. 2921–2929.
| 1cs.CV
|
Higher-order compatible discretization on
hexahedrals
arXiv:1304.7018v1 [math-ph] 25 Apr 2013
Jasper Kreeft and Marc Gerritsma
Abstract We derive a compatible discretization method that relies heavily on the
underlying geometric structure, and obeys the topological sequences and commuting properties that are constructed. As a sample problem we consider the vorticityvelocity-pressure formulation of the Stokes problem. We motivate the choice for
a mixed variational formulation based on both geometric as well as physical arguments. Numerical tests confirm the theoretical results that we obtain a pointwise
divergence-free solution for the Stokes problem and that the method obtains optimal
convergence rates.
1 Introduction
As sample problem we consider the Stokes flow problem in its vorticity-velocitypressure formulation,
ω − curl u = 0
in Ω ,
(1a)
curl ω + grad p = f
in Ω ,
(1b)
div u = 0
in Ω .
(1c)
In this article we consider prescribed velocity boundary conditions, u = 0 on ∂ Ω ,
but the method holds for all admissible types of boundary conditions, see [8].
Despite the simple appearance of Stokes flow model, there exists a large number
of numerical methods to simulate Stokes flow. They all reduce to two classes, that is,
Jasper Kreeft
Shell Global Solutions, Grasweg 31, 1031 HW, Amsterdam, e-mail: jasper.kreeft@shell.
com
Marc Gerritsma
Delft University of Technology, Kluyverweg 2, 2629 HT, Delft, e-mail: m.i.gerritsma@
tudelft.nl
1
2
Jasper Kreeft and Marc Gerritsma
either circumventing the LBB stability condition, like stabilized methods, e.g. [7],
or satisfying this condition, as in compatible or mixed methods, e.g. [4]. The last requires the construction of dedicated discrete vector spaces. Best known are the curl
conforming Nédélec and divergence conforming Raviart-Thomas spaces. Here, we
consider a subclass of compatible methods, i.e. mimetic methods. Mimetic methods
do not solely search for appropriate vector spaces, but aim to mimic structures and
symmetries of the continuous problem, see [2, 3, 10, 11]. As a consequence of this
mimicking, mimetic methods automatically preserve most of the physical and mathematical structures of the continuous formulation, among others the LBB condition
and, most important, a pointwise divergence-free solution, [8, 9].
At the heart of the mimetic method there are the well-known integral theorems of
Newton-Leibniz, Stokes and Gauss, which couple the operators grad, curl and div,
to the action of the boundary operator on a manifold. Therefore, obeying geometry
and orientation will result in satisfying exactly the mentioned theorems, and consequently performing the vector operators exactly in a finite dimensional setting. In 3D
we distinguish between four types of sub-manifolds, that is, points, lines, surfaces
and volumes, and two types of orientation, namely, outer- and inner-orientation.
Examples of sub-manifolds are shown in Figure 1 together with the action of the
boundary operator.
Outer Orientation
∂
∂
∂
∂
∂
∂
Inner Orientation
Fig. 1 The four geometric objects possible in R3 , point, line, surface and volume, with outer(above) and inner- (below) orientation. The boundary operator, ∂ , maps k-dimensional objects to
(k − 1)-dimensional objects.
By creating a quadrilateral or hexahedral mesh, we divide the physical domain
in a large number of these geometric objects, and to each geometric object we associate a discrete unknown. This implies that these discrete unknowns are integral
quantities. Since the three earlier mentioned theorems are integral equations, it follows for example that taking a divergence in a volume is equivalent to taking the
sum of the integral quantities associated to the surrounding surface elements, i.e.
the fluxes. So using integral quantities as degrees of freedom to perform a grad,
Higher-order compatible discretization on hexahedrals
3
curl or div, is equivalent to taking the sum of the degrees of freedom located at its
boundary.
These relations are of purely topological nature. They form a topological sequence or complex. This sequence is fundamental. It has a direct connection with
the complexes that are related to the physical domain, the computational domain,
the physical problem and the discretization.
Although the original work, [9, 10], was presented in terms of differential geometry and algebraic topology, here we will use vector calculus because it is the more
common mathematical language. Nevertheless, we will put emphasis on the distinction between topology and metric, on complexes and on commuting diagrams,
which drives the former two languages.
We make use of spectral element interpolation functions as basis functions. In
the past nodal spectral elements were mostly used in combination with Galerkin
projection (GSEM). The GSEM satisfies the LBB condition by lowering the polynomial degree of the pressure by two with respect to the velocity. This results in a
method that is only weakly divergence-free, meaning that the divergence of the velocity field only convergence to zero with mesh refinement. The present study uses
mimetic spectral element interpolation or basis functions, [10]. The mixed mimetic
spectral element method (MMSEM) satisfies the LBB condition and gives a pointwise divergence-free solution for all mesh sizes.
2 Can we really discretize exactly?
Since the Stokes flow model (1) should hold on a certain physical domain, we will
include geometry by means of integration. In that case we can relate every physical
quantity to a geometric object. Starting with the incompressibility constraint (1c)
we have due to Gauss’ divergence theorem,
Z
Z
div u dV =
V
u · n dS = 0,
∂V
and using Stokes’ circulation theorem the relation (1a) can be written as
Z
S
ω × n dS =
Z
S
curl u × n dS =
Z
u · t dl.
∂S
From the first relation it follows that div u is associated to volumes. The association
to a geometric object for velocity u is less clear. In fact it can be associated to
two different types of geometric objects. A representation of velocity compatible
with the incompressibility constraint is given in terms of the velocity flux, u · n,
through a surface that bounds the volume, while in the circulation relation velocity,
u · t, is represented along a line that bounds the surface. We will call the velocity
vector through a surface outer-oriented and the velocity along a line segment inneroriented. A similar distinction can be made for vorticity, see [9].
4
Jasper Kreeft and Marc Gerritsma
The last equation to be considered is (1b). This equation shows that classical
Newton-Leibniz, Stokes circulation and Gauss divergence theorems tell only half
the story. From the perspective of the classical Newton-Leibniz theorem, the gradient acting on the pressure relates line values to their corresponding end point, while
the Stokes circulation theorem shows that the curl acting on the vorticity vector relates surface values to the line segment enclosing it. So how does this fit into one
equation? In fact, from a geometric perspective, there exists two gradients, two curls
and two divergence operators. One of each is related to the mentioned integral theorems as explained above. The others are their formal adjoint operators. Let grad,
curl and div be the original differential operators associated to the mentioned integral theorems, then the formal Hilbert adjoint operators grad∗ , curl∗ and div∗ are
defined as,
a, −grad∗ b Ω := div a, b Ω , a, curl∗ b Ω := curl a, b Ω , a, −div∗ b Ω := grad a, b Ω .
From a geometric interpretation, the adjoint operators detours via the opposite type
of orientation. Where div relates a vector quantity associated to surfaces to a scalar
quantity associated to a volume enclosed by these surfaces. Its adjoint operator,
grad∗ , relates a scalar quantity associated with a volume to a vector quantity associated with its surrounding surfaces. This is illustrated in Figure 2. Following Figure 2,
the adjoint operator grad∗ consists of three consecutive steps: First, switch from an
outer oriented scalar associated to volumes to an inner oriented scalar associated to
points, then take the derivative and finally switch from an inner oriented vector associated to lines to an outer oriented vector associated to surfaces. In a similar way
we can describe the derivatives curl∗ and div∗ .
Outer Orientation
grad
∂
div *
curl
div
∂
∂
curl *
grad*
Inner Orientation
Fig. 2 Geometric interpretation of the action of the boundary operators, vector differential operators and their formal Hilbert adjoint operators.
Higher-order compatible discretization on hexahedrals
5
Since the horizontal relations are purely topological and the vertical relations
purely metric, the operators grad, curl and div are purely topological operators,
while grad∗ , curl∗ and div∗ are metric. This makes them much harder to discretize.
Now (1b) could then either be associated to an inner-oriented line segment by
rewriting it as
curl∗ ω + grad p = f,
or be associated to an outer-oriented surface by rewriting it as
curl ω + grad∗ p = f.
Without geometric considerations we could never make a distinction between grad,
curl and div and their associated Hilbert adjoints div∗ , curl∗ and grad∗ .
Since our focus is on obtaining a pointwise divergence-free discretization, we
decide to use the expression where the equations are associated to outer-oriented
geometric objects,
ω − curl∗ u = 0
in Ω ,
(2a)
curl ω + grad p = f
in Ω ,
(2b)
div u = 0
in Ω ,
(2c)
∗
where the first equation is associated to outer-oriented line segments, the second to
outer-oriented surfaces and the third to outer-oriented volumes.
3 Complexes
Figure 2 reveals already a number of sequence or complex structures. Starting from
geometry, we consider points, P, lines, L, surfaces, S, and volumes, V . They possess
a sequence in combination with the boundary operator, ∂ . The boundary of a volume
is a surface, the boundary of a surface is a line and the boundary of a line are its two
end points. This results in the following complex,
∂
∂
∂
∂
0 ←− P ←− L ←− S ←− V.
(3)
An important property of the complex is that if we apply the boundary operator
twice, we always find an empty set, e.g. if S = ∂V , then ∂ S = 0.
/ As follows directly
from the previously mentioned integral theorems, it follows, as a consequence of
∂ ∂ = 0,
/ that curl grad = 0 and div curl = 0. The derivatives themselves also form a
complex. In a Hilbert setting this becomes,
grad
curl
div
H 1 (Ω ) −→ H(Ω , curl) −→ H(Ω , div) −→ L2 (Ω ),
(4)
and using the Hilbert adjoint relations we also obtain the adjoint complex with properties curl∗ grad∗ = 0 and div∗ curl∗ = 0,
6
Jasper Kreeft and Marc Gerritsma
−div∗
−grad∗
curl∗
L02 (Ω ) ←− H0 (Ω , div∗ ) ←− H0 (Ω , curl∗ ) ←− H01 (Ω )
(5)
In the Hilbert setting, the variables of the Stokes problem are in the following spaces, ω ∈ H(Ω , curl) ∩ H0 (Ω , div∗ ), u ∈ H(Ω , div) ∩ H0 (Ω , curl∗ ) and p ∈
L2 (Ω )∩H0 (Ω , grad∗ ). It is hard, if even possible at all, to find discrete vector spaces
that are subsets of these function spaces and simultaneously satisfy the complex
properties. Instead, the Stokes problem can be cast into an equivalent variational
or mixed formulation where we make use of the Hilbert adjoint properties. This
simplifies the function spaces of the flow variables. The mixed formulation reads;
Find (ω, u, p) ∈ {H(Ω , curl)×H(Ω , div)×L2 (Ω )} with f ∈ H(Ω , div) given,
for all (σ , v, q) ∈ {H(Ω , curl) × H(Ω , div) × L2 (Ω )}, such that,
σ , ω Ω − curl σ , u Ω = 0,
(6a)
v, curl ω Ω − div v, p Ω = v, f Ω ,
(6b)
q, div u Ω = 0.
(6c)
With the formulation and corresponding function spaces, we are able to construct
compatible discrete vector spaces. Note that we now completely avoid the metric
dependent derivatives grad∗ and curl∗ , and their corresponding complex.
4 Discretization of Stokes problem
Degrees of freedom. In many numerical methods, especially in finite difference and
finite element methods, the discrete coefficients are point values. In the proposed
mimetic structure, the discrete unknowns represent integral values on k-dimensional
submanifolds, ranging from points to volumes, so 0 ≤ k ≤ 3. These k-dimensional
submanifolds are oriented, constitute the computational domain and span the physical domain. The concept of orientation shown in Figures 1 and 2 gave rise to the
boundary operator, ∂ , which can be represented by connectivities consisting only of
-1, 0 and 1, see also [9].
The space of degrees of freedom are given by P, L , S and V . These spaces
form a duality pairing with the geometric spaces P, L, S and V . The degrees of
freedom are integral values, i.e.
Z
l
w · t dl ∈ L ,
Z
S
u · n dS ∈ S ,
Z
V
p dV ∈ V .
(7)
By the definition of the degrees of freedom spaces and the previously mentioned
integral theorems, we can define the formal adjoint of the boundary operator, i.e.
the coboundary operator, δ . The coboundary operator is the discrete representation
of the topological derivatives grad, curl and div. Since ∂ ∂ = 0,
/ it follows from a
discrete Newton-Leibniz, Stokes and Gauss theorem that applying the coboundary
Higher-order compatible discretization on hexahedrals
7
operator twice is always zero, δ δ = 0/ (see [2, 9]). The coboundary operator also
has matrix representations, G, C and D, that are the transpose of the connectivity
matrices. We obtain the following topological sequence,
G
C
D
P −→ L −→ S −→ V ,
(8)
where CG = 0 and DC = 0. These matrices will explicitly appear in the final matrix
system. An illustration of DC = 0 is given in Figure 3. More details on the structure
ω
DC ω
Cω
Fig. 3 The action of twice the coboundary operator δ on a vorticity d.o.f. has a zero net result on
its surrounding volumes, because they all have both a positive and a negative contribution from its
neighboring velocity faces.
of geometry, orientation and degrees of freedom can be found in Gerritsma et al,
[6].
Mimetic Operators. Let W = H(Ω , curl), V = H(Ω , div) and Q = L2 (Ω ). The
discretization of the flow variables involves a projection operator, πh , from the complete vector spaces W , V and Q, to the discrete vector spaces Wh , Vh and Qh . Here the
flow variables are expressed in terms of d.o.f. defined on k-cells, and corresponding
interpolation functions (also called basis-functions). The projection operator actually consists of two steps, a reduction operator, R, that integrates the flow variables
on k-cells, and a reconstruction operator, I , that interpolates the d.o.f. using the
appropriate basis-functions. These mimetic operators were defined in [2, 10]. A
composition of the two operators gives the projection operator πh = I ◦ R 1 .
Reduction operator R is simply defined by integration. It possesses the following
commutation relations,
Rgrad = GR,
Rcurl = CR,
Rdiv = DR.
(9)
The treatment of the reconstruction operator leaves some freedom, as long as it
satisfies the following properties: be the right inverse of the reduction, RI = Id,
be the approximate left inverse of the reduction, I R = Id + O(h p ), and it should
possess the following commutation relations,
grad I = I G,
1
curl I = I C,
div I = I D.
(10)
For completeness, in a Hilbert setting the projection needs an additional smoothing argument.
This step is ignored here to increase readibility. See [8] for more details.
8
Jasper Kreeft and Marc Gerritsma
When both the reduction and reconstruction operators commute with continuous and discrete differentiation, than also the projection operator πh possesses
a commutation relation with differentiation. In case of the divergence operator,
which is relevant to obtain a pointwise divergence-free solution, the commutation relation is given by,
div πh = div I R = I DR = I Rdiv = πh div .
(11)
The commutation relations in case of divergence are illustrated below,
div
V −−−−→
yR
D
Q
+
yR
S −−−−→ V .
D
S −−−−→
yI
V
yI
div
=
div
Vh −−−−→ Qh .
V −−−−→
π
y h
Q
π
y h
div
Vh −−−−→ Qh .
Since property (11) also holds for the grad and curl, we obtain the following
complex for discrete vector spaces,
grad
curl
div
Φh −→ Wh −→ Vh −→ Qh .
(12)
In practice we use I DR from (11) in computations. Relation (11) implies
among others that it satisfies the discrete LBB condition,
qh , div vh Ω
βh := inf sup
> β > 0,
(13)
qh ∈Qh vh ∈Vh kqh kQ kvh kV
where β is the inf-sup constant of the continuous problem, (1). Whereas the LBB
condition is a measurement for numerical stability, the commutation relation indicates physical correctness of the numerical method. This last is a much stronger
statement, which includes also the former.
The conditions on the reconstruction operator have led to the construction of
mimetic spectral element basis-functions, [5, 10]. Since we use a tensor-based construction of point, line, surface and volume corresponding basis-functions, we only
need nodal and edge interpolation functions. The nodal interpolation functions are
the well-known Lagrange polynomials. The edge polynomials were derived from
the Lagrange polynomials, based on the given conditions. For a set of Lagrange
polynomials, li (x), i = 0, . . . , N, the edge polynomials, ei (x), i = 1, . . . , N, are given
by,
i−1
dlk (x)
ei (x) = − ∑
.
(14)
k=0 dx
The Lagrange and edge polynomials possess the condition RI = Id, i.e.,
Higher-order compatible discretization on hexahedrals
9
Z xj
li (x j ) = δi, j ,
x j−1
ei (x) dx = δi, j ,
(15)
where δi, j is the Kronecker delta. The interpolation function
for a variable associated
to a surface, for example, is given by, si, j,k (x, y, z) = li (x)e j (y)ek (z), ei (x)l j (y)ek (z),
ei (x)e j (y)lk (z) .
Example 1 (Divergence operator in 2D). One of the most interesting properties of the mimetic method presented in this paper, is that within our weak
formulation, the divergence-free constraint is satisfied pointwise. Let uh ∈ Vh
be the velocity flux defined as
N
∑i=0 ∑Nj=1 ui, j li (x)e j (y)
.
(16)
uh =
∑Ni=1 ∑Nj=0 vi, j ei (x)l j (y)
Then the change of mass, mh ∈ Qh , is equal to the divergence of uh ,
N
N
mh = div uh = ∑ ∑ (ui, j − ui−1, j + vi, j − vi, j−1 )ei (x)e j (y).
i=1 j=1
N
N
= ∑ ∑ mi, j ei (x)e j (y),
(17)
i=1 j=1
where mi, j = ui, j − ui−1, j + vi, j − vi, j−1 can be compactly written as m = Du.
Note that if the mass production is zero, as in our model problem (1c), the
incompressibility constraint can already be satisfied at the discrete level. Interpolation using ei (x)e j (y) then results in a solution of velocity uh that is
pointwise divergence-free.
5 A priori error estimates
By standard interpolation theory it follows that we obtain the following h-convergence
rates for the interpolation errors of the flow variables,
kω − πh ωkH(curl) = O(hN ),
ku − πh ukH(div) = O(hN ),
kp − πh pkL2 = O(hN ),
(18)
and that kdiv u − div πh ukL2 = 0 due to the commuting property.
In cases with empty harmonic vector spaces, we have that the discrete vector
spaces are conforming, i.e., Wh ⊂ W , Vh ⊂ V and Qh ⊂ Q. Moreover, due to the commuting property, it follows that these spaces are compatible, i.e., curlWh ⊂ Vh and
divVh = Qh . Finally they possess a Helmholtz-Hodge decomposition, σ = grad φ +
10
Jasper Kreeft and Marc Gerritsma
⊥
curl∗ v and v = curl σ + grad∗ q. In terms of vector spaces, this is, Wh = ZWh ⊕ ZW
h
and Vh = ZVh ⊕ ZV⊥h , where Z refers to the kernel or nullspace and Z ⊥ to its orthogonal complement. Having all these properties, a priori error estimates are derived in
[8] that show optimal convergence rates for all admissible boundary conditions, including the no-slip boundary condition, which is non-trivial in mixed finite element
methods. The a priori error estimates are given by
kω − ω h kW ≤ C inf kω − σ h kW ,
(19)
ku − uh kV ≤ C inf ku − vh kV +C inf kω − σ h kW ,
(20)
σ h ∈Wh
vh ∈Vh
σ h ∈Wh
kp − ph kQ ≤ C inf kp − qh kQ +C inf ku − vh kV +C inf kω − σ h kW ,
qh ∈Qh
vh ∈Vh
σ h ∈Wh
(21)
where the constants C will differ in each case and are independent of h. It shows
that the rate of convergence of the approximation errors are the same as those of the
interpolation errors.
6 Numerical Results
For many years, the lid-driven cavity flow was considered one of the classical benchmark cases for the assessment of numerical methods and the verification of incompressible (Navier)-Stokes codes. The 3D lid-driven cavity test case deals with a flow
in a unit box with five solid boundaries and moving lid as the top boundary, moving with constant velocity equal to minus one in x-direction. Especially the two line
singularities make the lid-driven cavity problem a challenging test case.
The left plot in Figure 4 shows slices of the magnitude of the velocity field in a
three dimensional lid-driven cavity Stokes problem, obtained on a 2 × 2 × 2 element
mesh, where each element contains a Gauss-Lobatto mesh of N = 8. The slices are
taken at 10%, 50% and 90% of the y-axis. The right plot in Figure 4 shows slices of
divergence of the velocity field. Figure 4 confirms that the mixed mimetic spectral
element method leads to an accurate result with a divergence-free solution.
The second testcase shows the optimal convergence behavior for a 2D Stokes
problem with no-slip boundary conditions. The testcase originates from a recent
paper by Arnold et al [1], where sub-optimal convergence is shown and proven for
no-slip boundary conditions when using Raviart-Thomas elements. Since RaviartThomas elements are the most popular H(div, Ω) conforming elements, we compare
our method to these results.
Figure 5 shows the results of the Stokes problem on a unit square with velocity
and pressure fields given by u = − 2x2 (x − 1)2 y(2y − 1)(y − 1), 2y2 (y − 1)2 x(2x −
T
1)(x − 1) , p = (x − 21 )5 + (y − 21 )5 . While for velocity both methods show optimal
convergence, for pressure a difference of 12 is noticed in the rate of convergence and
for vorticity a difference in rate of convergence of 32 is revealed.
Higher-order compatible discretization on hexahedrals
Z
11
Z
Y
X
Y
X
div u
-14
5x10
3
1
-1
-3
-5
Fig. 4 Left: slices of magnitude of the velocity field of a three dimensional lid-driven cavity Stokes
problem obtained on a 2 × 2 × 2 element mesh with N = 8. Right: slices of the divergence of
velocity. Is confirms a divergence-free velocity field.
0
10
−2
||u-uh ||H(div) ||u-uh ||L 2
||ω-ωh||H(curl)
Raviart−Thomas, N=2
Mimetic Spectral, N=2
Mimetic Spectral, N=3
0
10
1
_
2
10
−4
10
3
−6
10
−3
−2
10
−4
10
2
10
−2
10
−1
10
h
2
−6
10
0
10
−3
10
−1
10
10
−2
h
3
−1
10
0
10
0
10
−2
||p-ph ||L2
10
3
_
2
−4
10
2
3
−6
10
−3
10
−2
10
h
10
0
Fig. 5 Comparison of the h-convergence between Raviart-Thomas and Mimetic spectral element
projections for the 2D Stokes problem with no-slip boundary conditions.
12
Jasper Kreeft and Marc Gerritsma
References
1. D. Arnold, R. Falk, and J. Gopalakrishnan. Mixed finite element approximation of the vector
Laplacian with Dirichlet boundary conditions. Mathematical Models & Methods in Applied
Sciences, 22(9):1250024, 2012.
2. P. Bochev and J. Hyman. Principles of mimetic discretizations of differential operators. In
D. Arnold, P. Bochev, R. Lehoucq, R. Nicolaides, and M. Shashkov, editors, Compatible Discretizations, volume 142 of IMA Volumes in Mathematics and its Applications, pages 89–119.
Springer, 2006.
3. F. Brezzi and A. Buffa. Innovative mimetic discretizations for electromagnetic problems.
Journal of computational and applied mathematics, 234:1980–1987, 2010.
4. F. Brezzi and M. Fortin. Mixed and hybrid finite element methods, volume 15 of Springer
Series in Computational Mathematics. Springer-Verlag, New York, 1991.
5. M. Gerritsma. Edge functions for spectral element methods. In J. Hesthaven and E. Rønquist,
editors, Spectral and High Order Methods for Partial Differential Equations, pages 199–208.
Springer, 2011.
6. M. Gerritsma, R. Hiemstra, J. Kreeft, A. Palha, P. Rebelo, and D. Toshniwal. The geometric
basis of mimetic spectral approximations. In (this issue), 2013.
7. T. Hughes, P. Franca, and M. Balestra. A new finite element formulation for computational
fluid dynamics: V. Circumventing the Babuška-Brezzi condition: A stable Petrov-Galerkin
formulation of the Stokes problem accomodating equal-order interpolations. Computer Methods in Applied Mechanics and Engineering, 59:85–99, 1986.
8. J. Kreeft and M. Gerritsma. A priori error estimates for compatible spectral discretization
of the Stokes problem for all admissible boundary conditions. submitted, arXiv:1206.2812,
2013.
9. J. Kreeft and M. Gerritsma. Mixed mimetic spectral element method for Stokes flow: a pointwise divergence-free solution. Journal Computational Physics, 240:284–309, 2013.
10. J. Kreeft, A. Palha, and M. Gerritsma. Mimetic framework on curvilinear quadrilaterals of
arbitrary order. submitted, arXiv:1111.4304, 2013.
11. J. B. Perot. Discrete conservation properties of unstructured mesh schemes. Annual review of
fluid mechanics, 43:299–318, 2011.
| 5cs.CE
|
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
1
A Novel Pre-processing Scheme to Improve the
Prediction of Sand Fraction from Seismic
Attributes using Neural Networks
Soumi Chaki, Student Member, IEEE, Aurobinda Routray, Member, IEEE, William K. Mohanty
Abstract— This paper presents a novel pre-processing scheme
to improve the prediction of sand fraction from multiple seismic
attributes such as seismic impedance, amplitude and frequency
using machine learning and information filtering. The available
well logs along with the 3-D seismic data have been used to
benchmark the proposed pre-processing stage using a
methodology which primarily consists of three steps: preprocessing, training and post-processing. An Artificial Neural
Network (ANN) with conjugate-gradient learning algorithm has
been used to model the sand fraction. The available sand fraction
data from the high resolution well logs has far more information
content than the low resolution seismic attributes. Therefore,
regularization schemes based on Fourier Transform (FT),
Wavelet Decomposition (WD) and Empirical Mode
Decomposition (EMD) have been proposed to shape the high
resolution sand fraction data for effective machine learning. The
input data sets have been segregated into training, testing and
validation sets. The test results are primarily used to check
different network structures and activation function
performances. Once the network passes the testing phase with an
acceptable performance in terms of the selected evaluators, the
validation phase follows. In the validation stage, the prediction
model is tested against unseen data. The network yielding
satisfactory performance in the validation stage is used to predict
lithological properties from seismic attributes throughout a given
volume. Finally, a post-processing scheme using 3-D spatial
filtering is implemented for smoothing the sand fraction in the
volume. Prediction of lithological properties using this framework
is helpful for Reservoir Characterization.
Index Terms— Artificial Neural Network (ANN), wavelets,
Empirical Mode Decomposition (EMD), entropy, Fourier
Transform, Normalized Mutual Information (NMI), preprocessing, regularization, Reservoir Characterization (RC), sand
fraction, 3-D median filtering.
I. INTRODUCTION
H
YDROCARBONS migrate from source rock through
porous medium to reach reservoir rock for temporary
preservation [1]. Finally, the mobile hydrocarbons get seized in
the cap rocks. As such, the identification of hydrocarbon–
enriched–formations by characterization of each layer in the
borehole is of enormous importance to the explorers.
Soumi Chaki and Aurobinda Routray are with the Department of Electrical
Engineering, Indian Institute of Technology Kharagpur, Kharagpur- 721302,
India (e-mail: [email protected], [email protected]).
Recognition of potential hydrocarbon–enriched zones in a
prospective oil exploration field can be carried out using well
logs which can categorize layers into different sections such as
dry, water containing, and hydrocarbon bearing layers. The
lithological properties in the neighborhood of a borehole can be
estimated from well logs, whereas their distributions become
difficult to predict away from the wells. In such cases, available
seismic attributes can be used as a guidance to predict
lithological information at all traces of the area of interest [2].
Well logs and seismic attributes are integrated at available well
locations to design a reservoir model with the least uncertainty.
However, mapping between well logs and seismic attributes is
governed by nonlinear relationship and characterized by
mismatch in information content. Such nonlinear problems can
be approached using state-of-art computer–based methods like
hybrid systems [3], [4], multiple regression, neural networks
[5], Neuro-fuzzy Systems [6] etc.
It has been found from literature that ANN has been widely
used by researchers and engineers from diverse backgrounds to
model single or multiple target properties from predictor
variables in different research problems because of its accurate
prediction and generalization capability. For example, ANN has
been used in climatological studies [7], ocean engineering [8],
telecommunications [9], text recognition [10], financial time
series [11], reservoir characterization [12]–[17], etc. A diverse
dataset containing information assembled from multiple
domains can be used for learning and validation of ANN.
However, it has some inherent limitations. Firstly, performance
of ANN is dependent on the selection of network structure and
associated parameters. Secondly, training a complex
multilayered network is a time intensive process. Furthermore,
a complex network trained with relatively smaller number of
learning patterns may lead to over fitting. It is equally important
to assess the possibility of modelling the target property from
predictor variables using ANN or any other nonlinear
modelling approach. Sometimes the model performance can be
improved by applying suitable filtering techniques to the
predictor/target variables in the pre-processing stage. Several
studies have contributed to the performance analysis of ANN
along with other machine learning algorithms to model a target
variable from single or multiple predictors with respect to RC
William K. Mohanty is with the Department of Geology and Geoscience,
Indian Institute of Technology Kharagpur, Kharagpur- 721302, India (e-mail:
[email protected]).
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
II. STUDY AREA AND WORKFLOW
A. Study Area
In this study, the working dataset has been acquired from a
western onshore hydrocarbon field in India. The hydrocarbon
field is located at an intra-cratonic basin, spread along the
western periphery of central India. It is surrounded by the
Aravalli range. Deccan craton and Saurashtra craton are located
in the east-west direction along the basin. Structurally, the field
is located as a broad nosing feature; thus, housing the
hydrocarbons between two major synclines. The hydrocarbon
is present within a series of vertically stacked sandstone
reservoirs individually separated by intervening shale. The
average thickness of a sand layer is in the order of 5-6 m. Due
to the discrete sand depositions in thin layers, imaging of the
seismic data and obtaining a mapping between seismic and
borehole dataset are challenging tasks.
A spatial database containing seismic attributes and well logs
has been acquired from the study area in .sgy and .las format
respectively. The SEG-Y file format is one of the numerous
standards developed by the Society of Exploration
Geophysicists (SEG) for storing geophysical data [18]. The Log
ASCII Standard (LAS) format has been developed by the
Canadian Well Logging Society to standardize the organization
of digital log curves information in 1989 [19]. As the workflow
is developed on MATLAB platform, the .sgy data files are
converted in .mat format (MATLAB platform compatible
format) for MATLAB compatibility.
Fig. 1 represents the placements of the four wells across the
study area. These wells will be hereafter referred as A, B, C,
and D in terms of inlines and crosslines (xlines). Additionally,
well A (XXX3645, XXX7387), well B (XXX2107, XXX7285),
well C (XXX2768, XXX7073), and well D (XXX0916,
XXX7807) are the locations in terms of Universal Transverse
Mercator (UTM) coordinates. The depth of the each well is
around 3000 meter from ground, whereas the zone of interest is
from 2750 meter to 2975 meter in subsurface for well A. The
zone of interests are 2720 meter – 2950 meter for well B, C, and
D.
150
140
Inline
problem; however, the following aspects still remain
unexplored:
Design of an appropriate pre-processing stage for
effectiveness of machine learning algorithms
Proper choice of structure and parameters associated with
selected machine learning algorithms (here, ANN model
parameters- e.g. activation function type, number of hidden
layers etc.)
Suitable post-processing methods for the predicted output
This paper proposes a novel pre-processing scheme and
demonstrates the use of the said scheme in an appropriately
designed framework, consisting of three stages- pre-processing,
modelling, and post-processing, to estimate sand fraction (SF)
from multiple seismic attributes. Here the target variable SF
represents per unit sand volume within the rock.
The workflow starts with pre-processing stage, which uses
three alternative approaches for target variable regularization
based on Fourier Transform (FT), Wavelet Decomposition
(WD), and Empirical Mode Decomposition (EMD). Next, a
functional relationship between the regularized target sand
fraction and the seismic attributes is calibrated using ANN. In
this study, a simple network structure consisting a single hidden
layer is selected over relatively complex networks. The network
parameters and training algorithm are decided empirically. The
results obtained, are evaluated in terms of four performance
indicators– Correlation Coefficient (CC), Root Mean Square
Error (RMSE), Absolute Error Mean (AEM) and Scatter Index
(SI) [7]. The inclusion of the pre-processing stage in the
workflow is found to produce remarkable improvement in the
prediction results as opposed to use of the original sand fraction
using the same ANN structure. After learning and validation,
post processing is carried out to improve the prediction result
over the study area. In this way, a complete workflow is devised
for modelling a lithological property (here, sand fraction) from
given seismic attributes.
The paper is organized as follows. Section II describes the
study area and the complete workflow adopted in this paper.
Section III explains the pre-processing steps. Section IV
illustrates the ANN model building and validation. In Section
V, the post-processing is described. Finally, Section VI
concludes the paper with a discussion on achievements in this
study and possible future directions of research.
2
D
130
B
120
C
A
110
100
250
300
350
Xline
400
450
Fig. 1. Location of four wells in the study area in terms of inlines and
crosslines (xlines)
The borehole dataset contains basic logs such as gamma ray,
resistivity, density along with derived geo-scientific logs such
as sand fraction, permeability, porosity, water saturation, etc.
These well logs are treated as one dimensional signals for
further processing in this study. On the other hand, the seismic
dataset contains seismic impedance, amplitude, and
instantaneous frequency throughout the volume.
B. Workflow
The systematic workflow consisting three stages e.g. preprocessing, model building–validation and post-processing is
demonstrated in Fig. 2. The workflow is implemented on a 64bit MATLAB platform installed in Dell Precision M6800 with
Intel® core™ i7-4700MQ CPU @2.40 GHz having 32GB
RAM. The regularization approach in the pre-processing stage
is highlighted using a small dotted rectangle in Fig. 2.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
3
Fig. 2. The proposed workflow to model sand fraction from multiple seismic attributes
Three alternative regularization methods based on FT, WD, and
EMD are implemented in this study. In the model building
stage, training, testing, and validation are carried out
consecutively. If the validation performance does not improve
as expected even after implementing either of the three
regularization techniques, then the parameters associated with
the selected regularization method are modified before carrying
out training, testing, and validation again. This tuning is carried
out using a trade-off between loss of high frequency
information from target attribute and acceptable performance
evaluators.
The improvement in the ANN performance with the
regularization technique is quantified using the performance
evaluators presented in Tables-III and IV later. First, the ANN
is trained with the original sand fraction as target variable
without involving any pre–processing. The same procedure is
repeated with the regularized sand fraction as target variable of
the ANN. This is helpful to reassure the improvement in ANN
performance while using the regularized target variables. The
validated network is used to obtain sand fraction over the study
area from the three seismic attributes. The post-processing stage
includes further smoothing and filtering of the estimated output
to remove unwanted predictions.
III. PRE-PROCESSING
Pre-processing plays a crucial role on the performance of a
machine learning algorithm. In this paper, an efficient preprocessing approach is proposed as part of the adopted
methodology to obtain a functional relationship between
seismic attributes and lithological properties.
A. Signal Reconstruction
The borehole data are recorded at specific well locations
along depth with a high vertical resolution. Conversely, the
seismic data are collected spatially in time domain with a
sampling interval of two milliseconds. Signal reconstruction is
an important pre-processing step and has been discussed in
existing literature. In [20], the up-scaling of an acoustic
impedance log has been carried out by a wavelet transform
based method as well as a geostatistical technique. The acoustic
impedance log is computed from the recorded logs of the Pwave velocity and the density at the corresponding welllocation. The reconstructed acoustic impedance logs obtained
from the raw impedance log by wavelet transform and
geostatistical methods have turned out to be similar. In [21], the
seismic attributes and well logs have been integrated to predict
facies using a data-driven methodology. The complexities
regarding resolution of individual datasets (seismic and well
logs) and difference in the individual sampling intervals have
been discussed in detail in [21]. For the present work, the well
logs are converted from depth to time domain at 0.15
milliseconds sampling interval using the given velocity profile
obtained from well-seismic-tie. It can be noted that the
sampling intervals of the two datasets (seismic attributes and
well logs) are different. Hence in order to integrate the two, the
band limited seismic attributes are reconstructed at each time
instant corresponding to the well logs by a sinc interpolator
while adhering to the Nyquist–Shannon sampling theorem [22].
Fig. 3 shows the three seismic attributes and the sand
fraction, along well A. The red dots on the seismic attributes
represent their original values at time interval of two
milliseconds and the green curves represent the corresponding
reconstructed signals at the time instants marked on the well
logs. The blue high frequency curve in Fig. 3(d) represents sand
fraction along the same well.
Time(ms)
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
2240
2240
2240
2240
2260
2260
2260
2260
2280
2280
2280
2280
2300
2300
2300
2300
2320
2320
2320
2320
2340
2340
2340
2340
2360
2360
2360
2360
2380
6
2380
0
5000
2380
10 -5000 0
Seismic 6
Seismic
x 10
Impedance
Amplitude
8
Reconstructed seismic attributes
(a)
(b)
100 200
Inst.
Frequency
2380
0
0.5
Sand
Fraction
1
Original seismic attributes
(c)
(d)
Fig. 3. Seismic attributes and sand fraction along well A (The red dots on
the seismic attributes are the original values, the green line is the
reconstructed signal)
B. Signal Regularization
It can again be observed from Fig. 3 that the frequencies
present in seismic signals are much lower compared to that of
the sand fraction. In other words, the sand fraction carries much
more information as compared to the seismic attributes.
According to laws of information theory, a higher informationcarrying signal cannot be modelled using single or multiple
lower information-carrying predictor signals [23]. Only a part
of the target variable that is dependent on the predictor variables
can be modelled. For the present case, the spatially distributed
seismic attributes have low vertical resolution; whereas the high
resolution borehole data are recorded at the specific well
locations and carry high information content [24]. Hence, the
estimation of subsurface lithological properties from seismic
attributes deals with uncertainty due to the mismatch in the
information content. In order to circumvent this disparity, the
seismic attributes are re-sampled at each time instants
corresponding to the well logs to integrate with the lithological
properties. However, this re-sampling does not contribute any
additional information [23]. The seismic information would not
be able to successfully delineate the rock properties completely,
which is a key step for RC. As no new information can be added
to the seismic attributes, some amount of information has to be
filtered from the high information carrying SF signal in order to
calibrate a functional relationship between them. Thus, the
necessity of information filtering for the SF prediction through
regularization is established.
In this paper, three different signal processing approaches are
selected and implemented in order to filter the target signal. The
parameters belonging to this stage are tuned following the
changes in entropy before and after filtering along with visual
inspection of the output signal with respect to that of the
original target signal. The entropy has been computed from the
Power Spectral Density (PSD) of the signal. The average
amount of information gained from a measurement that
specifies X is defined to be the entropy H ( X ) of a system. It
can be formally defined as
(1)
H ( X ) p( xi ) log 2 p( xi )
i
i th
where p ( x i ) is the probability of X having the
value x i
in the dataset. This is known as Shannon entropy [23], [25]. If
4
A is another random variable described on the same dataset
then the mutual information between the two can be expressed
as
(2)
I ( X ; A) H ( X ) H ( X | A)
where, H ( X | A) is the conditional entropy of X after A has
been observed. A reservoir property (SF for the present case)
can be represented by X and seismic attribute e.g. seismic
impedance can be represented by A . The statistical property of
I ( X ; A ) can be interpreted as the reduction in the uncertainty
of the reservoir property, due to observation of attribute A . In
[26], Normalized Mutual Information (NMI) is defined as the
mutual information normalized by minimum entropy of both
the variables.
(3)
NMI ( X ; A ) I ( X ; A ) / min( H ( X ), H ( A))
In this study, the NMI computed between predictor and target
signal has been used to adjust the parameters of the information
filtering algorithms.
Before starting the regularization, the seismic inputs and sand
fraction values are normalized by Z-score and min-max
methods respectively using the data from all four wells taken
together. In this paper, the modelling of the target SF from the
three seismic attributes is carried out by a single hidden layer
ANN trained using a back propagation algorithm. The network
structure and training procedure are discussed in detail in the
following section. The target variable is normalized within the
range of output activation function keeping some offset from
limiting value of the activation function. Otherwise, the back
propagation algorithm tends to drive the free network
parameters to infinity. As a result, the learning process may
slow down [27]. Hence, the target variables are normalized
between 0.1 and 0.9 so that it does not overlap with the
saturation region of the log-sigmoid function.
1)
Fourier Transform (FT)
The first regularization approach is based on Fourier
transform (Algorithm I). Here, the spectrums of target and
predictor variables are compared and higher frequency
components of the target signal are truncated. Then, the target
signal is reconstructed using Inverse Fourier transform (IFT).
ALGORITHM I: SF regularization Based on Fourier Transform
Task : Regularizing target sand fraction based on Fourier Transform
Input : Predictor signal x (t ) and target signal y(t )
a) The sand fraction y(t ) extracted from a well-log by proper depth to time
conversion using the given velocity profile. Let the seismic amplitude
along the same well-log be x (t ) .
N
b) Compute Fourier Transform of x (t ) : X (k )
( j 1)( k 1)
N
x( j)ω
j 1
where,
ω N e 2 πi / N
is an N th root of unity
N
Similarly, FT of target y(t ) is computed as: Y ( k )
( j 1)( k 1)
N
y ( j )ω
j 1
where,
ω N e 2 πi / N
is an N th root of unity
c) Compare the spectrums of target and predictor signals
d) Select the bandwidth parameter ξ max Hz
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
e) The part of the target spectrum exceeding ξ max Hz is truncated to zero.
Modified Target : Ymod (k )
f) Construct regularized target signal y r ( t ) by carrying out IFT of the
truncated
spectrum
:
yr (t )
1
N
N
Y
( j 1)( k 1)
mod ( k )ωN
,
where,
k 1
th
ω N e 2 πi / N is an N root of unity
g) Calculate entropies of predictors (seismic attributes here) as well as
original and regularized target signals (sand fraction here).
h) If entropy of regularized target is comparable with that of the predictor
signal and the regularization result is satisfactory, then regularization is
completed else go to step d).
Output : Regularized target signal y r (t )
Comparison between the spectrums of the sand fraction (Fig.
4 (a)) and that of the seismic impedance (Fig. 4 (b)), reveals the
presence of higher order frequencies in the former.
5
research to find the prediction capability of a given nonlinear
mapping process. The original and regularized SF signals are
presented in Fig. 4(c) by the blue and red curves respectively.
The FT based regularization method is more sophisticated
than bandpass filtering in the time domain. In order to carry out
bandpass filtering, two cut-off frequencies of the ideal bandpass
filter are defined beyond which all frequency components are
removed. In case of ideal bandpass filters, zero phase shift is an
important requirement. However, ideal bandpass filter is not
realizable while working with practical datasets having finite
number of samples. Moreover, in case of practical time-domain
bandpass filtering, a transition takes place between the pass
band and the stop band i.e. the frequency component beyond
the cut-off frequencies does not diminish abruptly. In case of
FT based regularization, the frequency components of the SF
spectrum beyond the selected parameter range (-0.2:+0.2 Hz)
are completely truncated before reconstruction of regularized
SF log. Additionally, the phase shift resulting from bandpass
filtering can yield lead-lag relationships between actual and
filtered logs [28]. On the other hand, FT based regularization
does not incur phase shift between actual and regularized logs
as shown in Fig. 4(c).
TABLE I
ENTROPY ( IN BIT) OF PSD OF SIGNALS FOR WELL A
Variables
Entropy Value
0.15
Seismic Amplitude
0.16
Inst. frequency
0.12
Original SF
0.28
Regularized
SF
Seismic Impedance
FT
0.17
WD
0.20
EMD
0.24
TABLE II
NORMALIZED MUTUAL INFORMATION (NMI) AMONG PREDICTORS AND
TARGET SAND FRACTION FOR WELL A
Seismic
Seismic
Inst.
Predictor Variable
Impedance
Amplitude
Frequency
Regularized
SF
Original Sand Fraction
Fig. 4. Regularization based on FT to reconstruct target signal along Well
A; (a) SF spectrum, (b) seismic impedance spectrum, and (c) superimposed
plot of actual and regularized SF
It can be observed from Fig. 4(b) that the spectrum of bandlimited seismic impedance diminishes beyond a certain
frequency range (-0.2: +0.2 Hz). Then, the part of the sand
fraction spectrum belonging to a slightly wider frequency range
(green curve, Fig. 4(a)) is reconstructed to obtain the
regularized target. The wider range of frequency is chosen with
the assumption that ANN as a nonlinear predictor is capable of
mapping input signals of lower frequencies to output signals
with higher frequencies. Of course it needs an entirely different
0.12
0.11
0.09
FT
0.16
0.16
0.14
WD
0.15
0.14
0.11
EMD
0.14
0.14
0.12
As shown in Table I, the information content of the original
sand fraction is higher as compared to that of the seismic
predictor variables which makes it difficult to model the target
(sand fraction) from predictor attributes. The regularization
process decreases the information content in the sand fraction
as seen in Table I. The dependency between predictor and target
variables in terms of NMI (Table II) also improves as a result
of regularization.
2)
Wavelet Decomposition (WD)
Wavelet decomposition has been extensively used in
different fields of research. For example, it is applied in EEG
signal for artifact removal [29], [30] and study of geomagnetic
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
signals [31]–[33] etc. A time-frequency representation of a one
dimensional non-stationary signal is obtained using wavelet
analysis. Wavelets are oscillating functions localized in time
and frequency [29], [34], [35]. A finite energy time domain
signal can be decomposed and expressed in terms of scaled and
shifted versions of a mother wavelet ( t ) and a corresponding
scaling function ( t ) after decomposition. The scaled and
shifted form of the mother wavelet l , k (t ) and the
corresponding scaling function l , k ( t ) are mathematically
represented as
l , k (t ) 2l / 2 (2l t k ), l , k R
(4)
l/2
initial levels of original target signal are made zero and
regularized signal is constructed by performing Inverse
Discrete Wavelet Transform (IDWT) from the modified
coefficients.
Signal
G(k)
H(k)
2
2
Level 1
G(k)
H(k)
2
2
l
l , k (t ) 2 (2 t k ), l , k R
(5)
The original signal X (t ) is first decomposed into high
frequency and low frequency components using high pass and
low pass filters. After each filtering step, the output time series
is down-sampled by two. The low frequency part approximates
the signal while the high frequency part denotes residuals
between original and approximate signal. At successive levels
the approximate component is further decomposed using the
same set of high-pass and low-pass filters. A time domain signal
X (t ) can be expressed in terms of the aforementioned mother
wavelet l , k (t ) and corresponding scaling function l , k ( t ) at
level l as
X (t ) al (k )l , k (t ) dl (k ) l , k (t )
k
6
k
(6)
where al (k) and d l ( k ) are the approximate and detailed
coefficients at level l . These coefficients are computed using
filter bank approach as in [36]. Fig. 5 describes the steps of
wavelet decomposition for three levels. Here, a signal is
decomposed into approximate and detailed coefficients using
low pass H (k ) and high pass G(k ) filters respectively. After
decomposition, the coefficients can be modified. In case of
signal reconstruction, the modified approximate and detailed
coefficients are up sampled by two and then convolved with
respective synthesis filters and then the resulting pair is
summed. Finally, modified signal is acquired following l level
synthesis.
The performance of wavelet analysis is dependent on the
mother wavelet selection and decomposition level. The
Daubechies family of wavelets has a compact support with
relatively more number of vanishing moments [34]. Therefore,
in most of the cases different variants of Daubechies family
wavelets are used for signal analysis. The initial choice for
mother wavelet and decomposition level can be db4 and six
respectively for this kind of study. However, the initial choice
of wavelet and decomposition level can be modified if the
regularization results are not satisfactory. In this paper, the
choice of mother wavelet and decomposition level has been
carried out empirically based on the regularization result and
performance of the ANN. Algorithm II describes the steps
associated with WD based regularization.
Fig. 6(a)-(b) represent the results of WD–based
regularization of target sand fraction with predefined wavelet
type, and decomposition level. The first three detailed
coefficients of original sand fraction signal are demonstrated in
Fig. 6 (a). After the decomposition, detailed coefficients of
Level 2
G(k)
H(k)
2
2
Level 3
Fig. 5. Demonstration of wavelet decomposition of a signal for level 3
Algorithm II: SF Regularization Based on WD
Task : Regularizing target SF based on Wavelet Decomposition
Input : Predictor signal x (t ) and target signal y(t )
a) Same as Algorithm-I.
b) Select the wavelet type and number of decomposition levels.
c) Apply the procedure as in Fig. 5 to the target signal.
d) Decide: detailed coefficients to be truncated for regularization
e) The selected detailed coefficients are made zero.
f) The regularized target signal is reconstructed from the modified
coefficients.
g) Calculate entropies of predictors as well as original and
regularized target signals.
h) If entropy of regularized target is comparable with that of the
predictor signal and the regularization result is satisfactory, then
regularization is completed, else go to step d).
Output : Regularized target signal y r (t )
Fig. 6. Regularization based on WD to reconstruct the sand fraction along
Well A; (a) decomposition result of the actual SF (d1-d3), (b) superimposed
plot of actual and regularized SF
For Well A, first five detailed coefficients are truncated and
regularized target is reconstructed from approximate and
detailed coefficients of the sixth level by IDWT. The
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
regularization result for Well A is presented in Fig. 6 (b), where
blue and red curves represent original and regularized target
sand fraction signals respectively. Table I-II reveal the changes
in information content of the original and regularized sand
fraction by WD and increase of dependency between target and
predictor variables as a result of regularization.
3) Empirical Mode Decomposition (EMD)
Seismic and well log signals are non-stationary signals.
Reports suggest that in most of the cases the frequency analysis
of signals are carried out in selected windows with respect to a
given orthogonal basis [37]–[39]. The disadvantage of basis
decomposition techniques is the mismatch between signal trend
and constant basis functions. These necessitate a new
decomposition
method,
namely
Empirical
Mode
Decomposition (EMD). EMD is an algorithmic decomposition
method which decomposes the input signal into a set of Intrinsic
Mode Functions (IMFs) and a residue signal [40]. There are two
properties associated with IMFs such that (1) the numbers of
zero–crossings and extrema present in IMFs are same, and (2)
IMFs are symmetric with respect to the local mean [40]. In other
words, EMD detects and extracts the highest frequency
component in the signal [34], [41]–[47]; such that, in step
(k+1), the extracted IMF contains lower frequency component
compared to that extracted in step k. Moreover, being an
adaptive data-driven method, EMD decomposes an input signal
into a variable number of components. Thus, EMD overcomes
the inherent limitation of deciding a priori the number of
decomposition levels as in WD. Algorithm III describes the
detailed steps associated with EMD based regularization of
target SF.
7
g) The number of IMFs and distribution of IMFs are observed for target
y(t ) and predictor x (t ) : p n
h) Decide p1 : number of IMFs truncated from the EMD of y(t ) for
regularization where, p1 p
i) Construct regularized target signal y r ( t ) :
p1
yr (t )
imf (t ) r (t)
i
p
i 1
j) Calculate entropies of predictors as well as original and regularized
target signals.
k) If entropy of regularized target is comparable with that of the predictor
signal and the regularization result is satisfactory, then regularization
is completed else go to step h).
Output : Regularized target signal y r (t )
In Fig. 7 (a)-(b), the first four IMFs of the sand fraction log
and the IMFs and residue of the decomposed seismic
impedance along the Well A are plotted.
Algorithm III SF Regularization Based on EMD
Task : Regularizing target sand fraction based on EMD
Input : Predictor signal x (t ) and target signal y(t )
a) Same as Algorithm I.
b) Initialize: r0 (t ) x(t ) , i 1
c) Extract the
ith IMF:
i. Initialize: h0 (t ) ri (t ) , j 1
ii. Extract the local minima and maxima of h j 1 ( t )
iii. Create upper envelope emax (t ) and lower envelope emin (t ) of
h j 1 ( t ) by interpolating local maxima and minima
e
(t ) emin (t )
iv. Calculate mean envelope: m j 1 (t ) max
2
v. h j (t ) h j 1 (t ) m j 1(t )
vi. If h j ( t ) is an IMF, then, imf i (t ) h j (t )
else go to step- (ii). with j j 1
d) ri (t ) ri 1(t ) imfi (t )
e) If ri (t) has at least two extrema, then, go to c) with i i 1
n
else, x (t )
imf (t ) r (t ) is decomposed into n numbers of IMFs
i
n
i 1
and residue signal.
f) EMD of target signal y(t ) is carried out following steps b)-e)
p
y (t )
imf (t ) r (t )
i
i 1
p
Fig. 7. EMD–based regularization to reconstruct sand fraction signal along
Well A; (a) IMF1-IMF4 of the decomposed SF, (b) IMF1-residue of the
decomposed seismic impedance (c) superimposed plot of actual and
regularized SF
The comparison between Fig. 7 (a)-(b) reveals that the number
of IMFs obtained is higher in case of target signal (sand
fraction) than its predictor counterpart (seismic-impedance).
The first IMF component is suppressed and the other IMFs are
used to reconstruct the regularized sand fraction. The
superimposed plots of actual (blue curve) and regularized sand
fraction (red curve) signals are presented in Fig. 7 (c). The user
decides the regularization result is satisfactory or not based on
visual inspection of original and regularized target variables.
The regularized target is smoother compared to original signal;
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
nevertheless, the trend of the original signal is preserved even
after information filtering based on either of the three proposed
regularization methods. The number p1 is selected such that it
is less than that of the decomposed IMFs of the original signal
to be regularized. It is usually selected as one less than the
number of decomposed IMFs of the original signal. If the
regularization result is satisfactory in terms of entropy criterion
and superimposed plots of actual and regularized SF, the initial
p1 value is retained as in Algorithm III. Otherwise, the value
of p1 is to be modified. The initial selection of p1 would be
one less than the number of decomposed IMFs of the original
target signal for the working datasets. Table I represents the
entropies of predictor and target attributes for Well A. The
improvement in mutual dependency between predictor and
target variables in terms of NMI is evident from Table II.
The FT is a linear transform where the frequency spectrum
of a time-domain signal is obtained. Based on the comparison
between the frequency spectrums of the predictor seismic
impedance and target SF, the regularization parameter is
selected. The result of FT based regularization changes
according to the value of the selected regularization parameter.
If the value of the selected regularization parameter were very
small, then the regularized signal would become smooth by
shedding the finer trends of the original target signal. Again,
regularization with large parameter value would not carry out
sufficient information filtering to enable effective modelling of
the target variable. The basic assumption in FT is stationarity of
a signal. However, seismic and SF signals are non-stationary.
Therefore, the next two approaches based on WD and EMD are
opted to cross-validate the performance of FT based
regularization. The selection of the mother wavelet and the
decomposition level determines the granularity of the
decomposed target signal. The regularized target signal is
reconstructed from the modified coefficients. These parameters
(wavelet type, decomposition level, and detailed coefficients to
be truncated) are finalized empirically based on the
regularization result. However, in case of EMD, the number of
parameters is less. Only, the number of IMFs to be used for
reconstruction of the target property needs to be decided. In all
the three cases, it is ensured that the regularized target signal
retains the trend of the original signal. In case of FT based
approach, the regularization parameter is decided based on the
FT results of both the predictor as well as the target signal. The
number of IMFs of the seismic attributes is less compared to
those of SF in case of EMD. While in both the cases of WD and
EMD based regularization, the number of detailed coefficients
and IMFs are decided irrespective of the decomposition results
of the predictor seismic signals. Instead these parameters are
decided based on the regularization results and entropy criterion
as mentioned in Algorithm II and Algorithm III respectively.
IV. MODEL BUILDING AND VALIDATION
In order to establish the efficacy of the proposed
regularization method, the task of model building and
validation stage is carried out using both the original and
regularized sand fraction as target variable. The training dataset
is created by aggregating 70% sample patterns from each of the
wells. The training patterns are scrambled to remove any
8
possible trend along the depths. The remaining 30% samples
from each of the four wells are combined and then divided into
two parts to create the testing and the validation datasets. First,
the network is trained using training patterns with initial
parameter values. Then, the network structure and activation
functions are tuned using testing patterns. The testing phase is
important for evaluating the generalization capability of the
trained network [11]. The network which performs
satisfactorily in terms of CC, RMSE, AEM and SI is then
selected for use in the validation stage. The statistical analysis
of the errors involved in the model are important for proper
understanding of the performance. The initial network structure
is decided intuitively depending on nature of the problem and
amount of available training patterns. In this study several runs
of training, testing and validation of neural network structures
with varying number of neurons, layers, activation functions
and learning methods have been carried out to decide the best
structure as well as the most effective learning algorithm.
Finally in the hidden layer, hyperbolic tangent sigmoid transfer
function has been used. The tangent sigmoid transfer function
is an automatic choice for researchers for use in hidden layer to
achieve the bi-directional swing [48], [49]. The activation
function used in the output layer is log-sigmoid which is nonsymmetric in nature and facilitates faster learning rate [27]. The
number of nodes in the input layer is same as the number of
predictor attributes to be used to model the ANN. For example,
in case of predicting sand fraction from three predictor
attributes– namely, seismic impendence, instantaneous
frequency and amplitude, the number of input and output nodes
will be three and one respectively. Finally, a network with
single hidden layer is trained using the Scaled Conjugate
Gradient (SCG) Back propagation Algorithm [50]. The backpropagation learning using the SCG method is given in the
Appendix. SCG proceeds in the conjugate direction of the
previous step instead of following the gradient direction as in
other gradient descent based algorithms. In case of the
conjugate gradient algorithms, the step size can be selected by
a line search without evaluating the Hessian matrix. However,
the line search method is itself computationally expensive due
to the evaluation of multiple error functions. Moreover, the line
search deals with multiple parameters crucial for its
performance. Sometimes the Hessian matrix becomes non
positive definite which leads to increase in the error in weight
updates due to ill conditioning in the matrix inversion. In SCG,
the Hessian matrix is made positive definite. The
computationally inefficient line search in every iteration is
avoided using a step size scaling mechanism in SCG. Another
advantage of SCG is that it does not involve any user dependent
parameters. The SCG method has been used to solve problems
in different research domains such as prediction of groundwater
level [51], and telemarketing success [52], etc.
Table II reveals that the NMIs between instantaneous
frequency and SF (original/regularized) are relatively lower
compared to other cases. In the first attempt, two input
attributes (seismic impedance and amplitude) have been used to
build the prediction model and the corresponding results are
documented in Table III in terms of the performance evaluators.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
Though variation of instantaneous frequency is comparatively
lower, it is an important attribute. Therefore, all three seismic
attributes have been used as inputs to the prediction model
(Table IV) in the second attempt. The performance comparison
of the method of averaging and the proposed regularization
procedures has been carried out. The original SF is filtered by a
moving average filter with a span of nine samples, before being
passed to the ANN training module. Table V represents the
validation performance of the ANN in modelling the averaged
SF from the seismic impedance, amplitude and instantaneous
frequency. The prediction results improve in terms of
performance evaluators in case of filtered SF by averaging as
compared to the original SF as target. It can be observed from
Table IV and Table V that the performance of the trained ANN
is superior while working with the processed SF using the
proposed regularization as target compared to the original and
averaged SF.
The results reported in Table III, Table IV, and Table V lead
to the following important observations:
The performance of the trained networks is improved with
the use of regularized target signals. This can be quantified
in terms of higher CCs and lower error values.
In all cases, inclusion of the instantaneous frequency as the
third predictor improves the prediction.
The proposed regularization technique is superior to the
averaging method.
Further it can be observed from Table III and Table IV that the
network performance is superior in case of regularization based
on WD with two predictor variables. On the other hand, with
three predictors FT based regularization outperformed the other
two regularization approaches in terms of performance
evaluators. However, for all cases, the performance is improved
while using regularized sand fraction as target instead of
original log. The extent of advantage of each scheme over the
others may vary among different datasets. Therefore a given
scheme cannot be universally recommended. A user is provided
with multiple choices among which the user can select the
regularization technique that best suits with the working dataset
before application of machine learning algorithm.
Literature study reveals that pre-processing is an important
step prior to modelling. For example, normalization, relevant
attributes selection, importance sampling etc. have been part of
9
the pre-processing technique in different fields such as
environmental modelling [53], chemistry [54], biomedical [55],
[56], reservoir engineering [17], [57], [58]. Generally, CC
values above 80%; RMSE, AEM values below 0.15 and SI
value below 0.35 can be considered as good fit in such studies.
If the prediction model attains the values of the respective
performance indicators beyond the above mentioned limits,
then the model should be retrained after adjusting associated
parameters. Baziar et al. [57] has compared the performance of
three machine learning techniques such as SVM, multilayer
perceptron (MLP), and Co-Active Neuro-Fuzzy Inference
System (CANFIS) to predict permeability from well logs.
Further, the normalization has been carried out as the preprocessing technique before modelling. However, integration
of seismic and borehole dataset are not required since only well
logs are used as the predictor variables in [57]. The
performances of the three approaches (SVM, MLP, CANFIS)
have been quantified in terms of CC, average absolute error
(AAE, equivalent to AEM), and mean squared error (MSE i.e.
square of RMSE). The methodology used to model the original
SF is almost similar to the methodology adopted for the
permeability prediction using multilayer perceptron as in [57].
Comparison between the validation performance (prediction
using data not part of training) reported in [57] and those
achieved in this study has revealed that the regularization stage
improves the prediction result in terms of CC, AEM, and
RMSE. The pre-processing step involving a single ANN in [17]
is similar to the case in the present paper while modelling with
original SF as target. Comparison among the ANN
performances with and without the regularization step reveals
that the results of ANN has improved in terms of evaluators in
the former case.
In the case of unsatisfactory validation performance, the user
may modify the ANN structures and network parameters and
retrain the modified ANN. If the validation performance still
does not improve, the user can modify the regularization
parameters and carry out the modelling with regularized target
signal. By changing the regularization parameters the
information retention in the filtered/reconstructed changes e.g.
by choosing a narrower bandwidth in the FT based method the
well logs become smoother with a reduced entropy.
Well
Name
A
B
C
D
TABLE III
STATISTICS OF VALIDATION PERFORMANCE (TWO PREDICTORS)
Original Target Sand Fraction(SF)
Filtered by Fourier
Filtered by EMD
CC
RMSE AEM
SI
CC
RMSE AEM
SI
CC
RMSE AEM
0.63
0.21
0.17
0.7
0.69
0.15
0.12
0.52 0.71
0.17
0.13
0.57
0.19
0.15
0.6
0.63
0.15
0.12
0.5
0.62
0.16
0.12
0.68
0.16
0.12
0.53 0.78
0.12
0.09
0.37 0.76
0.12
0.08
0.54
0.18
0.14
0.58 0.63
0.14
0.11
0.46 0.64
0.12
0.09
SI
0.64
0.49
0.37
0.41
CC
0.74
0.65
0.77
0.66
Filtered by Wavelets
RMSE
AEM
0.15
0.12
0.15
0.12
0.11
0.08
0.12
0.09
SI
0.53
0.5
0.36
0.39
Well
Name
A
B
C
D
TABLE IV
STATISTICS OF VALIDATION PERFORMANCE (THREE PREDICTORS)
Filtered by Fourier
Filtered by EMD
SI
CC
RMSE AEM
SI
CC
RMSE AEM
0.62 0.94
0.07
0.05
0.23 0.92
0.09
0.06
0.56 0.86
0.09
0.07
0.30 0.88
0.09
0.06
0.45 0.91
0.07
0.05
0.24 0.87
0.08
0.06
0.49 0.83
0.1
0.07
0.33 0.80
0.09
0.06
SI
0.31
0.28
0.25
0.32
CC
0.93
0.87
0.89
0.84
Filtered by Wavelets
RMSE
AEM
0.08
0.06
0.10
0.08
0.08
0.05
0.08
0.06
SI
0.28
0.34
0.25
0.28
Original Target Signal
CC
RMSE AEM
0.76
0.17
0.13
0.65
0.17
0.13
0.76
0.14
0.11
0.68
0.15
0.12
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
TABLE V
STATISTICS OF VALIDATION PERFORMANCE (THREE PREDICTORS, TARGET:
AVERAGED SF)
Filtered Sand Fraction (SF) by Averaging
Well
Name
CC
RMSE
AEM
SI
A
0.84
0.13
0.10
0.45
B
0.77
0.14
0.10
0.44
C
0.86
0.11
0.08
0.33
D
0.73
0.13
0.10
0.43
The prediction in the entire volume along each inline and cross
line has been carried out using the network selected after
acceptable performance in the validation step. Fig. 8 represents
the variation of the seismic amplitude at a specific inline (inline
136 containing Well D). The sand fraction variation at the same
inline shown in Fig. 9 is obtained by prediction of the sand
fraction over the study area from available seismic attributes
using validated network parameters. The network used here for
prediction has been calibrated using EMD–regularized–sand
fraction as target.
10
area contained by the selected window [59], [60]. Thus, the
neighborhood values of the particular pixel play an important
role in the determination of the modified pixel value as a result
of order filtering. In case of order statistics filtering the window
size is selected to define the neighborhood around the centered
pixel. Selection of window value is crucial for the degree of
smoothing. The predicted sand fraction in the volume is used as
input to the post-processing operation. Every element in the
volume is considered as a pixel and is smoothened using 3-D
median filter with respect to its neighborhood within a 3x3x3
window size. The missing values along the boundaries are
ignored. In order to carry out median filtering, the values of the
pixel and its neighbors within the selected window are first
sorted. Then, the centered pixel value is replaced by the median
value determined from the sorted pixel values. For example, in
case of 3-D median filtering with a 3x3x3 window, the 14th
largest value of the neighborhood replaces the pixel value at the
center of the neighborhood. The SF logs over the area are
predicted from seismic attributes and variation of the predicted
SF at a particular inline is presented in Fig. 9. The detailing of
the SF variation is revealed by the color code in Fig. 9. The
smoothened SF values attained by median filtering may lose the
exactness of the predicted SF logs; however, the volumetric
representation of the filtered SF collectively offers better
understanding of the SF variation. Fig. 10 represents the result
of median filtering along inline 136. The effect of localizing
different levels of sand fraction values can be observed by
comparing Fig. 9 and Fig. 10.
Fig. 8. Variation of the seismic amplitude at inline 136
Fig. 10. Result of 3-D median filtering with a 3x3x3 window on the
predicted sand fraction variation at inline 136
Fig. 9. Variation of the predicted sand fraction at inline 136
V. POST-PROCESSING
It can be observed from Fig. 8 that the variation of the seismic
attributes over the study area is smooth. However, the sand
fraction across the study area as shown in Fig. 9 changes
abruptly. The transition of the sand fraction values should be
smoother and more or less agree to the patterns of seismic data.
Thereby rises the need for the post–processing stage in order to
obtain a smoother sand fraction variation across the volume. To
incorporate this rationale, the predicted values are filtered
through a 3-D median filter.
Median filter is an order statistic filter i.e. non-linear spatial
filter. In case of order statistics filters, the filtered output is
dependent on the ordering (ranking) of the pixels in the image
Post-processing with larger window sizes (for example
5x5x5 or higher order) would further smoothen out the
predicted SF volume at the cost of detailed SF variation. In our
subsequent research attempts, adaptive post processing method
having the potential to further improve the prediction result will
be worked out. This would help in subsurface characterization.
Different types of spatial filters would be experimented to
observe the changes in filtered SF.
Thus, the complete framework including pre-processing,
learning and validation, and finally post-processing,
successfully carries out mapping between seismic attributes and
the sand fraction.
VI. DISCUSSION AND CONCLUSION
This paper brings out a complete workflow consisting of an
elegant regularization step in pre-processing to enhance the
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
learning capability of ANN in carrying out mapping between
seismic attributes and lithological property (sand fraction)
successfully. The improvement in mapping with introduction of
the regularization step is observed from the performance
analysis. The comparison among the results achieved in this
work and existing literatures [17], [57] reveals the superiority
of the proposed regularization step to obtain improved
performance in terms of the performance evaluators. In the
present study, synthetic SF logs are generated over the study
area from available seismic information using the validated
network parameters. After obtaining SF volume, postprocessing is carried out to improve visualization by means of
median filtering. Similarly, other important petrophysical
properties such as porosity, permeability can be modelled from
seismic attributes following the proposed workflow. The
volume-wise prediction results of SF and other petrophysical
properties (i.e. porosity, permeability) enables the user to
identify zones with high sand content and porosity, and higher
probability of hydrocarbon presence. Thus, this study would
help to identify potential drilling locations of a new well in a
study area.
The present work differs from the work done in [17] in terms
of pre-processing stage, division of training-testing dataset,
adopted machine learning technique, and post-processing
algorithm. In addition the well tops and horizon information are
used [17] to carry out zone wise division of the overall dataset
and modular ANN is applied to model SF from multiple seismic
attributes. The performance attained in [17] would be improved
with the use of regularized SF log as target instead of original
one.
The selection of initial parameters are crucial for achieving
acceptable performance of ANN. In the present work, the ANN
structure and activation functions in hidden and output layers
have been empirically adjusted. Also the initial values of weight
and bias matrices have been chosen randomly. As a possible
future extension, the domain of metaheuristic algorithms can be
explored in order to automate such selections to further
strengthen the framework. Additionally, while in the postprocessing stage, the use of spatial filtering has provided a
significant improvement in the sand fraction variation over the
study area; adaptive post–processing method can be
investigated in future.
APPENDIX
THE BACK-PROPAGATION LEARNING OF AN ANN USING SCALED C ONJUGATE
GRADIENT LEARNING
The back-propagation learning of a multilayer perceptron is
carried out in two phases. The synoptic weights of the network
are constant and the input signal is propagated through the
hidden layers to the output layer in the forward phase. The
changes only take place in the activation potentials and output
of the neurons [27]. In contrast, an error signal is computed as
the difference between network output and actual (desired)
target (response). The error signal is propagated through layers
in backward direction from output layer to input layer in the
backward phase. The training procedure is described in details
as follows.
Assume,
{x(n), d (n)}nN1
11
(7)
to be the set of training samples used for learning of a back
propagation multilayer perceptron. The input vector x ( n ) is
applied to the input layer nodes and the desired output vector
d (n) is present at the output node to compute the error between
the desired and the actual output. The forward signal is
propagated from the input layer to the output layer through
single or multiple hidden layers. The induced local field for
neuron j in layer l can be computed from the output of neuron
i in the previous layer (l 1) at iteration n i.e. yil 1 (n) and the
synaptic weight wlji (n) that is connected from neuron i in
( l 1) layer and is expressed as
v (jl ) (n)
w(jil ) (n) yi(l 1) (n)
(8)
i
In case of output layer (here, l L and L is the network depth),
the output of neuron j is written as
y(jL) (n) o j (n)
Therefore, the error is computed as
e j ( n) d j ( n ) o j ( n)
(9)
(10)
The supervised learning of a multilayer ANN can be viewed as
a problem of numerical optimization. The error surface of a
multilayer ANN is a nonlinear function of weight vector w .
Assume that, the error energy averaged over the training
samples or the empirical risk be avg . Using (10), avg can be
computed as
1 N
2
e j (n)
(11)
2 N n 1 jC
where, the set C contains all the neurons in the output layer.
The second derivative of the cost function avg with respect to
avg
the weight vector w is called the Hessian matrix and denoted
by H so that,
2 avg
(12)
H
w2
The Hessian matrix is considered as positive definite unless
mentioned otherwise. There are several algorithms to train an
ANN. In case of conjugate gradient methods, the computational
complexity and memory usage are large because of calculation
and storage of the Hessian matrix at each stage. The
indefiniteness of H is handled by a scaling coefficient k 0
in case of the scaled conjugate gradient (SCG). The other
parameters r k and p k represent the search direction and the
steepest descent direction respectively. Different training
algorithms have been experimented to train the single layer
ANN to model SF from multiple seismic attributes. Finally, the
SCG algorithm is selected over other algorithms to train the
ANN due to its superiority in terms of obtained performances
evaluators. The steps associated with the SCG can be presented
as follows [50]:
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
Algorithm IV: Scaled Conjugate Gradient (SCG) to train a network
Task: Obtain weight vector after training
Input : Training dataset, weight vector, and parameters associated with
training
1 and 0 104 ,
a) Selection of parameters: Weight vector w
1 0 . Assume,
0 1 10 4 ,
[5]
[6]
[7]
p r E '(w
1 ), k 1 and success = true
1 1
b) If success = true , then compute the second-order information:
/ p
k
[8]
k
k p ) E '( w
k )) /
s k ( E '( w
k k
k
[9]
T
k p k s k
k ) p 2
c) Modify k : k k (k
k
[10]
d) If k 0 , then make the Hessian matrix positive definite
2
k 2(k k / pk )
2
k k k p k
k
[11]
[12]
k
T
pk r k
e) Evaluate step size: k
k k / k
f) Compute the comparison parameter:
[13]
[14]
2
k ) E (w
k
k 2 k [ E ( w
k p k )] / k
g) If k 0 , then error can be reduced:
k 1 w
k p
w
k k
k 1 )
r k 1 E '( w
k 0, success = true
[15]
[16]
[17]
If k mod N 0 then
p
k 1 r k 1
2
else:
T
k ( r k 1 r k 1 r k ) / k
p
k 1 r k 1 k p k
If k 0.75 , modify the scale parameter, k 0.25k
else
k k , success = false
h) If k 0.25 , increase the scale parameter
2
k k (k (1 k ) / pk )
k 1 as the
i) Check: If the steepest descent direction r k 0, then w
desired minimum. Otherwise, k k 1 and go to b).
k 1
Output: Trained weight vector w
REFERENCES
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[1]
[2]
[3]
[4]
G. Álvarez, B. Sansó, R. J. Michelena, and J. R. Jiménez, “Lithologic
Characterization of a Reservoir Using Continuous-Wavelet Transforms,”
IEEE Trans. Geosci. Remote Sens., vol. 41, no. 1, pp. 59–65, Jan. 2003.
M. Bosch, T. Mukerji, and E. F. Gonzalez, “Seismic inversion for
reservoir properties combining statistical rock physics and geostatistics:
A review,” Geophysics, vol. 75, no. 5, pp. 75A165–75A176, Sep. 2010.
M. Ali Ahmadi, S. Zendehboudi, A. Lohi, A. Elkamel, and I. Chatzis,
“Reservoir permeability prediction by neural networks combined with
hybrid genetic algorithm and particle swarm optimization,” Geophys.
Prospect., vol. 61, no. 3, pp. 582–598, May 2013.
S. Yu, X. Guo, K. Zhu, and J. Du, “A neuro-fuzzy GA-BP method of
seismic reservoir fuzzy rules extraction,” Expert Syst. with Appl., vol. 37,
no. 3, pp. 2037–2042, Mar. 2010.
[25]
[26]
[27]
[28]
12
D. P. Hampson, J. S. Schuelke, and J. a. Quirein, “Use of multiattribute
transforms to predict log properties from seismic data,” Geophysics, vol.
66, no. 1, pp. 220–236, Jan. 2001.
P. M. Wong, A. G. Bruce, and T. T. D. Gedeon, “Confidence bounds of
petrophysical predictions from conventional neural networks,” IEEE
Trans. Geosci. Remote Sens., vol. 40, no. 6, pp. 1440–1444, 2002.
N. Sharma and M. M. Ali, “A neural network approach to improve the
vertical resolution of atmospheric temperature profiles from geostationary
satellites,” IEEE Geosci. Remote Sens. Lett., vol. 10, no. 1, pp. 34–37, Jan.
2013.
M. M. Ali, P. S. V. Jagadeesh, I.-I. Lin, and J.-Y. Hsu, “A neural network
approach to estimate tropical cyclone heat potential in the Indian ocean,”
IEEE Geosci. Remote Sens. Lett., vol. 9, no. 6, pp. 1114–1117, Nov. 2012.
E. Ostlin, H.-J. Zepernick, and H. Suzuki, “Macrocell path-loss prediction
using artificial neural networks,” IEEE Trans. Veh. Technol., vol. 59, no.
6, pp. 2735–2747, Jul. 2010.
S. España-Boquera, M. J. Castro-Bleda, J. Gorbe-Moya, and F. ZamoraMartinez, “Improving offline handwritten text recognition with hybrid
HMM/ANN models.,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 33,
no. 4, pp. 767–779, Apr. 2011.
I. Kaastra and M. Boyd, “Designing a neural network for forecasting
financial and economic time series,” Neurocomputing, vol. 10, pp. 215–
236, 1996.
C. C. Fung, K. W. Wong, and H. Eren, “Modular artificial neural network
for prediction of petrophysical properties from well log data,” IEEE
Trans. Instrum. Meas., vol. 46, no. 6, pp. 1295–1299, Dec. 1997.
P. M. Wong, T. D. Gedeon, and I. J. Taggart, “An improved technique in
porosity prediction: a neural network approach,” IEEE Trans. Geosci.
Remote Sens., vol. 33, no. 4, pp. 971–980, Jul. 1995.
J.-S. Lim, “Reservoir properties determination using fuzzy logic and
neural networks from well data in offshore Korea,” J. Pet. Sci. Eng., vol.
49, no. 3–4, pp. 182–192, Dec. 2005.
M. Nikravesh and F. Aminzadeh, “Past, present and future intelligent
reservoir characterization trends,” J. Pet. Sci. Eng., vol. 31, no. 2–4, pp.
67–79, Nov. 2001.
M. Nikravesh and M. Hassibi, “Intelligent reservoir characterization (
IRESC),” in Proc. IEEE Int. Conference on Ind. Informatics, 2003, pp.
369–373.
S. Chaki, A. K. Verma, A. Routray, W. K. Mohanty, and M. Jenamani,
“Well tops guided prediction of reservoir properties using modular neural
network concept: a case study from western onshore, India,” J. Pet. Sci.
Eng., vol. 123, pp. 155 – 163, Jul. 2014.
K. M. Barry, D. A. Cavers, and C. W. Kneale, “Report on recommended
standard for digital tape formats,” Geophysics, vol. 40, no. 2, pp. 344–
352, 1975.
J.
R.
Victorine,
“Kansas
Geological
Survey,”
http://www.kgs.ku.edu/stratigraphic/PROFILE/HELP/Help-PCSaveLASFile.html. .
A. Douillard, D. Schlumberger, and O. Balz, “Upscaling well logs to
seismic scale – comparison of a wavelet-transform based method with a
geostatistical technique,” in Petrophysics meets Geophysics, 2000, no.
November, pp. 1–9.
L. Stright, A. Boucher, T. Mukherji, and R. Derksen, “Revisiting the use
of seismic attributes as soft data for subseismic facies prediction:
proportions versus probabilities,” Lead. Edge, vol. 28, no. 12, pp. 1460–
1679, 2009.
A. V Oppenheim, R. W. Schafer, and J. R. Buck, Discrete-Time Signal
Processing. Prentice Hall, 1999.
T. M. Cover and J. A. Thomas, Elements of Information Theory. New
York, NY, USA: John Wiley & Sons, Inc., 1991.
T. Mukerji, P. Avseth, G. Mavko, I. Takahashi, and E. F. González,
“Statistical rock physics: combining rock physics, information theory, and
geostatistics to reduce uncertainty in seismic reservoir characterization,”
Lead. Edge, vol. 20, no. 3, pp. 313–319, Mar. 2001.
G. Balasis, R. Donner, S. Potirakis, J. Runge, C. Papadimitriou, I. Daglis,
K. Eftaxias, and J. Kurths, “Statistical Mechanics and InformationTheoretic Perspectives on Complexity in the Earth System,” Entropy, vol.
15, no. 11, pp. 4844–4888, Nov. 2013.
P. a Estévez, M. Tesmer, C. a Perez, and J. M. Zurada, “Normalized
mutual information feature selection,” IEEE Trans. Neural Netw., vol. 20,
no. 2, pp. 189–201, Feb. 2009.
S. Haykin, Neural networks: a comprehensive foundation. New Jersey,
USA: Prentice hall, 1999.
H. Steehouwer and K. M. Lee, “A Zero Phase Shift Band Pass Filter,”
2012.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
[29] S. Kar, A. Routray, and B. P. Nayak, “Functional network changes
associated with sleep deprivation and fatigue during simulated driving:
validation using blood biomarkers.,” Clin. Neurophysiol., vol. 122, no. 5,
pp. 966–974, May 2011.
[30] A. Chaudhuri, A. Routray, and S. Kar, “Effect of sleep deprivation on
estimated distributed sources for scalp EEG signals : A case study on
human drivers,” in IEEE 4th Int. Conf. Intelligent Human Comput.
Interaction, 2012.
[31] G. Balasis, I. A. Daglis, M. Georgiou, C. Papadimitriou, and R.
Haagmans, “Magnetospheric ULF wave studies in the frame of Swarm
mission: a time-frequency analysis tool for automated detection of
pulsations in magnetic and electric field observations,” Earth, Planets Sp.,
vol. 65, no. 11, pp. 1385–1398, Nov. 2013.
[32] G. Balasis, I. a. Daglis, E. Zesta, C. Papadimitriou, M. Georgiou, R.
Haagmans, and K. Tsinganos, “ULF wave activity during the 2003
Halloween superstorm: multipoint observations from CHAMP, Cluster
and Geotail missions,” Ann. Geophys., vol. 30, no. 12, pp. 1751–1768,
Dec. 2012.
[33] N. Zaourar, M. Hamoudi, M. Mandea, G. Balasis, and M. Holschneider,
“Wavelet-based multiscale analysis of geomagnetic disturbance,” Earth,
Planets Sp., vol. 65, no. 12, pp. 1525–1540, Dec. 2013.
[34] S. Mallat, A Wavelet Tour of Signal Processing, 2nd Ed. New York, NY,
USA: Academic Press, 1999.
[35] I. Daubechies, “The wavelet transform, time-frequency localization and
signal analysis,” IEEE Trans. Inf. Theory, vol. 36, no. 5, pp. 961–1005,
Sep. 1990.
[36] O. Rioul and M. Vetterli, “Wavelets and Signal Processing,” IEEE Signal
Processing Mag., pp. 14–38, 1991.
[37] L. Cohen, Time-frequency analysis theory and applications. Prentice
Hall, 1995.
[38] P. Flandrin, Temps-frequence. Hermes, 1993.
[39] I. Daubechies, Ten lectures on wavelets. SIAM, 1992.
[40] I. Magrin-chagnolleau and R. G. Baraniuk, “Empirical mode
decomposition based time-frequency attributes,” in 69th Annual
International Meeting, SEG, Expanded Abstracts, 1999, pp. 1949–1953.
[41] N. E. Huang, Z. Shen, S. R. Long, M. C. Wu, H. H. Shih, Q. Zheng, N.C. Yen, C. C. Tung, and H. H. Liu, “The empirical mode decomposition
and the Hilbert spectrum for non-linear and non-stationary time series
analysis,” in Proc. R. Soc. London. Ser. A Math. Phys. Eng. Sci., vol. 454,
no. 1971, pp. 903–995, 1998.
[42] D. Labate, F. La Foresta, G. Occhiuto, F. C. Morabito, A. Lay-ekuakille,
and P. Vergallo, “Empirical Mode Decomposition vs . Wavelet
Decomposition for the Extraction of Respiratory Signal from SingleChannel ECG : A Comparison,” IEEE Sensors J., vol. 13, no. 7, pp. 2666–
2674, Jul. 2013.
[43] N. Mammone, F. La Foresta, F. C. Morabito, and S. Member, “Automatic
artifact rejection from multichannel scalp EEG by wavelet ICA,” IEEE
Sensors J., vol. 12, no. 3, pp. 533–542, Mar. 2012.
[44] D. B. Perciva and A. T. Walden, Wavelet Methods for Time Series
Analysis. U.K.: Cambridge Univ. Press, 2000.
[45] N. ur Rehman and D. P. Mandic, “Filter Bank Property of Multivariate
Empirical Mode Decomposition,” IEEE Trans. Signal Process., vol. 59,
no. 5, pp. 2421–2426, May 2011.
[46] N. Rehman and D. P. Mandic, “Multivariate empirical mode
decomposition,” Proc. R. Soc. London. Ser. A Math. Phys. Eng. Sci., vol.
466, no. 2117, pp. 1291–1302, 2010.
[47] C. D. Blakely, A Fast Empirical Mode Decomposition Technique for
Nonstationary Nonlinear Time Series, New York, NY, USA: Elsevier,
2005.
[48] B. L. Kalman and S. C. Kwasny, “Why Tanh : chossing a sigmoidal
function,” in Int. Joint Conf. on Neural Networks, 1992, vol. 4, pp. 578–
581.
[49] M. Leshno, V. Y. Lin, A. Pinkus, and S. Schocken, “Multilayer
feedforward networks with a nonpolynomial activation function can
approximate any function,” Neural Networks, vol. 6, pp. 861–867, 1993.
[50] M. F. Moller, “A scaled conjugate gradient algorithm for fast supervised
learning,” Neural Networks, vol. 6, pp. 525–533, 1993.
[51] S. Maiti and R. K. Tiwari, “A comparative study of artificial neural
networks, Bayesian neural networks and adaptive neuro-fuzzy inference
system in groundwater level prediction,” Environ. Earth Sci., vol. 71, no.
7, pp. 3147 –3160, 2014.
[52] S. Moro, P. Cortez, and P. Rita, “A data-driven approach to predict the
success of bank telemarketing,” Decis. Support Syst., vol. 62, pp. 22–31,
2014.
13
[53] H. R. Maier and G. C. Dandy, “Neural networks for the prediction and
forecasting of water resources variables : a review of modelling issues and
applications,” Environ. Model. Softw., vol. 15, pp. 101–124, 2000.
[54] Å. Rinnan, F. Van Den Berg, and S. B. Engelsen, “Review of the most
common pre-processing techniques for near-infrared spectra,” Trends
Anal. Chem., vol. 28, no. 10, pp. 1201–1222, Nov. 2009.
[55] A. Smolinska, A.-C. Hauschild, R. R. R. Fijten, J. W. Dallinga, J.
Baumbach, and F. J. van Schooten, “Current breathomics--a review on
data pre-processing techniques and machine learning in metabolomics
breath analysis.,” J. Breath Res., vol. 8, no. 2, p. 027105, Jun. 2014.
[56] M. Rathi and C. Gupta, “An approach to predict breast cancer and drug
suggestion using machine learning techniques,” ACEEE Int. J. Inf.
Technol., vol. 4, no. 1, pp. 23 – 31, 2014.
[57] S. Baziar, M. Tadayoni, M. Nabi-Bidhendi, and M. Khalili, “Prediction
of permeability in a tight gas reservoir by using three soft computing
approaches: A comparative study,” J Nat. Gas Sci. Eng., vol. 21, pp. 718–
724, Nov. 2014.
[58] S. Chaki, A. K. Verma, A. Routray, W. K. Mohanty, and M. Jenamani,
“A one-class classification framework using SVDD : application to an
imbalanced geological dataset,” in Proc. IEEE Students’ Technology
Symp. (TechSym), 2014, pp. 76–81.
[59] R. Gonzalez and R. Woods, Digital image processing, Second. New
Jersey, USA: Prentice hall, 2002.
[60] T. S. Huang, G. J. Yang, and G. Y. Tang, “A Fast Two-Dimensional
Median Filtering Algorithm,” IEEE Trans. Acoust., Speech, Signal
Process., vol. ASSP-27, no. 1, pp. 13–18, Feb. 1979.
Soumi Chaki (S’2014) received the B.E.
degree from IIEST Shibpur (formerly
BESU, Shibpur), India in 2010. She is
currently working toward the M.S. degree
in Electrical Engineering from Indian
Institute of Technology Kharagpur in India.
Her research interests include signal
processing, machine learning algorithms.
Aurobinda Routray (M’2012) is a
Professor with the Department of
Electrical Engineering, IIT Kharagpur in
India.
His research interests include nonlinear
and statistical signal processing, real-time
and embedded signal processing,
numerical lineal algebra and data-driven diagnostics.
William K. Mohanty is a Professor with
the Department of Geology and
Geophysics, IIT Kharagpur in India.
His
research
interests
include
seismology, reservoir characterization.
| 5cs.CE
|
LOVÁSZ-SAKS-SCHRIJVER IDEALS AND COORDINATE SECTIONS OF
DETERMINANTAL VARIETIES
arXiv:1801.07916v1 [math.AC] 24 Jan 2018
ALDO CONCA AND VOLKMAR WELKER
Abstract. Motivated by questions in algebra and combinatorics we study two ideals associated
to a simple graph G:
• the Lovász-Saks-Schrijver ideal defining the d-dimensional orthogonal representations of
the graph complementary to G and
• the determinantal ideal of the (d + 1)-minors of a generic symmetric with 0s in positions
prescribed by the graph G.
In characteristic 0 these two ideals turn out to be closely related and algebraic properties such
as being radical, prime or a complete intersection transfer from the Lovász-Saks-Schrijver ideal
to the determinantal ideal. For Lovász-Saks-Schrijver ideals we link these properties to combinatorial properties of G and show that they always hold for d large enough. For specific classes
of graphs, such a forests, we can give a complete picture and classify the radical, prime and
complete intersection Lovász-Saks-Schrijver ideals.
Introduction
Let k be a field, n ≥ 1 be an integer and set [n] := {1, . . . , n}. For a simple graph G = ([n], E)
with vertex set [n] and edge set E we study the following two classes of ideals associated to G.
• Lovász-Saks-Schrijver ideals:
For an integer d ≥ 1 we consider the polynomial ring k[yi,` | i ∈ [n], ` ∈ [d]]. For every
edge e = {i, j} ∈ [n]
we set
2
fe(d) =
d
X
yi` yj` .
`=1
The ideal
LkG (d) = ( fe(d) | e ∈ E ) ⊆ k[yi,` | i ∈ [n], ` ∈ [d]]
is called the Lovász-Saks-Schrijver ideal, LSS-ideal for short, of G with respect to k. The
ideal LkG (d) defines the variety of orthogonal representations of the graph complementary
to G (see [28, 27]).
• Coordinate sections of generic (symmetric) determinantal ideals:
Consider the polynomial ring k[xij | 1 ≤ i ≤ j ≤ n] and let X be the generic n × n
symmetric matrix, that is, the (i, j)-th entry of X is xij if i ≤ j and xji if i > j. Let
sym
XG
be the matrix obtained from X by replacing the entries in positions (i, j) and (j, i)
sym
for {i, j} ∈ E with 0. For an integer d let Idk (XG
) ⊆ k[xij | 1 ≤ i ≤ j ≤ n] be the ideal
sym
sym
k
of d-minors of XG . The ideal Id (XG ) defines a coordinate hyperplane section of the
generic symmetric determinantal variety. Similarly, we consider ideals defining coordinate
hyperplane sections of the generic determinantal varieties and the generic skew-symmetric
Pfaffian varieties.
sym
k
We observe in Section 7 that the ideal Id+1
(XG
) and the ideal LkG (d) are closely related.
Indeed, if k has characteristic 0 classical results from invariant theory can be employed to show
sym
k
that Id+1
(XG
) is radical (resp. is prime, resp. has the expected height) provided LkG (d) is
Date: January 25, 2018.
2010 Mathematics Subject Classification. Primary .
Key words and phrases.
The work on this paper was partly supported by a DAAD Vigoni project and by INdAM.
1
2
A. CONCA AND V. WELKER
radical (resp. is prime, resp. is a complete intersection). We also exhibit similar relations between
variants of LkG (d) and ideals defining coordinate sections of determinantal and Pfaffian ideals.
These facts turn the focus on algebraic properties of the LSS-ideals LkG (d). In particular, we
analyze the questions: when is LkG (d) a radical ideal? when is it a complete intersection? when
is it a prime ideal? We show that for d large enough all three properties hold for LkG (d). Among
others, we are able to give necessary conditions that lead to a full classification of graphs for
which LkG (d) is a complete intersection or prime in case of small d. In characteristic 0 we deduce
sym
k
sufficient conditions for Id+1
(XG
) to be radical, prime or of expected height. To our knowledge
coordinate sections of determinantal varieties have been systematically studied only in the case of
maximal minors, for example the results in [5, 17, 18].
The study of the properties of LSS-ideals has its roots in the work Lovász on orthogonal graph
embeddings (see [27] for references, motivation and an overview) and we think it is interesting in
its own. An orthogonal embedding of a graph G in Rd is a map φ : V → Rd where φ(i) ⊥ φ(j)
if {i, j} is an edge in the edge set Ē of the graph Ḡ complementary to G. Thus by definition
the real variety associated to the LSS-ideal LRḠ (d) of the complementary graph Ḡ coincides with
the set of all orthogonal graph embeddings of G in Rd . Note that the variety includes degenerate
embeddings that are not injective or send vertices to the zero vector. Since the geometry of the
variety of orthogonal graph embeddings was first studied in [28] the ideals LkG (d) carry the name
Lovász-Saks-Schrijver ideals. Indeed, many of our algebraic results are inspired by results from
[27, 28] about the geometry of the real variety of (general position) orthogonal embeddings.
For d = 1, 2 LSS-ideals are well understood objects. For d = 1 the LSS-ideal LkG (1) is called
edge ideal of the graph G. As a squarefree monomial ideal it is clearly a radical ideal with respect
to every field. It is prime only when E is empty and a complete intersection if and only if G is
a matching, i.e. any two edges in E have empty intersection. Starting from d = 2 the properties
of being radical, prime and complete intersection become more subtle. For the results in this case
see [22]. For d > 2 we know of no general results beyond the ones described in Section 2.
In Section 1.2 we generalize LSS-ideals to hypergraphs. We are able to state a few of the
results from Section 2 in this generality. But most questions on hypergraph LSS-ideals remain
unanswered. Nevertheless, extending the link between LSS-ideals for graphs and ideals of minors,
hypergraph LSS-ideals for uniform hypergraphs can be related to the closure of the space of
symmetric tensors of bounded rank with prescribed 0s in the their expansion in the standard basis
(see Proposition 8.10).
The paper is organized as follows. In Section 1 we introduce basic concepts and notation from
graph theory and Gröbner bases theory. Then in Section 2 we formulate our main results on LSSideals and sketch some of the proofs. In Section 3 we provide the proofs for the results showing
persistence of the properties complete intersection and primeness. In particular, it follows that
for fixed d there are graphs which are minimal obstructions to these properties. In Section 4
we exhibit some of these obstructions and prove their necessity. For small d we give complete
characterizations of graphs with prime or complete intersection LSS-ideals. In Section 5 we define
a new combinatorial invariant for graphs. We use it to prove that LkG (d) is radical, complete
intersection or prime for d large enough. In Section 6 we define our notation for ideals of minors
and Pfaffians of generic matrices and recall classical results about their relation to invariant theory.
Then in Section 7 we use this connection to invariant theory to prove that if LkG (d) is radical or
prime then same property hold for the ideal of (d + 1) minors of generic matrices with positions of
0s prescribed by a graph G. In addition, we give obstructions on G preventing the corresponding
ideals of minors to be prime. Finally, in Section 8 we pose questions and state open problems.
1. Notations and generalities
1.1. Graph and Hypergraph Theory. In the following we introduce graph theory notation.
We mostly follow the conventions from [12]. For us a graph G = (V, E) is a simple
graph on a
finite vertex set V . In particular, E is a subset of the set of 2-element subsets V2 of V . In most
of the cases we assume that V = [n] := {1, . . . , n}. A subgraph of a graph G = (V, E) is a graph
G0 = (V, E 0 ) such that E 0 ⊆ E. More generally, a hypergraph H = (V, E) is a pair consisting
LSS-IDEALS AND COORDINATE SECTIONS
3
of a finite set of vertices V and a set E of subsets of V . We are only interested in the situation
when the sets in E are inclusionwise incomparable. Such a set of subsets is called a clutter. In
particular, if G = (V, E) is a graph then G is a hypergraph and E is a clutter.
For d, m, n > 0 we will use the following notations:
• Kn denotes the complete graph on n vertices ([n], [n]
2 ),
• Km,n denotes the complete bipartite graph ([m] ∪ [ñ], {{i, j̃} : i ∈ [m], j̃ ∈ [ñ] } with
bipartition [m] and [ñ] = {1̃, . . . , ñ}.
• Bn denotes the subgraph of Kn,n obtained by removing the edges {i, ĩ} with i = 1, . . . , n.
• For n > 2 we denote by Cn the n-cycle, i.e. the subgraph of Kn with edges {1, 2}, {2, 3}, . . . ,
{n − 1, n}, {n, 1}.
• For n > 1 we denote by Pn the path of length n − 1, i.e. the subgraph of Kn with edges
{1, 2}, {2, 3}, . . . , {n − 1, n}.
We denote by Ḡ = (V, Ē) with Ē = V2 \ E the graph complementary to G = (V, E). Let
W ⊆ V . We write GW = (W, {e ∈ E : e ⊆ W }) for the graph induced by G on vertex set W
and G − W for the subgraph induced by G on V \ W . In case W = {v} for some v ∈ V we simply
write G − v for G − {v}.
A graph G = (V, E) on a vertex set V of size ≥ k + 1 is called k-(vertex)connected if for every
W ⊂ V with |W | = k − 1 the graph G − W is connected. By deg(v) := {e ∈ E | v ∈ e} we
denote the degree of the vertex v and by ∆(G) = maxv∈V deg(v) the maximal degree of a vertex
in G. Clearly, if G is k-connected then every vertex has degree at least k and ∆(Ḡ) ≤ n − k − 1,
where n = |V |. Finally, we denote by ω(G) the clique number of G, i.e. the largest a such that G
contains a copy of a complete subgraph Ka .
The following well known fact follows directly from the definitions.
Lemma 1.1. Given a graph G = ([n], E) and an integer 1 ≤ d ≤ n the following conditions are
equivalent:
(1) Ḡ is (n − d)-connected.
(2) G does not contain a subgraph Ka,b with a + b = d + 1.
1.2. Basics on LSS-ideals and their generalization to hypergraphs. Let H = (V, E) be a
hypergraph. We define for e ∈ E
d Y
X
fe(d) =
yi` .
`=1 i∈e
If E is a clutter we call the ideal
LkH (d) = ( fe(d) | e ∈ E ) ⊆ k[yi,` | i ∈ [n], ` ∈ [d]]
the LSS-ideal of the hypergraph H.
It will sometimes be useful to consider LkH (d) as a multigraded ideal. For that we equip
k[yi,` | (i, `) ∈ [n] × [d]] with the multigrading induced by deg(yi,` ) = ei for the i-th unit vector ei
P
(d)
in Zn and (i, `) ∈ V × [d]. Clearly, for e ∈ E the polynomial fe is multigraded of degree i∈e ei .
In particular, LkH (d) is Zn -multigraded. The following remark is an immediate consequence of
(d)
(d)
the fact that if E is a clutter the two polynomials fe and fe0 corresponding to distinct edges
e, e0 ∈ E have incomparable multidegrees.
(d)
Remark 1.2. Let H = ([n], E) be a hypergraph such that E is clutter. The generators fe , e ∈ E,
of LkH (d) form a minimal system of generators. In particular, LkH (d) is a complete intersection if
(d)
and only if the polynomials fe , e ∈ E, form a regular sequence.
The following alternative description of LkG (d) for a graph G turns out to be helpful in some
places.
Remark 1.3. Let G = ([n], E) be a graph. Consider the n × d matrix Y = (yi,` ). Then LkG (d) is
the ideal generated by the entries of the matrix Y Y T in positions (i, j) with {i, j} ∈ E. Here Y T
denotes the transpose of Y .
4
A. CONCA AND V. WELKER
Similarly, for a bipartite graph G, say a subgraph of Km,n , one considers two sets of variables
yij with (i, j) ∈ [m] × [d], yĩj with (i, j) ∈ [ñ] × [d] and the matrices Y = (yij ) and Ye = (yĩj ). Then
LkG (d) coincides with the ideal generated by the entries of the product matrix Y Ye T . in positions
(i, j) for {i, j̃} ∈ E
1.3. Gröbner Bases. We use the following notations and facts from Gröbner bases theory, see
for example [4] for proofs and details. Consider the polynomial ring k[yij | (i, j) ∈ [n] × [d]]. For
a vector w = (wij )(i,j)∈[n]×[d] ∈ Rnd and a polynomial
X
f (y) =
aα yα
α∈N[n]×[d]
we set mw (f ) = maxaα 6=0 {α · w} and
inw (f ) =
X
aα · yα .
α·w=mw (f )
The latter is called the initial term of f with respect to w. For a fixed term order ≺ and w ∈ R[n]×[d]
we set yα ≺ω yβ if and only if either α · w < β · w or α · w = β · w and yα ≺ yβ .
The following will allows us to deduce properties of ideals from properties of their initial ideals.
Proposition 1.4. Let I be a homogeneous ideal in the polynomial ring S, ≺ a term order on
S and w ∈ Rnd . If in≺ (I) or inw (I) is radical (resp. a complete intersection, prime) then
so is I. Moreover, if for generators f1 , . . . , fr of I the initial terms in≺ (f1 ), . . . , in≺ (fr ) (resp.
inw (f1 ), . . . , inw (fr )) form a regular sequence then f1 , . . . , fr form a regular sequence and are a
Gröbner basis for I.
2. Results and counterexamples for Lovász-Saks-Schrijver ideals
For the first part of this section let G = (V, E) be a graph. We start by studying radicality of
LkG (d). As mentioned in the Introduction LkG (1) is always radical for trivial reasons. For d = 2
the following result from [22] gives a complete answer.
Theorem 2.1 (Thm. 1.1 [22]). Let G = ([n], E) be a graph. If char k 6= 2 then the ideal LkG (2) is
radical. If char k = 2 then LkG (2) is radical if and only if G is bipartite.
The next examples show that LkG (3) need not be radical. In the examples we assume that k
has characteristic 0 but we consider it very likely that the ideals are not radical over any field.
A quick criterion implying that an ideal J in a polynomial ring S is not radical is to identify
an element g ∈ S such that J : g 6= J : g 2 . We call such a g a witness (of the fact that J is not
radical). Of course the potential witnesses must be sought among the elements that are somehow
“closely related” to J. Alternatively, one can try to compute the radical of J or even its primary
decomposition directly and read off whether J is radical. But these direct computations are
extremely time consuming for LSS-ideals and did not terminate on our computers in the examples
below. Nevertheless, in all examples we have quickly identified witnesses.
Example 2.2. We present three examples of graphs G such that LkG (3) is not radical over any field
k of characteristic 0. The first example has 6 vertices and 9 edges and it is the smallest example
we have found (both in terms of edges and vertices). The second example has 7 vertices and 10
edges and it is a complete intersection. This shows that LkG (3) can be a complete intersection
without being radical. The third example is bipartite, a subgraph of K5,4 with 12 edges, and is
the smallest bipartite example we have found. In all cases, since the LSS-ideal LkG (3) has integral
coefficients, we may assume that k = Q and exhibit a witness g, i.e. a polynomial g such that
LkG (3) : g 6= LkG (3) : g 2 . The latter inequality can be checked with the help of CoCoA [1] or
Macaulay 2 [19].
(1) Let G be the graph with 6 vertices and 9 edges depicted in Figure 1(1), i.e. with edges
E = {{1, 2}, {1, 3}, {1, 4}, {1, 5}, {2, 3}, {2, 4}, {2, 6}, {3, 5}, {4, 6}}.
LSS-IDEALS AND COORDINATE SECTIONS
1
2
5
1
6
4
3
(1)
2
5
6
5
1
2
3
4
5
7
4
3
1̃
(2)
2̃
3̃
4̃
(3)
Figure 1. Graphs G with non-radical LkG (3)
Here the witness can be chosen as follows. Denote by Y = (yij ) a generic 6 × 3
T
matrix. As discussed in Remark 1.3 the ideal LQ
G (3) is generated by the entries of Y Y
corresponding to the positions in E. Now g can be taken as the 3-minor of Y with row
indices 1, 5, 6.
(2) Let G be the graph with 7 vertices and 10 edges depicted in Figure 1(2), i.e. with edges
E = {{1, 2}, {1, 4}, {1, 5}, {2, 3}, {2, 7}, {3, 4}, {3, 7}, {4, 5}, {5, 6}, {6, 7}}.
Here the witness can be chosen as follows. Denote by Y = (yij ) a generic 7 × 3 matrix.
T
Again as discussed in Remark 1.3 the ideal LQ
G (3) is generated by the entries of Y Y
corresponding to the positions in E. Now g can be taken as the 3-minor of Y with row
indices 1, 2, 4. The fact that LQ
G (3) is a complete intersection can be checked quickly with
CoCoA [1] or Macaulay 2 [19].
(3) Let G be the subgraph of the complete bipartite graph K5,4 depicted in Figure 1(3), i.e.
with edges
E = {{1, 1̃}, {1, 2̃}, {1, 3̃}, {1, 4̃}, {2, 1̃}, {2, 2̃}, {3, 2̃}, {3, 3̃}, {4, 3̃}, {4, 4̃}, {5, 1̃}, {5, 4̃}}.
Denote by X = (xij ) a generic 5 × 3 matrix and by Y = (yij ) a generic 3 × 4 matrix. As
explained in Remark 1.3 the ideal LQ
G (3) is generated by the entries of XY corresponding
to the positions in E. Now the witness g can be taken to be the 3-minor of X corresponding
to the column indices 1, 2, 4.
The following result shows that the properties complete intersection and prime for LkG (d) are
closely linked and persist once they occur.
Theorem 2.3. Let G = ([n], E) be a graph. Then:
(1) If LkG (d) is prime then LkG (d) is a complete intersection.
(2) If LkG (d) is a complete intersection then LkG (d + 1) is prime.
(3) If LkG (d) is prime then LkG (d + 1) is prime.
(4) If LkG (d) is a complete intersection then LkG (d + 1) is a complete intersection.
(5) If LkG (d) is a complete intersection then LkG0 (d) is a complete intersection for every subgraph G0 of G.
(6) If LkG (d) is a prime then LkG0 (d) is prime for every subgraph G0 of G.
The proof of the theorem consists of several steps that we first briefly sketch and then present
in full detail in Section 3.
Sketch of the proof. To prove (1) and (2) we interpret LkG (d) as the defining ideal of the symmetric
algebra of a module over the quotient of the polynomial ring by LkG−n (d). Then we show that the
statement follows by induction on n employing a result of Avramov [2, Prop. 3] characterizing
complete intersection symmetric algebras and a result of Huneke [24, Thm 1.1] characterizing
symmetric algebras that are domains.
6
A. CONCA AND V. WELKER
To prove (3) and (4) we consider the vector v = (vij )(i,j)∈[n]×[d+1] ∈ Rn(d+1) with entries vij = 1
(d+1)
(d)
if (i, j) ∈ [n] × [d] and vi,d+1 = 0 for every i ∈ [n]. Observe that inv (fe
) = fe for all e ∈ E.
Therefore, inv (LkG (d + 1)) ⊆ LkG (d). Either LkG (d) is a complete intersection by assumption or by
(1) in case the assumption is that LkG (d) is prime. This implies that inv (LkG (d+1)) = LkG (d). Then
the assertions follow by the transfer of properties from inv (J) to J as recalled in Proposition 1.4.
Assertion (5) is obvious. For (6) one observes that by (1) LkG (d) is also a complete intersection.
It is a general fact that if a regular sequence of homogeneous polynomials generates a prime ideal
then so does every subsequence.
Remark 2.4. There is no persistence result for the property of being radical. Indeed, we already
have seen that LkG (1) is always radical and LkG (2) is always radical in case char k 6= 2. On the
other hand Example 2.2 gives examples of a non-radical LkG (3). Simple examples also show that
radicality is not inherited by subgraphs.
On the other hand radicality is inherited by induced subgraphs. This follows from the fact that
for every subset W ⊆ V one has
LkGW (d) = LkG (d) ∩ k[yij : i ∈ W, j ∈ [d]].
as can be checked using the multigraded structure.
We will now see that LkG (d) is a complete intersection and prime (and hence radical) for d large
enough.
We prove this fact in Section 5. Indeed, more generally we show that for d 1 and a hypergraph
H = (V, E), where E is a clutter, the hypergraph LSS-ideals LkH (d) is radical and a complete
intersection.
As a vehicle we define a purely (hyper)graph theoretic invariant pmd(H) ∈ N called the positive
matching decomposition of H and show the following.
Theorem 2.5. Let H = (V, E) be a hypergraph for a clutter E. Then:
(1) For all d ≥ pmd(H) the ideal LkH (d) has a radical complete intersection initial ideal.
(2) For all d ≥ pmd(H) the ideal LkH (d) is a radical complete intersection.
(3) If G = (V, E) is graph then for all d ≥ pmd(G) + 1 the ideal LkG (d) is prime.
(4) If G = (V, E) is a graph then pmd(G) ≤ min{2|V |−3, |E|} Furthermore, if G is a bipartite
graph then pmd(G) ≤ min{|V | − 1, |E|}.
The proof of the theorem consists of several steps that we first briefly sketch and then present
in full details in Section 5.
Sketch of the proof. We show that for every d ≥ pmd(G) there exists a vector v = (vij )(i,j)∈[n]×[d] ∈
(d)
Rnd such that the set {inv (fe ) : e ∈ E} consists of pairwise coprime monomials. It follows that
(d)
inv (LkG (d)) = (inv (fe ) : e ∈ E) and hence inv (LkG (d)) is radical and a complete intersection. This
complete the proof of (1). By Proposition 1.4 then (2) is implied. Therefore, in the graph case
Theorem 2.3(2) implies (3). The claim (4) is derived by simple estimates using the combinatorial
structure of the graph.
For complete graphs G = Kn and char k = 0 we provide asymptotic (in terms of n) results on
Kn
when LG
(d) is radical, complete intersection or prime in Proposition 7.6 and Corollary 7.7 using
sym
k
n
the transfer of properties from LK
G (d) to Id+1 (XG ) and bounds derived using Gröbner basis
arguments in Corollary 5.5.
For the case of complete bipartite graphs Km,n results by De Concini and Strickland [13] or
Musili and Seshadri [31] imply (1)-(3) of the following theorem.
Theorem 2.6. Let G = Km,n . Then:
(1) LkG (d) is radical for every d.
(2) LkG (d) is a complete intersection if and only if d ≥ m + n − 1.
(3) LkG (d) is prime if and only if d ≥ m + n.
(4) pmd(G) = m + n − 1.
LSS-IDEALS AND COORDINATE SECTIONS
7
Proof. Taking into account Remark 1.3, the assertions (1), (2), and (3) follow form general results
on the variety of complexes proved from [13] and, with different techniques, from [31]. It has been
observed by Tchernev [33] that the assertions in [13] that refer to the so-called Hodge algebra
structure of the variety of complexes in [13] are not correct. However, those assertions can be
replaced with statements concerning Gröbner bases as it is done, for example, in a similar case in
[33]. Hence, (1),(2) and (3) can still be deduced from the arguments in [13]. Alternative proofs of
(2) and (3) are given also in Section 4. Assertion (4) is proved in Section 5.
Seeing Theorem 2.6(2) and (3) one may wonder if the assertion (2) in Theorem 2.3 can be
reversed. The next example shows that in general this is not the case:
Example 2.7. Let G = Cn be the n-cycle. If n is even and n 6= 4 then the ideal LkG (d) is prime
if and only if it is a complete intersection if and only if d ≥ 3. These assertions are special cases
of the subsequent Theorem 2.9.
In view of Theorem 2.3 for fixed d and k the graphs G for which LkG (d) is a complete intersection
or prime define (downward) monotone graph properties. Thus by persistence there are graphs G
and numbers d such that LkG (d) is not prime, LkG0 (d) is prime for each proper subgraph G0 of
G and LkG (d + 1) is prime. Such a pair G and d can be considered as a “minimal obstruction”
to primeness. Similarly, we have minimal obstructions to being a complete intersection. The
next results are first steps towards a classification of minimal obstructions. The results are partly
inspired by theorems from Lovász’s book [27, Ch. 9.4].
Proposition 2.8. Let G = ([n], E). Then we have:
(1) If LkG (d) is prime then G does not contain a subgraph isomorphic to Ka,b with a+b = d+1,
i.e. Ḡ is (n − d)-connected. Furthermore, if d > 3 and char k = 0 then G does not contain
a subgraph isomorphic to Bd .
(2) If LkG (d) is a complete intersection then G does not contain a subgraph isomorphic to Ka,b
with a + b = d + 2, i.e. Ḡ is (n − d + 1)-connected. Furthermore, if d > 2 and char k = 0
then G does not contain a subgraph isomorphic to Bd+1 .
Further obstructions are derived from Proposition 7.6 in Corollary 7.8. For example, in characteristic 0, if LkG (6) is prime then G cannot contain K6 . But for small values of d the implications
of Proposition 2.8 are actually equivalences.
Theorem 2.9. Let G be a graph.
(1) For d ≤ 3 then the following are equivalent:
(a) LkG (d) is prime.
(b) G does not contain a subgraph isomorphic to Ka,b with a + b = d + 1.
(c) Ḡ is (n − d)-connected.
(2) For d ≤ 2 the following conditions are equivalent:
(a) LkG (d) is a complete intersection
(b) G does not contain a subgraph Ka,b with a + b = d + 2 and when d = 2 the graph G
does not contain an even cycle.
(c) Ḡ is (n − d + 1)-connected and when d = 2 the graph G does not contain an even
cycle.
For forests (i.e. graphs without cycles) we can give a quite complete picture.
Theorem 2.10.
(1) LkG (d) is
(2) LkG (d) is
(3) LkG (d) is
Let G be a forest and k any field. Then:
radical for all d.
a complete intersection if and only if d ≥ ∆(G).
prime if and only if d ≥ ∆(G) + 1.
The main tool in the proof of Theorem 2.10 is the notion of Cartwright-Sturmfels ideals developed in [9, 10, 11] and inspired by [7, 8]. Indeed it turns out that for a forest G the ideal LkG (d)
is a Cartwright-Sturmfels ideal.
8
A. CONCA AND V. WELKER
3. Stabilization of algebraic properties of LkG (d)
In this section we prove Theorem 2.3 and state some of its consequences. Before embarking in
the proofs we need to recall important results on properties of the symmetric algebra of a module.
We will state the results in the way that suit our needs best.
Recall that, given a ring R and an R-module M presented as the cokernel of an R-linear map
f : Rm → Rn
the symmetric algebra SymR (M ) of M is (isomorphic to) the quotient of SymR (Rn ) = R[x1 , . . . , xn ]
by the ideal J generated by the entries of A (x1 , . . . , xn )T where A is the m×n matrix representing
f . Vice versa every quotient of R[x1 , . . . , xn ] by an ideal J generated by homogeneous elements
of degree 1 in the xi ’s is the symmetric algebra of an R-module.
Part (1) of the following is a special case of [2, Prop. 3] and part (2) a special case of [24, Thm
1.1]. Here and in the rest of the paper we denote for a matrix A with entries in a ring R and a
number t by It (A) the ideal generated by the t-minors of A in R.
Theorem 3.1. Let R be a complete intersection. Then
(1) SymR (M ) is a complete intersection if and only if height It (A) ≥ m − t + 1 for all t =
1, . . . , m.
(2) SymR (M ) is a domain and Im (A) 6= 0 if and only if R is a domain, and height It (A) ≥
m − t + 2 for all t = 1, . . . , m.
The equivalent conditions of (2) imply those of (1).
Remark 3.2. Let G = ([n], E) be a graph. The (multigraded) ideal LkG (d) ⊆ S = k[yi,j : i ∈
[n], j ∈ [d]] is generated by elements that have degree at most one in each block of variables.
Hence LkG (d) can be seen as an ideal defining a symmetric algebra in various ways.
For example, set G1 = G − n, U = {i ∈ [n − 1]|{i, n} ∈ E}, u = |U |, S 0 = k[yi,j : i ∈ [n − 1], j ∈
[d]] and R = S 0 /LkG1 (d). Then S/LkG (d) is the symmetric algebra of the cokernel of the R-linear
map
Ru → Rd
associated to the u × d matrix A = (yij ) with i ∈ U and j = 1, . . . , d.
Remark 3.3. In order to apply Theorem 3.1 to the case described in Remark 3.2 it is important
to observe that for every G no minors of the matrix (yij )(i,j)∈[n]×[d] vanish modulo LkG (d). This
is because LkG (d) is contained in the ideal J generated by the monomials yik yjk and the terms in
the minors of (yij ) do not belong to J for obvious reasons.
We will formulate our next results in terms of the following algebraic invariants. Given an
algebraic property P of ideals and a graph G we set
asymk (P, G) = inf{d : LkG (d0 ) has property P for all d0 ≥ d}.
Here we are of course interested in the properties P ∈ {radical, c.i., prime}. Other properties
of ideals such as defining a normal ring or a UFD are interesting as well but will not be treated
here.
Before we use this new notation we provide the proof of Theorem 2.3(4) and (5).
Proof of Theorem 2.3(4) and (5). Assume LkG (d) is a complete intersection. Then each minimal
(d)
generating set is a regular sequence. By Remark 1.2 the fe for e ∈ E form a minimal generating
set and hence a regular sequence. In particular, each subset is a regular sequence as well. From
this (5) follows.
For (4) we consider the vector w assigning weight 1 to all variables yi` for (i, `) ∈ V × [d] and
(d0 )
(d)
weight 0 to all other variables. Then inw (fe ) = fe for every d0 ≥ d and e ∈ E. Hence the
(d0 )
initial forms of the generators of LG (d0 ) form a regular sequence. It follows that the fe , e ∈ E,
form a regular sequence and hence LkG (d0 ) is a complete intersection.
In terms of asymk (c.i., G) Theorem 2.3(4) yields the following corollary.
LSS-IDEALS AND COORDINATE SECTIONS
9
Corollary 3.4. Let G = (V, E) be a graph. Then
asymk (c.i., G) = inf{d : LkG (d) is c.i.}.
Now we are in position to prove Theorem 2.3 (1) and (6).
Proof of Theorem 2.3 (1) and (6). First we show that (1) implies (6). From Remark 1.2 we know
(d)
that the fe , e ∈ E, form a minimal system of generators. Thus if LkG (d) is a complete intersection
then these generators form a regular sequence. If a regular sequence generates a prime ideal in
a standard graded algebra or in a local ring then so does every subset of the sequence. Now (6)
follows.
To prove (1) we argue by induction on the number n of vertices. As usual we assume V = [n].
The case n ≤ 2 is trivial. We use the notation from Remark 3.2. Note, that S 0 /LkG1 (d) is an
algebra retract of S/LkG (d). Therefore LkG1 (d) = LkG (d) ∩ S 0 and so LkG1 (d) is prime. By induction
it follows that LkG1 (d) is a complete intersection. Since u is the degree of the vertex n in G we have
that K1,u ⊂ G and hence, by Proposition 2.8, 1 + u < d + 1, i.e. u < d. By virtue of Remark 3.3
we have that the minors of the matrix A are non-zero in S 0 /LkG1 (d). In particular, Iu (A) 6= 0 in
S 0 /LkG1 (d) and hence (2) in Theorem 3.1 holds. Then (1) in Theorem 3.1 holds as well, i.e. LkG (d)
is a complete intersection.
As an immediate corollary of Theorem 2.3(1) we obtain.
Corollary 3.5. Let G = (V, E) be a graph. Then asymk (c.i., G) ≤ asymk (prime, G).
Before we can proceed to the proof of Theorem 2.3(2), we need another technical lemma.
Lemma 3.6. Let A be an m × n matrix with entries in a Noetherian ring R. Assume m ≤ n.
Let S = R[x] = R[x1 , . . . , xm ] be a polynomial ring over R and let B be the m × (n + 1) matrix
with entries in S obtained by adding the column (x1 , . . . , xm )T to A. Then we have height I1 (B) =
height I1 (A) + m and
height It (B) ≥ min{height It−1 (A), height It (A) + m − t + 1}
for all 1 < t ≤ m.
Proof. Set u = min{height It−1 (A), height It (A)+m−t+1}. Let P be a prime ideal of S containing
It (B). We have to prove that height P ≥ u. If P ⊇ It−1 (A) then height P ≥ height It−1 (A) ≥ u.
If P 6⊇ It−1 (A) then we may assume that the (t−1)-minor F corresponding to the first (t −1) rows
and column of A is not in P . Hence, height P = height P RF [x] and P RF [x] contains It (A)RF [x]
and (xj − F −1 Gj : j = t, . . . , m) with Gj ∈ R[x1 , . . . , xt−1 ]. Since the elements xj − F −1 Gj are
algebraically independent over RF we have
height P RF [x] ≥ height It (A)RF + (m − t + 1) ≥ height It (A) + (m − t + 1).
Now we turn to the proof of Theorem 2.3(2) and (3).
Proof or Theorem 2.3 (2) and (3). First, we show that (2) implies (3). If LkG (d) is prime then by
(1) LkG (d) is a complete intersection. Now by (2) it follows that LkG (d+1) is prime. This completes
the proof of (3).
For the proof of (2) we argue by induction on the number n of vertices.
If n ≤ 2 the assertion is obvious. Assume n > 2. Set G1 = G − n, U = {i ∈ [n]|{i, n} ∈ E},
u = |U |, Y = (yij )(i,j)∈U ×[d+1] , S = k[yij : i ∈ [n], j ∈ [d + 1]], S 0 = k[yij : i ∈ [n − 1], j ∈ [d + 1]]
and R = S 0 /LkG1 (d + 1). By construction, S/LkG (d + 1) is the symmetric algebra of the R-module
presented as the cokernel of the map Ru → Rd+1 associated to Y .
By assumption LkG (d) is a complete intersection and hence LkG1 (d) is a complete intersection
as well. It then follows by induction that LkG1 (d + 1) is prime and hence R is a domain. By
Proposition 2.8 we have u ≤ d and by Remark 3.3 Iu (Y ) 6= 0 in R. Therefore, by Theorem 3.1(2)
we have
LkG (d + 1) is prime ⇔ heightIt (Y ) ≥ u − t + 2 in R for every t = 1, . . . , u.
10
A. CONCA AND V. WELKER
Equivalently, we have to prove that
height It (Y ) + LkG1 (d + 1) ≥ u − t + 2 + g in S 0 for every t = 1, . . . , u
where g = height LkG1 (d + 1) = |E| − u.
Consider the weight vector w ∈ Rn×d+1 defined by wij = 1 and wi d+1 = 0 for all j ∈ [d] and
i ∈ [n]. By construction the initial terms of the standard generators of inw (LkG1 (d + 1)) are the
standard generators of LkG1 (d).
Since the standard generators of It (Y ) coincide with their initial terms with respect to inw it
follows that inw (It (Y )) ⊇ It (Y ) (indeed equality holds but we do not need this fact).
Therefore, inw (It (Y ) + LkG1 (d + 1)) ⊇ It (Y ) + LkG1 (d) and it is enough to prove that
height It (Y ) + LkG1 (d) ≥ u − t + 2 + g in S 0 for every t = 1, . . . , u
or, equivalently,
heightIt (Y ) ≥ u − t + 2 in R0 for every t = 1, . . . , u
where R0 = S 0 /LkG1 (d).
The variables y1 d+1 , . . . , yn−1 d+1 do not appear in the generators of LkG1 (d). Hence R0 =
00
R [y1 d+1) , . . . , yn−1 d+1 ] with R00 = k[yij : (i, j) ∈ [n − 1] × [d]]/LkG−n (d). Let Y 0 be the matrix Y
with the d + 1-st column removed. Then S/LkG (d) can be regarded as the symmetric algebra of
the R00 -module presented as the cokernel of
0 → (R00 )u
(1)
Y0
−→
(R00 )d .
By assumption S/LkG (d) is a complete intersection. Hence by Theorem 3.1(1) we know
height It (Y 0 ) ≥ u − t + 1 in R00 for every t = 1, . . . , u
Since Y is obtained from Y 0 by adding a column of variables over R00 by Lemma 3.6 we have:
height, It (Y ) ≥ min{height, It−1 (Y 0 ), height, It (Y 0 ) + u − t + 1} ≥ u − t + 2
in R0 and for all t = 1, . . . , m.
Theorem 2.3(2) and (3) together with Corollary 3.5 directly imply the following.
Corollary 3.7. Let G = (V, E) be a graph.
Then
asymk (prime, G) = inf{d : LkG (d) is prime}
and
asymk (c.i., G) ≤ asym(prime, G) ≤ asymk (c.i., G) + 1.
The following proposition is an immediate consequence of Theorem 2.3 and Theorem 2.5.
Proposition 3.8. Let G = (V, E) be a graph. Then
asymk (radical, G) ≤ 2 |V | − 3,
asymk (c.i., G) ≤ 2 |V | − 3 and asymk (prime, G) ≤ 2 |V | − 2.
If G is bipartite then
asymk (radical, G) ≤ |V | − 1,
asymk (c.i., G) ≤ |V | − 1 and asymk (prime, G) ≤ |V |.
These bounds are not tight in general as the following example shows.
Example 3.9. Using CoCoA [1] or Macaulay 2 [19] one can check for fields k of characteristic 0
that LkK4 (2) is not a complete intersection while LkK4 (3) is. Hence by Theorem 2.3 (4) we have
asymk (c.i., K4 ) = 3. Similarly, one can check that LkK4 (d) is not prime for d = 3 and hence
Theorem 2.3 implies that asymk (prime, K4 ) = 4. Finally, one checks that LkK4 (d) is radical for all
1 ≤ d ≤ 3 and hence asymk (radical, K4 ) = 1.
In Corollary 7.7 we will be able to analyze the asymptotic behavior of asymk (prime, Kn ) for
n → ∞.
LSS-IDEALS AND COORDINATE SECTIONS
11
4. Obstructions to algebraic properties
In this section we prove Theorem 2.9 and study necessary and sufficient conditions for LkG (d)
to be radical, complete intersections or prime. First, we turn to necessary conditions which yield
lower bounds on asymk (radical, G), asymk (prime, G) and asymk (c.i., G).
We start with the proof of Proposition 2.8.
Proof of Proposition 2.8. First, we show that (1) implies (2). If LkG (d) is complete intersection
then by Theorem 2.3 (2) it follows that LkG (d + 1) is prime. Hence if G contains Ka,b with
a + b > d + 1 or Bd+1 (in case char k = 0 and d > 2) then G violates the conditions from (1) for
primeness of LkG (d + 1). This implies (2).
For (1) we first prove that if LkG (d) is prime then G does not contain Ka,b for a+b > d. Suppose
for contradiction that LkG (d) is prime and G contains Ka,b for some a + b > d. We may decrease
either a or b or both and assume right away that a + b = d + 1, a, b ≥ 1. In particular a, b ≤ d.
We assume that V = [n] and that Ka,b is a subgraph of G, where after renaming vertices here
Ka,b is the complete bipartite graph with edges {i, a + j} for i ∈ [a] and j ∈ [b]. Set R = S/LkG (d)
and Y = (yi ` ) ∈ Ra×d and Z = (z`,i ) ∈ Rd×b with z`,i = yi+a,` . Since Ka,b is a subgraph of G we
have Y Z = 0 in R
By assumption R is a domain and Y Z = 0 can be seen as a matrix identity over the field of
fractions of R.
Hence
rank(Y ) + rank(Z) ≤ d.
From a + b = d + 1 it follows that rank(Y ) < a or rank(Z) < b. This implies that any a-minor of
Y or any b-minor of Z is zero, i.e., Ia (Y ) = 0 or Ib (Z) = 0 as ideals of R.
But by Remark 3.3 none of the minors of Y and Z are in LkG (d). This is a contradiction and
hence LkG (d) is not prime.
It remains to be shown that if char k = 0 and d > 3 then G does not contain a copy of Bd .
Here, we unfortunately have to resort to Proposition 7.4 and Lemma 7.13(iii) from Section 7.
But it is easily seen that its derivation is independent of results from preceding sections.
By Theorem 2.3(6) we know that if LkG (d) is prime then so is LkBd (d). Then Proposition 7.4
gen
implies that Id+1 (XG
) is prime for a generic matrix X of arbitrary size and this contradicts
Lemma 7.13(iii).
Next we provide the proof of Theorem 2.9.
Proof of Theorem 2.9 part (1). By Lemma 1.1 conditions (b) and (c) in Theorem 2.9(1) are equivalent. Hence it suffices to prove the equivalence of (a) and (b).
For d = 1 the statements are obvious: LkG (1) is prime if and only if G has no edges set which
is equivalent to not containing K11 .
For d = 2 we know by Proposition 2.8 that (a) implies (b). When (b) holds then the edges of
G are pairwise disjoint. It follows that the monomial ideal LkG (1) is a complete intersection. Then
by Theorem 2.3 (2) assertion (a) follows.
For d = 3 again by Proposition 2.8 condition (a) implies (b). To prove that (b) implies (a) we
may assume that k is algebraically closed. Then, since the tensor product over k of k-algebras that
are domains is a domain (see Corollary to Proposition 1 in Bourbaki’s Algebra [6, Chapter v, 17]
or [30, Prop. 5.17]) we may assume that the graph is connected. A connected graph satisfying (b)
is either an isolated vertex or a path Pn with n ≥ 2 vertices or a cycle Cn of length n 6= 4. Hence
we have to prove that LkG (3) is prime when G = Pn with n ≥ 2 or G = Cn with n ≥ 3 and n 6= 4.
If G = Pn then pmd(Pn ) ≥ 2 (indeed pmd(Pn ) ≥ 2 for n > 2). This can be seen easily form the
definition or by using Lemma 5.2(iv) to check that a maximal matching and its complement form
a positive matching decomposition. Hence by Theorem 2.5(3) it follows that LkPn (3) is prime.
Now let G = Cn be the n-cycle for n ≥ 3 and n 6= 4 and set m = n − 1. To prove that LkCn (3)
is prime we use the symmetric algebra perspective. Observe that Cn − n is Pm for m = n − 1.
Set J = LkPm (3), S = k[yij : i ∈ [m] j ∈ [3]] and R = S/J. We have already proved that J is a
12
A. CONCA AND V. WELKER
prime complete intersection of height m − 1. We have to prove that the symmetric algebra of the
cokernel of the R-linear map:
y11 y12 y13
2 Y
3
R → R with Y =
ym1 ym2 ym3
is a domain. Since by Remark 3.3 I2 (Y ) 6= 0 in R, taking into consideration Remark 3.2 we may
apply Theorem 3.1. Therefore, it is enough to prove that
height I1 (Y ) ≥ 3 and height I2 (Y ) ≥ 2 in R.
Equivalently, it is enough to prove that
(2)
height I1 (Y ) + J
≥ m + 2 and
(3)
height I2 (Y ) + J
≥ m + 1 in S.
We prove first (2). Since height I1 (Y ) = 6 in S then (2) is obvious for m ≤ 4. For m > 4 observe
that I1 (Y ) + J can be written as I1 (Y ) + H where H is the LSS-ideal of the path of length m − 2
on vertices 2, 3, . . . , m − 1. Because I1 (Y ) and H use disjoint set of variables, we have
height I1 (Y ) + H = 6 + m − 3 = m + 3
and this proves (2). Now we note, en passant, that the condition height I2 (Y ) ≥ 1 holds in R
because R is a domain and I2 (Y ) 6= 0. Hence we deduce from Theorem 3.1(1) that LkCn (3) is a
complete intersection for all n ≥ 3.
It remains to prove (3). Since height I2 (Y ) = 3 in S the assertion in obvious for m = 2,
i.e. n = 3. Hence we may assume m ≥ 4 (here we use n 6= 4). Then let P be a prime ideal
of S containing I2 (Y ) + J. We have to prove that height P ≥ m + 1. If P contains I1 (Y ) then
height P ≥ m+2 by (2). So we may assume that P does not contain I1 (Y ), say y11 6∈ P , and prove
that height P Sx ≥ m + 1 where x = y11 . Since I2 (Y )Sx = (ym2 − x−1 ym1 y12 , ym3 − x−1 ym1 y13 )
we have
(3)
fm−1,m
= ym−1,1 ym1 + ym−1,2 ym2 + ym−1,3 ym3
= ym−1,1 ym1 + ym−1,2 x−1 ym1 y12 + ym−1,3 x−1 ym1 y13
(3)
= x−1 ym1 f1,m−1
Since
(3)
fm−1,m
∈ J, we have
(3)
ym1 f1,m−1
mod I2 (Y )Sx
∈ P Sx . Then we have that either ym1 ∈ P Sx or
(3)
f1,m−1
∈ P Sx . In the first case P Sx contains ym1 , ym2 , ym3 and the LSS-ideal associated to the
path of length m − 2 with vertices 1, . . . , m − 1. Hence height P Sx ≥ 3 + m − 2 = m + 1 as
(3)
desired. Finally, if f1,m−1 ∈ P Sx we have that P Sx contains the ideal LkCm−1 (3) associated to
the (m − 1)-cycle with vertices 1, . . . , m − 1 and we have already observed that this ideal is a
complete intersection. Since ym2 − x−1 ym1 y12 , ym3 − x−1 ym1 y13 are in P Sx as well it follows that
height P Sx ≥ 2 + m − 1 = m + 1.
Proof of Theorem 2.9 part (2). By Lemma 1.1 conditions (b) and (c) in Theorem 2.9 (2) are equivalent. Hence it suffices to prove the equivalence of (a) and (b). We prove first that (a) implies (b).
By Proposition 2.8 that (a) implies that G does not contain Ka,b with a + b = d + 1. Suppose then
by contradiction that G does contain C2m . Hence LkC2m (2) is a complete intersection of height
2m. But the generators of LkC2m (2) are (up to sign) among the 2-minors of the matrix:
y11 −y22 y31 . . . y2m−1,1 −y2m,2
y12 y21 y32 . . . y2m−1,2 y2m,1
and the ideal of 2-minors of such a matrix has height 2m − 1, a contradiction.
Now we prove that (b) implies (a). We may assume that k is algebraically closed. Since the
tensor product over a perfect field k of reduced k-algebras is reduced [6, Thm 3, Chapter V, 15],
we may assume that G is connected. A connected graph satisfying (b) is either an isolated vertex,
or a path or an odd cycle. We have already observed that pmd(Pn ) ≤ 2. By Theorem 2.5 it
LSS-IDEALS AND COORDINATE SECTIONS
13
follows that LkPn (2) is a complete intersection. It remains to prove that LkC2m+1 (2) is a complete
intersection (of height 2m + 1). Note that LkP2m+1 (2) ⊂ LkC2m+1 (2) and we know already that
(2)
LkP2m+1 (2) is a complete intersection of height 2m. Hence it remains to prove that f1,2m+1 does
not belong to any minimal prime of LkP2n+1 (2). The generators of LkP2n+1 (2) are (up to sign) the
adjacent 2-minors of the matrix:
Y =
y11
y12
−y22
y21
y31
y32
...
...
y2m−1,1
y2m−1,2
−y2m,2
y2m,1
y2m+1,1
y2m+1,2
The minimal primes of LkP2n+1 (2) are described in the proof of [15, Thm.4.3], see also [23] and
[21]. By the description given in [15] it is easy to see that all minimal primes of LkP2n+1 (2) with
the exception of I2 (Y ) are contained in the ideal Q = (yij : 2 < i < 2m + 1 1 ≤ j ≤ 2). Clearly
(2)
(2)
f1,2m+1 6∈ Q. Finally one has f1,2m+1 6∈ I2 (Y ) since the monomial y11 y2m+1,1 is divisible by no
monomials in the support of the generators of I2 (Y ).
We proceed with the proof of Theorem 2.10. For the proof we first formulate a result that is a
special case of a more general statement. For this we need to introduce the concept of SturmfelsCartwright ideals. This concept was coined in [9] inspired by earlier work in [8] and [7]. It was
further developed and applied to various classes of ideals in [10] and [11].
Consider for d1 , . . . , dn ≥ 1 the polynomial ring S = k[yij : i ∈ [n], j ∈ [di ]] with multigrading
degyij = ei ∈ Zn . The group G = GLd1 (k) × · · · × GLdn (k) acts naturally on S as the group of
Zn -graded K-algebra automorphism. The Borel subgroup of G is B = Ud1 (k) × · · · × Udn (k) where
Ud (k) denotes the subgroup of GLd (k) of the upper triangular matrices. A Zn -graded ideal J is
Borel fixed if g(J) = J for every g ∈ B. A Zn -graded ideal I of S is called a Cartwright-Sturmfels
ideal if there exists a radical Borel fixed ideal J with the same multigraded Hilbert-series.
Theorem 4.1. For d1 , . . . , dn ≥ 1 let S = k[yij : i ∈ [n], j ∈ [di ]] be the polynomial ring with
Zn multigrading induced by degyij = ei ∈ Zn . and G = (V, E) be a forest (i.e. a graph without
cycles). For each e = {i, j} ∈ E let fe ∈ S be a Zn -graded polynomial of degree ei + ej . Then
I = (fe : e ∈ E) is a Cartwright-Sturmfels ideal. In particular, I and all its initial ideals are
radical.
Proof. First, we observe that we may assume that the generators fe of I form a regular sequence.
To this end we introduce new variables and add to each fe a monomial me in the new variables of
degree e so that me and me0 are coprime if e 6= e0 . The new polynomials fe + me with e ∈ E form
a regular sequence by Proposition 1.4 since their initial terms with respect to an appropriate term
order are the pairwise coprime monomials me . The ideal I arises as a multigraded linear section
of the ideal (fe + me : e ∈ E) by setting all new variables to 0. By [8, Thm. 1.16(5)] the family of
Cartwright-Sturmfels ideals is closed under any multigraded linear section. Hence it is enough to
prove the statement for the ideal (fe + me : e ∈ E). Equivalently we may assume right away that
the generators fe of I form a regular sequences.
The multigraded Hilbert series of a multigraded S-module M can by written as
KM (z1 , . . . , zn )
Qn
.
di
i=1 (1 − zi )
The numerator KM (z1 , . . . , zn ) is a Laurent polynomial polynomial with integral coefficients called
the K-polynomial of M . Since the fe ’s form a regular sequence the K-polynomial of S/I is the
polynomial:
Y
F (z) = F (z1 , . . . , zn ) =
(1 − zi zj ) ∈ Q[z1 , . . . , zn ].
{i,j}∈E
To prove that I is Cartwright-Sturmfels we have to prove that there is a Borel-fixed radical
ideal J such that the K-polynomial of S/J is F (z). Taking into consideration the duality between
Cartwright-Sturmfels ideals and Cartwright-Sturmfels∗ ideals discussed in [9], it is enough to
exhibit a monomial ideal J whose generators are in the polynomial ring S 0 = k[y1 , y2 , . . . , yn ]
14
A. CONCA AND V. WELKER
equipped with the (fine) Zn -grading deg yi = ei ∈ Zn such that the K-polynomial of J regarded
as an S 0 -module is F (1 − z1 , . . . , 1 − zn ), that is,
Y
(zi + zj − zi zj ).
{i,j}∈E
We claim that, under the assumption that ([n], E) is a forest, the ideal
Y
J=
(yi , yj )
{i,j}∈E
has the desired property. In other words, we have to prove that the tensor product
O
TE =
T{i,j}
{i,j}∈E
of the truncated Koszul complexes:
T{i,j} : 0 → S 0 (−ei − ej ) → S 0 (−ei ) ⊕ S 0 (−ej ) → 0
associated to yi , yj resolves the ideal J. Consider a leaf {a, b} of E. Set E 0 = E \ {{a, b}},
Y
J0 =
(yi , yj )
{i,j}∈E 0
and J 00 = (ya , yb ). Then by induction on the number of edges we have that TE 0 resolves the
0
ideal J 0 . Then the homology of TE is TorS∗ (J 0 , J 00 ). Since since {a, b} is a leaf one of the two
variables ya , yb does not appear at all in the generators of J 0 . Hence ya , yb forms a regular J 0 0
sequence. Then TorS≥1 (J 0 , J 00 ) = 0 and hence TE resolves J 0 ⊗ J 00 . Finally, J 0 ⊗ J 00 = J 0 J 00
0
since TorS1 (J 0 , S/J 00 ) = 0. This concludes the proof that the ideal I is a Cartwright-Sturmfels
ideal. Every initial ideal of a Cartwright-Sturmfels ideal is a Cartwright-Sturmfels ideal as well
because the property just depends on the Hilbert series. In particular, every initial ideal of a
Cartwright-Sturmfels ideal is radical.
Now we are ready to prove Theorem 2.10:
(d)
Proof of Theorem 2.10.
(1) Setting d1 = · · · = dn = d and fe = fe in Theorem 4.1 we have
that LkG (d) is a Cartwright-Sturmfels ideal and hence radical.
(2) If LkG (d) is a complete intersection then, by Proposition 2.8, G does not contain a copy
of K1,d+1 as a subgraph, that is, ∆(G) ≤ d. Vice versa if d ≥ ∆(G) one proves that
LkG (d) is a complete intersection by using induction on the number of vertices of G and
the symmetric algebra point of view. As G is a forest we may assume that {n − 1, n} is a
leaf. Then G1 = G − n is a forest with ∆(G1 ) ≤ ∆(G) ≤ d. Hence by induction LkG1 (d) is
a complete intersection. Set
R = k[yi,j : i ∈ [n − 1], j ∈ [d]]/LkG1 (d).
We may interpret LkG (d) as the ideal defining the symmetric algebra of the R-module
defined as the cokernel of the R-linear map R → Rd associated to the matrix A =
(yn−1,1 , . . . , yn−1,d ). Hence, by virtue of Theorem 3.1(1), it is enough to prove that
height (yn−1,1 , . . . , yn−1,d ) ≥ 1 in R
equivalently,
height (yn−1,1 , . . . , yn−1,d ) + LkG1 (d) > height LkG1 (d)
which is true because at most ∆(G) − 1 of the generators of LkG1 (d) are contained in
(yn−1,1 , . . . , yn−1,d ).
(3) If LkG (d) is prime then, by Proposition 2.8, G does not contain a copy of K1,d as a subgraph,
that is, ∆(G) ≤ d − 1. Vice versa, we know by (2) that LkG (d) is a complete intersection
for d ≥ ∆(G). Hence by Theorem 2.3 LkG (d) is prime for d ≥ ∆(G) + 1.
LSS-IDEALS AND COORDINATE SECTIONS
15
Hence for a forest G we have a complete picture of asymptotic behaviour:
Corollary 4.2. Let G be a forest and k any field. Then: asymk (radical, G) = 1, asymk (c.i., G) =
∆(G) and asymk (prime, G) = ∆(G) + 1.
5. Positive matching decompositions
In this section we introduce the concepts of positive matchings and positive matching decomposition and prove Theorem 2.5.
Definition 5.1. Given a hypergraph H = (V, E) a positive matching of G is a subset M ⊂ E
of pairwise disjoint sets (i.e., a matching) such that there exists a weight function w : V → R
satisfying:
X
w(i) > 0 if
A∈M
i∈A
(4)
X
w(i) < 0 if
A ∈ E \ M.
i∈A
We illustrate the definition for subgraphs of K4 . The edge set M = {{1, 2}, {3, 4}} is a matching
in K4 but it is not positive. Nevertheless, the matching M is positive in the graph with edge set
{{1, 2}, {1, 3}, {2, 3}, {3, 4}} with respect to the weight function w(1) = 0, w(2) = 1, w(3) = −2
and w(4) = 3.
The next lemma summarizes some elementary properties of positive matchings.
S
Lemma 5.2. Let H = (V, E) be a hypergraph such that E is a clutter, M ⊆ E and VM = A∈M A.
(i) M is a positive matching for H if and only if M is a positive matching for the induced
hypergraph HVM = (VM , {A ∈ E | A ⊆ VM }).
(ii) If M is a positive matching on H and A ∈ E is such that M ∪ {A} is a matching and
there is no B ∈ M with B ∩ VM 6= ∅ =
6 B ∩ A then M ∪ {A} is a positive matching on H.
(iii) If H is a bipartite graph with bipartition V = V1 ∪ V2 then M is a positive matching if and
only if M is a matching and directing the edges e ∈ E from V1 to V2 if e ∈ M and from
V2 to V1 if e ∈ E \ M yields an acyclic orientation.
Proof.
(i) Clearly a weight function on V for which M is a positive matching restricts to VM
making M a positive matching on HVM . Conversely, setting w(i) = − max{|A| : A ∈
M } · max{w(j) : j ∈ VM } for i ∈ V \ VM extends a weight function on VM to V and
shows that M is a positive matching on H.
(ii) Let w we a weight function for which M is positive. Set T = max{1, w(`) : ` ∈ VM } > 0.
We define w0 : V → R as follow. If ` ∈ VM we set w0 (`) = w(`). If ` ∈ A we set w0 (`) = T
and for ` ∈ V \ (VM ∩ A) we set w0 (`) = −| max{|A|, |B| : B ∈ M } · T − 1. One easily
checks that M ∪ {A} is a positive matching with respect to w0 .
(iii) We change the coordinates w(i) to −w(i) for i ∈ V2 in the inequalities defining a positive
matchings. As a simple reformulation of (4) we get that in these coordinates a matching
M is positive if and only if there is a weight function such that for {i, j} ∈ E, i ∈ V1 ,
j ∈ V2 we have
(5)
w(i) > w(j) if
{i, j} ∈ M,
w(i) < w(j) if
{i.j} ∈ E \ M.
This is equivalent to the existence of a region in the arrangement of hyperplanes w(i) =
w(j) for {i, j} ∈ E in RV satisfying (5). But it is well known that the regions in this
arrangement are in one to one correspondence with the acyclic orientations of G (see [20,
Lemma 7.1]).
Now we are in position to introduce the key concept of this section.
16
A. CONCA AND V. WELKER
Definition 5.3. Let H = (V, E) be a hypergraph for whichSE is a clutter. A positive matching
p
decomposition (or pm-decomposition) of G is a partition E = i=1 Ei into pairwise disjoint subsets
such that Ei is a positive matching on (V, E \ ∪i−1
j=1 Ej ) for i = 1, . . . , p. The Ei are called the parts
of the pm-decomposition. The smallest p for which G admits a pm-decomposition with p parts
will be denoted by pmd(H).
S Note that by definition one has pmd(H) ≤ |E| because of the obvious pm-decomposition
A∈E {A}. On the other hand pmd(G) is smaller than |E| for most clutters. Next we prove
the bound from Theorem 2.5(4).
Proof of Theorem 2.5(4). First, consider the case when G = (V, E) is an arbitrary graph. Set
n = |V |. We may assume that G is the complete graph Kn because any pm-decomposition of Kn
induces a pm-decomposition on its subgraphs. For ` = 1, . . . , 2n − 3 we set E` = {{i, j} : i + j =
` + 2}. For example, for n = 7 we have
E1 = {12},
E2 = {13},
E3 = {14, 23},
E4 = {15, 24},
E5 = {16, 25, 34}, E6 = {17, 26, 35}, E7 = {27, 36, 45}, E8 = {37, 46},
E9 = {47, 56},
E10 = {57},
E11 = {67}
where for simplicity we have written ij for the edge {i, j}.
Clearly one has E = ∪2n−3
`=1 E` and this is a pm-decomposition of Kn since when a new edge
is inserted the smallest index involved in that edge satisfies the condition of Lemma 5.2 (ii) with
respect to the current data. For example for n = 7 when we insert the edge 27 in E7 the vertex
2 satisfies the condition of Lemma 5.2 (ii) with respect to the matching {36, 45}, i.e. the edges
23, 24, 25, 26 are already used in an earlier step of the construction.
Now consider the case when G = (V, E) is a bipartite graph. Let V = V1 ∪ V2 be a the
bipartition for G. We may assume V1 = [m] and V2 = [ñ] =
Sp{1̃, . . . , ñ} for numbers m, n ≥ 1
and we may assume that G = Km,n . We show that E = `=1 E` with p = m + n − 1 and
E` = {{i, j̃} : i + j = ` + 1} is positive matching decomposition of Km,n . That is, we show that
S`−1
E` is a positive matching on E \ k=1 Ek for ` = 1, . . . , m + n − 1.
For ` = 1 the assertion is obvious since E1 contains a single edge. Now assume ` ≥ 2. By
Lemma 5.2(iii) it suffices to show that directing the edges in E` from [m] to [ñ] and the edges in
S`
E \ k=1 Ek in the other direction yields an acyclic orientation. Assume the resulting directed
graph has a cycle. As a cycle in a bipartite graph is of even length this cycle must contain at least
two edges of type i → j̃ for {i, j̃} ∈ E` or equivalently i + j = ` + 1. We assume {i, j̃} ∈ E` is
chosen with this property so that j is maximal. The next edge in the directed cycle is an edge
j̃ → i0 . We must have i0 < i or i0 > i. If i0 < i then the following edge in the cycle i0 → j˜0 must
again satisfy i0 + j 0 = ` + 1. But by i0 < i and i + j = ` + 1 it follows that j 0 > j a contradiction.
Analogously, consider the edge j˜00 → i preceding i → j̃. Again by construction j 00 + i > ` + 1.
But then j 00 > j again a contradiction. Hence there is no cycle and E` is a positive matching on
S`−1
E \ k=1 Ek .
In the bipartite case, Theorem 2.6(4) shows that pmd(Km,n ) = m + n − 1 and computer
experiments show that pmd(Kn ) = 2n − 3 holds for small value of n.
Next we connect positive matching decompositions to algebraic properties of LSS-ideals.
Lemma
Sp 5.4. Let H = (V, E) be a hypergraph such that E is a clutter, d ≥ p := pmd(H) and
E = `=1 E` a positive matching decomposition. Then there exists a term order < on S such that
for every ` and every A ∈ E` we have
Y
(d)
(6)
in< (fA ) =
yi` .
i∈A
Proof. To define < we first define weight vectors w1 , . . . , wp ∈ RV ×[d] . For that purpose we use
the weight functions w` : V → R, associated to each matching E` , ` = 1, . . . , p. The weight vector
w` is defined as follows:
• w` (yik ) = 0 if k 6= ` and
LSS-IDEALS AND COORDINATE SECTIONS
17
• w` (yi` ) = w` (i).
Q
P By definition the weight of the monomial i∈A yik withQrespect to w` is 0 if k 6= ` and
i∈A w` (i) if k = `. Hence, by definition, the weight of
i∈A yi` is positive if A ∈ E` and
negative if A ∈ E \ ∪`v=1 Ev .
It follows that:
(7)
(d)
inw1 (fA )
=
Q
di∈A yi1
XY
yik
if A ∈ E1
if A ∈ E \ {E1 }.
k=2 i∈A
We define the term order < as follows: y α < y β if
(1) |α| < |β| or
(2) |α| = |β| and w` (y α ) < w` (y β ) for the smallest ` such that w` (y α ) 6= w` (y β ) or
(3) |α| = |β| and w` (y α ) = w` (y β ) for all ` and y α <0 y β for an arbitrary but fixed term order
<0 .
(d)
Q A simple induction using (7) now shows that for all ` and for all A ∈ E we have in< (fA ) =
i∈A yi` .
We conclude this section with the proofs of Theorem 2.5(1)-(3) and Theorem 2.6(4) and a
simple corollary.
Sp
Proof of Theorem 2.5. Let p = pmd(H) and E = `=1 E` a pm-decomposition of H.
Hence by Lemma 5.4 there is a term order < satisfying (6). Since each E` , ` = 1, . . . , p, is
(d)
a matching (6) implies that the initial monomials of the generators fA , A ∈ E, of LH (d) are
(d)
pairwise coprime. Hence in< (LkH (d)) = (in< (fe : e ∈ E) is a radical complete intersection. This
prove (1). Then (2) follows from (1) and Proposition 1.4 and (3) follows from Theorem 2.3.
Now we can also complete the proof of Theorem 2.6(4).
Proof of Theorem 2.6(4). From Theorem 2.5(4) we know that pmd(Km,n ) ≤ m + n − 1. From
Theorem 2.5(3) we know that LkKm,n (d) is prime for d ≥ pmd(Km,n ) + 1. From Theorem 2.9(3)
we know that if LkKm,n (d) is prime then Km,n does not contain a subgraph Ka,b with a + b = d + 1.
The latter then implies d ≥ n + m. Hence pmd(Km,n ) ≥ m + n − 1 and therefore pmd(Km,n ) =
m + n − 1.
Using the fact Theorem 2.3(6) that primeness is inherited by subgraphs the following is an
immediate consequence of Theorem 2.5.
Corollary 5.5.
(i) Let G be a subgraph of Kn then LK
G (d) is a radical complete intersection
for d ≥ min{2n − 3, |E|} and prime for d ≥ min{2n − 3, |E|} + 1.
(ii) Let G be a subgraph of Km,n then LK
G (d) is a radical complete intersection for d ≥ min{m+
n − 1, |E|} and prime for d ≥ min{m + n − 1, |E|} + 1.
6. Determinantal rings from the point of view of invariant theory
The goal of this section is to recall some classical results in invariant theory (see for example
[34]) that have been treated in modern terms by De Concini and Procesi in [14]. In particular,
we recall how determinantal/Pfaffian rings arise as invariant ring of group actions. We assume
throughout that the base field k is of characteristic 0.
6.1. Generic determinantal rings as rings of invariants (gen). We take an n × m matrix
gen
k
gen
of variables Xm,n
= (xij )(i,j)∈[m]×[n] and consider the ideal Id+1
(Xm,n
) of S gen = k[xij : (i, j) ∈
gen
[m] × [n]] generated by the (d + 1)-minors of Xm,n . Consider two matrices of variables Y and Z of
size m × d and d × n and the following action of G = GLd (k) on the polynomial ring k[Y, Z]: The
matrix A ∈ G acts by the k-algebra automorphism of k[Y, Z] that sends Y → Y A and Z → A−1 Z.
The entries of the product matrix Y Z are clearly invariant under this action. Hence the ring of
18
A. CONCA AND V. WELKER
invariants k[Y, Z]G contains the subalgebra k[Y Z] generated by the entries of the product Y Z.
The First Main Theorem of Invariant Theory for this action says that k[Y, Z]G = k[Y Z]. We have
a surjective k-algebra map:
φ : S gen → k[Y, Z]G = k[Y Z]
Pd
sending X to Y Z, that is xij to (Y Z)ij = `=1 yi ` z` j . Clearly the product matrix Y Z has
k
gen
rank d and hence we have Id+1
(Xm,n
) ⊆ Ker φ. The Second Main Theorem of Invariant Theory
k
gen
says that Id+1 (Xm,n ) = Ker φ. Hence
k
gen
S/Id+1
(Xm,n
) ' k[Y Z]
(8)
6.2. Generic symmetric determinantal rings as rings of invariants (sym). We take an
n × n symmetric matrix of variables
Xnsym
x11
x12
x13
..
.
..
.
..
.
=
x1n−1
x1n
x12
x22
x23
..
.
..
.
..
.
x13
x23
x33
..
.
..
.
..
.
···
···
···
..
.
..
.
..
.
x1 n−1
x2 n−1
x3 n−1
..
.
..
.
..
.
x1n
x2n
x3n
..
.
..
.
..
.
x2n−1
x2n
x3n−1
x3n
···
···
xn−1 n−1
xn−1 n
xn−1 n
xnn
k
and consider the ideal Id+1
(Xnsym ) generated by the (d + 1)-minors of Xnsym in the polynomial
sym
ring S
= k[xij : 1 ≤ i ≤ j ≤ n]. Consider a matrix of variables Y of size n×d and the following
action of the orthogonal group G = Od (k) = {A ∈ GLt (k) : A−1 = AT } on the polynomial ring
k[Y ] = k[yij |Yij : (i, j) ∈ [n] × [t]]: Any A ∈ G acts by the k-algebra automorphism of k[Y ] that
sends Y to Y A. The entries of the product matrix Y Y T are invariant under this action and hence
the ring of invariants contains the subalgebra k[Y Y T ] generated by the entries of Y Y T . The First
Main Theorem of Invariant Theory for this action asserts that k[Y ]G = k[Y Y T ]. Then we have a
surjective presentation:
φ : S sym → k[Y Y T ]
sending X to Y Y T . Since the product matrix Y Y T has rank t we have Id+1 (X) ⊆ Ker φ. The
Second Main Theorem of Invariant Theory then says that Id+1 (X) = Ker φ. Hence
k
S sym /Id+1
(Xnsym ) ' k[Y Y T ].
(9)
6.3. Generic Pfaffian rings as rings of invariants (skew). We take an n × n skew-symmetric
matrix of variables
Xnskew
0
−x12
−x13
..
.
..
.
..
.
=
−x1n−1
−x1n
x12
0
−x23
..
.
..
.
..
.
x13
x23
0
..
.
..
.
..
.
···
···
···
..
.
..
.
..
.
x1 n−1
x2 n−1
x3 n−1
..
.
..
.
..
.
x1n
x2n
x3n
..
.
..
.
..
.
−x2n−1
−x2n
−x3n−1
−x3n
···
···
0
−xn−1 n
−xn−1 n
0
LSS-IDEALS AND COORDINATE SECTIONS
19
and consider the ideal Pf k2t+2 (X) generated by the Pfaffians of size (2d + 2) in k[xij | 1 ≤ i <
j ≤ n] of Xnskew in the polynomial ring S skew = k[xij : 1 ≤ i ≤ j ≤ n]. Consider a matrix of
variables Y of size n × 2d and let J2t be the 2d × 2d block matrix with d blocks
0
−1
1
0
on the diagonal and 0 in the other positions. The sympletic group G = Sp2d (k) = {A ∈ GL2t (k) :
AJAT = J} acts on the polynomial ring k[Y ] = k[yij ] as follows: an A ∈ G acts on k[Y ] by
the automorphism that sends Y → Y A. The entries of the product matrix Y J2d Y T are invariant
under this action and hence the ring of invariants contains the subalgebra k[Y J2t Y T ] generated by
the entries of Y J2d Y T . The First Main Theorem of Invariant Theory for the current action says
that k[Y ]G = k[Y J2t Y T ]. Then we have a surjective presentation: φ : S skew → k[Y Y T ] sending X
to Y JY T . The product matrix Y JY T has rank 2d and hence we have Pf k2d+2 (X) ⊆ Ker φ. The
Second Main Theorem of Invariant Theory for this action says that Pf k2d+2 (X) = Ker φ. Hence
(10)
S skew / Pf k2d+2 (Xnskew ) ' k[Y JY T ].
7. Determinantal ideals of matrices with 0’s and their relation to LSS-ideals
The classical invariant theory point of view described in Section 6 shows that the generic
determinantal and Pfaffian ideals are prime as they are kernels of ring maps whose codomains
are integral domains. Their height is also well know (see for example [3] and the references given
there):
gen
) of d-minors of a m × n matrix of variables is (n + 1 −
(gen) The height of the ideal Idk (Xm,n
d)(m + 1 − d).
(sym) The height of the ideal Idk (Xnsym ) of d-minors of a symmetric n × n matrix of variables is
n−d+2
.
2
(skew) The height of the ideal of Pfaffians Pf k2d (Xnskew
) of size 2d (and degree d) of an n × n
skew-symmetric matrix of variables is n−2d+2
.
2
If one replaces the entries of the matrices with general linear forms in, say, u variables, then
Bertini’s theorem in combination with the fact that the generic determinantal/Pfaffian rings are
Cohen-Macaulay implies that the determinantal/Pfaffian ideals remain prime as long as u ≥
2+height and radical if u ≥ 1+height.
But what about the case of special linear sections of determinantal ideals of matrices? And
what about the case of coordinate sections? Are the corresponding ideals prime or radical? To
describe coordinate sections we employ the following notation.
(gen) In the generic case we take – as described in the introduction – a bipartite graph G =
gen
([m] ∪ [ñ], E) where E ⊆ [m] × [ñ] and denote by XG
the matrix obtained from the m × n
matrix of variables X = (xij )(i,j)∈[m]×[n] by replacing the entries in position (i, j) with 0
for {i, j̃} ∈ E.
(sym) In the generic symmetric case we take a subgraph G = ([n], E) of Kn and denote by
sym
XG
the matrix obtained from the generic symmetric matrix Xnsym by replacing for all
{i, j} ∈ E the entries in row i, column j and row j, column i with 0.
(skew) In the generic skewsymmetric case where we take a subgraph G = ([n], E) of Kn and
skew
denote by XG
the matrix obtained from the generic skewsymmetric matrix Xnskew by
replacing for all {i, j} ∈ E the entries in row i, column j and row j, column i with 0.
gen
sym
gen
sym
In this terminology Idk (XG
) (resp. Idk (XG
)) is the ideal of d-minors of XG
resp. XG
) in
k
gen
sym
skew
skew
skew
S
(resp. S
). We write Pf 2d (XG ) for the ideal of Pfaffians of size 2d of XG
in S
.
gen
sym
skew
We ask for conditions on G that imply that Idk (XG
) (resp. Idk (XG
) or Pf k2d (XG
)) is radical
or prime.
20
A. CONCA AND V. WELKER
Simple examples show that special linear sections of relatively small height of generic determinantal ideals can give non-prime and non-radical ideals. On the positive side, for maximal minors,
we have the following results:
Remark 7.1.
(1) Eisenbud [17] proved that the ideal of maximal minors of a 1-generic m × n
matrix of linear forms is prime and remains prime even after modding out any set of
≤ m − 2 linear forms. In particular, the ideal of maximal minors of an m × n matrix of
linear forms is prime provided the ideal generated by the entries of the matrix has at least
m(n − 1) + 2 generators.
(2) Giusti and Merle in [18] studied the ideal of maximal minors of coordinate sections in
the generic case. One of their main results, [18, Thm.1.6.1] characterizes, in combinatorial terms, the subgraph graphs G of Km,n , m ≤ n, such that the variety associated to
gen
gen
k
k
Im
(XG
) is irreducible, i.e. the radical of Im
(XG
) is prime.
gen
k
(3) Boocher proved in [5] that for any subgraph G of Km,n , m ≤ n, the ideal Im
(XG
) is
radical. Combining his result with the result of Giusti and Merle, one obtains a characgen
k
) is prime.
terization of the graphs G such that Im
(XG
(4) Generalizing the result of Boocher, in [8] and [9] it is proved that ideals of maximal minors
of a matrix of linear forms that is either row or column multigraded is radical.
gen
In the generic case every non-zero minor of a matrix of type XG
has no multiple factors
gen
might
because its multidegree is square-free. This suggests that the determinantal ideals of XG
always be radical. The following example shows that this is not the case:
gen
Example 7.2. Let XG
be the 6 × 6 matrix associated to the graph from Example 2.2(3). That
is, in the 6 × 6 generic matrix we set to 0 the entries in positions
(1, 1), (1, 2), (1, 3), (1, 4), (2, 1), (2, 2), (3, 2), (3, 3), (4, 3), (4, 4), (5, 1), (5, 4).
gen
I4k (XG
)
Then
is not radical over a field of characteristic 0 and very likely over any field. Here
gen
gen
the “witness” is g = x1,5 , i.e. I4k (XG
) : g 6= I4k (XG
) : g 2 . Since G is contained in K5,4 one can
gen
k
consider as well I4 (XG ) in the 5 × 5 matrix but that ideal turns out to be radical.
Similarly for symmetric matrices we have:
sym
Example 7.3. Let XG
be the 7 × 7 generic symmetric matrix associated to the graph from
Example 2.2(1). That is, in the 7 × 7 generic symmetric matrix we set to 0 the entries in positions
{1, 2}, {1, 3}, {1, 4}, {1, 5}, {2, 3}, {2, 4}, {2, 6}, {3, 5}, {4, 6}
sym
as well as in the symmetric positions. Then I4k (XG
) is not radical over a field of characteristic
sym
0 The witness here is g = x1,6 . Since G is contained in K6 one can consider as well I4k (XG
) in
the 6 × 6 matrix but that ideal turns out to be radical.
The examples Example 2.2, Example 7.2 and Example 7.3 are indeed closely related as we now
explain.
Let G = ([m] ∪ [ñ], E) be a subgraph of the complete bipartite graph Km,n . In view of the
isomorphism (8) we have that
gen
k
S gen / Id+1
(XG
) + (xij | {i, j̃} ∈ E) ' k[Y Z]/JG (d)
where Y = (yij ), Z = (zij ) are respectively m × d and d × n matrices of variables and JG (d)
is the ideal of k[Y Z] generated by (Y Z)i,j with {i, j̃} ∈ E. The LSS-ideal LkG (d) ⊂ k[Y, Z] is
indeed equal to JG (d)k[Y, Z]. Now it is a classical result in invariant theory, (derived from the
fact that linear groups are reductive in characteristic 0) that k[Y Z] is a direct summand of k[Y, Z]
in characteristic 0. This implies that
JG (d)
= LkG (d) ∩ k[Y Z].
The next proposition is an immediate consequence.
LSS-IDEALS AND COORDINATE SECTIONS
21
Proposition 7.4. Let k be a field of characteristic 0, d ≥ 1 and G = ([m] ∪ [ñ], E) be a subgraph
of Km,n . If the LSS-ideal LkG (d) is radical (resp. is a complete intersection, resp. is prime) then
gen
k
the coordinate section Id+1
(XG
) of the generic determinantal ideal is radical (resp. has maximal
height, resp. is prime).
Now we start from a subgraph E ⊆ {{i, j} : 1 ≤ i < j ≤ n} of Kn . For d + 1 ≤ n we may
sym
k
k
consider the coordinate section Id+1
(XG
) of Id+1
(Xnsym ). Using the isomorphism (9) we obtain:
Proposition 7.5. Let k be a field of characteristic 0 and G = ([n], E) a subgraph of Kn . If the
LSS-ideal LkG (d) is radical (resp. is a complete intersection, resp. is prime) then the coordinate
sym
k
section Id+1
(XG
) of the generic determinantal ideal is radical (resp. has maximal height, resp.
is prime).
Now we can go back to LSS-ideals.
Proposition 7.6. Let
k be a field of characteristic 0 and n ∈ N. Let wn be the largest positive
integer such that w2n ≤ n. Then:
(a) LkKn (d) is not prime for d = n + wn2−2 − 1.
−2
(b) LkKn (d) is not a complete intersection for d = n + wn+1
− 2.
2
Proof.
(a) We set hn = w2n and mn = wn + d − 1. The numbers are chosen so that, using the
formulas mentioned at the beginning of the section, the ideal Id+1 (X) of (d + 1)-minors
of a generic
symmetric mn × mn matrix X has height hn . Consider Kn as the graph
([mn ], [n]
2 ) on mn vertices where the vertices n + 1, . . . , mn do not appear in an edge. If,
sym
k
by contradiction, the ideal LkKn (d) is prime then by Proposition 7.5 the ideal Id+1
(XK
)
n
is prime and of height hn . But one has
sym
k
Id+1
(XK
) ⊂ (x11 , x22 , . . . , xhn hn )
n
(11)
which is a contradiction. To check (11) it is enough to prove that the rank of the matrix
sym
XK
n
mod (x11 , x22 , . . . , xhn hn )
is at most d. That is, we have to check that the rank of an (mn × mn )-matrix with block
decomposition
0 A
B C
where 0 is the zero matrix of size (hn × n), is at most d. Since d = mn − n + mn − hn
the latter is obvious.
(b) We set hn = wn+1
and mn = wn+1 + d − 1. As above, the numbers are chosen so that
2
the ideal Id+1 (X) of (d + 1)-minors of a generic symmetric mn × mn matrix X has height
hn .
Assume, by contradiction, that LkKn (d) is a complete intersection. From Proposition 7.5
sym
k
it follows that Id+1
(XK
) has height hn . But
n
sym
k
Id+1
(XK
) ⊂ (x11 , x22 , . . . , xhn −1,hn −1 )
n
(12)
which is a contradiction. As above (12) boils down to an obvious statement about the
rank of a matrix with a zero submatrix of a certain size.
Using this result we can now analyze the asymptotic behavior of both asymk (c.i., Kn ) and
asymk (prime, Kn ).
Corollary 7.7. Let k be a field of characteristic 0. Then
(13)
asymk (prime, Kn )
asymk (c.i., Kn )
= lim
= 2.
n→∞
n→∞
n
n
lim
22
A. CONCA AND V. WELKER
Proof. By Corollary 5.5(i) we have asymk (prime, Kn ) ≤ 2n − 2. By Proposition 7.6 we have
wn+1 − 2
(14)
n+
− 1 ≤ asymk (c.i., Kn ) ≤ asymk (prime, Kn )
2
Hence the equalities in (13) follow from the fact that
wn+1 −2
lim
n→∞
2
n
= 1.
Using Proposition 7.6 and Theorem 2.3 we obtain further obstructions.
Corollary 7.8. Let G be a graph on n vertices and k a
field of characteristic 0. Then LkG (d)
wω(G)+1
− 2 and LkG (d) is not prime for d ≤
is not a complete intersection for d ≤ ω(G) +
2
wω(G)
− 1 where wω(G) is defined as in Proposition 7.6.
ω(G) +
2
To get an actual feeling of the obstruction. We list a few explicit example of new obstructions
derived from Corollary 7.8.
d
d≤2
d≤3
d≤4
d≤5
d≤6
d≤7
d≤8
d ≤ 10
d ≤ 11
d ≤ 12
obstruction to complete intersection
K4
obstruction to primeness
K3
K4
K5
K6
K7
K8
K5
K6
K7
K8
K9
K11
K12
K12
K13
skew
) of Pf k2d+2 (Xnskew ). We
For 2d + 2 ≤ n we may consider the coordinate section Pf k2d+2 (XG
may as well consider the same graph G = ([n], E) and the associated twisted LSS-ideal L̂kG (d)
defined as follows. For every i ∈ [n] we consider 2d indeterminates yi 1 , . . . , yi 2d . For e = {i, j},
(d)
1 ≤ i < j ≤ n we set fˆe to be the entry of the matrix Y JY T in row i and column j, i.e.
fˆe(d) =
d
X
yi 2k−1 yj 2k − yi 2k yj 2k−1 .
k=1
Then we call
L̂kG (d) = (fˆe(d) : e ∈ E).
the twisted LSS-ideal associated to G. For d = 1 the twisted LSS-ideal coincides with the
so-called binomial edge ideal defined and studied in [21, 25, 29, 32].
√
Remark 7.9. Assume −1 ∈ k and G is bipartite with bipartition [n] = V1 ∪ V2 then the
coordinate transformation
√
• yi 2k−1 7→
√ yi 2k−1 and yi 2k 7→ −1yi 2k−1 for i ∈ V1 ,
• yj 2k 7→ −1 · yi 2k−1 and yj 2k−1 7→ yi 2k for j ∈ V2 ,
sends L̂kG (d) to LkG (2d). In particular, for a bipartite graph G we have that L̂kG (d) is radical
(resp. prime) if and only if LkG (2d) radical (resp. prime).
Using the isomorphism (10) we obtain:
Proposition 7.10. Let k be a field of characteristic 0 and G = ([n], E) a subgraph of Kn . If
the twisted LSS-ideal L̂kG (d) is radical (resp. is c.i., resp. is prime) then the coordinate section
skew
Pf k2d+2 (XG
) of the generic Pfaffian ideal is radical (resp. has maximal height, resp. is prime).
LSS-IDEALS AND COORDINATE SECTIONS
23
We have:
Theorem 7.11. Let k be a field of characteristic 0.
gen
gen
(gen) If G is a subgraph of Km,n then the ideals I2k (XG
) and I3k (XG
) are radical.
sym
sym
(sym) If G is a subgraph of Kn then the ideals I2k (XG
) and I3k (XG
) are radical.
skew
(skew) If G is a subgraph of Kn then the ideal Pf k4 (XG
) is radical.
Furthermore if G is a forest then
gen
sym
skew
(1) Idk (XG
), Idk (XG
) and Pf k2d (XG
) are radical for all d.
gen
sym
k
k
(2) Id (XG ) and Id (XG ) have maximal height if d ≥ ∆(G) + 1.
gen
sym
(3) Idk (XG
) and Idk (XG
) are prime if d ≥ ∆(G) + 2.
Proof. The statements for ideals of 2-minors in the cases (gen) and (sym) follow from Proposition 7.4 and Proposition 7.5 using the fact that the edge ideal of a graph is radical. Indeed these
results hold over a field of arbitrary characteristic as the corresponding ideals are “toric.”
By [22, Thm. 1.1] the ideal LkG (2) is radical for all graphs G. Using Proposition 7.4 and
gen
sym
Proposition 7.5 this implies that I3k (XG
) is radical for bipartite graphs G and I3k (XG
) is radical
for arbitrary graphs.
By [21, Cor. 2.2] the ideal L̂kG (1) is radical for all graphs G. Using Proposition 7.10 this implies
skew
) is radical for arbitrary graphs.
the Pf k4 (XG
Finally, for a forest G the results in the case of minors are derived from Proposition 7.4,
Proposition 7.5 and Theorem 2.10. In the Pfaffian case they follows using Theorem 4.1 and
Proposition 7.10.
The following corollary is an immediate consequence of the assertion (skew) in Theorem 7.11.
Corollary 7.12. Let G(2, n) be the coordinate ring of the Graßmannian of 2-dimension subspaces
in kn equipped in its standard Plücker coordinates. Then any subset of the Plücker coordinates
generates a radical ideal in G(2, n).
We note that there are subsets of 2-minors of a generic matrix that define non-radical ideals.
For example the ideal generated by the four 2-minors [12|12], [12|23], [23|12], [23|23] of a generic
3 × 3 is not radical.
A statement analogous to Corollary 7.12 for higher order Graßmannians is not true. Indeed, the
point is that a set of m-minors of a generic matrix m×n does not generate a radical ideal in general
(as it does for m = 2). For example in the Graßmannian G(3, 6) modulo [123], [124], [135], [236]
the class of [125][346] is a non-zero nilpotent.
gen
sym
Next we look into necessary conditions for Idk (XG
) and Idk (XG
) to be prime. The condition
tie in with Proposition 2.8.
Lemma 7.13. Let G = ([n], G) be a graph.
sym
k
(i) If Id+1
(XG
) is prime then G does not contain a subgraph isomorphic to Ka,b for a+b > d
(i.e. Ḡ is (n − d)-connected).
gen
k
(ii) Assume G is bipartite with bipartition [n] = V1 ∪ V2 and d + 1 ≤ |V1 |, |V2 |. If Id+1
(XG
)
is prime then deleting any |V1 | − d − 1 vertices from V1 and |V2 | − d − 1 vertices from V2
yields a connected graph.
gen
k
(iii) If G = Bd with d ≥ 4 and X is the generic (d + 2) × (d + 2) matrix then Id+1
(XG
) is
not prime.
(i) Assume Ḡ is not (n − d)-connected. Then there are n − d − 1 vertices such that
the graph obtained from Ḡ by deleting the vertices is disconnected. This implies that
sym
selecting in XG
the rows and columns corresponding to the remaining d + 1 vertices
yields a matrix which after reordering the vertices is block-diagonal with at least two
blocks. Hence its determinant is non-zero and reducible. Since the determinant is among
sym
k
a minimal generating set, it follows that Id+1
(XG
) cannot be prime.
(ii) One easily checks that similar arguments as for the proof of the first part of (i) verify the
assertion.
Proof.
24
A. CONCA AND V. WELKER
gen
(iii) Set Yd = XB
, i.e.,
d
x11
0
..
.
Yd =
0
xd+1,1
xd+2,1
0
x22
..
.
···
···
..
.
0
0
..
.
···
xd+1,2
xd+2,2
0
···
···
xdd
···
···
x1,d+1
x2,d+1
..
.
..
.
xd+1,d+1
xd+2,d+1
x1,d+2
x2,d+2
..
.
..
.
xd+1,d+2
xd+2,d+2
.
and J = Id+1 (Yd ) and let S be the polynomial ring whose indeterminates are the non-zero
entries of Yd . First, we prove that for every d ≥ 1 the ideal J has the expected height,
i.e. height J = 4. For d = 1, 2, 3 the ideal J is indeed prime of height 4: for d = 1 this is
obvious because Y1 is the generic 3 × 3 matrix while for d = 2 and d = 3 it follows from
the fact that the corresponding LSS-ideal is prime by virtue of Proposition 7.4. For d > 3
let P be a prime containing J. If P contains (x11 , x22 , x33 , x44 ) then height P ≥ 4. If P
does not contain (x11 , x22 , x33 , x44 ) we may assume x11 6∈ P . Inverting x11 and using the
standard localization trick for determinantal ideals one sees that P Sx11 contains, up to
a change of variables, Id (Yd−1 ). Hence height P = height P Sx ≥ 4. Now that we know
that J has height 4 to prove that J is not prime for d ≥ 4 it is enough to observe that
J ⊂ (x11 , x22 , x33 , x44 ). The latter is straightforward since mod (x11 , x22 , x33 , x44 ) the
submatrix of Y consisting of the first 4-rows as rank 2.
8. Questions and open problems
In Corollary 3.4 and Corollary 3.4 we have seen that for the properties c.i. and prime of LkG (d)
there is persistence along the parameter d but Example 2.2 shows persistence need not to hold for
the property of being radical.
Question 8.1. What patterns
{d < asymk (radical, G) : LkG (d) radical }
can occur for graphs G?
We expect that erratic behavior can occur. For example we believe that there exists a graph G
and a number d ≥ 3 such that LkG (d) and LkG (d + 2) are non-radical while LkG (d + 1) is radical.
We have seen in Theorem 2.9, [22, Cor. 1.4] and [22, Thm 1.1] that for certain fixed d we can
combinatorially classify the graphs G for which LkG (d) is radical, complete intersection or prime.
These classifications are based on rather simple graph theoretic properties of G.
Question 8.2. Fix a number d. Are there a (simple) combinatorially defined classes Gradical,d ,
Gc.i.,d and Gprime,d such that (say for fields k of characteristic 0):
LkG (d) is radical ⇔ G ∈ Gradical,d
and
LkG (d) is a complete intersection ⇔ G ∈ Gc.i.,d
and
LkG (d) is prime ⇔ G ∈ Gprime,d ?
As said above, we expect that the pattern of numbers d for which LkG (d) is radical can be quite
erratic. Therefore, let us concentrate on the properties prime and complete intersection. Here the
fact that the property is inherited by subgraphs supports hope for the classifications asked for in
Question 8.2. After Theorem 2.9 d = 3 for complete intersection and d = 4 for primeness are the
first parameters for which the classification is open.
For d = 3 we do not even have a conjecture when LkG (d) is a complete intersection. The graph
G from Figure 1(1) gives a graph G for which it can be checked that LkG (3) is not a complete
LSS-IDEALS AND COORDINATE SECTIONS
25
intersection for char k = 0 while G still satisfies the necessary conditions from Proposition 2.8, i.e.
has no subgraphs Ka,b with a + b = 5 and B4 .
For d = 4 geometric results from Lovász’s [27, Ch 9.4] book indicate that Proposition 2.8 still
carries the essential obstacles for LkG (4) being prime.
Question 8.3. Is it true that LkG (4) is prime if and only if G does not contain a subgraph
isomorphic to Ka,b for a + b = 5 and no subgraph isomorphic to B4 ?
k
Via the fact that primeness of LkG (d) implies primeness of Id+1
(XG ) a result by Giusti and
Merle [18, Thm. 1.6.1] guides the intuition behind the following question.
Question 8.4. Let G be a subgraph of Km,n graph and assume m ≤ n. Is it true that LkG (m − 1)
is prime if and only if G does not contain a subgraph isomorphic to Ka,b for a + b ≥ m?
By Proposition 7.4 and Proposition 7.5 we know if LkG (d) is radical or prime then so are
sym
k
and Id+1
(XG
) respectively. But our general bounds for asymk (radical, G) and
asymk (prime, G) from Corollary 5.5 are not good enough to make use of this implication. Indeed, Corollary 7.7 shows that for the properties complete intersection and prime and n large
enough there are graphs G for which Proposition 7.5 does not prove primality of an interesting
ideal. On the other hand the use of Theorem 2.10 in Theorem 7.11 shows that one can take
advantage of this connection in some cases. It would be interesting to exhibit classes different
from forests where this is possible.
gen
k
Id+1
(XG
)
Question 8.5. Are there more interesting classes of graphs G = ([n], E) for which asymk (c.i., G) ≤
n or asymk (prime, G) ≤ n − 1 ?
Despite the fact that Proposition 7.6 destroys the hope for using Theorem 7.11 for general
graphs, it would be interesting replace the asymptotic result by an actual value. By Corollary 7.7
for n large we have asymk (prime, Kn ) = 2n −
cn for some numbers cn ∈ o(n) which using the
notation of Proposition 7.6 satisfy n − wn2−2 + 1 ≥ cn ≥ 2. But we have no conjecture for an
actual formula for cn .
Question 8.6. What is the value of asymk (prime, Kn )?
For radicality we have a concrete conjecture in the case G = Kn .
Conjecture 8.7. We conjecture that asymk (radical, Kn ) = 1 (at least if char k = 0). In other
words, given a matrix of variables X of size n×d we conjecture the ideal of the off-diagonal entries
of XX T is radical for all n, d.
It would also be interesting to study the ideal generated by all the entries of XX T . We note
that the symplectic version of this problem has been investigated by De Concini in [12].
Next we turn to open problems about hypergraph LSS-ideals. We know from Theorem 2.5(2)
that for a hypergraph H = (V, E) for which E is a clutter the ideal LkH (d) is a radical complete
intersection for d ≥ pmd(G). But we prove in Theorem 2.5(3) that LH (k ) is prime for d ≥
pmd(H) + 1 only in the case that H is a graph.
Question 8.8. Is it true that for a hypergraph H = (V, E), where E is a clutter, we have LkH (d)
is prime for d ≥ pmd(H) + 1?
Similarly, the persistence results from Theorem 2.3 ask for generalizations.
Question 8.9. Let H = (V, E) be a hypergraph, where E is a clutter. Is it true that if LkH (d) is
a complete intersection (resp. prime) then so is LkH (d + 1)?
For a number r ≥ 1 we call a hypergraph H = (V, E) an r-uniform graph if every element of E
has cardinality r. In particular, E then it is a clutter. For example, graphs are 2-uniform graphs.
We say that an r-uniform graph H = (V, E) is r-partite if there is a partition V = V1 ∪ · · · ∪ Vk
such that #(A ∩ Vi ) = 1 for all i ∈ [r]. For r = 2 a 2-uniform hypergraph is 2-partite if and only
if the hypergraph considered as a graph is bipartite. Now we connect the study of ideal LkH (d) for
r-uniform (r-partite) graphs with the study of coordinate sections of the space of tensors with a
given rank. We consider two mappings:
26
A. CONCA AND V. WELKER
(φ) Let e1 , . . . , en be the standard basis vectors of kn . For vectors vi = (vi,j )j∈[d] ∈ kd , i ∈ [r],
consider the map φ that sends (v1 , . . . , vr ) ∈ (kd )n to the tensor
d X
X
j=1 σ∈Sr
vσ(i),j · · · vσ(r),j eσ(1) ⊗ · · · ⊗ eσ(r) ∈ kn ⊗ · · · ⊗ kn .
|
{z
}
r
We take the sums over the different tensors arising from ei1 ⊗ · · · ⊗ eir , for numbers
1 ≤ i1 ≤ · · · ≤ ir ≤ n, by permuting the positions as standard basis of the space of
symmetric tensors.
(j)
(ψ) Let n = n1 + · · · + nr for natural numbers n1 , . . . , nr ≥ 1. Let ei ∈ knj be the i-th
(j)
standard basis vector of knj , i ∈ [nj ] and j ∈ [r]. For vectors vi = (vi,j )j∈[d] ∈ kd for
(j)
i ∈ [nj ] and j ∈ [r] consider the map ψ that sends (vi )(i,j)∈[nj ]×[r] to
X
(1)
(r) (1)
(r)
vi1 · · · vir ei1 ⊗ · · · ⊗ eir ∈ kn1 ⊗ · · · ⊗ knr .
(i1 ,...,ir )∈[n1 ]×···×[nr ]
(r)
(1)
We take the tensors ei1 ⊗ · · · ⊗ eir for numbers ij ∈ [nj ], j ∈ [r] as the standard basis of
k n1 ⊗ · · · ⊗ k nr .
Recall that a (symmetric) tensor has (symmetric) rank ≤ d it can written as a sum of ≤ d
decomposable (symmetric) tensors. For more details on tensor rank and the geometry of bounded
rank tensors we refer the reader to [26]. Let H = (V, E) be a hypergraph. We write V(LkH (d))
for the vanishing locus of LkH (d). The the definition of the maps φ and ψ immediately implies the
following proposition.
Proposition 8.10. Let H = ([n], E) be an r-uniform hypergraph and k an algebraically closed
field.
(i) Then the restriction of the map φ to V(LkH (d)) is a parametrization of the space of symmetric tensors in kn ⊗ · · · ⊗ kn of rank ≤ d which when expanded in the standard basis has zero
{z
}
|
r
coefficient for the basis elements indexed by 1 ≤ i1 < · · · < ir ≤ n and {i1 , . . . , ir } ∈ E.
In particular, the Zariski-closure of the image of the restriction is irreducible if LkH (d) is
prime.
(ii) If H is r-partite with respect to the partition V = V1 ∪· · ·∪Vr where |Vi | = ni , i ∈ [r]. Then
the restriction of the map ψ to V(LkH (d)) is a parametrization of the space of tensors in
kn1 ⊗ · · · ⊗ knr of rank ≤ d which when expanded in the standard basis have zero coefficient
for the basis elements indexed by i1 , . . . , ir where {i1 , . . . , ir } ∈ E. In particular, the
Zariski-closure of the image of the restriction is irreducible if LkH (d) is prime.
Proposition 8.10 gives further motivation to Question 8.8. Indeed, it suggests to strengthen
Question 8.5.
Question 8.11. Let k be an algebraically closed field. Can one describe classes of r-uniform
hypergraphs H for which LkH (d) is prime for some d bounded from above by the maximal symmetric
rank of a symmetric sensor in kn ⊗ · · · ⊗ kn .
|
{z
}
r
An analogous question can be asked for r-partite r-uniform hypergraphs and tensors of bounded
rank.
References
[1] J.Abbott, A.M.Bigatti, G.Lagorio, CoCoA-5: a system for doing Computations in Commutative Algebra,
Available at http://cocoa.dima.unige.it.
[2] L.L.Avramov, Complete intersections and symmetric algebras, J. Algebra 73 (1981) 248–263.
[3] W.Bruns, U.Vetter, Determinantal rings, Lecture Notes in Math. 1327, Springer-Verlag, Berlin, 1988.
[4] W.Bruns, A.Conca, Gröbner bases and determinantal ideals, in: Commutative Algebra, Singularities and
Computer Algebra (Sinaia, 2002), 9–66, NATO Sci. Ser. II Math. Phys. Chem. 115, Kluwer Acad. Publ.,
Dordrecht, 2003.
LSS-IDEALS AND COORDINATE SECTIONS
27
[5] A.Boocher, Free resolutions and sparse determinantal ideals, Math. Res. Lett. 19 (2012) 805–821.
[6] Bourbaki, Algebra II, Chapters 4–7, Springer-Verlag, Berlin, 2003.
[7] D. Cartwright, B.Sturmfels, The Hilbert scheme of the diagonal in a product of projective spaces, Int. Math.
Res. Not., no. 9, (2010) 1741–1771.
[8] A.Conca, E. De Negri, E.Gorla, Universal Gröbner bases for maximal minors, Int. Math. Res. Not. IMRN,
no. 11, (2015) 3245–3262.
[9] A.Conca, E.De Negri, E.Gorla, Universal Gröbner bases and Cartwright-Sturmfels ideals, arXiv:1608.08942.
[10] A.Conca, E.De Negri, E.Gorla, Multigraded generic initial ideals of determinantal ideals, in: Homological
and Computational Methods in Commutative Algebra, A. Conca, J. Gubeladze, Joseph, T. Römer (Eds.),
81–96, INdAM Series 20, Springer-Verlag, Heidelberg, 2017.
[11] A.Conca, E.De Negri, E.Gorla, Cartwright-Sturmfels ideals associated to graphs and linear spaces,
arXiv:1705.00575.
[12] C.De Concini, Symplectic standard tableaux, Adv. Math. 34 (1979) 1–27.
[13] C.De Concini, E.Strickland, On the variety of complexes, Adv. Math. 41 (1981) 57–77.
[14] C.De Concini, C.Procesi, A characteristic free approach to invariant theory, Adv. Math. 21 (1976) 330–354.
[15] P.Diaconis, D.Eisenbud, B.Sturmfels, Lattice walks and primary decomposition, in: Mathematical Essays in
Honor of Gian-Carlo Rota, 173–93, Prog. Math. 161, Birkhäuser Boston, Boston, MA, 1998.
[16] R.Diestel, Graph Theory, 3rd Ed., Graduate Texts in Math. 173, Springer-Verlag, Heidelberg, 2000.
[17] D.Eisenbud, Linear sections of determinantal varieties, Amer. J. Math. 110 (1988) 541–575.
[18] M.Giusti, M.Merle, Singularités isolées et sections planes de variétés déterminantielles. II. Sections de
variétés déterminantielles par les plans de coordonnées. Algebraic geometry (La Rábida, 1981), 103118,
Lecture Notes in Math. 961, Springer-Verlag, Berlin, 1982.
[19] D.R.Grayson, M.E.Stillman, Macaulay2, a software system for research in algebraic geometry, Available at
http://www.math.uiuc.edu/Macaulay2/.
[20] C.Greene, T.Zaslavsky, On the interpretation of Whitney numbers through arrangements of hyperplanes,
zonotopes, non-Radon partitions, and orientations of graphs, Trans. Amer. Math. Soc. 280 (1983) 97–126.
[21] J.Herzog, T.Hibi, F.Hreinsdóttir, T.Kahle, J.Rauh, Binomial edge ideals and conditional independence statements, Adv. in Appl. Math. 45 (2010) 317–333.
[22] J.Herzog, A. Macchia, S.Madani, V.Welker, On the ideal of the orthogonal representations of a graph in R2 ,
Adv. in Appl. Math. 71 (2015) 146–173.
[23] S.Hoşten, S.Sullivant, Ideals of adjacent minors. J. Algebra 277 (2004) 615–642.
[24] C.Huneke, On the symmetric algebra of a module, J. Algebra 69 (1981) 113-119.
[25] D.Kiani, S.Saeedi Madani, The Castelnuovo-Mumford regularity of binomial edge ideals, J. Combin. Theory
Ser. A 139 (2016) 80–86.
[26] J.M. Landsberg, Tensors: geometry and applications, Graduate Studies in Mathematics 128, American
Mathematical Society, Providence, RI, 2012
[27] L.Lovász, Geometric Representations of Graphs, http://www.cs.elte.hu/˜lovasz/geomrep.pdf, 2016.
[28] L.Lovász, M.Saks, A.Schrijver, Orthogonal representations and connectivity of graphs, Linear Alg. Appl.
114/115 (1989) 439–454.
[29] K.Matsuda, S.Murai, Regularity bounds for binomial edge ideals, J. Commut. Algebra 5 (2013) 141–149.
[30] J.S. Milne, Algebraic Geometry, Version 6.02, http://www.jmilne.org/math/CourseNotes/AG.pdf, 2017.
[31] C.Musili, C.S.Seshadri, Schubert varieties and the variety of complexes, in: Arithmetic and Geometry, Vol.
II, 329–359, Progr. Math. 36, Birkhäuser Boston, Boston, MA, 1983.
[32] M.Ohtani, Graphs and ideals generated by some 2-minors, Comm. Alg. 39 (2011) 905–917.
[33] A.Tchernev, Universal complexes and the generic structure of free resolutions, Michigan Math. J. 49 (2001)
65–96.
[34] H.Weyl, The Classical Groups, their Invariants and Representations, Princeton Univ. Press, Princeton, N.J.,
1939.
Dipartimento di Matematica, Università di Genova, Via Dodecaneso 35, I-16146 Genova, Italy
E-mail address: [email protected]
Philipps-Universität Marburg, Fachbereich Mathematik und Informatik, D-35032 Marburg, Germany
E-mail address: [email protected]
| 0math.AC
|
Link Adaptation Algorithms for Dual
Polarization Mobile Satellite Systems
Anxo Tato1 , Pol Henarejos2 , Carlos Mosquera1 , and Ana Pérez-Neira2,3
arXiv:1803.07318v1 [eess.SP] 20 Mar 2018
1
2
Universidade de Vigo
[email protected], [email protected]
www.gpsc.uvigo.es
Centre Tecnològic de Telecomunicacions de Catalunya
Castelldefels, Barcelona
[email protected], [email protected]
www.cttc.cat
3
Universitat Politècnica de Catalunya, Barcelona
Abstract. The use of dual polarization in mobile satellite systems is
very promising as a means for increasing the transmission capacity. In
this paper we study a system which uses simultaneously two orthogonal polarizations in order to communicate with the users. The application of MIMO signal processing techniques along with Adaptive Coding
and Modulation in the forward link can provide remarkable throughput
gains up to 100 % when compared with the single polarization system.
The gateway is allowed to vary the MIMO and Modulation and Coding
Schemes for each frame. The selection is done by means of a link adaptation algorithm which uses a tunable margin to achieve a predefined
target Frame Error Rate.
Key words: link adaptation, Adaptive Coding and Modulation, MIMO,
Dual-polarization, Satellite communications, Mobile Satellite Systems
1 Introduction
In recent years, the spectrum saturation and the increasing demand for higher
data rates in a ubiquitous way encourages the engineers to design new techniques
in order to increase the capacity of communication systems without resorting to
expand the occupied bandwidth. Two of these techniques are the leverage of
multiple antennas at both transmitter and receiver by means of MIMO (Multiple Input Multiple Output) signal processing techniques and also the so-called
Adaptive Coding and Modulation (ACM) or link adaptation. Both are part of
many current terrestrial wireless communication standards such as LTE [1] and
IEEE 802.16 [2] for cellular technologies and IEEE 802.11 [3] for wireless local
area networks.
In this paper we propose to apply MIMO techniques to mobile satellite communication systems by exploiting the polarization domain. Satellite links operating at low frequency bands such as L- and S-bands, usually rely on a single
circular polarization, either RHCP (Right Hand Circular Polarization) or LHCP
2
Anxo Tato, Pol Henarejos,Carlos Mosquera, Ana Pérez-Neira
(Left Hand Circular Polarization), to avoid the effects of the Faraday rotation [4].
Here we propose the simultaneous use of both orthogonal polarizations within the
same beam to communicate with the users following up some preliminary studies presented in [5] within the framework of SatNEx-IV, the European Satellite
Network of Experts. An algorithm to switch among different MIMO schemes will
be presented. In particular, SISO, Orthogonal Polarization-Time Block Codes
(OPTBC), Polarization Modulation (PMod) or V-BLAST (Vertical Bell Laboratories Layer Space-Time) will be chosen as a function of a set of effective
Signal to Noise Ratio (SNR) metrics.
On the other hand, the use of ACM techniques allows the system to adapt its
instantaneous rate to the current channel capacity without designing the system
for the channel worst case. In this paper, following our previous works [6], [7] and
[8], we propose the use of a link adaptation algorithm with an adaptive margin
for selecting the Modulation and Coding Scheme of each frame (MODCOD or
MCS). This algorithm operates together with the MIMO mode selection scheme
to increase the spectral efficiency whilst trying to maintain the Frame Error
Rate (FER) at a level specified by the Quality of Service (QoS) parameter of
the connection.
The remainder of the article is structured as follows. Section 2 provides a
description of the satellite communication system, including the link adaptation
algorithms, an overview of the channel model and how the channel series are
generated. Section 3 describes the link adaptation algorithms for selecting both
MIMO mode and MCS. Section 4 provides the simulation results of the algorithms in a maritime mobile scenario. Lastly, in Section 5 we collect the main
conclusions and note the future work on this topic.
2 System model
We consider a satellite communications system which serves mobile users operating in L- or S-band. A link adaptation algorithm is proposed for the forward link
from the gateway towards the mobile terminal (MT). Typically, at these bands
only one polarization is used to protect from high cross-polarization leakage. For
example, the service BGAN (Broadband Global Area Network), standardized by
the ETSI as TS 102 744, only employs RHCP [9]. In higher frequency bands, like
Ku and Ka, the high XPD (cross polar discrimination) would make it possible
to use two independent streams in both polarizations. The system we propose
here, and for which we develop the adaptive algorithms, uses simultaneously
both polarizations, yielding a 2x2 MIMO system.
The system that we analyse is narrow-band and single-carrier, and to some
extent it is inspired by BGAN. The baseline Single Input Single Output (SISO)
system is complemented by additional MIMO modes, namely OPTBC, Polarization Modulation (PMod) and V-BLAST:
– OPTBC (Orthogonal Polarization-Time Block Code) is based on Alamouti
space-time coding, used for achieving transmit diversity and introduced in
Link Adaptation Algorithms for Dual Polarization Mobile Satellite Systems
3
[10]. The spatial components in Alamouti are replaced by the two available
polarizations in OPTBC.
– PMod is analogous to Spatial Modulation (SM) techniques [11], with only one
polarization per channel use transmitting the symbols from a given constellation. One extra bit of information is conveyed indicating which polarization is
chosen. For a QPSK constellation this implies a 50 % gain in spectral efficiency.
– In V-BLAST [12] two independent symbols are transmitted per channel use
and the receiver is in charge of reducing the cross-stream interference to perform the detection.
All these three MIMO schemes do not require Channel State Information (CSI)
at the transmitter and therefore are suitable for the satellite scenario where the
CSIT is outdated due to the long round trip time.
In addition to the forward link, where our algorithms operate, we assume that
there is a feedback channel in the return link used by the Mobile Terminal (MT)
to inform the gateway about the CSI, the result of the frames decoding (in the
form of ACK/NACK) and the optimum transmission mode (which is calculated
by the receiver). All these information is used by the gateway, our transmitter, in
the selection of the preferred MCS by the link adaptation algorithm we propose.
The signal model for a given time instant n is
√
(1)
yn = P Hn xn + wn
where P is the transmitted power, Hn ∈ C2x2 is the channel matrix, yn ∈ C2
is the vector of the received signal which has a component per each polarization,
xn ∈ C2 is the transmitted signal and wn ∼ CN (0, σ 2 I2 ) represents the
additive complex white Gaussian noise (AWGN). Therefore, the resultant Signal
to Noise Ratio is SNR = P/σ 2 .
The transmitted symbols are grouped into blocks or packets (codewords)
which span N = 2560 symbols, transmitted with a baud rate of 33600 symbols/s,
which gives a frame length of 80 ms. As we use frames of just one block length,
hereafter we will employ the term frame for referring to a codeword. In order to
speed up the simulations, we do not implement the entire signal processing block
chain and thus, we use Physical Layer Abstraction techniques, in particular the
Mutual-Information Effective SNR [6]. The comparison of the effective SNR of
a received frame with a threshold SNR of the MCS used to transmit that frame
allows us to decide if the frame can be decoded (the effective SNR is higher than
the corresponding threshold) or not.
The effective SNR of a frame, SNRef f , given the set of the N SNRs of each
symbol γn can be obtained with
!
N
1 X
−1
SNRef f = Φ
Φ(γn ) ,
(2)
N n=1
procedure called SNR compression in [13]. The SNR γn for each symbol period
is calculated for the different MIMO modes as:
4
Anxo Tato, Pol Henarejos,Carlos Mosquera, Ana Pérez-Neira
– SISO:
γn = SNR|[Hn ]11 |2
– OPTBC:
γn =
(3)
SNR
kHn k2F
2
(4)
– PMod:
|[Hn ]11 |2 + |[Hn ]22 |2
2
– V-BLAST (MMSE (Minimum Mean Square Error) receiver):
(5)
γn = SNR
γn,k =
1
−1
SNR H
[(I2 +
Hn Hn )−1 ]k,k
2
(6)
In the previous equations [Hn ]ij denotes the coefficient (i, j) of the channel
matrix for time instant n, k indicates the number of the stream (1 or 2), I2 is
the 2 × 2 identity matrix and kHn kF denotes the Frobenius norm of the matrix.
We should say that in the case of V-BLAST the SNR expression depends on the
type of receiver. Throughout this work we assume the MMSE receiver is used
for V-BLAST reception, given its simplicity and robustness against noise.
Table 1 collects all the available transmission modes and MCS, including the
required effective SNR for correct decoding (obtained from the curve of SNR vs
Mutual Information) and corresponding spectral efficiency in bps/Hz.
Table 1. Coding rate options for the F80T1Q-1B bearer [9], QPSK constellation.
L8
L7
L6
L5
L4
L3
L2
L1
R
Coding rate
0.34 0.40 0.48 0.55 0.63 0.70 0.77 0.83 0.87
Threshold SNR (SNRth ) (dB)
-2.15 -1.21 -0.09 0.83 1.84 2.74 3.67 4.54 5.19
Spectral efficiency SISO/OPTBC 0.68 0.80 0.96 1.10 1.26 1.40 1.54 1.66 1.74
Spectral efficiency PMod
1.02 1.20 1.44 1.65 1.89 2.10 2.31 2.49 2.61
Spectral efficiency V-BLAST
1.36 1.60 1.92 2.20 2.52 2.80 3.08 3.32 3.48
2.1 Channel model
The simulation of the mobile satellite dual polarized channel has been done following the work of [14]. The channel is obtained as the sum of three different
components: the Line-of-Sight (LoS), the specular reflected signal and the diffuse
components, produced by the scatterers near the MT. LoS and specular components are modelled as Rice random variables whereas the diffuse component,
which causes the fast fading, is Rayleigh distributed. These three components
are grouped into the following channel matrix:
Link Adaptation Algorithms for Dual Polarization Mobile Satellite Systems
H = βejφ KL + ξejφ KS + DKD
5
(7)
KL , KS and KD are the K-factor matrices which collect the Rice factors of each
polarization for the LoS, specular and diffuse components, respectively. The two
matrices β and ξ are related to how the channel mixes the two polarizations and
matrix D entries are complex Gaussian random variables with a given covariance
matrix. The parameters for building all these matrices depend on the considered
environment and can be found in [14].
2.2 Channel generation
Fig. 1 shows the block diagram of the channel time series generator. One important aspect of the generated channel series is the time correlation and the
Doppler spread. Assuming the Clarke’s model, the coherence time can be approximated by
3λ
(8)
τc = √ ,
4v π
being λ the wavelength and v the mobile terminal speed. We first generate Q =
dN/(τc fs )e independent realizations of the channel matrices from the uniformly
distributed random phases and the complex Gaussian random matrix D. Then
we make a linear interpolation to obtain the N channel matrices and a low-passfilter with cut-off frequency equal to the Doppler spread, Ds = v/λ, is applied.
Lastly, the channel matrices are scaled to yield the given average SNR of the
simulation.
Fig. 1. Block diagram of the channel time series generator
3 Algorithm for mode and MCS selection
In this paper, the link adaptation procedure has to select the MCS and the
MIMO mode (SISO, OPTBC, PMod or V-BLAST). We split these tasks between
6
Anxo Tato, Pol Henarejos,Carlos Mosquera, Ana Pérez-Neira
the gateway and the MT. On the one hand, the MT selects the MIMO mode
that achieves the highest throughput given the current SNR. The receiver needs
to compute the effective SNR for all modes to choose the most efficient. This
task is performed by the receiver to reduce the feedback load. In addition, the
transmitter (gateway) is informed by the MT of the ACK/NACK, the preferred
MIMO mode and the effective SNR for that mode, and selects consequently the
MCS of the following frame.
In order to formalize the procedure let us define the following matrices. S
is the matrix with the spectral efficiency of each combination of MIMO mode
- MCS. For example, the element Sij is the spectral efficiency of MIMO mode
i and MCS number j. We also define the vector SNRef f , which represents the
CSI and whose elements are the effective SNR for each one of the transmission
modes. Lastly, we have the vector with the threshold SNR of each MCS SNRth ;
note that this threshold is the same for all MIMO modes.
The optimization problem to choose the optimum mode T̂ can be written as
T̂ = i
s. to
Rij = max{Rij },
i,j
(9)
with the corresponding matrix of rates given by Rij = Sij [SNRef f,i ≥ SNRth,j ].
If no MCS verifies the required effective SNR, i.e., all Rij are zero, or in case
of a tie, the MIMO mode with the highest effective SNR is chosen and reported
back to the gateway.
The choice of the MCS at the gateway follows our previous work in [7], which
exploits an adaptive margin. Here, for lack of space, we do not replicate the
derivation of the algorithm; those interested readers are referred to [6],[7]. The
gateway selects an MCS mi using a Lookup Table (LUT), represented by means
of a function Π(·) which maps SNR intervals to MCSs: mi = Π (SNRi−d + ci ).
All this is shown graphically in the diagram of Fig. 2. The value of the SNR
reported back by the MT, SNRi−d , plus a margin ci is introduced in the LUT
for selecting the MCS; this margin is updated when new feedback comes in.
We assume a round trip time which amounts to the duration of d frames. The
recursive equation for updating the value of the margin is
ci+1 = ci −
µ
(i−d − p0 )θ.
θ2 + SNR2i−2d
(10)
This recursion is derived in [6] to solve the following optimization problem
2
min J(c) = min |E[] − p0 | .
c
c
(11)
The involved variables are SNRi−2d , the effective SNR used by the LUT to decide
the MCS of frame i−d, and the acknowledgement i−d (1 for ACK, 0 for NACK).
µ and θ are two constants which take the values 1 and 10, respectively. In this
work we suppose the objective FER, p0 , has a fixed value, although there are
some studies like [15] which show the benefits of having a variable p0 in order to
maximize the throughput.
Link Adaptation Algorithms for Dual Polarization Mobile Satellite Systems
7
Fig. 2. Block diagram of the MCS selection algorithm with the LUT.
4 Simulation results
Several simulations are performed to evaluate the spectral efficiency gain from
the use of two polarizations simultaneously, and also to understand the potential
robustness coming from the use of an adaptive margin in the link adaptation
algorithm. All the simulations results presented here are obtained for a maritime
scenario with a vessel moving at a constant speed of 50 km/h. The parameters of
the channel generator were taken from [14]. The carrier frequency in the simulations is 1.6 GHz, a typical frequency of some Mobile Satellite Systems operating
in the L-band. QPSK modulation with frames of 80 ms and 2560 symbols (baud
rate of 33600 symbols/s) are used in the physical layer, similarly to the bearer
F80T1Q-1B of BGAN [9]. Each simulation comprises the transmission and reception of N = 60, 000 frames for a specific average SNR, ranging from -5 dB to
25 dB in steps of 2.5 dB. Average spectral efficiency, defined as (1 − i )rmi /N ,
and cumulative FER during the whole transmission is also computed for each
simulation. In the previous expression rj is the rate of the jth MCS, mi is the
MCS selected for frame i and i is the corresponding ACK. Lastly, the RTT is
set to 7 frames (560 ms) to approximate the feedback delay for GEO satellites.
Contrary to other publications which assume instantaneous feedback, here it has
a reasonable value and the proposed scheme is robust even with this delay.
The objective of the first set of simulations is to show the benefits of using
two orthogonal polarizations simultaneously to serve a mobile user in low bands
of the spectrum. Despite the XPD being lower in these bands than in Ku and Ka
bands, it provides a significant throughput gain even with the available power
split between the two polarizations. In these simulations the SISO case (transmit
only in one polarization) is compared with the use of the two polarizations
(MIMO) based on the algorithm for choosing dynamically the MIMO mode of
Equation (9). In both cases ACM is employed with a fixed margin c = −1 dB
in the LUT (see Fig. 2).
The results collected in Table 2 show that use of dual polarization provides
a gain higher than 50% for most cases. Interestingly, OPTBC makes it possible
to extend the operation range at low SNR below SISO. As the SNR increases
the selected mode turns to be PMod instead of OPTBC with gains of up to
8
Anxo Tato, Pol Henarejos,Carlos Mosquera, Ana Pérez-Neira
Table 2. Results of the comparison of single polarization (SISO) and adaptive dual
polarization (MIMO).
Average SNR
Selected modes
-5 dB
OPTBC
Efficiency (SISO) Efficiency (MIMO)
0 bps/Hz
0.68 bps/Hz
Gain
Inf
-2.5 to 7.5 dB
PMod
0.68 - 1.50 bps/Hz
1.02 - 2.32 bps/Hz
50 - 55 %
10 dB
PMod (30 %), BLAST (70 %)
1.69 bps/Hz
2.47 bps/Hz
47 %
12.5 to 25 dB
BLAST
1.73 - 1.74 bps/Hz
2.77 - 3.48 bps/Hz
60 - 100 %
55%. At an average SNR of 10 dB both PMod and V-BLAST are employed and
for higher SNRs V-BLAST is mainly used. The V-BLAST MIMO mode, which
sends two simultaneous streams of symbols, offers up to 100% gain in terms of
throughput, i.e., it can double the capacity.
Next we let the margin to adapt in an effort to match a target FER. Figure
3 shows the results obtained in terms of average spectral efficiency and FER. An
ACM scheme with a fixed margin of -1 dB is compared with a link adaptation
algorithm with adaptive margin for two different values of target FER, 10% and
1%.
0
10
3
−1
10
2.5
−2
10
FER
Spectral Efficiency (bps/Hz)
3.5
2
−3
10
1.5
−4
0.5
−5
10
Fixed margin (−1 dB)
Adaptive margin (target FER 10 %)
Adaptive margin (target FER 1 %)
1
Fixed margin (−1 dB)
Adaptive margin (target FER 10 %)
Adaptive margin (target FER 1 %)
−5
0
5
10
15
Average SNR (dB)
20
25
10
−5
0
5
10
15
20
25
Average SNR (dB)
Fig. 3. Average spectral efficiency (left) and FER (right)
In Fig. 3 we observe the behaviour of the algorithm in terms of spectral efficiency and FER for a wide range of LOS SNRs. The spectral efficiency grows
linearly with the SNR in dB until it reaches its maximum value of 3.48 bps/Hz.
When looking only at the throughput there are not significant differences between using a fixed and an adaptive margin. Furthermore, we can observe how
the target FER p0 influences the final spectral efficiency. For example, the simulations with the target FER 1 % have a slightly better efficiency than those
with target FER 10 %. This dependence of the spectral efficiency with the target
error rate was already studied in [15] where the authors show that there is an
optimal FER which maximizes the throughput.
Link Adaptation Algorithms for Dual Polarization Mobile Satellite Systems
9
In general, the margin ci converges to a value and oscillates around it, except
for the three higher SNRs where it grows indefinitely. We have observed that
the value of the final margin depends on the operation point (LOS SNR), the
target FER p0 , the channel conditions and the mobile speed. This last fact is
exemplified in [7] where we show the evolution of the margin in an Intermediate
Tree Shadow environment for several MT speeds.
On the other side, the FER behaves very differently when adapting the margin
with respect to the fixed case; as shown on the right plot in Figure 3, a fixed
margin cannot guarantee a prescribed FER for different SNR values. As opposed
to this, the adaptive algorithm exposed earlier matches the target FER, 0.1 and
0.01 in the example, for a wide range of LOS SNRs. Therefore, the combination
of the transmission mode selection and the adaptive margin achieves remarkable
throughput gains when compared with the SISO case, whilst guaranteeing a
physical layer FER suited to the prescribed QoS.
5 Conclusions and future work
The simultaneous use of dual orthogonal circular polarizations in L/S-bands for
mobile satellite communication systems is proposed in this paper. The application of MIMO techniques to this satellite scenario, by switching among several
MIMO modes, generates significant increments in the spectral efficiency that
can get close to 100% for high values of SNR, without using extra bandwidth
or power. Moreover, we have proposed link adaptation algorithms to select the
MIMO mode and the MCS of each frame based on an adaptive margin. This
helps the system to adapt itself to the channel conditions and guarantee a prefixed target Frame Error Rate for a wide range of SNRs. Lastly, in the future
we plan to improve the algorithm through the online calculation of the optimal
FER which maximizes the spectral efficiency.
6 Acknowledgements
This work was partially funded by the Agencia Estatal de Investigación (Spain)
and the European Regional Development Fund under project MYRADA (TEC201675103-C2-2-R). Also funded by the Xunta de Galicia (Secretarı́a Xeral de Universidades) under a predoctoral scholarship (co-funded by the European Social
Fund) and under Agrupación Estratéxica Consolidada de Galicia accreditation
2016-2019 (co-funded by the European Regional Development Fund - ERDF).
Part of the research was done during a stay at the Centre Tecnològic de Telecomunicacions de Catalunya (CTTC) supported by the project SatNEx-IV, cofunded by the European Space Agency (ESA). This work has also received funding from the Spanish Ministry of Economy and Competitiveness (Ministerio de
Economı́a y Competitividad) under project TEC2014-59255-C3-1-R and from
the Catalan Government (2014SGR1567).
10
Anxo Tato, Pol Henarejos,Carlos Mosquera, Ana Pérez-Neira
References
1. “LTE; evolved universal terrestrial radio access (E-UTRA); physical layer procedures,” ETSI TS 136 213 V8.8.0 (2009-10).
2. “IEEE standard for air interface for broadband wireless access systems,” IEEE Std
802.16-2012 (Revision of IEEE Std 802.16-2009), pp. 1 –2542, 17 2012.
3. “IEEE standard for information technology–telecommunications and information
exchange between systems local and metropolitan area networks–specific requirements part 11: Wireless lan medium access control (MAC) and physical layer
(PHY) specifications,” IEEE Std 802.11-2012 (Revision of IEEE Std 802.11-2007),
pp. 1–2793, 29 2012.
4. M. Richharia, Mobile Satellite Communications: Principles and Trends, 2nd ed.
Wiley, 2014.
5. P. Henarejos, A. Perez-Neira, N. Mazzali, and C. Mosquera, “Advanced signal
processing techniques for fixed and mobile satellite communications,” in 2016 8th
Advanced Satellite Multimedia Systems Conference and the 14th Signal Processing
for Space Communications Workshop (ASMS/SPSC), Sept 2016, pp. 1–8.
6. A. Rico-Alvariño, J. Arnau, and C. Mosquera, “Link adaptation in mobile satellite
links: schemes for different degrees of csi knowledge,” International Journal of
Satellite Communications and Networking, vol. 34, no. 5, pp. 679–694, 2016,
sat.1164. [Online]. Available: http://dx.doi.org/10.1002/sat.1164
7. A. Rico-Alvariño, A. Tato, and C. Mosquera, “Robust adaptive coding and modulation scheme for the mobile satellite forward link,” in Signal Processing Advances
in Wireless Communications (SPAWC), 2015 IEEE 16th International Workshop
on, June 2015, pp. 530–534.
8. A. Tato, C. Mosquera, and I. Gomez, “Link adaptation in mobile satellite links:
field trials results,” in 2016 8th Advanced Satellite Multimedia Systems Conference and the 14th Signal Processing for Space Communications Workshop
(ASMS/SPSC), Sept 2016, pp. 1–8.
9. “Satellite component of UMTS (S-UMTS); family SL satellite radio interface,”
ETSI TS 102 744, Oct. 2015.
10. S. M. Alamouti, “A simple transmit diversity technique for wireless communications,” IEEE Journal on Selected Areas in Communications, vol. 16, no. 8, pp.
1451–1458, Oct 1998.
11. P. Henarejos and A. I. Perez-Neira, “Dual polarized modulation and reception for
next generation mobile satellite communications,” IEEE Transactions on Communications, vol. 63, no. 10, pp. 3803–3812, Oct 2015.
12. P. W. Wolniansky, G. J. Foschini, G. D. Golden, and R. A. Valenzuela, “V-blast:
an architecture for realizing very high data rates over the rich-scattering wireless channel,” in 1998 URSI International Symposium on Signals, Systems, and
Electronics. Conference Proceedings (Cat. No.98EX167), Sep 1998, pp. 295–300.
13. F. Kaltenberger, I. Latif, and R. Knopp, “On scalability, robustness and accuracy
of physical layer abstraction for large-scale system-level evaluations of LTE networks,” in 2013 Asilomar Conference on Signals, Systems and Computers, Nov
2013, pp. 1644–1648.
14. M. Sellathurai, P. Guinand, and J. Lodge, “Space-time coding in mobile satellite
communications using dual-polarized channels,” IEEE Transactions on Vehicular
Technology, vol. 55, no. 1, pp. 188–199, Jan 2006.
15. S. Park, R. C. Daniels, and R. W. Heath, “Optimizing the target error rate for link
adaptation,” in 2015 IEEE Global Communications Conference (GLOBECOM),
Dec 2015, pp. 1–6.
| 7cs.IT
|
arXiv:1508.07552v1 [math.RA] 30 Aug 2015
Canonical and n-canonical modules on a
Noetherian algebra
Mitsuyasu Hashimoto
Department of Mathematics, Okayama University
Okayama 700–8530, JAPAN
[email protected]
Dedicated to Professor Shiro Goto on the occasion of his 70th birthday
Abstract
We define canonical and n-canonical modules on a module-finite algebra over a Noether commutative ring and study their basic properties.
Using n-canonical modules, we generalize a theorem on (n, C)-syzygy
by Araya and Iima which generalize a well-known theorem on syzygies
by Evans and Griffith. Among others, we prove a non-commutative
version of Aoyama’s theorem which states that a canonical module descends with respect to a flat local homomorphism. We also prove the
codimension two-argument for modules over a coherent sheaf of algebras with a 2-canonical module, generalizing a result of the author.
Contents
1 Introduction
2
2 Preliminaries
5
3 Xn,m -approximation
10
2010 Mathematics Subject Classification. Primary 16E05; Secondary 16E65. Key Words
and Phrases. canonical module; n-canonical module; (n, C)-syzygy; n-C-spherical approximation.
1
4 (n, C)-TF property
17
5 Canonical module
22
6 n-canonical module
26
7 Non-commutative Aoyama’s theorem
34
8 Evans–Griffith’s theorem for n-canonical modules
37
9 Symmetric and Frobenius algebras
40
10 Codimension-two argument
46
1.
Introduction
(1.1) In [EvG], Evans and Griffith proved a criterion of a finite module over
a Noetherian commutative ring R to be an nth syzygy. This was generalized
to a theorem on (n, C)-syzygy for a semidualizing module C over R by Araya
and Iima [ArI]. The main purpose of this paper is to prove a generalization
of these results in the following settings: the ring R is now a finite R-algebra
Λ, which may not be commutative; and C is an n-canonical module.
(1.2) The notion of n-canonical module was introduced in [Has] in an algebrogeometric situation. The criterion for a module to be an nth syzygy for
n = 1, 2 by Evans–Griffith was generalized using n-canonical modules there,
and the standard ‘codimension-two argument’ (see e.g., [Hart4, (1.12)]) was
also generalized to a theorem on schemes with 2-canonical modules [Has,
(7.34)]. We also generalize this result to a theorem on modules over noncommutative sheaves of algebras (Proposition 10.5).
(1.3) Let (R, m) be a complete semilocal Noetherian ring, and Λ 6= 0 a
module-finite R-algebra. Let I be a dualizing complex of R. Then RHomR (Λ, I)
is a dualizing complex of Λ. Its lowest non-vanishing cohomology is denoted
by KΛ , and is called the canonical module of Λ. If (R, m) is semilocal but not
complete, then a Λ-bimodule is called a canonical module if it is the canonical module after completion. An n-canonical module is defined using the
canonical module. A finite right (resp. left, bi-)module C of Λ is said to be ncanonical over R if (1) C satisfies Serre’s (Sn′ ) condition as an R-module, that
is, for any P ∈ Spec R, depthRP CP ≥ min(n, dim RP ). (2) If P ∈ SuppR C
2
cP is isomorphic to K c as a right (left, bi-) module
with dim RP < n, then C
ΛP
c
c
of ΛP , where ΛP is the P RP -adic completion of ΛP .
(1.4) In order to study non-commutative n-canonical modules, we study
some non-commutative analogue of the theory of canonical modules developed
by Aoyama [Aoy], Aoyama–Goto [AoyG], and Ogoma [Ogo] in commutative
algebra. Among them, we prove an analogue of Aoyama’s theorem [Aoy] which
states that the canonical module descends with respect to flat homomorphisms
(Theorem 7.5).
(1.5) Our main theorem is the following.
Theorem 8.4 (cf. [EvG, (3.8)], [ArI, (3.1)]). Let R be a Noetherian commutative ring, and Λ a module-finite R-algebra, which may not be commutative.
Let n ≥ 1, and C be a right n-canonical Λ-module. Set Γ = EndΛop C. Let
M ∈ mod C. Then the following are equivalent.
1 M ∈ TF(n, C).
2 M ∈ UP(n, C).
3 M ∈ Syz(n, C).
4 M ∈ (Sn′ )C .
Here M ∈ (Sn′ )C means that SuppR M ⊂ SuppR C, and for any P ∈
Spec R, depth MP ≥ min(n, dim RP ), and this is a (modified) Serre’s condition. M ∈ Syz(n, C) means M is an (n, C)-syzygy. M ∈ UP(n, C) means
existence of an exact sequence
0 → M → C 0 → C 1 → · · · → C n−1
which is still exact after applying (?)† = HomΛop (?, C).
(1.6) The condition M ∈ TF(n, C) is a modified version of Takahashi’s condition “M is n-C-torsion free” [Tak]. Under the assumptions of the theorem,
let (?)† = HomΛop (?, C), Γ = EndΛop C, and (?)‡ = HomΓ (?, C). We say that
M ∈ TF(1, C) (resp. M ∈ TF(2, C)) if the canonical map λM : M → M †‡ is injective (resp. bijective). If n ≥ 3, we say that M ∈ TF(n, C) if M ∈ TF(2, C),
and ExtiΓ (M † , C) = 0 for 1 ≤ i ≤ n − 2, see Definition 4.5. Even if Λ is
a commutative ring, a non-commutative ring Γ appears in a natural way, so
even in this case, the definition is slightly different from Takahashi’s original
one. We prove that TF(n, C) = UP(n, C) in general (Lemma 4.7). This is a
modified version of Takahashi’s result [Tak, (3.2)].
3
(1.7) As an application of the main theorem, we formulate and prove a
different form of the existence of n-C-spherical approximations by Takahashi
[Tak], using n-canonical modules, see Corollary 8.5 and Corollary 8.6. Our
results are not strong enough to deduce [Tak, Corollary 5.8] in commutative
case. For related categorical results, see below.
(1.8) Section 2 is preliminaries on the depth and Serre’s conditions on modules. In Section 3, we discuss Xn,m -approximation, which is a categorical abstraction of approximations of modules appeared in [Tak]. Everything is done
categorically here, and Theorem 3.16 is an abstraction of [Tak, (3.5)], in view
of the fact that TF(n, C) = UP(n, C) in general (Lemma 4.7). In Section 4,
we discuss TF(n, C), and prove Lemma 4.7 and related lemmas. In Section 5,
we define the canonical module of a module-finite algebra Λ over a Noetherian
commutative ring R, and prove some basic properties. In Section 6, we define
the n-canonical module of Λ, and prove some basic properties, generalizing
some constructions and results in [Has, Section 7]. In Section 7, we prove a
non-commutative version of Aoayama’s theorem which says that the canonical
module descends with respect to flat local homomorphisms (Theorem 7.5). As
a corollary, as in the commutative case, we immediately have that a localization of a canonical module is again a canonical module. This is important
in Section 8. In Section 8, we prove Theorem 8.4, and the related results on
n-C-spherical approximations (Corollary 8.5, Corollary 8.6) as its corollaries.
Before these, we prove non-commutative analogues of the theorems of Schenzel and Aoyama–Goto [AoyG, (2.2), (2.3)] on the Cohen–Macaulayness of the
canonical module (Proposition 8.2 and Corollary 8.3). In section 9, we define
and discuss non-commutative, higher-dimensional symmetric, Frobenius, and
quasi-Frobenius algebras and their non-Cohen–Macaulay versions. In commutative algebra, the non-Cohen–Macaulay version of Gorenstein ring is known
as quasi-Gorenstein rings. What we discuss here is a non-commutative version
of such rings. Scheja and Storch [SS] discussed a relative notion, and our definition is absolute in the sense that it is independent of the choice of R. If R is
local, our quasi-Frobenius property agrees with Gorensteinness discussed by
Goto and Nishida [GN], see Proposition 9.7 and Corollary 9.8. In Section 10,
we show that the codimension-two argument using the existence of 2-canonical
modules in [Has] is still valid in non-commutative settings.
(1.9) Acknowledgments: Special thanks are due to Professor Osamu Iyama
for valuable advice and discussion. Special thanks are also due to Professor
Tokuji Araya. This work was motivated by his advice, and Proposition 8.2 is
4
an outcome of discussion with him.
The author is also grateful to Professor Kei-ichiro Iima, Professor Takesi
Kawasaki, Professor Ryo Takahashi, Professor Kohji Yanagawa, and Professor
Yuji Yoshino for valuable advice.
2.
Preliminaries
(2.1) Unless otherwise specified, a module means a left module. Let B be a
ring. HomB or ExtB mean the Hom or Ext for left B-modules. B op denotes
the opposite ring of B, so a B op -module is nothing but a right B-module. Let
B Mod denote the category of B-modules. B op Mod is also denoted by Mod B.
For a left (resp. right) Noetherian ring B, B mod (resp. mod B) denotes the
full subcategory of B Mod (resp. Mod B) consisting of finitely generated left
(resp. right) B-modules.
(2.2) For derived categories, we employ standard notation found in [Hart].
For an abelian category A, D(A) denotes the unbounded derived category
of A. For a plump subcategory (that is, a full subcategory which is closed
under kernels, cokernels, and extensions) B of A, DB (A) denotes the triangulated subcategory of D(A) consisting of objects F such that H i (F) ∈ B for
any i. For a ring B, We denote D(B Mod) by D(B), and DB mod (B Mod) by
Dfg (B) (if B is left Noetherian).
(2.3) Throughout the paper, let R denote a commutative Noetherian ring.
If R is semilocal (resp. local) and m its Jacobson radical, then we say that
(R, m) is semilocal (resp. local). We say that (R, m, k) is semilocal (resp. local)
if (R, m) is semilocal (resp. local) and k = R/m.
(2.4) We set R̂ := R ∪ {∞, −∞} and consider that −∞ < R < ∞. As
a convention, for a subset Γ of R̂, inf Γ means inf(Γ ∪ {∞}), which exists
uniquely as an element of R̂. Similarly for sup.
(2.5) For an ideal I of R and M ∈ mod R, we define
depthR (I, M) := inf{i ∈ Z | ExtiR (R/I, M) 6= 0},
and call it the I-depth of M [Mat, section 16]. It is also called the M-grade of I
[BS, (6.2.4)]. When (R, m) is semilocal, we denote depth(m, M) by depthR M
or depth M, and call it the depth of M.
Lemma 2.6. The following functions on M (with valued in R̂) are equal for
an ideal I of R.
5
1 depthR (I, M);
2 inf P ∈V (I) depthRP MP , where V (I) = {P ∈ Spec R | P ⊃ I};
3 inf{i ∈ Z | HIi (M) 6= 0};
4 ∞ if M = IM, and otherwise, the length of any maximal M-sequence
in I.
5 Any function φ such that
a φ(M) = ∞ if M = IM.
b φ(M) = 0 if HomR (R/I, M) 6= 0.
c φ(M) = φ(M/aM) + 1 if a ∈ I is a nonzerodivisor on M.
Proof. We omit the proof, and refer the reader to [Mat, section 16], [BS,
(6.2.7)].
(2.7) For a subset F of X = Spec R, we define codim F = codimX F , the
codimension of F in X, by inf{ht P | P ∈ F }. So ht I = codim V (I) for
an ideal I of R. For M ∈ mod R, we define codim M := codim SuppR M =
ht ann M, where ann denotes the annihilator. For n ≥ 0, we denote the set
ht−1 (n) = {P
∈ Spec R | ht P = n} by Rhni . For a subset Γ of Z, RhΓi means
S
ht−1 (Γ) = n∈Γ Rhni . Moreover, we use notation such as Rh≤3i , which stands
for Rh{n∈Z|n≤3}i . For M ∈ mod R, the set of minimal primes of M is denoted
by Min M.
We define M [n] := {P ∈ Spec R | depth MP = n}. Similarly, we use
notation such as M [<n] (= {P ∈ Spec R | depth MP < n}).
(2.8) Let M, N ∈ mod R. We say that M satisfies the (SnN )R -condition or
(SnN )-condition if for any P ∈ Spec R, depthRP MP ≥ min(n, dimRP NP ). The
(SnR )R -condition or (SnR )-condition is simply denoted by (Sn′ )R or (Sn′ ). We say
that M satisfies the (Sn )R -condition or (Sn )-condition if M satisfies the (SnM )condition. (Sn ) (resp. (Sn′ )) is equivalent to say that for any P ∈ M [<n] , MP
is a Cohen–Macaulay (resp. maximal Cohen–Macaulay) RP -module. That is,
depth MP = dim MP (resp. depth MP = dim RP ). We consider that (SnN )R is
a class of modules, and also write M ∈ (SnN )R (or M ∈ (SnN )).
Lemma 2.9. Let 0 → L → M → N → 0 be an exact sequence in mod R, and
n ≥ 1.
6
1 If L, N ∈ (Sn′ ), then M ∈ (Sn′ ).
′
2 If N ∈ (Sn−1
) and M ∈ (Sn′ ), then L ∈ (Sn′ ).
Proof. 1 follows from the depth lemma:
∀P
depthRP MP ≥ min(depthRP LP , depthRP NP ),
and the fact that maximal Cohen–Macaulay modules are closed under extensions. 2 is similar.
Corollary 2.10. Let
0 → M → Ln → · · · → L1
be an exact sequence in mod R, and assume that Li ∈ (Si′ ) for 1 ≤ i ≤ n.
Then M ∈ (Sn′ ).
Proof. This is proved using a repeated use of Lemma 2.9, 2.
Lemma 2.11 (Acyclicity Lemma, [PS, (1.8)]). Let (R, m) be a Noetherian
local ring, and
(1)
∂
∂s−1
∂
s
1
L : 0 → Ls −
→
Ls−1 −−→→ · · · → L1 −
→
L0
be a complex of mod R such that
1 For each i ∈ Z with 1 ≤ i ≤ s, depth Li ≥ i.
2 For each i ∈ Z with 1 ≤ i ≤ s, Hi (L) 6= 0 implies that depth Hi (L) = 0.
Then L is acyclic (that is, Hi (L) = 0 for i > 0).
Lemma 2.12 (cf. [IW, (3.4)]). Let (1) be a complex in mod R such that
1 For each i ∈ Z with 1 ≤ i ≤ s, Li ∈ (Si′ ).
2 For each i ∈ Z with 1 ≤ i ≤ s, codim Hi (L) ≥ s − i + 1.
Then L is acyclic.
Proof. Using induction on s, we may assume that Hi (L) = 0 for i > 1. Assume
that L is not acyclic. Then H1 (L) 6= 0, and we can take P ∈ AssR H1 (L).
By assumption, ht P ≥ s. Now localize at P and considering the complex LP
over RP , we get a contradiction by Lemma 2.11.
7
Example 2.13. Let f : M → N be a map in mod R.
1 If M ∈ (S1′ ) and fP is injective for P ∈ Rh0i , then f is injective. Indeed,
consider the complex
f
0→M −
→ N = L0
and apply Lemma 2.12.
2 ([LeW, (5.11)]) If M ∈ (S2′ ), N ∈ (S1′ ), and fP is bijective for P ∈ Rh≤1i ,
then f is bijective. Consider the complex
f
0→M −
→ N → 0 = L0
this time.
Lemma 2.14. Let (R, m) be a Noetherian local ring, and N ∈ (Sn )R . If
P ∈ Min N with dim R/P < n, then we have
dim R/P = depth N = dim N < n.
If, moreover, N ∈ (Sn′ )R , then depth N = dim R.
Proof. Ischebeck proved that if M, N ∈ mod R and i < depth N − dim M,
then ExtiR (M, N) = 0 [Mat, (17.1)]. As Ext0R (R/P, N) 6= 0, we have that
depthR N ≤ dim R/P < n. The rest is easy.
Corollary 2.15. Let M ∈ (Sn )R and N ∈ (Sn′ )R . If Min M ⊂ Min N, then
M ∈ (Sn′ )R .
Proof. Let P ∈ M [<n] . As M ∈ (Sn ), depth MP = dim MP . Take Q ∈ Min M
such that Q ⊂ P and dim RP /QRP = dim MP < n. As Min M ⊂ Min N,
we have that QRP ∈ Min NP . By Lemma 2.14, dim RP = dim RP /QRP =
depth MP , and hence M ∈ (Sn′ ).
Corollary 2.16. Let n ≥ 1, and R ∈ (Sn ). Then for M ∈ mod R, we have
that (Sn′ )R = (Sn )R ∩ (S1′ ).
Proof. Obviously, (Sn′ )R ⊂ (Sn )R ∩(S1′ ). For the converse, apply Corollary 2.15
for N = R.
8
(2.17) Let M, N ∈ mod R. We say that M satisfies the (Sn′ )N -condition, or
′
M ∈ (Sn′ )N = (Sn′ )R
N , if M ∈ (Sn ) and SuppR M ⊂ SuppR N.
Lemma 2.18. Let n ≥ 1, N ∈ (Sn′ ), and M ∈ mod R. Then the following
are equivalent.
1 M ∈ (Sn′ )N .
2 M ∈ (Sn ) and Min M ⊂ Min N.
Proof. 1⇒2. As (Sn′ ) ⊂ (Sn ), M ∈ (Sn ). As M ∈ (Sn′ ) with n ≥ 1, Min M ⊂
Min R. By assumption, Min M ⊂ Supp N. So Min M ⊂ Min R ∩ Supp N ⊂
Min N.
2⇒1. M ∈ (Sn′ ) by Corollary 2.15. Supp M ⊂ Supp N follows from
Min M ⊂ Min N.
(2.19) There is another case that (Sn ) implies (Sn′ ). An R-module N is said
to be full if SuppR N = Spec R. A finitely generated faithful R-module is full.
Lemma 2.20. Let M, N ∈ mod R. If N is a full R-module, then M satisfies
(Sn′ ) condition if and only if M satisfies (SnN ) condition. If annR N ⊂ annR M,
then M satisfies the (SnN )R condition if and only if M satisfies the (Sn′ )R/ annR N
condition.
Proof. The first assertion is because dim NP = dim RP for any P ∈ Spec R.
The second assertion follows from the first, because for an R/ annR N-module,
(SnN )R and (SnN )R/ annR N are the same thing.
Lemma 2.21. Let I be an ideal of R, and S a module-finite commutative
R-algebra. For M ∈ mod S, we have that depthR (I, M) = depthS (IS, M). In
particular, if R is semilocal, then depthR M = depthS M.
i
Proof. Note that HIi (M) ∼
(M) by [BS, (4.2.1)]. By Lemma 2.6, we get
= HIS
the lemma immediately.
Lemma 2.22. Let ϕ : R → S be a finite homomorphism of rings, M ∈ mod S,
and n ≥ 0.
1 If M ∈ (Sn′ )R , then M ∈ (Sn′ )S .
2 Assume that for any Q ∈ Min S, ϕ−1 (Q) ∈ Min R (e.g., S ∈ (S1′ )R ).
If M ∈ (Sn′ )S , and RP is quasi-unmixed for any P ∈ R[<n] , then M ∈
(Sn′ )R .
9
Proof. 1. Let Q ∈ M [<n] . Then depthRP MP = depthSP MP ≤ depthSQ MQ <
n by Lemma 2.21 and Lemma 2.6, where P = ϕ−1 (Q). So MP is a maximal
Cohen–Macaulay RP -module by the (Sn′ )R -property, and hence ht Q ≤ ht P =
depthRP MP ≤ depthSQ MQ , and hence MQ is a maximal Cohen–Macaulay
SQ -module, and M ∈ (Sn′ )S .
2. Let P ∈ Spec R, and depthRP MP < n. Then by Lemma 2.21 and
Lemma 2.6, there exists some Q ∈ Spec S such that ϕ−1 (Q) = P and
depthSQ MQ =
inf
ϕ−1 (Q′ )=P
depthSQ′ MQ′ = depthSP MP = depthRP MP < n.
Then ht Q = depth RP MP . So it suffices to show ht P = ht Q. By assumption,
RP is quasi-unmixed. So RP is equi-dimensional and universally catenary
[Mat, (31.6)]. By [Gro4, (13.3.6)], ht P = ht Q, as desired.
(2.23) We say that R satisfies (Rn ) (resp. (Tn )) if RP is regular (resp. Gorenstein) for P ∈ Rh≤ni .
Lemma 2.24. Let ϕ : R → S be a flat morphism between Noetherian rings,
and M ∈ mod R.
1. If M ∈ (Sn′ )R and the ring SP /P SP satisfies (Sn ) for P ∈ Spec R, then
S ⊗R M ∈ (Sn′ )S .
2. If ϕ is faithfully flat and S ⊗R M ∈ (Sn′ )S , then M ∈ (Sn′ )R .
3. If R satisfies (Sn ) (resp. (Tn ), (Rn )) and SP /P SP satisfies (Sn ) (resp.
(Tn ), (Rn )) for P ∈ Spec R, then S satisfies (Sn ) (resp. (Tn ), (Rn )).
Proof. Left to the reader (see [Mat, (23.9)]).
3.
Xn,m -approximation
(3.1) Let A be an abelian category, and C its additive subcategory closed
under direct summands. Let n ≥ 0. We define
⊥n
C := {a ∈ A | ExtiA (a, c) = 0 1 ≤ i ≤ n}.
Let a ∈ A. A sequence
(2)
C : 0 → a → c0 → c1 → · · · → cn−1
10
is said to be an (n, C)-pushforward if it is exact with ci ∈ C. If in addition,
C† : 0 ← a† ← (c0 )† ← (c1 )† ← · · · ← (cn−1 )†
is exact for any c ∈ C, where (?)† = HomA (?, c), we say that C is a universal
(n, C)-pushforward.
If a ∈ A has an (n, C)-pushforward, we say that a is an (n, C)-syzygy, and
we write a ∈ Syz(n, C). If a ∈ A has a universal (n, C)-pushforward, we say
that a ∈ UPA (n, C) = UP(n, C). Obviously, UPA (n, C) ⊂ SyzA (n, C).
(3.2) We write Xn,m (C) = Xn,m :=
a 6= 0, we define
⊥n
C ∩ UP(m, C) for n, m ≥ 0. Also, for
Cdim a = inf{m ∈ Z≥0 | there is a resolution
0 → cm → cm−1 → · · · → c0 → a → 0}.
We define Cdim 0 = −∞. We define Yn (C) = Yn := {a ∈ A | Cdim a < n}. A
sequence E is said to be C-exact if it is exact, and A(E, c) is also exact for each
c ∈ C. Letting a C-exact sequence an exact sequence, A is an exact category,
which we denote by AC in order to distinguish it from the abelian category A
(with the usual exact sequences).
(3.3) Let C0 ⊂ A be a subset. Then ⊥n C0 , UP(n, C0 ), Xn,m (C0 ), C0 dim, and
Yn (C0 ) = Yn mean ⊥n C, UP(n, C), Xn,m (C), Cdim, and Yn (C), respectively,
where C = add C0 , the smallest additive subcategory containing C0 and closed
under direct summands. If c ∈ C, ⊥n c, UP(n, c) and so on mean ⊥n add c,
UP(n, add c) and so on. A C0 -exact sequence means an add C0 -exact sequence.
A sequence E in A is C0 -exact if and only if for any c ∈ C0 , A(E, c) is exact.
(3.4) By definition, any object of C is an injective object in AC .
(3.5) Let E be an exact category, and I an additive subcategory of E. Then
for e ∈ E, we define
PushE (n, I) := {e ∈ E | There exists an exact sequence
0 → e → c0 → c1 → · · · → cn−1 with ci ∈ I}.
Note that PushE (0, I) is the whole E. Thus PushAC (n, C) = UPA (n, C).
If a ∈ E is a direct summand of an object of I, then a ∈ Push(∞, I).
11
Lemma 3.6. Let E be an exact category. Let I be an additive subcategory of
E consisting of injective objects. Let
f
g
0→a−
→ a′ −
→ a′′ → 0
be an exact sequence in E and m ≥ 0. Then
1 If a ∈ Push(m, I) and a′′ ∈ Push(m, I), then a′ ∈ Push(m, I).
2 If a′ ∈ Push(m + 1, I) and a′′ ∈ Push(m, I), then a ∈ Push(m + 1, I).
3 If a ∈ Push(m + 1, I), a′ ∈ Push(m, I), then a′′ ∈ Push(m, I).
Proof. Let i : E ֒→ A be the Gabriel–Quillen embedding [TT]. We consider
that E is a full subcategory of A closed under extensions, and a sequence in
E is exact if and only if it is so in A.
We prove 1. We use induction on m. The case that m = 0 is trivial, and
so we assume that m > 0. Let
0→a→c→b→0
be an exact sequence such that c ∈ I and b ∈ Push(m − 1, I). Let
0 → a′′ → c′′ → b′′ → 0
be an exact sequence such that c′′ ∈ I and b′′ ∈ Push(m − 1, I). As C(a′ , c) →
C(a, c) is surjective, we can form a commutative diagram with exact rows and
columns
0
0
0
0
a
/
0
/
c
0
/
b
f
a′
/
1
0
/ c ⊕ c′′
g
(1 0)
a′′
/
c′′
/
/
b′
0
/
0
/
/
b′′
/
0
0
0
0
in A. As E is closed under extensions in A, this diagram is a diagram in E.
By induction assumption, b′ ∈ Push(m − 1, I). Hence a′ ∈ Push(m, I).
12
We prove 2. Let 0 → a′ → c → b′ → 0 be an exact sequence in E such
that c ∈ I and b′ ∈ Push(m, I). Then we have a commutative diagram in E
with exact rows and columns
0
0
0
0
/
f
a
/
/
/
a′′
/
c
/
0
/
0
/
0
a′′
g
a′
/
1c
c
/
.
0
b
/
b′
0
0
Applying 1, which we have already proved, b ∈ Push(m, I), since a′′ and b′ lie
in Push(m, I). So a ∈ Push(m + 1, I), as desired.
We prove 3. Let 0 → a → c → b → 0 be an exact sequence in E such that
c ∈ I and b ∈ Push(m, I). Taking the push-out diagram
0
0
/a
0
f
a′
/
.
g
/
a′′
/
0
/
0
1a′′
0
/
c
b
0
u
/
1b
/
a′′
/
b
0
Then u ∈ Push(m, I) by 1, which we have already proved. Since c ∈ I, the
middle row splits. Then by the exact sequence 0 → a′′ → u → c → 0 and 2,
we have that a′′ ∈ Push(m, I), as desired.
Corollary 3.7. Let E and I be as in Lemma 3.6. Let m ≥ 0, and a, a′ ∈ E.
Then a ⊕ a′ ∈ Push(m, I) if and only if a, a′ ∈ Push(m, I).
13
Proof. The ‘if’ part is obvious by Lemma 3.6, 1, considering the exact sequence
0 → a → a ⊕ a′ → a′ → 0.
(3)
We prove the ‘only if’ part by induction on m. If m = 0, then there is
nothing to prove. Let m > 0. Then by induction assumption, a′ ∈ Push(m −
1, I). Then applying Lemma 3.6, 2 to the exact sequence (3), we have that
a ∈ Push(m, I). a′ ∈ Push(m, I) is proved similarly.
Corollary 3.8. Let
f
g
0→a−
→ a′ −
→ a′′ → 0
be a C-exact sequence in A and m ≥ 0. Then
1 If a ∈ UP(m, C) and a′′ ∈ UP(m, C), then a′ ∈ UP(m, C).
2 If a′ ∈ UP(m + 1, C) and a′′ ∈ UP(m, C), then a ∈ UP(m + 1, C).
3 If a ∈ UP(m + 1, C), a′ ∈ UP(m, C), then a′′ ∈ UP(m, C).
(3.9) We define ⊥ C = ⊥∞ C :=
Obviously, C ⊂ UP(∞, C).
T
i≥0
⊥i
C and UP(∞, C) :=
T
j≥0
UP(j, C).
Lemma 3.10. We have
UP(∞, C) = {a ∈ A | There exists some C-exact sequence
0 → a → c0 → c1 → c2 → · · · with ci ∈ C for i ≥ 0}.
Proof. Let a ∈ UP(∞, C), and take any C-exact sequence
0 → a → c0 → a1 → 0
with c0 ∈ C. Then a1 ∈ UP(∞, C) by Corollary 3.8, and we can continue
infinitely.
S
(3.11) We define Y∞ := i≥0 Yi . We also define Xi,j := ⊥i C ∩ UP(j, C) for
0 ≤ i, j ≤ ∞.
(3.12) Let 0 ≤ i, j ≤ ∞. We say that a ∈ A lies in Zi,j if there is a short
exact sequence
0→y→x→a→0
in A such that x ∈ Xi,j and y ∈ Yi .
14
(3.13) We define ∞ ± r = ∞ for r ∈ R.
Lemma 3.14. Let 0 ≤ i, j ≤ ∞ with j ≥ 1. Assume that C ⊂ ⊥i+1 C. Let
f
g
0 → z −
→ x −
→ z ′ → 0 be a short exact sequence in A with z ∈ Zi,j and
x ∈ Xi+1,j−1 . Then z ′ ∈ Zi+1,j−1 .
Proof. By assumption, there is an exact sequence
j
ϕ
0→y−
→ x′ −
→z→0
such that Cdim y < i and x′ ∈ Xi,j . As j ≥ 1, there is an C-exact sequence
h
0 → x′ −
→ c → x′′′ → 0
such that c ∈ C. Then we have a commutative diagram with exact rows and
columns
0
0
0
0
hj
0
y
/
/
j h
fϕ
′
/
0
/
z
1
0
f
/x
0
0
g
/
x′′
/
(0 1)
ϕ
y′
/
c⊕x
x
/
c
/
0
/
z′
/
0
0
0
As the top row is exact, y ∈ Yi , and c ∈ C, y ′ ∈ Yi+1 . By assumption,
c ∈ Xi+1,∞ and x ∈ Xi+1,j−1. So c ⊕ x ∈ Xi+1,j−1 . As the middle row is C-exact
and x′ ∈ Xi,j , we have that x′′ ∈ Xi+1,j−1 by Corollary 3.8. The right column
shows that z ′ ∈ Zi+1,j−1 , as desired.
Lemma 3.15. Let 0 ≤ i, j ≤ ∞, and assume that i ≥ 1 and C ⊂ ⊥i C. Let
(4)
f
g
0→z−
→x−
→ z′ → 0
be a short exact sequence in A with z ′ ∈ Zi,j and x ∈ Xi,j+1 . Then z ∈ Zi−1,j+1 .
15
h
Proof. Take an exact sequence 0 → y ′ → x′′ →
− z ′ → 0 such that x′′ ∈ Xi,j
′
and y ∈ Yi . Taking the pull-back of (4) by h, we get a commutative diagram
with exact rows and columns
0
j
0
z
/
y′
/
/
1y ′
y′
/
/
0
/
a
x′′
/
1z
0
.
0
0
/
0
h
f
z
/
x
g
/
z′
0
0
0
/
0
By induction, we can prove easily that ⊥i C ⊂ ⊥i+1−l Yl . In particular, ⊥i C ⊂
⊥1
Yi , and Ext1A (x, y ′) = 0. Hence the middle column splits, and we can replace
a by x ⊕ y ′ . By the definition of Yi , there is an exact sequence
0 → y → c → y′ → 0
of A such that y ∈ Yi−1 and c ∈ C. Then adding 1x to this sequence, we get
0 → y → x ⊕ c → x ⊕ y′ → 0
is exact. Pulling back this exact sequence with j : z → a = x ⊕ y ′, we get a
commutative diagram with exact rows and columns
0
0
0
y
/
0
/
x′
/
z
1y
0
/
y
0
/
j
/
x⊕c
0
.
/
/
1x′′
x′′
x ⊕ y′
/ x′′
0
0
16
0
/
/
0
As x′′ ∈ ⊥1 C, the middle column is C-exact. As x′′ ∈ Xi,j and x ⊕ c ∈ Xi,j+1,
we have that x′ ∈ Xi−1,j+1 . As the top row shows, z ∈ Zi−1,j+1, as desired.
Theorem 3.16. Let 0 ≤ n, m ≤ ∞, and assume that C ⊂
the following are equivalent.
⊥n
C. For z ∈ A,
1 z ∈ Zn,m .
2 There is an exact sequence
(5)
dn−1
d
ε
n
0 → xn −→
xn−1 −−−→ x0 −
→z→0
such that xi ∈ Xn−i,m+i .
If, moreover, for each a ∈ A, there is a surjection x → a with x ∈ Xn,n+m ,
then these conditions are equivalent to the following.
3 For each exact sequence (5) with xi ∈ Xn−i,m+i+1 for 0 ≤ i ≤ n − 1, we
have that xn ∈ X0,n+m .
ε
Proof. 1⇒2. There is an exact sequence 0 → y → x0 −
→ z → 0 with x0 ∈ Xn,m
and y ∈ Yn . So there is an exact sequence
d
dn−1
d
n
2
0 → xn −→
xn−1 −−−→ · · · −
→
x1 → y → 0
with xi ∈ C for 1 ≤ i ≤ n. As C ⊂ Xn,∞ , we are done.
2⇒1. Let zi = Im di for i = 1, . . . , n, and z0 := z. Then by descending
induction on i, we can prove zi ∈ Zn−i,m+i for i = n, n − 1, . . . , 0, using
Lemma 3.14 easily.
1⇒3 is also proved easily, using Lemma 3.15.
3⇒2 is trivial.
4.
(n, C)-TF property
(4.1) In the rest of this paper, let Λ be a module-finite R-algebra, which
may not be commutative. A Λ-bimodule means a Λ ⊗R Λop -module. Let
C ∈ mod Λ be fixed. Set Γ := EndΛop C. Note that Γ is also a modulefinite R-algebra. We denote (?)† := HomΛop (?, C) : mod Λ → (Γ mod)op , and
(?)‡ := HomΓ (?, C) : Γ mod → (mod Λ)op .
(4.2) We denote Syzmod Λ (n, C), UPmod Λ (n, C), and Cdimmod Λ M respectively by SyzΛop (n, C), UPΛop (n, C), and CdimΛop M.
17
(4.3) Note that for M ∈ mod Λ and N ∈ Γ mod, we have standard isomorphisms
HomΛop (M, N ‡ ) ∼
= HomΓ (N, M † ).
= HomΓ⊗R Λop (N ⊗R M, C) ∼
(6)
The first isomorphism sends f : M → N ‡ to the map (n ⊗ m 7→ f (m)(n)). Its
inverse is given by g : N ⊗R M → C to (m 7→ (n 7→ g(n ⊗ m))). This shows
that (?)† has ((?)‡ )op : (Γ mod)op → mod Λ as a right adjoint. Hence ((?)† )op
is right adjoint to (?)‡ . We denote the unit of adjunction Id → (?)†‡ = (?)‡ (?)†
by λ. Note that for M ∈ mod Λ, the map λM : M → M †‡ is given by
λM (m)(ψ) = ψ(m) for m ∈ M and ψ ∈ M † = HomΛop (M, C). We denote the
unit of adjunction N → N ‡† by µ = µN for N ∈ Γ mod. When we view µ
as a morphism N ‡† → N (in the opposite category (Γ mod)op ), then it is the
counit of adjunction.
Lemma 4.4. (?)† and (?)‡ give a contravariant equivalence between add C ⊂
mod Λ and add Γ ⊂ Γ mod.
Proof. It suffices to show that λ : M → M †‡ is an isomorphism for M ∈ add C,
and µ : N → N ‡† is an isomorphism for N ∈ add Γ. To verify this, we may
assume that M = C and N = Γ. This case is trivial.
Definition 4.5 (cf. [Tak, (2.2)]). Let M ∈ mod Λ. We say that M is (1, C)-TF
or M ∈ TFΛop (1, C) if λM : M → M †‡ is injective. We say that M is (2, C)-TF
or M ∈ TFΛop (2, C) if λM : M → M †‡ is bijective. Let n ≥ 3. We say that M
is (n, C)-TF or M ∈ TFΛop (n, C) if M is (2, C)-TF and ExtiΓ (M † , C) = 0 for
1 ≤ i ≤ n − 2. As a convention, we define that any M ∈ mod Λ is (0, C)-TF.
Lemma 4.6. Let Θ : 0 → M → L → N → 0 be a C-exact sequence in mod Λ.
Then for n ≥ 0, we have the following.
1 If M ∈ TF(n, C) and N ∈ TF(n, C), then L ∈ TF(n, C).
2 If L ∈ TF(n + 1, C) and N ∈ TF(n, C), then M ∈ TF(n + 1, C).
3 If M ∈ TF(n + 1, C) and L ∈ TF(n, C), then N ∈ TF(n, C).
Proof. We have a commutative diagram
0
/
M
h
/
L
λM
0
/
M †‡
/
N
λL
h†‡
L†‡
0
λN
/
/
/
N †‡
/
Ext1Γ (M † , C)
18
/ Ext1 (L† , C)
Γ
/
···
with exact rows.
We only prove 3. We may assume that n ≥ 1. So λM is an isomorphism
and λL is injective. By the five lemma, λN is injective, and the case that n = 1
has been done. If n ≥ 2, then λL is also an isomorphism and Ext1Γ (M † , C) = 0,
and so λN is an isomorphism. Moreover, for 1 ≤ i ≤ n − 2, ExtiΓ (L† , C) and
i
†
†
Exti+1
Γ (M , C) vanish. so ExtΓ (N , C) = 0 for 1 ≤ i ≤ n − 2, and hence
N ∈ TF(n, C).
1 and 2 are also proved similarly.
Lemma 4.7 (cf. [Tak, Proposition 3.2]).
UPΛop (n, C).
1 For n = 0, 1, SyzΛop (n, C) =
2 For n ≥ 0, TFΛop (n, C) = UPΛop (n, C).
Proof. If n = 0, then SyzΛop (n, C) = TFΛop (0, C) = UPΛop (0, C) = mod Λ. So
we may assume that n ≥ 1.
Let M ∈ SyzΛop (1, C). Then there is an injection ϕ : M → N with
N ∈ add C. Then
ϕ /
M
N
∼
= λN
λM
M †‡
ϕ†‡
/
N †‡
is a commutative diagram. So λM is injective, and M ∈ TFΛop (1, C). This
shows UPΛop (1, C) ⊂ SyzΛop (1, C) ⊂ TFΛop (1, C). So 2⇒1.
We prove 2. First, we prove UPΛop (n, C) ⊂ TFΛop (n, C) for n ≥ 1. We
use induction on n. The case n = 1 is already done above.
Let n ≥ 2 and M ∈ UPΛop (n, C). Then by the definition of UPΛop (n, C),
there is a C-exact sequence
0→M →L→N →0
such that L ∈ add C and N ∈ UPΛop (n − 1, C). By induction hypothesis,
N ∈ TFΛop (n − 1, C). Hence M ∈ TFΛop (n, C) by Lemma 4.6. We have
proved that UPΛop (n, C) ⊂ TFΛop (n, C).
Next we show that TFΛop (n, C) ⊂ UPΛop (n, C) for n ≥ 1. We use induction
on n.
Let n = 1. Let ρ : F → M † be any surjective Γ-linear map with F ∈ add Γ.
Then the map ρ′ : M → F ‡ which corresponds to ρ by the adjunction (6) is
ρ‡
λ
→ M †‡ −
ρ′ : M −−M
→ F ‡,
19
which is injective by assumption. Then ρ is the composite
(ρ′ )†
µF
ρ : F −→ F ‡† −−→ M † ,
which is a surjective map by assumption. So (ρ′ )† is also surjective, and hence
ρ′ : M → F ‡ gives a (1, C)-universal pushforward.
Now let n ≥ 2. By what we have proved, M has a (1, C)-universal pushforward h : M → L. Let N = Coker h. Then we have a C-exact sequence
0→M →L→N →0
with L ∈ add C. As M ∈ TF(n, C), N ∈ TF(n − 1, C) by Lemma 4.6. By
induction assumption, N ∈ UP(n − 1, C). So by the definition of UP(n, C),
we have that M ∈ UP(n, C), as desired.
Lemma 4.8. For any N ∈ Γ mod, N ‡ ∈ Syz(2, C).
Proof. Let
h
F1 −
→ F0 → N → 0
be an exact sequence in Γ mod such that Fi ∈ add Γ. Then
h‡
0 → N ‡ → F0‡ −→ F1‡
is exact, and Fi‡ ∈ add C. This shows that N ‡ ∈ Syz(2, C).
op
(4.9) We denote by (Sn′ )C = (Sn′ )ΛC ,R the class of M ∈ mod Λ such that M
viewed as an R-module lies in (Sn′ )R
C , see (2.17).
Lemma 4.10. Assume that C satisfies (Sn′ ) as an R-module. Then Syz(r, C) ⊂
op
(Sr′ )ΛC ,R for r ≥ 1.
Proof. This follows easily from Corollary 2.10.
(4.11) For an additive category C and its additive subcategory X , we denote
by C/X the quotient of C divided by the ideal consisting of morphisms which
factor through objects of X .
20
(4.12) For each M ∈ mod Λ, take a presentation
∂
ε
F(M) : F1 (M) −
→ F0 (M) −
→M →0
(7)
with Fi ∈ add ΛΛ . We denote
Coker(∂ † ) = Coker(1C ⊗ ∂ t ) = C ⊗Λ Tr M
by TrC M, where (?)t = HomΛop (?, Λ) and Tr is the transpose, see [ASS,
(V.2)], and we call it the C-transpose of M. TrC is an additive functor from
mod Λ := mod Λ/ add ΛΛ to ΓC mod := Γ mod / add C.
Proposition 4.13. Let n ≥ 0, and assume that ExtiΓ (C, C) = 0 for i =
1, . . . , n. Then for M ∈ mod Λ, we have the following.
0 For 1 ≤ i ≤ n, ExtiΓ (TrC ?, C) is a well-defined additive functor mod Λ →
mod Λ.
1 If n = 1, there is an exact sequence
λ
0 → Ext1Γ (TrC M, C) → M −−M
→ M †‡ → Ext2Γ (TrC M, C).
If n = 0, then there is an injective homomorphism Ker λM ֒→ Ext1Γ (TrC M, C).
2 If n ≥ 2, then
i There is an exact sequence
λ
→ M †‡ → Ext2Γ (TrC M, C) → 0.
0 → Ext1Γ (TrC M, C) → M −−M
ii There are isomorphisms Exti+2 Γ(TrC M, C) ∼
= ExtiΓ (M † , C) for
1 ≤ i ≤ n − 2.
iii There is an injective map ExtΓn−1 (M † , C) ֒→ Extn+1
Γ (TrC M, C).
Proof. 0 is obvious by assumption.
We consider that F(M) is a complex with M at degree zero. Then consider
∂†
ε†
Q(M) := F(M)† [2] : F1 (M)† ←− F0 (M)† ←
− M† ← 0
where F1 (M)† is at degree zero. As this complex is quasi-isomorphic to
TrC (M), there is a spectral sequence
E1p,q = ExtqΓ (Q(M)−p , C) ⇒ Extp+q
Γ (TrC M, C).
21
1,0
⊂ E 1 . If n ≥ 1, then E10,1 = 0, and
In general, Ker λM = E21,0 ∼
= E∞
2,0
2,0
1,0
⊂ E 2 . So 1
E∞
= E 1 . Moreover, as E10,1 = 0, Coker λM ∼
= E∞
= E2 ∼
follows.
2,0
If n ≥ 2, then E10,2 = E11,1 = 0 by assumption, so E∞
= E 2 , and i of 2
p,q
p,q
follows. Note that E1 = 0 for p ≥ 3. Moreover, E1 = 0 for p = 0, 1 and
1 ≤ q ≤ n. So for 1 ≤ i ≤ n − 1, we have
2,i
֒→ E i+2 ,
E12,i ∼
= E∞
and the inclusion is an isomorphism if 1 ≤ i ≤ n − 2. So ii and iii of 2
follow.
Corollary 4.14. Let n ≥ 1. If ExtiΓ (C, C) = 0 for 1 ≤ i ≤ n, then M is
(n, C)-TF if and only if ExtiΓ (TrC M, C) = 0 for 1 ≤ i ≤ n. If ExtiΓ (C, C) = 0
for 1 ≤ i < n and ExtiΓ (TrC M, C) = 0 for 1 ≤ i ≤ n, then M is (n, C)-TF.
5.
Canonical module
(5.1) Let R = (R, m) be semilocal, where m is the Jacobson radical of R.
(5.2) We say that a dualizing complex I over R is normalized if for any
maximal ideal n of R, Ext0R (R/n, I) 6= 0. We follow the definition of [Hart2].
(5.3) For a left or right Λ-module M, dim M or dimΛ M denotes the dimension dimR M of M, which is independent of the choice of R. We call
depthR (m, M), which is also independent of R, the global depth, Λ-depth, or
depth of M, and denote it by depthΛ M or depth M. M is called globally
Cohen–Macaulay or GCM for short, if dim M = depth M. M is GCM if and
only if it is Cohen–Macaulay as an R-module, and all the maximal ideals of R
have the same height. This notion is independent of R, and depends only on Λ
and M. M is called a globally maximal Cohen–Macaulay (GMCM for short)
if dim Λ = depth M. We say that the algebra Λ is GCM if the Λ-module Λ
is GCM. However, in what follows, if R happens to be local, then GCM and
Cohen–Macaulay (resp. GMSM and maximal Cohen–Macaulay) (over R) are
the same thing, and used interchangeably.
(5.4) Assume that (R, m) is complete semilocal, and Λ 6= 0. Let I be a
normalized dualizing complex of R. The lowest non-vanishing cohomology
i
group Ext−s
R (Λ, I) (ExtR (Λ, I) = 0 for i < −s) is denoted by KΛ , and is called
the canonical module of Λ. Note that KΛ is a Λ-bimodule. Hence it is also a
Λop -bimodule. In this sense, KΛ = KΛop . If Λ = 0, then we define KΛ = 0.
22
(5.5) Let S be the center of Λ. Then S is module-finite over R, and
IS = RHomR (S, I) is a normalized dualizing complex of S. This shows that
RHomR (Λ, I) ∼
= RHomS (Λ, IS ), and hence the definition of KΛ is also independent of R.
Lemma 5.6. The number s in (5.4) is nothing but d := dim Λ. Moreover,
AssR KΛ = AsshR Λ := {P ∈ MinR Λ | dim R/P = dim Λ}.
Proof. We may replace R by R/ annR Λ, and may assume that Λ is a faithful
module. We may assume that I is a fundamental dualizing complex of R.
That is, for each P ∈ Spec R, E(R/P ), the injective hull of R/P , appears
exactly once (at dimension − dim R/P ). If Ext−i
R (Λ, I) 6= 0, then there exists
−i
some P ∈ Spec R such that ExtRP (ΛP , IP ) 6= 0. Then P ∈ SuppR Λ and
dim R/P ≥ i. On the other hand, Ext−d
RP (ΛP , IP ) has length l(ΛP ) and is
nonzero for P ∈ AsshR Λ. So s = d.
The argument above shows that each P ∈ AsshR Λ = Assh R supports KΛ .
So AsshR Λ ⊂ MinR KΛ . On the other hand, as the complex I starts at degree
−d, KΛ ⊂ I−d , and Ass KΛ ⊂ Ass I−d ⊂ Assh R = AsshR Λ.
Lemma 5.7. Let (R, m) be complete semilocal. Then KΛ satisfies the (S2Λ )R condition.
Proof. It is easy to see that (KΛ )n is either zero or KΛn for each maximal ideal
n of R. Hence we may assume that R is local. Replacing R by R/ annR Λ, we
may assume that Λ is a faithful R-module, and we are to prove that KΛ satisfies
(S2′ )R by Lemma 2.20. Replacing R by a Noether normalization, we may
further assume that R is regular by Lemma 2.22, 1. Then KΛ = HomR (Λ, R).
So KΛ ∈ Syz(2, R) ⊂ (S2′ )R by Lemma 4.8 (consider that Λ there is R here,
and C there is also R here).
(5.8) Assume that (R, m) is semilocal which may not be complete. We say
that a finitely generated Λ-bimodule K is a canonical module of Λ if K̂ is
isomorphic to the canonical module KΛ̂ as a Λ̂-bimodule. It is unique up
to isomorphisms, and denoted by KΛ . We say that K ∈ mod Λ is a right
canonical module of Λ if K̂ is isomorphic to KΛ̂ in mod Λ̂, where ˆ? is the
m-adic completion. If KΛ exists, then K is a right canonical module if and
only if K ∼
= KΛ in mod Λ.
These definitions are independent of R, in the sense that the (right) canonical module over R and that over the center of Λ are the same thing. The right
23
canonical module of Λop is called the left canonical module. A Λ-bimodule ω
is said to be a weakly canonical bimodule if Λ ω is left canonical, and ωΛ is
right canonical. The canonical module KΛop of Λop is canonically identified
with KΛ .
(5.9) If R has a normalized dualizing complex I, then Î is a normalized
dualizing complex of R̂, and so it is easy to see that KΛ exists and agrees
with Ext−d (Λ, I), where d = dim Λ(:= dimR Λ). In this case, for any P ∈
Spec R, IP is a dualizing complex of RP . So if R has a dualizing complex and
(KΛ )P 6= 0, then (KΛ )P , which is the lowest nonzero cohomology group of
RHomRP (ΛP , IP ), is the RP -canonical module of ΛP . See also Theorem 7.5
below.
Lemma 5.10. Let (R, m) be local, and assume that KΛ exists. Then we have
the following.
1 AssR KΛ = AsshR Λ.
2 KΛ ∈ (S2Λ )R .
3 R/ ann KΛ is quasi-unmixed, and hence is universally catenary.
Proof. All the assertions are proved easily using the case that R is complete.
(5.11) A Λ-module M is said to be Λ-full over R if SuppR M = SuppR Λ.
Lemma 5.12. Let (R, m) be local. If KΛ exists and Λ satisfies the (S2 )R condition, then R/ ann KΛ is equidimensional, and KΛ is Λ-full over R.
Proof. The same as the proof of [Ogo, Lemma 4.1] (use Lemma 5.10, 3).
(5.13) Let (R, m) be local, and I be a normalized dualizing complex. By
the local duality,
KΛ∨ = Ext−d (Λ, I)∨ ∼
= Hmd (Λ)
(as Λ-bimodules), where ER (R/m) is the injective hull of the R-module R/m,
and (?)∨ is the Matlis dual HomR (?, ER (R/m)).
24
(5.14) Let (R, m) be semilocal, and I be a normalized dualizing complex.
Note that RHomR (?, I) induces a contravariant equivalence between Dfg (Λop )
and Dfg (Λ). Let J ∈ Dfg (Λ ⊗R Λop ) be RHomR (Λ, I).
RHomR (?, I) : Dfg (Λop ) → Dfg (Λ)
is identified with
RHomΛop (?, RHomR (Λ ΛR , I)) = RHomΛop (?, J)
and similarly,
RHomR (?, I) : Dfg (Λ) → Dfg (Λop )
is identified with RHomΛ (?, J). Note that a left or right Λ-module M is maximal Cohen–Macaulay if and only if RHomR (M, I) is concentrated in degree
−d, where d = dim Λ.
(5.15) J above is a dualizing complex of Λ in the sense of Yekutieli [Yek,
(3.3)].
(5.16) Λ is GCM if and only if KΛ [d] → J is an isomorphism. If so, M ∈
mod Λ is GMCM if and only if RHomR (M, I) is concentrated in degree −d
if and only if ExtiΛop (M, KΛ ) = 0 for i > 0. Also, in this case, as KΛ [d] is a
dualizing complex, it is of finite injective dimension both as a left and a right
Λ-module. To prove these, we may take the completion, and may assume that
R is complete. All the assertions are independent of R, so taking the Noether
normalization, we may assume that R is local. By (5.14), the assertions follow.
(5.17) For any M ∈ mod Λ which is GMCM,
M∼
= RHomR (Ext−d
= RHomR (RHomR (M, I), I) ∼
Λop (M, KΛ [d]), I)[−d].
Hence M † := HomΛop (M, KΛ ) is also a GMCM Λ-module, and hence
HomΛ (M † , KΛ ) → RHomΛ (M † , J) = RHomR (M † , I)
is an isomorphism (in other words, ExtiΛ (M † , KΛ ) = 0 for i > 0). So the
canonical map
(8)
M → HomΛ (HomΛop (M, KΛ ), KΛ ) = HomΛ (M † , KΛ )
m 7→ (ϕ 7→ ϕm) is an isomorphism. This isomorphism is true without assuming that R has a dualizing complex (but assuming the existence of a canonical
25
module), passing to the completion. Note that if Λ = R and KR exists and
Cohen–Macaulay, then KR is a dualizing complex of R.
Similarly, for N ∈ Λ mod which is GMCM,
N → HomΛop (HomΛ (N, KΛ ), KΛ )
n 7→ (ϕ 7→ ϕn) is an isomorphism.
(5.18) In particular, letting M = Λ, if Λ is GCM, we have that KΛ =
HomΛop (Λ, KΛ ) is GMCM. Moreover,
Λ → EndΛop KΛ
is an R-algebra isomorphism, where a ∈ Λ goes to the left multiplication by
a. Similarly,
Λ → (EndΛ KΛ )op
is an isomorphism of R-algebras.
(5.19) Let (R, m) be a d-dimensional complete local ring, and dim Λ = d.
Then by the local duality,
−d
∼
∼
Hmd (KΛ )∨ ∼
= Ext−d
R (KΛ , I) = ExtΛop (KΛ , J) = EndΛop KΛ ,
where J = HomR (Λ, I) and (?)∨ = HomR (?, ER (R/m)).
6.
n-canonical module
(6.1) We say that ω is an R-semicanonical right Λ-module (resp. R-semicanonical
left Λ-module, weakly R-semicanonical Λ-bimodule, R-semicanonical Λ-bimodule)
if for any P ∈ Spec R, RP ⊗R ω is the right canonical module (resp. left
canonical module, weakly canonical module, canonical module) of RP ⊗R
Λ for any P ∈ suppR ω. If we do not mention what R is, then it may
mean R is the center of Λ. An R-semicanonical right Λop -module (resp.
R-semicanonical left Λop -module, weakly R-semicanonical Λop -bimodule, Rsemicanonical Λop -bimodule) is nothing but an R-semicanonical left Λ-module
(resp. R-semicanonical right Λ-module, weakly R-semicanonical Λ-bimodule,
R-semicanonical Λ-bimodule).
26
(6.2) Let C ∈ mod Λ (resp. Λ mod, (Λ ⊗R Λop ) mod, (Λ ⊗R Λop ) mod). We
say that C is an n-canonical right Λ-module (resp. n-canonical left Λ-module,
weakly n-canonical Λ-bimodule, n-canonical Λ-bimodule) over R if C ∈ (Sn′ )R ,
and for each P ∈ Rh<ni , we have that CP is an RP -semicanonical right ΛP module (resp. RP -semicanonical left ΛP -module, weakly RP -semicanonical
ΛP -bimodule, RP -semicanonical ΛP -bimodule). If we do not mention what
R is, it may mean R is the center of Λ.
Example 6.3.
0 The zero module 0 is an R-semicanonical Λ-bimodule.
1 If R has a dualizing complex I, then the lowest non-vanishing cohomology group K := Ext−s
R (Λ, I) is an R-semicanonical Λ-bimodule.
2 By Lemma 5.10, any left or right R-semicanonical module K of Λ satisfies the (S2Λ )R -condition. Thus a (right) semicanonical module is 2canonical over R/ annR Λ.
3 If K is (right) semicanonical (resp. n-canonical) and L is a projective
R-module such that LP is rank at most one, then K ⊗R L is again (right)
semicanonical (resp. n-canonical).
4 If R is a normal domain and C its rank-one reflexive module of R, then
C is a 2-canonical R-module (here Λ = R).
5 The R-module R is n-canonical if and only if for P ∈ R[<n] , RP is
Gorenstein. This is equivalent to say that R satisfies (Tn−1 ) + (Sn ).
(6.4) As in section 4, let C ∈ mod Λ, and set Γ = EndΛop C, (?)† =
HomΛop (?, C), and (?)‡ = HomΓ (?, C). Moreover, we set Λ1 := (EndΓ C)op .
The R-algebra map Ψ1 : Λ → Λ1 is induced by the right action of Λ on C.
Lemma 6.5. Let C ∈ mod Λ be a 1-canonical Λop -module over R. Let M ∈
mod Λ. Then the following are equivalent.
1 M ∈ TF(1, C).
2 M ∈ UP(1, C).
3 M ∈ Syz(1, C).
4 M ∈ (S1′ )R
C.
27
Proof. 1⇔2 is Lemma 4.7. 2⇒3 is trivial. 3⇒4 follows from Lemma 4.10
immediately.
We prove 4⇒1. We want to prove that λM : M → M †‡ is injective. By
Example 2.13, localizing at each P ∈ Rh0i , we may assume that (R, m) is
zero-dimensional local. We may assume that M is nonzero. By assumption,
C is nonzero, and hence C = KΛ by assumption. As R is zero-dimensional,
Λ is GCM, and hence Λ → Γ = EndΛop KΛ is an isomorphism by (5.18). As
Λ is GCM and M is GMCM, (8) is an isomorphism. As Λ = Γ, the result
follows.
Lemma 6.6. Let C be a 1-canonical right Λ-module over R, and N ∈ Γ mod.
Then N ‡ ∈ TFΛop (2, C). Similarly, for M ∈ mod Λ, we have that M † ∈
TFΓ (2, C).
Proof. Note that λN ‡ : N ‡ → N ‡†‡ is a split monomorphism. Indeed, (µN )‡ :
N ‡†‡ → N ‡ is the left inverse. Assume that N ‡ ∈
/ TF(2, C), then W :=
Coker λN ‡ is nonzero. Let P ∈ AssR W . As W is a submodule of N ‡†‡ ,
P ∈ AssR N ‡†‡ ⊂ AssR C ⊂ Min R. So CP is the right canonical module KΛP .
So ΓP = ΛP , and (λN ‡ )P is an isomorphism. This shows that WP = 0, and
this is a contradiction. The second assertion is proved similarly.
Lemma 6.7. Let (R, m) be local, and assume that KΛ exists. Let C := KΛ .
If Λ is GCM, Ψ1 : Λ → Λ1 is an isomorphism.
Proof. As C possesses a bimodule structure, we have a canonical map Λ →
Γ = EndΛop C, which is an isomorphism as Λ is GCM by (5.18). So Λ1 is identified with ∆ = (EndΛ C)op . Then Ψ1 : Λ → (EndΛ C)op is an isomorphism
again by (5.18).
Lemma 6.8. If C satisfies the (S1′ )R condition, then Γ ∈ (S1′ )R
C and Λ1 ∈
(S1′ )R
.
Moreover,
Ass
Γ
=
Ass
Λ
=
Ass
C
=
Min
C.
R
R 1
R
R
C
Proof. The first assertion is by Γ = HomΛop (C, C) ∈ SyzΓ (2, C), and Λ1 =
HomΓ (C, C) = SyzΛ1 (2, C). We prove the second assertion. AssR Γ ⊂ AssR EndR C =
AssR C. AssR Λ1 ⊂ AssR EndR C = AssR C = MinR C. It remains to show
that SuppR C = SuppR Γ = SuppR Λ1 . Let P ∈ Spec R. If CP = 0, then
ΓP = 0 and (Λ1 )P = 0. On the other hand, if CP 6= 0, then the identity map
CP → CP is not zero, and hence ΓP 6= 0 and (Λ1 )P 6= 0.
28
Q
(6.9) Let C be a 1-canonical right Λ-module over R. Define Q := P ∈MinR C RP .
If P ∈ MinR C, then CP = KΛP . Hence ΦP : ΛP → (Λ1 )P is an isomorphism
by Lemma 6.7. So 1Q ⊗ Ψ1 : Q ⊗R Λ → Q ⊗R Λ1 is also an isomorphism. As
AssR Λ1 = MinR C, we have that Λ1 ⊂ Q ⊗R Λ1 .
Lemma 6.10. Let C be a 1-canonical right Λ-module over R. If Λ is commutative, then so are Λ1 and Γ.
Proof. As Λ1 ⊂ Q ⊗R Λ1 = Q ⊗R Λ and Q ⊗R Λ is commutative, Λ1 is a
commutative ring. We prove that Γ is commutative. As AssR Γ ⊂ MinR C, Γ
is a subring of Q ⊗ Γ. As
Y
Y
EndΛP (KΛP )
EndΛP CP ∼
Q ⊗R Γ ∼
=
=
P
P ∈MinR C
and ΛP → EndΛP (KΛP ) is an isomorphism (as ΛP is zero-dimensional), Q⊗R Γ
is, and hence Γ is also, commutative.
Lemma 6.11. Let C be a 1-canonical right Λ-module over R. Let M and N
be left (resp. right, bi-) modules of Λ1 , and assume that N ∈ (S1′ )Λ1 ,R . Let
ϕ : M → N be a Λ-homomorphism of left (resp. right, bi-) modules. Then ϕ
is a Λ1 -homomorphism of left (resp. right, bi-) modules.
Q
Proof. Let Q = P ∈MinR C RP . Then we have a commutative diagram
M
ϕ
/
N
iM
Q ⊗R M
,
iN
1⊗ϕ
/
Q ⊗R N
where iM (m) = 1 ⊗ m and iN (n) = 1 ⊗ n. Clearly, iM and iN are Λ1 -linear.
As ϕ is Λ-linear, 1 ⊗ ϕ is Q ⊗ Λ-linear. Since Λ1 ⊂ Q ⊗ Λ1 = Q ⊗ Λ, 1 ⊗ ϕ is
Λ1 -linear. As iN is injective, it is easy to see that ϕ is Λ1 -linear.
Lemma 6.12. Let C be a 1-canonical right Λ-module over R. Then the restriction M 7→ M is a full and faithful functor from (S1′ )Λ1 ,R to (S1′ )Λ,R
C . Simiop ,R
op
Λ
′ Λop
,R
′
larly, it gives a full and faithful functors (S1 ) 1 → (S1 )C
and (S1′ )Λ1 ⊗R Λ1 ,R →
op
R Λ ,R
(S1′ )Λ⊗
.
C
Proof. We only consider the case of left modules. If M ∈ Λ1 mod, then it is a
homomorphic image of Λ1 ⊗R M. Hence suppR M ⊂ suppR Λ1 ⊂ suppR C. So
the functor is well-defined and obviously faithful. By Lemma 6.11, it is also
full, and we are done.
29
(6.13) Let C be a 1-canonical Λ-bimodule over R. Then the left action
of
Q Λ on C induces an R-algebra map Φ : Λ → Γ = EndΛop C. Let Q =
P ∈MinR C RP . Then Γ ⊂ Q ⊗R Γ = Q ⊗R Λ. From this we get
Lemma 6.14. Let C be a 1-canonical Λ-bimodule over R. Let M and N
be left (resp. right, bi-) modules of Γ, and assume that N ∈ (S1′ )Γ,R . Let
ϕ : M → N be a Λ-homomorphism of left (resp. right, bi-) modules. Then ϕ
is a Γ-homomorphism of left (resp. right, bi-) modules.
Proof. Similar to Lemma 6.11, and left to the reader.
Corollary 6.15. Let C be as above. (?)†‡ = HomΓ (HomΛop (?, C), C) is canonically isomorphic to (?)†⋆ = HomΛ (HomΛop (?, C), C), where (?)⋆ = HomΛ (?, C).
Proof. This is immediate by Lemma 6.14.
Lemma 6.16. Let C be a 1-canonical Λ-bimodule over R. Then Φ induces a
op
′ Λ,R
′ Γop ,R
′ Λ ,R
full and faithful functor (S1′ )Γ,R
→
(S
)
.
Similarly,
(S
)
→
(S
)
1
1
1
C
C
op
op
R Λ ,R
and (S1′ )Γ⊗R Γ ,R → (S1′ )Λ⊗
are also induced.
C
Proof. Similar to Lemma 6.12, and left to the reader.
Corollary 6.17. Let C be a 1-canonical Λ-bimodule. Set ∆ := (EndΛ C)op .
Then the canonical map Λ → Γ induces an equality
Λ1 = (EndΓ C)op = (EndΛ C)op = ∆.
Similarly, we have
Λ2 := End∆op C = EndΛop C = Γ.
Proof. As C ∈ (S1′ )Γ,R , the first assertion follows from Lemma 6.16. The
second assertion is proved by left-right symmetry.
Lemma 6.18. Let C be a 1-canonical right Λ-module over R. Set Λ1 :=
(EndΓ C)op . Let Ψ1 : Λ → Λ1 be the canonical map induced by the right action
of Λ on C. Then Ψ1 is injective if and only if Λ satisfies the (S1′ )R condition
and C is Λ-full over R.
Proof. Ψ1 : Λ → Λ1 is nothing but λΛ : Λ → Λ†‡ , and the result follows from
Lemma 6.5 immediately.
Lemma 6.19. Let C be a 1-canonical Λ-bimodule over R. Then the following
are equivalent.
30
1 The canonical map Ψ : Λ → ∆ is injective, where ∆ = (EndΛ C)op , and
the map is induced by the right action of Λ on C.
2 Λ satisfies the (S1′ )R condition, and C is Λ-full over R.
3 The canonical map Φ : Λ → Γ is injective, where the map is induced by
the left action of Λ on C.
Proof. By Corollary 6.17, we have that Λ1 = (EndΓ C)op = ∆. So 1⇔2 is a
consequence of Lemma 6.18.
Reversing the roles of the left and the right, we get 2⇔3 immediately.
Lemma 6.20. Let C be a 1-canonical right Λ-module over R. Then the canonical map
(9)
HomΛop (Λ1 , C) → HomΛop (Λ, C) ∼
=C
induced by the canonical map Ψ1 : Λ → Λ1 is an isomorphism of Γ ⊗R Λop
1 modules.
Proof. The composite map
C∼
=C
= HomΛ1 (Λ1 , C) = HomΛ (Λ1 , C) → HomΛ (Λ, C) ∼
is the identity. The map is a Γ ⊗R Λop -homomorphism. It is also Λop
1 -linear
by Lemma 6.12.
(6.21) When (R, m) is local and C = KΛ , then Λ1 = ∆, and the map
(9) is an isomorphism of Γ ⊗R ∆op -modules from K∆ and KΛ , where ∆ =
(EndΛ KΛ )op . Indeed, to verify this, we may assume that R is complete regular
local with annR Λ = 0, and hence C = HomR (Λ, R), and C is a 2-canonical
Λ-bimodule over R, see (6.3). So (6.17) and Lemma 6.20 apply. Hence we
have
Corollary 6.22. Let (R, m) be a local ring with a canonical module C = KΛ
of Λ. Then K∆ = HomΛop (∆, KΛ ) is isomorphic to KΛ as a Γ ⊗R ∆op -module,
where ∆ = (EndΛ KΛ )op .
Lemma 6.23. Let n ≥ 1. If C is an n-canonical right Λ-module over R, then
1 C is an n-canonical right Λ1 -module over R.
2 C is an n-canonical left Γ-module over R.
31
Proof. 1. As the (Sn′ )-condition holds, it suffices to prove that for P ∈ Rh<ni ,
CP ∼
= (KΛ1 )P as a right (Λ1 )P -module. After localization, replacing R by RP ,
we may assume that R is local and C = KΛ . Then C ∼
= KΛ1 as right
= KΛ ∼
′ Λop
,R
Λ-modules. Both C and KΛ1 are in (S1 ) 1 , and isomorphic in mod Λ. So
they are isomorphic in mod Λ1 by Lemma 6.12.
2. Similarly, assuming that R is local and C = KΛ , it suffices to show that
∼
C = KΓ as left Γ-modules. Identifying Γ = End∆op C = Λ2 and using the
left-right symmetry, this is the same as the proof of 1.
Lemma 6.24. Let C ∈ mod Λ be a 2-canonical right Λ-module over R. Let
M ∈ mod Λ. Then the following are equivalent.
1 M ∈ TF(2, C).
2 M ∈ UP(2, C).
3 M ∈ Syz(2, C).
4 M ∈ (S2′ )R
C.
Proof. We may assume that Λ is a faithful R-module. 1⇔2⇒3⇒4 is easy. We
show 4⇒1. By Example 2.13, localizing at each P ∈ Rh≤1i , we may assume
that R is a Noetherian local ring of dimension at most one. So the formal
, where ˆ? denotes the
fibers of R are zero-dimensional, and hence M̂ ∈ (S2′ )R̂
Ĉ
completion. So we may further assume that R = (R, m) is complete local.
We may assume that M 6= 0 so that C 6= 0 and hence C = KΛ . The case
dim R = 0 is similar to the proof of Lemma 6.5, so we prove the case that
dim R = 1. Note that I = Hm0 (Λ) is a two-sided ideal of Λ, and any module
op
in (S1′ )Λ ,R is annihilated by I. Replacing Λ by Λ/I, we may assume that Λ
is a maximal Cohen–Macaulay R-module. Then (8) is an isomorphism. As
C = KΛ and
Λ → EndΛop KΛ = EndΛop C = Γ
is an R-algebra isomorphism, we have that λM : M → M †‡ is identified with
the isomorphism (8), as desired.
Corollary 6.25. Let C be a 2-canonical right Λ-module over R. Then the
canonical map Φ : Λ → Λ1 is an isomorphism if and only if Λ satisfies (S2′ )R
and C is full.
Proof. Follows immediately by Lemma 6.24 applied to M = Λ.
32
(6.26) Let C be a 2-canonical Λ-bimodule. Let Γ = EndΛop C and ∆ =
(EndΛ C)op . Then by the left multiplication, an R-algebra map Λ → Γ is induced, while
Q by the right multiplication, an R-algebra map Λ → ∆ is induced.
Let Q = P ∈MinR C RP . Then as Γ ⊂ Q ⊗R Γ = Q ⊗R Λ = Q ⊗R ∆ ⊃ ∆, both
Γ and ∆ are identified with Q-subalgebras of Q ⊗R Λ. As ∆ = Λ1 = Λ†‡ , we
have a commutative diagram
Λ
λΛ
Λ†‡
/
ν
Γ
λΓ
/
=∆.
ν †‡
Γ†‡
As Γ = HomΛop (C, C) = C † , Γ ∈ SyzΛ (2, C) by Lemma 4.8. By Lemma 6.24,
we have that Γ ∈ (S2′ )C . Hence by Lemma 6.24 again, λΓ : Γ → Γ†‡ is an
isomorphism. Hence ∆ ⊂ Γ. By symmetry ∆ ⊃ Γ. So ∆ = Γ. With this
identification, Γ acts on C not only from left, but also from right. As the
actions of Γ extend those of Λ, C is a Γ-bimodule. Indeed, for a ∈ Λ, the
left multiplication λa : C → C (λa (c) = ac) is right Γ-linear. So for b ∈ Γ,
ρb : C → C (ρb (c) = cb) is left Λ-linear, and hence is left Γ-linear.
Theorem 6.27. Let C be a 2-canonical right Λ-module. Then the restriction
op
Λop ,R
M 7→ M gives an equivalence ρ : (S2′ )Λ1 ,R → (S2′ )C
.
Proof. The functor is obviously well-defined,
and is full and faithful by Lemma 6.12.
op
′ Λ ,R
On the other hand, given M ∈ (S2 )C , we have that λM : M → M †‡ is an
op
isomorphism. As M †‡ has a Λop
1 -module structure which extends the Λ †‡
∼
module structure of M = M , we have that ρ is also dense, and hence is an
equivalence.
Corollary 6.28. Let C be a 2-canonical Λ-bimodule. Then the restriction
M 7→ M gives an equivalence
op ,R
RΓ
ρ : (S2′ )Γ⊗
C
RΛ
→ (S2′ )Λ⊗
C
op ,R
.
Proof. ρ is well-defined, and is obviously faithful. If h : M → N is a morphism
op
op
R Λ ,R
R Γ ,R
of (S2 )Λ⊗
between objects of (S2 )Γ⊗
, then h is Γ-linear Γop -linear
C
C
by Theorem 6.27 (note that Λ1 = ∆ = Γ here). Hence ρ is full.
op
R Λ ,R
Let M ∈ (S2 )Λ⊗
, the left (resp. right) Λ-module structure of M is
C
extendable to that of a left (resp. right) Γ-module structure by Theorem 6.27.
It remains to show that these structures make M a Γ-bimodule. Let a ∈ Λ.
33
Then λa : M → M given by λa (m) = am is a right Λ-linear, and hence is
right Γ-linear. So for b ∈ Γ, ρb : M → M given by ρb (m) = mb is left Λ-linear,
and hence is left Γ-linear, as desired.
Proposition
6.29. Let C be a 2-canonical opright Λ-module. Then (?)† :
op
′ Λ ,R
(S2 )C
→ (S2′ )Γ,R and (?)‡ : (S2′ )Γ,R → (S2′ )ΛC ,R give a contravariant equivalence.
Proof. As we know that (?)† and (?)‡ are contravariant adjoint each other, it
suffices to show that the unit λM : M → M †‡ and the (co-)unit µN : N → N ‡†
are isomorphisms. λM is an isomorphism by Lemma 6.24. Note that C is
a 2-canonical left Γ-module by Lemma 6.23. So µN is an isomorphism by
Lemma 6.24 applied to the right Γop -module C.
Corollary 6.30. Let C be a 2-canonical Λ-bimodule. Then (?)† = HomΛop (?, C)
op
and HomΛ (?, C) give a contravariant equivalence between (S2′ )ΛC ,R and (S2′ )Λ,R
C .
op
′ Λ⊗Λ ,R
They also give a duality of (S2 )C
.
Proof. The first assertion is immediate by Proposition 6.29 and Theorem 6.27.
The second assertion follows easily from the first and Corollary 6.28.
7.
Non-commutative Aoyama’s theorem
Lemma 7.1. Let (R, m, k) → (R′ , m′ , k ′ ) be a flat local homomorphism between Noetherian local rings.
1 Let M be a Λ-bimodule such that M ′ := R′ ⊗R M is isomorphic to
Λ′ := R′ ⊗R Λ as a Λ′ -bimodule. Then M ∼
= Λ as a Λ-bimodule.
2 Let M be a right Λ module such that M ′ := R′ ⊗R M is isomorphic to
Λ′ := R′ ⊗R Λ as a right Λ′ -module. Then M ∼
= Λ as a right Λ-module.
Proof. Taking the completion, we may assume that both R and R′ are complete. Let 1 = e1 + · · · + er be the decomposition of 1 into the mutually
orthogonal primitive idempotents of the center S of Λ. Then replacing R
by Sei , Λ by Λei , and R′ by the local ring of R′ ⊗R Sei at any maximal
ideal, we may further assume that S = R. This is equivalent to say that
R → EndΛ⊗R Λop Λ is isomorphic. So R′ → EndΛ′ ⊗R′ (Λ′ )op Λ′ is also isomorphic,
and hence the center of Λ′ is R′ .
P
1. Let ψ : M ′ → Λ′ be an isomorphism. Then we can write ψ = P m
i=1 ui ψi
n
−1
′
op
with ui ∈ R and ψi ∈ HomΛ⊗R Λ (M, Λ). Also, we can write ψi = j=1 vj ϕj
34
P
with vj ∈ R′ and ϕj ∈ HomΛ⊗R Λop (Λ, M). As i,j ui vj ψi ϕj = ψψ −1 = 1 ∈
EndΛ′ ⊗R′ (Λ′ )op Λ′ ∼
= R′ and R′ is local, there exists some i, j such that ui vj ψi ϕj
is an automorphism of Λ′ . Then ψi : M ′ → Λ′ is also an isomorphism. By
faithful flatness, ψi : M → Λ is an isomorphism.
2. It is easy to see that M ∈ mod Λ is projective. So replacing Λ by Λ/J,
where J is the radical of J, and changing R and R′ as above, we may assume
that R is a field and Λ is central simple. Then there is only one simple right
Λ-module, and M and Λ are direct sums of copies of it. As M ′ ∼
= Λ′ , by
dimension counting, the number of copies are equal, and hence M and Λ are
isomorphic.
Lemma 7.2. Let (R, m, k) → (R′ , m′ , k ′ ) be a flat local homomorphism between Noetherian local rings.
1 Let C be a 2-canonical bimodule of Λ over R. Let M be a Λ-bimodule
such that M ′ := R′ ⊗R M is isomorphic to C ′ := R′ ⊗R C as a Λ′ bimodule. Then M ∼
= C as a Λ-bimodule.
2 Let C be a 2-canonical right Λ-module over R. Let M be a right Λmodule such that M ′ := R′ ⊗R M is isomorphic to C ′ := R′ ⊗R C as a
right Λ′ -module. Then M ∼
= C as a right Λ-module.
Proof. 1. As M ′ ∼
= C ′ and C ∈ (S2′ )C , it is easy to see that M ∈ (S2′ )C .
Hence M is a Γ-bimodule, where Γ = EndΛop C = EndΛ C, see (6.26) and
Corollary 6.28. Note that (M † )′ ∼
= Γ′ as Γ′ -bimodules. By Lemma 7.1,
= (C † )′ ∼
† ∼
1, we have that M = Γ as a Γ-bimodule. Hence M ∼
= M †‡ ∼
= Γ‡ ∼
= C.
† ′ ∼
† ′ ∼
′
′
† ∼
2. As (M ) = (C ) = Γ as Γ -modules, M = Γ as Γ-modules by
Lemma 7.1, 2. Hence M ∼
= C.
= Γ‡ ∼
= M †‡ ∼
Proposition 7.3. Let (R, m, k) → (R′ , m′ , k ′ ) be a flat local homomorphism
between Noetherian local rings. Assume that R′ /mR′ is zero-dimensional, and
M ′ := R′ ⊗R M is the right canonical module of Λ′ := R′ ⊗R Λ. Then R′ /mR′
is Gorenstein.
Proof. We may assume that both R and R′ are complete. Replacing R by
R/ annR Λ and R′ by R′ ⊗R R/ annR Λ, we may assume that Λ is a faithful
R-module. Let d = dim R = dim R′ .
Then
R′ ⊗R Hmd (M) ∼
= HomR′ (Γ′ , E ′ ),
= Hmd ′ (KΛ′ ) ∼
= Hmd ′ (R′ ⊗R M) ∼
35
where Λ′ = R′ ⊗R Λ, E ′ = ER′ (R′ /m′ ) is the injective hull of the residue field,
Γ = EndΛop M, Γ′ = R′ ⊗R Γ ∼
= EndΛ′ KΛ′ , and the isomorphisms are those of
′
Γ -modules. The last isomorphism is by (5.19). So R′ ⊗R Hmd (M) ∈ Mod Γ′ is
injective. Considering the spectral sequence
E2p,q = ExtpR′ ⊗R (Γ⊗R k) (W, ExtqΓ′ (R′ ⊗R (Γ ⊗R k), R′ ⊗R Hmd (M)))
′
d
⇒ Extp+q
Γ′ (W, R ⊗R Hm (M))
1,0
for W ∈ Mod(R′ ⊗R (Γ ⊗R k)), E21,0 = E∞
⊂ Ext1Γ′ (W, R′ ⊗R Hmd (M)) = 0 by
the injectivity of R′ ⊗R Hmd (M). It follows that
HomΓ′ (R′ ⊗R (Γ ⊗R k), R′ ⊗R Hmd (M)) ∼
= (R′ /mR′ ) ⊗k HomR (k, Hmd (M))
is an injective (R′ /mR′ ) ⊗k (Γ ⊗R k)-module. However, as an R′ /mR′ -module,
this is a free module. Also, this module must be an injective R′ /mR′ -module,
and hence R′ /mR′ must be Gorenstein.
Lemma 7.4. Let (R, m, k) → (R′ , m′ , k ′ ) be a flat local homomorphism between Noetherian local rings such that R′ /mR′ is Gorenstein. Assume that
the canonical module KΛ of Λ exists. Then R′ ⊗R KΛ is the canonical module
of R′ ⊗R Λ.
Proof. We may assume that both R and R′ are complete. Let I be the normalized dualizing complex of R. Then R′ ⊗R I[d′ − d] is a normalized dualizing
complex of R′ , where d′ = dim R′ and d = dim R, since R → R′ is a flat local
homomorphism with the d′ − d-dimensional Gorenstein closed fiber, see [AvF,
(5.1)] (the definition of a normalized dualizing complex in [AvF] is different
from ours. We follow the one in [Hart2, Chapter V]). So
′
′
′
−d′
∼
∼
R′ ⊗R KΛ ∼
= R′ ⊗R Ext−d
R (Λ, I) = ExtR (R ⊗R Λ, R ⊗R I[d − d]) = KΛ′ .
Theorem 7.5 ((Non-commutative Aoyama’s theorem) cf. [Aoy, Theorem 4.2]).
Let (R, m) → (R′ , m′ ) be a flat local homomorphism between Noetherian local
rings.
1 If M is a Λ-bimodule and M ′ = R′ ⊗R M is the canonical module of
Λ′ = R′ ⊗R Λ, then M is the canonical module of Λ.
2 If M is a right Λ-module such that M ′ is the right canonical module of
Λ′ , then M is the right canonical module of Λ.
36
Proof. We may assume that both R and R′ are complete. Then the canonical
module exists, and the localization of a canonical module is a canonical module, and hence we may localize R′ by a minimal element of {P ∈ Spec R′ |
P ∩ R = m}, and take the completion again, we may further assume that the
fiber ring R′ /mR′ is zero-dimensional. Then R′ /mR′ is Gorenstein by Proposition 7.3. Then by Lemma 7.4, M ′ ∼
= R′ ⊗R KΛ . By Lemma 7.2,
= KΛ ′ ∼
M∼
= KΛ . In 1, the isomorphisms are those of bimodules, while in 2, they are
of right modules. The proofs of 1 and 2 are complete.
Corollary 7.6. Let (R, m) be a Noetherian local ring, and assume that K
is the canonical (resp. right canonical) module of Λ. If P ∈ SuppR K, then
the localization KP is the canonical (resp. right canonical) module of ΛP . In
particular, K is a semicanonical bimodule (resp. right module), and hence is
2-canonical over R/ annR Λ.
Proof. Let Q be a prime ideal of R̂ lying over P . Then (K̂)Q ∼
= R̂Q ⊗RP KP
is nonzero by assumption, and hence is the canonical (resp. right canonical)
module of R̂Q ⊗R Λ. Using Theorem 7.5, KP is the canonical (resp. right
canonical) module of ΛP . The last assertion follows.
(7.7) Let (R, m) be local, and assume that KΛ exists. Assume that Λ is
a faithful R-module. Then it is a 2-canonical Λ-bimodule over R by Corollary 7.6. Letting Γ = EndΛop KΛ , KΓ ∼
= KΛ as Λ-bimodules by Corollary6.22.
So by Corollary 6.28, there exists some Γ-bimodule structure of KΛ such that
KΓ ∼
= KΛ as Γ-bimodules. As the left Γ-module structure of KΛ which extends the original left Λ-module structure is unique, and it is the obvious
action of Γ = EndΛop KΛ . Similarly the right action of Γ is the obvious action
of Γ = ∆ = (EndΛ KΛ )op , see (6.26).
8.
Evans–Griffith’s theorem for n-canonical modules
Lemma 8.1 (cf. [Aoy, Proposition 2], [Ogo, Proposition 4.2], [AoyG, Proposition 1.2]). Let (R, m) be local and assume that Λ has a canonical module
C = KΛ . Then we have
1 λR : Λ → EndΛop KΛ is injective if and only if Λ satisfies the (S1 )R
condition and SuppR Λ is equidimensional.
2 λR : Λ → EndΛop KΛ is bijective if and ony if Λ satisfies the (S2 )R
condition.
37
Proof. Replacing R by R/ annR Λ, we may assume that Λ is a faithful Rmodule. Then KΛ is a 2-canonical Λ-bimodule over R by Corollary 7.6. KΛ
is full if and only if SuppR Λ is equidimensional by Lemma 5.10, 1.
Now 1 is a consequence of Lemma 6.19. 2 follows from Corollary 6.25 and
Lemma 5.12.
Proposition 8.2 (cf. [AoyG, (2.3)]). Let (R, m) be a local ring, and assume
that there is an R-canonical module KΛ of Λ. Assume that Λ ∈ (S2 )R , and KΛ
is a Cohen–Macaulay R-module. Then Λ is Cohen–Macaulay. If, moreover,
KΛ is maximal Cohen–Macaulay, then so is Λ.
Proof. The second assertion follows from the first. We prove the first assertion.
Replacing R by R/ annR Λ, we may assume that Λ is faithful. Let d = dim R.
So Λ satisfies (S2′ ), and KΛ is maximal Cohen–Macaulay. As KΛ is the lowest
non-vanishing cohomology of J := RHomR (Λ, I), there is a natural map σ :
KΛ [d] → J which induces an isomorphism on the −dth cohomology groups.
Then the diagram
Λ
λ
/
HomΛop (KΛ [d], KΛ [d])
σ∗
λ
RHomΛop (J, J)
σ∗
/
RHomΛop (KΛ [d], J)
is commutative. The top horizontal arrow λ is an isomorphism by Lemma 8.1.
Note that
RHomΛop (J, J) ∼
= RHomR (J, I) = RHomR (RHomR (Λ, I), I) = Λ,
and the left vertical arrow is an isomorphism. As KΛ is maximal Cohen–
Macaulay, RHomΛop (KΛ [d], J) is concentrated in degree zero. As H i (J) = 0
for i < −d, we have that the right vertical arrow σ∗ is an isomorphism. Thus
the bottom horizontal arrow σ ∗ is an isomorphism. Applying RHomΛ (?, J)
to this map, we have that KΛ [d] → J is an isomorphism. So Λ is Cohen–
Macaulay, as desired.
Corollary 8.3 (cf. [AoyG, (2.2)]). Let (R, m) be a local ring, and assume
that there is an R-canonical module KΛ of Λ. Then KΛ is a Cohen–Macaulay
(resp. maximal Cohen–Macaulay) R-module if and only if Γ = EndΛop KΛ is
so.
38
Proof. As KΛ and Γ has the same support, if both of them are Cohen–
Macaulay and one of them are maximal Cohen–Macaulay, then the other is
also. So it suffices to prove the assertion on the Cohen–Macaulay property.
To verify this, we may assume that Λ is a faithful R-module. Note that Γ
satisfies (S2′ ). By Corollary 6.22, KΛ is Cohen–Macaulay if and only if KΓ is.
If Γ is Cohen–Macaulay, then KΓ is Cohen–Macaulay by (5.18). Conversely,
if KΓ is Cohen–Macaulay, then Γ is Cohen–Macaulay by Proposition 8.2.
Theorem 8.4 (cf. [EvG, (3.8)], [ArI, (3.1)]). Let R be a Noetherian commutative ring, and Λ a module-finite R-algebra, which may not be commutative.
Let n ≥ 1, and C be a right n-canonical Λ-module. Set Γ = EndΛop C. Let
M ∈ mod C. Then the following are equivalent.
1 M ∈ TF(n, C).
2 M ∈ UP(n, C).
3 M ∈ Syz(n, C).
4 M ∈ (Sn′ )C .
Proof. 1⇒2⇒3⇒4 is easy. We prove 4⇒1. By Lemma 6.5, we may assume
that n ≥ 2. By Lemma 6.24, M ∈ TF(2, C). Let
F : 0 ← M † ← F0 ← F1 ← · · · ← Fn−1
be a resolution of M † in Γ mod with each Fi ∈ add Γ. It suffices to prove its
dual
‡
F‡ : 0 → M → F0‡ → F1‡ → · · · → Fn−1
is acyclic. By Lemma 2.12, we may localize at P ∈ Rh<ni , and may assume
that dim R < n. If M = 0, then F is split exact, and so F‡ is also exact. So we
may assume that M 6= 0. Then by assumption, C ∼
= KΛ in mod Λ, and C is a
maximal Cohen–Macaulay R-module. Hence Γ is Cohen–Macaulay by Corollary 8.3. So by (5.16) and Lemma 6.22, RHomΓ (M † , C) = RHomΓ (M † , KΓ ) =
M, and we are done.
Corollary 8.5. Let the assumptions and notation be as in Theorem 8.4. Let
n ≥ 0. Assume further that
1 ExtiΛop (C, C) = 0 for 1 ≤ i ≤ n;
2 C is Λ-full.
39
3 Λ satisfies the (Sn′ )R condition.
Then for 0 ≤ r ≤ n,
⊥r
C is contravariantly finite in mod Λ.
Proof. For any M ∈ mod Λ, the nth syzygy module Ωn M satisfies the (Sn′ )R
Cn
op
condition by 2 and 3. By Theorem 8.4, Ω M ∈ TFΛ (n, C). By Theorem 3.16, M ∈ Zr,0 , and there is a short exact sequence
g
0→Y →X −
→M →0
with X ∈ Xr,0 = ⊥r C and Y ∈ Yr . As Ext1Λop (X, Y ) = 0, we have that g is a
right ⊥r C-approximation, and hence ⊥r C is contravariantly finite.
Corollary 8.6. Let the assumptions and notation be as in Theorem 8.4. Let
n ≥ 0, and C a Λ-full (n + 2)-canonical Λ-bimodule over R. Assume that Λ
′
satisfies the (Sn+2
)R condition. Then ⊥n C is contravariantly finite in mod Λ.
Proof. By Corollary 8.5, it suffices to show that ExtiΛop (C, C) = 0 for 1 ≤ i ≤
n. Let ∆ = (EndΛ C)op . Then the canonical map Λ → ∆ is an isomorphism
by Lemma 6.25, since C is a Λ-full 2-canonical Λ-bimodule over R. As Λ ∈
′
(Sn+2
)R and C is a Λ-full (n + 2)-canonical left Λ-module over R, applying
Theorem 8.4 to Λop , we have that Exti∆op (C, C) = 0 for 1 ≤ i ≤ n. As we have
Λop → ∆op is an isomorphism, we have that ExtiΛop (C, C) = 0, as desired.
9.
Symmetric and Frobenius algebras
(9.1) Let (R, m) be a Noetherian semilocal ring, and Λ a module-finite Ralgebra. We say that Λ is quasi-symmetric if Λ is the canonical module of Λ.
That is, Λ ∼
= KΛ as Λ-bimodules. It is called symmetric if it is quasi-symmetric
and GCM. Note that Λ is quasi-symmetric (resp. symmetric) if and only if Λ̂
is so, where ˆ? denotes the m-adic completion. Note also that quasi-symmetric
and symmetric are absolute notion, and is independent of the choice of R in
the sense that the definition does not change when we replace R by the center
of Λ.
(9.2) For (non-semilocal) Noetherian ring R, we say that Λ is locally quasisymmetric (resp. locally symmetric) over R if for any P ∈ Spec R, ΛP is a
quasi-symmetric (resp. symmetric) RP -algebra. This is equivalent to say that
for any maximal ideal m of R, Λm is quasi-symmetric (resp. symmetric). In
the case that (R, m) is semilocal, Λ is locally quasi-symmetric (resp. locally
symmetric) over R if it is quasi-symmetric (resp. symmetric), but the converse
is not true in general.
40
Lemma 9.3. Let (R, m) be a Noetherian semilocal ring, and Λ a module-finite
R-algebra. Then the following are equivalent.
1 ΛΛ is the right canonical module of Λ.
2
ΛΛ
is the left canonical module of Λ.
Proof. We may assume that R is complete. Then replacing R by a Noether
normalization of R/ annR Λ, we may assume that R is regular and Λ is a
faithful R-module.
We prove 1⇒2. By Lemma 5.10, Λ satisfies (S2′ )R . As R is regular and
dim R = dim Λ, KΛ = Λ∗ = HomR (Λ, R). So we get an R-linear map
ϕ : Λ ⊗R Λ → R
such that ϕ(ab ⊗ c) = ϕ(a ⊗ bc) and that the induced map h : Λ → Λ∗
given by h(a)(c) = ϕ(a ⊗ c) is an isomorphism (in mod Λ). Now ϕ induces a
homomorphism h′ : Λ → Λ∗ in Λ mod given by h′ (c)(a) = ϕ(a ⊗ c). To verify
that this is an isomorphism, as Λ and Λ∗ are reflexive R-modules, we may
localize at P ∈ Rh<2i , and then take a completion, and hence we may further
assume that dim R ≤ 1. Then Λ is a finite free R-module, and the matrices of
h and h′ are transpose each other. As the matrix of h is invertible, so is that
of h′ , and h′ is an isomorphism.
2⇒1 follows from 1⇒2, considering the opposite ring.
Definition 9.4. Let (R, m) be semilocal. We say that Λ is a pseudo-Frobenius
R-algebra if the equivalent conditions of Lemma 9.3 are satisfied. If Λ is GCM
in addition, then it is called a Frobenius R-algebra. Note that these definitions
are independent of the choice of R. Moreover, Λ is pseudo-Frobenius (resp.
Frobenius) if and only if Λ̂ is so, where ˆ? is the m-adic completion. For a
general R, we say that Λ is locally pseudo-Frobenius (resp. locally Frobenius)
over R if ΛP is pseudo-Frobenius (resp. Frobenius) for P ∈ Spec R.
Lemma 9.5. Let (R, m) be semilocal. Then the following are equivalent.
1 (KΛ̂ )Λ̂ is projective in mod Λ̂.
2
Λ̂ (KΛ̂ )
is projective in Λ̂ mod,
where ˆ? denotes the m-adic completion.
41
Proof. We may assume that (R, m, k) is complete regular local and Λ is a
faithful R-module. Let ¯? denote the functor k⊗R ?. Then Λ̄ is a finite dimensional k-algebra. So mod Λ̄ and Λ̄ mod have the same number of simple
modules, say n. An indecomposable projective module in mod Λ is nothing
but the projective cover of a simple module in mod Λ̄. So mod Λ and Λ mod
have n indecomposable projectives. Now HomR (?, R) is an equivalence between add(KΛ )Λ and add Λ Λ. It is also an equivalence between add Λ (KΛ ) and
add ΛΛ . So both add(KΛ )Λ and add Λ (KΛ ) also have n indecomposables. So 1
is equivalent to add(KΛ )Λ = add ΛΛ . 2 is equivalent to add Λ (KΛ ) = add Λ Λ.
So 1⇔2 is proved simply applying the duality HomR (?, R).
(9.6) Let (R, m) be semilocal. If the equivalent conditions in Lemma 9.5
are satisfied, then we say that Λ is pseudo-quasi-Frobenius. If it is GCM
in addition, then we say that it is quasi-Frobenius. These definitions are
independent of the choice of R. Note that Λ is pseudo-quasi-Frobenius (resp.
quasi-Frobenius) if and only if Λ̂ is so.
Proposition 9.7. Let (R, m) be semilocal. Then the following are equivalent.
1 Λ is quasi-Frobenius.
2 Λ is GCM, and dim Λ = idim Λ Λ, where idim denotes the injective dimension.
3 Λ is GCM, and dim Λ = idim ΛΛ .
Proof. 1⇒2. By definition, Λ is GCM. To prove that dim Λ = idim Λ Λ, we
may assume that R is local. Then by [GN, (3.5)], we may assume that R
is complete. Replacing R by the Noetherian normalization of R/ annR Λ, we
may assume that R is a complete regular local ring of dimension d, and Λ
its maximal Cohen–Macaulay module. As add Λ Λ = add Λ (KΛ ) by the proof
of Lemma 9.5, it suffices to prove idim Λ (KΛ ) = d. Let IR be the minimal
injective resolution of the R-module R. Then J = HomR (Λ, IR ) is an injective
resolution of KΛ = HomR (Λ, R). As the length of J is d and
Extd (Λ/mΛ, KΛ ) ∼
= Extd (Λ/mΛ, R) 6= 0,
R
Λ
we have that idim Λ (KΛ ) = d.
2⇒1. We may assume that R is complete regular local and Λ is maximal
Cohen–Macaulay. By [GN, (3.6)], we may further assume that R is a field.
Then Λ Λ is injective. So (KΛ )Λ = HomR (Λ, R) is projective, and Λ is quasiFrobenius, see [SkY, (IV.3.7)].
1⇔3 is proved similarly.
42
Corollary 9.8. Let R be arbitrary. Then the following are equivalent.
1 For any P ∈ Spec R, ΛP is quasi-Frobenius.
2 For any maximal ideal m of R, Λm is quasi-Frobenius.
3 Λ is a Gorenstein R-algebra in the sense that Λ is a Cohen–Macaulay
R-module, and idimΛP ΛP ΛP = dim ΛP for any P ∈ Spec R.
Proof. 1⇒2 is trivial.
2⇒3. By Proposition 9.7, we have idim Λm Λm = dim Λm for each m. Then
by [GN, (4.7)], Λ is a Gorenstein R-algebra.
3⇒1 follows from Proposition 9.7.
(9.9) Let R be arbitrary. We say that Λ is a quasi-Gorenstein R-algebra if
ΛP is pseudo-quasi-Frobenius for each P ∈ Spec R.
Definition 9.10 (Scheja–Storch [SS]). Let R be general. We say that Λ is
symmetric (resp. Frobenius) relative to R if Λ is R-projective, and Λ∗ :=
HomR (Λ, R) is isomorphic to Λ as a Λ-bimodule (resp. as a right Λ-module).
It is called quasi-Frobenius relative to R if the right Λ-module Λ∗ is projective.
Lemma 9.11. Let (R, m) be local.
1 If dim Λ = dim R, R is quasi-Gorenstein, and Λ∗ ∼
= Λ as Λ-bimodules
∗
∗ ∼
(resp. Λ = Λ as right Λ-modules, Λ is projective as a right Λ-module),
then Λ is quasi-symmetric (resp. pseudo-Frobenius, pseudo-quasi-Frobenius).
2 If R is Gorenstein and Λ is symmetric (resp. Frobenius, quasi-Frobenius)
relative to R, then Λ is symmetric (resp. Frobenius, quasi-Frobenius).
3 If Λ is nonzero and R-projective, then Λ is quasi-symmetric (resp. pseudoFrobenius, pseudo-quasi-Frobenius) if and only if R is quasi-Gorenstein
and Λ is symmetric (resp. Frobenius, quasi-Frobenius) relative to R.
4 If Λ is nonzero and R-projective, then Λ is symmetric (resp. Frobenius,
quasi-Frobenius) if and only if R is Gorenstein and Λ is symmetric (resp.
Frobenius, quasi-Frobenius) relative to R.
Proof. We can take the completion, and we may assume that R is complete
local.
43
1. Let d = dim Λ = dim R, and let I be the normalized dualizing complex
of R. Then
−d
∼
KΛ = Ext−d
(I)) ∼
= Hom(Λ, KR ) ∼
= Hom(Λ, R) = Λ∗
R (Λ, I) = HomR (Λ, H
as Λ-bimodules, and the result follows.
2. We may assume that Λ is nonzero. As R is Cohen–Macaulay and Λ is
a finite projective R-module, Λ is a maximal Cohen–Macaulay R-module. By
1, the result follows.
3. The ‘if’ part follows from 1. We prove the ‘only if’ part. As Λ
is R-projective and nonzero, dim Λ = dim R. As Λ is R-finite free, KΛ ∼
=
HomR (Λ, KR ) ∼
= Λ∗ ⊗R KR . As KΛ is R-free and Λ∗ ⊗R KR is nonzero and is
isomorphic to a direct sum of copies of KR , we have that KR is R-projective,
and hence R is quasi-Gorenstein, and KR ∼
= R. Hence KΛ ∼
= Λ∗ , and the
result follows.
4 follows from 3 easily.
(9.12) Let (R, m) be semilocal. Let a finite group G act on Λ by R-algebra
automorphisms.
L Let Ω = Λ ∗ G, the twisted group algebra. That is, Ω =
Λ ⊗R RG =
g∈G Λg as an R-module, and the product of Ω is given by
(ag)(a′ g ′) = (a(ga′ ))(gg ′) for a, a′ ∈ Λ and g, g ′ ∈ G. This makes Ω a modulefinite R-algebra.
(9.13) We simply call an RG-module a G-module. We say that M is a
(G, Λ)-module if M is a G-module, Λ-module, the R-module structures coming
from that of the G-module structure and the Λ-module structure agree, and
g(am) = (ga)(gm) for g ∈ G, a ∈ Λ, and m ∈ M. A (G, Λ)-module and an
Ω-module are one and the same thing.
(9.14) By the action (a⊗a′ )g)a1 = a(ga1 )a′ , we have that Λ is a (Λ⊗Λop )∗Gmodule in a natural way. So it is an Ω-module by the action (ag)a1 = a(ga1 ).
It is also a right Ω-module by the action a1 (ag) = g −1(a1 a). If the action of
G on Λ is trivial, then these actions make an Ω-bimodule.
(9.15) Given an Ω-module M and an RG-module V , M ⊗R V is an Ωmodule by (ag)(m ⊗ v) = (ag)m ⊗ gv. HomR (M, V ) is a right Ω-module by
(ϕ(ag))(m) = g −1 ϕ(a(gm)). It is easy to see that the standard isomorphism
HomR (M ⊗R V, W ) → HomR (M, HomR (V, W ))
is an isomorphism of right Ω-modules for a left Ω-module M and G-modules
V and W .
44
(9.16) Now consider the case Λ = R. Then the pairing φ : RG ⊗R RG → R
given by φ(g ⊗ g ′ ) = δgg′ ,e (Kronecker’s delta) is non-degenerate, and induces
an RG-bimodule isomorphism Ω = RG → (RG)∗ = Ω∗ . As Ω = RG is a finite
free R-module, we have that Ω = RG is symmetric relative to R.
Lemma 9.17. If Λ is quasi-symmetric (resp. symmetric) and the action of G
on Λ is trivial, then Ω is quasi-symmetric (resp. symmetric).
Proof. Taking the completion, we may assume that R is complete. Then
replacing R by a Noether normalization of R/ annR Λ, we may assume that R
is a regular local ring, and Λ is a faithful R-module. As the action of G on Λ
is trivial, Ω = Λ ⊗R RG is quasi-symmetric (resp. symmetric), as can be seen
easily.
(9.18) In particular, if Λ is commutative quasi-Gorenstein (resp. Gorenstein)
and the action of G on Λ is trivial, then Ω = ΛG is quasi-symmetric (resp.
symmetric).
(9.19) In general, Ω Ω ∼
= Λ ⊗R RG as Ω-modules.
Lemma 9.20. Let M and N be right Ω-modules, and let ϕ : M → N be a
homomorphism of right Λ-modules. Then ψ : M ⊗ RG → N ⊗ RG given by
ψ(m ⊗ g) = g(ϕ(g −1m)) ⊗ g is an Ω-homomorphism. In particular,
1 If ϕ is a Λ-isomorphism, then ψ is an Ω-isomorphism.
2 If ϕ is a split monomorphism in mod Λ, then ψ is a split monomorphism
in mod Ω.
Proof. Straightforward.
Proposition 9.21. Let G be a finite group acting on Λ. Set Ω := Λ ∗ G.
1 If the action of G on Λ is trivial and Λ is quasi-symmetric (resp. symmetric), then so is Ω.
2 If Λ is pseudo-Frobenius (resp. Frobenius), then so is Ω.
3 If Λ is pseudo-quasi-Frobenius (resp. quasi-Frobenius), then so is Ω.
Proof. 1 is Lemma 9.17. To prove 2 and 3, we may assume that (R, m) is
complete regular local and Λ is a faithful module.
45
2.
(KΩ )Ω ∼
= HomR (Λ ⊗R RG, R) ∼
= HomR (Λ, R) ⊗ (RG)∗ ∼
= KΛ ⊗ RG
as right Ω-modules. It is isomorphic to ΛΩ ⊗ RG ∼
= ΩΩ by Lemma 9.20, 1,
∼
since KΛ = Λ in mod Λ. Hence Ω is pseudo-Frobenius. If, in addition, Λ is
Cohen–Macaulay, then Ω is also Cohen–Macaulay, and hence Ω is Frobenius.
3 is proved similarly, using Lemma 9.20, 2.
Note that the assertions for Frobenius and quasi-Frobenius properties also
follow easily from Lemma 9.11 and [SS, (3.2)].
10.
Codimension-two argument
(10.1) Let X be a locally Noetherian scheme, U its open subscheme, and Λ
a coherent OX -algebra. Assume the (S2′ ) condition on Λ. Let i : U ֒→ X be
the inclusion. In what follows we use the notation for rings and modules to
schemes and coherent algebras and modules in an obvious manner.
(10.2) Let M ∈ mod Λ. That is, M is a coherent right Λ-module. Then by
restriction, i∗ M ∈ mod i∗ Λ.
(10.3) For a quasi-coherent i∗ Λ-module N , we have an action
u⊗1
a
i∗ N ⊗OX Λ −−→ i∗ N ⊗OX i∗ i∗ Λ → i∗ (N ⊗OU i∗ Λ) −
→ i∗ N .
So we get a functor i∗ : Mod i∗ Λ → Mod Λ, where Mod i∗ Λ (resp. Mod Λ)
denote the category of quasi-coherent i∗ Λ-modules (resp. Λ-modules).
Lemma 10.4. Let the notation be as above. Assume that U is large in X
(that is, codimX (X \ U) ≥ 2). If M ∈ (S2′ )Λ, , then the canonical map u :
M → i∗ i∗ M is an isomorphism.
Proof. Follows immediately from [Has, (7.31)].
Proposition 10.5. Let the notation be above, and let U be large in X. Assume
that there is a 2-canonical right Λ-module. Then we have the following.
1 If N ∈ (S2′ )i
∗ Λ,U
, then i∗ N ∈ (S2′ )Λ,X .
2 i∗ : (S2′ )Λ,X → (S2′ )i
each other.
∗ Λ,U
and i∗ : (S2′ )i
46
∗ Λ,U
→ (S2′ )Λ,X are quasi-inverse
Proof. The question is local, and we may assume that X is affine.
1. There is a coherent subsheaf Q of i∗ N such that i∗ Q = i∗ i∗ N = N by
[Hart2, Exercise II.5.15]. Let V be the Λ-submodule of i∗ N generated by Q.
That is, the image of the composite
Q ⊗OX Λ → i∗ N ⊗OX Λ → i∗ N .
Note that V is coherent, and i∗ Q ⊂ i∗ V ⊂ i∗ i∗ N = i∗ Q = N .
Let C be a 2-canonical right Λ-module. Let ?† := HomΛop (?, C), Γ =
EndΛ C, and ?‡ := HomΓ (?, C). Let M be the double dual V †‡ . Then M ∈
(S2′ )Λ,X , and hence
M∼
= i∗ i∗ M ∼
= i∗ i∗ (V †‡ ) ∼
= i∗ (i∗ V)†‡ ∼
= i∗ (N †‡ ) ∼
= i∗ N .
So i∗ N ∼
= M lies in (S2′ )Λ,X .
2 follows from 1 and Lemma 10.4 immediately.
References
[Aoy]
Y. Aoyama, Some basic results on canonical modules, J. Math. Kyoto
Univ. 23 (1983), 85–94.
[AoyG]
Y. Aoyama and S. Goto, On the endomorphism ring of the canonical
module, J. Math. Kyoto Univ. 25 (1985), 21–30.
[ArI]
T. Araya and K.-i. Iima, Locally Gorensteinness over Cohen–
Macaulay rings, arXiv:1408.3796v1
[ASS]
I. Assem, D. Simson, and A. Skowroński, Elements of the Representation Theory of Associative Algebras, Cambridge (2006).
[AvF]
L. L. Avramov and H.-B. Foxby, Locally Gorenstein homomorphisms, Amer. J. Math. 114 (1992), 1007–1047.
[BS]
M. P. Brodmann and R. Y. Sharp, Local Cohomology: An Algebraic
Introduction with Geometric Applications, Cambridge (1998).
[EvG]
E. G. Evans and P. Griffith, Syzygies, London Math. Soc. Lecture
Note Series 106, Cambridge (1985).
47
[GN]
S. Goto and K. Nishida, Towards a theory of Bass numbers with
applications to Gorenstein algebras, Colloq. Math. 91 (2002), 191–
253.
[Gro4]
A. Grothendieck, Eléments de Géométrie Algébrique IV, 3e Partie,
IHES Publ. Math. 28 (1966).
[Hart]
R. Hartshorne, Residues and Duality, Lecture Notes in Math. 20,
Springer (1966).
[Hart2]
R. Hartshorne, Algebraic Geometry, Graduate Texts in Math. 52,
Springer Verlag (1977).
[Hart4]
R. Hartshorne, Generalized divisors on Gorenstein schemes, KTheory 8 (1994), 287–339.
[Has]
M. Hashimoto, Equivariant class group. III. Almost principal fiber
bundles, arXiv:1503.02133v1
[IW]
O. Iyama and M. Wemyss, On the noncommutative Bondal-Orlov
conjecture, J. Reine Angew. Math. 683 (2013), 119–128.
[LeW]
G. J. Leuschke and R. Wiegand, Cohen–Macaulay Representations,
AMS (2012).
[Mat]
H. Matsumura, Commutative Ring Theory, First paperback edition,
Cambridge (1989).
[Ogo]
T. Ogoma, Existence of dualizing complexes, J. Math. Kyoto Univ.
24 (1984), 27–48.
[PS]
C. Peskine and L. Szpiro, Dimension projective finie et cohomologie
locale, Publ. IHES 42 (1973), 47–119.
[SS]
G. Scheja and U. Storch, Quasi-Frobenius-Algebren und lokal
vollständige Durchschnitte, Manuscripta Math. 19 (1976), 75–104.
Addendum: 20 (1977), 99–100.
[SkY]
A. Skowroński and K. Yamagata, Frobenius Algebras. I., Basic Representation Theory, European Mathematical Society (2011).
48
[Tak]
R. Takahashi, A New approximation theory which unifies spherical
and Cohen–Macaulay approximations, J. Pure Appl. Algebra 208
(2007), 617–634.
[TT]
R. W. Thomason and T. Trobaugh, Higher algebraic K-theory of
schemes and of derived categories, in The Grothendieck Festschrift.
III. (Birkhäuser) (1990), pp. 247–435.
[Yek]
A. Yekutieli, Dualizing complexes over noncommutative graded algebras, J. Algebra 153 (1992), 41–84.
49
| 0math.AC
|
Angle-based Shape Determination Theory of Planar
Graphs with Application to Formation Stabilization
Gangshan Jinga , Guofeng Zhangb , Heung Wing Joseph Leeb , Long Wangc,∗
arXiv:1803.04276v1 [cs.SY] 12 Mar 2018
a Center
for Complex Systems, School of Mechano-electronic Engineering, Xidian
University, Xi’an 710071, China
b Department of Applied Mathematics, Hong Kong Polytechnic University, Hong Kong,
China
c Center for Systems and Control, College of Engineering, Peking University, Beijing
100871, China
Abstract
This paper presents an angle-based approach for distributed formation shape
stabilization of multi-agent systems in the plane. We develop an angle rigidity
theory to study whether a planar framework can be determined by angles between segments uniquely up to translations, rotations, scalings and reflections.
The proposed angle rigidity theory is applied to the formation stabilization
problem, where multiple single-integrator modeled agents cooperatively achieve
an angle-constrained formation. During the formation process, the global coordinate system is unknown for each agent and wireless communications between
agents are not required. Moreover, by utilizing the advantage of high degrees of
freedom, we propose a distributed control law for agents to stabilize a desired
formation shape with desired orientation and scale. Two simulation examples
are performed for illustrating effectiveness of the proposed control strategies.
1. Introduction
A multi-agent formation stabilization problem is to design a decentralized
control law for a group of mobile agents to stabilize a prescribed formation
shape. An associated fundamental problem is: how to determine the geometric
shape of a graph embedded in a space, based on some local constraints such as
displacements, distances and bearings.
A straightforward approach for determining a shape is constraining the location of each vertex in the graph. A position-based formation strategy usually takes large costs and is unnecessary when the position of each agent is
∗ Corresponding
author: Long Wang
Email addresses: [email protected] (Gangshan Jing),
[email protected] (Guofeng Zhang), [email protected] (Heung Wing
Joseph Lee), [email protected] (Long Wang )
1
not strictly required. For reduction of information exchange and improvement of robustness of the control strategy, a displacement-constrained formation method, which determines the target formation shape by relative positions
between agents, has been extensively studied [1, 2, 3, 4, 5]. This method is
also called consensus-based formation since the formation problem can often be
transformed into a consensus problem, which is a hot topic being widely studied
[6, 7, 8, 5]. The investigations of displacement-based formation show that the
shape of a graph can be determined by inter-agent displacements uniquely if the
graph is connected. A disadvantage of displacement-based formation control is
the requirement of the global coordinate system.
During the last decade, distance-based shape control gained a lot of attention
since it has no requirement of the global coordinate system for each agent [9, 10,
11, 12, 13, 14, 15, 16, 17]. Different from the displacement-based approach, for
a noncomplete graph embedded in a space, it is not straightforward to answer
that whether its shape can be determined by edge lengths uniquely. A tool of
great utility to deal with this problem is the traditional graph rigidity theory
(we will refer to this theory as distance rigidity theory in this paper) [18, 19, 20],
which has been studied intensively in the area of mathematics.
In more recent years, bearing-constrained formation control attracted many
interests due to the low costs of bearing measurements [21, 22, 23, 24, 25]. In
this issue, the formation shape is constrained by inter-agent bearings. To distinguish what kind of shapes can be uniquely determined by inter-agent bearings,
the authors in [22] and [24] proposed the bearing rigidity theory. Compared to
distance-constrained formation control, an advantage of bearing-constrained formation strategy is the fact that no restriction on scale of the target formation is
imposed. As a result, it is simpler to control the scale of a bearing-based formation, which will benefit for obstacle avoidance, see [25]. Unfortunately, similar
to the displacement-based approach, a bearing-constrained formation requires
either the global coordinate system for each agent or developing observers based
on inter-agent communications, [24].
Besides the above-mentioned investigations, there are some other issues associated with formation control and formation strategies, for more details, we
refer the readers to [26, 27, 28, 29].
This paper studies the angle-constrained formation problem in the plane, in
which the target formation shape is the shape of a planar graph, and will be
encoded by angles between pair of edges joining a common vertex. Similar issues
have been reported in the literature. In [21], the authors discussed the possibility
of an angle-based formation approach and presented some initial results. In
another relevant reference [30], the authors solved the cyclic formation problem
by constraining the angle subtended at each vertex by its two neighbors. In this
case, the cyclic formation can be stretched while preserving invariance of each
angle, thus the target formation cannot be accurately stabilized. In contrast to
[30], we study how to stabilize a formation shape via angle constraints, such that
the stabilized formation is congruent to the target formation up to translations,
rotations, scalings and reflections.
Our contributions can be summarized as follows. (i). Enlightened by dis2
tance rigidity theory and bearing rigidity theory, we propose an angle rigidity
theory to study whether the shape of a planar graph can be uniquely determined
by angles only; see Section 3. (ii). We prove that for a planar framework, infinitesimal angle rigidity is equivalent to infinitesimal bearing rigidity; Theorem
3. From [31], infinitesimal angle rigidity is also a generic property of the graph.
(iii). We show that for a framework embedded by a triangulated Laman graph,
once it is strongly rigid, it can always be determined by angles uniquely up to
translations, rotations, scalings and reflections; see Theorem 7. (iv). We propose a distributed control law for achieving formation shape stabilization based
on the angle rigidity theory. It is shown that our control strategy can locally
exponentially stabilize multiple agents to form an infinitesimally angle rigid formation in the plane; see Theorem 8. (v). We design a distributed control law,
which can steer all agents to form a target formation shape with prescribed
orientation and scale; see Theorem 9. Note that in the formation maneuver
control literature [4, 16, 25], simultaneously controlling orientation and scale of
a formation usually cannot be achieved.
The advantages of angle-based formation approach are threefold. (i). Each
agent only has to measure relative displacements from neighbors with respect
to its local coordinate system. (ii). No wireless communications between agents
are required. (iii). Compared to displacement-, distance- and bearing-based
approaches, an angle-constrained shape has higher degrees of freedom. More
precisely, angles are invariant to motions including translations, rotations and
scalings, while inter-agent displacements, distances and bearings are only invariant to a subset of these motions. As a result, it is more convenient to achieve
formation maneuver control by using angle constraints.
The paper is structured as follows. Section 2 introduces some preliminaries
of distance- and bearing rigidity theory. Section 3 presents the angle rigidity
theory. Section 4 firstly proposes a distributed control law for achieving formation stabilization based on angle rigidity theory, then proposes a distributed
maneuver control law for stabilizing a formation shape with prescribed orientation and scale. Section 5 presents an application example to verify validity of
the formation strategy. Section 6 concludes the whole paper.
Notations: Throughout this paper, R denotes the set of real numbers; Rn
is the n−dimensional Euclidean space; || · || stands for the Euclidean norm;
X T means the transpose of matrix X; ⊗ is the Kronecker product; range(X),
null(X) and rank(X) denote the image space, null space, and the rank of matrix
X; In represents the n×n identity matrix; A\B is the set of those elements of A
not belonging to B; A vector p = (pT1 , · · · , pTs )T with pi ∈ R2 , i = 1, · · · , s is said
to be degenerate
if p1 , · ·
· , ps are collinear; O(n) is the orthogonal group in Rn ;
cos θ − sin θ
Ro (θ) =
is the 2-dimensional rotation matrix associated with
sin θ
cos θ
1 0
θ ∈ [0, 2π); Re (θ) = Ro (θ)I¯ with I¯ =
is the 2-dimensional reflection
0 −1
matrix associated with θ ∈ [0, 2π); x⊥ = Ro ( π2 )x for x ∈ R2 . For Xi ∈ Ra×b ,
i = 1, · · · , q, we denote diag(Xi ) = blockdiag{X1 , · · · , Xq } ∈ Rqa×qb .
3
An undirected graph with n vertices and m edges is denoted as G = (V, E),
where V = {1, · · · , n} and E ⊂ V × V denote the vertex set and the edge set,
respectively. Here we do not distinguish (i, j) and (j, i) in E. The incidence
matrix is represented by H = [hij ], which is a matrix with rows and columns
indexed by edges and vertices of G with an orientation. hij = 1 if the ith edge
sinks at vertex j, hij = −1 if the ith edge leaves vertex j, and hij = 0 otherwise.
It is well-known that rank(H) = n − 1 if and only if graph G is connected. Let
K denote a complete graph with n vertices.
2. Preliminaries of Graph Rigidity Theory
In this section, we introduce some preliminaries of distance and bearing
rigidity theory in the plane, which are taken from [18, 19, 24]. Distance rigidity
theory is to answer whether p can be uniquely determined up to translations,
rotations, and reflections, by partial length constraints on edges of G, while
bearing rigidity theory is to answer whether p can be uniquely determined up
to translations and scalings by partial bearing constraints on edges of G. In the
following, we will introduce these two theories in a unified approach.
We refer to a pair (G, p) as a framework, where G is a graph and p =
(pT1 , · · · , pTn )T ∈ R2n is called a configuration, pi is the coordinate of vertex i,
i = 1, · · · , n. To define rigidity of a framework (G, p), a smooth rigidity function
rG (·) : R2n → Rs should be first given, where s is some positive integer. By
the given rigidity function rG , several definitions associated with rigidity can be
induced as follows.
A framework (G, p) is said to be rigid if there exists a neighborhood Up
−1
of p such that rG−1 (rG (p)) ∩ Up = rK
(rK (p)) ∩ Up . (G, p) is globally rigid if
−1
−1
rG (rG (p)) = rK (rK (p)).
An infinitesimal motion is an assignment of velocities that guarantees the
invariance of rG (p), i.e.,
ṙG (p) =
∂rG (p)
v = 0,
∂p
(1)
where v = (v1T , · · · , vnT )T , vi = ṗi is the velocity of vertex i. We say a motion is
trivial if it satisfies equation (1) for any framework with n vertices. A framework
is infinitesimally rigid if every infinitesimal motion is trivial. Denote the rigidity
G (p)
matrix ∂r∂p
by R(p). Then equation (1) is equivalent to ṙG (p) = R(p)ṗ = 0.
Let T be the dimension of the space formed by all trivial motions, then a
framework (G, p) is infinitesimally rigid if and only if rank(R(p)) = 2n − T .
In the traditional graph rigidity theory, the above-mentioned rigidity function is commonly set by the following distance rigidity function:
DG (p) = (· · · , ||eij (p)||2 , · · · )T , (i, j) ∈ E,
where eij (p) = pi − pj .
4
(2)
Recently, the authors in [24] developed bearing rigidity theory by using the
following bearing rigidity function:
T
BG (p) = (· · · , gij
(p), · · · )T , (i, j) ∈ E,
(3)
p −p
where gij (p) = ||pii −pjj || .
For a framework in the plane, there are totally 2 independent translations,
1 independent rotation, 1 independent scaling. The trivial motions for a framework determined by distances can only be translations and rotations, thus the
dimension of trivial motion space should be TD = 2 + 1 = 3. The trivial
motions for a framework determined by bearings are translations and scalings,
accordingly, the dimension of trivial motion space is TB = 2 + 1 = 3.
The following two lemmas will be used in our paper.
Lemma 1. [24] A framework in R2 is infinitesimally bearing rigid if and only
if it is infinitesimally distance rigid.
Lemma 2. [16] If a framework in R2 is infinitesimally distance rigid, then for
any vertex i, the relative position vectors pi − pj , j ∈ Ni cannot be all collinear.
It is worth noting that infinitesimal bearing rigidity implies global bearing
rigidity [24], whereas infinitesimal distance rigidity cannot induce global distance
rigidity.
3. Angle rigidity
In this section, we develop an angle rigidity theory to investigate how to
encode geometric shapes of graphs embedded in the plane through angles only.
T
For a framework (G, p) in R2 , we will employ gij
gik as the constraint on the angle
between edges eij and eik , which is actually the cosine of this angle. Let TG =
T
{(i, j, k) ∈ V 3 : (i, j), (i, k) ∈ E, j < k}, then {gij
gik ∈ [−1, 1] : (i, j, k) ∈ TG } is
the set of constraints on all angles in (G, p). We should note that a framework
often has redundant angle information for shape determination. For example,
T
T
T
in Fig. 1 (a), once g12
g13 and g21
g23 are available, it holds that g31
g32 =
T
T
cos(π − arccos(g12 g13 ) − arccos(g21 g23 )). That is, the information of partial
angles in the graph is often sufficient to recognize a framework. Therefore, by
employing a subset TG∗ ⊂ TG with |TG∗ | = w, we will try to study whether (G, p)
T
can be uniquely determined by {gij
gik ∈ [−1, 1] : (i, j, k) ∈ TG∗ } based on the
angle rigidity theory to be developed in this paper. Note that although TG∗ is
only a subset of TG , the elements in TG∗ should involve all vertices in G, otherwise
the shape of (G, p) can never be determined.
For a framework (G, p), the angle rigidity function corresponding to TG∗ can
be written as
T
fTG∗ (p) = (· · · , gij
(p)gik (p), · · · )T , (i, j, k) ∈ TG∗ .
For the sake of notational simplicity, we denote fG (p) = fTG (p).
5
(4)
3
2
1
2
2
1
2
1
1
4
(a)
3
4
(b)
3
(c)
(d)
Figure 1:
(a) A globally and infinitesimally angle rigid framework with TG∗ =
{(1, 2, 3), (2, 1, 3)}. (b) A framework that is not angle rigid. (c) A globally and infinitesimally angle rigid framework with TG∗ = {(1, 2, 4), (2, 1, 4), (2, 3, 4), (3, 2, 4), (4, 1, 3)}. (d) A
globally angle rigid framework with TG = {(1, 2, 3), (2, 1, 3), (2, 1, 3)}.
It is easy to see that whether fTG∗ (p) can determine a unique shape congruent
to p is determined by the choice of TG∗ . As a result, the definitions of angle
rigidity must be associated with TG∗ . We present the following definitions.
Definition 1. A framework (G, p) is angle rigid if for some TG∗ , there exists a
−1
neighborhood Up of p such that fT−1∗ (fTG∗ (p)) ∩ Up = fK
(fK (p)) ∩ Up .
G
Definition 2. A framework (G, p) is globally angle rigid if for some TG∗ , there
−1
holds fT−1∗ (fTG∗ (p)) = fK
(fK (p)).
G
Definition 3. A framework (G, p) is minimally angle rigid if (G, p) is angle
rigid, and deletion of any edge will make (G, p) not angle rigid.
By these definitions, the frameworks (a) and (c) in Fig. 1 are both globally
angle rigid. For the framework (b), by moving the vertices along the blue arrows,
fG is invariant but the shape is deformed, thus (b) is not angle rigid. For the
framework (d), since the graph is complete, it obviously holds fG−1 (fG (p)) =
−1
fK
(fK (p)), thus (d) is globally angle rigid. Note that the shape of (d) still
cannot be determined by angles uniquely.
Similar to distance and bearing rigidity theory, we define the infinitesimal
angle motion as a motion preserving the invariance of fTG∗ (p). The velocity
v = ṗ corresponding to an infinitesimal motion should satisfy f˙TG∗ (p) = 0, which
is equivalent to the following equation
T
T
ġij
gik + gij
ġik = 0, (i, j, k) ∈ TG∗ .
(5)
∂g
ij
From [24], ∂eij
= ||e1ij || Pij , where Pij , P (gij ), P (·) : R2 → R2×2 is a projection
matrix defined as P (x) = I2 − xxT , x ∈ R2 is a unit vector. Then we have ġij =
1
||eij || Pij ėij .
T
Let g(p) = (· · · , gij
(p), · · · )T , where (i, j) ∈ E, and Rg ,
follows from the chain rule that
f˙TG∗ =
∂fTG∗ ∂g ∂e
Pij
ṗ = Rg (p) diag(
)H̄ ṗ = RTG∗ (p)ṗ,
∂g ∂e ∂p
||eij ||
6
3
∂fT ∗
G
∂g
. It
P
where H̄ = H ⊗ I2 , RTG∗ (p) , Rg (p) diag( ||eijij|| )H̄ = Rg (p)RB (p) ∈ Rw×2n is
termed the angle rigidity matrix, RB = ∂g(p)
is actually the bearing rigidity
∂p
matrix. Therefore, equation (5) is equivalent to RTG∗ (p)ṗ = 0.
Next we define infinitesimal angle rigidity, to do this, we should distinguish
all trivial motions for an angle-constrained geometric shape. By an intuitive
observation, the motions always preserving invariance of angles in the framework
are translations, rotations, and scalings. Therefore, the dimension of the trivial
motion space is 2 + 1 + 1 = 4. Note that the trivial motion space is always
a subspace of null(RTG∗ ), implying that dim(null(RTG∗ )) ≥ 4. We present the
following definition.
Definition 4. A framework (G, p) is infinitesimally angle rigid if there exists
a TG∗ such that every possible motion satisfying (5) is trivial, or equivalently,
dim(null(RTG∗ )) = 4.
By this definition, the frameworks (a) and (c) in Fig. 1 are both infinitesimally angle rigid. The frameworks (b) and (d) are not infinitesimally angle
rigid since they both have nontrivial infinitesimal angle motions, which are interpreted by the arrows in blue.
The following lemma gives the specific form of trivial motions preserving
invariance of angles.
Lemma 3. The trivial motion space for angle rigidity in R2 is S = Sr ∪Ss ∪St ,
where Sr = {(In ⊗ Ro ( π2 ))p} is the space formed by rotations, Ss = span{p} is
the space formed by scalings, St = null(H̄) = {1n ⊗ (1, 0)T , 1n ⊗ (0, 1)T } is the
space formed by translations.
Proof: In [24], the authors showed that Ss and St are scaling and translational
motion spaces, respectively, and always belong to null(RB (p)). Since RTG∗ (p) =
Rg (p)RB (p), it is straightforward that Ss ∪ St ⊆ null(RTG∗ (p)). Next we show
Sr ⊆ null(RTG∗ (p)).
Let η T =
T
∂gij
gik
∂g
be an arbitrary row of Rg . It suffices to show η T RB (p)(In ⊗
P
Ro ( π2 ))p = 0. Note that η = (0, gik , 0, gij , 0)T , which follows η T diag( ||eijij|| ) =
T
T
(0, gik
Pij /||eij ||, 0, gij
Pik /||eik ||, 0). Note also that H̄(In ⊗ Ro ( π2 ))p = (H ⊗
π
I2 )(In ⊗ Ro ( 2 ))p = (Im ⊗ Ro ( π2 ))(H ⊗ I2 )p = (Im ⊗ Ro ( π2 ))e, where e =
(· · · , eTij , · · · )T , the order of eij in the vector e is the same as the one of gij in
the vector g. It follows that
π
Pij
π
η T RB (p)(In ⊗ Ro ( ))p = η T diag(
)H̄(In ⊗ Ro ( ))p
2
||eij ||
2
P
π
P
π
ij
ik
T
T
= gik
Ro ( )eij + gij
Ro ( )eik
||eij ||
2
||eik ||
2
π
π
T
T
T
T
= gik (I − gij gij )Ro ( )gij + gij (I − gik gik
)Ro ( )gik
2
2
π
π
T
T
= gik Ro ( )gij + gij Ro ( )gik
2
2
π
T
T π
= gik (Ro ( ) + Ro ( ))gij = 0.
2
2
7
This completes the proof.
A direct consequence of Lemma 3 is the following result.
Lemma 4. A framework (G, p) is infinitesimally angle rigid if and only if null(RTG (p)) =
S.
−1
In [18], the authors showed that the set DK
(DK (p)), which includes all
configurations having inter-distance congruent to p, is always a manifold of
dimension 3. In fact, since an angle-constrained shape has at least 4 degrees of
−1
freedom, fK
(fK (p)) is a manifold of dimension 4 when (K, p) is infinitesimally
angle rigid (i.e., p is a regular point). See the following theorem.
Theorem 1. Let Sp , {q ∈ R2n : q = c(In ⊗ R)p + 1n ⊗ ξ, R ∈ O(2), c ∈
−1
R \ {0}, ξ ∈ R2 }. If (K, p) is infinitesimally angle rigid, then fK
(fK (p)) = Sp ,
and Sp is a 4-dimensional manifold.
The proof will be presented in later subsections.
With the aid of Theorem 1, we can derive the relationship between infinitesimal angle rigidity and angle rigidity for a given TG∗ , which is as follows.
Theorem 2. If (G, p) is infinitesimally angle rigid, then (G, p) is angle rigid.
Proof: By [18, Proposition 2], there is a neighborhood U of p, such that
−1
fT−1∗ (fTG∗ (p)) ∩ U is a manifold of dimension 4. From Theorem 1, fK
(fK (p)) is
G
−1
also a 4-dimensional manifold. As a result, fT−1∗ (fTG∗ (p)) and fK
(fK (p)) coincide
G
in U , implying that (G, p) is angle rigid.
The converse of Theorem 2 is not true. A typical counter-example is the
framework (K, p) with p being a degenerate configuration. In this case, (K, p)
is globally angle rigid but not infinitesimally angle rigid.
3.1. The Relation to Bearing Rigidity
In this subsection, we will establish some connections between angle rigidity
and bearing rigidity [24, 31]. The following theorem shows the equivalence of
infinitesimal angle rigidity and infinitesimal bearing rigidity in the plane, which
also implies the feasibility of angle-based approach for determining a planar
framework.
Theorem 3. A framework (G, p) is infinitesimally angle rigid if and only if it
is infinitesimally bearing rigid.
Proof: Necessity. Since null(RTG∗ ) = (null(Rg )∩range(RB ))∪null(RB ), dim(null(RTG∗ ))
reaches its minimum only if dim(null(RB )) is minimal. Recall that it always
holds that null(RB ) ⊇ Ss ∪ St , once (G, p) is infinitesimally angle rigid, it must
hold that null(RB (p)) = Ss ∪ St . That is, (G, p) is infinitesimally bearing rigid.
Sufficiency. Note that RTG∗ = Rg RB , and the infinitesimal bearing rigidity
implies null(RB ) = St ∪ Ss . To show null(RTG∗ ) = S, it suffices to show that for
any η ∈ null(Rg ) ∩ range(RB ), we have η = RB q for some q ∈ Sr .
8
Suppose η = RB v and Rg η = Rg RB v = 0 for some v = (v1T , · · · , vnT )T ∈ R2n .
T
Let gij
gik be a component of fG such that gij and gik are not collinear. Then
RTG∗ v = 0 implies that
T
∂gij
gik
∂g
P
diag( ||eijij|| )H̄v = 0, which is equivalent to
eTik Pij (vi − vj ) + eTij Pik (vi − vk ) = 0.
(6)
Note that for any nonzero vectors x, y ∈ R2 , P (x)y is perpendicular to x.
Therefore, there always exist cij , cik ∈ R such that
π
π
Pij (vi − vj ) = cij Ro ( )gij , Pik (vi − vk ) = cik Ro ( )gik .
2
2
(7)
It follows that
π
π
vi − vj = cij Ro ( )gij + c0ij gij , vi − vk = cik Ro ( )gik + c0ik gik
2
2
(8)
for some c0ij , c0ik ∈ R. Substituting (7) into (6), we have
π
π
cij eTik Ro ( )gij + cik eTij Ro ( )gik = 0.
2
2
Note also that RoT ( π2 ) = −Ro ( π2 ), then we have
π
T
(cij ||eik || − cik ||eij ||)gij
Ro ( )gik = 0.
2
T
Since gij and gik are not collinear, gij
Ro ( π2 )gik 6= 0. It follows that cij ||eik || =
cik ||eij ||. That is, cij = cijk ||eij ||, cik = cijk ||eik || for some cijk ∈ R. Together
with (8), we have
π
π
vi − vj = cijk Ro ( )eij + c̄ij eij , vi − vk = cijk Ro ( )eik + c̄ik eik ,
2
2
(9)
where c̄ij = c0ij /||eij ||, c̄ik = c0ik /||eik ||.
So far we have proved that if (i, j, k) ∈ TG∗ and gij is not collinear with
gik , then (9) holds for some cijk ∈ R. In the following, by constructing a TG∗ ,
we will show that there exists a common constant c ∈ R such that vi − vj =
cRo ( π2 )eij + c̄ij eij for all (i, j) ∈ E.
Now we construct a set TG∗ ⊆ TG such that gij and gik are not collinear for
all (i, j, k) ∈ TG∗ . Since (G, p) is infinitesimally bearing rigid, from Lemma 1 and
Lemma 2, for any vertex i, there exist at least two neighbors j, k ∈ Ni such
that gij and gik are not collinear. As a result, we can divide Ni into two sets
N̂i and Ňi , such that for any j ∈ N̂i and k ∈ Ňi , gij and gik are not collinear.
We construct a set i TG∗ by the following two steps:
Step 1. Select a vertex j1 ∈ N̂i randomly, let (i, j1 , k)(if j1 < k) or (i, k, j1 )(if
j1 > k) for all k ∈ Ňi be an element of i TG∗ .
Step 2. Select a vertex k1 ∈ Ňi randomly, let (i, j, k1 )(if j < k1 ) or (i, k1 , j)(if
j > k1 ) for all j ∈ N̂i \ {j1 } be an element of i TG∗ .
An example of constructing i TG∗ is shown in Fig. 2.
9
&i
&i
k1
k1
j1
k2
k2
i
j1
j2
j2
k3
k3
(a)
(b)
Figure 2: An example to illustrate how to construct i TG∗ . (a) The subgraph composed of
vertex i and its neighbors j1 , j2 , k1 , k2 , k3 . Note that i, k2 , k3 are collinear, i, j1 , j2 are
collinear. (b) N̂i = {j1 , j2 }, Ňi = {k1 , k2 , k3 }. If js and kl are connected by a red line, it
implies that the angle between edge (i, js ) and edge (i, kl ) is selected to be an angle constraint.
This also implies that (i, js , kl ) (if js < kl ) or (i, kl , js ) (if kl < js ) is an element of i TG∗ .
Let TG∗ = ∪i∈V i TG∗ . It is obvious that for any i, j, k ∈ TG∗ , gij and gik are not
collinear. Now we regard each edge (i, j) of G as a vertex of G 0 , (i, j) and (i, k) are
adjacent if (i, j, k) or (i, k, j) belongs to TG∗ . By our approach for construction
of i TG∗ , it is easy to see that for any i ∈ V and j, k ∈ Ni , (i, j) and (i, k) are
either adjacent or both neighbors of (i, j1 ) or (i, k1 ). Therefore, the graph G 0
corresponding to TG∗ is connected. We regard cij as the state corresponding to
(i, j) if vi − vj = cij Ro ( π2 )eij + c̄ij eij . Note that (9) implies that if (i, j) and
(i, k) are adjacent, they share a common state cijk ∈ R. Since G 0 is connected,
all edges in G 0 have a consensus state c ∈ R. That is, vi −vj = cRo ( π2 )eij + c̄ij eij
for all (i, j) ∈ E.
This implies that H̄v = c(Im ⊗ Ro ( π2 ))H̄p + C̄ H̄p, where C̄ = diag(c̄ij ) ⊗ I2 .
Then
Pij
)H̄v
||eij ||
Pij
π
Pij
= diag(
)c(Im ⊗ Ro ( ))(H ⊗ I2 )p + diag(
)C̄ H̄p
||eij ||
2
||eij ||
Pij
π
Pij
= diag(
)(H ⊗ I2 )c(In ⊗ Ro ( ))p + diag(
) diag(c̄ij eij )
||eij ||
2
||eij ||
π
= RB c(In ⊗ Ro ( ))p.
2
η = RB v = diag(
Since c(In ⊗ Ro ( π2 ))p ∈ Sr , the proof is completed.
Remark 1. In [31], the authors proved that infinitesimal bearing rigidity is a
generic property of the graph. That is, if (G, p) is infinitesimally bearing rigid,
then (G, q) is infinitesimally bearing rigid for almost all configuration q. From
Theorem 3, it is easy to obtain that infinitesimal angle rigidity is also a generic
property of the graph, thus is primarily determined by the graph, rather than the
configuration.
10
3
2
2
1
1
4
3
4
(a) (G, p)
(b) (G, q)
Figure 3:
Both (G, p) and (G, q) are infinitesimally angle rigid for
{(1, 2, 4), (4, 1, 2), (1, 3, 4), (4, 1, 3)}, globally angle rigid for T̄G∗ = {(1, 2, 3)} ∪ TG∗ .
TG∗
=
Remark 2. From Definition 4, we can conclude that the minimal number of
angle constraints for achieving infinitesimal angle rigidity is 2n − 4. This fact
has also been shown in [21]. On the other hand, it has been shown in [24] that
the minimal number of edges for a framework to be infinitesimally bearing rigid
is 2n − 3. By Theorem 3, the same is true for infinitesimal angle rigidity.
Remark 3. In the proof of sufficiency in Theorem 3, by constructing an appropriate TG∗ , we show that if η ∈ null(Rg ) ∩ range(RB ), then η must correspond to a rotational motion, i.e., η = RB v, v ∈ Sr . Nevertheless, the
TG∗ constructed by us may not be suitable to determine a unique shape. As
shown in Fig. 3 (a), a TG∗ constructed by the approach in Theorem 3 is TG∗ =
{(1, 2, 4), (4, 1, 2), (1, 3, 4), (4, 1, 3)}. Although (G, p) with TG∗ is infinitesimally
angle rigid, fTG∗ (p) may determine an incorrect shape in Fig. 3 (b). However, if
we let T̄G∗ = {(1, 2, 3)} ∪ TG∗ , then fT̄G∗ (p) can always determine a correct shape.
This implies that (G, p) in Fig 3 (a) is both infinitesimally and globally angle
rigid for T̄G∗ (p). Moreover, we can conclude that for given TG∗ , infinitesimal
angle rigidity cannot induce global angle rigidity directly.
Consider a framework (G, p) in the plane. For distance rigidity theory, it
is obvious that the shape of (G, p) can be uniquely determined by DG (p) if
G = K. For bearing rigidity theory, the authors of [24] showed that BG (p)
uniquely determines a shape if G is infinitesimally bearing rigid. However, for
angle rigidity theory, it cannot be immediately answered that whether the shape
can be uniquely determined by angles between edges. On one hand, angles are
only constraints on relationships between those edges joining a common vertex.
Even for a complete graph, if n > 3, there always exist disjoint edges, the angle
between each pair of disjoint edges cannot be constrained directly. On the other
hand, it has been shown in Remark 3 that an infinitesimal angle rigid framework
with specified TG∗ cannot determine a unique shape in particular cases.
−1
−1
In the following theorem, the connection between fK
(fK (p)) and BK
(BK (p))
is established.
−1
Theorem 4. Given configurations p, q ∈ R2n , q ∈ fK
(fK (p)) if and only if
−1
−1
(In ⊗ R) q ∈ BK (BK (p)), where R ∈ O(2).
11
1
2
1
3
6
5
4
2
5
4
6
(a) (G, p)
3
(b) (G, q)
Figure 4: fG (p) = fG (q), but there does not exist R ∈ O(2) such that BG (p) = (Im ⊗R)BG (q).
The angles in red are all constrained angles determined by TG .
Remark 4. It is easy to see that the complete graph K in Theorem 4 can be
replaced with G provided (G, p) is both globally angle rigid and globally bearing
rigid. Note that once K is replaced with a general graph G, Theorem 4 will not
hold any more. As shown in Fig. 4, although q ∈ fG−1 (fG (p)), there does not
exist R ∈ O(2) such that q ∈ (In ⊗ R)−1 q ∈ BG−1 (BG (p)).
Next we present some lemmas that are required to prove Theorem 4.
In [32], the authors showed that for a positive semi-definite matrix A ∈
Rn×n with rank(A) = r, if ΠT AΠ = RT R for a specified permutation matrix
Π ∈ Rn×n , where R ∈ Rr×n , then this Cholesky decomposition is unique. Here
the uniqueness of Cholesky decomposition implies that if R̄T R̄ = ΠT AΠ for
some R̄ ∈ Rr×n , then R = R R̄ for some R ∈ O(2). It is straightforward to
obtain the following lemma.
Lemma 5. For a matrix R ∈ Rr×n with rank(R) = r, if RT R = R̄T R̄ for some
R̄ ∈ Rr×n , then R = R R̄ for some R ∈ O(r).
Let Hx = H (x) , I − 2xxT be a Householder transformation, here x ∈ R2
is a unit vector. Geometrically, Hx y with y ∈ R2 is a reflection of y about the
vector which is perpendicular to x. We list some easily checked properties of
Hx in the following lemma.
Lemma 6. For any given unit vectors x, y ∈ R2 , Hx has the following properties:
(i) HxT = H , Hx2 = I2 ;
(ii) Hx = Re (θ) for some θ ∈ [0, 2π);
(iii) For any θ ∈ [0, 2π), there exists a unit vector z ∈ R2 such that Hz =
Re (θ);
(iv) The eigenspace of Hx associated with the eigenvalue 1 is span{x⊥ }.
With the aid of Lemma 6, we can establish the following result.
Lemma 7. If Aη = Bη, where A, B ∈ O(2), and η ∈ R2 is a unit vector, then
A = B or A = BHη⊥ .
12
Proof: Note that a 2-dimensional orthogonal matrix is either a rotation matrix
or a reflection matrix. Without loss of generality, we discuss the problem in
three cases:
Case 1. A = Ro (α), B = Ro (β) for some α, β ∈ [0, 2π). Then Ro (α)η =
Ro (β)η, implying Ro (α − β)η = η. Hence α − β = 0. That is, A = B.
Case 2. A = Re (α), B = Re (β) for some α, β ∈ [0, 2π). Following the same
procedure in Case 1, one can also obtain A = B.
Case 3. A = Ro (α), B = Re (β) = Ro (β)I¯ for some α, β ∈ [0, 2π). Then
¯ It follows that η = Ro (β − α)Iη
¯ = Re (β − α)η. From
Ro (α)η = Ro (β)Iη.
2
Lemma 6 (iii), there exists some x ∈ R such that Hx = Re (β − α). That is,
η = Hx η. Using (iv) in Lemma 6, we have η ∈ span{x⊥ }. Then x = ±η ⊥ ,
¯ implying that
Hx = Hη⊥ . As a result, Hη⊥ = Re (β − α) = Ro (β − α)I,
Ro (α)Hη⊥ = Re (β). By (i) in Lemma 6, we have A = Ro (α) = Re (β)Hη⊥ =
BHη⊥ .
Let F denote a graph with 4 vertices and 5 edges, then the following lemma
holds.
Lemma 8. (F, p) is infinitesimally bearing rigid if and only if p is nondegenerate.
The necessity of Lemma 8 is obvious. The sufficiency of Lemma 8 can be
obtained from Lemma 9 to be given in the next subsection.
With Lemmas 5, 6, 7 and 8 at hand, we now give the proof for Theorem 4.
−1
Proof of Theorem 4: Note that (In ⊗ R)−1 q ∈ BK
(BK (p)) is equivalent to
−1
BK ((In ⊗R) q) = BK (p), which is also equivalent to BK (q) = (Im ⊗R)BK (p).
Therefore, it suffices to show that fK (q) = fK (p) if and only if BK (q) = (Im ⊗
R)BK (p).
Sufficiency. For any i, j, k ∈ V, it is straightforward that
T
T
T
gij
(q)gik (q) = gij
(p)R T Rgik (p) = gij
(p)gik (p).
To prove the necessity, we consider the following two cases.
Case 1. The configuration p is degenerate. Let g̃ be a unit vector such
that g̃ is collinear with gij (p) for all i, j ∈ V, then gij (q) = Rgij (p) if and
only if gij (q) = RHg̃⊥ (p) gij (p). For any i, j ∈ V, let Rij ∈ O(2) such that
gij (q) = Rij gij (p). To prove the necessity, it suffices to show that for any
distinct vertices i, j, k ∈ V, if gij (q) = Rij gij (p) and gik (q) = Rik gik (p), there
always holds Rij = Rik or Rij = Rik Hg̃⊥ (p) . Without loss of generality,
T
T
suppose gij (p) = gik (p). Then gij
(q)gik (q) = gij
(p)gik (p) = 1, which holds
if and only if gij (q) = gik (q), i.e., Rij gij (p) = Rik gik (p) = Rik gij (p). By
Lemma 7, Rij = Rik or Rij = Rik Hgij
⊥ (p) . Since Hg ⊥ (p) = Hg̃ ⊥ (p) , the proof
ij
is completed.
Case 2. The configuration p is nondegenerate. Note that K is complete,
hence each vertex i has at least two neighbors j and k such that gij (p) and
gik (p) are not collinear. As a result, we can divide Ni into two sets N̂i and
Ňi , such that for any j ∈ N̂i and k ∈ Ňi , gij (p) and gik (p) are not collinear.
13
We first show that given i ∈ V, for any j ∈ N̂i , k ∈ Ňi , l ∈ V \ {i, j, k},
it always holds that Gijkl (q) = Rijkl Gijkl (p) for some Rijkl ∈ O(2), where
Gijkl = (gij , gik , gil , gjk , gjl , gkl ) ∈ R2×6 .
Since K is complete, we have l ∈ Ni . Without loss of generality, we consider
l ∈ Ňi . For the triangle composed of i, j, k, let Gijk = (gij , gik , gjk ) ∈ R2×3 .
Since fK (q) = fK (p), we have GTijk (q)Gijk (q) = GTijk (p)Gijk (p). Note that
gij (p) and gik (p) are not collinear, thus we have rank(Gijk (p)) = 2. By virtue of
Lemma 5, the Cholesky decomposition of GTijk (p)Gijk (p) determines Gijk (p) up
to a 2 × 2 orthogonal matrix Rijk . That is, Gijk (q) = Rijk Gijk (p). Similarly,
we have Gijl (q) = Rijl Gijl (p) for Rijl ∈ O(2). For vertices j, k, l, it follows
from Case 1 that Gjkl (q) = Rjkl Gjkl (p) for Rjkl ∈ O(2) no matter j, k, l are
collinear or not. Since Rijk gij (p) = Rijl gij (p) = gij (q), according to Lemma 7,
Rijk = Rijl or Rijk = Rijl Hgij
⊥ (p) .
Suppose that Rijk 6= Rijl , then
T
T
gjk
(p)gjl (p) = gjk
(q)gjl (q)
T
T
= gjk
(p)Rijk
Rijl gjl (p)
T
= gjk
(p)Hgij
⊥ (p) gjl (p)
T
T
⊥
⊥T
= gjk
(p)gjl (p) − 2gjk
(p)gij
(p)gij
(p)gjl (p).
T
⊥
⊥T
(p)gij
(p)gij
(p)gjl (p) = 0. Since gij (p) and gik (p) are not
This implies gjk
collinear, gij (p) and gjk (p) are also not collinear. Similarly, gij (p) and gjl (p) are
not collinear. Thus a contradiction arises. We then have Rijk = Rijl , Rijkl ,
which implies that Ḡijkl (q) = Rijkl Ḡijkl (p), where Ḡijkl = (gij , gik , gil , gjk , gjl ) ∈
R2×5 . Consider the framework (F, q), where F is a graph with vertex set
{i, j, k, l} and edge set {(i, j), (i, k), (i, l), (j, k), (k, l)}. Since these four vertices
are not collinear, according to Lemma 8, (F, q) is infinitesimally bearing rigid,
thus is globally bearing rigid. This implies that gkl (q) can be uniquely determined by Ḡijkl (q). As a result, Gijkl (q) = Rijkl Gijkl (p).
The above proof implies that given i ∈ V, it holds that
R , Rijk = Rijl = Rikl = Rjkl
for any j ∈ N̂i , k ∈ Ňi , l ∈ V \ {i, j, k}. Note that any edge in graph K is
involved in a triangle including vertex i. Therefore, gij (q) = Rgij (p) for any
(i, j) ∈ E.
It is important to note that Theorem 4 cannot induce equivalence of global
angle rigidity and global bearing rigidity. Some examples show that this equivalence holds, but we still have no idea of how to prove it. Nonetheless, we are
able to establish the following result.
Theorem 5. If a framework (G, p) is (globally) angle rigid, then it is (globally)
bearing rigid.
Proof: Suppose (G, p) is angle rigid. Then there exists a neighborhood Up of p
−1
such that fG−1 (fG (p)) ∩ Up = fK
(fK (p)) ∩ Up . For this Up , consider any q ∈
14
BG−1 (BG (p))∩Up . It follows from BG (p) = BG (q) that fG (p) = fG (q). Therefore,
fK (p) = fK (q). By Theorem 4, BK (p) = (Im ⊗ R)BK (q) for some R ∈ O(2).
Recall that BG (p) = BG (q), we have R = I2 . As a result, BK (p) = BK (q), i.e.,
−1
q ∈ BK
(BK (p)). That is, (G, p) is bearing rigid.
From [24], bearing rigidity is equivalent to global bearing rigidity. Since
global angle rigidity obviously leads to angle rigidity, it can also induce global
bearing rigidity.
To prove Theorem 1, we introduce the following theorem in [33].
Theorem 6. ([33] Constant-Rank Level Set Theorem) Let M and N be smooth
manifolds, and let Φ : M → N be a smooth map, the Jacobian matrix of Φ has
constant rank r. Each level set of Φ is a properly embedded submanifold of
codimension r in M .
Proof of Theorem 1: From Theorem 3, (K, p) is infinitesimally bearing rigid.
−1
[24] shows that BK
(BK (p)) = {q ∈ R2n : q = cp + 1n ⊗ ξ, c ∈ R \ {0}, ξ ∈ R2 }.
−1
Together with Theorem 4, there must hold fK
(fK (p)) = Sp .
−1
Next we show Sp is a manifold. For any q ∈ fK
(fK (p)), it is obvious that
q = (In ⊗ R)(cp + 1n ⊗ ξ) for some R ∈ O(2), scalar c and vector ξ ∈ R2 . From
the chain rule, we have
rank
∂fK (p) 1
∂fK (p)
∂fK (q)
= rank
= rank
(In ⊗ R −1 ) = 2n − 4.
∂q
∂c(In ⊗ R)p
∂p c
Note that fK : R2n → R|TK | is a smooth map, according to Theorem 6,
−1
fK
(fK (p)) is a properly embedded submanifold of dimension 2n − (2n − 4) = 4.
In fact, even for a complete graph, it is possible that the geometric shape
cannot be determined by angle-only or bearing-only information. A typical
example is the degenerate configuration shown in Fig. 1 (d). Generally, we
hope to determine a framework (G, p) by angles uniquely up to translations,
rotations, scalings and reflections in the plane. That is, fG−1 (fG (p)) = Sp . In
the next subsection, we will introduce a specific class of frameworks satisfying
this condition.
3.2. A class of frameworks uniquely determined by angles
In [17], the authors introduced a particular class of Laman graphs termed
triangulated Laman graphs, which are constructed by a modified Henneberg
insertion procedure. In what follows, we will show that the shape of such frameworks can always be determined by angles uniquely. Let Ln = (Vn , En ) be an
n−point(n ≥ 3) triangulated Laman graph, its definition is as follows.
Definition 5. Let L2 be the graph with vertex set V2 = {1, 2} and edge set
E2 = {(1, 2)}. Ln (n ≥ 3) is a graph obtained by adding a vertex n and two
edges (n, i), (n, j) into graph Ln−1 for some i and j satisfying (i, j) ∈ En−1 .
15
Note that the triangulated Laman graph considered here is always an undirected graph, that is, if (i, j) ∈ En , then it holds (j, i) ∈ En .
For frameworks associated with triangulated Laman graphs, [17] also presents
the following result.
Lemma 9. [17] If (Ln , p) is strongly rigid, i.e., pi , pj and pk are not collinear
for any three vertices i, j, k satisfying (i, j), (j, k), (i, k) ∈ En , then (Ln , p) is
infinitesimally distance rigid.
For n = 4, it is easy to see that (L4 , p) is strongly rigid if and only if p
is nondegenerate. Therefore, (L4 , p) is infinitesimally distance rigid. It follows
from Lemma 1 that (L4 , p) is infinitesimally bearing rigid. Note also that a
graph with 4 vertices and 5 edges is always a triangulated Laman graph. The
sufficiency of Lemma 8 then follows.
The following theorem shows that the shape of a strongly rigid framework
in the plane can always be uniquely determined by angles.
Theorem 7. If (Ln , p) is strongly rigid, then
(i) (Ln , p) is minimally infinitesimally angle rigid for
TL∗n = {(i, j, k) ∈ Vn3 : (i, j), (j, k), (i, k) ∈ En , i, j < k};
(10)
(ii) (Ln , p) is globally angle rigid for TL†n = TL∗n ∪ ∆TLn , where ∆TLn =
{(i, k, l) : k = min{Ni ∩ Nj ∩ Vl−1 }, i, j ∈ Nl , i < j < l, l = 4, · · · , n} if n ≥ 4,
and ∆TLn = ∅ otherwise;
(iii) fL−1
(fLn (p)) = Sp .
n
Proof: (i) From Lemma 9, (Ln , p) is infinitesimally bearing rigid. Then null(RB ) =
Ss ∪ St . It suffices to show that for any η ∈ null(Rg ) ∩ range(RB ), there always
exists q ∈ Sr such that η = RB q. Suppose that RTL∗n v = Rg RB v = 0 and
RB v 6= 0, where v = (v1T , · · · , vnT ) ∈ R2n . In the proof of Theorem 3, we have
shown that for any (i, j, k) ∈ TL∗n , if gij is not collinear with gik , then (9) holds
for some cijk . Recall that (Ln , p) is strongly rigid, then for any (i, j, k) ∈ TG∗ , (9)
holds for some cijk . Without loss of generality, suppose i < j < k. Due to the
definition in (10), for each triangle in Ln formed by vertices i, j and k, we have
(i, j, k), (j, i, k) ∈ TL∗n . Now we regard (i, j) as a vertex of G 0 for all (i, j) ∈ En ,
two vertices in G 0 are adjacent if they belong to a same triangle in G. Let cij
be the state of (i, j) if vi − vj = cij Ro ( π2 )eij + c̄ij eij for some c̄ij ∈ R. It is easy
to see that (i, j), (i, k) and (k, j) have a common state, implying that adjacent
vertices in G 0 must have a common state. Note that in every step during the
generation of graph Ln , a new triangle is generated based on an existing edge.
Therefore, G 0 must be connected. As a result, there exists a constant c ∈ R such
that vi − vj = cRo ( π2 )eij + c̄ij eij for all (i, j) ∈ En . By similar analysis to the
proof of Theorem 3, we can obtain v ∈ Sr , which implies that (Ln , p) is infinitesimally angle rigid for TG∗ . Moreover, observe that |En | = 1 + 2(n − 2) = 2n − 3,
we conclude that (Ln , p) is minimally infinitesimally angle rigid.
(ii). We prove the statement by induction.
16
For n = 3, it is obvious that (L3 , p) with TL†3 = TL∗3 = {(1, 2, 3), (2, 1, 3)} is
globally angle rigid.
For n ≥ 4, suppose that (Ln−1 , p) with TL†n−1 is globally angle rigid. Next
we show that (Ln , p) is globally angle rigid with TL†n . Without loss of generality,
let i and j be the neighbors of n and i < j. Note that (i, j) ∈ En−1 , and i, j
must have at least one common neighbor vertex in Ln−1 , let k be minimum
index among them, it is easy to see TL†n = {(i, j, n), (j, i, n), (i, k, n)} ∪ TL†n−1 .
It suffices to show that for any q such that fT † (p) = fT † (q), it always holds
Ln
Ln
fKn (p) = fKn (q). Since Ln−1 is globally angle rigid, by Theorem 4, there
exists a matrix Rn−1 ∈ O(2) such that gi0 j 0 (p) = Rn−1 gi0 j 0 (q) for all i0 , j 0 ∈
T
T
T
T
Vn−1 . From gij
(p)gin (p) = gij
(q)gin (q), and gji
(p)gjn (p) = gji
(q)gjn (q), we
T
T
T
have gni (p)gnj (p) = gni (q)gnj (q). That is, Gijn (p)Gijn (p) = GTijn (q)Gijn (q),
where Gijn = (gij , gnj , gni ) ∈ R2×3 . Using strong rigidity of (Ln , p), we have
rank(Gijn (p)) = 2. By Lemma 5, Gijn (p) = Rijn Gijn (q) for some Rijn ∈ O(2).
It follows that gij (p) = Rn−1 gij (q) = Rijn gij (q). According to Lemma 7,
Rn−1 = Rijn or Rn−1 = Rijn Hgij
⊥ (q) .
Suppose that Rn−1 6= Rijn , then gik (p) = Rn−1 gik (q) = Rijn Hgij
⊥ (q) gik (q),
and gin (p) = Rijn gin (q). It follows that
T
T
T
T
gik
(p)gin (p) = gik
(q)Hgij
⊥ (q) Rijn Rijn gin (q) = gik (q)Hg ⊥ (q) gin (q).
ij
(11)
T
T
Recall that (i, k, n) ∈ TL†n , we have gik
(p)gin (p) = gik
(q)gin (q). Together
T
T
with (11), it follows gik (q)Hgij
(q)g
(q),
which
holds if and only
⊥ (q) gin (q) = g
in
ik
if gij (q) is collinear with either gik (q) or gin (q), i.e., either (qiT , qjT , qkT )T or
(qiT , qjT , qnT )T is degenerate, implying that either (pTi , pTj , pTk )T or (pTi , pTj , pTn )T
is degenerate. This conflicts with strong rigidity of (Ln , p). Therefore, Rn−1 =
Rijn , Rn . That is, gi0 j 0 (p) = Rn gi0 j 0 (q) for any (i0 , j 0 ) ∈ En . It follows that
fK (p) = fK (q). Hence, (Ln , p) with TL∗n is globally angle rigid.
−1
(iii). By (i), (ii) and Theorem 1, fT−1
(fTLn (p)) = fK
(fK (p)) = Sp .
Ln
An example of strongly rigid framework embedded by a triangulated Laman
graph is shown in Fig. 5 (a). The angles in red are constrained angles determined
by TL†5 . The graph in Fig. 5 (b) contains a triangulated Laman graph but itself
is not a triangulated Laman graph.
4. Application to formation control
In this section, we will apply the angle-based approach for framework recognition to distributed formation control in the plane. The target formation will
be characterized by some constraints on angles. In order to form a desired
shape, the multiple mobile agents are required to meet these constraints via a
distributed controller.
17
1
1
2
5
3
4
2
5
3
4
(a)
(b)
Figure 5: (a) A framework embedded by a triangulated Laman graph L5 . (L5 , p) is infinitesimally angle rigid for TL∗5 = {(1, 2, 3), (1, 3, 4), (1, 4, 5), (2, 1, 3), (3, 1, 4), (4, 1, 5)}, and globally
angle rigid for TL†5 = TL∗5 ∪ {(1, 2, 4), (1, 3, 5)}. The angles in red are constrained angles de-
termined by TL†5 . (b) A framework embedded by a graph containing L5 . The framework is
globally and infinitesimally angle rigid for TGF , where TGF is in form (14). The angles in red
are constrained angles determined by TGF .
4.1. The Formation Stabilization Problem
Consider n agents moving in the plane, each agent i has a simple kinematic
point dynamics
ṗi = ui , i ∈ V,
(12)
where pi ∈ R2 and ui ∈ R2 are the position and control input of agent i, respectively, in the global coordinate frame. We consider that the global coordinate
system is absent for the agents, each agent i has its own local coordinate system. Let pij be the coordinate of agent j’s position with respect to agent i’s
local coordinate system. Agent i can measure the relative position state pii − pik
if k ∈ Ni .
In this paper, we employ an infinitesimally angle rigid framework (G, p∗ )
to describe the target formation shape. Each agent is viewed as a vertex of
the framework. An interaction link between two agents is regarded as an edge
in graph G. That is, G is also the sensing graph interpreting the interaction
relationship between agents.
By virtue of Theorem 1, the target formation shape can be defined as the
following manifold:
E = Sp∗ = {p ∈ R2n : p = c(In ⊗ R)p∗ + 1n ⊗ ξ, R ∈ O(2), c ∈ R \ {0}, ξ ∈ R2 }.
(13)
For the target formation (G, p∗ ), we make the following assumption:
Assumption 1. G contains a triangulated Laman graph Ln as a subgraph, and
(Ln , p∗ ) is strongly rigid.
The set determining all angle constraints is given by
TGF = {(i, j, k) ∈ V 3 : (i, j), (j, k), (i, k) ∈ E, i, j < k},
18
(14)
Remark 5. Assumption 1 is a graphical condition for (G, p∗ ), and will be the
only condition for achieving stability of the target formation. Once Assumption
1 holds, it is easy to see that TL∗n ⊂ TGF , where TL∗n is in form (10). Since
we have shown in Theorem 7 that (Ln , p∗ ) is infinitesimally angle rigid for TL∗n ,
together with En ⊂ E, it follows that (G, p∗ ) is infinitesimally angle rigid for TGF .
It is also worth noting that strongly rigid configurations form a dense subset of
R2n , which is shown in [17].
A framework (G, p) is said to be realizable if there exists some q ∈ R2n such
that fG (q) = fG (p). Throughout this paper, we always assume that the target
formation shape is realizable. Then the angle-based formation stabilization
problem can be summarized as follows.
Problem 1. Given a realizable formation (G, p∗ ) satisfying Assumption 1, design a distributed control law for each agent i based on the relative position
measurements {pii − pij , j ∈ Ni }, such that E is asymptotically stable.
4.2. A Steepest Descent Formation Controller
∗
T
∗T ∗
Denote gij = gij (p), gij
= gij (p∗ ) for all (i, j) ∈ E, δ(i,j,k) = gij
gik − gij
gik ,
∗
(i, j, k) ∈ TG . To form a desired geometric shape, the multi-agent system should
minimize the following cost function:
VF (p) =
1
2
X
T
∗T ∗ 2
(gij
gik − gij
gik ) =
(i,j,k)∈TG∗
1
2
X
2
δ(i,j,k)
.
(15)
(i,j,k)∈TG∗
On the basis of function (15), a gradient-based control strategy can be derived
as
X
X
T
uF
y1 (eij , eik )−
y2 (eji , ejk ), i ∈ V (16)
i = −(∇pi VF (p)) = −
(j,k)∈NTi
(j,k)∈NT i
where NTi = {(j, k) ∈ V 2 : (i, j, k) ∈ TGF }, NT i = {(j, k) ∈ V 2 : (j, i, k) or (j, k, i) ∈
P
T
∗T ∗
T
ik
gik − gij
gik )( ||eijij|| gik + ||ePik
TGF }, y1 (eij , eik ) = (gij
|| gij ), y2 (eji , ejk ) = (gji gjk −
P
∗T ∗
gji
gjk ) ||eijij|| gkj .
Observe that if (j, i, k) ∈ TGF , the control input of agent i includes a term
associated with ejk . This can be obtained by simple subtraction eik − eij . From
the form of TGF in (14), we have k, j ∈ Ni . Therefore, it is reasonable for agent
i to use ejk .
We define the desired equilibrium for the formation system as
T
T
EF = {p ∈ R2n : gij
(p)gik (p) = gij
(p∗ )gik (p∗ ), (i, j, k) ∈ TGF }.
(17)
It is easy to see EF is only a subset of E . EF = E if and only if (G, p∗ )
is globally angle rigid for TGF . For example, in Fig. 5 (b), the framework is
globally and infinitesimally angle rigid for TGF , thus it holds EF = E . However,
the framework in Fig. 5 (a) is infinitesimally angle rigid but not globally angle
19
rigid. In this scenario, even if all the angle constraints determined by TGF are
satisfied, it is possible that the target formation shape is not formed. Nonetheless, from the definition of infinitesimal angle rigidity, for any q ∈ E , there exists
a neighborhood U of q, such that E ∩ U = EF ∩ U . Hence, stability of EF can
still be sufficient for local stability of E .
Let δ(p) = (· · · , δ(i,j,k) , · · · )T = fTGF (p) − fTGF (p∗ ), (i, j, k) ∈ TG∗ . Then
δ(p) = 0 if and only if p ∈ EF , and (15) can be written as V (p) = 21 ||δ||2 . By the
chain rule, the multi-agent system (12) with control input (16) can be written
in the following compact form
ṗ = −(∇p VF (p))T = −(δ T (p) ·
∂fTGF (p)
∂p
)T = −RTT F (p)δ(p).
G
(18)
Note that uF
i only requires agent i to sense real-time relative displacements
from its neighbors. Therefore, uF
i is a distributed control strategy. Moreover,
by implementing uF
i , the formation system has the following properties.
Lemma 10. Under the control law (16), the following statements hold:
(i) The global coordinate system is not required for each agent.
(ii) If p(0) is degenerate, then p(t) = p(0) for t ≥ 0.
q P
P
(iii) The centroid po (t) = n1 i∈V pi (t) and the scale s(t) = n1 i∈V ||pi (t) − po (t)||2
are both invariant.
The proof of (i) is straightforward by a similar approach to those of [13, 16],
the validity of (ii) is also easy to verify. Thus their proofs are omitted here. For
(iii), observe that po (t) = n1 pT (t)1n , according to (18), we have
ṗo =
1 T
1
ṗ 1n = δ T RTGF 1n = 0.
n
n
P
To show ṡ = 0, we first note that ṗT p = δ T RTGF p = δ T Rg diag( ||eijij|| )H̄p = 0.
It follows that
2
Pij
2
)H̄(1n ⊗ po ) = 0.
ṡ = √ ṗT (p − 1n ⊗ po ) = − √ δ T diag(
||eij ||
n
n
4.3. Stability analysis
Theorem 8. For a group of n ≥ 3 agents with dynamics (12) and controller
(16) moving in the plane. Under Assumption 1, for any q ∈ E , there is a
neighborhood Uq of q, such that if p(0) ∈ Uq , then limt→∞ p(t) = p† for some
p† ∈ E .
Proof: For any q ∈ E , let ρ = p − q, f (p) = −RTT F (p)δ(p), expanding f (p)
G
in Taylor series about q, we have f (p) = f (q) + ∂f∂p(q) ρ + g(ρ). Then (18) is
equivalent to
∂f (q)
ρ̇ =
ρ + g(ρ) = Jf (q)ρ + g(ρ),
(19)
∂p
20
where Jf (q) =
∂f (q)
∂p
=−
T
∂RT
F (q)
G
∂p
T
δ(q) − RTT F (q) ∂δ(q)
∂p = −RT F (q)RTGF (q). From
G
G
Lemma 9, the validity of Assumption 1 implies that (G, q) is infinitesimally angle
rigid. Therefore, Jf (q) has 4 zero eigenvalues and the rest are negative real
numbers. There must exist an orthonormal transformation Q ∈ R2n×2n such
˜ where J˜ is Hurwitz. Then (19) is equivalent
that QJf (q)QT = diag(04×4 , J),
to
ϕ̇ = g1 (ϕ, ψ),
˜ + g2 (ϕ, ψ),
ψ̇ = Jψ
(20)
where (ϕT , ψ T )T = Qρ, (g1T , g2T )T = Qg(ρ). Note that ρ = 0 is an equilibrium
point of (19), hence, g1 (0, 0) = 0 and g2 (0, 0) = 0. Since g(ρ) = f (p) − Jf (q)ρ,
we have Jg |ρ=0 = 0. It follows that Jg1 (0, 0) = 0 and Jg2 (0, 0) = 0. Observe that
M = {(ϕT , ψ T )T : (ϕT , ψ T )T = Qρ, ρ + q ∈ E } is a 4-dimensional manifold.
We next show M is a center manifold. Note that M is invariant since E ⊂
˜ + g2 (ϕ, ψ) = 0,
EF ⊂ {p ∈ R2n : f (p) = 0}. Any equilibria must satisfy Jψ
by implicit function theorem, there is a neighborhood U of the origin, such
that ψ = h(ϕ) for ψ ∈ U , where h(·) : R4 → R2n−4 is smooth and h(0) = 0,
Jh (0) = 0. Since M is a 4-dimensional manifold, there must exist an open
set in R4 that is diffeomorphic to a neighborhood U 0 of origin in M. Because
M can be represented by (ϕT , hT (ϕ))T in U ∩ U 0 , we conclude that M is a
center manifold. The flow on the manifold M is governed by the 4-dimensional
system ξ˙ = g1 (ξ, h(ξ)) for sufficiently small ξ. Recall that M is a manifold of
equilibria, we have ξ˙ = 0. By center manifold theory [35], for any sufficiently
small ϕ(0), ψ(0), we have ϕ(t) = ξ(0) + O(e−γt ), ψ = h(ξ(0)) + O(e−γt ) for
some γ > 0. This implies that limt→∞ (ϕT , ψ T )T = (ξ T (0), hT (ξ(0)))T ∈ M. It
follows that limt→∞ p(t) = limt→∞ QT (ϕT , ψ T )T + q = QT (ξ T (0), hT (ξ(0)))T +
q = p† ∈ E . The proof is completed.
Observe that the equilibrium set of system (18) is E = {p ∈ R2n : RTT F (p)δ =
G
F
0}. If RTGF ∈ R|TG |×2n is of full row rank, then p ∈ E implies δ = 0, it follows
that p ∈ EF . However, RTGF (p) varies as the formation system evolves, it is
difficult to determine its rank. Moreover, once |TGF | > 2n − 4, RTGF can never
be of full row rank. As a result, undesired equilibria often exist for system
(18). Hence we can only obtain local stability of EF , which further implies local
stability of E . Theorem 8 actually means that by implementing the control law
(16), the agents can cooperatively restore the desired formation shape under a
small perturbation from any q ∈ E , and the convergence rate is as fast as e−γt
for some γ > 0 dependent on q. However, it is uncertain that whether there
exists a uniform exponent γ for all q ∈ E . This is because E is not compact,
there does not exist a finite subcover including E .
4.4. Orientation and Scale Control
We have shown that the angle-constrained formation has 4 degrees of freedom, which is higher than that of displacement-, distance-, and bearing-based
21
formations. This ensures that one advantage of the angle-based formation approach is the convenience of orientation and scale control. In this subsection, we
will propose an angle-based control scheme to steer all agents to form a target
formation shape with prescribed orientation and scale.
Given (G, p∗ ) as the target formation shape satisfying Assumption 1, a configuration forming the target formation with desired orientation and scale can
be written by p = c∗ (In ⊗ Ro (θ∗ ))p∗ + 1n ⊗ ξ for some constant θ∗ ∈ [0, 2π),
c∗ ∈ R \ {0} and an arbitrary translational vector ξ ∈ R2 . It is worth noting
that p∗i here denotes the position of agent i in the global coordinate frame. Let
p̃ = c∗ (In ⊗ Ro (θ∗ ))p∗ , then the target equilibrium can be described as
EM = {p ∈ R2n : p = p̃ + 1n ⊗ ξ, ξ ∈ R2 }.
(21)
To control the orientation of the formation, it is obviously necessary that
some agents should have access to the global coordinate system. To keep the
target shape in a precise orientation, we will try to constrain the displacement
between two adjacent agents, which is similar to [16]. Since orientation and
scale of the ultimate formation is determined by these two agents, we call them
leaders. It is noteworthy that any two adjacent agents can be selected as leaders,
and controlling their relative position is sufficient to control the orientation and
scale of the formation (this fact will be shown later). Moreover, different from
[16], using angle-based approach, the target displacement between leaders can
be artificially specified and does not have to satisfy a fixed length constraint.
That is, not only orientation, but also scale of the target formation can be
controlled.
Suppose agents l1 and l2 are leaders, l1 , l2 ∈ V. Then p̃l1 − p̃l2 is the displacement of l1 and l2 in the formation with target orientation and scale. Now
we summarize the problem that we will deal with in this subsection as below.
Problem 2. Given a realizable target formation (G, p∗ ) satisfying Assumption
1, and the target displacement p̃l1 − p̃l2 known by agents l1 and l2 , design a distributed control law for each agent i based on the relative position measurements
{pii − pij , j ∈ Ni }, such that EM is asymptotically stable.
To solve Problem 2, we consider the following set containing the target
equilibrium EM
El = {p ∈ EF : pl1 − pl2 = p̃l1 − p̃l2 }.
(22)
The following lemma shows that once (G, p∗ ) is infinitesimally angle rigid,
El and EM coincide near each point in EM .
Lemma 11. If (G, p∗ ) is infinitesimally angle rigid, then for any q ∈ EM , there
exists a neighborhood Uq of q, such that EM ∩ Uq = El ∩ Uq .
F
f (p)
Proof: Let fl (p) =
∈ R|TG |+2 , fM (p) = (· · · , (pi − pj )T , · · · )T ∈
pl1 − pl2
−1
R2m , it follows that El = fl−1 fl (p̃), EM = fM
fM (p̃). Since G must be connected,
22
we have rank( ∂f∂pM ) = rank(H̄) = 2n − 2, according to Theorem 6, EM is a 2dimensional manifold.
Next we show El is also a 2-dimensional manifold near each q ∈ EM . Without
loss of generality, suppose pl1 − pl2 is consisted of the (2k − 1)-th row and 2k-th
row of H̄p. Let S = [Sij ] ∈ R2×2m be a matrix with S1,2k−1 = S2,2k = 1,
∂(pl1 −pl2 )
= S H̄.
and Sij = 0 for other i, j. Then S H̄p = pl1 − pl2 and
∂p
For any q ∈ EM , define Rl (q) =
∂fl
∂p |p=q
= (RTT F (q), (S H̄)T )T , it is easy
G
to obtain null(Rl (q)) = null(RTGF (q)) ∩ null(S H̄). We first notice that (G, q)
must be infinitesimally angle rigid, implying null(RTGF (q)) = S, where S is
the trivial motion space shown in Lemma 3. We also note that null(S H̄) =
(null(S) ∩ range(H̄)) ∪ null(H̄). It can be verified that null(Rl (q)) = S ∩
l
null(S H̄) = null(H̄) = span{1n ⊗I2 }. Then we obtain rank( ∂f
∂p |p=q ) = 2n−2 =
2n
l
max{rank( ∂f
∂p ) : p ∈ R }, i.e., q is a regular point of fl . From [18, Proposition 2], there exists a neighborhood U of q, such that El ∩ U is a 2-dimensional
manifold. Together with EM ⊂ El , we have EM ∩ U ⊂ El ∩ U . It follows that
El ∩ Uq = EM ∩ Uq for some Uq ⊂ U .
By virtue of Lemma 11, when the initial positions of agents are close to EM ,
to drive the agents into EM , it suffices to constrain pl1 − pl2 to be p̃l1 − p̃l2 while
steering the agents to meet angle constraints determined by TGF . Therefore, we
wish the agents to cooperatively minimize the following cost function
V = VF + VM ,
(23)
where VF is in form (15), VM = 21 ||p̃l1 − p̃l2 − (pl1 − pl2 )||2 .
As a result, the gradient-based control law is induced as
M
ui = uF
i + ui = −∇pi VF − ∇pi VM ,
i∈V
(24)
M
where uF
i in form (16) is for driving agents to maintain the target shape, ui is
for controlling formation orientation and scale.
It is easy to see that the control law (24) is distributed and uM
= 0 for
i
i ∈ V \{l1 , l2 }. Under (24), property (i) in Lemma 10 also holds for the formation
system, while (ii) in Lemma 10 becomes invalid. Moreover, during the evolution,
the centroid is still invariant, but the scale may be changed.
Define the graph Gl = (V, El ), where El = {(pl1 , pl2 )} (we do not distinguish
(pl1 , pl2 ) and (pl2 , pl1 )). Let Hl ∈ R1×n be the incidence matrix and Ll = HlT Hl
be the Laplacian matrix, corresponding to graph Gl . Denote p̄ = p − p̃, by using
control law (24), the formation system can be written in the following compact
form
ṗ = hM (p) = −RTT F (p)δ(p) − (Ll ⊗ I2 )p̄(p).
(25)
G
The Jacobian matrix of hM at the desired equilibrium p̃ ∈ El is
JhM (p)|p=p̃ = −
∂RTT F (p)
G
∂p
δ(p)|p=p̃ − RTGF (p)
∂δ(p)
|p=p̃ − (Ll ⊗ I2 )|p=p̃
∂p
= −(RTT F (p̃)RTGF (p̃) + Ll ⊗ I2 ) , JM .
G
23
The following theorem shows the effectiveness of our control strategy.
Theorem 9. For a group of n ≥ 3 agents with dynamics (12) and controller
(24) moving in the plane. Under Assumption 1, EM is locally exponentially
stable.
Proof: By Lemma 11, we only have to show local exponential stability of El .
Note that system (25) has a similar form to [16, Equation (9)]. Moreover,
null(JM ) = null(Rl (p̃)) = span{1n ⊗ I2 }, where Rl is the matrix defined in the
proof of Lemma 11. Through a process similar to the proof of [16, Theorem 3],
it can be shown that El is locally exponentially stable. By Lemma 11, EM is
also locally exponentially stable.
Remark 6. Besides orientation and scale, we can also control translation of
the formation by pre-specifying ξ in EM to be ξ ∗ . To do this, it suffices to
add an extra control term uξ = p̃l1 + ξ ∗ − pl1 into the control law of agent
l1 . In this scenario, the desired equilibrium is a point p† = p̃ + 1n ⊗ ξ ∗ . Let
ET = {p ∈ R2n : pl1 = p̃l1 + ξ ∗ }, then ET ∩ EM = {p† }. By Lemma 11, there
is a neighborhood Up† of p† such that EM ∩ Up† = El ∩ Up† . It follows that
{p† } = ET ∩ El ∩ Up† . By linearization approach in [34, Theorem 4.13], it is
easy to prove local exponential stability of ET ∩ El . Therefore, p = p† is also
locally exponentially stable.
Most of methods for formation control in the literature do not apply to the
problem in this subsection. More specifically, by a displacement-based approach,
neither orientation nor scale can be controlled; by a distance-based approach,
the formation scale cannot be controlled; by a bearing-based approach, the
formation orientation cannot be controlled.
5. Simulations
In this section, we present two numerical examples to illustrate the effectiveness of the theoretical findings. In the first example, the target formation
shape is formed exponentially fast under the formation control law (16), while
in the second example, by implementing the control law (24), the formation is
transformed into another one with desired orientation and scale.
Example 1. Consider a group of 5 autonomous agents moving in the plane.
The target formation shape is a regular pentagon described by the framework
∗T ∗
(G, p∗ ) in Fig. 5 (a). The set of desired angle information should be {g12
g13 =
∗T ∗
∗T ∗
∗T ∗
∗T ∗
∗T ∗
g45 =
0.8090, g13 g14 = 0.8090, g14 g15 = 0.8090, g21 g23 = −0.3090, g31 g34 = 0.3090, g41
∗
0.8090}. Note that G is a triangulated Laman graph, and (G, p ) is strongly
rigid. That is, Assumption 1 holds. Without loss of generality, choose qi =
2π
T
T
T T
(cos( 2π
5 i), sin( 5 i)) , i = 1, · · · , 5. Then q = (q1 , · · · , qn ) ∈ E . Set the initial
position vector of the agents as p(0) = q + r, where r ∈ R10 is a perturbation, each component of r is a pseudorandom value drawn from the uniform
distribution on (−0.5, 0.5). By implementing the control law (16), Fig. 6 (a)
24
2
1.4
1.2
Initial positions
Initial positions
Final positions
Final positions
1.5
VF(t)
e-0.1tVF(0)
1
y-axis
1
0.8
0.5
0.6
0
0.4
-0.5
-1
-1
0.2
-0.5
0
0.5
1
1.5
2
0
2.5
x-axis
0
10
20
30
40
50
60
Time
(a)
(b)
Figure 6: (a) Under control law (16), the agents asymptotically form a regular pentagon. (b)
VF (t) vanishes to zero in an exponential speed.
is obtained, which shows that the desired formation shape can be formed by our
formation strategy. Fig. 6 (b) describes the evolution of VF (t), where VF (t) is
in form (15). It can be observed that VF (t) ≤ e−0.1t VF (0) for all t ≥ 0, implying
exponential convergence of the formation system. In conclusion, the simulation
result illustrates Theorem 8.
In fact, when we repeat the simulation by choosing other values of r in the
same way as above, it can always be obtained that VF vanishes to zero exponentially and the target formation shape is eventually formed. Moreover, when
we select each component of r from the uniform distribution on (−1, 1), the
target formation shape can still be formed in most cases. In other cases, the
angle constraints can usually be satisfied with an exponentially fast speed, i.e.,
VF vanishes to zero exponentially, whereas the target formation shape is not
eventually formed. This is because (G, p∗ ) with TGF is not globally angle rigid.
Note that the edge length of the pentagon formed by q is 1.176, therefore, the
attraction region is sizable.
Example 2. In this example, we control orientation and scale of the formation
formed in Example 1 by implementing the control input (24). Let agents 3 and 4
be the two leaders. Now we aim to drive the direction of p3 − p4 to be horizontal
with respect to the global coordinate system, while setting the length of each
edge as 0.5. It suffices to set the target displacement between two leaders as
p̃3 − p̃4 = (−0.5, 0)T . Fig. 7 shows the trajectories of agents and the evolution
of V (t) in (23), in which we can observe the validity of Theorem 9.
6. Conclusion
In this paper, we have developed an angle rigidity theory to study when
planar frameworks can be determined by angles uniquely up to translations,
rotations, scalings and reflections. We have also proved that the shape of a
25
&
&
2
0.35
Initial positions of followers
Final positions of followers
Initial positions of leaders
Final positions of leaders
1.5
0.3
0.25
y-axis
1
V(t)
e-0.03tV(0)
0.2
0.5
0.15
0
0.1
-0.5
-1
0.05
-0.5
0
0.5
1
1.5
2
2.5
0
3
x-axis
0
20
40
60
80
100
120
140
160
180
200
Time
(b)
(a)
Figure 7: (a) Under control law (24), the regular pentagon formed by all agents is asymptotically transformed into another regular pentagon with desired orientation and scale. (b) V (t)
vanishes to zero exponentially.
triangulated framework can always be uniquely determined by angles. On the
basis of the proposed angle rigidity theory, a distributed formation controller
has been designed for formation shape stabilization. We have proved that by
implementing our control strategy, a formation containing a strongly rigid triangulated framework is locally exponentially stable. Taking the advantage of
high degrees of freedom, we have proposed a distributed control strategy, which
can drive agents to stabilize a target formation shape with prescribed orientation and scale. Two simulations have been given to show effectiveness of our
formation strategies. The future work includes global stability of angle-based
formation maneuver control.
References
[1] Fax, J. A., and Murray, R. M., 2004. Information flow and cooperative control of vehicle formations. IEEE Transactions on Automatic Control, 49(9),
1465-1476.
[2] Ren, W., and Atkins, E. 2007. Distributed multi-vehicle coordinated control via local information exchange. International Journal of Robust and
Nonlinear Control, 17(10-11), 1002-1033.
[3] Xiao, F., Wang, L., Chen, J., and Gao, Y., 2009. Finite-time formation
control for multi-agent systems. Automatica, 45(11), 2605-2611.
[4] Coogan, S., and Arcak, M. 2012. Scaling the size of a formation using relative
position feedback. Automatica, 48(10), 2677-2685.
[5] Jing, G., Zheng, Y., and Wang, L. 2017. Consensus of Multiagent Systems
with Distance-Dependent Communication Networks. IEEE transactions on
neural networks and learning systems, 28(11), 2712-2726.
26
[6] Olfati-Saber, R., Fax, J. A., and Murray, R. M. 2007. Consensus and cooperation in networked multi-agent systems. Proceedings of the IEEE, 95(1),
215-233.
[7] Wang, L., and Xiao, F. 2007. A new approach to consensus problems in
discrete-time multiagent systems with time-delays. Science in China Series
F: Information Sciences, 50(4), 625-635.
[8] Wang, L., and Xiao, F. 2010. Finite-time consensus problems for networks of
dynamic agents. IEEE Transactions on Automatic Control, 55(4), 950-955.
[9] Anderson, B. D. O., Yu, C., Fidan, B., and Hendrickx, J., 2008. Rigid
graph control architectures forautonomous formations. IEEE Control Systems Magazine, 28(6), 48-63.
[10] Krick, L., Broucke, M. E., and Francis, B. A., 2009. Stabilization of infinitesimally rigid formations of multi-robot networks. International Journal
of Control 82(3), 423-439.
[11] Yu, C., Anderson, B. D. O., Dasgupta, S., and Fidan, B., 2009. Control of
minimally persistent formations in the plane. SIAM Journal on Control and
Optimization, 48(1), 206-233.
[12] Summers, T. H., Yu, C., Dasgupta, S., and Anderson, B. D. O., 2011. Control of minimally persistent leader-remote-follower and coleader formations
in the plane. IEEE Transactions on Automatic Control, 56(12), 2778-2792.
[13] Oh, K. K., and Ahn, H. S., 2011. Formation control of mobile agents based
on inter-agent distance dynamics. Automatica, 47(10), 2306-2312.
[14] Zelazo, D., Franchi, A., Bülthoff, H. H., and Robuffo Giordano, P., 2015.
Decentralized rigidity maintenance control with range measurements for
multi-robot systems. The International Journal of Robotics Research, 34(1),
105-128.
[15] Mou, S., Belabbas, M. A., Morse, A. S., Sun, Z., and Anderson, B. D. O.,
2016. Undirected rigid formations are problematic. IEEE Transactions on
Automatic Control, 61(10), 2821-2836.
[16] Sun, Z., Park, M. C., Anderson, B. D. O., and Ahn, H. S., 2017. Distributed
stabilization control of rigid formations with prescribed orientation. Automatica, 78, 250-257.
[17] Chen, X., Belabbas, M. A., and Başar, T. 2017. Global stabilization of
triangulated formations. SIAM Journal on Control and Optimization, 55(1),
172-199.
[18] Asimow, L., and Roth, B., 1978. The rigidity of graphs. Transactions of
the American Mathematical Society, 245, 279-289.
27
[19] Hendrickson, B., 1992. Conditions for unique graph realizations. SIAM
Journal on Computing 21 (1), 65-84.
[20] Liberti, L., Lavor, C., Maculan, N., and Mucherino, A., 2014. Euclidean
distance geometry and applications. Siam Review, 56(1), 3-69.
[21] Eren, T., Whiteley, W., Morse, A. S., Belhumeur, P. N., and Anderson,
B. D. O., 2003. Sensor and network topologies of formations with direction,
bearing, and angle information between agents. In Proceedings of the 42nd
IEEE Conference on Decision and Control, pp. 3064-3069.
[22] Zelazo, D., Franchi, A., and Giordano, P. R. 2004. Rigidity theory in SE(2)
for unscaled relative position estimation using only bearing measurements,
in Proceedings of European Control Conference, pp. 2703-2708.
[23] Bishop, A. N., Shames, I., and Anderson, B. D. O., 2011. Stabilization of
rigid formations with direction-only constraints. In Proceedings of the 50th
IEEE Conference on Decision and Control and European Control Conference
(CDC-ECC), pp. 746-752.
[24] Zhao, S., Zelazo, D., 2016. Bearing rigidity and almost global bearingonly formation stabilization. IEEE Transactionson Automatic Control 61
(5), 1255-1268.
[25] Zhao, S., and Zelazo, D. 2017. Translational and scaling formation maneuver control via a bearing-based approach. IEEE Transactions on Control of
Network Systems 4(3), 429-438.
[26] Wang, L., Shi, H., Chu, T., Zhang, W., and Zhang L. 2004. Aggregation of
foraging swarms. Lecture Notes in Artificial Intelligence, vol. 3339, Springer,
pp. 766-777.
[27] Oh, K. K., Park, M. C., and Ahn, H. S., 2015. A survey of multi-agent
formation control. Automatica 53, 424-440.
[28] Lin, Z., Wang, L., Chen, Z., Fu, M., and Han, Z. 2016. Necessary and sufficient graphical conditions for affine formation control. IEEE Transactions
on Automatic Control, 61(10), 2877-2891.
[29] Aranda, M., López-Nicolás, G., Sagüés, C., and Zavlanos, M. M. 2016.
Distributed formation stabilization using relative position measurements in
local coordinates. IEEE Transactions on Automatic Control, 61(12), 39253935.
[30] Zhao, S., Lin, F., Peng, K., Chen, B. M., and Lee, T. H. 2014. Distributed
control of angle-constrained cyclic formations using bearing-only measurements. Systems & Control Letters, 63, 12-24.
[31] Zhao, S., Sun, Z., Zelazo, D., Trinh, M., Ahn, H. 2017. Laman graphs
are generically bearing rigid in arbitrary dimensions. Arxiv preprint:
1703.04035.
28
[32] Dongarra, J. J., Moler, C. B., Bunch, J. R., and Stewart, G. W. 1979.
LINPACK users’ guide. Society for Industrial and Applied Mathematics.
[33] Lee, J. 2000. Introduction to smooth manifolds (Vol. 218). Springer.
[34] Khalil, H. K. 2002. Noninear Systems, 3rd edition. Prentice-Hall.
[35] Carr, J. 2012. Applications of centre manifold theory (Vol. 35). Springer
Science & Business Media.
29
| 3cs.SY
|
Type Targeted Testing
Eric L. Seidel, Niki Vazou, and Ranjit Jhala
arXiv:1410.5370v2 [cs.PL] 16 Jan 2015
UC San Diego
Abstract. We present a new technique called type targeted testing, which translates precise refinement types into comprehensive test-suites. The key insight behind our approach is that through the lens of SMT solvers, refinement types can
also be viewed as a high-level, declarative, test generation technique, wherein
types are converted to SMT queries whose models can be decoded into concrete
program inputs. Our approach enables the systematic and exhaustive testing of
implementations from high-level declarative specifications, and furthermore, provides a gradual path from testing to full verification. We have implemented our
approach as a Haskell testing tool called TARGET , and present an evaluation that
shows how TARGET can be used to test a wide variety of properties and how it
compares against state-of-the-art testing approaches.
1 Introduction
Should the programmer spend her time writing better types or thorough tests? Types
have long been the most pervasive means of describing the intended behavior of code.
However, a type signature is often a very coarse description; the actual inputs and outputs may be a subset of the values described by the types. For example, the set of ordered integer lists is a very sparse subset of the set of all integer lists. Thus, to validate
functions that produce or consume such values, the programmer must painstakingly
enumerate these values by hand or via ad-hoc generators for unit tests.
We present a new technique called type targeted testing, abbreviated to TARGET,
that enables the generation of unit tests from precise refinement types. Over the last
decade, various groups have shown how refinement types – which compose the usual
types with logical refinement predicates that characterize the subset of actual type inhabitants – can be used to specify and formally verify a wide variety of correctness
properties of programs [29,7,23,27]. Our insight is that through the lens of SMT solvers,
refinement types can be viewed as a high-level, declarative, test generation technique.
TARGET tests an implementation function against a refinement type specification
using a query-decode-check loop. First, TARGET translates the argument types into a
logical query for which we obtain a satisfying assignment (or model) from the SMT
solver. Next, TARGET decodes the SMT solver’s model to obtain concrete input values
for the function. Finally, TARGET executes the function on the inputs to get the corresponding output, which we check belongs to the specified result type. If the check fails,
the inputs are returned as a counterexample, otherwise TARGET refutes the given model
to force the SMT solver to return a different set of inputs. This process is repeated for a
given number of iterations, or until all inputs up to a certain size have been tested.
TARGET offers several benefits over other testing techniques. Refinement types provide a succinct description of the input and output requirements, eliminating the need
to enumerate individual test cases by hand or to write custom generators. Furthermore,
TARGET generates all values (up to a given size) that inhabit a type, and thus does not
skip any corner cases that a hand-written generator might miss. Finally, while the above
advantages can be recovered by a brute-force generate-and-filter approach that discards
inputs that do not meet some predicate, we show that our SMT-based method can be
significantly more efficient for enumerating valid inputs in a highly-constrained space.
TARGET paves a gradual path from testing to verification, that affords several advantages over verification. First, the programmer has an incentive to write formal specifications using refinement types. TARGET provides the immediate gratification of an
automatically generated, exhaustive suite of unit tests that can expose errors. Thus, the
programmer is rewarded without paying, up front, the extra price of annotations, hints,
strengthened inductive invariants, or tactics needed for formally verifying the specification. Second, our approach makes it possible to use refinement types to formally verify
some parts of the program, while using tests to validate other parts that may be too difficult to verify TARGET integrates the two modes by using refinement types as the uniform specification mechanism. Functions in the verified half can be formally checked
assuming the functions in the tested half adhere to their specifications. We could even
use refinements to generate dynamic contracts [9] around the tested half if so desired.
Third, even when formally verifying the type specifications, the generated tests can act
as valuable counterexamples to help debug the specification or implementation in the
event that the program is rejected by the verifier.
Finally, TARGET offers several concrete advantages over previous property-based
testing techniques, which also have the potential for gradual verification. First, instead
of specifying properties with arbitrary code [4,21] which complicates the task of subsequent formal verification, with TARGET the properties are specified via refinement
types, for which there are already several existing formal verification algorithms [27].
Second, while symbolic execution tools [12,22,28] can generate tests from arbitrary
code contracts (e.g. assertions) we find that highly constrained inputs trigger path explosion which precludes the use of such tools for gradual verification.
In the rest of this paper, we start with an overview of how TARGET can be used and
how its query-decode-check loop is implemented (§ 2). Next, we formalize a general
framework for type-targeted testing (§ 3) and show how it can be instantiated to generating tests for lists (§ 4), and then automatically generalized to other types (§ 4.6).
All the benefits of TARGET come at a price; we are limited to properties that can be
specified with refinement types. We present an empirical evaluation that shows TAR GET is efficient and expressive enough to capture a variety of sophisticated properties,
demonstrating that type-targeted testing is a sweet spot between automatic testing and
verification (§ 5).
2
2 Overview
We start with a series of examples pertaining to a small grading library called Scores.
The examples provide a bird’s eye view of how a user interacts with TARGET, how
TARGET is implemented, and the advantages of type-based testing.
Refinement Types A refinement type is one where the basic types are decorated with
logical predicates drawn from an efficiently decidable theory. For example,
type Nat
= {v:Int | 0 <= v}
type Pos
= {v:Int | 0 < v}
type Rng N = {v:Int | 0 <= v && v <
N}
are refinement types describing the set of integers that are non-negative, strictly positive,
and in the interval [0, N) respectively. We will also build up function and collection
types over base refinement types like the above. In this paper, we will not address the
issue of checking refinement type signatures [27]. We assume the code is typechecked,
e.g. by GHC, against the standard type signatures obtained by erasing the refinements.
Instead, we focus on using the refinements to synthesize tests to execute the function,
and to find counterexamples that violate the given specification.
2.1 Testing with Types
Base Types Let us write a function rescale that takes a source range [0,r1), a target
range [0,r2), and a score n from the source range, and returns the linearly scaled score
in the target range. For example, rescale 5 100 2 should return 40. Here is a first
attempt at rescale
rescale :: r1:Nat -> r2:Nat -> s:Rng r1 -> Rng r2
rescale r1 r2 s = s * (r2 ‘div‘ r1)
When we run TARGET, it immediately reports
Found counter-example: (1, 0, 0)
Indeed, rescale 1 0 0 results in 0 which is not in the target Rng 0, as the latter is
empty! We could fix this in various ways, e.g. by requiring the ranges are non-empty:
rescale :: r1:Pos -> r2:Pos -> s:Rng r1 -> Rng r2
Now, TARGET accepts the function and reports
OK. Passed all tests.
Thus, using the refinement type specification for rescale, TARGET systematically
tests the implementation by generating all valid inputs (up to a given size bound) that
respect the pre-conditions, running the function, and checking that the output satisfies
the post-condition. Testing against random, unconstrained inputs would be of limited
value as the function is not designed to work on all Int values. While in this case we
could filter invalid inputs, we shall show that TARGET can be more effective.
Containers Let us suppose we have normalized all scores to be out of 100
type Score = Rng 100
3
Next, let us write a function to compute a weighted average of a list of scores.
average
average []
average wxs
where
total
n
:: [(Int, Score)] -> Score
= 0
= total ‘div‘ n
= sum [w * x | (w, x) <- wxs ]
= sum [w
| (w, _) <- wxs ]
It can be tricky to verify this function as it requires non-linear reasoning about an unbounded collection. However, we can gain a great degree of confidence by systematically testing it using the type specification; indeed, TARGET responds:
Found counter-example: [(0,0)]
Clearly, an unfortunate choice of weights can trigger a divide-by-zero; we can fix this
by requiring the weights be non-zero:
average :: [({v:Int | v /= 0}, Score)] -> Score
but now TARGET responds with
Found counter-example: [(-3,3),(3,0)]
which also triggers the divide-by-zero! We will play it safe and require positive weights,
average :: [(Pos, Score)] -> Score
at which point TARGET reports that all tests pass.
Ordered Containers The very nature of our business requires that at the end of the
day, we order students by their scores. We can represent ordered lists by requiring the
elements of the tail t to be greater than the head h:
data OrdList a = [] | (:) {h :: a, t :: OrdList {v:a | h <= v}}
Note that erasing the refinement predicates gives us plain old Haskell lists. We can now
write a function to insert a score into an ordered list:
insert :: (Ord a) => a -> OrdList a -> OrdList a
TARGET automatically generates all ordered lists (up to a given size) and executes
insert to check for any errors. Unlike randomized testers, TARGET is not thwarted
by the ordering constraint, and does not require a custom generator from the user.
Structured Containers Everyone has a few bad days. Let us write a function that takes
the best k scores for a particular student. That is, the output must satisfy a structural
constraint – that its size equals k. We can encode the size of a list with a logical measure
function [27]:
measure len :: [a] -> Nat
len []
= 0
len (x:xs) = 1 + len xs
Now, we can stipulate that the output indeed has k scores:
best
:: k:Nat -> [Score] -> {v:[Score] | k = len v}
best k xs = take k $ reverse $ sort xs
4
Now, TARGET quickly finds a counterexample:
Found counter-example: (2,[])
Of course – we need to have at least k scores to start with!
best :: k:Nat -> {v:[Score]|k <= len v} -> {v:[Score]|k = len v}
and now, TARGET is assuaged and reports no counterexamples. While randomized testing would suffice for best, we will see more sophisticated structural properties such as
height balancedness, which stymie random testers, but are easily handled by TARGET.
Higher-order Functions Perhaps instead of taking the k best grades, we would like
to pad each individual grade, and, furthermore, we want to be able to experiment with
different padding functions. Let us rewrite average to take a functional argument, and
stipulate that it can only increase a Score.
padAverage
:: (s:Score -> {v:Score | s <= v})
-> [(Pos, Score)] -> Score
padAverage f [] = f 0
padAverage f wxs = total ‘div‘ n
where
total
= sum [w * f x | (w, x) <- wxs ]
n
= sum [w
| (w, _) <- wxs ]
TARGET automatically checks that padAverage is a safe generalization of average.
Randomized testing tools can also generate functions, but those functions are unlikely
to satisfy non-trivial constraints, thereby burdening the user with custom generators.
2.2 Synthesizing Tests
Next, let us look under the hood to get an idea of how TARGET synthesizes tests from
types. At a high-level, our strategy is to: (1) query an SMT solver for satisfying assigments to a set of logical constraints derived from the refinement type, (2) decode
the model into Haskell values that are suitable inputs, (3) execute the function on the
decoded values to obtain the output, (4) check that the output satisfies the output type,
(5) refute the model to generate a different test, and repeat the above steps until all tests
up to a certain size are executed. We focus here on steps 1, 2, and 4 – query, decode,
and check – the others are standard and require little explanation.
Base Types Recall the initial (buggy) specification
rescale :: r1:Nat -> r2:Nat -> s:Rng r1 -> Rng r2
TARGET encodes input requirements for base types directly from their corresponding
refinements. The constraints for multiple, related inputs are just the conjunction of the
constraints for each input. Hence, the constraint for rescale is:
.
C0 = 0 ≤ r1 ∧ 0 ≤ r2 ∧ 0 ≤ s < r1
In practice, C0 will also contain conjuncts of the form −N ≤ x ≤ N that restrict Intvalued variables x to be within the size bound N supplied by the user, but we will omit
these throughout the paper for clarity.
5
Note how easy it is to capture dependencies between inputs, e.g. that the score s be
in the range defined by r1. On querying the SMT solver with the above, we get a model
[r1 7→ 1, r2 7→ 1, s 7→ 0]. TARGET decodes this model and executes rescale 1 1 0
to obtain the value v = 0. Then, TARGET validates v against the post-condition by
checking the validity of the output type’s constraint:
r2 = 1 ∧ v = 0 ∧ 0 ≤ v ∧ v < r2
As the above is valid, TARGET moves on to generate another test by conjoining C0 with
a constraint that refutes the previous model:
.
C1 = C0 ∧ (r1 6= 1 ∨ r2 6= 1 ∨ s 6= 0)
This time, the SMT solver returns a model: [r1 7→ 1, r2 7→ 0, s 7→ 0] which, when
decoded and executed, yields the result 0 that does not inhabit the output type, and so is
reported as a counterexample. When we fix the specification to only allow Pos ranges,
each test produces a valid output, so TARGET reports that all tests pass.
Containers Next, we use TARGET to test the implementation of average. To do so,
TARGET needs to generate Haskell lists with the appropriate constraints. Since each list
is recursively either “nil” or “cons”, TARGET generates constraints that symbolically
represent all possible lists up to a given depth, using propositional choice variables
to symbolically pick between these two alternatives. Every (satisfying) assignment of
choices returned by the SMT solver gives TARGET the concrete data and constructors
used at each level, allowing it to decode the assignment into a Haskell value.
For example, TARGET represents valid [(Pos, Score)] inputs (of depth up to 3),
required to test average, as the conjunction of Clist and Cdata :
.
Clist = (c00 ⇒ xs 0 = []) ∧ (c01 ⇒ xs 0 = x1 : xs 1 ) ∧ (c00 ⊕ c01 )
∧ (c10 ⇒ xs 1 = []) ∧ (c11 ⇒ xs 1 = x2 : xs 2 ) ∧ (c01 ⇒ c10 ⊕ c11 )
∧ (c20 ⇒ xs 2 = []) ∧ (c21 ⇒ xs 2 = x3 : xs 3 ) ∧ (c11 ⇒ c20 ⊕ c21 )
∧ (c30 ⇒ xs 3 = []) ∧ (c21 ⇒ c30 )
.
Cdata = (c01 ⇒ x1 = (w1 , s1 ) ∧ 0 < w1 ∧ 0 ≤ s1 < 100)
∧ (c11 ⇒ x2 = (w2 , s2 ) ∧ 0 < w2 ∧ 0 ≤ s2 < 100)
∧ (c21 ⇒ x3 = (w3 , s3 ) ∧ 0 < w3 ∧ 0 ≤ s3 < 100)
The first set of constraints Clist describes all lists up to size 3. At each level i, the choice
variables ci0 and ci1 determine whether at that level the constructed list xs i is a “nil” or
a “cons”. In the constraints [] and ( : ) are uninterpreted functions that represent “nil”
and “cons” respectively. These functions only obey the congruence axiom and hence,
can be efficiently analyzed by SMT solvers [19]. The data at each level xi is constrained
to be a pair of a positive weight wi and a valid score si .
The choice variables at each level are used to guard the constraints on the next
levels. First, if we are generating a “cons” at a given level, then exactly one of the
choice variables for the next level must be selected; e.g. c11 ⇒ c20 ⊕ c21 . Second, the
constraints on the data at a given level only hold if we are generating values for that
6
level; e.g. c21 is used to guard the constraints on x3 , w3 and s3 . This is essential to avoid
over-constraining the system which would cause TARGET to miss certain tests.
To decode a model of the above into a Haskell value of type [(Int, Int)], we
traverse constraints and use the valuations of the choice variables to build up the list
appropriately. At each level, if ci0 7→ true, then the list at that level is [], otherwise
ci1 7→ true and we decode xi+1 and xs i+1 and “cons” the results.
We can iteratively generate multiple inputs by adding a constraint that refutes each
prior model. As an important optimization, we only refute the relevant parts of the
model, i.e. those needed to construct the list (§ 4.5).
Ordered Containers Next, let us see how TARGET enables automatic testing with
highly constrained inputs, such as the increasingly ordered OrdList values required
by insert. From the type definition, it is apparent that ordered lists are the same as the
usual lists described by Clist , except that each unfolded tail must only contain values that
are greater than the corresponding head. That is, as we unfold x1:x2:xs :: OrdList
– At level 0, we have OrdList {v:Score| true}
– At level 1, we have OrdList {v:Score| x1 <= v}
– At level 2, we have OrdList {v:Score| x2 <= v && x1 <= v}
and so on. Thus, we encode OrdList Score (of depth up to 3) by conjoining Clist with
Cscore and Cord , which capture the valid score and ordering requirements respectively:
.
Cord = (c11 ⇒ x1 ≤ x2 ) ∧ (c21 ⇒ x2 ≤ x3 ∧ x1 ≤ x3 )
.
Cscore = (c01 ⇒ 0 ≤ x1 < 100) ∧ (c11 ⇒ 0 ≤ x2 < 100) ∧ (c21 ⇒ 0 ≤ x3 < 100)
Structured Containers Recall that best k requires inputs whose structure is constrained – the size of the list should be no less than k. We specify size using special
measure functions [27], which let us relate the size of a list with that of its unfolding,
and hence, let us encode the notion of size inside the constraints:
.
Csize = (c00 ⇒ len xs 0 = 0) ∧ (c01 ⇒ len xs 0 = 1 + len xs 1 )
∧ (c10 ⇒ len xs 1 = 0) ∧ (c11 ⇒ len xs 1 = 1 + len xs 2 )
∧ (c20 ⇒ len xs 2 = 0) ∧ (c21 ⇒ len xs 2 = 1 + len xs 3 )
∧ (c30 ⇒ len xs 3 = 0)
At each unfolding, we instantiate the definition of the measure for each alternative of the
datatype. In the constraints, len · is an uninterpreted function derived from the measure
definition. All of the relevant properties of the function are spelled out by the unfolded
constraints in Csize and hence, we can use SMT to search for models for the above
constraint. Hence, TARGET constrains the input type for best as:
0 ≤ k ∧ Clist ∧ Cscore ∧ Csize ∧ k ≤ len xs 0
where the final conjunct comes from the top-level refinement that stipulates the input
have at least k scores. Thus, TARGET only generates lists that are large enough. For
example, in any model where k = 2, it will not generate the empty or singleton list, as
in those cases, len xs 0 would be 0 (resp. 1), violating the final conjunct above.
7
-- Manipulating Refinements
refinement :: RefType -> Refinement
subst
:: RefType -> [(Var, Var)] -> RefType
-- Manipulating Types
unfold
:: Ctor -> RefType -> [(Var, RefType)]
binder
:: RefType -> Var
proxy
:: RefType -> Proxy a
Fig. 1. Refinement Type API
Higher-order Functions Finally, TARGET’s type-directed testing scales up to higherorder functions using the same insight as in QuickCheck [4], namely, to generate a
function it suffices to be able to generate the output of the function. When tasked with
the generation of a functional argument f, TARGET returns a Haskell function that
when executed checks whether its inputs satisfy f’s pre-conditions. If they do, then
f uses TARGET to dynamically query the SMT solver for an output that satisfies the
constraints imposed by the concrete inputs. Otherwise, f’s specifications are violated
and TARGET reports a counterexample.
This concludes our high-level tour of the benefits and implementation of TARGET.
Notice that the property specification mechanism – refinement types – allowed us to
get immediate feedback that helped debug not just the code, but also the specification
itself. Additionally, the specifications gave us machine-readable documentation about
the behavior of functions, and a large unit test suite with which to automatically validate
the implementation. Finally, though we do not focus on it here, the specifications are
amenable to formal verification should the programmer so desire.
3 A Framework for Type Targeted Testing
Next, we describe a framework for type targeted testing, by formalizing an abstract
representation of refinement types (§ 3.1), describing the operations needed to generate
tests from types (§ 3.2), and then using the above to implement TARGET via a querydecode-check loop (§ 3.3). Subsequently, we instantiate the framework to obtain tests
for refined primitive types, lists, algebraic datatypes and higher-order functions (§ 4).
3.1 Refinement Types
A refinement type is a type, where each component is decorated with a predicate from
a refinement logic. For clarity, we describe refinement types and refinements abstractly
as RefType and Refinement respectively. We write Var as an alias for Refinement
that is typically used to represent logical variables appearing within the refinement.
Notation In the sequel, we will use double brackets [[]] to represent the various entities
in the meta-language used to describe TARGET. For example, [[k]], [[k ≤ len v]], and
8
[[{v : [Score] | k ≤ len v}]] are the Var, Refinement, and RefType representing the
corresponding entities written in the brackets.
Next, we describe the various operations over them needed to implement TARGET.
These operations, summarized in Figure 1, fall into two categories: those which manipulate the refinements and those which manipulate the types.
Operating on Refinements To generate constraints and check inhabitation, we use the
function refinement which returns the (top-level) refinement that decorates the given
refinement type. We will generate fresh Vars to name values of components, and will
use subst to replace the free occurrences of variables in a given RefType. Suppose
that t is the RefType represented by [[{v : [Score] | k ≤ len v}]]. Then,
– refinement t evaluates to [[k ≤ len v]] and
– subst t [([[k]], [[x0 ]])] evaluates to [[{v : [Score] | x0 ≤ len v}]].
Operating on Types To build up compound values (e.g. lists) from components (e.g. an
integer and a list), unfold breaks a RefType (e.g. a list of integers) into its constituents
(e.g. an integer and a list of integers) at a given constructor (e.g. “cons”). binder simply extracts the Var representing the value being refined from the RefType. To write
generic functions over RefTypes and use Haskell’s type class machinery to query and
decode components of types, we associate with each refinement type a proxy representing the corresponding Haskell type (in practice this must be passed around as a
separate argument). For example, if t is [[{v : [Score] | k ≤ len v}]],
– unfold [[:]] t evaluates to [([[x]], [[Score]]), ([[xs]], [[[Score]]])],
– binder t evaluates to [[v]], and
– proxy t evaluates to a value of type Proxy [Int].
3.2 The Targetable Type Class
Following QuickCheck, we encapsulate the key operations needed for type-targeted
testing in a type class Targetable (Figure 2). This class characterizes the set of types
class Targetable a where
query :: Proxy a -> Int -> RefType -> SMT Var
decode :: Var -> SMT a
check :: a -> RefType -> SMT (Bool, Var)
toReft :: a -> Refinement
Fig. 2. The class of types that can be tested by TARGET
which can be tested by TARGET. All of the operations can interact with an external
SMT solver, and so return values in an SMT monad.
– query takes a proxy for the Haskell type for which we are generating values, an
integer depth bound, and a refinement type describing the desired constraints, and
generates a set of logical constraints and a Var that represents the constrained value.
9
– decode takes a Var, generated via a previous query and queries the model returned by the SMT solver to construct a Haskell value of type a.
– check takes a value of type a, translates it back into logical form, and verifies that
it inhabits the output type t.
– toReft takes a value of type a and translates it back into logical form (a specialization of check).
3.3 The Query-Decode-Check Loop
Figure 3 summarizes the overall implementation of TARGET, which takes as input a
function f and its refinement type specification t and proceeds to test the function
against the specification via a query-decode-check loop: (1) First, we translate the refined inputTypes into a logical query. (2) Next, we decode the model (i.e. satisfying
assignment) for the query returned by the SMT solver to obtain concrete inputs. (3)
Finally, we execute the function f on the inputs to get the corresponding output,
which we check belongs to the specified outputType. If the check fails, we return
the inputs as a counterexample. After each test, TARGET, refutes the given test to
force the SMT solver to return a different set of inputs, and this process is repeated
until a user specified number of iterations. The checkSMT call may fail to find a model
meaning that we have exhaustively tested all inputs upto a given testDepth bound.
If all iterations succeed, i.e. no counterexamples are found, then TARGET returns Ok,
indicating that f satisfies t up to the given depth bound.
target f t = do
let txs = inputTypes t
vars <- forM txs $ \tx ->
query (proxy tx) testDepth tx -- Query
forM [1 .. testNum] $ \_ -> do
hasModel <- checkSMT
when hasModel $ do
inputs <- forM vars decode
-- Decode
output <- execute f inputs
let su = zip (map binder txs) (map toReft inputs)
let to = outputType t ‘subst‘ su
(ok,_) <- check output to
-- Check
if ok then
refuteSMT
else
throw (CounterExample inputs)
return Ok
Fig. 3. Implementing TARGET via a query-decode-check loop
10
4 Instantiating the TARGET Framework
Next, we describe a concrete instantiation of TARGET for lists. We start with a constraint generation API (§ 4.1). Then we use the API to implement the key operations query (§ 4.2), decode (§ 4.3), check (§ 4.4), and refuteSMT (§ 4.5), thereby
enabling TARGET to automatically test functions over lists. We omit the definition of
toReft as it follows directly from the definition of check. Finally, we show how the
list instance can be generalized to algebraic datatypes and higher-order functions (§ 4.6).
4.1 SMT Solver Interface
Figure 4 describes the interface to the SMT solvers that TARGET uses for constraint generation and model decoding. The interface has functions to (a) generate logical variables
of type Var, (b) constrain their values using Refinement predicates, and (c) determine
the values assigned to the variables in satisfying models.
fresh
:: SMT Var
guard
:: Var -> SMT a
-> SMT a
constrain :: Var -> Refinement -> SMT ()
apply
unapply
:: Ctor -> [Var] -> SMT Var
:: Var -> SMT (Ctor, [Var])
oneOf
whichOf
:: Var -> [(Var, Var)] -> SMT ()
:: Var -> SMT Var
eval
:: Refinement -> SMT Bool
Fig. 4. SMT Solver API
– fresh allocates a new logical variable.
– guard b act ensures that all the constraints generated by act are guarded by the
choice variable b. That is, if act generates the constraint p then guard b act
generates the (implication) constraint b ⇒ p.
– constrain x r generates a constraint that x satisfies the refinement predicate r.
– apply c xs generates a new Var for the folded up value obtained by applying
the constructor c to the fields xs, while also generating constraints from the measures. For example, apply [[:]] [[[x1 ]], [[xs1 ]]] returns [[x1 : xs 1 ]] and generates the
constraint len (x1 : xs 1 ) = 1 + len xs 1 .
– unapply x returns the Ctor and Vars from which the input x was constructed.
– oneOf x cxs generates a constraint that x equals exactly one of the elements of
cxs. For example, oneOf [[xs0 ]] [([[c00 ]],[[[]]]),([[c01 ]],[[x1 : xs1 ]])] yields:
(c00 ⇒ xs 0 = []) ∧ (c01 ⇒ xs 0 = x1 : xs 1 ) ∧ (c00 ⊕ c01 )
11
query p d t = do
let cs = ctors d
bs <- forM cs (\_ -> fresh)
xs <- zipWithM (queryCtor (d-1) t) bs cs
x <- fresh
oneOf x
(zip bs xs)
constrain x (refinement t)
return x
queryCtor d
let fts =
fs’
<x
<return x)
t b c = guard b (do
unfold c t
scanM (queryField d) [] fts
apply c fs’
queryField d su (f, t) = do
f’ <- query (proxy t) d (t ‘subst‘ su)
return ((f, f’) : su, f’)
ctors d
| d > 0
= [ [[:]], [[[]]] ]
| otherwise = [ [[[]]] ]
Fig. 5. Generating a Query
– whichOf x returns the particular alternative that was assigned to x in the current
model returned by the SMT solver. Continuing the previous example, if the model
sets [[c00 ]] (resp. [[c01 ]]) to true, whichOf [[xs0 ]] returns [[[]]] (resp [[x1 : xs1 ]]).
– eval r checks the validity of a refinement with no free variables. For example,
eval [[len (1 : []) > 0]] would return True.
4.2 Query
Figure 5 shows the procedure for constructing a query from a refined list type, e.g. the
one required as an input to the best or insert functions from § 2.
Lists query returns a Var that represent all lists up to depth d that satisfy the logical
constraints associated with the refined list type t. To this end, it invokes ctors to obtain
all of the suitable constructors for depth d. For lists, when the depth is 0 we should only
use the [[[]]] constructor, otherwise we can use either [[:]] or [[[]]] . This ensures that query
terminates after encoding all possible lists up to a given depth d. Next, it uses fresh
to generate a distinct choice variable for each constructor, and calls queryCtor to
generate constraints and a corresponding symbolic Var for each constructor. The choice
variable for each constructor is supplied to queryCtor to ensure that the constraints
are guarded, i.e. only required to hold if the corresponding choice variable is selected
in the model and not otherwise. Finally, a fresh x represents the value at depth d and is
constrained to be oneOf the alternatives represented by the constructors, and to satisfy
the top-level refinement of t.
12
decode x = do
x’
<- whichOf x
(c,fs’) <- unapply x’
decodeCtor c fs’
decodeCtor [[[]]] [] = return []
decodeCtor [[:]] [x,xs] = do
v <- decode x
vs <- decode xs
return (v:vs)
Fig. 6. Decoding Models into Haskell Values
Constructors queryCtor takes as input the refined list type t, a depth d, a particular
constructor c for the list type, and generates a query describing the unfolding of t at the
constructor c, guarded by the choice variable b that determines whether this alternative
is indeed part of the value. These constraints are the conjunction of those describing the
values of the individual fields which can be combined via c to obtain a t value. To do
so, queryCtor first unfolds the type t at c, obtaining a list of constituent fields and
their respective refinement types fts. Next, it uses
scanM :: Monad m => (a -> b -> m (a, c)) -> a -> [b] -> m [c]
to traverse the fields from left to right, building up representations of values for the fields
from their unfolded refinement types. Finally, we invoke apply on c and the fields fs’
to return a symbolic representation of the constructed value that is constrained to satisfy
the measure properties of c.
Fields queryField generates the actual constraints for a single field f with refinement
type t, by invoking query on t. The proxy enables us to resolve the appropriate typeclass instance for generating the query for the field’s value. Each field is described by
a new symbolic name f’ which is substituted for the formal name of the field f in
the refinements of subsequent fields, thereby tracking dependencies between the fields.
For example, these substitutions ensure the values in the tail are greater than the head
as needed by OrdList from § 2.
4.3 Decode
Once we have generated the constraints we query the SMT solver for a model, and if
one is found we must decode it into a concrete Haskell value with which to test the
given function. Figure 6 shows how to decode an SMT model for lists.
Lists decode takes as input the top-level symbolic representation x and queries the
model to determine which alternative was assigned by the solver to x, i.e. a nil or a cons.
Once the alternative is determined, we use unapply to destruct it into its constructor c
and fields fs’, which are recursively decoded by decodeCtor.
Constructors decodeCtor takes the constructor c and a list of symbolic representations for fields, and decodes each field into a value and applies the constructor to obtain
the Haskell value. For example, in the case of the [[[]]] constructor, there are no fields, so
we return the empty list. In the case of the [[:]] constructor, we decode the head and the
tail, and cons them to return the decoded value. decodeCtor has the type
Targetable a => Ctor -> [Var] -> SMT [a]
13
check v t = do
let (c,vs) = splitCtor v
let fts
= unfold c t
(bs, vs’) <- fmap unzip (scanM checkField [] (zip vs fts))
v’
<- apply c vs’
let t’
= t ‘subst‘ [(binder t, v’)]
b’
<- eval (refinement t’)
return (and (b:bs), v’)
checkField su (v, (f, t)) = do
(b, v’) <- check v (t ‘subst‘ su)
return ((f, v’) : su, (b, v’))
splitCtor []
= ([[[]]], [])
splitCtor (x:xs) = ([[:]], [x,xs])
Fig. 7. Checking Outputs
i.e. if a is a decodable type, then decodeCtor suffices to decode lists of a. Primitives
like integers that are directly encoded in the refinement logic are the base case – i.e. the
value in the model is directly translated into the corresponding Haskell value.
4.4 Check
The third step of the query-decode-check loop is to verify that the output produced by
the function under test indeed satisfies the output refinement type of the function. We
accomplish this by encoding the output value as a logical expression, and evaluating the
output refinement applied to the logical representation of the output value.
check, shown in Figure 7, takes a Haskell (output) value v and the (output) refinement type t, and recursively verifies each component of the output type. It converts
each component into a logical representation, substitutes the logical expression for
the symbolic value, and evaluates the resulting Refinement.
4.5 Refuting Models
Finally, TARGET invokes refuteSMT to refute a given model in order to force the
SMT solver to produce a different model that will yield a different test input. A naı̈ve
implementation of refutation is as follows. Let X be the set of all variables appearing in
the constraints. Suppose that in the current model, each variable x is assigned the value
σ(x). Then, to refute the model, we add a refutation constraint ∨x∈X x 6= σ(x). That
is, we stipulate that some variable be assigned a different value.
The naı̈ve implementation is extremely inefficient. The SMT solver is free to pick
a different value for some irrelevant variable which was not even used for decoding.
As a result, the next model can, after decoding, yield the same Haskell value, thereby
blowing up the number of iterations needed to generate all tests of a given size.
14
TARGET solves this problem by forcing the SMT solver to return models that yield
different decoded tests in each iteration. To this end TARGET restricts the refutation
constraint to the set of variables that were actually used to decode the Haskell value.
We track this set by instrumenting the SMT monad to log the set of variables and choicevariables that are transitively queried via the recursive calls to decode. That is, each
call to decode logs its argument, and each call to whichOf logs the choice variable
corresponding to the alternative that was returned. Let R be the resulting set of decoderelevant variables. TARGET refutes the model by using a relevant refutation constraint
∨x∈R x 6= σ(x) which ensures that the next model decodes to a different value.
4.6 Generalizing TARGET To Other Types
The implementation in § 4 is for List types, but ctors, decodeCtor, and splitCtor
are the only functions that are List-specific. Thus, we can easily generalize the implementation to:
– primitive datatypes, e.g. integers, by returning an empty list of constructors,
– algebraic datatypes, by implementing ctors, decodeCtor, and splitCtor for
that type.
– higher-order functions, by lifting instances of a to functions returning a.
Algebraic Datatypes Our List implementation has three pieces of type-specific logic:
– ctors, which returns a list of constructors to unfold;
– decodeCtor, which decodes a specific Ctor; and
– splitCtor, which splits a Haskell value into a pair of its Ctor and fields.
Thus, to instantiate TARGET on a new data type, all we need is to implement these
three operations for the type. This implementation essentially follows the concrete template for Lists. In fact, we observe that the recipe is entirely mechanical boilerplate, and
can be fully automated for all algebraic data types by using a generics library.
Any algebraic datatype (ADT) can be represented as a sum-of-products of component types. A generics library, such as GHC.Generics [15], provides a univeral sum-ofproducts type and functions to automatically convert any ADT to and from the universal
representation. Thus, to obtain Targetable instances for any ADT it suffices to define
a Targetable instance for the universal type.
Once the universal type is Targetable we can automatically get an instance for
any new user-defined ADT (that is an instance of Generic) as follows: (1) to generate
a query we simply create a query for GHC.Generics’ universal representation of the
refined type, (2) to decode the results from the SMT solver, we decode them into the
universal representation and then use GHC.Generics to map them back into the userdefined type, (3) to check that a given value inhabits a user-defined refinement type,
we check that the universal representation of the value inhabits the type’s universal
counterpart.
The Targetable instance for the universal representation is a generalized version
of the List instance from § 4, that relies on various technical details of GHC.Generics.
15
Higher Order Functions Our type-directed approach to specification makes it easy to
extend TARGET to higher-order functions. Concretely, it suffices to implement a typeclass instance:
instance (Targetable input, Targetable output)
=> Targetable (input -> output)
In essence, this instance uses the Targetable instances for input and output to
create an instance for functions from input -> output, after which Haskell’s type
class machinery suffices to generate concrete function values.
To create such instances, we use the insight from QuickCheck, that to generate (constrained) functions, we need only to generate output values for the function. Following
this route, we generate functions by creating new lambdas that take in the inputs from
the calling context, and use their values to create queries for the output, after which we
can call the SMT solver and decode the results to get concrete outputs that are returned
by the lambda, completing the function definition. Note that we require input to also
be Targetable so that we can encode the Haskell value in the refinement logic, in
order to constrain the output values suitably. We additionally memoize the generated
function to preserve the illusion of purity. It is also possible to, in the future, extend our
implementation to refute functions by asserting that the output value for a given input
be distinct from any previous outputs for that input.
5 Evaluation
We have built a prototype implementation of TARGET1 and next, describe an evaluation on a series of benchmarks ranging from textbook examples of algorithms and data
structures to widely used Haskell libraries like CONTAINERS and XMONAD. Our goal
in this evaluation is two-fold. First, we describe micro-benchmarks (i.e. functions) that
quantitatively compare TARGET with the existing state-of-the-art, property-based testing tools for Haskell – namely SmallCheck and QuickCheck – to determine whether
TARGET is indeed able to generate highly constrained inputs more effectively. Second,
we describe macro-benchmarks (i.e. modules) that evaluate the amount of code coverage that we get from type-targeted testing.
5.1 Comparison with QuickCheck and SmallCheck
We compare TARGET with QuickCheck and SmallCheck by using a set of benchmarks
with highly constrained inputs. For each benchmark we compared TARGET with SmallCheck and QuickCheck, with the latter two using the generate-and-filter approach,
wherein a value is generated and subsequently discarded if it does not meet the desired constraint. While one could possibly write custom “operational” generators for
each property, the point of this evaluation is compare the different approaches ability
to enable “declarative” specification driven testing. Next, we describe the benchmarks
and then summarize the results of the comparison (Figure 8).
1
http://hackage.haskell.org/package/target-0.1.1.0
16
L IST. INSERT
RBT REE . ADD
XM ONAD . FOCUS
LEFT
Time (sec)
104
100
10−4
5
10
15
20
5
10
2
4
6
8
Depth
M AP. DELETE
M AP. DIFFERENCE
4
Time (sec)
10
100
10−4
TARGET
2
4
SmallCheck
6
8
10
2
Depth
4
Lazy SmallCheck
6
8
Lazy SmallCheck (slow)
Fig. 8. Results of comparing TARGET with QuickCheck, SmallCheck, and Lazy SmallCheck on a
series of functions. TARGET , SmallCheck, and Lazy SmallCheck were both configured to check
the first 1000 inputs that satisfied the precondition at increasing depth parameters, with a 60
minute timeout per depth; QuickCheck was run with the default settings, i.e. it had to produce
100 test cases. TARGET , SmallCheck, and Lazy SmallCheck were configured to use the same
notion of depth, in order to ensure they would generate the same number of valid inputs at each
depth level. QuickCheck was unable to successfully complete any run due to the low probability
of generating valid inputs at random.
Inserting into a sorted List Our first benchmark is the insert function from the
homonymous sorting routine. We use the specification that given an element and a
sorted list, insert x xs should evaluate to a sorted list. We express this with the
type
type Sorted a = List <{\hd v -> hd < v}> a
insert :: a -> Sorted a -> Sorted a
where the ordering constraint is captured by an abstract refinement [25] which states
that each list head hd is less than every element v in its tail.
Inserting into a Red-Black Tree Next, we consider insertion into a Red-Black tree.
data RBT a = Leaf | Node Col a (RBT a) (RBT a)
data Col
= Black | Red
Red-black trees must satisfy three invariants: (1) red nodes always have black children,
(2) the black height of all paths from the root to a leaf is the same, and (3) the elements
17
in the tree should be ordered. We capture (1) via a measure that recursively checks each
Red node has Black children.
measure isRB :: RBT a -> Prop
isRB Leaf
= true
isRB (Node c x l r) = isRB l && isRB r &&
(c == Red => isBlack l && isBlack r)
We specify (2) by defining the Black height as:
measure bh :: RBT a -> Int
bh Leaf
= 0
bh (Node c x l r) = bh l + (if c == Red then 0 else 1)
and then checking that the Black height of both subtrees is the same:
measure isBH :: RBT a -> Prop
isBH Leaf
= true
isBH (Node c x l r) = isBH l && isBH r && bh l == bh r
Finally, we specify the (3), the ordering invariant as:
type OrdRBT a = RBT <{\r v -> v < r}, {\r v -> r < v}> a
i.e. with two abstract refinements for the left and right subtrees respectively, which state
that the root r is greater than (resp. less than) each element v in the subtrees. Finally, a
valid Red-Black tree is:
type OkRBT a = {v:OrdRBT a | isRB v && isBH v}
Note that while the specification for the internal invariants for Red-Black trees is tricky,
the specification for the public API – e.g. the add function – is straightforward:
add :: a -> OkRBT a -> OkRBT a
Deleting from a Data.Map Our third benchmark is the delete function from the
Data.Map module in the Haskell standard libraries. The Map structure is a balanced
binary search tree that implements purely functional key-value dictionaries:
data Map k a = Tip | Bin Int k a (Map k a) (Map k a)
A valid Data.Map must satisfy two properties: (1) the size of the left and right subtrees must be within a factor of three of each other, and (2) the keys must obey a binary
search ordering. We specify the balancedness invariant (1) with a measure
measure isBal :: Map k a -> Prop
isBal (Tip)
= true
isBal (Bin s k v l r) = isBal l && isBal r &&
(sz l + sz r <= 1 ||
sz l <= 3 * sz r <= 3 * sz l)
and combine it with an ordering invariant (like OrdRBT) to specify valid trees.
type OkMap k a = {v : OrdMap k a | isBal v}
18
We can check that delete preserves the invariants by checking that its output is an
OkMap k a. However, we can also go one step further and check the functional correctness property that delete removes the given key, with a type:
delete :: Ord k => k:k -> m:OkMap k a
-> {v:OkMap k a | MinusKey v m k}
where the predicate MinusKey is defined as:
predicate MinusKey M1 M2 K
= keys M1 = difference (keys M2) (singleton K)
using the measure keys describing the contents of the Map:
measure keys :: Map k a -> Set k
keys (Tip)
= empty ()
keys (Bin s k v l r) = union (singleton k)
(union (keys l) (keys r))
Refocusing XMonad StackSets Our last benchmark comes from the tiling window
manager XMonad. The key invariant of XMonad’s internal StackSet data structure
is that the elements (windows) must all be unique, i.e. contain no duplicates. XMonad
comes with a test-suite of over 100 QuickCheck properties; we select one which states
that moving the focus between windows in a StackSet should not affect the order of
the windows.
prop_focus_left_master n s =
index (foldr (const focusUp) s [1..n]) == index s
With QuickCheck, the user writes a custom generator for valid StackSets and then
runs the above function on test inputs created by the generator, to check if in each case,
the result of the above is True.
With TARGET, it is possible to test such properties without requiring custom generators. Instead the user writes a declarative specification:
type OkStackSet = {v:StackSet | NoDuplicates v}
(We refer the reader to [26] for a full discussion of how to specify NoDuplicates).
Next, we define a refinement type:
type TTrue = {v:Bool | Prop v}
that is only inhabited by True, and use it to type the QuickCheck property as:
prop_focus_left_master :: Nat -> OkStackSet -> TTrue
This property is particularly difficult to verify; however, TARGET is able to automatically generate valid inputs to test that prop_focus_left_master always returns
True.
Results Figure 8 summarizes the results of the comparison. QuickCheck was unable
to successfully complete any benchmark to the low probability of generating properly
constrained values at random.
List Insert TARGET is able to test insert all the way to depth 20, whereas Lazy
SmallCheck times out at depth 19.
19
Red-Black Tree Insert TARGET is able to test add up to depth 12, while Lazy SmallCheck times out at depth 6.
Map Delete TARGET is able to check delete up to depth 10, whereas Lazy SmallCheck times out at depth 7 if it checks ordering first, or depth 6 if it checks balancedness first.
StackSet Refocus TARGET and is able to check this property up to depth 8, while Lazy
SmallCheck times out at depth 7.
TARGET sees a performance hit with properties that require reasoning with the theory of Sets e.g. the no-duplicates invariant of StackSet. While Lazy SmallCheck times
out at a higher depths, when it completes e.g. at depth 6, it does so in 0.7s versus TAR GET ’s 9 minutes. We suspect this is because the theory of sets are a relatively recent
addition to SMT solvers [18], and with further improvements in SMT technology, these
numbers will get significantly better.
Overall, we found that for small inputs Lazy SmallCheck is substantially faster as
exhaustive enumeration is tractable, and does not incur the overhead of communicating
with an external general-purpose solver. Additionally, Lazy SmallCheck benefits from
pruning predicates that exploit laziness and only force a small portion of the structure
(e.g. ordering). However, we found that constraints that force the entire structure (e.g.
balancedness), or composing predicates in the wrong order, can force Lazy SmallCheck
to enumerate the entire exponentially growing search space.
TARGET, on the other hand, scales nicely to larger input sizes, allowing systematic
and exhaustive testing of larger, more complex inputs. This is because TARGET eschews
explicit enumeration-and-filtering (which results in searching for fewer needles in larger
haystacks as the sizes increas), in favor of symbolically searching for valid models via
SMT, making TARGET robust to the strictness or ordering of constraints.
5.2 Measuring Code Coverage
The second question we seek to answer is whether TARGET is suitable for testing entire libraries, i.e. how much of the program can be automatically exercised using our
system? Keeping in mind the well-known issues with treating code coverage as an indication of test-suite quality [16], we consider this experiment a negative filter.
To this end, we ran TARGET against the entire user-facing API of Data.Map, our
RBTree library, and XMonad.StackSet – using the constrained refined types (e.g.
OkMap, OkRBT, OkStackSet) as the specification for the exposed types – and measured
the expression and branch coverage, as reported by hpc [11]. We used an increasing
timeout ranging from one to thirty minutes per exported function.
Results The results of our experiments are shown in Figure 9. Across all three libraries,
TARGET achieved at least 70% expression and 64% alternative coverage at the shortest
timeout of one minute per function. Interestingly, the coverage metrics for RBTree and
Data.Map remain relatively constant as we increase the timeouts, with a small jump
in expression coverage between 10 and 20 minutes. XMonad on the other hand, jumps
from 70% expression and 64% alternative coverage with a one minute timeout, to 96%
expression and 94% alternative with a ten minute timeout.
20
DATA .M AP
XM ONAD .S TACK S ET
RBT REE
% Coverage
1
0.8
0.6
0.4
0.2
0
0
10
20
30
expressions
always-false
0
10
20
30
Timeout (min)
booleans
alternatives
0
10
20
30
always-true
local-functions
Fig. 9. Coverage-testing of Data.Map.Base, RBTree, and XMonad.StackSet using TAR GET . Each exported function was tested with increasing depth limits until a single run hit a timeout ranging from one to thirty minutes. Lower is better for “always-true” and “always-false”,
higher is better for everything else.
There are three things to consider when examining these results. First is that some
expressions are not evaluated due to Haskell’s laziness (e.g. the values contained in
a Map). Second is that some expressions should not be evaluated and some branches
should not be taken, as these only happen when an unexpected error condition is triggered (i.e. these expressions should be dead code). TARGET considers any inputs that
trigger an uncaught exception a valid counterexample; the pre-conditions should rule
out these inputs, and so we expect not to cover those expressions with TARGET.
The last remark is not intrinsically related to TARGET, but rather our means of
collecting the coverage data. hpc includes otherwise guards in the “always-true”
category, even though they cannot evaluate to anything else. Data.Map contained 56
guards, of which 24 were marked “always-true”. We manually counted 21 otherwise
guards, the remaining 3 “always-true” guards compared the size of subtrees when rebalancing to determine whether a single or double rotation was needed; we were unable
to trigger the double rotation in these cases. XMonad contained 9 guards, of which 4
were “always-true”. 3 of these were otherwise guards; the remaining “always-true”
guard dynamically checked a function’s pre-condition. If the pre-condition check had
failed an error would have been thrown by the next case, we consider it a success of
TARGET that the error branch was not triggered.
5.3 Discussion
To sum up, our experiments demonstrate that TARGET generates valid inputs: (1) where
QuickCheck fails outright, due to the low probability of generating random values satisfying a property; (2) more efficiently than Lazy SmallCheck, which relies on lazy
pruning predicates; and (3) providing high code coverage for real-world libraries with
no hand-written test cases.
Of course our approach is not without drawbacks; we highlight five classes of pitfalls the user may encounter.
21
Laziness in the function or in the output refinement can cause exceptions to go unthrown if the output value is not fully demanded. For example, TARGET would decide
that the result [1, undefined] inhabits [Int] but not [Score], as the latter would
have to evaluate 0 <= undefined < 100. This limitation is not specific to our system, rather it is fundamental to any tool that exercises lazy programs. Furthermore,
TARGET only generates inductively-defined values, it cannot generate infinite or cyclic
structures, nor will the generated values ever contain ⊥.
Polymorphism Like any other tool that actually runs the function under scrutiny, TAR GET can only test monomorphic instantiations of polymorphic functions. For example,
when testing XMonad we instantiated the “window” parameter to Char and all other
type parameters to (), as the properties we were testing only examined the window.
This helped drastically reduce the search space, both for TARGET and SmallCheck.
Advanced type-system features such as GADTs and Existential types may prevent GHC
from deriving a Generic instance, which would force the programmer to write her
own Targetable instance. Though tedious, the single hand-written instance allows
TARGET to automatically generate values satisfying disparate constraints, which is still
an improvement over the generate-and-filter approach.
Refinement types are less expressive than properties written in the host language. If
the pre-conditions are not expressible in TARGET’s logic, the user will have to use the
generate-and-filter approach, losing the benefits of symbolic enumeration.
Input explosion TARGET excels when the space of valid inputs is a sparse subset of
the space of all inputs. If the input space is not sufficiently constrained, TARGET may
spend lose its competitive advantage over other tools due to the overhead of using a
general-purpose solver.
6 Related Work
TARGET is closely related to a number of lines of work on connecting formal specifications, execution, and automated constraint-based testing. Next, we describe the closest
lines of work on test-generation and situate them with respect to our approach.
6.1 Model-based Testing
Model-based testing encompasses a broad range of black-box testing tools that facilitate
generating concrete test-cases from an abstract model of the system under test. These
systems generally (though not necessarily) model the system at a holistic level using
state machines to describe the desired behavior [6], and may or may not provide fully
automatic test-case generation. In addition to generating test-cases, many model-based
testing tools, e.g. Spec Explorer [28] will produce extra artifacts like visualizations to
help the programmer understand the model. One could view property-based testing, including our system, as a subset of model-based testing focusing on lower-level properties of individual functions (unit-testing), while using the type-structure of the functions
under scrutiny to provide fully automatic generation of test-cases.
22
6.2 Property-based Testing
Many property-based testing tools have been developed to automatically generate testsuites. QuickCheck [4] randomly generates inputs based on the property under scrutiny,
but requires custom generators to consistently generate constrained inputs. [3] extends
QuickCheck to randomly generate constrained values from a uniform distribution. In
contrast SmallCheck [21] enumerates all possible inputs up to some depth, which allows it to check existential properties in addition to universal properties; however, it too
has difficulty generating inputs to properties with complex pre-conditions. Lazy SmallCheck [21] addresses the issue of generating constrained inputs by taking advantage
of the inherent laziness of the property, generating partially-defined values (i.e. values
containing ⊥) and only filling in the holes if and when they are demanded. Korat [2]
instruments a programmer-supplied repOk method, which checks class invariants and
method pre-conditions, to monitor which object fields are accessed. The authors observe
that unaccessed fields cannot have had an effect on the return value of repOk and are
thereby able to exclude from the search space any objects that differ only in the values
of the unaccessed fields. While Lazy SmallCheck and Korat’s reliance on functions in
the source language for specifying properties is convenient for the programmer (specification and implementation in the same language), it makes the method less amenable
to formal verification, the properties would need to be re-specified in another language
that is restricted enough to facilitate verification.
6.3 Symbolic Execution and Model-checking
Another popular technique for automatically generating test-cases is to analyze the
source code and attempt to construct inputs that will trigger different paths through
the program. DART [12], CUTE [22], and Pex [24] all use a combination of symbolic
and dynamic execution to explore different paths through a program. While executing
the program they collect path predicates, conditions that characterize a path through a
program, and at the end of a run they negate the path predicates and query a constraint
solver for another assignment of values to program variables. This enables such tools
to efficiently explore many different paths through a program, but the technique relies on the path predicates being expressible symbolically. When the predicates are not
expressible in the logic of the constraint solver, they fall back to the values produced
by the concrete execution, at a severe loss of precision. Instead of trying to trigger all
paths through a program, one might simply try to trigger erroneous behavior. Check ’n’
Crash [5] uses the ESC/Java analyzer [10] to discover potential bugs and constructs concrete test-cases designed to trigger the bugs, if they exist. Similarly, [1] uses the BLAST
model-checker to construct test-cases that bring the program to a state satisfying some
user-provided predicate.
In contrast to these approaches, TARGET (and more generally, property-based testing) treats the program as a black-box and only requires that the pre- and post-conditions
be expressible in the solver’s logic. Of course, by expressing specifications in the source
language, e.g. as contracts, as in PEX [24], one can use symbolic execution to generate tests directly from specifications. One concrete advantage of our approach over the
symbolic execution based method of PEX is that the latter generates tests by explicitly
23
enumerating paths through the contract code, which suffers from a similar combinatorial problem as SmallCheck and QuickCheck. In contrast, TARGET performs the same
search symbolically within the SMT engine, which performs better for larger input sizes.
6.4 Integrating Constraint-solving and Execution
TARGET is one of many tools that makes specifications executable via constraint solving. An early example of this approach is TestEra [17] that uses specifications written
in the Alloy modeling language [13] to generate all non-isomorphic Java objects that
satisfy method pre-conditions and class invariants. As the specifications are written in
Alloy, one can use Alloy’s SAT-solver based model finding to symbolically enumerate
candidate inputs. Check ’n’ Crash uses a similar idea, and SMT solvers to generate
inputs that satisfy a given JML specification [5]. Recent systems such as SBV [8] and
Kaplan [14] offer a monadic API for writing SMT constraints within the program, and
use them to synthesize program values at run-time. SBV provides a thin DSL over the
logics understood by SMT solvers, whereas Kaplan integrates deeply with Scala, allowing the use of user-defined recursive types and functions. Test generation can be viewed
as a special case of value-synthesis, and indeed Kaplan has been used to generate testsuites from preconditions in a similar manner to TARGET.
However, in all of the above (and also symbolic execution based methods like PEX
or JCrasher), the specifications are assertions in the Floyd-Hoare sense. Consequently,
the techniques are limited to testing first-order functions over monomorphic data types.
In contrast, TARGET shows how to view types as executable specifications, which yields
several advantages. First, we can use types to compositionally lift specifications about
flat values (e.g. Score) over collections (e.g. [Score]), without requiring special recursive predicates to describe such collection invariants. Second, the compositional nature of types yields a compositional method for generating tests, allowing us to use
type-class machinery to generate tests for richer structures from tests for sub-structures.
Third, (refinement) types have proven to be effective for verifying correctness properties in modern modern languages that make ubiquitous use of parametric polymorphism
and higher order functions [29,7,20,23,26] and thus, we believe TARGET’s approach of
making refinement types executable is a crucial step towards our goal of enabling gradual verification for modern languages.
Acknowledgements
This work was supported by NSF grants CCF-1422471, CNS-0964702, CNS-1223850,
CCF-1218344, CCF-1018672, and a generous gift from Microsoft Research. We thank
Lee Pike and the reviewers for their excellent feedback on a draft of this paper.
References
1. Beyer, D., Chlipala, A.J., Henzinger, T.A., Jhala, R., Majumdar, R.: Generating tests from
counterexamples. In: ICSE 04: Software Engineering. pp. 326–335 (2004)
2. Boyapati, C., Khurshid, S., Marinov, D.: Korat: Automated testing based on Java predicates.
In: ISSTA 02: Software Testing and Analysis. pp. 123–133. ACM (2002)
24
3. Claessen, K., Duregård, J., Palka, M.H.: Generating constrained random data with uniform
distribution. pp. 18–34. FLOPS ’14 (2014)
4. Claessen, K., Hughes, J.: QuickCheck: a lightweight tool for random testing of haskell programs. In: ICFP. ACM (2000)
5. Csallner, C., Smaragdakis, Y.: Check ’n’ crash: combining static checking and testing. In:
ICSE. pp. 422–431 (2005)
6. Dias Neto, A.C., Subramanyan, R., Vieira, M., Travassos, G.H.: A survey on model-based
testing approaches: A systematic review. In: WEASELTech ’07. ACM (2007)
7. Dunfield, J.: Refined typechecking with Stardust. In: PLPV (2007)
8. Erkök, L.: SBV: SMT based verification in haskell, http://leventerkok.github.io/sbv/
9. Findler, R.B., Felleisen, M.: Contract soundness for object-oriented languages. In: OOPSLA.
pp. 1–15 (2001)
10. Flanagan, C., Leino, K., Lillibridge, M., Nelson, G., Saxe, J.B., Stata, R.: Extended static
checking for Java. In: PLDI (2002)
11. Gill, A., Runciman, C.: Haskell program coverage. pp. 1–12. Haskell ’07, ACM (2007)
12. Godefroid, P., Klarlund, N., Sen, K.: Dart: directed automated random testing. In: PLDI. pp.
213–223 (2005)
13. Jackson, D.: Alloy: a lightweight object modelling notation. ACM Transactions on Software
Engineering and Methodology (TOSEM) 11(2), 256–290 (2002)
14. Köksal, A.S., Kuncak, V., Suter, P.: Constraints as control. pp. 151–164. POPL ’12, ACM,
New York, NY, USA (2012)
15. Magalhães, J.P., Dijkstra, A., Jeuring, J., Löh, A.: A generic deriving mechanism for haskell.
In: Haskell Symposium. ACM (2010)
16. Marick, B.: How to misuse code coverage. In: Proceedings of the 16th Interational Conference on Testing Computer Software. pp. 16–18 (1999)
17. Marinov, D., Khurshid, S.: Testera: A novel framework for automated testing of java programs. ASE ’01, IEEE Computer Society, Washington, DC, USA (2001)
18. de Moura, L., Bjørner, N.: Generalized, efficient array decision procedures. In: FMCAD
(2009)
19. Nelson, G.: Techniques for program verification. Tech. Rep. CSL81-10, Xerox Palo Alto
Research Center (1981)
20. Nystrom, N., Saraswat, V.A., Palsberg, J., Grothoff, C.: Constrained types for object-oriented
languages. In: OOPSLA. pp. 457–474 (2008)
21. Runciman, C., Naylor, M., Lindblad, F.: Smallcheck and lazy smallcheck: Automatic exhaustive testing for small values. In: Haskell Symposium. ACM (2008)
22. Sen, K., Marinov, D., Agha, G.: Cute: A concolic unit testing engine for c. In: ESEC/FSE.
ACM (2005)
23. Swamy, N., Chen, J., Fournet, C., Strub, P.Y., Bhargavan, K., Yang, J.: Secure distributed
programming with value-dependent types. In: ICFP (2011)
24. Tillmann, N., Halleux, J.d.: Pex–White box test generation for .NET. In: Tests and Proofs,
pp. 134–153 (2008)
25. Vazou, N., Rondon, P., Jhala, R.: Abstract refinement types. In: ESOP (2013)
26. Vazou, N., Seidel, E.L., Jhala, R.: Liquidhaskell: Experience with refinement types in the
real world. In: Haskell Symposium (2014)
27. Vazou, N., Seidel, E.L., Jhala, R., Vytiniotis, D., Peyton-Jones, S.: Refinement types for
haskell. In: ICFP (2014)
28. Veanes, M., Campbell, C., Grieskamp, W., Schulte, W., Tillmann, N., Nachmanson, L.:
Model-based testing of object-oriented reactive systems with spec explorer. In: Formal Methods and Testing. Springer-Verlag (2008)
29. Xi, H., Pfenning, F.: Eliminating array bound checking through dependent types. In: PLDI
(1998)
25
| 6cs.PL
|
On freeness of the random fundamental group
arXiv:1601.07520v3 [math.AT] 10 Oct 2017
Andrew Newman
∗
October 12, 2017
Abstract
Let Y (n, p) denote the probability space of random 2-dimensional simplicial complexes in the
Linial–Meshulam model, and let Y ∼ Y (n, p) denote a random complex chosen according to this
distribution. In a paper of Cohen, Costa, Farber, and Kappeler, it is shown that for p = o(1/n)
with high probability π1 (Y ) is free. Following that, a paper of Costa and Farber shows that for
values of p which satisfy 3/n < p ≪ n−46/47 , with high probability π1 (Y ) is not free. Here we
improve on both of these results to show that there are explicit constants γ2 < c2 < 3, so that
for p < γ2 /n with high probability Y has free fundamental group and that for p > c2 /n, with
high probability Y has fundamental group which either is not free or is trivial.
1
Introduction
For positive integers n and d and p = p(n) ∈ [0, 1], the space of Linial–Meshulam random ddimensional simplicial complexes, first introduced in [10] and [13] and denoted Yd (n, p), is defined
to be the probability space of d-dimensional simplicial
complexes on n vertices with complete
n
(d − 1)-skeleton where each of the possible d+1
d-dimensional faces is included independently
with probability p. Here we are primarily interested in the d = 2 case and so we suppress the
dimension parameter and write Y (n, p) for Y2 (n, p). Now the question of the fundamental group
of Y ∼ Y (n, p) is nontrivial and has been studied in [4, 6, 7, 8]. Additionally the series of papers
[1, 2, 3, 12], study Yd (n, p) in the regime p = c/n. We will describe these results below, but we
introduce two constants first introduced in [3] and [1] that are needed to state our main theorem.
− log y
Let γ2 = (2x(1−x))−1 where x is the unique nonzero solution to exp(− 1−x
2x ) = x and let c2 = (1−y)2
where y is the unique root in (0, 1) of 3(1 − y) + (1 + 2y) log y = 0. Here we build on the work of
[3], [7], and [12] to prove the following result about the fundamental group of a random 2-complex.
Note that most of the theorems stated here are asymptotic results and we use the phrase “with high
probability”, abbreviated w.h.p., to mean that a property holds for Y ∼ Yd (n, p) with probability
tending to 1 as n tends to infinity. The following theorem is the main result of this paper.
Theorem 1. If c < γ2 and Y ∼ Y (n, c/n), then with high probability π1 (Y ) is a free group and if
c > c2 and Y ∼ Y (n, c/n) then with high probability π1 (Y ) is not a free group.
Now γ2 is first defined in [3] and c2 is first defined in [1] and approximations are computed as
γ2 ≈ 2.455407 and c2 ≈ 2.753806.
∗
The Ohio State University
1
2
The lower bound
In this section we prove the first part of theorem that for c < γ2 one has π1 (Y ) is a free group with
high probability for Y ∼ Y (n, c/n). This result will follow by adapting the argument of [3] used to
prove the following result.
Theorem 2 (2-dimensional case of Theorem 1.4 from [3]). Let γ2 be as above. If c < γ2 then w.h.p.
Y ∼ Y (n, c/n) is 2-collapsible or contains ∂∆3 as a subcomplex.
We first define what it means for a simplicial complex to be d-collapsible. For a d-dimensional
simplicial complex Y , we say that a (d− 1)-dimensional face τ is free if it is contained in exactly one
d-dimensional face σ ∈ Y . For a free (d − 1)-face τ an elementary collapse at τ is defined to be the
simplicial complex Y ′ obtained from Y be removing τ and the unique d-face σ in Y containing τ .
If there is a sequence of elementary collapses that removes all d-dimensional faces of Y we say that
Y is d-collapsible. Observe that elementary collapses are homotopy equivalences, so if a 2-complex
is 2-collapsible (to a graph) then in particular it has free fundamental group. Therefore theorem 2
above almost proves the lower bound except for the problem of tetrahedron boundaries. Note that
it is impossible to rule out ∂∆3 appearing as a subcomplex of Y ∼ Y (n, c/n) for any c > 0 since
the expected number of copies of ∂∆3 in Y ∼ Y (n, c/n) approaches a Poisson distribution with
mean c4 /24. Additionally, [3] does not state any result about partial collapsibility in the presence
of a few copies of ∂∆3 and indeed it is not clear that any partial collapsibility result would hold.
However such a result is not needed to imply that the fundamental group of Y ∼ Y (n, c/n) is free
for c < γ2 as we will see below.
Following the convention of [3] define a core to be a 2-dimensional simplicial complex in which
every edge is contained in at least two faces. Also for a 2-complex Y , let R(Y ) denote the simplicial
complex obtained by collapsing all the free edges of Y and let R∞ (Y ) denote the simplicial complex
obtained after repeatedly collapsing at all free edges until no free edges remain. The two key results
of [3] that we will use are the following.
Theorem 3 (2-dimensional case of Theorem 4.1 from [3]). For every c > 0 there exists a constant
δ = δ(c) > 0 such that w.h.p. every core subcomplex K of Y ∼ Y (n, c/n) with f2 (K) ≤ δn2 must
contain the boundary of a tetrahedron.
Theorem 4 (2-dimensional case of Theorem 5.3 from [3]). Let δ > 0 and 0 < c < γ2 be fixed and
suppose Y ∼ Y (n, c/n). Then w.h.p. f2 (R∞ (Y )) ≤ δn2 .
Now to bound the probability that π1 (Y ) is not a free group for Y ∼ Y (n, c/n), we will bound
the probability that Y ∼ Y (n, c/n) for c < γ2 has a core which contains no tetrahedron boundary
or has a pair of tetrahedron boundaries that are not face disjoint. This will be an upper bound to
the probability that π1 (Y ) is not free by the following proposition.
Proposition 5. Let Y be a 2-dimensional simplicial complex. If every core of Y contains a tetrahedron boundary and all the tetrahedron boundaries are face-disjoint then π1 (Y ) is free
Proof. Let Ỹ be the 3-dimensional simplicial complex obtained from Y be adding a 3-simplex inside
all the tetrahedron boundaries of Y . Now Ỹ (2) = Y so π1 (Y ) = π1 (Ỹ ). Now let Z be obtained from
Ỹ by collapsing at a free 2-dimensional face at every 3-dimensional face, such a collapse will remove
2
all the tetrahedra from Ỹ as the tetrahedron boundaries in Y are face disjoint so every tetrahedron
in Ỹ has that all of its faces are free. Equivalently, Z is obtained from Y by deleting one face
from every tetrahedron boundary of Y . Now collapsing at free faces is a homotopy equivalence so
π1 (Z) = π1 (Ỹ ). Furthermore Z is 2-collapsible. Indeed Z has no cores as a core K of Z would be
a core in Y as well since Z is obtained from Y by removing faces. But every core of Y contains a
tetrahedron boundary and Z has no tetrahedron boundaries. Since Z does not have a core it must
be 2-collapsible, otherwise deleting all the isolated edges of R∞ (Z) would give us a subcomplex
of Z that has no faces of degree zero or one, so such a subcomplex would be a core. Thus Z is
2-collapsible, in particular Z is homotopy equivalent to a graph so π1 (Z) is a free group.
Now we are ready to prove the first part of Theorem 1, that is for c < γ2 and Y ∼ Y (n, c/n),
π1 (Y ) is a free group with high probability.
Proof of lower bound on Theorem 1. Let c < γ2 be fixed and suppose Y ∼ Y (n, c/n), by proposition 5, the probability that π1 (Y ) is not free is bounded above by the sum of the probability that
Y contains tetrahedron boundaries that share a face and the probability that Y has a core with
no tetrahedron boundary. First it is easy to bound the probability that Y contains tetrahedron
boundaries that share a face. Two tetrahedron boundaries in a simplicial complex sharing a face
must meet in exactly one face. Two tetrahedron boundaries meeting at one face is a simplicial
complex with 5 vertices and 7 faces, the expected number of such subcomplexes in Y ∼ Y (n, p)
is O(n5 p7 ) which in this case is O(c7 /n2 ) = o(1). So by Markov’s inequality the probability that
there are tetrahedron boundaries in Y that are not face disjoint is o(1).
We will now use the two theorems from [3] above to show that the probability that Y has a core
with no tetrahedron boundary is o(1). Let δ = δ(c) be the δ given by Theorem 3. Let F denote
the collection of 2-complexes on n vertices containing a core with no tetrahedron boundary and
let G denote the collection of 2-complexes on n vertices for which all cores have size at most δn2 .
Note that if Y ∈
/ G , then f2 (R∞ (Y )) > δn2 since cores are unaffected by elementary collapses, so
P r(Y ∈
/ G ) = o(1) by Theorem 4. Now we bound P r(Y ∈ F ).
P r(Y ∈ F ) = P r(Y ∈ F ∩ G ) + P r(Y ∈ F \ G )
≤ P r(Y ∈ F ∩ G ) + P r(Y ∈
/ G)
≤ P r(Y ∈ F ∩ G ) + o(1)
≤ P r(Y ∈ {X : X has a core K with at most δn2 faces and no ∂∆3 }) + o(1)
Now by the choice of δ and Theorem 3, we know that the probability that Y has a core which has
at most δn2 faces but no tetrahedron boundary is o(1). Thus we have that P r(Y ∈ F ) = o(1)
which completes the proof.
3
The upper bound
We now turn our attention to proving that when c > c2 and Y ∼ Y (n, c/n), with high probability
π1 (Y ) is not a free group. In fact relevant results by Costa and Farber ([7]) will prove that the
cohomological dimension is 2. We refer the reader to [5] for background on group cohomology
theory. The main result of [7] is the following:
3
Theorem 6 (Theorem 2 of [7]). Assume that p ≪ n−46/47 , then w.h.p. a random 2-complex
Y ∼ Y (n, p) is asphericable. That is the complex Z obtained from Y by removing one face from
each tetrahedron of Y is aspherical (i.e. the universal cover of Z is contractible).
From here Costa and Farber prove the following result.
Theorem 7 (Theorem 3B of [7]). For any constants c > 3 and 0 < ǫ < 1/47 and p satisfying
c/n < p < n−1+ǫ , the cohomological dimension of Y ∼ Y (n, p) equals 2 with high probability.
To prove our upper bound from Theorem 1 we will use the following result of Linial and Peled
[12] to reduce the constant 3 in Theorem 7 to c2 , the argument will follow exactly the argument
of Costa and Farber in their proof of Theorem 7, but with the current state-of-the-art (and bestpossible, also by [12]) threshold for emergence of homology in degree 2 for a random 2-complex.
Theorem 8 (Special case of Theorem 1.3 from [12]). Suppose c > c2 , then w.h.p. Y ∼ Y (n, c/n)
has
dim H2 (Y ; R) = Θ(n2 ).
In [12], the constant implicit in Θ(n2 ) is given explicitly, but we do not need it here. We are
now ready to prove the second part of theorem 1.
Proof of upper bound on Theorem 1. Fix c > c2 and suppose that Y is a simplicial complex drawn
from Y (n, c/n). Now let Z be obtained from Y by removing one face from every tetrahedron
boundary. With high probability π1 (Y ) = π1 (Z), and by theorem 6, Z is aspherical. Therefore
showing that β2 (Z) 6= 0 would imply that the cohomological dimension of π1 (Y ) is at least two.
Now by theorem 8 we know that with high probability β2 (Y ) = Θ(n2 ). Also we have by a first
moment argument that the expected number of tetrahedron boundaries is bounded above by c4 /24.
Therefore by Markov’s inequality with high probability Y has no more than, say, n tetrahedron
boundaries. Now given a 2-dimensional simplicial complex, removing a face can drop β2 by at most
one. Therefore when we remove one face from from each tetrahedron boundary of Y to obtain Z
we drop β2 (Y ) by at most n, then w.h.p. β2 (Z) = Θ(n2 ) > 0. Thus the cohomological dimension
of π1 (Y ) is at least two (actually equality holds by theorem 6), in particular π1 (Y ) is not a free
group.
4
Concluding Remarks
The statement of Theorem 1 perhaps implicitly suggests a sharp threshold for the property that
a random 2-complex has fundamental group which is not free. However, it is worth mentioning
that the property that the fundamental group of a simplicial complex is free is not a monotone
property, so it is not obvious at all that a sharp threshold should exist. However, by theorem 1
and theorem 7, for c2 /n < p < 3 log n/n and Y ∼ Y (n, p), with high probability π1 (Y ) is not free.
Combining this with a result from [8] that for p > (2 log n + ω(n))/n (with ω(n) → ∞ as n → ∞)
and Y ∼ Y (n, p), π1 (Y ) has property (T) with high probability, we have that for p > c2 /n, the
fundamental group of Y ∼ Y (n, p) is with high probability free only if it is trivial as the only free
group with property (T) is the trivial group. On the other hand, [6] proves that for p = o(1/n),
Y ∼ Y (n, p) collapses to a graph with high probability. Thus we do have at least a coarse threshold
of 1/n for the fundamental group of a random 2-complex to be either not free or trivial.
4
It remains to discover the fundamental group for Y (n, p) for γ2 ≤ c ≤ c2 and p = c/n. Right
now, there does not seem to be enough evidence to establish a conjecture. There are the following
three possibilites for what happens to the fundamental group of Y (n, p) in this intermediate regime:
1. γ2 /n is the sharp threshold for the fundamental group of Y (n, p) to go from a free group to
a non-free group.
2. c2 /n is the sharp threshold for the fundamental group of Y (n, p) to go from a free group to
a non-free group.
3. Neither γ2 /n nor c2 /n is the sharp threshold for the fundamental group of Y (n, p) to go from
a free group to a non-free group.
Any of these three would be interesting in their own way.
If (1) holds, then by [7], Y (n, c/n) in the regime c ∈ (γ2 , c2 ) has cohomological dimension equal
to 2 and Y (n, c/n) is asphericable. Thus when we remove a face from every tetrahedron we have that
Y (n, c/n) is a K(G, 1) for a group G which has cohomological dimension 2. However, the reason for
it to have cohomological dimension 2 must be different than the reason for π1 (Y (n, c/n)) to have
cohomological dimension 2 in the regime c > c2 . Indeed, in the regime c > c2 , H2 (Y (n, c/n), R) 6= 0
is enough to imply that the cohomological dimension of the fundamental group is at least 2. For
c < c2 the second homology group of Y (n, c/n) with coefficients in R is trivial, after the removal
of a face from each tetrahedron boundary.
Moreover, there is an apparent lack of torsion in H1 (Y (n, c/n)) if c < c2 . This has not been
proved, but extensive experiments conducted in [9] provide evidence to support this, and [14] state
the following conjecture regarding torsion in homology:
Conjecture (2-dimensional case of the conjecture from [14]). For every p = p(n) such that |np−c2 |
is bounded away from 0, H1 (Yd (n, p); Z) is torsion-free with high probability.
Torsion in homology is observed experimentally for p close to cd /n. For more about this torsion
see [9]. All of this is to say that it is likely that one will not be able to prove that the cohomological
dimension of π1 (Y (n, c/n)) ≥ 2 for γ2 < c < c2 by proving H2 (Y (n, c/n), Z/qZ) 6= 0 for some prime
q.
On the other hand if (2) holds, then Y (n, c/n) in the regime γ2 < c < c2 is homotopy equivalent
to a wedge of circles after the removal of one face from each tetrahedron boundary. This follows
from the fact that Y (n, c/n) is asphericable and that an aspherical space which is a CW-complex
is unique up to homotopy equivalence. Now for c < γ2 , as we show above following the results of
[3], Y (n, c/n) is homotopy to a wedge of circles after the removal of a face from each tetrahedron
boundary. However, this homotopy equivalence is given by a sequence of elementary collapses which
reduces the complex to a graph. It is proved in [2] that such a series of elementary collapses is not
possible for c > γ2 . Furthermore, [11] points out that in the regime γ2 < c < c2 , Y (n, c/n) is far
from being 2-collapsible in the sense that a constant fraction of the faces must be deleted to arrive
at a complex which is 2-collapsible. Thus Y (n, c/n) in the regime γ2 < c < c2 if (2) holds would
be homotopy equivalent to a wedge of circles, but not via the same type of homotopy equivalence
5
which exists for smaller values of c.
In summary, regardless of whether the truth is (1) or (2), new techniques will almost certainly
be required to prove which is correct. Of course, (3) is a possibility as well. Indeed it is possible
that no sharp threshold exists for the property that π1 (Y (n, p)) as we discuss above. It could also
be that there is some c∗ ∈ (γ2 , c2 ) so that c∗ /n is the sharp threshold for π1 (Y (n, p)) to be non-free,
or that within this intermediate regime there is a positive probability Y (n, p) is free and a positive
probability that it is not.
References
[1] L. Aronshtam and N. Linial, When does top homology in a random simplicial complex vanish?,
Random Structures & Algorithms 46 (2013), 26–35.
[2]
, The threshold for d-collapsibility in random complexes, Random Structures & Algorithms 48 (2015), 260–269.
[3] L. Aronshtam, N. Linial, T. Luczak, and R. Meshulam, Collapsibility and vanishing of top
homology in random simplicial complexes., Discrete & Computational Geometry 49 (2013),
317–334.
[4] E. Babson, C. Hoffman, and M. Kahle, The fundamental group of random 2-complexes, Journal
of the American Mathematical Society 24 (2011), 1–28.
[5] K. S. Brown, Cohomology of groups, Springer-Verlag New York Inc., 1982.
[6] D. Cohen, A. Costa, M. Farber, and T. Kappeler, Topology of random 2-complexes, Discrete
& Computational Geometry 47 (2012), 117–149.
[7] A. Costa and M. Farber, The asphericity of random 2-dimensional complexes, Random Structures & Algorithms 46 (2015), 261–273.
[8] C. Hoffman, M. Kahle, and E. Paquette, Spectral gaps of random graphs and applications to
random topology, arXiv: 1201.0425.
[9] M. Kahle, F. H. Lutz, A. Newman, and K. Parsons, Cohen–Lenstra heuristics for torsion in
homology of random complexes, In preparation.
[10] N. Linial and R. Meshulam, Homological connectivity of random 2-complexes, Combinatorica
26 (2006), 475–487.
[11] N. Linial and Y. Peled, Random simplicial complexes – around the phase transition, arXiv:
1609.00914.
[12]
, On the phase transition in random simplicial complexes, Annals of Mathematics 184
(2016), 745–773.
[13] R. Meshulam and N. Wallach, Homological connectivity of random k-dimensional complexes,
Random Structures & Algorithms 34 (2009), 408–417.
[14] T. Luczak and Y. Peled, Integral homology of random simplicial complexes, arXiv: 1607.06985.
6
| 4math.GR
|
Reducing bias in nonparametric density estimation via bandwidth
dependent kernels: L1 view1
arXiv:1611.10203v1 [math.ST] 30 Nov 2016
Kairat Mynbaev
International School of Economics
Kazakh-British Technical University
Tolebi 59
Almaty 050000, Kazakhstan
email: kairat [email protected]
and
Carlos Martins-Filho
Department of Economics
University of Colorado
Boulder, CO 80309-0256, USA
email: [email protected]
&
IFPRI
2033 K Street NW
Washington, DC 20006-1002, USA
email: [email protected]
November, 2016
Abstract. We define a new bandwidth-dependent kernel density estimator that improves existing convergence rates for the bias, and preserves that of the variation, when the error is measured in L1 . No additional
assumptions are imposed to the extant literature.
Keywords and phrases. Kernel density estimation, higher order kernels, bias reduction.
AMS-MS Classification. 62G07, 62G10, 62G20.
1 We
thank an anonymous referee and an Associate Editor for excellent comments that improved this note significantly.
1
Introduction
Given a sequence of n ∈ N independent realizations {Xj }nj=1 of the random variable X, having density f on
R, the Rosenblatt-Parzen kernel estimator (Rosenblatt (1956), Parzen (1962)) of f is given by
n
1X
(Shn K)(x − Xj ),
n j=1
fn (x) =
(1.1)
where Shn is an operator defined by
1
(Shn K)(x) =
K
hn
K is a kernel, i.e., a function on R such that
that hn → 0 as n → ∞.1
R
x
hn
,
(1.2)
K(x)dx = 1 and hn > 0 is a non-stochastic bandwidth such
One of the most natural and mathematically sound (Devroye and Györfi (1985), Devroye (1987)) criteria
to measure the performance of fn as an estimator of f is the L1 distance
R
|fn − f |. In particular, given that
this distance is a random variable (measurable function of {Xj }nj=1 ) it is convenient to focus on E
R
|fn − f | ,
where E denotes the expectation taken using f . For this criterion, there is a simple bound (Devroye, 1987,
p. 31)
E
Z
|fn − f |
≤
Z
|(f ∗ Shn K) − f | + E
where for arbitrary f, g ∈ L1 , (f ∗ g)(x) =
R
|f ∗ Shn K − f | is called bias over R and E
R
R
Z
|fn − f ∗ Shn K| ,
g(y)f (x − y)dy is the convolution of f and g. The term
|fn − f ∗ Shn K| is called the variation over R. There exists
a large literature devoted to establishing conditions on f and K that assure suitable rates of convergence
of the bias to zero as n → ∞ (see, inter alia, Silverman (1986), Devroye (1987) and Tsybakov (2009)). In
particular, if K is of order s, i.e., αj (K) = 0 for j = 1, ..., s − 1 and αs (K) 6= 0, where αj (K) =
is the jth moment of K, and f has an integrable derivative f (s) , then
R
R
tj K(t)dt
|f ∗ Shn K − f | is of order O(hsn )
and this order cannot be improved, see, e.g., (Devroye, 1987, Theorem 7.2). In this note, we show that if in
(1.2) the kernel is allowed to depend on n, then the order O(hsn ) can be replaced by the order o(hsn ), without
increasing the order of the kernel or the smoothness of the density. In addition, another result from Devroye
1 Throughout
this note, integrals are over R, unless otherwise specified.
1
(1987) states that if K is a kernel of order greater than s and the derivative f (s) is a-Lipschitz then the bias
is of order O(hs+a
n ). We achieve the same rate of convergence with kernels of order s.
2
Main results
Let L1 and C denote the spaces of integrable and (bounded) continuous functions on R with norms kf k1 =
R
|f | and kf kC = sup |f |, and βs (K) =
R
|t|s |K(t)| dt. Let {Kn } be a sequence of kernels and define
n
1X
fˆn (x) =
(Shn Kn )(x − Xj ).
n j=1
In the following Theorem 1, the density f has the same degree of smoothness and the kernels Kn are of the
same order as in (Devroye, 1987, Theorem 7.2), but the bias is of order o(hsn ) instead of O(hsn ). This results
because the kernels depend on n and have “disappearing” moments of order s.
Theorem 1. Let {Kn } be a sequence of kernels of order s such that: 1. αs (Kn ) → 0; 2. {us Kn (u)} is
uniformly integrable. For all f with absolutely continuous f (s−1) and f (s) ∈ L1 , we have kf ∗ Shn Kn − f k1 =
o(hsn ).
Proof. Note that since Kn is a kernel
f ∗ Shn Kn (x) − f (x) =
Z
Kn (t)[f (x − hn t) − f (x)]dt.
(2.1)
Since f is s-times differentiable, by Taylor’s Theorem,
f (x − hn t) − f (x) =
s−1 (j)
X
f (x)
j!
j=1
j
(−hn t) +
x−hn t
Z
x
(x − hn t − u)s−1 (s)
f (u)du.
(s − 1)!
Furthermore, given that Kn is of order s,
1
f ∗ Shn Kn (x) − f (x) =
(s − 1)!
Z Z
x−hn t
(x − hn t − u)s−1 f (s) (u)duKn (t)dt.
x
(2.2)
Letting λ = − u−x
hn t we have
Z
x−hn t
x
s−1 (s)
(x − hn t − u)
f
s
(u)du = (−hn t)
Z
0
2
1
f (s) (x − hn λt)(1 − λ)s−1 dλ.
(2.3)
Substituting (2.3) into (2.2) we obtain
f ∗ Shn Kn (x) − f (x) =
Since
R1
0
(−hn )s
s!
Z Z
1
0
f (s) (x − hn λt)s(1 − λ)s−1 dλts Kn (t)dt.
(2.4)
(1 − λ)s−1 dλ = 1s , we have that
(−hn )s
(s − 1)!
Z Z
1
f
(s)
0
s−1
(x)(1 − λ)
(−hn )s (s)
f (x)
dλt Kn (t)dt =
s!
s
Z
ts Kn (t)dt.
(2.5)
Then, adding and subtracting (2.5) to the right-hand side of (2.4) gives
f ∗ Shn Kn (x) − f (x) =
(−hn )s
s!
Z Z 1
f (s) (x)αs (Kn ) +
[f (s) (x − hn λt) − f (s) (x)]s(1 − λ)s−1 dλts Kn (t)dt .
0
Since f (s) ∈ L1 we write its continuity modulus as ω(δ) = sup|t|≤δ
R
f (s) (x − t) − f (s) (x) dx. It is well-
known (see properties M.2, M.6 and M.7 in Zhuk and Natanson (2003)) that
ω(δ) ≤ 2 f (s)
1
, ω is nondecreasing and lim ω(δ) = 0.
(2.6)
δ→0
Then,
kf ∗ Shn Kn − f k1 ≤
≤
=
≤
hsn
s!
hsn
s!
f (s)
f (s)
hsn (s)
f
s!
hsn (s)
f
s!
1
1
1
|αs (Kn )| +
Z Z
1
0
1
Z
f (s) (x − hn λt) − f (s) (x) dx s(1 − λ)s−1 dλ|ts Kn (t)|dt
ω(λhn |t|) | ts Kn (t)| dt s(1 − λ)s−1 dλ
0
Z
Z 1 Z
ω(λhn |t|)|ts Kn (t)|dt s(1 − λ)s−1 dλ
+
|αs (Kn )| +
|αs (Kn )| +
Z
0
Z
|t|≤ √1
hn
|t|> √1
hn
(2.7)
1
|αs (Kn )| + ω
p
hn βs (Kn ) + 2 f (s)
1
Z
|t|> √1
hn
|ts Kn (t)| dt .
(2.8)
Given that αs (Kn ) → 0 as n → ∞, {ts Kn (t)} is uniformly integrable, which implies sup βs (Kn ) < ∞, and
n
using (2.6) and (2.8) we have
kf ∗ Shn Kn − f k1 = o(hsn ).
(2.9)
Remark 1. Kernel sequences {Kn } that satisfy the restrictions imposed by Theorem 1 can be easily
constructed. To this end, denote by Bs the space of functions with bounded norm kKkBs = β0 (K) + βs (K).
3
Take functions K(0) , K(s) ∈ Bs such that
α0 (K(0) ) = 1, αj (K(0) ) = 0 for j = 1, ..., s; αj (K(s) ) = 0 for j = 0, 1, ..., s − 1, αs (K(s) ) = 1.
(2.10)
We define the n-dependent kernel Kn = K(0) + hn K(s) with 0 < hn ≤ 1. Note that Kn is a kernel of order
s with αs (Kn ) = hn which tends to zero as n → ∞. It is clear that any kernel K of order s can be written
as K = K(0) + αs (K)K(s) , so that the conventional s-order kernels obtain from ours with αs (K) = hn .
Furthermore, it follows from (2.10) that {ts (K(0) (t) + hn K(s) (t))} is uniformly integrable.
Now, to obtain K(0) and K(s) , assume that for a nonnegative kernel K we have β2s (K) < ∞. Then, we
can associate with K a symmetric matrix
α0 (K) α1 (K)
α1 (K) α2 (K)
As =
...
...
αs (K) αs+1 (K)
...
...
...
...
αs (K)
αs+1 (K)
,
...
α2s (K)
such that det As 6= 0 (see Mynbaev et al. (2014)). For an arbitrary vector b ∈ Rs+1 let a = A−1
s b and
define a polynomial transformation of K by (Ta K)(t) =
Ps
i=0
ai ti K(t). Then, we put K(0) = Ta K with
b = (1, 0, ..., 0)′ and K(s) = Ta K with b = (0, ..., 0, 1)′ , which satisfy (2.10). Thus, we have the following
corollary to Theorem 1.
Corollary 1. Let Kn = K(0) + hn K(s) where K(0) and K(s) are as defined in Remark 1. Then, for all f
with absolutely continuous f (s−1) and f (s) ∈ L1 , we have kf ∗ Shn Kn − f k1 = o(hsn ).
Remark 2. If Kn are supported on [−M, M ] for some M > 0 and for all n, then in (2.7), instead of splitting
√
√
R = |t| ≤ 1/ hn ∪ |t| > 1/ hn we can use R = {|t| ≤ M } ∪ {|t| > M } and then instead of (2.8) we get
kf ∗ Shn Kn − f k1 ≤
hsn h (m)
f
s!
1
i
|αs (Kn )| + ω(hn M )βs (Kn ) .
Hence, selecting {Kn } in such a way that αs (Kn ) = O(ω(hn )), sup βs (Kn ) < ∞ and using the fact that
n
2
ω(hn M ) ≤ (M + 1) ω(hn ) we get a result that is more precise than (2.9), i.e.,
kf ∗ Shn Kn − f k1 = O(hsn ω(hn )).
2 See
property M.5 in Zhuk and Natanson (2003).
4
Remark 3. By Young’s inequality, the variation of fˆn using Kn = K(0) + hn K(s) is such that
E
(0)
Letting fn
Z
|fˆn − f ∗ Shn Kn | ≤ E
Z
|fˆn − f ∗ Shn K(0) | + hn
|f ∗ Shn K(s) |
Z
Z
Z
≤ E |fˆn − f ∗ Shn K(0) | + hn |f | |K(s) |.
be the estimator in (1.1) with K = K(0) , we have E
Shn K(0) | + hn
R
Z
R
|fˆn − f ∗ Shn K(0) | ≤ E
R
(0)
|fn − f ∗
R
|f | |K(s) |. Hence,
E
Z
|fˆn − f ∗ Shn Kn | ≤ E
Z
|fn(0)
− f ∗ Shn K(0) | + 2 hn
Z
|f |
Z
|K(s) |.
Since, hn → 0 as n → ∞, the variation of fˆn is asymptotically bounded by the variation of the conventional
estimator fn using K(0) , i.e., E
R
(0)
|fn − f ∗ Shn K(0) |.
R
Under the assumptions that f has a variance, (1 + t2 )(K(0) (t))2 dt < ∞, (Devroye, 1987, Theorem 7.4)
showed that E
R
(0)
|fn − f ∗ Shn K(0) | = O((nhn )−1/2 ). Thus,
Z
p
nhn E |fˆn − f ∗ Shn Kn | = (1 + (nh3n )1/2 )O(1) = O(1),
where the last equality follows if nh3n ≤ c < ∞.
We now provide an analog for Theorem 7.1 in Devroye (1987). There, the bias order O(hs+a ) is achieved
for kernels with orders greater than s, while in the following theorems we obtain the same order of bias for
kernels of order s.
Theorem 2. Let {Kn } be a sequence of kernels of order s such that: 1. αs (Kn ) = O(han ); 2. sup βs+a (Kn ) <
n
∞, for some a ∈ (0, 1]. For f with absolutely continuous f
(s−1)
and f
(s)
∈ L1 assume that for some
0<c<∞
ω(δ) ≤ c |δ|a .
(2.11)
Then, kf ∗ Shn Kn − f k1 = O(hs+a
n ).
Proof. As in the proof of Theorem 1, we have
kf ∗ Shn Kn − f k1 ≤
hsn
s!
f (s)
1
|αs (Kn )| +
Z
1
0
5
Z
ω(λhn |t|)s(1 − λ)s−1 | ts Kn (t)| dtdλ
(2.12)
By (2.11)
hs
kf ∗ Shn Kn − f k1 ≤ n
s!
f
(s)
1
|αs (Kn )| +
chan
Z Z
0
1
s−1
s(1 − λ)
dλ t
s+a
Kn (t) dt .
Hence, under conditions 1. and 2. in the statement of the theorem we have kf ∗ Shn Kn − f k1 = O(hs+a
n ).
Remark 4. Practitioners may find condition (2.11) too general, preferring more primitive conditions on
f (s) . To this end, we say that a function g defined on R satisfies a global Lipschitz condition of order a ∈ (0, 1]
if there exist positive functions l(x), r(x) such that
|g(x − h) − g(x)| ≤ l(x)|h|a for |h| ≤ r (x) , x ∈ R.
(2.13)
The function l is called a Lipschitz constant and the function r is called a Lipschitz radius. The class
Lip(a, δ), for δ > 1, is defined as the set of functions g which satisfy (2.13) with l and r such that
Z
(l(x) + r(x)−δ )dx < ∞.
(2.14)
In the next lemma we give two sufficient sets of conditions for g ∈ Lip(a, δ). In the first case g is compactly
supported, and in the second it is not.
Lemma 1. a) Suppose g has compact support, supp g ⊆ [−N, N ] for some N > 0, and g satisfies the usual
Lipschitz condition |g(x − h) − g(x)| ≤ c|h|a for any x, h and some a ∈ (0, 1]. Set l(x) = c, r(x) = 1 for
|x| < N and l(x) = 0, r(x) = |x| − N for |x| ≥ N . Then, g ∈ Lip(a, δ) with any δ > 1.
b) Suppose that |g (1) (t)| ≤ ce−|t| , t ∈ R. Let l(x) = c exp(−|x|/2 + 1/2), r(x) = (1 + |x|)/2, x ∈ R. Then,
g ∈ Lip(1, δ) with any δ > 1.
Proof. a) If |x| < N , then |g(x − h) − g(x)| ≤ |h|a l(x) for all h (and not only for |h| ≤ r(x)). If |x| ≥ N , then
|h| ≤ r(x) = |x| − N implies |x − h| ≥ |x| − |h| ≥ N , so that |g(x − h) − g(x)| = 0 = |h|a l(x) for |h| ≤ r(x).
b) Let |x| ≥ 1. We have
|g(x − h) − g(x)| ≤ |h|
sup
|t−x|≤|h|
|g (1) (t)| ≤ |h|c
sup
e−|t| .
(2.15)
|t−x|≤|h|
|t − x| ≤ |h| ≤ r(x) = (1 + |x|)/2 implies |t| = |x + t − x| ≥ |x| − |t − x| ≥ |x|/2 − 1/2 ≥ 0 and (2.15) gives
|g(x − h) − g(x)| ≤ |h|c exp(1/2 − |x|/2) = |h|l(x) for |h| ≤ r(x). Now let |x| < 1. Then, e−|x|/2 ≥ e−1/2 so
6
that by (2.15), |g(x − h) − g(x)| ≤ |h|c ≤ |h|c exp(1/2 − |x|/2) = |h|l(x) for all h and not only for |h| ≤ r(x).
Condition (2.14) is obviously satisfied in both cases.
By part a) of Lemma 1, compactly supported densities with derivative f (s) that satisfies the usual aLipschitz condition are such that f (s) ∈ Lip(a, δ) for any δ > 1. This corresponds to the case treated in
Theorem 7.1 of Devroye (1987). Part b) shows that for densities with unbounded domains, not covered by
Theorem 7.1, if f (s) (x) has derivative that decays exponentially as |x| → ∞, then f (s) ∈ Lip(1, δ) for any
δ > 1. Next we provide a version of Theorem 2 for densities with derivative f (s) ∈ Lip(a, δ).
Theorem 3. Suppose that the density f is such that its derivative f (s) belongs to L1 , C (the respective
norms are finite) and Lip(a, δ). Let {Kn } be a sequence of kernels of order s such that: 1. αs (Kn ) = O(han );
2. sup max{βs+a (Kn ), βs+δ (Kn )} < ∞. Then, kf ∗ Shn Kn − f k1 = O(hs+a
n ).
n
Proof. As in the proof of Theorem 1, we have
hs
kf ∗Shn Kn −f k1 ≤ n
s!
f
(s)
1
|αs (Kn )| +
Z
1
0
Z Z
f
(s)
(x − hn λt) − f
(s)
s
s−1
(x) |t Kn (t)|dt dx s(1 − λ)
dλ .
f (s) (x − hn λt) − f (s) (x) |ts Kn (t)|dt and note that since f (s) ∈ Lip(a, δ)
Z
Z
(s)
(s)
s
+
I(x) =
f (x − hn λt) − f (x) |t Kn (t)|dt
Let I(x) =
R
λhn |t|≤r(x)
≤
Z
λhn |t|>r(x)
Z
l(x)λa han |t|a+s |Kn (t)|dt +
λhn |t|≤r(x)
Z
≤ han βs+a (Kn )l(x) +
|f (s) (x − hn λt) − f (s) (x)||ts Kn (t)|dt
λhn |t|>r(x)
|f (s) (x − hn λt) − f (s) (x)||ts Kn (t)|dt.
λhn |t|>r(x)
Letting I1 (x) =
R
|f (s) (x − hn λt) − f (s) (x)||ts Kn (t)|dt we have
λhn |t|>r(x)
I1 (x) ≤
Z
λhn |t|>r(x)
1 (s)
|f (x − hn λt)| + |f (s) (x)| |t|s+δ |Kn (t)|dt.
δ
|t|
Noting that |t|−δ < λδ hδn r(x)−δ and given that kf (s) kC < ∞, we obtain I1 (x) ≤ 2kf (s) kC
λδ hδn
β (Kn ).
r(x)δ s+δ
Consequently,
I(x) ≤ han max{βs+a (Kn ), βs+δ (Kn )} l(x) + 2hnδ−a kf (s) kC
7
1
r(x)δ
.
(2.16)
Since
R1
0
s(1 − λ)s−1 ds =
kf ∗Shn Kn −f k1 ≤
hsn
s!
1
s
and given (2.14)
f (s)
1
|αs (Kn )| + han max{βs+a (Kn ), βs+δ (Kn )}
s
1
Z
l(x) + 2kf (s) kC
1
r(x)δ
dx .
(2.17)
Thus, using conditions 1. and 2. in the statement of the theorem, we have kf ∗ Shn Kn − f k1 = O(hs+a
n ).
Remark 5. As in the case of Theorem 1, Theorems 2 and 3 do not address the construction of the kernel
sequence {Kn }. The following corollary to Theorem 3 shows that Kn = K(0) + han K(s) is a suitable kernel
sequence, where K(0) and K(s) are as defined above.
Corollary 2. Suppose the density f is such that its derivative f (s) belongs to L1 , C and to Lip(a, δ),
where a ∈ (0, 1], δ > 1. Let K(0) , K(s) satisfy (2.10) and belong to the intersection Bs+a ∩ Bs+δ . Put
Kn = K(0) +han K(s) , 0 ≤ hn ≤ 1. Then Kn is a kernel of order s for hn > 0 and kf ∗Shn Kn −f k1 = O(hs+a
n ).
The condition K(0) ∈ Bs+a and the definition Kn = K(0) + han K(s) can be replaced by K(0) ∈ Bs+1 and
Kn = K(0) + hn K(s) , respectively, without affecting the conclusion.
References
Devroye, L., 1987. A course in density estimation. Birkhäuser, Boston, MA.
Devroye, L., Györfi, L., 1985. Nonparametric density estimation: the L1 view. John Wiley and Sons, New
York, NY.
Mynbaev, K., Nadarajah, S., Withers, C., Aipenova, A., 2014. Improving bias in kernel density estimation.
Statistics and Probability Letters 94, 106–112.
Parzen, E., 1962. On estimation of a probability density and mode. Annals of Mathematical Statistics 33,
1065–1076.
Rosenblatt, M., 1956. Remarks on some nonparametric estimates of a density function. Annals of Mathematical Statistics 27, 832–837.
Silverman, B. W., 1986. Density estimation for statistics and data analysis. Chapman and Hall, London.
Tsybakov, A. B., 2009. Introduction to nonparametric estimation. Springer-Verlag, New York, NY.
Zhuk, V. V., Natanson, G. I., 2003. Seminorms and continuity modules of functions defined on a segment.
Journal of Mathematical Sciences 118, 4822–4851.
8
| 10math.ST
|
Ext and Tor on two-dimensional cyclic
quotient singularities
arXiv:1601.05673v2 [math.AG] 5 May 2016
Lars Kastner ∗
Freie Universität Berlin
Given two torus invariant Weil divisors D and D0 on a two-dimensional cyclic quotient
singularity X, the groups ExtiX (O(D), O(D0 )), i > 0, are naturally Z2 -graded. We
interpret these groups via certain combinatorial objects using methods from toric geometry.
In particular, it is enough to give a combinatorial description of the Ext1 -groups in
the polyhedra of global sections of the Weil divisors involved. Higher Exti -groups are
then reduced to the case of Ext1 via a quiver. We use this description to show that
Ext1X (O(D), O(K − D0 )) = Ext1X (O(D0 ), O(K − D)), where K denotes the canonical
0
divisor on X. Furthermore, we show that Exti+2
X (O(D), O(D )) is the Matlis dual of
X
0
Tori (O(D), O(D )).
1 Preliminaries
Let us start by recalling the basic definitions and notation. For toric geometry we will
follow [CLS11] closely.
Given two coprime positive integers 0 < q < n, we can define a 2-dimensional cyclic
quotient singularity X by taking the quotient of C2 by the action of the cyclic subgroup
of GL(2, C) generated by
ξn 0
,
0 ξnq
where ξn denotes an n-th root of unity. Any 2-dimensional cyclic quotient singularity
arises in this manner.
The equivalent definition in terms of toric geometry is as follows: Let N be a twodimensional lattice and let M = HomZ (N, Z) be the dual lattice. Now let NQ := N ⊗Z Q
and MQ be the associated Q-vector spaces. We identify N and M with Z2 by choosing
the usual scalar product as pairing.
Definition 1.1. Given two integers 0 < q < n such that gcd(q, n) = 1, we define the
cyclic quotient singularity X to be
X := Spec C[σ ∨ ∩ M ],
∗
The author is supported by the DFG (German research foundation) priority program SPP 1489.
1
where σ ∨ denotes the dual cone of
σ := cone ρ0 := (1, 0), ρ1 := (−q, n)
⊆ NQ .
By R we denote the coordinate ring C[σ ∨ ∩ M ] of X.
Let us introduce a running example.
Example 1.2. Let n = 7 and q = 3. Then the Hilbert basis of the dual cone σ ∨ has
four elements, indicated by the dots in the picture.
σ∨
Thus R = C[x[0,1] , x[1,1] , x[2,1] , x[7,3] ] or, if we label the axes x and y, R = C[y, xy, x2 y, x7 y 3 ].
Remark 1.3. There is a close relationship between the Hilbert basis of σ ∨ and the
n
continued fraction expansion of n−q
discovered by Riemenschneider ([Rie81]). This has
already lead to a very fruitful discussion of their deformation theory in terms of chains
representing zero ([Ste91; Chr91]) and so-called p-resolutions ([Alt98]). The connection
n
of the Ext functor with the continued fraction expansion of n−q
is part of [Kas15].
Any torus invariant Weil divisor is an integer linear combination of the orbits corresponding to the rays of σ, denoted as [ρ0 ] and [ρ1 ]. We write D = a0 [ρ0 ] + a1 [ρ1 ],
ai ∈ Z.
Notation. Throughout this article, we will omit the O, for example we will write
ExtiX (D, D0 ) for ExtiX (O(D), O(D0 )).
In order to study the modules ExtiX (D, D0 ) we may instead study the modules
ExtiR (Γ(X, D), Γ(X, D0 )), since X is affine. In the toric language the global sections of a
torus invariant Weil divisor are given by the section polyhedron.
Definition 1.4. For a torus invariant Weil divisor D = a0 [ρ0 ] + a1 [ρ1 ], the section
polyhedron is given by
PD := {u ∈ MQ | hu, ρi i ≥ −ai }.
The lattice points of the section polyhedron correspond to the homogeneous global
sections of O(D). We use this to describe H 0 (X, O(D)) as a divisorial ideal over R:
M
HD := H 0 (X, O(D)) =
C · χu ⊆ C[M ].
u∈PD ∩M
Keep in mind that D is not necessarily Cartier. Hence the summation D + D0 of divisors
does not translate to the multiplication of the ideals, and in general we have
HD+D0 ( HD · HD0 .
2
Because σ ∨ is simplicial, every PD will just be the Minkowski sum of σ ∨ and a rational
vector ν(D) ∈ Q2 . We write
PD = ν(D) + σ ∨ .
The (minimal) homogeneous generators of the R-module correspond exactly to the
lattice points on the compact edges of conv(PD ∩ M ). Denote these lattice points as
G(D) := {u ∈ M | u lies on a compact edge of conv(PD ∩ M )} ⊆ M.
Remark 1.5. If G(D) = {u0 , . . . , ur } we can sort these lattice points from left to right,
such that
hu0 , ρ0 i < hu1 , ρ0 i < . . . < hur , ρ0 i
and
hur , ρ1 i < hur−1 , ρ1 i < . . . < hu0 , ρ1 i.
Notation. Throughout this paper we will use the following shorthand notation: Let P
be a subset of MQ , then we denote by C{P } the R-module
M
C{P } :=
C · χu ,
u∈P ∩M
with the multiplication
w
x ·χ
u
(
χu+w
:=
0
u+w ∈P
.
else
Note that only with this multiplication C{P } becomes a M -graded R-module. Furthermore, this is not well-defined for every subset P of Q2 , rather one needs “convexity” of P
with respect to σ ∨ . Thus, we require for p ∈ P with u + p ∈ P for u ∈ σ ∨ that λu + p ∈ P
for 0 ≤ λ ≤ 1, which is actually a little stronger than needed.
As an example note that HD = C{PD }.
The class group of X is Z/nZ. By 0 = E 0 , . . . , E n−1 we denote the torus invariant
divisors E i := −i · [ρ0 ]. These divisors form a system of representatives for the class
group of X. Furthermore, let K = −[ρ0 ] − [ρ1 ] be the canonical divisor. One can now
calculate the vertex of the canonical divisor to be
q+1
].
ν(K) = [1,
n
Example 1.6. We will consider the divisors E 1 and E 3 . First we draw the corresponding
polyhedra of global sections.
[7, 3]
ν(E 3 )
ν(E 2 )
3
Then we compute the generators of HE 2 and HE 3 . They are depicted as dots of the
respective colors.
Our approach is to resolve such a divisorial ideal HD projectively in a Taylor-like
fashion ([Tay66]). Choosing a minimal generating set of syzygies, we arrive at a short
exact sequence, which may be seen as a slight generalization of cellular resolutions ([MS05,
Ch. 4]) to the coordinate rings of cyclic quotient singularities.
2 Resolving torus invariant divisors on CQS
A major obstacle, stemming from the non-regularity of R, is the infiniteness of the
complexes. It turns out that the syzygies of the generators of HD are isomorphic to
a direct sum of divisorial ideals, i.e. global sections of other torus invariant divisors.
Finiteness of the class group yields that we can encode all the information needed to
freely resolve any HD in a finite quiver Q with edges labelled by elements of M , effectively
overcoming the previously mentioned obstacle. This results in a recursive formula for
both Exti and Tori , i > 1, in terms of Ext1 and Tor1 , respectively.
Denote by {u0 , . . . , ur } the generators G(D) of D. Now take π to be the canonical
surjection
r
M
i
R[−ui ] HD , ei 7→ xu .
i=0
ui
ui+1
For every pair
and
of consecutive generators, we can build an injective map into
the kernel of π in the following way:
i−1
(xu
i
i
i−1
) ∩ (xu ) → ⊕ri=0 R[−ui ], xu 7→ xu−u ei − xu−u
i−1
ei−1 .
i
Denote by ι the map from the direct sum ⊕ri=1 (xu ) ∩ (xu ) of these ideals into
⊕ri=0 R[−ui ].
The following proposition may be seen as a generalization of [MS05, Prop 3.1], which
states that every monomial ideal in the polynomial ring with two variables can be resolved
freely in a short exact sequence, to the singular case.
Proposition 2.1. The sequence
0→
r
M
i−1
((xu
ι
i
) ∩ (xu )) ,→
i=1
r
M
π
R[−ui ] HD → 0
i=0
i−1
is exact. Furthermore, the fractional ideals (xu
Let us illustrate this proposition in the example.
4
i
) ∩ (xu ) are divisorial.
Example 2.2. In our example D = E 3 is an ideal of R, generated by x3 y 2 , x4 y 2 and
x7 y 3 . We obtain the following sequence
[7, 3]
ν(E 3 )
0 → [(x[3,2] ) ∩ (x[4,2] )]⊕[(x[4,2] ) ∩ (x[7,3] )] ,→ R[−[3, 2]]⊕R[−[4, 2]]⊕R[−[7, 3]] HE 3 → 0
One immediately recognizes the summands of the first term as divisorial ideals, i.e.
[(x[3,2] ) ∩ (x[4,2] )] ⊕ [(x[4,2] ) ∩ (x[7,3] )] = HE 1 [−[3,2]] ⊕ HE 3 [−[4,2]] .
We begin the proof of Proposition 2.1 by showing that all homogeneous elements of
ker π are in the image of ι.
Lemma 2.3. Let a ∈ ker π ⊆ ⊕ri=0 R[−ui ] be a homogeneous element of degree u ∈ M .
Then there are bi ∈ C, i = 1, . . . , r such that
a=
r
X
i
i−1
bi · (xu−u ei − xu−u
ei−1 ),
i=1
with u − ui ∈ σ ∨ whenever bi 6= 0. In particular, a ∈ im ι.
Proof. For the cases r = 1 there is nothing to prove. We prove the lemma explicitly for
the case r = 2. The methods used in that case are the same as for the induction step in
the general case, since we just show how to split off the last summand.
Because a is homogeneous, every entry of it is a multiple of a monomial. Hence, take
a to be
0
1
2
a = (a0 xu−u , a1 xu−u , a2 xu−u ).
Furthermore, assume a0 6= 0 6= a2 , otherwise we are done. This implies u − u0 ∈ σ ∨ and
u − u2 ∈ σ ∨ . Together with the inequalities of Remark 1.5 we get
0 ≤ hu − u0 , ρ0 i < hu − u1 , ρ0 i < hu − u2 , ρ0 i
0 ≤ hu − u2 , ρ1 i < hu − u1 , ρ1 i < hu − u0 , ρ1 i.
Therefore, u − u1 ∈ σ ∨ and we split a into the following
0
1
2
a = (a0 xu−u , a1 xu−u , a2 xu−u )
0
1
1
2
= (a0 xu−u , (a1 − a2 )xu−u , 0) + (0, a2 xu−u , −a2 xu−u ).
5
The second summand is clearly an element of both ker π and im ι. We conclude a1 − a2 =
−a0 and hence, we are done.
Proof of Proposition 2.1. It is already clear that π is surjective.
We already know that ι is injective on the direct summands. Hence, an homogeneous
element of degree u of the kernel of ι gives rise to an equation
r
X
i−1
ai (xu−u
i
ei−1 − xu−u ei ),
i=1
i−1
in R#G(D) , with ai ∈ C and ai = 0 for xu ∈
/ ((xu
i
) ∩ (xu )). Now we factor out
I = (xu − 1 | u ∈ σ ∨ ∩ M ),
and consider the equation over R/I ∼
= C. But the vectors ei−1 − ei ∈ C#G(D) are linearly
independent and hence, ai = 0 for i = 1, . . . , r. Factoring out I did not change the ai ,
thus, ι must be injective.
All modules are M -graded and the maps are homogeneous of degree 0. Therefore
ker π is a M -graded submodule of ⊕ri=0 R[−ui ]. In particular, ker π can be generated by
homogeneous elements and thus, Lemma 2.3 implies im ι = ker π.
i−1
i
The final claim is that ((xu ) ∩ (xu )) is divisorial. Take the polyhedron
(ui−1 + σ ∨ ) ∩ (ui + σ ∨ ).
i−1
i
Its lattice points correspond exactly to the monomials in ((xu ) ∩ (xu )). Furthermore
it is the polyhedron of the divisor −hui−1 , ρ0 i[ρ0 ] − hui , ρ1 i[ρ1 ].
Remark 2.4. An alternative proof of Proposition 2.1 can be given by using a modification
of the criterion for exactness of [BPS98, Lemma 2.2]. One just needs to replace the
expression of the least common multiple of two monomials xu and xv by the intersection
of the corresponding principal ideals (xu ) ∩ (xv ). Using this approach, the above sequence
is a subcomplex of the Taylor resolution for the divisorial ideal HD .
Applying Proposition 2.1 recursively, we can construct free resolutions of D up to any
desired length. Taking into account the finiteness of the class group of X, we can encode
the information of Proposition 2.1 for different divisors in a quiver.
For each E i , Proposition 2.1 gives a sequence
r(E i )
0→
M
j=1
r(E i )
kj
j
E [−v ] ,→
M
R[−uj ] E i → 0,
j=0
i
where {u0 , . . . , ur(E ) } = G(E i ) and v j ∈ M .
Let us introduce the labelled quiver Q: It consists of an ordinary quiver, i.e. a set of
vertices Q0 , a set of arrows Q1 and two functions s, t : Q1 → Q0 returning the source
and the target of an arrow. Additionally we have a function l : Q1 → M equipping each
arrow with a label.
6
Definition 2.5. Take Q to be the quiver with vertices Q0 := {E 0 , . . . , E n−1 }. For every
direct summand E kj [−v j ] in the above sequence we add an arrow a to Q1 such that
s(a) := E kj , t(a) := E i , l(a) := v j .
We do this for all such exact sequences for i = 0, . . . , n − 1.
Furthermore, for an arbitrary Weil divisor D we define the sources of the incoming
arrows in Q. This definition takes care of shifting all divisors in the right way.
Definition 2.6. Let E i be the divisor linearly equivalent to D, i.e. D = E i [−u] for
u ∈ M . Then we define
inQ D := {E j [−v j − u] | ∃ a ∈ Q1 with s(a) = E j , t(a) = E i , l(a) = v j }.
The sequence of Proposition 2.1 now becomes
M
0→
HG ,→ R#G(D) D → 0,
G∈inQ D
with the grading imposed by the generators of D on the middle module.
Example 2.7. In the running example with n = 7 and q = 3 the quiver Q looks as
follows:
[1, 1]
[4, 2]
E1
[3, 2]
E7
[6, 3]
E3
[4, 2]
E6
[6, 3]
[5, 3]
[2, 1]
E4
[2, 1]
E2
E5
One immediately recognizes the first term in the sequence for E 3 given in Example 2.2
via
inQ (E 3 ) = {E 3 [−[4, 2]], E 1 [−[3, 2]]}.
One can use this construction to define a free resolution of E 3 recursively. This resolution
will even be minimal. However, the quiver Q is the more elegant way of dealing with the
infiniteness of these resolutions.
For the case of X being Gorenstein we recover a well-known result by Eisenbud:
7
Remark 2.8. Building the quiver Q for the case q = n − 1 one notices that it consists of
b n−1
2 c disjoint cycles of length 2. For n being even, one of the cycles will have length
one, i.e. it is a loop. In particular, for every non-trivial divisor D the direct sum at the
beginning of the exact sequence of Proposition 2.1 has exactly one summand. This means
that the resolution of the divisorial ideal HD is 2-periodic, which was already observed
by Eisenbud in [Eis80] and is due to X being Gorenstein for this choice of n and q and
HD being MCM.
As a final remark, we can use the quiver Q to construct higher Exti and Tori recursively.
Remark 2.9. For Ext and Tor we obtain the following formulas using Proposition 2.1,
applying Hom(•, D0 ) and taking the long exact sequence of cohomology:
M
Extn+1 (D, D0 ) =
Extn (G, D0 )
G∈inQ D
and
M
Torn+1 (D, D0 ) =
Torn (G, D0 )
G∈inQ D
for n > 0.
Thus, we only need to know how to compute Ext1 and Tor1 .
3 Ext1
Applying Hom(•, D0 ) to the short exact sequence of Proposition 2.1 for D, we may
consider the long exact sequence of cohomology. The first part gives a formula for
Ext1 (D, D0 ). In this section we will rephrase this formula in combinatorial terms.
To a Weil divisor D define the following set
Definition 3.1.
[
E(D) := int(PD )\
u + int(σ ∨ ).
u∈G(D)
Example 3.2. We draw the sets E(E i ) for E 3 and E 2 in the running example with
n = 7 and q = 3:
Remember that the leftmost and the bottom boundary do not belong to the E sets, this
is indicated by the dashed lines. Since E 2 only has two generators, its E set looks like a
parallelepiped. On the other hand, E 3 is generated by three elements and hence, its E
set has one ’dent’.
8
Let us now establish the link of E(D) with Ext1 (D, D0 ).
Proposition 3.3. For two Weil divisors D and D0 define
ext(D, D0 ) := −(E(D) − ν(D0 )).
Then
Ext1 (D, D0 ) = C{ext(D, D0 )}.
We start by proving a lemma about the support of the M -graded module Hom(D, D0 ).
Lemma 3.4. Denote by
hom(D, D0 ) := − ν(D0 ) + PD0 .
Then
Hom(D, D0 ) = C{hom(D, D0 )}.
In particular, the module Hom(D, D0 ) is a divisorial ideal.
Proof. We know that Hom(D, D0 ) is M -graded and thus, generated by homogeneous
elements. Furthermore, we note that any non-zero map D → D0 must be injective,
implying that the image of any non-zero element of D completely determines the map.
Hence, non-zero homogeneous maps D → D0 are given as C-multiples of multiplication
of D by monomials xu ∈ C[M ] such that xu · D ⊆ D0 . This leaves us with determining
all the lattice points
{u ∈ M | PD + u ⊆ PD0 }.
These are exactly the lattice points of hom(D, D0 ).
Lemma 3.5. Denote by {u0 , . . . , ur } the generators G(D) of D. Furthermore write
(ui−1 + σ ∨ ) ∩ (ui + σ ∨ ) = v i + σ ∨ , v i ∈ MQ , i = 1, . . . , r
for the elements of inQ D. Then
0
ext(D, D ) =
r
[
−v i + ν(D0 ) + σ ∨ \ (−u0 + ν(D0 ) + σ ∨ ) ∪ (−ur + ν(D0 ) + σ ∨ ) .
i=1
Proof. Using the assumption we can write E(D) as
E(D) = int PD \
r
[
ui + int σ ∨ .
i=0
Furthermore we note that
(u0 − σ ∨ ) ∩ (ur − σ ∨ ) = ν(D) − σ ∨ .
Now use the equations giving the v i to obtain
E(D) =
r
[
v i − σ ∨ \ (u0 − σ ∨ ) ∪ (ur − σ ∨ ) .
i=1
Multiplying this with (−1) and adding ν(D0 ) on both sides yields the desired formula.
9
Proof of Proposition 3.3. Take the sequence of Proposition 2.1 and apply Hom(•, D0 ) to
it. Considering the first part of the long exact sequence of cohomology we have
0 → Hom(D, D0 ) → Hom(R#G(D) , D0 )
M
→ Hom(
G, D0 ) → Ext1 (D, D0 ) → 0.
G∈inQ D
Hence, we need to understand the quotient of Hom(⊕G∈inQ D G, D0 ) by the image of
Hom(R#G(D) , D0 ). Now we use Lemma 3.4:
M
M
M
Hom
G, D0 =
Hom(G, D0 ) =
C{hom(G, D0 )}.
G∈inQ D
G∈inQ D
G∈inQ D
Next we determine the image of Hom(R#G(D) , D0 ) in this direct sum. Take ui to be a
generator of D, not at the boundary, i.e. 0 6= i 6= r. Then Di and Di+1 are the only
summands of ⊕G∈inQ D G mapping non-trivially to R[−ui ] via the following map
Di ⊕ Di+1 →
R[−ui−1 ] ⊕ R[−ui ] ⊕ R[−ui+1 ],
.
i−1
i
i
i+1
(xu , xw ) 7→ (−xu−u , xu−u − xw−u , xw−u )
Applying Hom(•, D0 ) to this map gives
Hom(R[−ui ], D0 ) → Hom(Di , D0 ) ⊕ Hom(Di+1 , D0 )
.
i
xu−u
7→
(xu , −xu )
Hence, dividing by the image of Hom(R[−ui ], D0 ) means that the elements xu ∈ Hom(Di , D0 )
and xu ∈ Hom(Di+1 , D0 ) get identified. This explains why we take the union of the
−v i + σ ∨ .
Now consider u0 . Then the map D1 → R[−u0 ] ⊕ R[−u1 ] yields
Hom(R[−u0 ], D0 ) → Hom(D1 , D0 )
.
0
xu−u
7→
xu
Thus, all xu ∈ Hom(D1 , D0 ) such that u + u0 ∈ PD0 are set to zero. One proceeds
analogously for ur . This explains the part which is cut off and we are done.
Let us briefly remark on the MCM’ness and sMCM’ness as mentioned in the introduction.
Remark 3.6. We claim that Exti (D, K) = 0 for i > 0 and any D. By the recursion
formula it is enough to show that all Ext1 (D, K) = 0. Recall that the vertex of K is
ν(K) = [1, q+1
n ]. Now proceed to compute ext(D, K) = E(D) − ν(K) in two steps: First
subtract [1, nq ], second subtract [0, n1 ]. One can then show
q
(E(D) − [1, ]) ∩ M ⊆ E(D) ∩ M.
n
10
Now we argue that G(D) = E(D) ∩ M , otherwise G(D) would not generate HD . Hence,
the closure of E(D) − [1, nq ] can only have lattice points on its top and rightmost edges.
In particular, the lattice points of E(D) − [1, nq ] do not lie in the ‘valleys’ of E(D) − [1, nq ],
since they were at the ‘valleys’ of E(D). Subtracting [0, n1 ] from E(D) − [1, nq ] yields the
desired result.
Remark 3.7. Assume D to be non-trivial. For the simplest case of H0 = R in the second
argument we see
dimC Ext1R (D, R) = #G(D) − 2.
Thus, Ext1R (D, R) vanishes whenever HD is generated by exactly two elements, thereby
relating our construction to the result of Wunram ([Wun87]).
4 Tor1
Using the same strategy as for Ext, we can derive a combinatorial description of Tor1 as
well. While Ext1 depended on the polyhedra of global sections of the involved divisors
exclusively, Tor1 needs the quiver Q as an additional datum.
Definition 4.1.
[
T(D) := PD \
u + σ∨.
u∈G(D)
Proposition 4.2. For two Weil divisors D and D0 let
tor(D, D0 ) := T(D) + ν(D0 ).
Then
Tor1 (D, D0 ) =
M
M
W i ∈inQ D
C{tor(W ij , D0 )} .
W ij ∈inQ W i
Proof. For a Weil divisor W , denote by F0 (W ) the M -graded module RG(W ) in the
middle of the sequence of Proposition 2.1. Then we can build a free resolution of D as
follows:
M
M
M
d
d
F0 (W ij ) →2
F0 (W i ) →1 F0 (D) D → 0,
W i ∈inQ D
W ij ∈inQ W i
W i ∈inQ D
resulting from repeatedly applying Proposition 2.1. Tensorizing this sequence with D0
and then taking cohomology yields
Tor1 (D, D0 ) =
ker(d1 ⊗ idD0 )
.
im(d2 ⊗ idD0 )
11
Tensorizing R[−u] with D0 yields D0 [−u]. Inserting this, we see that the kernel of d1 ⊗idD0
is exactly
M
M
ker(d1 ⊗ idD0 ) =
C{PW ij + ν(D0 )} .
W i ∈inQ D
W ij ∈inQ W i
By construction it is enough to consider the single summands. In each of these we have to
remove the image of F0 (W ij )⊗D0 . Assume W ij to be generated by G(W ij ) = {u0 , . . . , ur }.
Then the image of F0 (W ij ) ⊗ D0 is exactly
r
X
k
xu D 0 .
k=0
Thus for the support we have
(PW ij + ν(D0 ))\
r
[
(uk + D0 ) = ν(D0 ) +
PW ij \
k=0
r
[
!
(uk + σ ∨ )
k=0
and this equals tor(W ij , D0 ), hence, finishing the proof.
5 The Matlis dual
Definition 5.1. We define the Matlis dual (H)∨ of an R-module H to be
(H)∨ := HomR (H, C[−σ ∨ ∩ M ]).
One can check that in our setting, C[−σ ∨ ∩ M ] is exactly the injective hull of C as an
R-module.
Remark 5.2. As stated in [MS05], this means that for a M -graded module H,
((H)∨ )−u = HomC (Hu , C).
The R-multiplication on (H)∨ is then the transpose of the R-multiplication on H. In
particular, this means that (C{P })∨ = C{−P }.
Proposition 5.3. The Matlis dual of Ext1 (D, K − D0 ) is
(Ext1 (D, K − D0 ))∨ = C{tor(D, D0 )}.
In order to prove this proposition, we will first have a closer look at the sets E(D)
and T(D). By construction they only differ at the boundaries. We will introduce a set
link(D), which is the common core of E(D) and T(D), meaning that it links these sets
when intersecting with the lattice M . This set and its behaviour under shifts by vertices
ν(D0 ) is the key to understanding Theorem 6.2, the Matlis duality of Ext and Tor.
12
Definition 5.4. For a Weil divisor D, define the link of D:
h
i
link(D) := E(D) ∩ E(D) + ν(K)] .
Example 5.5. Let us construct link(E 3 ). Again we have E(E 3 ) with the dashed lower
and left edge. Now we shift it by
4
ν(K) = [1, ]
7
and take the closure. The intersection link(E 3 ) is indicated in green.
Alternatively one can define link(D) using the T set. One just has to close the first set
in the intersection. This is the first statement in the following proposition.
Proposition 5.6. Given D and D0 two Weil divisors on X, we have
1. link(D) = T(D) ∩ [T(D) + ν(K)];
2. [E(D) + ν(D0 )] ∩ M = [link(D) + ν(D0 )] ∩ M ; and
3. [T(D) + ν(D0 )] ∩ M = [link(D) + ν(D0 ) − ν(K)] ∩ M .
Proof. It is already clear that the closures of E and T are equal. The complementary
boundary containments then yield the first formula.
The proofs of the second and third claim are very similar, hence, we will only prove
the second claim.
The containment of link(D) + ν(D0 ) in E(D) + ν(D0 ) is trivial.
By construction, E(D) is contained in PD , even in the interior. Let
PD = {u ∈ MQ | hu, ρ0 i ≥ aD , hu, ρ0 i ≥ bD },
with aD , bD ∈ Z, i.e. D = −aD · [ρ0 ] − bD · [ρ1 ]. Similarly, let D0 = −aD0 · [ρ0 ] − bD0 · [ρ1 ]
for some aD0 , bD0 ∈ Z. Hence
hu, ρ0 i ≥ aD + aD0 ,
E(D) + ν(D0 ) ⊆ PD + ν(D0 ) = u ∈ MQ |
.
hu, ρ1 i ≥ bD + bD0
Now we want lattice points in the interior of the right hand side. Taking u ∈ int(PD +
ν(D0 )), u evaluates to an integer with both ρ0 and ρ1 . Since it is not on the boundary of
PD + ν(D0 ), we have
hu, ρ0 i ≥ aD + aD0 + 1, and hu, ρ1 i ≥ bD + bD0 + 1.
13
This corresponds exactly to adding the vertex of the canonical divisor K = −[ρ0 ] − [ρ1 ].
We obtain
[E(D) + ν(D0 )] ∩ M ⊆ PD + ν(D0 ) + ν(K).
Intersecting the right hand side with E(D) + ν(D0 ) preserves this relation. As a final
step, we note that we can replace PD by E(D). One then recognizes this intersection as
link(D) + ν(D0 ).
Proof of Proposition 5.3. Inserting the formula for tor and ext in terms of E(D) and
T(D), we use the formula of Proposition 5.6 to obtain:
tor(D, D0 ) ∩ M = [T(D) + ν(D0 )] ∩ M
= [link(D) + ν(D0 ) − ν(K)] ∩ M
= [E(D) − ν(K − D0 )] ∩ M
= − ext(D, K − D0 ) ∩ M.
Inserting Remark 5.2 finishes the proof.
6 Main theorems
Theorem 6.1. Let D and D0 be two Weil divisors on a cyclic quotient singularity. Then
Ext1 (D, K − D0 ) = Ext1 (D0 , K − D).
Proof. We want to prove the equality of the lattice points of the respective supports, i.e.
ext(D, K − D0 ) ∩ M = ext(D0 , K − D) ∩ M.
The trick is to show that there are no lattice points in the respective complements. Since
this is symmetric we will just show
ext(D0 , K − D)\ ext(D, K − D0 ) ∩ M = ∅.
We reverse the sign and insert the definition of E(D) to get
− ext(D, K − D0 ) = ν(−K + D0 ) + E(D)
= (ν(−K + D + D0 ) + int(σ ∨ ))\
r
[
(ν(D0 − K) + ui + int σ ∨ ),
i=0
where {u0 , . . . , ur } = G(D). Both − ext(D, K − D0 ) and − ext(D0 , K − D) result from
ν(−K + D + D0 ) + int(σ ∨ ) by cutting off certain pieces at the top. Inserting the equation
for − ext(D, K − D0 ), we can rephrase this as an intersection, i.e.
− ext(D0 , K − D)\ − ext(D, K − D0 )
= − ext(D0 , K − D) ∩
r
[
(ν(D0 ) − ν(K) + ui + int σ ∨ ).
i=0
14
The set − ext(D0 , K − D) is a shift of E(D0 ). Thus, we can replace ν(D0 ) + ui + int σ ∨
by ui + E(D0 ) in the above intersection. Hence,
0
0
− ext(D , K − D)\ − ext(D, K − D ) ⊆
=
r
[
(− ν(K) + ui + E(D0 ))
i=0
r
[
− ext(D0 [−ui ], K).
i=0
Since D0 is Q-Cartier and hence, MCM, we know that Ext(D0 , K) = 0, which stays true
under M -shifts of D0 . Thus, −(ext(D0 [−ui ], K) ∩ M ) = ∅ for all i = 0, . . . , r, and we are
done.
Finally, we show the duality of Ext and Tor.
Theorem 6.2. Let D and D0 be two Weil divisors on a cyclic quotient singularity. Then
Exti+2 (D, K − D0 ) = (Tori (D, D0 ))∨
for i > 0.
Proof. Using the recursion of Remark 2.9 on both sides this is a consequence of Lemma 6.3
below.
Lemma 6.3. Given two Weil divisors D and D0 , we have the following equality of
R-modules:
Ext3 (D, K − D0 ) = (Tor1 (D, D0 ))∨ .
Proof. First, we use the description of Tor1 developed in Proposition 4.2. Then we insert
the formula for the Matlis dual of Ext1 of Proposition 5.3.
M
M
Tor1 (D, D0 ) =
C{tor(W ij , D0 )}
W i ∈inQ D
W ij ∈inQ W i
∼
=
M
M
W i ∈inQ D
(Ext1 (W ij , K − D0 ))∨ .
W ij ∈inQ W i
Applying the recursion formula of Remark 2.9 for Ext3 , we obtain the desired result.
Note that although Theorem 6.2 shows the symmetry of Theorem 6.1 for i > 2, it does
not generalize Theorem 6.1, as it does not imply the i = 1 case. In particular, the case
i = 2 remains open.
15
References
[Alt98]
Klaus Altmann. “P-resolutions of cyclic quotients from the toric viewpoint.”
English. In: Singularities. The Brieskorn anniversary volume. Proceedings of
the conference dedicated to Egbert Brieskorn on his 60th birthday, Oberwolfach,
Germany, July 1996. Basel: Birkhäuser, 1998, pp. 241–250.
[BPS98]
Dave Bayer, Irena Peeva and Bernd Sturmfels. “Monomial resolutions.” English. In: Math. Res. Lett. 5.1-2 (1998), pp. 31–46.
[Chr91]
Jan Arthur Christophersen. “On the components and discriminant of the
versal base space of cyclic quotient singularities.” English. In: Symmetric
Lagrangian singularities and Gauss maps of theta divisors. 1991, pp. 81–92.
[CLS11]
David A. Cox, John B. Little and Henry K. Schenck. Toric varieties. English.
Providence, RI: American Mathematical Society (AMS), 2011, pp. xxiv + 841.
[Eis80]
David Eisenbud. “Homological algebra on a complete intersection, with an
application to group representations”. In: Transactions of the American
Mathematical Society 260.1 (1980), pp. 35–64.
[Kas15]
Lars Kastner. “Ext on affine toric varieties”. available at http://www.diss.
fu-berlin.de/diss/receive/FUDISS_thesis_000000101520. PhD thesis.
Freie Universität Berlin, 2015.
[MS05]
Ezra Miller and Bernd Sturmfels. Combinatorial commutative algebra. English.
New York, NY: Springer, 2005, pp. xiv + 417.
[Rie81]
Oswald Riemenschneider. “Zweidimensionale Quotientensingularitaeten: Gleichungen und Syzygien.” German. In: Arch. Math. 37 (1981), pp. 406–417.
[Ste91]
Jan Stevens. “On the versal deformation of cyclic quotient singularities.”
English. In: Symmetric Lagrangian singularities and Gauss maps of theta
divisors. 1991, pp. 302–319.
[Tay66]
Diana Kahn Taylor. “Ideals generated by monomials in an R-sequence”. PhD
thesis. University of Chicago, Department of Mathematics, 1966.
[Wun87] Jürgen Wunram. “Reflexive modules on cyclic quotient surface singularities”.
In: Singularities, representation of algebras, and vector bundles. Springer,
1987, pp. 221–231.
16
| 0math.AC
|
Seeding the Initial Population of Multi-Objective
Evolutionary Algorithms: A Computational Study
Tobias Friedricha,, Markus Wagner*b
arXiv:1412.0307v1 [cs.NE] 30 Nov 2014
a
Friedrich-Schiller-Universität Jena, Germany
b
University of Adelaide, Australia
Abstract
Most experimental studies initialize the population of evolutionary algorithms with random genotypes. In practice, however, optimizers are typically
seeded with good candidate solutions either previously known or created according to some problem-specific method. This seeding has been studied
extensively for single-objective problems. For multi-objective problems, however, very little literature is available on the approaches to seeding and their
individual benefits and disadvantages. In this article, we are trying to narrow
this gap via a comprehensive computational study on common real-valued
test functions. We investigate the effect of two seeding techniques for five
algorithms on 48 optimization problems with 2, 3, 4, 6, and 8 objectives. We
observe that some functions (e.g., DTLZ4 and the LZ family) benefit significantly from seeding, while others (e.g., WFG) profit less. The advantage of
seeding also depends on the examined algorithm.
Keywords: Multi-objective optimization, approximation, comparative
study, limited evaluations
1. Introduction
In many real-world applications trade-offs between conflicting objectives
play a crucial role. As an example, consider route planning, where one objective might be travel time and another fuel consumption. For such problems,
we need specialized optimizers that determine the Pareto front of mutually
∗
Corresponding author. Phone: +61 8 8313 5405, fax: +61 8 8313 4366,
email: [email protected].
Preprint submitted to Elsevier
December 2, 2014
non-dominated solutions. There are several established evolutionary multiobjective evolutionary algorithms (MOEA) and many comparisons on various
test functions. However, most of them start with random initial solutions.
In practice, however, a good initial seeding can make problem solving approaches competitive that would otherwise be inferior.
If prior knowledge exists or can be generated at a low computational cost,
good initial estimates may generate better solutions with faster convergence.
For single-objective evolutionary algorithms, methods such as seeding have
been studied for about two decades; see, e.g., [17, 20, 23, 29, 39] for studies and examples (see [26] for a recent categorization). For example, the
effects of seeding for the Traveling Salesman Problem (TSP) and the jobshop scheduling problem (JSSP) were investigated in [31]. The algorithms
were seeded with known good solutions in the initial population, and it was
found that the results were significantly improved on the TSP but not on the
JSSP. To investigate the influence of seeding on the optimisation, a varying
percentage of seeding was used, ranging from 25 to 75%. Interestingly, it
was also pointed out that a 100% seed is not necessarily very successful on
either problems [27]. This is one of the very few reported that seeding can in
some cases beneficial to the optimisation process, but not necessarily always
is. In [21] a seeding technique for dynamic environments was investigated.
There, the population was seeded when a change in the objective landscape
arrived, aiming at a faster convergence to the new global optimum. Again,
some of the investigated seeding approaches were more successful than others.
One of the very few studies that can be found on seeding techniques
for MOEAs is the one performed by Hernandez-Diaz et al. [22]. There,
seeds were created using gradient-based information. These were then fed
into NSGA-II [10] and the quality was assessed on the benchmark family
ZDT [42]. The results indicate that the proposed approach can produce a
significant reduction in the computational cost of the approach.
In general, seeding is not well documented for multi-objective problems,
even for real-world problems. If seeding is done, then typically the approach is outlined and used with the comment that it worked in “preliminary
experiments”—the reader is left in the dark on the design process behind the
used seeding approach. This is quite striking as one expects that humans can
construct a few solutions by hand, even if they do not represent the ranges of
the objectives well. The least that one should be able to do is to reuse existing designs, and to modify these iteratively towards extremes. Nevertheless,
even this manual seeding is rarely reported.
2
In this paper, we are going to investigate the effects of two structurally
different seeding techniques for five algorithms on 48 multi-objective optimization (MOO) problems.
Seeding
As seeding we use the weighted-sum method, where the trade-off preferences are specified by non-negative weights for each objective. Solutions
to these weighted-sums of objectives can be found with an arbitrary classical single-objective evolutionary algorithm. In our experiments we use the
CMA-ES [18]. Details of the two studied weighting schemes are presented in
Section 2.1.
Quality measure
There are different ways to measure the quality of the solutions. A recently very popular measure is the hypervolume indicator, which measures
the volume of the objective space dominated by the set of solutions relative
to a reference point [41]. Its disadvantage is its high computational complexity [3, 4] and the arbitrary choice of the reference point. We instead
consider the mathematically well founded approximation constant. In fact, it
is known that the worst-case approximation obtained by optimal hypervolume distributions is asymptotically equivalent to the best worst-case additive
approximation constant achievable by all sets of the same size [6]. For a rigorous definition, see Section 2. This notion of multi-objective approximation
was introduced by several authors [15, 19, 30, 34, 35] in the 80’s and its
theoretical properties have been extensively studied [9, 12, 32, 33, 36].
Algorithms
We use the jMetal framework [13] and its implementation of NSGA-II [10],
SPEA2 [43], SMS-EMOA [14], and IBEA [40]. Additionally to these more
classical MOEAs, we also study AGE [7], which aims at directly minimizing
the approximation constant and has shown to perform very well for larger
dimensions [37, 38]. For each of these algorithms we compare their regular
behavior after a certain number of iterations with their performance when
initialized with a certain seeding.
Benchmark families
We compare the aforementioned algorithms on four common families
of benchmark functions. These are DTLZ [11], LZ09 [28], WFG [24],
3
and ZDT [42]. While the last three families only contain two- and threedimensional problems, DTLZ can be scaled to an arbitrary number of dimensions.
2. Preliminaries
We consider minimization problems with d objective functions, where
d ≥ 2 holds. Each objective function fi : S 7→ R, 1 ≤ i ≤ d, maps from
the considered search space S into the real values. In order to simplify the
presentation we only work with the dominance relation on the objective space
and mention that this relation transfers to the corresponding elements of S.
For two points x = (x1 , . . . , xd ) and y = (y1 , . . . , yd ), with x, y ∈ Rd we
define the following dominance relation:
x y :⇔ xi ≤ yi for all 1 ≤ i ≤ d,
x ≺ y :⇔ x y and x 6= y.
We assess the seeding schemes and algorithms by their achieved additive
approximation of the (known) Pareto front. We use the following definition.
Definition 1. For finite sets S, T ⊂ Rd , the additive approximation of T
with respect to S is defined as
α(S, T ) := max min max (si − ti ).
s∈S
t∈T 1≤i≤d
We measure the approximation constant with respect to the known Pareto
front of the test functions. The better an algorithm approximates a Pareto
front, the smaller the additive approximation value is. Perfect approximation
is achieved if the additive approximation constant becomes 0. However, the
approximation constant achievable for a (finite) population with respect to a
continuous Pareto front (consisting of an infinite number of points) is always
strictly larger than 0. It depends on the fitness function what is the smallest
possible approximation constant achievable with a population of bounded
size.
2.1. Seeding
For the task of computing the seeds, we employ an evolutionary strategy
(ES), because it “self-adapts” the extent to which it perturbs decision variables when generating new solutions based on previous ones. The Covariance
4
Matrix Adaptation based evolutionary strategy (CMA-ES) [18] self-adapts
the covariance matrix of a multivariate normal distribution. This normal
distribution is then used to sample from the multidimensional search space
where each variate is a search variable. The co-variance matrix allows the
algorithm to respect the correlations between the variables making it a powerful evolutionary search algorithm.
P
To compute a seed, a (2,4)-CMA-ES minimizes di=1 ai fi (x), where the
fi (x) are the objective values of the solution x. In preliminary testing, we noticed that larger population values for CMA-ES tended to result in seeds with
better objective values. This came at the cost of significantly increased evaluation budgets, as the learning of the correlations takes longer. Our choice
does not necessarily represent the optimal choice across all 48 benchmark
functions, however, it is our take on striking a balance between (1) investing evaluations in the seeding and (2) investing evaluations in the regular
multi-objective optimization. Note that large computational budgets for the
seeding have the potential to put the unseeded approaches at a disadvantage,
if the final performance assessment is not done carefully.
The number of seeds, the coefficients used, and the budget of evaluations is determined by the seeding approaches, which we will describe in the
following.
CornersAndCentre: A total of 10,000 evaluations is equally distributed over the generation of d + 1 seeds. The rest of the population
is generated randomly. For the i-th seed, 1 ≤ i ≤ d, the coefficients aj
(1 ≤ j ≤ d) are set in the following way:
(
10 if i = j,
ai =
1 otherwise.
Thus, we prevent the seeding mechanism from treating the optimization problem in a purely single-objective way by entirely neglecting any trade-off relationships between the objectives.1 Lastly, the (d + 1)-th weight vector uses
equal weights of 1 per objective. This way, we aim at getting a seed that is
relatively central with respect to the others.
LinearCombinations: Here a total of 100 seeds is generated, where
each seed is the result of running CMA-ES for 1,000 evaluations. The coeffi1
If the ranges of the objective values differ significantly, then the coefficients should be
adjusted accordingly.
5
cients of the linear combinations are integer values and we construct them in
the following way. First, we consider all “permutations” of coefficients with
ai = 1 for one coefficient and aj6=i = 0 for all others. Then, we consider all
permutations where two coefficients have the value 1, then those where three
coefficients have the value 1, and so on. When all such permutations that
are based on {0, 1} are considered, we consider all permutations based on
{0, 1, 2}, then based on {0, 1, 3}, then based on {0, 2, 3}, and so on.
Consequently, we achieve a better distribution of points in the objective
space. This comes, however, at the increased initial computational cost.
Furthermore, the budget per seed is lower than in the CornersAndCentre
approach, which typically results in less optimized seeds.
NoSeed: All solutions of the initial population are generated randomly.
This is the approach that is typically used for the generation of the initial
population.
2.2. Evolutionary Multi-Objective Optimization Algorithms
In the following, we outline the five optimization algorithms for which we
will later-on investigate the benefits of seeding the initial populations.
Many approaches try to produce good approximations of the true Pareto
front by incorporating different preferences. For example, the environmental
selection in NSGA-II [10] first ranks the individuals using non-dominated
sorting. Then, in order to distinguish individuals with the same rank, the
crowding distance metric is used, which prefers individuals from less crowded
sections of the objective space. The metric value for each solution is computed by adding the edge lengths of the cuboids in which the solutions reside,
bounded by the nearest neighbors.
SPEA2 [43] works similarly. The raw fitness of the individuals according
to Pareto dominance relations between them is calculated, and then a density
measure to break the ties is used. The individuals that reside close together
in the objective space are less likely to enter the archive of best solutions.
In contrast to these two algorithms, IBEA [40] is a general framework,
which uses no explicit diversity preserving mechanism. The fitness of individuals is determined solely based on the value of a predefined indicator.
Typically, implementations of IBEA come with the epsilon indicator or the
hypervolume indicator, where the latter measures the volume of the dominated portion of the objective space.
SMS-EMOA [14] is a frequently used IBEA, which uses the hypervolume
indicator directly in the search process. It is a steady-state algorithm that
6
uses non-dominated sorting as a ranking criterion, and the hypervolume as
the selection criterion to discard the individual that contributes the least hypervolume to the worst-ranked front. While SMS-EMOA often outperforms
its competition, its runtime unfortunately increases exponentially with the
number of objectives. Nevertheless, with the use of fast approximation algorithms (e.g., [2, 5, 25]), this algorithm can be applied to solve problems with
many objectives as well.
Recently, approximation-guided evolution (AGE) [7] has been introduced,
which allows to incorporate a formal notion (such as Definition 1) of approximation into a multi-objective algorithm. This approach is motivated by
studies in theoretical computer science studying multiplicative and additive
approximations for given multi-objective optimization problems [8, 9, 12, 36].
As the algorithm cannot have complete knowledge about the true Pareto
front, it uses the best knowledge obtained so far during the optimization process. It stores an archive A consisting of the non-dominated objectives vectors
found so far. Its aim is to minimize the additive approximation α(A, P ) of
the population P with respect to the archive A. The experimental results
presented in [7] show that given a fixed time budget it outperforms current
state-of-the-art algorithms in terms of the desired additive approximation,
as well as the covered hypervolume on standard benchmark functions.
3. Experimental Setup
We use the jMetal framework [13], and our code for the seeding as well
all used seeds are available online2 . As test problems we used the benchmark families DTLZ [11], ZDT [42], LZ09 [28], and WFG [24], We used the
functions DTLZ 1-4, each with 30 function variables and with d ∈ {2, 4, 6, 8}
objective values/dimensions.
In order to investigate the benefits of seeding even in the long run, we limit
the calculations of the algorithms to a maximum of 106 fitness evaluations
and a maximum computation time of four hours per run. Note that the
time restriction had to be used as the runtime of some algorithms increases
exponentially with respect to the size of the objective space.
AGE uses random parent selection; in all other algorithms parents are
selected via a binary tournament. As variation operators, the polynomial
2
http://cs.adelaide.edu.au/~markus/publications.html
7
mutation and the simulated binary crossover [1] were applied, which are both
used widely in MOEAs [10, 16, 43]. The distribution parameters associated
with the operators were ηm = 20.0 and ηc = 20.0. The crossover operator
is biased towards the creation of offspring that are close to the parents,
and was applied with pc = 0.9. The mutation operator has a specialized
explorative effect for MOO problems, and was applied with pm = 1/(number
of decision variables). Population size was set to µ = 100 and λ = 100.
Each setup was repeated 100 times. Note that these parameter settings are
the default settings in the jMetal framework, and they can often be found
in the literature, which makes a cross-comparison easier. To the best of our
knowledge, this parameter setting does not favor any particular algorithm
or put one at a disadvantage, even though individual algorithms can have
differing optimal settings for individual problems.
In a real-world scenario, if an algorithm is run several times (e.g. because
of restarts), the seeding might be only calculated once. In this case, it might
make sense to compare the unseeded and seeded variant of an algorithm with
the same number of fitness evaluations. However, we observed the expected
outcome that in this case seeding is almost always beneficial. We therefore
consider a more difficult scenario where the optimization is only run once and
the number of fitness function evaluations used for the seeding is deduced
from the number of fitness evaluations available for the MOEA.
As pointed out earlier, we assess the seeding schemes and algorithms using
the additive approximation of the Pareto front. However, as it is difficult to
compute the exact achieved approximation constant of a known Pareto front,
we approximate it. For the quality assessment on the LZ, WFG and ZDT
functions, we compute the achieved additive approximations with respect to
the Pareto fronts given in the jMetal package. For the DTLZ functions, we
draw one million points of the front uniformly at random. and then compute
the additive approximation achieved for this set.
We also measure the hypervolume for all experiments. As the behaviors of
the five algorithms differ significantly, there is no single reference point that
allows for a meaningful comparison of all functions. However, we observe
the same qualitative comparison with the hypervolume as we do with the
additive approximation. Therefore, we omit all hypervolume values in this
paper, because the additive approximation constant gives a much better way
to compare the results for these benchmark functions, where the Pareto fronts
are known in advance.
In addition to calculating the average ratio of the achieved approximation
8
constant with and without seeding, we also perform a non-parametric test
on the significance of the observed behavior. For this, we compare the final
approximation of the 100 runs without seeding and the 100 runs with seeding
using the Wilcoxon-Mann-Whitney two-sample rank-sum test at the 95%
confidence level.
4. Experimental Results
Our results are summarized in Tables 1 and 2. They compare the approximation constant achieved with CornersAndCentre seeding (Table 1) and
LinearCombinations seeding (Table 2) with the same number of iterations
without seeding. As the seeding itself requires a number of fitness function
evaluations (104 for CornersAndCentre and 105 for LinearCombinations), we allocate the seeded algorithms fewer fitness function evaluations.
This makes it harder for the seeded algorithms to outperform its unseeded
variant, as discussed above.
Figures 1 and 2 show some representative charts how the approximation
constant behaves over the runtime of the algorithms. First note that the
approximation constant is mostly monotonically decreasing. As a smaller
approximation constant corresponds to a better approximation of the Pareto
front, this means that most algorithms achieve a better approximation over
time. Exceptions are SPEA2 ( ), which is unable to handle the six dimensional variants of DTLZ, and NSGA-II ( ), which sometimes gets worse
after a certain time. For most problems and algorithms, the total maximal number of fitness function evaluations (106 ) was enough such that the
algorithms have converged.
The small black circles in Figures 1 and 2 indicate the average approximation constant of the initial seeding after the number of fitness function
evaluations needed to calculate it. Note that because of their specific selection
schemes, some algorithms like SPEA2 ( ) and NSGA-II ( ) sometimes increase the approximation constant after the initial seeding (e.g. on LZ09 F1).
Another surprising effect can be observed on LZ09 F2. There, seeding is disadvantageous to NSGA-II ( ), while it is advantageous to SPEA2 ( ).
For all considered test problems (not only the ones shown in Figures 1
and 2), either AGE ( ) or SMS-EMOA ( ) reach the best approximation
constant. However, for test problems with more than two or three dimensions
(cf. Figure 1), SMS-EMOA ( ) fails due to the high computational cost of
calculating the hypervolume. On some problems AGE ( ) does not finish all
9
iterations within 106 steps, but still achieves the best approximation constant
(e.g. DTLZ1 6D and DTLZ4 6D).
For all algorithms, both seedings are beneficial on some test functions.
However, the generally more performant algorithms (AGE and SMS-EMOA)
typically gain the most from both seedings. On some functions, these algorithms not only achieve a better approximation faster with seeding, but
it seems that all best approximations can only be achieved with seeding
(e.g. DTLZ4). On DTLZ4 2D the gap between the approximation constant
achieved by SMS-EMOA ( ) with and without seeding is about two orders
of magnitude, which is the difference between a very good approximation of
the Pareto front and basically no approximation of the Pareto front.
Tables 1 and 2 give a numerical comparison assuming that about 10%
of the fitness evaluations are used for seeding. The shown numbers are the
ratios of the median approximation constant without seeding to the median
approximation constant with seeding. Values >1.00 indicate where seeding is
beneficial in the median. We additionally show statistically significance based
on the Wilcoxon-Mann-Whitney two-sample rank-sum test at the 95% confidence level. “>” marks statistically significant improvements, “<” marks
statistically significant worsenings, “=” marks statistically insignificant findings. The ratios in Table 1 correspond to the approximation constants after
105 function evaluations in the left column of Figures 1 and 2. The ratios in
Table 2 correspond to the approximation constants after 106 function evaluations in the right column of Figures 1 and 2.
Counting only the statistically significant results over all functions and
seedings, the tables show that the majority profits from the seeding. The algorithms which benefit the most are SPEA2 ( ) with 20ד>” and 11ד<”,
AGE ( ) with 28ד>” and 21ד<”, and IBEA ( ) with 22ד>” and
17ד<”. There are significant differences depending on the test function.
The LZ09 benchmark family profits the most: Summing up the significant
results for all algorithms, there are 55ד>” and 18ד<”. Also for DTLZ4
there are 23ד>” and 8ד<”. The worst performance of the seeding is
achieved on the rather difficult WFG functions: While the CornersAndCentre seeding achieves over all algorithms 10ד>” and 5ד<’, the LinearCombinations seeding only achieves 10ד>” and 28ד<”.
We can do a similar analysis to assess the benefits of the investigated
seeding approaches. We observe that over all algorithms the CornersAndCentre seeding yields in total 51ד>” and 36ד<’, which is a bit better
than the LinearCombinations seeding which yields in total 62ד>” and
10
55ד<”. In order to answer the question whether this is statistically significant, we calculate the average rank of with and without seeding for each of
the 100 runs, each of the 48 functions, and each of the 5 algorithms. With this
combined data from all runs, functions and algorithms, the Wilcoxon-MannWhitney two-sample rank-sum test shows significance at the 95% confidence
level that both seedings improve upon no seeding.
5. Conclusions
Seeding can result in a significant reduction of the computational cost and
the number of fitness function evaluations needed. We observe that there is
an advantage on many common real-valued fitness functions even if computing an initial seeding reduces the number of fitness function evaluations
available for the MOEA. For some functions we observe a dramatic improvement in quality and needed runtime (e.g., DTLZ4 and the LZ09 family).
For practitioners, our results show that it can be worthwhile to apply
some form of seeding (especially when evaluations are expensive), but also
to investigate different MOEAs as well, as they have proven to benefit differently from seeding. While we observed that seeding can be very beneficial,
our experiments could not reveal why this is the case for a particular combination of seeding, algorithm, and function landscape. To answer this, many
parts have to be studied: the mappings that the benchmark functions create
from the search spaces into the objective spaces, the connectedness between
different local Pareto fronts, the adequacy of using CMA-ES in the seeding
procedure, and much more. As a next step towards this goal, we propose to
investigate seeding for combinatorial optimization problems.
Acknowledgements
The research leading to these results has received funding from the Australian Research Council (ARC) under grant agreement DP140103400 and
from the European Union Seventh Framework Programme (FP7/2007-2013)
under grant agreement no 618091 (SAGE).
11
References
[1] R. B. Agrawal and K. Deb. Simulated binary crossover for continuous
search space. Technical report, 1994.
[2] J. Bader, K. Deb, and E. Zitzler. Faster hypervolume-based search
using Monte Carlo sampling. In Multiple Criteria Decision Making for
Sustainable Energy and Transportation Systems (MCDM ’10), Vol. 636
of Lecture Notes in Economics and Mathematical Systems, pp. 313–326.
Springer, 2010.
[3] K. Bringmann and T. Friedrich. Parameterized average-case complexity
of the hypervolume indicator. In 15th Annual Conference on Genetic
and Evolutionary Computation Conference (GECCO ’13), pp. 575–582.
ACM Press, 2013.
[4] K. Bringmann and T. Friedrich. Approximating the volume of unions
and intersections of high-dimensional geometric objects. Computational
Geometry: Theory and Applications, 43:601–610, 2010.
[5] K. Bringmann and T. Friedrich. Approximating the least hypervolume
contributor: NP-hard in general, but fast in practice. Theoretical Computer Science, 425:104–116, 2012.
[6] K. Bringmann and T. Friedrich. Approximation quality of the hypervolume indicator. Artificial Intelligence, 195:265–290, 2013.
[7] K. Bringmann, T. Friedrich, F. Neumann, and M. Wagner.
Approximation-guided evolutionary multi-objective optimization. In
Proc. 22nd International Joint Conference on Artificial Intelligence (IJCAI ’11), pp. 1198–1203, Barcelona, Spain, 2011. IJCAI/AAAI.
[8] T. C. E. Cheng, A. Janiak, and M. Y. Kovalyov. Bicriterion single
machine scheduling with resource dependent processing times. SIAM J.
on Optimization, 8:617–630, 1998.
[9] C. Daskalakis, I. Diakonikolas, and M. Yannakakis. How good is the
Chord algorithm? In 21st Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA ’10), pp. 978–991, 2010.
12
[10] K. Deb, A. Pratap, S. Agrawal, and T. Meyarivan. A fast and elitist
multiobjective genetic algorithm: NSGA-II. IEEE Trans. Evolutionary
Computation, 6:182–197, 2002.
[11] K. Deb, L. Thiele, M. Laumanns, and E. Zitzler. Scalable test problems
for evolutionary multiobjective optimization. In Evolutionary Multiobjective Optimization, Advanced Information and Knowledge Processing,
pp. 105–145. 2005.
[12] I. Diakonikolas and M. Yannakakis. Small approximate Pareto sets for
biobjective shortest paths and other problems. SIAM Journal on Computing, 39:1340–1371, 2009.
[13] J. J. Durillo, A. J. Nebro, and E. Alba. The jMetal framework for multiobjective optimization: Design and architecture. In IEEE Congress on
Evolutionary Computation (CEC ’10), pp. 4138–4325, 2010.
[14] M. T. M. Emmerich, N. Beume, and B. Naujoks. An EMO algorithm using the hypervolume measure as selection criterion. In 3rd International
Conference on Evolutionary Multi-Criterion Optimization (EMO ’05),
pp. 62–76. Springer, 2005.
[15] Y. G. Evtushenko and M. Potapov. Methods of numerical solution of
multicriterion problem. In Soviet mathematics – doklady, Vol. 34, pp.
420–423, 1987.
[16] M. Gong, L. Jiao, H. Du, and L. Bo. Multiobjective immune algorithm
with nondominated neighbor-based selection. Evolutionary Computation, 16:225–255, 2008.
[17] J. J. Grefenstette. Incorporating problem specific knowledge into genetic
algorithms. Genetic algorithms and simulated annealing, 4:42–60, 1987.
[18] N. Hansen. The CMA evolution strategy: a comparing review. In Towards a new evolutionary computation. Advances in estimation of distribution algorithms, pp. 75–102. Springer, 2006.
[19] P. Hansen. Bicriterion path problems. In Multiple Criteria Decision
Making: Theory and Applications, Vol. 177 of Lecture Notes in Economics and Mathematical Systems, pp. 109–127, 1980.
13
[20] G. R. Harik and D. E. Goldberg. Linkage learning through probabilistic
expression. Computer Methods in Applied Mechanics and Engineering,
186:295–310, 2000.
[21] I. Hatzakis and D. Wallace. Dynamic multi-objective optimization with
evolutionary algorithms: a forward-looking approach. In Proc. 8th Annual Conference on Genetic and Evolutionary Computation Conference
(GECCO ’06), pp. 1201–1208. ACM Press, 2006.
[22] A. Hernandez-Diaz, C. A. Coello Coello, F. Perez, R. Caballero,
J. Molina, and L. Santana-Quintero. Seeding the initial population of
a multi-objective evolutionary algorithm using gradient-based information. In Proc. Congress on Evolutionary Computation (CEC ’08), pp.
1617–1624. IEEE Press, 2008.
[23] E. Hopper and B. Turton. An empirical investigation of meta-heuristic
and heuristic algorithms for a 2d packing problem. European Journal of
Operational Research, 128:34–57, 2001.
[24] S. Huband, L. Barone, R. L. While, and P. Hingston. A scalable multiobjective test problem toolkit. In 3rd International Conference on Evolutionary Multi-Criterion Optimization (EMO ’05), Vol. 3410 of LNCS,
pp. 280–295. Springer, 2005.
[25] H. Ishibuchi, N. Tsukamoto, Y. Sakane, and Y. Nojima. Indicator-based
evolutionary algorithm with hypervolume approximation by achievement scalarizing functions. In 12th Annual Conference on Genetic
and Evolutionary Computation Conference (GECCO ’10), pp. 527–534.
ACM Press, 2010.
[26] B. Kazimipour, X. Li, and A. Qin. A review of population initialization
techniques for evolutionary algorithms. In Proc. Congress on Evolutionary Computation (CEC ’14), pp. 2585–2592, 2014.
[27] E. Keedwell and S.-T. Khu. A hybrid genetic algorithm for the design
of water distribution networks. Engineering Applications of Artificial
Intelligence, 18:461–472, 2005.
[28] H. Li and Q. Zhang. Multiobjective optimization problems with complicated pareto sets, MOEA/D and NSGA-II. IEEE Trans. on Evolutionary Computation, 13:284–302, 2009.
14
[29] C.-F. Liaw. A hybrid genetic algorithm for the open shop scheduling
problem. European Journal of Operational Research, 124:28–42, 2000.
[30] P. Loridan. ε-solutions in vector minimization problems. Journal of Optimization Theory and Applications, 43:265–276, 1984.
10.1007/BF00936165.
[31] S. Oman and P. Cunningham. Using case retrieval to seed genetic algorithms. International Journal of Computational Intelligence and Applications, 1:71–82, 2001.
[32] C. H. Papadimitriou and M. Yannakakis. On the approximability of
trade-offs and optimal access of web sources. In 41st Annual Symposium on Foundations of Computer Science (FOCS ’00), pp. 86–92. IEEE
Press, 2000.
[33] C. H. Papadimitriou and M. Yannakakis. Multiobjective query optimization. In 20th ACM Symposium on Principles of Database Systems
(PODS ’01), pp. 52–59, 2001.
[34] H. Reuter. An approximation method for the efficiency set of multiobjective programming problems. Optimization, 21:905–911, 1990.
[35] G. Ruhe and B. Fruhwirth. ε-optimality for bicriteria programs and its
application to minimum cost flows. Computing, 44:21–34, 1990.
[36] S. Vassilvitskii and M. Yannakakis. Efficiently computing succinct tradeoff curves. Theor. Comput. Sci., 348:334–356, 2005.
[37] M. Wagner and T. Friedrich.
Efficient parent selection for
approximation-guided evolutionary multi-objective optimization. In
Proc. IEEE Congress on Evolutionary Computation (CEC ’13), pp.
1846–1853. IEEE, 2013.
[38] M. Wagner and F. Neumann. A fast approximation-guided evolutionary
multi-objective algorithm. In Proc. 15th Annual Conference on Genetic
and Evolutionary Computation Conference (GECCO ’13), pp. 687–694.
ACM, 2013.
[39] M. Yang, X. Zhang, X. Li, and X. Wu. A hybrid genetic algorithm
for the fitting of models to electrochemical impedance data. Journal of
Electroanalytical Chemistry, 519:1–8, 2002.
15
[40] E. Zitzler and S. Künzli. Indicator-based selection in multiobjective
search. In 8th International Conference on Parallel Problem Solving
from Nature (PPSN VIII), Vol. 3242 of LNCS, pp. 832–842. Springer,
2004.
[41] E. Zitzler and L. Thiele. Multiobjective evolutionary algorithms: A
comparative case study and the strength Pareto approach. IEEE Trans.
Evolutionary Computation, 3:257–271, 1999.
[42] E. Zitzler, K. Deb, and L. Thiele. Comparison of Multiobjective Evolutionary Algorithms: Empirical Results. Evolutionary Computation, 8:
173–195, 2000.
[43] E. Zitzler, M. Laumanns, and L. Thiele. SPEA2: Improving the strength
Pareto evolutionary algorithm for multiobjective optimization. In Evolutionary Methods for Design, Optimisation and Control with Application
to Industrial Problems (EUROGEN 2001), pp. 95–100, 2002.
16
DTLZ1 2D
102
102
101
101
100
100
10−1
10−1
10−2
10−2
DTLZ1 6D
10−3 3
10
103
DTLZ4 2D
105
106
10−3 3
10
103
102
102
101
101
100
100
10−1
10−1
103
104
105
106
103
100
100
10−1
10−1
10−2
10−2
103
DTLZ4 6D
104
104
105
106
100
10−1 3
10
103
104
105
106
104
105
106
104
105
106
104
105
106
100
104
105
106
10−1 3
10
Figure 1: Comparison of seeding with CornersAndCentre (left column) and LinearCombinations
(right column) on DTLZ1 and DTLZ4 for two and six dimensions. The approximation constant of the
Pareto front (y-axis) is shown as a function of the number of fitness function evaluations (x-axis) for
the seeded (×) and unseeded (+) versions of AGE ( ), IBEA ( ), NSGA-II ( ), SMS-EMOA ( ),
and SPEA2 ( ). The figures show the average of 100 repetitions each. Smaller approximation constants
indicate a better approximation of the front. The plots for the seeded versions are shifted by the number of
iterations required by the CornersAndCentre seeding (104 iterations) and the LinearCombinations
seeding (105 iterations); circles indicate the approximation of the initial seeding. The shaded areas illustrate
the difference between seeding and no seeding for a specific algorithm. Plots end prematurely if the time
limit of four hours was reached.
ZDT1
100
100
10−1
10−1
10−2
10−2
ZDT2
10−3 3
10
105
106
10−3 3
10
100
100
10−1
10−1
10−2
10−2
10−3 3
10
LZ09 F1
104
104
105
106
10−3 3
10
100
100
10−1
10−1
10−2
10−2
103
104
105
106
103
100
10−1
10−1
105
106
104
105
106
104
105
106
104
105
106
LZ09 F2
100
104
103
104
105
106
103
Figure 2: Comparison of seeding with CornersAndCentre (left column) and LinearCombinations
(right column) on ZDT1, ZDT2, LZ09 F1 and LZ09 F2. The approximation constant of the Pareto front
(y-axis) is shown as a function of the number of fitness function evaluations (x-axis) for the seeded (×)
and unseeded (+) versions of AGE ( ), IBEA ( ), NSGA-II ( ), SMS-EMOA ( ), and SPEA2
( ). The figures show the average of 100 repetitions each. Smaller approximation constants indicate a
better approximation of the front. The plots for the seeded versions are shifted by the number of iterations
required by the CornersAndCentre seeding (104 iterations) and the LinearCombinations seeding
(105 iterations); circles indicate the approximation of the initial seeding. The shaded areas illustrate the
difference between seeding and no seeding for a specific algorithm. Plots end prematurely if the time limit
of four hours was reached.
Function
DTLZ1 2D
DTLZ1 4D
DTLZ1 6D
DTLZ1 8D
DTLZ2 2D
DTLZ2 4D
DTLZ2 6D
DTLZ2 8D
DTLZ3 2D
DTLZ3 4D
DTLZ3 6D
DTLZ3 8D
DTLZ4 2D
DTLZ4 4D
DTLZ4 6D
DTLZ4 8D
LZ09 F1
LZ09 F2
LZ09 F3
LZ09 F4
LZ09 F5
LZ09 F6
LZ09 F7
LZ09 F8
LZ09 F9
AGE
(
)
0.99 =
0.94 <
0.18 <
1.02 =
0.96 <
1.08 >
0.95 <
0.85 <
0.85 <
0.97 <
0.09 <
0.86 <
0.96 <
4.73 >
1.08 >
1.03 >
3.56 >
1.47 >
1.06 >
6.36 >
1.13 >
1.14 >
1.52 >
0.99 =
2.16 >
IBEA NSGA-II SMS-EMOA SPEA2
(
)
1.00 =
0.98 =
1.00 =
0.98 =
1.01 =
1.00 =
1.00 =
1.00 =
1.00 =
1.00 =
1.00 =
1.00 =
1.14 >
3.79 >
2.22 >
1.79 >
1.50 >
0.97 <
1.03 >
4.11 >
1.07 >
1.00 >
1.80 >
1.47 >
1.30 >
(
)
1.00 =
0.99 =
1.00 =
1.00 =
1.00 =
1.00 =
1.00 =
0.98 =
1.00 =
1.00 =
0.99 =
1.00 =
1.00 >
1.00 >
1.00 =
1.01 =
1.05 >
0.73 <
0.99 =
0.96 <
0.98 =
1.00 =
0.88 <
1.09 =
0.81 <
(
)
1.00 =
1.44 >
–
–
0.99 <
1.01 >
–
–
1.00 =
0.99 =
–
–
1.02 >
5.24 >
1.91 >
–
1.04 >
1.47 >
1.05 >
5.98 >
1.06 >
1.09 >
1.75 >
1.17 >
2.23 >
(
)
1.02 =
1.05 =
1.01 =
0.98 <
1.00 =
1.00 =
1.00 =
1.00 >
1.02 =
0.95 =
1.00 =
1.03 >
1.05 >
2.91 >
0.98 <
1.01 =
0.97 =
1.50 >
1.01 =
6.49 >
1.04 >
1.02 >
1.58 >
0.99 =
2.22 >
Function
ZDT1
ZDT2
ZDT3
ZDT4
ZDT6
WFG1
WFG1
WFG2
WFG2
WFG3
WFG3
WFG4
WFG4
WFG5
WFG5
WFG6
WFG6
WFG7
WFG7
WFG8
WFG8
WFG9
WFG9
2D
3D
2D
3D
2D
3D
2D
3D
2D
3D
2D
3D
2D
3D
2D
3D
2D
3D
AGE
(
)
0.99 <
1.00 =
1.02 =
1.01 =
1.01 =
1.00 =
1.01 =
0.01 <
1.00 =
1.00 =
0.98 =
0.99 =
0.99 =
1.00 =
0.98 =
0.06 <
0.38 <
1.01 >
1.00 =
0.60 <
0.87 <
1.01 =
1.01 >
IBEA NSGA-II SMS-EMOA SPEA2
(
)
1.01 =
1.02 =
1.01 =
0.97 =
1.05 >
1.01 =
0.99 =
0.05 <
0.96 =
1.00 =
0.91 <
1.00 =
0.99 =
1.00 >
1.00 =
0.15 <
0.57 <
0.90 <
1.00 =
0.56 <
0.82 <
0.89 <
1.01 =
(
)
1.00 =
1.00 <
1.00 =
1.00 =
1.00 >
1.00 =
1.01 >
0.47 <
1.00 =
1.00 =
1.00 =
1.00 =
1.00 =
1.00 >
1.00 >
0.73 <
0.86 <
1.00 >
1.00 =
0.96 <
1.21 >
1.00 =
1.00 =
(
)
1.00 =
1.01 =
0.99 <
1.01 =
1.00 =
1.07 =
1.01 >
0.01 <
1.00 =
1.00 =
1.01 >
1.00 =
1.00 =
1.00 =
1.00 =
0.06 <
0.41 <
0.99 <
1.00 =
1.00 <
0.99 <
1.01 =
1.02 >
Table 1: Summary of our results for the improvement through LinearCombinations seeding. We
compare the default strategy NoSeed with 106 fitness function evaluations and LinearCombinations
seeding, which uses 105 fitness function evaluations, plus 9·105 fitness function evaluations. The table shows
the ratio of the median approximation constant of NoSeed divided by the median approximation constant
of LinearCombinations (100 independent runs each). Values >1.00 indicate where LinearCombinations achieves a better additive approximation, as the default strategy’s outcome is in the dividend. To
facilitate qualitative observations, we show only two decimal place. “>” marks statistically significant improvements, “<” marks statistically significant worsenings, “=” marks statistically insignificant findings.
In case a MOEA needed more than 4h time, the approximation constant after 4h is used. Dashes indicate
scenarios where not even the first iteration of the algorithm was completed within the allotted 4h.
(
)
0.98 =
1.02 =
1.00 =
0.99 =
1.01 =
1.00 =
1.01 =
0.02 <
1.02 =
1.00 =
0.97 =
0.99 =
1.01 =
1.00 =
0.98 =
0.09 <
0.63 <
1.00 =
1.00 =
1.00 <
0.98 <
1.02 =
1.00 =
Function
DTLZ1 2D
DTLZ1 4D
DTLZ1 6D
DTLZ1 8D
DTLZ2 2D
DTLZ2 4D
DTLZ2 6D
DTLZ2 8D
DTLZ3 2D
DTLZ3 4D
DTLZ3 6D
DTLZ3 8D
DTLZ4 2D
DTLZ4 4D
DTLZ4 6D
DTLZ4 8D
LZ09 F1
LZ09 F2
LZ09 F3
LZ09 F4
LZ09 F5
LZ09 F6
LZ09 F7
LZ09 F8
LZ09 F9
AGE
(
)
0.99 =
1.00 =
4.06 >
1.03 >
1.00 =
1.06 >
1.05 >
0.99 =
0.99 =
0.99 =
1.27 >
1.09 >
0.98 =
4.51 >
0.47 <
0.98 =
3.39 >
0.90 <
1.05 >
2.84 >
0.97 <
0.94 <
1.08 >
1.07 >
1.76 >
IBEA NSGA-II SMS-EMOA SPEA2
(
)
1.01 =
1.01 =
1.00 =
0.97 =
1.00 =
1.01 =
0.99 <
1.00 =
1.00 =
1.00 =
1.00 =
1.00 <
1.13 >
0.99 <
0.97 <
0.97 <
1.57 >
1.03 >
0.99 =
2.46 >
0.91 <
1.00 >
0.90 >
1.69 >
1.55 >
(
)
1.01 =
1.00 =
1.00 =
1.01 >
1.00 =
1.00 =
0.99 =
1.03 =
1.00 =
1.00 =
0.99 =
1.00 =
1.00 >
0.99 =
1.00 =
1.05 >
1.06 >
0.93 <
0.93 <
1.03 =
1.02 =
1.00 =
0.93 <
0.74 <
0.83 <
(
)
1.00 =
0.73 <
–
–
1.00 =
1.00 =
–
–
0.99 =
0.34 <
–
–
1.03 >
5.12 >
1.17 >
–
1.00 =
1.08 >
1.02 =
2.97 >
0.89 <
0.89 <
1.29 >
1.33 >
1.68 >
(
)
1.01 =
0.96 =
1.02 =
0.99 =
1.02 =
1.01 =
0.99 =
1.01 =
1.01 =
0.95 =
0.98 =
1.02 >
1.05 >
1.00 =
1.05 >
1.00 =
0.98 =
1.06 >
0.97 =
3.49 >
0.88 <
0.28 <
1.18 >
1.10 >
1.70 >
Function
ZDT1
ZDT2
ZDT3
ZDT4
ZDT6
WFG1
WFG1
WFG2
WFG2
WFG3
WFG3
WFG4
WFG4
WFG5
WFG5
WFG6
WFG6
WFG7
WFG7
WFG8
WFG8
WFG9
WFG9
2D
3D
2D
3D
2D
3D
2D
3D
2D
3D
2D
3D
2D
3D
2D
3D
2D
3D
AGE
(
)
0.99 <
1.02 =
1.01 =
0.99 =
1.01 =
1.02 >
1.00 =
0.98 =
1.00 =
1.00 =
1.01 =
1.00 =
1.00 =
1.00 =
0.99 =
0.93 =
1.01 =
1.01 =
1.00 =
1.02 =
1.18 =
1.00 =
1.00 =
IBEA NSGA-II SMS-EMOA SPEA2
(
)
1.01 =
1.02 =
0.99 =
0.98 =
1.03 =
1.01 =
0.98 <
1.00 =
0.93 =
1.00 =
1.00 =
1.00 =
0.98 =
1.02 >
1.00 =
0.95 =
0.99 =
1.00 =
1.02 >
0.70 <
1.00 =
1.03 =
1.01 =
(
)
1.00 =
1.00 =
1.00 =
1.00 =
1.00 >
0.87 =
1.00 =
1.00 =
1.00 =
1.00 =
1.00 =
1.00 =
1.00 =
1.00 >
1.00 >
1.00 =
1.00 <
1.00 =
1.00 =
0.96 >
1.21 >
1.00 =
1.00 =
(
)
0.99 <
1.01 >
0.99 =
1.01 =
1.00 =
1.06 >
1.00 =
1.01 =
1.00 =
1.00 =
1.00 =
1.01 >
1.00 =
1.00 =
1.00 =
0.90 =
1.00 =
0.99 <
1.00 =
1.00 =
1.00 =
0.99 =
1.00 =
Table 2: Summary of our results for the improvement through CornersAndCentre seeding. We
compare the default strategy NoSeed with 106 fitness function evaluations and CornersAndCentre
seeding, which uses 104 fitness function evaluations, plus 9 · 105 fitness function evaluations. The table
shows the ratio of the median approximation constant of NoSeed divided by the median approximation
constant of CornersAndCentre (100 runs independent each). Values >1.00 indicate where CornersAndCentre achieves a better additive approximation, as the default strategy’s outcome is in the dividend.
To facilitate qualitative observations, we show only two decimal place. “>” marks statistically significant
improvements, “<” marks statistically significant worsenings, “=” marks statistically insignificant findings. In case a MOEA needed more than 4h time, the approximation constant after 4h is used. Dashes
indicate scenarios where not even the first iteration of the algorithm was completed within the allotted 4h.
(
)
0.99 <
1.00 =
0.96 =
0.99 =
1.00 =
1.00 =
1.02 =
1.00 =
0.98 =
1.00 =
0.97 =
0.97 <
1.01 =
1.00 =
0.98 =
0.96 =
0.99 =
1.01 =
1.03 >
1.79 =
1.26 >
0.99 =
1.02 =
| 9cs.NE
|
arXiv:1802.08984v1 [cs.PL] 25 Feb 2018
Secure Serverless Computing Using Dynamic Information Flow
Control
KALEV ALPERNAS∗ , Tel Aviv University
CORMAC FLANAGAN, UC Santa Cruz
SADJAD FOULADI, Stanford University
LEONID RYZHYK, VMware Research
MOOLY SAGIV∗ , Tel Aviv University
THOMAS SCHMITZ, UC Santa Cruz
KEITH WINSTEIN, Stanford University
The rise of serverless computing provides an opportunity to rethink cloud security. We present an approach for securing
serverless systems using a novel form of dynamic information flow control (IFC).
We show that in serverless applications, the termination channel found in most existing IFC systems can be arbitrarily
amplified via multiple concurrent requests, necessitating a stronger termination-sensitive non-interference guarantee, which
we achieve using a combination of static labeling of serverless processes and dynamic faceted labeling of persistent data.
We describe our implementation of this approach on top of JavaScript for AWS Lambda and OpenWhisk serverless
platforms, and present three realistic case studies showing that it can enforce important IFC security properties with low
overhead.
1
INTRODUCTION
In May 2017, the Equifax credit reporting agency suffered a security breach, leaking social security numbers
and other personal information of 145.5 million consumers [34]. The breach, which exploited a code injection
vulnerability in Apache Struts [49], became the latest in a series of high-profile attacks on public and private
clouds compromising sensitive personal information of hundreds of million users [22, 23, 28, 33, 62, 65–68, 72, 73].
Most of these attacks can be traced down to two types of faults: misconfigurations and software vulnerabilities.
The former include issues like incorrect database security attributes [28, 53, 61], the choice of weak authentication
schemes [23], or the use of unpatched software [34]. The latter include code and SQL injections, file inclusions,
directory traversals, etc. [21, 22, 72, 73].
Simply put, the enormous Trusted Computing Base (TCB) of modern cloud applications makes it intractable to
enforce information security in these environments.
A promising avenue to a smaller TCB lies in the use of information flow control (IFC)-based security [25, 54]. In
the IFC world, information is protected by a global security policy that cannot be overridden by a misconfigured
application. The policy explicitly and concisely captures constraints on end-to-end information flow through the
system, e.g., “credit card numbers can only be exposed to appropriate card associations (e.g., Visa or MasterCard)”.
The IFC system enforces the policy even for buggy or malicious applications, thus removing application code
and configuration from the TCB of the cloud. In particular, an application that has been hijacked by a code
injection attack should not be able to bypass the enforcement mechanism. This is in contrast to security models
based on access control lists or capabilities, where, for instance, a compromised program running with database
administrator privileges can easily leak the entire database to a remote attacker.
Despite significant progress on IFC, it remains difficult to apply in real software. Dynamic IFC systems [8,
24, 29, 60] incur high runtime overhead. Static IFC [47, 48, 54, 71] systems shift the costs to development time,
∗ Work
done while visiting VMware Research.
usually via the use of type systems; however, they restrict the style of programming, which complicates their
adoption.
We demonstrate that IFC for cloud computing is not only feasible, but can be implemented with low overhead and
for essentially unmodified applications. We achieve these properties by leveraging recent developments in cloud
computing, namely, the rise of serverless computing [56]. Initially popularized by Amazon’s AWS Lambda [2],
serverless computing is rapidly gaining adoption by cloud providers [6, 32, 38, 41, 46] and tenants [1, 14, 19, 20,
30, 31, 36, 43, 51] due to its key benefits: elastic scalability, ease of deployment, and flexible pay-per-use pricing.
It achieves these benefits by decoupling application logic from resource management. In the serverless model,
users express their applications as collections of functions triggered in response to user requests or calls by other
functions. A function can be written in any language and may request a certain runtime environment, including,
e.g., specific versions of the Python interpreter and libraries. However, the function is agnostic of where this
environment is instantiated: a physical machine, a virtual machine or a container. The cloud platform manages
function placement and scheduling, automatically spawning new function instances on demand. This requires
application state to be decoupled from functions and placed in a shared data store (e.g., a database or a key-value
store), allowing all function instances to access the state regardless of their physical placement in the cloud.
We argue that serverless computing has fundamental implications for cloud security. In particular, it enables
practical IFC for the cloud. Our key observation is that a serverless function constitutes a natural unit of information
flow tracking. First, a serverless function activation handles a single request on behalf of a specific user and
only accesses secrets related to this request. Second, each invocation starts from a clean state and does not get
contaminated with sensitive data from previous invocations. Under a conservative assumption that all secrets
obtained during function execution propagate to all its outputs, we can track the global flow of information in the
system by monitoring inputs and outputs of functions comprising it.
Based on this observation, we develop the first IFC system
for serverless applications, called Trapeze. Trapeze encapsulates
cloud
sandbox
each unmodified serverless function in a sandbox, which intercepts all interactions between the function and the rest of the
F()
world, including other functions, shared data stores, and external
communication channels, and redirects them to the security shim
(a)
(b)
G()
shim
(Figure 1). The shim tracks information flow and enforces the
(d)
(c)
global security policy.
The class of supported policies, along with policy enforcement
datastore
rules, is defined by Trapeze’s dynamic IFC model. The model
addresses a weakness in most existing static and dynamic IFC Fig. 1. Trapeze architecture. Serverless function
systems, which leak information through the termination channel, F is encapsulated in a sandbox. All inputs and
whereby an adversary can infer secrets by observing termination outputs of F , including (a) invocations of F from
or non-termination of the program. The massively parallel nature within and from outside of the cloud, (b) calls to
of the serverless environment amplifies this weakness, allowing other serverless functions, (c) reads and writes to
the attacker to construct a high-bandwidth information channel, shared data stores, and (d) external communication channels, are monitored by the security shim.
effectively defeating the purpose of IFC (Section 2).
Our IFC model eliminates this channel by enforcing a strong
security property known as termination-sensitive non-interference (TSNI) [55]. Only a handful of previous IFC
models achieve this level of security; however these models are either too restrictive for most practical applications [15, 55, 58], or introduce significant overhead [27] (see Section 7).
Trapeze achieves TSNI through a novel combination of static program labeling with dynamic labeling of
the data store based on a faceted store semantics. Static program labeling restricts the sensitivity of data a
serverless function can observe ahead of time and is key to eliminating the termination channel. Dynamic data
2
labeling is crucial to securing unmodified applications that do not statically partition the data store into security
compartments, while the faceted store semantics eliminates implicit storage channels. We present a formal proof,
validated using the Coq proof assistant, that our model enforces TSNI.
We evaluate Trapeze on three real-world serverless applications: an online retail store [51], a parallel build
system [35], and an image feature extraction service [57]. We use Trapeze to secure these applications with
minimal changes to application code and with low runtime overhead.
Thus, our key contributions in this work are (1) a light-weight IFC shim architecture for serverless computing,
(2) a new IFC model that enforces TSNI, along with its formal semantics and proof of correctness, and (3) an
experimental evaluation of the architecture and the model on three serverless applications.
Finally, we point out that our IFC model is not limited to the serverless domain. Generally speaking, it applies to
any reactive system that decouples computation from state. Algorithmically, we show that TSNI can be enforced
in such systems efficiently. Examples of such systems include, e.g., Hadoop [5], Apache Spark [70] and Stateless
Network Functions [44].
2
WHY A NEW IFC MODEL FOR SERVERLESS?
As discussed above, a serverless function offers a convenient unit of information tracking, enabling practical IFC
for serverless applications. It seems natural that our next step should be adapting one of many existing IFC models
to the serverless environment. However, we do not take this path, as existing models do not provide adequate
security for serverless applications. Specifically, most previous IFC models, both dynamic and static, enforce a
security property known as termination-insensitive non-interference (TINI) [55]. Intuitively, TINI guarantees that
an attacker cannot deduce secrets stored in the system from its non-secret outputs. However, they may be able to
deduce part of a secret from the fact that the system stopped producing outputs.
This information channel, known as the termination channel, is often disregarded because it has low bandwidth,
typically leaking just a single bit. This is not true in serverless systems. Below, we construct an attack on a
serverless application, that amplifies the termination channel by spawning many parallel computations, each
leaking one bit.
Example 2.1. Consider a serverless system with two users: a benign user Bob and a malicious user Eve. We
introduce security labels b and e to tag Bob’s and Eve’s data respectively. Labels form a lattice, with labels higher
up in the lattice representing more secret data. Bob and Eve are mutually distrusting, therefore their labels are
incomparable (Figure 2a).
Eve launches a code injection attack against serverless function F, forcing the function to execute malicious
code in Figure 2b. This code is designed to leak Bob’s 64-bit secret stored under key 100 in the key-value store to
Eve.
We assume that the system is secured using a classical dynamic IFC model, where, upon reading a value
from the store, the function’s label gets bumped to the least upper bound of the function’s previous label and
the value’s label (note that, since the security shim cannot observe the data flow inside the function, it must
conservatively taint the entire function with this label). The function initially runs with Eve’s label e (every line
in the listing is annotated with the current label of the function). Before reading Bob’s secret, the function forks
64 instances of helper function H (lines 2–3). Each instance of H will leak one bit of the secret. Next, we read the
secret, bumping the function’s dynamic label to b ⊔ e = ⊤ and encode it using the first 64 keys in the store: we
store an arbitrary value under key i iff the ith bit of the secret is 1 (lines 5–7).
The helper function H also starts with initial label e. After a short delay (line 10), it reads the key equal to
the function’s argument i from the store (line 11). At this point the function’s label either bumps to ⊤ if the
corresponding bit of the secret is 1 or stays at e otherwise. Line 12 attempts to send a message to Eve (here, eve
is a handle to Eve’s HTTP session), which succeeds if the label of the channel (e) is higher than or equal to the
3
function F /* compromised */ ():
e for i = 0 to 63:
3 e
fork H(i)
4 e
secret = store . read (100)
5 ⊤
for i = 0 to 63:
6 ⊤
if secret [i ]:
7 ⊤
store . write (i ,1)
1
2
⟙
e
b
8
function H(i ):
e
delay (1 s)
11 e
store . read (i)
12 e/⊤
eve . send (i) /* diverges if label is ⊤*/
9
10
⟘
(a) Example security lattice. ⊤ is the most
secret label at the top of the lattice; ⊥ is the
bottom (least secret) label.
(b) High-bandwidth termination channel in the classical dynamic IFC model.
Fig. 2. A security lattice and a termination channel.
function’s current label and diverges otherwise. Eventually, the attacker at the other end of the channel receives
the list of bits of the secret equal to 0.
□
We implemented the above attack in AWS Lambda and were able to leak 170 bits per second by scaling
the number of threads. Thus, by elastically scaling the computation, the serverless architecture also scales the
termination channel. This is inline with theoretical results, which suggest that in a concurrent system, the
termination channel can leak a secret in time linear in the size of the secret [7].
We therefore aim for a stronger security guarantee, known as termination-sensitive non-interference (TSNI) [55],
which eliminates the termination channel. Note that the termination channel in the above example arises as the
function’s label, and hence its ability to send to an external channel, depends on the labels of values the function
reads from the store. In contrast, our proposed model assigns a static security label to each function activation.
To do so, we take advantage of the fact that a serverless function always runs on behalf of a specific user and
can be assigned a corresponding security label. The complete model, presented in Sections 3 and 4, also offers a
secure way to dynamically increase the function’s label without introducing a side channel.
The function’s label determines its view of the data store: the function can only observe the existence of
data whose label does not exceed the function’s label. For example, when reading a key that contains a secret
above function’s current label, the store returns the same result as if the key was not present in the store. This
information hiding semantics is somewhat tricky to maintain when multiple functions with incomparable labels
write to the same store location. We avoid information leaks in this situation by employing faceted store semantics,
where each record can contain several values (facets) with different security labels [10, 69].
To the best of our knowledge, Trapeze is the first IFC system to combine static program labeling with dynamic
labeling of data (using faceting). This combination eliminates termination and storage channels and enforces a
strong security property, TSNI.
3 INFORMAL DESIGN
3.1 Threat model and assumptions
We assume that the following entities are trusted, i.e., not malicious or compromised: (1) the cloud operator, (2)
physical hosts and network, (3) system software (OS, hypervisor, container manager, scheduler), (4) serverless
4
runtime, (5) shared data stores, (6) the sandboxing technology. Assumptions (1) through (5) can in the future be
relaxed with the help of a secure enclave technology such as Intel SGX [40, 42], data encryption, and software
verification [26, 39].
We further trust the serverless application administrator to enforce the following invariants on application
configuration: (1) all data stores used by the application are configured to only be accessible from serverless
functions, (2) all serverless functions in the system are sandboxed.
Finally, we trust the application developer to correctly define the application’s information flow policy and
declassifier functions (Section 3.4).
The rest of the application is untrusted. In particular, we assume that the attacker can compromise application
code running inside serverless functions, including any of the frameworks and libraries it uses.
In this paper we focus on data confidentiality, i.e., protecting sensitive data from being exposed to unauthorized
users. The complementary problem of enforcing data integrity, i.e., protecting data from unauthorized modification
is outside the scope of Trapeze, although it can also be enforced with the help of IFC techniques [54]. Finally,
we are not concerned with timing and other covert channels [17]. An example of such a channel is a function
running with a high label communicating with a function running with a low label by modulating its CPU or
memory usage.
3.2
Security lattice
We start the construction of our IFC model with the lattice of security labels. Labels represent security classes of
information flowing through the system. Trapeze does not assign any specific semantics to labels; however in
practice they typically represent users or roles of the system.
Trapeze relies on a trusted authentication gateway to tag all external input and output channels with correct
security labels. For example, when Eve establishes an HTTP session with the system, the session gets tagged
with Eve’s label e.
Given the labeling of inputs and outputs, Trapeze applies information flow rules presented below to enforce
that information received through an input channel labeled li can only be exposed through an output channel
labeled lo if li ⊑ lo .
3.3
Information flow rules
The choice of information flow rules determines two critical properties of an IFC system: security and transparency.
The former characterizes the class of insecure behaviors the system prevents. The latter characterizes the class
of secure programs that the system executes with unmodified semantics and that therefore do not need to be
modified to work with Trapeze. Trapeze’s enforces the strong security property of TSNI at the cost of some loss
of transparency that, we argue, is acceptable in serverless systems.
Trapeze assigns a runtime security label to every serverless function activation. This label is derived from the
event that triggered the function. In particular, if the function was invoked via an HTTP request from a user, it
obtains the user’s security label. Alternatively, when invoked by another function, it inherits the caller’s label.
The function’s label controls its ability to send to an output channel: a send is only allowed if the function’s label
is smaller than or equal to the channel label.
Trapeze also dynamically labels records in the data store. To this end, the security shim intercepts data store
operations issued by the function and modifies them to insert and check security labels. When a function creates
or updates a record in the store, the record inherits the function’s label (see detailed write semantics below).
When reading from the store, the function only observes values whose labels are below or equal to its own label.
From the function’s perspective, the store behaves as if it did not contain any data that the function may not
observe.
5
A function can upgrade its label to an arbitrary higher label using the raiseLabel operation. This operation
does not introduce an unauthorized information channel, as the decision to upgrade cannot depend on secrets
above function’s previous label (such secrets are simply invisible to the function). The upgrade mechanism is
useful, for example, when a function running on behalf of a regular user needs to update global statistics on
behalf of a superuser. Upgrade is a one-way operation: a function’s label can never be downgraded below its
current value.
Store semantics Trapeze’s security shim conceals the existence
of data whose security label is not less than or equal to the func1 function Fb ():
tion’s label. Maintaining this semantics is straightforward when
2 b
secret = store . read (100)
all writes to a data store location carry the same label. In the
3 b
for i = 0 to 63:
presence of conflicting writes, a resolution mechanism is required
4 b
if secret [i] == 1:
to avoid implicit storage channels [9]. When the conflicting labels
5 b
store . write (i ,1)
are comparable, conflict resolution can be performed, e.g., using
6
the no-sensitive-upgrade rule [9]. However, no such mechanism
7 function Fe ():
that avoids storage channels exists for writes with incomparable
8 e
x = 0
labels. The following example illustrates the problem:
9 e
for i = 0 to 63:
e
e
12 e
13 e
10
11
Example 3.1 (Implicit storage channel). Suppose we resolve conflicts by ignoring writes when an incomparable label exists in the
store (alternative strategies, e.g., silently overriding the existing
value or failing the conflicting write are similarly vulnerable).
Figure 3 shows two functions running with labels b and e respecFig. 3. Implicit storage channel via conflicting tively that collude to leak Bob’s 64-bit secret to Eve. Fb reads the
writes.
secret in line 2; however it does not have the authority to send
it to Eve directly. Instead it encodes each bit of the secret using
a record in the key-value store, as in Example 2.1 (lines 3–5). Fe
reconstructs the secret by attempting to write to locations 0 to 63 and then reading the value back in. Writes to
locations that correspond to 1s are ignored, indicating that the corresponding bit of the secret is 1.
□
store . write (i ,1234)
if store . read (i) == 1:
x [i] = 1
eve . send (x)
We eliminate such unauthorized flows using faceted store semantics, where each record can contain several
values (facets) with different labels [10, 69]. Facets are created dynamically: when a value with a new label is
stored in the record, a facet is created for it (see Section 4 for precise semantics). A read returns the most recent
write that is visible to the function. Thus, facets conceal writes with label b from a function running with label e,
unless b ⊑ e.
Example 3.2. We replay the example in Figure 3 with faceted store semantics. Since function labels b and e
are incomparable, their respective writes will go into different facets. The functions do not observe each other’s
writes either explicitly or indirectly, as in Example 3.1.
□
Faceted stores were previously introduced in IFC research in work on faceted execution [10, 12, 69]. The
fundamental difference from our approach is in the read semantics. Under faceted execution, a read conceptually
forks the program, creating a separate branch for each facet read from the store. If the program sends to an
external channel, only the branch whose label is compatible with that of the channel is allowed to send. Similar to
our design, faceted execution eliminates storage channels; however it does so at a potentially high runtime cost
and may become impractical in a system with a large security lattice. In contrast, Trapeze avoids faceted-execution
using the apriori knowledge of the function’s label by pruning all incompatible facets at read time. Existing
faceted execution systems expose the termination channel and therefore enforce TINI, whereas Trapeze enforces
the stronger TSNI property.
6
In the practical use of Trapeze, faceting is an exceptional situation. Trapeze is designed to run unmodified
applications that assume conventional store semantics. The moment multiple facets are created in some store
location, this semantics is violated, as different functions can now observe different values at the same location.
Trapeze treats such situations as attempted exploits and notifies the administrator, who can then take recovery
actions, e.g., remove the offending function from the system and rollback the store to the previous consistent
state. In the meanwhile Trapeze guarantees that the system continues running without exposing any sensitive
information to the attacker.
Faceted store semantics is emulated by the security shim on top of a conventional non-faceted store. In Section 6,
we implement facets on top of a key-value store. Yang et al. [69] present the design of a faceted SQL database.
Transparency The flip side of Trapeze’s strong TSNI security guarantees and light-weight protection is the
theoretical loss of transparency, i.e., the ability to run existing unmodified applications. By assigning a static
security label to a function, we restrict data that is visible to it. In particular, the function cannot access values
above its security level even if it does not send these values (or anything derived from them) through unauthorized
channels. On the other hand, all writes to the data store performed by the function are conservatively labeled
with the function’s label even if they do not carry any secrets.
Both problems can be addressed by refactoring the application. In particular, the function can gain access to
secret data via the raiseLabel operation. Conversely, one can avoid tainting data with excessively high labels by
splitting the offending function into several functions that run with lower labels. However, if many such changes
are required in order to adapt existing applications to work with Trapeze, this will create a barrier to Trapeze’s
practical adoption.
Our evaluation in Section 6 indicates that in practice the loss of transparency is not an issue in serverless
applications due to the common serverless design practice where every function only accesses values that are
related to a specific small task and are therefore likely to have compatible security labels.
3.4
Declassifiers
Many real-world applications allow limited flow of information down the security lattice. For example, a credit
reporting agency may make the distribution of consumers across credit score bands publicly available. This
statistics is computed based on the credit history of all consumers and must therefore be labeled with the least
upper bound of all their labels. However, since the aggregate statistics exposes negligible amount of information
about individual consumers, it can be safely declassified.
Similar to previous IFC models, Trapeze introduces declassifiers to support such scenarios. A declassifier is a
triple ⟨h, l, D⟩, where h and l are security labels, such that l ⊏ h, and D is a serverless function. A declassifier is
invoked just like any other serverless function; however its security label is computed using special rules. Let x
be the label of the calling function. Then the declassifier is assigned label x ′, such that:
(
l, if l ⊑ x ⊑ h
x′ =
x, otherwise
By design, declassifiers violate the non-interference property; therefore the formal model and proofs in Section 4
are given for the pure IFC model without declassifiers.
4
FORMAL SEMANTICS
In this section, we formalize our IFC semantics for serverless systems with an underlying persistent key-value
store. Since computations at different security labels might write to the same key, the key-value store maps each
key to a set of values (facets), each with a different label, and we order this set into a sequence according to the
temporal order of the writes. The initial store σ0 maps each key to the empty sequence: σ0 = λk.ϵ.
7
l
v
S
k
σ
t
p
ps
Σ
∈
∈
∈
∈
∈
∈
∈
∈
∈
Label
Value
LabeledValueSeq
Key
Store
Thread
Process
Processes
State
=
(Value × Label)∗
=
Key → LabeledValueSeq
e
∈
Event
ch
es
op
∈
∈
∈
OutputChannel
Events
Operation
f
∈
ReadContinuation
(t, l )
multisets of processes
(σ, ps)
::=
=
::=
(a) Serverless execution state.
::=
|
|
start p
output ch v
nop
=
::=
|
|
|
|
|
=
Event ∗
read k f
write k v t
send ch v t
fork t t
raiseLabel l t
stop
(Value × Label)⊥ → Thread
(b) External events and I/O operations.
e
Σ −→ Σ State transition relation
(σ, ps)
start p
−→
output ch v
(σ, ps ⊎ {p })
[s-start]
−→
(σ, ps ⊎ {(t ′, l )})
run(t ) = send ch v t ′
(σ, ps ⊎ {(t, l )})
−→
(σ, ps ⊎ {(f (last(σ (k )↓l )), l )})
run(t ) = read k f
(σ, ps ⊎ {(t, l )})
−→
(σ [k := write(σ (k), v, l )], ps ⊎ {(t ′, l )})
run(t ) = write k v t ′
[s-write]
run(t ) = fork t ′ t ′′
[s-fork]
(σ, ps ⊎ {(t, l )})
nop
nop
nop
(σ, ps ⊎ {(t, l )})
−→
(σ, ps ⊎
(σ, ps ⊎ {(t, l )})
−→
(σ, ps ⊎ {(t ′, l ′ )})
(σ, ps)
−→
(σ, ps)
nop
nop
{(t ′,
l )} ⊎
{(t ′′,
l )})
l ⊑ label(ch)
[s-send]
[s-read]
run(t ) = raiseLabel l ′ t ′
l ⊑ l′
[s-raise-label]
[s-skip]
es
Σ −→∗ Σ Multiple-step state transition relation
Σ
Σ
ϵ
−→∗
e ++e s
−→∗
Σ
[refl]
e
Σ −→ Σ′
Σ′′
es
Σ′ −→∗ Σ′′
[trans]
(c) Transition relations.
Fig. 4. Formal semantics.
A state Σ ∈ State of the system consists of a key-value store σ ∈ Store, plus a multiset of currently executing
serverless function activations called processes (see Figure 4(a)). Each process p = (t, l) consists of a thread t plus
its associated security label l ∈ Label.
Observable events e ∈ Event of the system include the input event start p, which starts a new process p,1 and
the output event output ch v, which sends the value v on output channel ch (see Figure 4(b)).
e
The state transition relation Σ −→ Σ ′ shown in Figure 4(c) describes how the system executes. The first
rule [s-start] handles an incoming event start p simply by adding p to the multiset of processes. The next
five transition rules all involve executing a particular process (t, l) until its next I/O operation op ∈ Operation.
For maximal generality, we do not formalize the computation language, but instead assume that the function
run : Thread → Operation executes the thread t and returns op, which includes a continuation for the rest of the
thread (analogous to the coinductive definitions used by Bohannon et al. [18]).
We describe each operation and its corresponding transition rule in turn:
Serverless systems spawn a new function/process p to handle each incoming event. Here, we assume that each incoming event contains
that new process, to simplify the formal development.
1
8
S↓l = S \ {(v ′, l ′) ∈ S | l ′ @ l }
ps↓l = multiset{(t, l ′) ∈ ps | l ′ ⊑ l }
σ↓l = λk.σ (k)↓l
e
∃t, l ′ . e = start (t, l ′) and l ′ ⊑ l
e↓l = e
∃ch, v. e = output ch v and label(ch) ⊑ l
nop otherwise.
(e 1 . . . en )↓l = e 1↓l . . . en ↓l
(σ , ps)↓l = (σ↓l , ps↓l )
Fig. 5. Definition of projection function.
• [s-send] for send ch v t ′: This rule checks that the process is permitted to output on channel ch (here,
label : OutputChannel → Label returns the security label of each channel). The process becomes stuck if
this check fails. Otherwise, it generates the output event e = output ch v and the new process state (t ′, l)
using the continuation t ′ returned from run.
• [s-read] for read k f : This rule reads the labeled value sequence σ (k) from the store; uses the projection
operation •↓l defined in Figure 5 to remove all values not visible to the current label; and passes the last
entry in this list (i.e. the most recent visible write) to the read continuation f .
Note that σ (k)↓l may be the empty sequence ϵ, either because key k was never written, or because no such
writes are visible to the current process; in this case, last(ϵ) returns ⊥, which is passed to f .
This rule (and the following four below) generates a dummy event nop, since it does not have any externally
visible behavior.
• [s-write] for write k v t ′: In a conventional data store, a new write would overwrite the previous value at
that key. In contrast, our faceted store semantics must ensure that a low-label process unable to see the
new write will still read an older write. Hence we represent σ (k) as a sequence of labeled values. At a new
write (v, l) of value v at label l, we can garbage-collect or remove all older writes (v ′, l ′) in this sequence
σ (k) that are no longer visible, namely those where l ⊑ l ′, since any process that could read (v ′, l ′) can
also read the more recent write (v, l). The following function performs this garbage collection, and then
appends the new labeled value:
write : LabeledValueSeq × Value × Label →
LabeledValueSeq
write(S, v, l) = (S \ {(v ′, l ′) ∈ S | l ⊑ l ′ }) ++ (v, l)
The symbol ++ denotes sequence concatenation.
• [p-fork] for fork t ′ t ′′: This rule forks a new thread t ′′, where t ′ is the continuation of the original thread.
Both threads inherit the security label of the original process.
• [s-raise-label] for raiseLabel l ′ t ′: This rule simply raises the label of the current process to a higher label
l ′, which, for example, permits the process to read more secret data.
• Finally, [s-skip] allows a state to perform stuttering nop steps at any time, as a technical device to facilitate
the non-interference proof below.
Note that no rule is needed for the stop operation. Instead we just leave the stopped processes in the process
multiset for simplicity.
9
5
TERMINATION SENSITIVE NON-INTERFERENCE
We use the notation •↓l on various domains to remove any information that is not visible to an observer at level
l (see Figure 5). For example: S↓l contains only values with labels visible to l, and ps↓l contains only processes
with labels visible to l. An event e is visible to l if it starts a process visible to l, or if it outputs on a channel visible
to l; otherwise, we say e↓l = nop. We write •1 ≈l •2 to denote that items appear equivalent to an observer at
level l, i.e., •1↓l = •2 ↓l .
Our proof is based on the projection lemma below, which relates the execution of the full system Σ and the
portion Σ↓l visible at level l. Every step of Σ has a corresponding step in Σ↓l (part 1) and vice versa (part 2).
Lemma (Projection).
Part 1. If
e1
Σ −→ Σ1′
then for some Σ2′ and e 2 ,
e2
Σ↓l −→ Σ2′
Σ1′ ≈l Σ2′
Part 2. If
e 1′ ≈l e 2′
e1
Σ↓l −→ Σ1
then for some Σ2′ and e 2 ,
e2
Σ −→ Σ2
Σ1 ≈l Σ2
e 1 ≈l e 2
Proof. See appendix.
Based on this lemma, our proof of single-step and multi-step termination sensitive non-interference is straightforward.2
Theorem (Single step Termination Sensitive Non-Interference).
If
e1
Σ1 ≈l Σ2
Σ1 −→ Σ1′
then for some Σ2′ and e 2 ,
e2
Σ2 −→ Σ2′
Σ1′ ≈l Σ2′
e 1 ≈l e 2
Proof. By Projection (Part 1):
e3
Σ1 ↓l −→ Σ3′
for some
Σ3′
e3
e2
for some
e 1 ≈l e 3
and e 3 . The l-equivalence assumption implies that Σ2 ↓l −→ Σ3′ . By Projection (Part 2):
Σ2 −→ Σ2′
Σ2′
Σ1′ ≈l Σ3′
Σ3′ ≈l Σ2′
e 3 ≈l e 2
and e 2 . Therefore, by transitivity of ≈l , we have Σ1′ ≈l Σ2′ and e 1 ≈l e 2 , as required.
Corollary (Termination Sensitive Non-Interference).
If
es 1
Σ1 ≈l Σ2
Σ1 −→∗ Σ1′
then for some Σ2′ and e 2 ,
es 2
Σ2 −→∗ Σ2′
Σ1′ ≈l Σ2′
2
es 1 ≈l es 2 .
When applied to a reactive system as we have here, this notion of termination-sensitive non-interference is often known as progress-sensitive
non-interference.
10
es 1
Proof. By induction on the derivation of Σ1 −→∗ Σ1′ .
Our Coq formalization of these semantics and proofs is available at https://github.com/anonymous-pldi2018-1/
faceted-tsni-coq.
Our TSNI result states that the set of observable outputs of the system under all possible schedules does not
depend on inputs that are not visible to the observer. This result does not prevent a malicious scheduler from
leaking secrets by prioritizing certain schedules, e.g., by scheduling low-security processes based on high-security
secrets. However, as mentioned earlier, we assume that the scheduler is not adversarial in this manner. Prior
work [59] has addressed this problem by assuming a round-robin scheduler, but this assumption is not realistic
for serverless computing.
6 IMPLEMENTATION AND EVALUATION
6.1 Implementation
In order to evaluate our proposed security architecture and IFC model, we developed an open-source prototype
implementation of Trapeze [4]. The implementation is portable and currently runs on the two most popular
serverless platforms—AWS Lambda [2] and OpenWhisk [6]. It consists of three components: the sandbox, the
security shim, and the authentication service.
Sandbox The Trapeze sandbox encapsulates application code, redirecting all its inputs and outputs through
the security shim (Figure 1). The exact sandboxing technology depends on the programming language used.
We currently support serverless functions written in JavaScript for the Node.js runtime, which is one of the
most common types of serverless functions on both AWS and IBM Cloud Functions, IBM’s public OpenWhisk
service. We encapsulate such functions using the VM2 JavaScript sandbox [63, 64]. Many serverless functions
rely on the ability to invoke external programs. We encapsulate such external executables using a ptrace-based
sandbox [35], which restricts program’s I/O activity to a temporary local directory that gets purged on every
serverless function invocation.
Security shim The security shim monitors all inputs and outputs of a function and enforces IFC rules. The shim
consists of multiple adapter modules, one for each supported input and output interface. There are three groups
of adapters: (1) data store adapters, (2) function call adapters, (3) external channel adapters.
A data store adapter implements faceted store semantics on top of a conventional cloud data store. Trapeze
currently supports a single type of data store—a faceted key-value store implemented on top of a relational
database. The key-value store implements a standard dictionary with the following operations: put(key, value),
get(key), del(key), and keys() (returns all keys in the store). The store is backed by a relational database table
with 3 columns, for the key, value, and label. The table contains an entry for each facet of each value in the
store. We used MySQL server available in AWS through Amazon Relational Database Service. The security shim
passes an additional parameter to every operation—the security label. The get operation performs an SQL query
that returns all entries that match the given key, and whose label is less than or equal to the given label. The
del operation deletes all facets with labels greater than or equal to the given label. The put operation deletes
the same elements a del does, and then inserts the given key-value pair, with the given label. In addition to
the faceted version, we also provide a conventional insecure key-value store implementation as a baseline for
performance evaluation.
Function call adapters support different ways to invoke serverless functions, making sure that the callee inherits
the caller’s label, as required by our IFC model. We support two invocation mechanisms: AWS Step Functions,
which run a workflow with multiple serverless functions, controlled by a finite automaton, and Amazon Kinesis,
which supports asynchronous communication via real-time event streams.
11
External channel adapters enable secure communication across the cloud boundary. The supported types of
channels are (1) user-initiated HTTP sessions, (2) email communication via the Nodemailer module[50], and (3)
connections to external S3 buckets (used to upload large data objects that do not fit in HTTP requests). HTTP
sessions obtain their security labels from the authentication service (see below). The Nodemailer adapter uses the
user database (see below) to map an email address to a user security label. The S3 adapter inherits the label of the
user who provides login credentials for the S3 bucket.
Authentication service The authentication service is responsible for associating a correct security label with
every external HTTP session. It is implemented on top of a user database that stores credentials, email addresses,
and security labels of all users in the system.
The entire Trapeze framework consists of 1174 lines of JavaScript code, including 649 lines in the the AWS and
OpenWhisk shim modules, 484 lines in the key-value store, and 41 lines in the authentication service.
6.2
Evaluation questions
Our evaluation aims to answer the following questions:
(1) Security: Can Trapeze enforce information security in real-world serverless applications? In particular, can
confidentiality requirements of such applications be captured in a security policy consisting of a security
lattice and trusted declassifiers? Can Trapeze enforce the policy in the presence of buggy or malicious
code?
(2) Transparency: Can Trapeze secure existing serverless applications with minimal modifications?
(3) Performance: Can Trapeze achieve the first two goals with low performance overhead?
6.3
Case studies
To answer the above questions, we carried out three case studies where we used Trapeze to add a security layer
to existing serverless applications. We outline each of the case studies below.
Case study 1: Hello, Retail! Hello, Retail! is a project from the serverless team at Nordstrom. The goal was to
produce a purely serverless, back-end for an e-commerce web site. It has since been open-sourced [51] and won
the architecture competition award at serverlessConf Austin’17[45].
We made several changes to Hello, Retail! before applying Trapeze to it. First, we replaced DynamoDB and S3
databases in Hello, Retail! with with calls to Trapeze’s key-value store. Second, we replaced calls to the Twilio
SMS messaging service, which is currently not supported by Trapeze, with e-mail communication. Third, we
extended the Hello, Retail! project with a product purchase subsystem, which manages online orders and credit
card payments. The resulting system consists of 21 serverless functions. Figure 6 shows the high-level architecture
of the system.
The system serves several types of users: (1) the store owner, who manages the online catalog and processes
orders, (2) photographers, who upload product images to the catalog, (3) customers, who navigate the catalog
and place orders, and (4) the VISA credit card authority, which authorizes card payments on behalf of customers.
The security lattice (Figure 7a) consists of labels, matching these user categories. Solid lines in the diagram show
the partial order of security labels; dashed arrows show declassifiers, with a declassifier ⟨h, l, D⟩ represented by
an arrow from label h to l. Table 1 summarizes the security labels in this case study.
The following scenario illustrates the flow of sensitive information through the Hello, Retail! system. Every
step in the scenario is annotated with security label(s) of data involved in this step.
1. [owner ] The owner creates a new product description in the catalog.
2. [photoдrapher i ] The owner sends an email to one of the photographers requesting a picture of
the product. The request includes information from the product description, declassified with the
⟨owner, photoдrapher i , photo_request()⟩ declassifier, which implements a trusted user interface that request
12
Email
Photo
Request
Photographer
Registry
Register
Photographer API
Receive
Photo API
Get Photos
Step Function
Photos
Event
Processor
Retail Event
Stream
(AWS Kinesis)
Create
Product
API
Photo
Storage
Catalog
API
Product
Release API
Credit Card
Registry
Credit Card
Authorization
Catalog
Builder
Product
Catalog
Purchase Product
Step Function
Purchase
API
Fig. 6. The architecture of the Hello, Retail! project. Circles labeled λ represent the main functions of this application.
Functions whose names end with ‘API’ can be directly invoked via client HTTP requests. Arrows show interactions between
different components. Red dashed arrows indicate interactions that carry declassified data.
label
owner
client j
clientCC j
V ISA
photoдrapher i
description
Sensitive information managed by the store owner, including product catalog and photographers’ email addresses. Items in the catalog are only visible to the owner until they
are released to the public via the release declassifier.
Labels online purchases of a specific customer. This information is visible to the customer
and the store owner (since client j ⊏ owner )
Customer credit card information that can only be released to a credit card authority
Labels the external communication channel to the VISA credit card authority
Product photos uploaded by a photographer
Table 1. Security classes in the Hello, Retail! case study.
owner’s confirmation of the declassification.
3. [photoдrapher i ] The photographer uploads a product image to the catalog.
4. [⊥] Once the owner is ready to make the product publicly available in the online catalog, she declassifies it
using the ⟨owner, ⊥, release()⟩ declassifier.
5. [client j /clientCC j ] A client orders a product from the catalog. Order information, labeled client j is visible to
the the client as well as the owner, since client j ⊏ owner . Client’s credit card details are labeled clientCC j , and
are hidden from the owner (clientCC j @ owner ).
13
⟙
VISA
owner
clientCC
photo_request()
release()
photographer
authorize()
client
⟙
⟙
tenant
user
⟘
⟘
(a) Hello, Retail!.
open_source()
(b) gg
⟘
(c) Feature Extraction.
Fig. 7. Security lattice and declassifiers in the different case studies.
6. [VISA] Before the order is finalized, credit card information is sent to VISA for payment authorization through
an external channel labeled V ISA.
7. [client j ] The response received through this channel consists of one bit of information indicating success or
failure, which gets declassified by the ⟨V ISA, client j , authorize()⟩ declassifier, making the outcome of the request
visible to the client and the owner.
Case study 2: gg gg [35, 37] is a system for running parallel software workflows, such as software compilation
and video processing, on serverless platforms. In gg, each unit of work, or thunk, specifies both the executable to
run and all its data dependencies. The workflow is synthesized as a direct acyclic graph (DAG) of thunks, and
is recursively executed on a serverless platform by gg’s execution engine. gg identifies each dependency in a
content-addressed way, using a key-value store as the storage backend. gg consists of a single serverless function,
which internally runs arbitrary user-provided executables. Each invocation of the function executes exactly one
thunk by fetching the dependencies from the object store, executing the thunk, and storing the output back into
the object store.
We use a parallel build framework implemented on top of gg [35] as a concrete use case. The framework
extracts a workflow DAG, where every thunk corresponds to an invocation of a build tool (e.g., a compiler or a
linker), from the project makefile.
The original gg implementation is single-tenant, with every authenticated user having access to all sources
and binaries in the system. We use Trapeze to introduce a secure multi-tenant mode to gg. In this mode, tenants
only have access to their own source and compiled code. A thunk running on behalf of the tenant taints all of its
outputs with the tenant’s label. A tenant may release some of their sources or compiled binaries to the public,
making them available to all other tenants. This is reflected in the security lattice in Figure 7b with mutually
incomparable tenant labels and a ⟨user i , ⊥, open_source()⟩ declassifier.
Prior to adding the multi-tenant mode to gg, we ported parts of it that were written in Python to JavaScript, as
well as modified it to use our key-value store.
Case study 3: Image feature extraction This serverless application gives its users access to Amazon’s AWS
Rekognition image analysis service [3]. It is based on the ‘Fetch File and Store in S3’ and ‘Analyse Image from S3’
examples from the Serverless Examples collection[57].
The application consists of the upload function that takes an image URL, fetches the image and stores it in
the key-value store, and the feature extraction function that uses AWS Rekognition to extract features from the
image and send them to the user.
We use Trapeze to add a security layer to this example, enforcing that every user can only access information
extracted from images they own. This policy is expressed in the simple security lattice in Figure 7c.
14
case study
app code modif. code declassifiers
Hello, Retail!
1,300
3
104
gg
8,800
2
94
Feature Extraction
95
0
0
Table 2. Size of the case studies in lines of code (LoCs): (1) LoCs comprising the application, (2) lines added or modified to
adapt the application to work with Trapeze, (3) LoCs in trusted declassifiers.
6.4
Security
We employ Trapeze to protect sensitive data in the three case studies. The original implementation of these
applications either offered coarse-grained protection, giving every authenticated user access to all data in
the system, or implemented ad hoc security policies embedded in application code. For example, the Hello,
Retail! system by design only exposes credit card details to the credit card authority. Such protection relies on
checks scattered around the application code and is easy to get wrong. Besides, it can be bypassed by any exploit
that subverts the application logic.
Trapeze captures the security requirements of each of the three case studies in a security policy consisting of a
security lattice and declassifiers, shown in Figures 7a, 7b, 7c. Our policies are simple and concise, consisting of
only several classes of labels and few declassifiers. The policies are decoupled from the application logic and its
software architecture. For instance, adding new functions to the application, changing its control flow, or even
refactoring the database schema, do not affect the security policy.
Furthermore, Trapeze is immune to malicious or compromised application logic. We simulated code injection
attacks in our case studies by replacing some of the original functions with malicious functions that attempt to
leak secrets to unauthorized users, similar to examples in Figures 2b and 3. As expected, these simulated attacks
failed when running the application with Trapeze.
6.5
Transparency
As discussed in Section 3, Trapeze enforces TSNI at the cost of reduced transparency, which may require the
developer to refactor the application to work with Trapeze.
Table 2 measures the loss of transparency in our case studies by reporting the size of changes to application
code in terms of lines of code (LoCs) required to adapt the application to work with Trapeze. This does not take
into account the changes needed to port the application to use our key-value store as well as other compatibility
changes that are not fundamental to the Trapeze architecture and can be made unnecessary with some additional
engineering effort.
None of the case studies required splitting a function into multiple functions. Hello, Retail! required calling
raiseLabel twice, once in order to upgrade the label of the purchase placement function from client j to clientCC j
before saving the customer’s credit card details in the data store, and once to upgrade the label of the payment
authorization function from client j to V ISA, in order to read credit card details from the store and send them
to the credit card authority. In addition, the gg case study required a minor change due to a technicality: the
existing code was not fully compatible with the VM2 sandbox.
These results indicate that in practice the loss of transparency is not an issue in serverless applications. This is
due to the common software design practices in the serverless world, where each function is assigned a single
small task and only accesses the values that are related to this task and are therefore likely to have compatible
security labels.
The last column of the table reports the total size of declassifiers used in each case study. This number
characterizes the amount of trusted application code in each example.
15
function runtime (ms)
insecure
Trapeze
Hello, Retail!
Update&Purchase 435
51,246
98,563
Build&Browse
955
257,799
391,900
gg
mosh
111
681,173
654,448
git
917 2,602,500
2,660,721
vim
234 1,242,873
1,338,128
openssh
654 1,626,223
1,649,139
Image Feature Extraction
Image Upload
1
475
525
Feature Extraction
1
1,882
2,114
Table 3. Trapeze performance.
scenario
6.6
#λ
∆(%)
92.33
52.03
-3.92
2.24
7.66
1.41
10.4
12.3
Performance
We measure the overhead of Trapeze by running case studies 1 and 3 on the AWS Lambda serverless platform.
We run case study 2 on the OpenWhisk platform in the IBM Cloud (formerly, IBM Bluemix), since AWS Lambda
does not support ptrace, which we use to sandbox binary executables in gg (Section 6.1).
Table 3 summarizes the runtime overhead of Trapeze. Since the overhead may depend on the exact workload, we
constructed several typical workflows for each case study. We test the Hello, Retail! case study on two workloads:
the ‘Build&Browse’ workload, that simulates the construction and browsing of the product catalog, and the
‘Update&Purchase’ workload, that simulates updates to the catalog followed by a series of online purchases.
We test gg by using it to compile four open source software packages: mosh, git, vim, and openssh. Finally, we
consider two scenarios for the Image Feature Extraction: the ‘Image Upload’ scenario, that uploads a single image
to the data store, and the ‘Feature Extraction’ scenario that performs feature extraction for a stored image.
For each workload, we report the total number of serverless function calls (the #λ column in Table 3). The
‘function runtime’ section of the table reports the total runtime of all functions in the scenario with and without
protection and the relative slowdown introduced by protection, averaged across 10 runs. These runtimes do not
include additional declassifier calls introduced in the protected execution. Accounting for declassifiers increased
the total runtime by another 10% in the Build&Browse workload, while not making any measurable impact on
other workloads. In our initial experiments we also measured the total time to execute each workload; however we
found that, due to the non-deterministic nature and varying resource availability in the serverless environment,
these times varied wildly across different runs and did not provide any insights into the performance of Trapeze.
For most workloads, Trapeze adds a modest overhead of up to 12.3% to function runtime. The negative overhead
in the gg mosh workload is due to the noisy IBM Cloud environment, where the runtime of a function varies
dramatically across different invocations based on the load on the node where the function is scheduled. The
overhead is higher in the Hello, Retail! case study, up to 92.33% in the Update&Purchase scenario. Further
benchmarking revealed that the bulk of this overhead is due to the startup time of the VM2 sandbox, which adds
100ms on average per function invocation. Since all functions in this example have short runtimes, the startup
time becomes a significant contributor to the total runtime. We measured that two thirds of the startup time is
spent loading libraries used by the application. In the future, much of this overhead can be eliminated by caching
preloaded libraries in memory.
Finally, we evaluate the storage overhead of our secure key-value store. Table 4 compares database sizes (in
kilobytes) for an insecure and secure versions of the store for each of the three case studies (we report results
16
case study
insecure (KB) Trapeze (KB)
Hello, Retail!
2,704
3,200
gg
15,054,000
15,118,496
Feature Extraction
8,256
8,416
Table 4. Storage overhead of Trapeze.
∆(%)
18.34
0.43
1.94
for one of the workloads in each case study, as the relative increase in the database size is independent of the
workload). The secure database requires more space, as it stores a security label with each value. This overhead
is low (under 2%) in examples where the database stores large objects, e.g., images or source files. It is more
significant (18.34%) in the Hello, Retail! case study where individual values stored in the database are small.
7
RELATED WORK
The modern technique for using security labels to dynamically monitor information flow was proposed by
Denning [25] as the Lattice Model. That work also defines the concept of implicit flow of information. Austin et al.
[11] describe three techniques for monitoring such flows, namely Failure Oblivious and No Sensitive Upgrade and
Permissive Upgrade. Our work exposes a fourth choice using faceted values specifically to eliminate implicit flows.
Austin and Flanagan [10] introduce faceted values as a full language-based enforcement technique: rather than
focusing specifically on controlling implicit flows, they combine the faceted value concept with the concept of
multiple executions, which enables more precise enforcement at the cost of more runtime overhead.
Trapeze enforces a strong form of non-interference, TSNI. Most existing IFC systems support only terminationinsensitive non-interference (TINI). The few previous systems that enforce TSNI are either too restrictive or
too costly for most practical applications. The multilevel security (MLS) model [15] achieves TSNI by statically
partitioning all the code and data in the system into security compartments. This model, designed primarily for
military-grade systems, is too restrictive for most applications and requires complete re-design of existing software.
Smith and Volpano [58] present a security type system that enforces TSNI by imposing a harsh restriction that
loop conditions may not depend on secret data. Heintze and Riecke [39] propose a secure typed Lambda calculus
called SLam. While SLam only enforces TINI, Sabelfeld and Sands [55] point out that a version of SLam with lazy
evaluation semantics would be termination-sensitive. This theoretical result has limited practical implications, as,
with the exception of Haskell, none of today’s major programming languages use lazy evaluation.
Indeed, Stefan et al. [59] implemented a Haskell library called LIO, which guarantees TSNI by requiring
programmers to decompose their programs into separate threads with floating labels. These are analogous to the
processes in our formalism. However, LIO uses the No Sensitive Upgrade rule to prevent implicit flows, while
Trapeze uses faceted values instead. They also assume a round-robin scheduler, which would be inappropriate
for our application to serverless computing.
Secure multi-execution [27] achieves TSNI by running multiple independent copies of the program, one for
each security class. This technique introduces CPU and memory overhead proportional to the number of security
classes. While acceptable in systems with few security classes (e.g., Devriese and Piessens [27] consider only
two classes, secret and non-secret), this becomes impractical in cloud-scale systems with potentially millions of
mutually untrusting users. Faceted execution has the potential to mitigate this drawback, and Bielova and Rezk
[16] have proposed a theoretical approach for extending the faceted execution model to enforce TSNI.
The Asbestos OS [29] applies dynamic IFC at the granularity of an OS process, similar to how Trapeze operates
at the granularity of a serverless function. Asbestos associates a static security label with each process; however
this label only serves as an upper bound on the label of data the process can access. Process’s effective label
changes dynamically, which enables the implicit termination channel.
17
To the best of our knowledge, Trapeze is the first system to apply IFC to serverless applications. Several
researchers advocate the use of IFC in the broader context of secure cloud computing [13, 52]; however we are
not aware of a practical implementation of these ideas.
8
CONCLUSION
The advent of serverless computing provides the opportunity to rebuild our cloud computing infrastructure
based on a rigorous foundation for information flow security. We present a novel and promising approach for
dynamic IFC in serverless systems. This approach combines (1) a sandbox and security shim that monitors all I/O
operations of each serverless function invocation; (2) static security labels for each serverless function invocation;
and (3) dynamic faceted labeling of data in the persistent store.
This combination of ideas provides the strong security guarantee of TSNI, which is necessary in serverless
settings to avoid high-bandwidth termination channel leaks via multiple concurrent requests.
Our Trapeze implementation of this approach is lightweight, requiring no new programming languages,
compilers, or virtual machines. The three case studies show that Trapeze can enforce important IFC properties
with low space and time overheads.
We believe Trapeze represents a promising approach for deploying serverless systems with rigorous security
guarantees that help prevent costly information leaks arising from buggy or mis-configured application code or
from code-injection attacks.
ACKNOWLEDGMENTS
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
Airbnb. 2017. StreamAlert: A serverless framework for real-time data analysis and alerting. (2017). http://airbnb.io/projects/streamalert/.
Amazon. 2017. AWS Lambda. (2017). https://aws.amazon.com/lambda/.
Amazon. 2017. AWS Rekognition. (2017). https://aws.amazon.com/rekognition/.
Anonymous authors. 2017. Trapeze source code repository. (2017). Link anonymized for double-blind reviewing.
Apache Software Foundation. [n. d.]. Apache Hadoop. ([n. d.]). https://hadoop.apache.org/.
Apache Software Foundation. 2017. OpenWhisk. (2017). https://openwhisk.apache.org/.
Aslan Askarov, Sebastian Hunt, Andrei Sabelfeld, and David Sands. 2008. Termination-Insensitive Noninterference Leaks More Than
Just a Bit. In Proc. of ESORICS 2008. Malaga, Spain, 333–348.
Thomas H. Austin and Cormac Flanagan. 2009. Efficient Purely-dynamic Information Flow Analysis. In Proc. of PLAS 2009. 113–124.
Thomas H. Austin and Cormac Flanagan. 2010. Permissive Dynamic Information Flow Analysis. In Proc. of PLAS 2010. 1–12.
Thomas H. Austin and Cormac Flanagan. 2012. Multiple Facets for Dynamic Information Flow. In Proc. of POPL 2012. 165–178.
Thomas H. Austin, Tommy Schmitz, and Cormac Flanagan. 2017. Multiple Facets for Dynamic Information Flow with Exceptions. ACM
Trans. Program. Lang. Syst. 39, 3, Article 10 (May 2017), 56 pages. https://doi.org/10.1145/3024086
Thomas H. Austin, Jean Yang, Cormac Flanagan, and Armando Solar-Lezama. 2013. Faceted Execution of Policy-agnostic Programs. In
Proc. of PLAS. Seattle, Washington, USA, 15–26.
Jean Bacon, David Eyers, Thomas F. J.-M. Pasquier, Jatinder Singh, Ioannis Papagiannis, and Peter Pietzuch. 2014. Information Flow
Control for Secure Cloud Computing. IEEE Transactions on Network and Service Management 11, 1 (Jan. 2014), 76–89.
Andrew Baird, Michael Connor, and Patrick Brandt. 2016. Coca-Cola: Running Serverless Applications with Enterprise Requirements.
(2016). https://aws.amazon.com/serverless/videos/video-lambda-coca-cola/.
D. Elliott Bell and Leonard J. LaPadula. 1973. Secure Computer Systems: Mathematical Foundations. Technical Report 2547. MITRE.
Nataliia Bielova and Tamara Rezk. 2016. Spot the difference: Secure multi-execution and multiple facets. In European Symposium on
Research in Computer Security. Springer, 501–519.
Arnab Kumar Biswas, Dipak Ghosal, and Shishir Nagaraja. 2017. A Survey of Timing Channels and Countermeasures. ACM Comput.
Surv. 50, 1 (March 2017), 6:1–6:39.
Aaron Bohannon, Benjamin C Pierce, Vilhelm Sjöberg, Stephanie Weirich, and Steve Zdancewic. 2009. Reactive noninterference. In
Proceedings of the 16th ACM conference on Computer and communications security. ACM, 79–90.
Mark Boyd. 2017. iRobot Confronts the Challenges of Running Serverless at Scale. (2017).
https://thenewstack.io/
irobot-confronts-challenges-running-serverless-scale/.
18
[20] Kuldeep Chowhan. 2016. Serverless Computing Patterns at Expedia. (2016). https://www.slideshare.net/AmazonWebServices/
aws-reinvent-2016-serverless-computing-patterns-at-expedia-svr306.
The PlayStation Network breach (FAQ).
(May 2011).
https://www.cnet.com/news/
[21] CNET Magazine. 2011.
the-playstation-network-breach-faq/.
[22] Computerworld. 2009. SQL injection attacks led to Heartland, Hannaford breaches. (Aug. 2009). https://www.computerworld.com/
article/2527185/security0/sql-injection-attacks-led-to-heartland--hannaford-breaches.html.
[23] Computerworld. 2014. Two-factor authentication oversight led to JPMorgan breach. (Dec. 2014). https://www.computerworld.com/
article/2862578/twofactor-authentication-oversight-led-to-jpmorgan-breach-investigators-reportedly-found.html.
[24] Willem De Groef, Dominique Devriese, Nick Nikiforakis, and Frank Piessens. 2012. FlowFox: a web browser with flexible and precise
information flow control. In Proc. of CCS 2012. 748–759.
[25] Dorothy E Denning. 1976. A lattice model of secure information flow. Comm, of the ACM 19, 5 (1976), 236–243.
[26] Dorothy E. Denning and Peter J. Denning. 1977. Certification of Programs for Secure Information Flow. Commun. ACM 20, 7 (July 1977),
504–513.
[27] Dominique Devriese and Frank Piessens. 2010. Noninterference Through Secure Multi-execution. In Proc. IEEE SSP 2010. 109–124.
[28] Digital Trends. 2016. The latest data breach involves the voting records of 93.4 million Mexican citizens. (April 2016). https:
//www.digitaltrends.com/computing/mexico-voting-breach/.
[29] Petros Efstathopoulos, Maxwell Krohn, Steve VanDeBogart, Cliff Frey, David Ziegler, Eddie Kohler, David Mazières, Frans Kaashoek,
and Robert Morris. 2005. Labels and event processes in the Asbestos operating system. In Proc. of SOSP 2005.
https://serverless.com/blog/
[30] Ken Ellis. 2017. How Reuters Replaced WebSockets with Amazon Cognito and SQS. (2017).
how-reuters-replaced-websockets-with-amazon-cognito-and-sqs/.
[31] Marius Eriksen. 2013. Your server as a function. In In Proc. of PLOS 2013.
[32] Fn Project. 2017. (2017). https://fnproject.io/.
[33] Forbes. 2014. eBay Suffers Massive Security Breach, All Users Must Change Their Passwords. (May 2014). https://www.forbes.com/
sites/gordonkelly/2014/05/21/ebay-suffers-massive-security-breach-all-users-must-their-change-passwords/#793467c57492.
[34] Forbes. 2017. How Hackers Broke Equifax: Exploiting A Patchable Vulnerability. (Sept. 2017). https://www.forbes.com/sites/
thomasbrewster/2017/09/14/equifax-hack-the-result-of-patched-vulnerability/#20abe9015cda.
[35] Sadjad Fouladi, Dan Iter, Shuvo Chatterjee, Christos Kozyrakis, Matei Zaharia, and Keith Winstein. 2017. A Thunk to Remember: make
-j1000 (and other jobs) on functions-as-a-service infrastructure (Under review). (2017). http://stanford.edu/~sadjad/gg-paper.pdf.
[36] Sadjad Fouladi, Riad S. Wahby, Brennan Shacklett, Karthikeyan Vasuki Balasubramaniam, William Zeng, Rahul Bhalerao, Anirudh
Sivaraman, George Porter, and Keith Winstein. 2017. Encoding, Fast and Slow: Low-Latency Video Processing Using Thousands of Tiny
Threads. In Proc. of NSDI 2017. Boston, MA, 363–376.
[37] gg project. 2017. gg source code repository. (2017). https://github.com/stanfordsnr/gg.
[38] Google. 2017. Google Cloud Functions. (2017). https://cloud.google.com/functions/.
[39] Nevin Heintze and Jon G. Riecke. 1998. The SLam Calculus: Programming with Secrecy and Integrity. In Proc. of POPL 1998. San Diego,
California, USA, 365–377.
[40] Tyler Hunt, Zhiting Zhu, Yuanzhong Xu, Simon Peter, and Emmett Witchel. 2016. Ryoan: A Distributed Sandbox for Untrusted
Computation on Secret Data. In Proc. of OSDI 2016. Savannah, GA, USA, 533–549.
[41] IBM. 2017. IBM Cloud Functions. (2017). https://console.bluemix.net/openwhisk/.
[42] Intel Corporation. 2014. Intel Software Guard Extensions Programming Reference. (2014).
[43] Eric Jonas, Shivaram Venkataraman, Ion Stoica, and Benjamin Recht. 2017. Occupy the Cloud: Distributed Computing for the 99%. CoRR
abs/1702.04024 (2017). http://arxiv.org/abs/1702.04024
[44] Murad Kablan, Azzam Alsudais, Eric Keller, and Franck Le. 2017. Stateless Network Functions: Breaking the Tight Coupling of State and
Processing. In Proc. of NSDI 2017. Boston, MA, 97–112.
[45] McKim, John. 2017. Announcing the Winners of the Inaugural ServerlessConf Architecture Competition. (2017). https://read.acloud.
guru/announcing-the-winners-of-the-inaugural-serverlessconf-architecture-competition-1dce2db6da3.
[46] Microsoft. 2017. Azure Functions. (2017). https://azure.microsoft.com/services/functions/.
[47] Andrew C Myers. 1999. JFlow: Practical mostly-static information flow control. In Proc. of POPL 1999. 228–241.
[48] Andrew C Myers and Barbara Liskov. 2000. Protecting privacy using the decentralized label model. TOSEM 9, 4 (2000), 410–442.
[49] National Vulnerability Database. 2017. CVE-2017-5638. (March 2017). https://nvd.nist.gov/vuln/detail/CVE-2017-5638.
[50] nodemailer 2017. nodemailer. (2017). https://github.com/nodemailer/nodemailer.
[51] Nordstrom Technology. 2017. Hello, Retail! (2017). https://github.com/Nordstrom/hello-retail.
[52] Thomas Pasquier, Jean Bacon, Jatinder Singh, and David Eyers. 2016. Data-Centric Access Control for Cloud Computing. In Proc. of
SACMAT 2016. Shanghai, China, 81–88.
[53] PCWorld. 2010. Microsoft Cloud Data Breach Heralds Things to Come. (Dec. 2010). https://www.pcworld.com/article/214775/microsoft_
cloud_data_breach_sign_of_future.html.
19
[54] Andrei Sabelfeld and Andrew C Myers. 2003. Language-based information-flow security. IEEE Journal on selected areas in communications
21, 1 (2003), 5–19.
[55] Andrei Sabelfeld and David Sands. 2001. A Per Model of Secure Information Flow in Sequential Programs. Higher Order Symbol. Comput.
14, 1 (March 2001), 59–91.
[56] Peter Sbarski. 2017. Serverless Architectures on AWS: With examples using AWS Lambda. Manning Publications, Shelter Island, NY.
[57] Serverless, Inc. 2017. Serverless Examples. (2017). https://github.com/serverless/examples.
[58] Geoffrey Smith and Dennis Volpano. 1998. Secure Information Flow in a Multi-threaded Imperative Language. In Proc. of POPL 1998. San
Diego, California, USA, 355–364.
[59] Deian Stefan, Alejandro Russo, Pablo Buiras, Amit Levy, John C Mitchell, and David Mazieres. 2012. Addressing covert termination and
timing channels in concurrent information flow systems. In ACM SIGPLAN Notices, Vol. 47. 201–214.
[60] Deian Stefan, Alejandro Russo, John C. Mitchell, and David Mazières. 2011. Flexible Dynamic Information Flow Control in Haskell. In
Proc. of Haskell 2011. 95–106.
[61] TechRepublic. 2017. Massive Amazon S3 leaks highlight user blind spots in enterprise race to the cloud. (July 2017). https://www.
techrepublic.com/article/massive-amazon-s3-breaches-highlight-blind-spots-in-enterprise-race-to-the-cloud/.
[62] The Register. 2011. RSA explains how attackers breached its systems. (April 2011). https://www.theregister.co.uk/2011/04/04/rsa_hack_
howdunnit/.
[63] Tom Van Cutsem and Mark S. Miller. 2013. Trustworthy Proxies: Virtualizing Objects with Invariants. In Proc. of ECOOP 2013. Montpellier,
France, 154–178.
[64] VM2 2017. VM2. (2017). https://github.com/patriksimek/vm2.
[65] Wikipedia. 2017. Anthem medical data breach. (2017). https://en.wikipedia.org/wiki/Anthem_medical_data_breach.
[66] Wikipedia. 2017. Sony Pictures hack. (2017). https://en.wikipedia.org/wiki/Sony_Pictures_hack.
[67] Wikipedia. 2017. Yahoo! data breaches. (2017). https://en.wikipedia.org/wiki/Yahoo!_data_breaches.
[68] Wired. 2016. Inside the Cyberattack That Shocked the US Government. (Oct. 2016).
https://www.wired.com/2016/10/
inside-cyberattack-shocked-us-government/.
[69] Jean Yang, Travis Hance, Thomas H. Austin, Armando Solar-Lezama, Cormac Flanagan, and Stephen Chong. 2016. Precise, Dynamic
Information Flow for Database-backed Applications. In Proc. of PLDI 2016. Santa Barbara, CA, USA.
[70] Matei Zaharia, Mosharaf Chowdhury, Tathagata Das, Ankur Dave, Justin Ma, Murphy McCauley, Michael J. Franklin, Scott Shenker, and
Ion Stoica. 2012. Resilient Distributed Datasets: A Fault-tolerant Abstraction for In-memory Cluster Computing. In Proc. of NSDI 2012.
San Jose, CA.
[71] Stephan Arthur Zdancewic. 2002. Programming languages for information security. Ph.D. thesis, Cornell University.
[72] ZDNet. 2015. Anatomy of the Target data breach: Missed opportunities and lessons learned. (Feb. 2015). http://www.zdnet.com/article/
anatomy-of-the-target-data-breach-missed-opportunities-and-lessons-learned/.
http://www.zdnet.com/article/
[73] ZDNet. 2016. AdultFriendFinder network hack exposes 412 million accounts. (Nov. 2016).
adultfriendfinder-network-hack-exposes-secrets-of-412-million-users.
A
SUMMARY OF AUXILIARY SEMANTIC DETAILS
B
PROOF DETAILS
ϵ is the empty sequence and ++ is concatenation.
label : OutputChannel → Label
run : Thread → Operation
write : LabeledValueSeq × Value × Label →
LabeledValueSeq
write(S, v, l) = (S \ {(v ′, l ′) ∈ S | l ⊑ l ′ }) ++ (v, l)
•1 ≈l •2 means •1 ↓l = •2↓l
Lemma (Invisibility).
If
(σ , ps ⊎
l′ @ l
e
−→ (σ ′, ps ⊎ ps ′)
{(t, l ′)})
20
then
σ ′↓l = σ↓l
e↓l = nop
ps ′↓l = {}
Proof. Omitted.
Lemma (Projection 1).
e1
(1) If Σ −→ Σ1′ ,
e2
then ∃Σ2′ .∃e 2 . Σ↓l −→ Σ2′ and Σ1′ ≈l Σ2′ and e 1′ ≈l e 2′ .
Proof.
Let (σ , ps) = Σ.
Let (σ1′, ps 1′ ) = Σ1′ .
Proceed by cases (ie inversion) on (1).
Case: [s-start]. Let t, l ′ be such that:
e 1 = start (t, l ′); and
σ1′ = σ ; and
ps 1′ = ps + (t, l ′).
Proceed by cases.
Case: where l ′ ⊑ l.
Pick Σ2′ = (σ↓l , ps↓l + (t, l ′)).
Pick e 2′ = start (t, l ′).
QED
by [s-start].
Case: where l ′ @ l.
Pick Σ2′ = Σ↓l and pick e 2′ = ϵ.
QED
by [s-skip].
Case: [s-skip]. e 1 = ϵ and Σ1′ = Σ.
Pick Σ2′ = Σ↓l and pick e 2′ = ϵ.
QED
by [s-skip].
Last case: any other rule. Let t, l ′, ps 1 , ps 2 be such that:
ps = ps 1 ⊎ {(t, l ′)}; and
ps 1′ = ps 1 ⊎ ps 2 ; and
Proceed by cases.
Case: where l ′ @ l.
σ1′↓l = σ↓l and e 1 ↓l = ϵ and ps 2 ↓l = {} by Invisibility.
QED
by [s-skip].
Last case: where l ′ ⊑ l.
Resume case analysis on (1).
Case: [s-send]. Let ch, v, t ′ be such that:
e 1 = output ch v; and
σ1′ = σ ; and
21
ps 2 = {(t ′, l ′)}; and
run(t) = send ch v t ′; and
l ′ ⊑ label(ch).
Pick Σ2′ = (σ↓l , ps 1 ↓l ⊎ {(t ′, l ′)}).
Pick e 2′ = output ch v.
QED
by [s-step].
Remaining cases omitted.
Lemma (Projection 2).
e1
(1) If Σ↓l −→ Σ1 ,
e2
then ∃Σ2 .∃e 2 . Σ −→ Σ2 and Σ1 ≈l Σ2 and e 1 ≈l e 2 .
Proof.
Let (σ , ps) = Σ.
Let (σ1 , ps 1 ) = Σ1 .
Proceed by cases (ie inversion) on (1).
Case: [s-start]. Let t, l ′ be such that:
e 1 = (t, l ′); and
σ1 = σ↓l ; and
ps 1 = ps↓l + (t, l ′); and
Pick Σ2 = (σ , ps + (t, l ′)) and e 2 = e 1 .
QED
by [s-start].
Case: [s-skip].
Pick Σ2 = Σ and e 2 = ϵ.
QED
by [s-skip].
Case: [s-send]. Let t, l ′, ps 3 , ch, v, t ′ be such that:
(2)
ps↓l = ps 3 ⊎ {(t, l ′)}; and
ps 1 = ps 3 ⊎ {(t ′, l ′)}; and
e 1 = output ch v; and
σ1 = σ↓l ; and
run(t) = send ch v t ′; and
l ′ ⊑ label(ch).
l′ ⊑ l
by (2).
ps 3 = (ps − (t, l ′))↓l
by (2).
Pick Σ2 = (σ , ps − (t, l ′) + (t ′, l ′)).
Pick e 2 = e 1 .
QED
by [s-send].
Remaining cases omitted.
22
| 6cs.PL
|
arXiv:1708.00638v1 [math.AG] 2 Aug 2017
ON SUBFINITENESS OF GRADED LINEAR SERIES
Huayi Chen & Hideaki Ikoma
Abstract. — Hilbert’s 14th problem studies the finite generation property
of the intersection of an integral algebra of finite type with a subfield of the
field of fractions of the algebra. It has a negative answer due to the counterexample of Nagata. We show that a subfinite version of Hilbert’s 14th problem
has a confirmative answer. We then establish a graded analogue of this result,
which permits to show that the subfiniteness of graded linear series does not
depend on the function field in which we consider it. Finally, we apply the
subfiniteness result to the study of geometric and arithmetic graded linear
series.
Contents
1. Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2. A weak form of Hilbert’s fourteenth problem. . . . . . . . . . . . . . 7
3. Graded linear series and subfiniteness. . . . . . . . . . . . . . . . . . . . . 9
4. A subfinite version of Zariski’s theorem. . . . . . . . . . . . . . . . . . . . 15
5. Applications. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1. Introduction
Let k be a field and X be an integral projective scheme over Spec k. If D is
a Cartier divisor on X, as a graded linear series of D, one refers to a graded
Huayi Chen is partially supported by ANR-14-CE25-0015, Hideaki Ikoma is supported
by JSPS Grant-in-Aid for Young Scientists (B) No. 16K17559 and partially by JSPS Grantin-Aid (S) No. 16H06335.
2
HUAYI CHEN & HIDEAKI IKOMA
L
0
sub-k-algebra of
n∈N H (X, nD). The graded linear series are closely related to the positivity of the divisor and are objects of central interest in
the study of the geometry of the underlying polarised scheme (X, D). Classically the asymptotic behaviour of graded linear series of finite type is well
understood through the theory of Hilbert polynomials. Several results in birational algebraic geometry, such as Fujita’s approximation theorem [9, 24],
show that certain graded linear series, even though not of finite type, still have
a similar asymptotic behaviour as in the finite generation case. More recently,
Lazarsfeld-Mustaţă [16] and Kaveh-Khovanskii [14, 13] have proposed, after
ideas of Okounkov [22, 23], a method to encode the asymptotic behaviour of
dimensions of the homogeneous components of a given graded linear series into
a convex body (called the Newton-Okounkov body) in an Euclidean space.
Note that a graded linear series of a Cartier divisor is always a graded
subalgebra of a graded algebra of finite type. It is then quite natural to ask
if there is a nice birational geometry for algebras of subfinite type (namely
subalgebras of an algebra of finite type) over a field.
From the point of view of birational geometry, it is more convenient to
consider graded linear series of a finitely generated field extension K/k without
specifying a polarised model of K. In this framework, as a graded linear series
of K/k, we refer to a graded sub-k-algebra V• of the polynomial algebra K[T ]
such that V0 = k and that Vn is a finite dimensional vector space over k for
any n ∈ N. In [6], a new construction of Newton-Okounkov bodies has been
proposed by using ideas from Arakelov geometry, which only depends on a
choice of a tower of successive field extensions k = K0 ⊂ K1 ⊂ · · · ⊂ Kd = K
such that each extension Ki+1 /Ki is transcendental and of transcendence
degree 1. The construction is valid for graded linear series of subfinite type
(namely contained in a graded linear series of finite type of K/k) whose field
of rational functions k(V• ) coincides with K (see Definition 3.1). One may
expect that the same method applies to general graded linear series of subfinite
type V• by considering V• as a graded linear series of k(V• )/k. However, the
main obstruction to this strategy is that a priori the condition of subfiniteness
depends on the extension K/k with respect to which we consider the graded
linear series. This leads to the following subfiniteness problem: given a graded
linear series V• of K/k of subfinite type, does there exist a graded linear series
W• of finite type of the extension k(V• )/k which contains V• ?
Note that the above problem is closely related to Hilbert’s fourteenth problem (1) . In fact, given a graded linear series V• of K/k which is contained in
a graded linear series of finite type V•′ . The intersection of V•′ with k(V• )[T ]
1. Let k be a field and k(x1 , . . . , xn ) be the field of rational functions of n variables.
Hilbert’s fourteenth problem asked whether the intersection of a subfield of k(x1 , . . . , xn )
and the polynomial algebra k[x1 , . . . , xn ] is finitely generated over k (as a k-algebra).
ON SUBFINITENESS OF GRADED LINEAR SERIES
3
gives a graded linear series of k(V• )/k containing V• , where k(V• ) is the field
of rational functions of V• . Unfortunately the intersection is not necessarily a
k-algebra of finite type, as is shown by Nagata’s counterexamples [21, 20] to
Hilbert’s fourteenth problem.
Note that the above subfiniteness problem actually asks for a weaker condition than the finite generation of the intersection of V•′ with k(V• )[T ]. It
suffices that the intersection is contained in a graded linear series of finite type
of k(V• ). Similarly, we can consider the following subfinite version of Hilbert’s
fourteenth problem, which actually has a positive answer (see Theorem 2.6 and
Corollary 2.7 infra).
Theorem 1.1. — Let k be a field, R be an integral k-algebra of finite type and
K be the field of fractions of R. Let K ′ be an extension of k which is contained
in K. Then there exists a finitely generated sub-k-algebra R′ of K ′ containing
R ∩ K ′ , such that Frac(R′ ) = Frac(R ∩ K ′ ).
The method of proof consists of an induction argument with respect to the
field extension K/k which permits to reduce the problem to the case where the
extension K/k is monogenerated. Similar method can be applied to the graded
case (but with more subtleties because of the grading structure), which leads
to the following result and gives a confirmative answer to the subfiniteness
problem of graded linear series. It shows that the subfiniteness of graded
linear series is an absolute condition, which does not depend on the choice of
field extension with respect to which the graded linear series is considered (see
Theorem 3.7 and Corollary 4.11 infra).
Theorem 1.2. — Let k be a field and K/k be a finitely generated field extension. Let V• be a graded linear series of K/k which is of subfinite type. Then
there exists a graded linear series of finite type W• of K/k such that V• ⊂ W•
and k(V• ) = k(W• ).
Recall that Hilbert’s fourteenth problem is reformulated in a geometric
setting by Zariski [25], see also [19] and the survey article [18]. Note that
Theorem 1.1 can be compared with the following result in [25].
Theorem 1.3 (Zariski). — Let k be a field, A an integrally closed k-algebra
of finite type, K := Frac(A), and K ′ /k a subextension of K/k. There then
exist an integrally closed k-algebra B of finite type and an ideal I of B such
that the fraction field of B is k-isomorphic to the fraction field of A ∩ K ′ and
that
[
A ∩ K′ =
(B : I n ),
n∈N
where (B :
I n)
:= {x ∈ Frac(B) :
xI n
⊂ B} denotes the ideal quotient.
4
HUAYI CHEN & HIDEAKI IKOMA
Inspired by this result, we establish the following projective version of
Zariski’s theorem and deduce an alternative proof for Theorem 1.2 (see Corollary 4.11 infra).
Theorem 1.4. — Let K/k be a finitely generated field extension and W• a
graded linear series of K/k that is generated over k by the homogeneous elements of degree 1. We assume that W1 contains 1 ∈ K and that the projective
′
spectrum P := Proj(WL
• ) is a normal scheme. Let K /k be a subextension of
′
′
K/k and W• ∩ K := n∈N (Wn ∩ K ). Then there exist a rational fibration
π : P 99K X of P over an integral normal projective k-scheme X and an
effective Q-Weil divisor D on X having the following properties.
(i) Rat(X) = k(W• ∩ K ′ ).
(ii) Wn ∩ K ′ ⊂ H 0 (X, nD) for every sufficiently positive integer n.
(iii) For every ξ ′ ∈ X with dim OX,ξ ′ = 1, there exists a ξ ∈ P with dim OP,ξ =
1 such that ξ ′ = cX (ordξ |Rat(X) ).
Moreover, if the transcendence degree of K ′ over k is 1, then we can replace
the property (ii) by (ii’) below.
(ii’) Wn ∩ K ′ = H 0 (X, nD) for every sufficiently positive integer n.
As an application of the above subfiniteness results, we establish a Fujita
approximation theorem for general graded linear series of subfinite type (see
Theorem 5.2 infra) and an upper bound for the Hilbert-Samuel function of
such graded linear series (see Theorem 5.3 infra). More precisely, we obtain
the following results.
Theorem 1.5. — Let K/k be a finitely generated field extension. For any
graded linear series V• of K/k of subfinite type, whose Kodaira-Iitaka dimension d is nonnegative, the limit
vol(V• ) =
lim
n∈N, Vn 6={0}, n→+∞
dimk (Vn )
nd /d!
exists in (0, +∞). Moreover, vol(V• ) is equal to the supremum of vol(W• ),
where W• runs over the set of all graded linear series contained in V• having d
as the Kodaira-Iitaka dimension. Finally, there exists a function f : N → R+
such that
nd
f (n) = vol(V• ) + O(nd−1 )
d!
and that dimk (Vn ) 6 f (n) for any n ∈ N.
We also apply the above results to the study of graded linear series in the
arithmetic setting (see Theorem 5.6 infra).
The article is organised as follows. In the second section, we prove a
weaker form of Hilbert’s 14th problem as the subfiniteness result stated in
ON SUBFINITENESS OF GRADED LINEAR SERIES
5
Theorem 1.1. In the third section, we prove a graded analogue of Theorem
1.1 in the setting of graded linear series. In the fourth section we consider
the subfiniteness problem in the geometric setting as a projective analogue
of Zariski’s result and establish Theorem 1.4. Finally in the fifth section, we
develop various applications.
Notation and conventions. —
1. The field of fractions of an integral domain A is denoted by Frac(A).
2. Let K/k be an extension of fields. We denote by tr.degk (K) the transcendence degree of K over k.
3. Let S be a scheme. For any i ∈ N, we denote by S (i) the set of points x
of S such that the local ring OS,x has i as its Krull dimension. If S is an
integral scheme, we denote by Rat(S) the field of rational functions on S.
4. Let k be a field and S be a projective normal scheme over Spec k. As Weil
divisor (resp. Q-Weil divisor ) on S one refers to an element
X
D=
nV V
V ∈S (1)
(1)
(1)
in Z⊕S (resp. Q⊕S ). If all coefficients nV are nonnegative, we say that
D is effective, denoted by D > 0. If φ is a nonzero rational function on S,
we denote by (φ) the principal Weil divisor associated with φ, namely
X
(φ) :=
ordV (φ)V.
V ∈S (1)
(1)
The map (.) : Rat(X)× → Z⊕S is a group homomorphism and induces a
(1)
Q-linear map from Rat(X)× ⊗Z Q to Q⊕S which we denote by (.)Q . If D
is a Q-Weil divisor on S, we define
(1.1)
H 0 (X, D) := φ ∈ Rat(S)× : D + (φ ⊗ 1)Q > 0 ∪ {0}
and
(1.2)
R(D)• :=
M
H 0 (X, nD)T n .
n>0
Note that R(D)• is a graded sub-k-algebra of the polynomial algebra
Rat(S)[T ].
5. Let K/k be a field extension. As discrete valuation of K over k, we refer
to a valuation ν : K → Z ∪ {+∞} such that ν(a) = 0 for any a ∈ k× .
Given such a valuation ν, we denote by Oν := {f ∈ K : ν(f ) > 0} its
valuation ring, mν the maximal ideal of Oν and κ(ν) := Oν /mν the residue
field. Note that κ(ν) is an extension of k and Oν is a k-algebra. If Oν is
equal to K, we say that the valuation ν is trivial (note that in this case
ν(a) = 0 for any a ∈ K × ). Two discrete valuations ν1 and ν2 of K over
HUAYI CHEN & HIDEAKI IKOMA
6
k are said to be equivalent if there exists an order-preserving isomorphism
ι : ν1 (K × ) → ν2 (K × ) such that ν2 = ι ◦ ν1 .
Let K/k be a finitely generated extension of fields. We say that a discrete
valuation ν of K over k is divisorial if
tr.degk (K) = tr.degk (κ(ν)) + rkZ (ν(K × )).
(1.3)
In the case where the valuation is not trivial, the divisorial condition above
is just
tr.degk (K) = tr.degk (κ(ν)) + 1
Let K ′ /k be a subextension of K/k and let ν be a discrete valuation of
K over k which is nontrivial. Then the restriction of ν to K ′ is a discrete
valuation of K ′ over k. We define the ramification index of ν with respect
to K ′ as the unique integer e(K ′ ,ν) ∈ N satisfying
ν(K ′× ) = e(K ′ ,ν)ν(K × ).
(1.4)
6. Let k be a field and S be an integral separated k-scheme. Given a discrete
valuation ν of Rat(S) over k, we say that a point x of S is the centre of ν
in S if
OS,x ⊂ Oν
(1.5)
and mx = mν ∩ OS,x ,
where mx denotes the maximal ideal of OS,x . By the valuative criterion
of separation, if the centre of ν in S exists, then it is unique. In the case
where the centre of ν in S exists, we denote it by cS (ν). If S is proper over
k, then by the valuative criterion of properness every discrete valuation of
Rat(S) over k has a centre in S.
A discrete valuation ν is trivial if and only if the centre of ν in S is
the generic point. Moreover, each regular point ξ ∈ S (1) defines a discrete
valuation L
ordξ : Rat(S) → Z ∪ {+∞} whose centre is ξ.
7. Let R• = n∈N Rn be a graded ring. We denote by Proj(R• ) the projective
f• the quasispectrum of R• . If M• is a graded R• -module, we denote by M
coherent OProj(R• ) -module associated with M• (see [10, §II.2.5]). For any
m ∈ N, we let M (m)• be the N-graded R• -module such that M (m)n =
Mn+m for any n ∈ N, and let M>m be the N-graded sub-R• -module of
M• such that (M>m )n = {0} if n < m and (M>m )n = Mn if n > m. In
^•
particular, one has M (m)• = M>m (m)• . The quasi-coherent sheaf R(m)
is denoted by OProj(R• ) (m). Note that if R• is generated as an R0 -algebra
by R1 , then OProj(R• ) (m) are invertible OProj(R• ) -modules for all m ∈ N,
and one has canonical isomorphisms
OProj(R ) (m) ⊗O
OProj(R ) (m′ ) ∼
= OProj(R ) (m + m′ )
Proj(R• )
•
for all
(m, m′ )
∈
N2 .
•
•
ON SUBFINITENESS OF GRADED LINEAR SERIES
7
L
8. Let R• = n∈N Rn be a graded ring. We say that R• is essentially integral
if the ideal R>1 of R• does not vanish and if the product of two nonzero
homogeneous elements of positive degree of R• is nonzero. Note that if
R• is essentially integral then the scheme Proj(R• ) is integral (see [10,
Proposition II.2.4.4])
2. A weak form of Hilbert’s fourteenth problem
Let k be a field, R be a finitely generated integral k-algebra and K be the
field of fractions of R. Clearly K is a finitely generated extension of k. Let
K ′ be a subextension of K/k, which is also a finitely generated extension (see
[4, Chapitre V, §14, n◦ 7, Corollaire 3]). We consider the intersection R ∩ K ′
and ask the following question which could be considered as a weaker form of
Hilbert’s fourteenth problem: does there exist a finitely generated sub-k-algebra
R′ of K ′ containing R ∩ K ′ such that Frac(R′ ) = Frac(R ∩ K ′ ). In this section,
we give a confirmative answer to this question.
Definition 2.1. — Let k be a field and A be a k-algebra. We say that A is
of subfinite type if it is a sub-k-algebra of a k-algebra of finite type.
Lemma 2.2. — An injective homomorphism of rings A → B yields a dominant morphism Spec(B) → Spec(A).
Proof. — Let p be a minimal prime ideal of A and S := A \ p. Since the
homomorphism of rings A → B is injective, also is the localised homomorphism
Ap → S −1 B. Hence S −1 B is nonzero. In particular, there exists a prime ideal
P of B such that P ∩ S = ∅, or equivalently, P ∩ A ⊂ p. Since P ∩ A is a
prime ideal of A and p is a minimal prime ideal of A, one has P ∩ A = p.
Proposition 2.3. — Let k be a field and A be a k-algebra of subfinite type.
We assume that A is an integral domain. Then there exists a k-algebra of finite
type containing A, which is also an integral domain.
Proof. — Let B be a k-algebra of finite type such that A ⊂ B. By Lemma
2.2, one can find a prime ideal p of B such that p ∩ A = {0}. Since B is a
k-algebra of finite type, also is B/p. The proposition is thus proved.
Lemma 2.4. — Let A be a k-algebra which is an integral domain, and K the
field of fractions of A. Let K ′ /K be a finite extension of K generated by one
element α and B ′ a sub-k-algebra of finite type of K ′ which contains A. Then
there exists a sub-k-algebra of finite type B of K which contains A.
Proof. — Let f ∈ K[T ] be the minimal polynomial of α over K, which
is assume to be monic. Let F1 , . . . , Fn be polynomials in K[T ] such that
8
HUAYI CHEN & HIDEAKI IKOMA
B ′ = k[F1 (α), . . . , Fn (α)]. Let S ⊂ K be the (finite) set of the coefficients of the polynomials f , F1 , . . . , Fn . We claim that A is contained
in k[S]. In fact, suppose that an element u of A is written in the form
ϕ(F1 (α), . . . , Fn (α)), where ϕ ∈ k[X1 , . . . , Xn ], then by Euclidean division
the polynomial ϕ(F1 , . . . , Fn ) ∈ k[S][T ] can be written as f g + u, where g is a
polynomial in k[S][T ]. Therefore, one has u ∈ k[S].
Lemma 2.5. — Let A be a k-algebra which is an integral domain, and K
the field of fractions of A. Let K ′ /K be a purely transcendental extension
of transcendence degree 1 and B ′ a sub-k-algebra of finite type of K ′ which
contains A. Then there exists a sub-k-algebra of finite type B of K which
contains A.
Proof. — Let α ∈ K ′ be a transcendental element over K such that K ′ =
K(α). Assume that B ′ = k[ϕ1 (α), . . . , ϕn (α)], where each ϕi is a rational
function of the form Fi /Gi , where Fi and Gi are polynomials of one variable
with coefficients in K and Gi 6= 0. Let β be an element in the algebraic
closure of the field K such that Gi (β) 6= 0 in K ′ (β) for any i ∈ {1, . . . , n}.
e := k[ϕ1 (β), . . . , ϕn (β)] ⊂ K(β). In fact, if an element
Then one has A ⊂ B
u of A can be written as P (ϕ1 (α), . . . , ϕn (α)), where P is a polynomial with
coefficients in k, then, since α is transcendental over K(β), by considering α
as the variable of rational functions and by specifying its value by β, we obtain
e and
that u = P (ϕ1 (β), . . . , ϕn (β)). Finally, by applying Lemma 2.4 to A ⊂ B
the finite extension K(β)/K, we obtain that there exists a k-algebra of finite
type B ⊂ K such that A ⊂ B.
Theorem 2.6. — Let k be a field and A be a k-algebra of subfinite type. We
assume in addition that A is an integral domain and we denote by K the field
of fractions of A. Then there exists a sub-k-algebra of finite type B of K such
that A ⊂ B.
Proof. — By Proposition 2.3, there exists a k-algebra of finite type B ′ which
is an integral domain containing A. Let K ′ be the field of fractions of B ′ , it
is a finitely generated extension of K. Therefore there exists a sequence of
extensions
K = K0 ( K1 ( . . . ( Kn = K ′
such that each extension Ki /Ki−1 is generated by one element, i ∈ {1, . . . , n}.
The extension Ki /Ki−1 is either generated by an algebraic element over Ki−1
or is purely transcendental of transcendence degree 1. By induction we obtain
that, for any i ∈ {0, . . . , n − 1}, there exists a sub-k-algebra of finite type Bi
of Ki such that Bi ⊃ A. The theorem is thus proved.
Corollary 2.7. — Let k be a field, R be an integral k-algebra of finite type
and K be the field of fractions of R. Let K ′ be an extension of k which is
ON SUBFINITENESS OF GRADED LINEAR SERIES
9
contained in K. Then there exists a finitely generated sub-k-algebra R′ of K ′
containing R ∩ K ′ , such that Frac(R′ ) = Frac(R ∩ K ′ ).
Proof. — By definition, R ∩ K ′ is an integral k-algebra of subfinite type. By
Theorem 2.6, there exists a sub-k-algebra of finite type R′ of Frac(R ∩ K ′ )
such that R ∩ K ′ ⊂ R′ . Clearly one has Frac(R′ ) = Frac(R ∩ K ′ ) since
R ∩ K ′ ⊂ R′ ⊂ Frac(R ∩ K ′ ). The assertion is thus proved.
3. Graded linear series and subfiniteness
Let k be a field and K/k be a finitely generated field extension. Let
M
K[T ] =
KT n
n∈N
be the graded ring of polynomials of one variable with coefficients in K.
Definition 3.1. — As a graded linear series of K/k we refer to a graded
sub-k-algebra
M
V• =
Vn T n
n∈N
of K[T ] such that V0 = k and that Vn is a finite dimensional k-vector subspace
of K for any n ∈ N>1 .
Let V• and V•′ be two graded linear series of K/k. If Vn ⊂ Vn′ for any n ∈ N,
we say that V• is contained in V•′ , or V• contains V•′ , and denote it by V• ⊂ V•′ .
Let V• be a graded linear series of K/k. If V• is finitely generated as a
k-algebra, we say that V• is of finite type. If V• is contained in a graded linear
series of finite type, we say that it is of subfinite type.
Let V• be a graded linear series of K/k. We denote by k(V• ) the subextension
of K/k generated by elements of the form f /g, where f and g are nonzero
elements of K such that there exists n ∈ N>1 with {f, g} ⊂ Vn . The field
k(V• ) is called the field of rational functions of V• .
Lemma 3.2. — Given any graded linear series V• of K/k, one has
k(Vn ) = k(V• )
for every sufficiently positive integer n with Vn 6= {0}, where k(Vn ) denotes the
subextension of K/k generated by the elements of the form f /g with {f, g} ⊂
Vn , g 6= 0.
Proof. — First, we note that if ℓ ∈ N>1 is an index such that Vℓ contains a
nonzero element h, then k(Vm ) ⊂ k(Vm+ℓn ) for any m, n ∈ N>1 . In fact, if
{f, g} ⊂ Vm and g 6= 0, then
f hn
f
=
and {f hn , ghn } ⊂ Vm+ℓn
g
ghn
Misuse of k as
the index since k
is referred to the
base field.
10
HUAYI CHEN & HIDEAKI IKOMA
for any n ∈ N>1 .
By changing the grading of V• , we may assume without loss of generality
that {n ∈ N : Vn 6= {0}} generates Z as a Z-module. There exist integers
{n1 , . . . , nr } ⊂ N>1 and nonzero elements {f1 , . . . , fr , g1 , . . . , gr } ⊂ K such
that {fi , gi } ⊂ Vni for any i ∈ {1, . . . , r} and that k(V• ) = k(f1 /g1 , . . . , fr /gr ).
Set p := lcm(n1 , . . . , nr ). By the above observation, we can assume {fi , gi } ⊂
Vp for any i, and one has
k(V• ) = k(f1 /g1 , . . . , fr /gr ) = k(Vp ).
Moreover, by the hypothesis that {n ∈ N : Vn 6= {0}} generates Z as a Zmodule, we can find a positive integer q such that p and q are coprime and
that k(Vp ) = k(Vq ) = k(V• ).
To conclude the proof, it suffices to show that {pm + qn : m, n ∈ N}
contains every sufficiently positive integer. Since p and q are coprime, we can
fix x, y ∈ Z such that px − qy = 1. Moreover, we can assume that both x and
y are positive. For any r with 0 6 r < q and any n with n > (q − 1)y,
qn + r = prx + q(n − ry) ∈ {pm + qn : m, n ∈ N}.
Hence {pm + qn : m, n ∈ N} contains every integer not less than q(q − 1)y.
Remark 3.3. — Let V• be a graded linear series of K/k and L
f be a nonzero
n
n
element of K. We denote by V• (f ) the graded linear series
n∈N f Vn T ,
n
n
where f Vn := {f g : g ∈ Vn }, called the twist of V• by f . Note that the twist
does not change the field of rational functions: one has k(V• (f )) = k(V• ) for
any f ∈ K \ {0}.
Proposition 3.4. — Let W• be a graded linear series of finite type of K/k.
Let n0 be an integer such that n0 > 1. There exist an integer r > 1 and a family
(fi T ni )ri=1 of homogeneous elements in W• such that the following conditions
are fulfilled:
(1) for any i ∈ {1, . . . , r}, one has ni > n0 ;
(2) for any integer n > n0 , the vector space Wn is generated by elements
of the form f1a1 · · · frar , where a1 , . . . , ar are natural numbers such that
a1 n1 + · · · + ar nr = n.
Proof. — Suppose that W• is generated by W1 T ⊕ · · · ⊕ Wd T d . We claim that
the graded linear series
M
k⊕
Wn T n
n>n0
is generated by Wn0 T n0 ⊕ · · · ⊕ W2n0 +d−2 T 2n0 +d−2 . Let n be an integer such
that n > 2n0 + d − 2. Since W• is generated by W1 T ⊕ · · · ⊕ Wd T d , we obtain
ON SUBFINITENESS OF GRADED LINEAR SERIES
that
X
Wn =
11
W1a1 · · · Wdad .
(a1 ,...,ad )∈Nd
a1 +2a2 +···+dad =n
Let (a1 , . . . , ad ) be an element in Nd such that a1 + 2a2 + · · · + dad = n. Since
n > 2n0 + d − 2, there exist an integer m > 1 and a family
(i)
(i)
(a1 , . . . , ad ) : i ∈ {1, . . . , m}
of elements in Nd such that
∀ j ∈ {1, . . . , d},
∀ i ∈ {1, . . . , m − 1},
(1)
(m)
a j + · · · + aj
(i)
= aj ,
(i)
(i)
n0 6 a1 + 2a2 + · · · + dad 6 n0 + d − 1,
and
(m)
n 0 6 a1
Therefore
Wn =
(m)
+ 2a2
(m)
+ · · · + dad
X
6 2n0 + d − 2.
bn
b2n
+d−2
Wn0 0 · · · W2n00+d−2 ,
(bn0 ,...,b2n0 +d−2 )∈Nn0 +d−1
n0 bn0 +···+(2n0 +d−2)b2n0 +d−2 =n
which concludes the claim (bj corresponds to the number of i ∈ {1, . . . , m}
(i)
(i)
(i)
such that a1 + 2a2 + · · · + dad = j). Finally it suffices to choose a family of homogeneous elements in W• which forms a basis of Wn0 T n0 ⊕ · · · ⊕
W2n0 +d−2 T 2n0 +d−2 .
Lemma 3.5. — Let K/k ′ /k be extensions of fields. We assume that the extension K/k is finitely generated and the extension k′ /k is finite. Let W•′ be a
graded linear series of finite type of K/k′ and let
M
W• = k ⊕
Wn′ T n .
n∈N>1
Then W• is a graded linear series of finite type of K/k.
Proof. — Let (fi T ni )ri=1 be a system of generators of W•′ . Let (θj )m
j=1 be a
′
basis of k over k. We claim that W• is generated by
(3.1)
(θj fi T ni )(i,j)∈{1,...,r}×{1,...,m} .
In fact, if ϕ is an element of Wn′ , then it can be written as
X
λa f1a1 · · · frar ,
a=(a1 ,...,ar )∈Nr
a1 n1 +···+ar nr =n
12
HUAYI CHEN & HIDEAKI IKOMA
where the coefficients λa belong to k′ . By writing λa as a linear combination
of (θj )m
j=1 , we obtain that ϕ lies in the graded linear series of K/k generated
by (3.1). The lemma is thus proved.
Definition 3.6. — Let V• be a graded linear series of K/k. We assume that
there exists n ∈ N>1 such that Vn 6= {0}. We define the Kodaira-Iitaka
dimension of V• as the transcendence degree of k(V• ) over k. We refer the
readers to [13, §3] and [7, §2] for the definition of Kodaira-Iitaka dimension
in the setting of graded linear series of Cartier divisors or line bundles. If
Vn = {0} for any n ∈ N>1 , then by convention the Kodaira-Iitaka dimension
of V• is defined to be −∞.
Theorem 3.7. — Let V• be a graded linear series of K/k. Assume that there
exists a graded linear series of finite type V•′ of K/k which contains V• . Then
there exists a graded linear series of finite type W• of K/k such that V• ⊂ W•
and k(V• ) = k(W• ).
Proof. — Step 1: reduction to the case where 1 ∈ V1 and k(V1′ ) = k(V•′ ). Let
Θ := {n ∈ N>1 : Vn 6= {0}}. The assertion of the theorem is trivial when
Θ = ∅. In the following, we assume that Θ is not empty, and hence it is
a subsemigroup of N
>1 . Let a ∈ N>1 be a generator of the subgroup of Z
L
′ T an is a k-algebra of finite type (see for example
generated by Θ. As n∈N Van
[10, Lemme II.2.1.6.(iv)]), by changing the grading we can reduce the problem
to the case where a = 1. In particular, there exists an m ∈ N>1 such that
the vector spaces Vm and Vm+1 are both nonzero. We pick x ∈ Vm \ {0} and
y ∈ Vm+1 \ {0}. By replacing V• by the graded linear series generated by
V• and (y/x)T and replacing V•′ by the graded linear series generated by V•′
and (y/x)T (this procedure does not change the fields of rational functions),
we reduce the problem to the case where V1 6= {0}. Finally, by replacing V•
by V• (f −1 ) and V•′ by V•′ (f −1 ) (see Remark 3.3 for the notation), where f is
a nonzero element of V1 (again this procedure does not change the fields of
fractions, see Remark 3.3), we reduce the problem to the case where 1 ∈ V1 .
Moreover, by replacing V•′ by the graded linear series generated by V•′ and
α1 T, . . . , αm T , where {α1 , . . . , αm } is a system of generators of k(V•′ ) over k,
we may assume that k(V1′ ) = k(V•′ ).
Step 2: reduction to the simple extension case by induction. As explained in
the previous step, we can assume 1 ∈ V1 and k(V1′ ) = k(V•′ ). Since k(V•′ )/k(V• )
is a finitely generated extension of fields (where V1 is assumed to contain 1),
there exist successive extensions of fields
k(V• ) = K0 ( K1 ( . . . ( Kb = k(V•′ )
such that each extension Ki /Ki−1 is generated by one element of V1′ .
ON SUBFINITENESS OF GRADED LINEAR SERIES
13
Assume that the theorem has been proved for the case where k(V•′ )/k(V• )
is generated by one element in V1′ . Then by induction we can show that, for
any i ∈ {0, . . . , b}, there exists a graded linear series of finite type W•(i) , which
contains V• and such that k(W•(i) ) = Ki . In fact, we can choose W•(r) = V•′ .
Assume that we have chosen a graded linear series of finite type W•(i+1) such
that W•(i+1) ⊃ V• and k(W•(i+1) ) = Ki+1 , where i ∈ {0, . . . , b − 1}. Let V•(i)
be the graded linear series generated by V• and a finite system of generators
(i)
of Ki /k in V1′ . The graded linear series V•(i) contains V• and Ki = k(V1 ).
Without loss of generality we may assume that V•(i) ⊂ W•(i+1) and that the
(i+1)
extension Ki+1 /Ki is generated by one element α in W1
, otherwise we just
(i)
(i+1)
replace W•
by the graded linear series generated by W•(i+1) , V1 and a
generator of the extension Ki+1 /Ki in V1′ . It is a graded linear series of finite
type which contains V• and has Ki+1 as its field of rational functions. If the
theorem has been proved for the simple extension case, then we obtain the
existence of a graded linear series of finite type W•(i) such that V• ⊂ W• and
k(W•(i) ) = Ki .
Note that the graded linear series W• = W•(0) satisfies the conditions V• ⊂
W• and k(V• ) = k(W• ). Therefore, to prove the theorem it suffices to prove the
particular case where the extension k(V•′ )/k(V• ) is generated by one element
in V1′ . Similarly, to prove the theorem under the supplementary condition that
the extension k(V•′ )/k(V• ) is algebraic, it suffices to prove the particular case
where the extension k(V•′ )/k(V• ) is generated by one element in V1′ which is
algebraic over k(V• ).
Step 3: algebraic extension case. In this step, we prove the theorem under
the assumption that the extension k(V•′ )/k(V• ) is algebraic. As explained in
the previous two steps, we may suppose without loss of generality that 1 ∈ V1 ,
k(V1′ ) = k(V•′ ) and the extension k(V•′ )/k(V• ) is generated by one element α
in V1′ which is algebraic over k(V• ).
Let
G(X) := X δ + ξ1 X δ−1 + · · · + ξδ ∈ k(V• )[X]
be the minimal polynomial of α over k(V• ). By Proposition 3.4, there exist
an integer r ∈ N>1 and homogeneous elements (fi T ni )ri=1 with ni > δ for any
i ∈ {1, . . . , r}, which generates the graded linear series
M
k⊕
Vn′ T n .
n>δ
Since 1 ∈ Vn ⊂ Vn′ for any n ∈ N>1 , for any i ∈ {1, . . . , r}, one has fi ∈ k(V•′ ).
Moreover, since the extension k(V•′ )/k(V• ) is generated by α (which is of degree
δ over k(V• )), there exist polynomials
Fi (X) := ηi,1 X δ−1 + · · · + ηi,δ ∈ k(V• )[X],
i ∈ {1, . . . , r}
HUAYI CHEN & HIDEAKI IKOMA
14
such that fi = Fi (α) for any i ∈ {1, . . . , r}. We introduce the following
polynomials in k(V• )[T, Y ]
e
G(T,
Y ) = Y δ + (ξ1 T )Y δ−1 + · · · + ξδ T δ ,
Fei (T, Y ) = (ηi,1 T ni −δ+1 )Y δ−1 + · · · + ηi,δ T ni .
e
Note that one has G(T,
T X) = G(X)T δ and Fe(T, T X) = Fi (X)T ni .
We let W• be the graded linear series generated by V1 T ⊕ · · · ⊕ Vδ−1 T δ−1
and the elements ξ1 T, . . . , ξδ T δ , ηi,1 T ni −δ+1 , . . . , ηi,δ T ni (i ∈ {1, . . . , r}). It is a
graded linear series of finite type of K/k such that k(W• ) ⊂ k(V• ). It remains
to prove that W• contains V• . Clearly Vn ⊂ Wn for n ∈ {1, . . . , δ − 1}. Let
n ∈ N>δ and ϕ be an element in Vn ⊂ Vn′ . By definition ϕ can be written in
the form
X
X
λa f1a1 · · · frar =
λa F1 (α)a1 · · · Fr (α)ar ,
a=(a1 ,...,ar )∈Nr
a1 n1 +···+ar nr =n
a=(a1 ,...,ar )∈Nr
a1 n1 +···+ar nr =n
where λa ∈ k. We consider the element
X
Fe(T, Y ) =
λa Fe1 (T, Y )a1 · · · Fer (T, Y )ar ∈ k(V• )[T, Y ].
a=(a1 ,...,ar )∈Nr
a1 n1 +···+ar nr =n
Viewed as a polynomial on Y with coefficients in k(V• )[T ], the coefficients of Fe(T, Y ) can be written as the values of certain polynomials on
ηi,1 T ni −δ+1 , . . . , ηi,δ T ni (i ∈ {1, . . . , r}). Note that one has
X
Fe(T, T X) =
λa F1 (X)a1 · · · Fr (X)ar T n .
a=(a1 ,...,ar )∈Nr
a1 n1 +···+ar nr =n
Therefore Fe(T, T α) − ϕT n = 0 in k(V•′ )[T ]. Since G is the minimal polynomial
of α, an Euclidean division argument shows that ϕT n can be written as a
polynomial of ξ1 T, . . . , ξδ T δ , ηi,1 T ni −δ+1 , . . . , ηi,δ T ni (i ∈ {1, . . . , r}). The
theorem is thus proved in the particular case where k(V•′ )/k(V• ) is an algebraic
extension.
Step 4: general case. In this step, we prove the theorem in the general case.
As explained in steps 1 and 2, we may assume that 1 ∈ V1 , k(V1′ ) = k(V•′ ) and
that the extension k(V•′ )/k(V• ) is generated by one element α in V1′ which is
transcendental over k(V• ) (the algebraic case has already been treated in Step
3).
Since V•′ is of finite type, there exist an integer r > 1 and homogeneous
elements (fi T ni )ri=1 which generate V•′ as a k-algebra. As k(V•′ )/k(V• ) is
generated by α, there exists rational functions Pi /Qi (i ∈ {1, . . . , r}), where
{Pi , Qi } ⊂ k(V• )[X], Qi 6= 0, such that fi = Pi (α)/Qi (α).
ON SUBFINITENESS OF GRADED LINEAR SERIES
15
Let θ be an element in the algebraic closure k of k, such that Qi (θ) 6= 0
b = K(θ). Then K/K
b
for any i ∈ {1, . . . , r}. Let b
k = k(θ) and K
is a finite
b
b
extension of field, and K/k is a purely transcendental extension generated by
b ] generated by V• and
α. Let Vb• and Vb•′ be the graded sub-b
k-algebra of K[T
′
′
b
b
V• respectively. Then V• is generated as a k-algebra by (fi T ni )ri=1 . We let
c• be the graded linear series of K/
b b
W
k generated by T and elements of the
n
i
form (Pi (θ)/Qi (θ))T , where i ∈ {1, . . . , r}. This is a graded linear series of
finite type. Note that Pi (θ)/Qi (θ) ∈ b
k(Vb• ) for any i ∈ {1, . . . , r}. Therefore
b
b
c
b
k(W• ) ⊂ k(V• ).
Let n ∈ N>1 and ϕ be an element of Vbn ⊂ Vbn′ . By definition ϕ can be written
in the form
X
ϕ=
λa f1a1 · · · frar ,
a=(a1 ,...,ar )∈Nr
a1 n1 +···+ar nr =n
where the coefficients λa belong to b
k. As α is transcendental over b
k(Vb• ), we
obtain that
r
Y
X
Pi (θ) ai
,
ϕ=
λa
Qi (θ)
r
a=(a1 ,...,ar )∈N
a1 n1 +···+ar nr =n
i=1
c• , which implies that
which shows that ϕ ∈ Wn . Therefore one has Vb• ⊂ W
b
b
b
b
c
c
k(V• ) = k(W• ) since we have already seen that k(W• ) ⊂ b
k(Vb• ).
Let
M
cn T n .
W
W ′ := k ⊕
•
n∈N> 1
b by Lemma 3.5 we
c• is a graded linear series of finite type of K/
b k,
Since W
′
b
obtain that W• is a graded linear series of K/k of finite type. Moreover,
c• ) = b
one has V• ⊂ W•′ and k(W•′ ) ⊂ b
k(W
k(Vb• ) is a finite extension of k(V• ).
Therefore, by the algebraic extension case of the theorem proved in Step 3 we
b such that
obtain the existence of a graded linear series of finite type W• of K/k
V• ⊂ W• and that k(V• ) = k(W• ). Moreover, the equality k(V• ) = k(W• ) and
the assumption 1 ∈ V1 ⊂ W1 imply that W• is a graded linear series of k(V• )/k
(and hence a graded linear series of K/k). The theorem is thus proved.
4. A subfinite version of Zariski’s theorem
4.1. Preliminaries. — In this section, we collect several basic facts on the
valuations and on the graded rings, which we use to show Theorem 1.4.
HUAYI CHEN & HIDEAKI IKOMA
16
4.1.1. — Let K/k be a finitely generated field extension. We set
D(K/k)
= {ν : divisorial valuations of K over k}
ν is a discrete valuation of K over k which either is trivial
= ν :
or satisfies ν(K × ) ∼
= Z and tr.degk (κ(ν)) = tr.degk (K) − 1
(see Notation and conventions 5).
Lemma 4.1. — Let K/k be a finitely generated field extension and K ′ /k be
a subextension of K/k. If ν ∈ D(K/k), then the restriction ν ′ := ν|K ′ belongs
to D(K ′ /k).
Proof. — Let ν ∈ D(K/k). The assertion is trivially true if the restriction of
ν to K ′ is a trivial valuation. In the rest of the proof, we assume that ν ′ is not
trivial (and consequently ν is not trivial). By applying [3, Chapitre VI, §10,
n◦ 3, Corollaire 1] to K ′ /k and K/K ′ , one has
(4.1)
tr.degk (κ(ν ′ )) + 1 = tr.degk (κ(ν ′ )) + rk(ν(K ′× )) 6 tr.degk (K ′ )
and
(4.2)
tr.degκ(ν ′ ) (κ(ν)) = tr.degκ(ν ′ ) (κ(ν)) + rk(ν(K × )/ν(K ′× )) 6 tr.degK ′ (K),
respectively. By taking the sum of these two inequalities, one obtains
(4.3)
tr.degk (κ(ν ′ )) + tr.degκ(ν ′ ) (κ(ν)) 6 tr.degk (K ′ ) + tr.degK ′ (K) − 1
= tr.degk (K) − 1.
Since ν belongs to D(K/k), one has
tr.degk (κ(ν)) = tr.degk (κ(ν ′ )) + tr.degκ(ν ′ ) (κ(ν)) = tr.degk (K) − 1,
which means that the inequality in (4.3) is actually an equality. Hence both inequalities (4.1) and (4.2) are equalities. In particular, one has tr.degk (κ(ν ′ )) =
tr.degk (K ′ ) − 1.
Lemma 4.2. — Let π : X → X ′ be a dominant morphism of integral separated k-schemes, K := Rat(X), K ′ := Rat(X ′ ), and ν a discrete valuation of
K/k. If the centre cX (ν) of ν in X exists, then π(cX (ν)) is the center of ν|K ′
in X ′ , namely π(cX (ν)) = cX ′ (ν|K ′ ).
Proof. — Since the morphism π is dominant, it induces an injective homomorphism of fields Rat(X ′ ) → Rat(X), which allows to consider K ′ as a subfield
of K. Recall that the centre cX (ν) is the unique point x ∈ X satisfying
OX,x ⊂ Oν and mx = mν ∩ OX,x (see Notation and conventions 6). Note that
Oν|K ′ = f ∈ K ′ : ν(f ) > 0 = Oν ∩ K ′ , and mν|K ′ = mν ∩ K ′ .
ON SUBFINITENESS OF GRADED LINEAR SERIES
17
Hence OX ′ ,π(cX (ν)) ⊂ Oν|K ′ and mπ(cX (ν)) ⊂ mν|K ′ (which implies mπ(cX (ν)) =
mv|K ′ ∩ OX ′ ,π(cX (ν)) since mπ(cX (ν)) is a maximal ideal).
4.1.2. — Let R• be a graded ring and let P := Proj(R• ). For each homogeneous element a ∈ R>1 , let
f
: deg f = p deg a
(4.4)
(R• )(a) :=
ap
be the degree 0 component of the localisation R• [1/a], and let DProj(R• )+ (a) :=
Spec((R• )(a) ) denote the affine open subscheme of Proj(R• ).
^• (see Notation and conventions 7). Given an s ∈ Rn , the
Set OP (n) := R(n)
local sections s/1 ∈ H 0 (DP + (a), OP (n)) = (R(n)• )(a) for a ∈ R1 glue up to a
global section αn (s) ∈ H 0 (P, OP (n)). The following lemmas are well-known.
Lemma 4.3 ([10, Proposition II.2.7.3]). — Suppose that the irrelevant ideal
R>1 is finitely generated, and let M• be a finitely generated graded R• -module.
f• = 0, then Mn = {0} for any sufficiently positive integer n.
If M
Lemma 4.4. — Let R• be a graded ring and P = Proj(R• ). If R• is essentially integral and is generated as an R0 -algebra by finitely many homogeneous
elements
in R1 , then the canonical homomorphism α• : R• → R(OP (1))• :=
L
0 (P, O (n)) is injective and any element of R(O (1)) is integral over
H
P
P
•
n∈N
R• .
Proof. — Suppose that R• is generated as an R0 -algebra by {a1 , . . . , ar } ⊂
R1 \{0}, where a1 , . . . , ar are all non zerodivisors in R>1 since R• is essentially
integral (see Notation and conventions 8). Given any p ∈ P , one can find an
ai such that ai ∈
/ p; hence (DP + (ai ))i∈{1,...,r} covers P . Thus, a section in
R(OP (1))• can naturally be identified with an element in
r
\
R• [1/ai ],
(4.5)
i=1
where the intersection is taken in R• [1/(a1 . . . ar )]. In particular, α• is injective.
Given any homogeneous element u ∈ R(OP (1))• , one can find an e > 1 such
that aei u ∈ R• for every i by (4.5). Since a1 , . . . , ar generates R>1 , one obtains
R>re u ⊂ R>re . Moreover, by induction,
R>re un ⊂ R>re un−1 ⊂ · · · ⊂ R>re u ⊂ R>re
for every n > 1. It implies that R• [u] ⊂ (1/a1 )re R• ; hence u is integral over
R• (see for example [17, Theorem 9.1]).
Lemma 4.5. — We keep the notation of Lemma 4.4. Suppose that R• is a
Noetherian integral domain and is generated as an R0 -algebra by finitely many
homogeneous elements in R1 .
HUAYI CHEN & HIDEAKI IKOMA
18
(1) If R• is an N -1 ring, then there exists an n0 > 0 such that αn is isomorphic
for every n > n0 .
(2) If R• is an integrally closed domain, then αn is isomorphic for every n > 0.
Proof. — (1) Recall that an integral domain is called an N -1 ring if its integral
closure in its fraction field is a finite generated module over itself. Note
that the graded rings R• and R•′ := R(OP (1))• have the same homogeneous
fraction field, which is the field of rational functions of the scheme Proj(R• ).
In particular, any homogeneous element of R•′ belongs to the homogeneous
fraction field of R• , which is contained in the fraction field of R• . By Lemma
4.4 we obtain that R•′ is contained in the integral closure of R• and hence is a
module of finite type over R• by the Noetherian and N -1 hypotheses.
We consider the exact sequence of OProj(R• ) -modules:
0
^• )
/ Ker(α
f•
/R
α
f•
f′
/R
•
^ •)
/ Coker(α
/ 0.
^• ) =
Since α
f• is isomorphic by [10, Proposition II.2.7.11], we have Ker(α
^ • ) = 0. Hence, by Lemma 4.3, we conclude.
Coker(α
(2) If R• is integrally closed, the above argument actually leads to R• = R•′
since R•′ is contained in the integral closure of R• .
Given a graded ring R• and a positive integer d, the Veronese subring R•(d)
of R• is defined as
(
Rn if d divides n, and
(d)
(4.6)
Rn :=
0
otherwise.
for any n ∈ N.
Lemma 4.6 ([10, Proposition II.2.4.7]). — For any d ∈ N>1 , the natural
inclusion u• : R•(d) → R• induces an isomorphism Proj(u• ) : Proj(R• ) →
Proj(R•(d) ) sending p to p ∩ R•(d) . Moreover, the open subscheme DProj(R• )+ (a)
is isomorphic to DProj(R(d) )+ (ad ) via Proj(u• ) for each homogeneous element
•
a ∈ R>1 .
Lemma 4.7. — If R• is of finite type over R0 , then Proj(R• ) is projective
over Spec(R0 ).
Proof. — By [2, Chapitre III, §1, n◦ 3, Proposition 3] and Lemma 4.6, there
exists a positive integer n0 such that R•(n0 ) = R0 [Rn0 ] and that the natural
morphism Proj(R• ) → Proj(R•(n0 ) ) is isomorphic over Spec(R0 ). Since the
L
(n0 )
′
graded ring R•′ :=
n∈N Rn0 n is finitely generated by R1 as an R0 -algebra,
′
R• can be written as a quotient of a suitable polynomial algebra R0 [T1 , . . . , Tr ]
with deg Ti = 1 for each i. Then Proj(R•′ ) → PrR0 is a closed embedding and
ON SUBFINITENESS OF GRADED LINEAR SERIES
19
Proj(R•′ ) is projective over Spec(R0 ) in the sense of Hartshorne [12, page 103].
(n )
The isomorphisms Rn′ → Rn00n yield an isomorphism (R•′ )(a) → (R•(n0 ) )(a) for
each homogeneous element a ∈ R1′ . Hence, Proj(R•(n0 ) ) → Proj(R•′ ) is an
isomorphism and Proj(R•(n0 ) ) is also projective over Spec(R0 ).
4.2. Proof of Theorem 1.4. — We begin with a reminder on the notation
and the hypotheses of the theorem. Let K/k be a finitely generated field
extension and W• a graded linear series of K/k (see Definition 3.1). We
assume that it is generated as a k-algebra by (a finite number of) homogeneous
elements of degree 1. Moreover, we assume that W1 contains 1 ∈ K and that
the projective spectrum P = Proj(W• ) is a normal scheme. Let K ′ /k be a
subextension of K/k.
Our purpose is to construct a rational fibration π : P 99K X of integral
normal projective k-schemes and an effective Q-Weil divisor D on X satisfying
the properties (i)–(iii) predicted
L in the theorem.
We set W•′ := W• ∩ K ′ = n∈N (Wn ∩ K ′ ) and Q := Proj(W•′ ). Note that
Q may not be of finite type over k; hence, may not be proper over k. Since
W• ∩ K ′ = W• ∩ k(W• ∩ K ′ ), we can assume without loss of generality that
k(W• ∩ K ′ ) = K ′ and k(W• ) = K. We divide the rest of the proof into six
steps.
Step 1: In this step, we give a valuation theoretic interpretation of the
required statement. Let H be the effective Cartier divisor on P defined by the
image of 1 via W1 → H 0 (P, OP (1)). By Lemma 4.5(1), one has
(4.7)
Wn = {φ ∈ K : nH + (φ) > 0}
o
n
= φ ∈ K : ordξ (φ) > −n ordξ (H), ∀ ξ ∈ P (1)
for every sufficiently positive integer n. Therefore,
o
n
(4.8)
Wn′ = φ ∈ K ′ : ordξ |K ′ (φ) > −n ordξ (H), ∀ ξ ∈ P (1)
for every sufficiently positive integer n.
Step 2: In this step, we show the following.
Claim 4.8. — The scheme Q is normal. Moreover, for any ξ ′ ∈ Q(1) , there
exists a ξ ∈ P (1) such that ordξ |K ′ is equivalent to ordξ ′ (see Notation and
conventions 5).
It is a consequence of the following.
Claim 4.9. — Let a be a homogeneous element of degree d > 1 in W•′ , R :=
(W• )(a) and R′ := (W•′ )(a) . Then R′ = R ∩ K ′ .
20
HUAYI CHEN & HIDEAKI IKOMA
Proof of Claim 4.9. — The inclusion R′ ⊂ R ∩ K ′ is obvious. If b is an
element of Wdn such that b/an ∈ K ′ , then b ∈ Wdn ∩ K ′ . Hence we obtain
b/an ∈ R′ .
For each homogeneous element a of positive degree in W•′ , we consider the
morphism
(4.9)
Spec (W• )(a) −→ Spec (W•′ )(a) .
Since P is normal, R := (W• )(a) is integrally closed; hence we have
\
R=
RP .
P∈Spec(R)(1)
By Claim 4.9, we know that
(4.10)
R′ = R ∩ K ′ =
\
RP ∩ K ′
P∈Spec(R)(1)
is a Krull ring; thus in particular, an integrally closed domain. Given any
ξ ′ = p ∈ Spec(R′ )(1) , one can find a P ∈ Spec(R)(1) such that
(4.11)
Rp′ = RP ∩ K ′
(see [17, Theorem 12.3]). To show Claim 4.8, it suffices to take ξ := P. In
fact, if x ∈ R′ \ p, then by (4.11)
a
1
= ∈ RP ∩ K ′
x
y
(a ∈ R, y ∈ R \ P).
Thus, if x ∈ P, then y = ax ∈ P, which is a contradiction. Hence, we know
p ⊃ P ∩ K ′ . Since p has height 1 and P ∩ K ′ 6= {0}, we have p = P ∩ K ′ .
Step 3: Fix an n0 > 1 such that
f
′
: f, g ∈ Wn ∩ K , g 6= 0
= K′
k
g
for every n divisible by n0 (see Lemma 3.2). Let V•0 be the graded sub-kalgebra of W•′ generated by Wn′ 0 and v• : V•0 → W•′ the natural inclusion
map.
Let ν0 : X0 → Proj(V•0 ) be a normalisation. Then X0 is an integral normal
projective k-scheme with Rat(X0 ) = k(V•0 ) = K ′ (Lemma 4.7). Set
0
G(v• ) := p ∈ Q : p 6⊃ v• (V>1
)
(see Notation and conventions 7) and let Proj(v• ) : G(v• ) → Proj(V•0 ) be the
morphism defined by v• . Since Q is normal, Proj(v• ) induces µ : G(v• ) → X0 .
Since
µ−1 ν0−1 (DProj(V•0 )+ (a)) = Spec((W•′ )(a) )
ON SUBFINITENESS OF GRADED LINEAR SERIES
21
0 , µ is an affine morphism ([10, Definition
for each homogeneous element a in V>1
II.1.2.1]).
(1)
Step 4: In this step, we show that, given any η ∈ X0 , one can find an
η ′ ∈ Q(1) such that η ′ ∈ G(v• ) and µ(η ′ ) = η. We suppose that one finds an
(1)
η ∈ X0 such that there is no η ′ ∈ G(v• )(1) with µ(η ′ ) = η, from which we are
going deduce a contradiction.
Since X0 is normal and
[
Proj(V•0 ) =
DProj(V•0 )+ (a),
a: homogeneous
0 and an l belonging to the integral
one can fix a homogeneous element a ∈ V>1
b of A := (V•0 )(a) in K ′ such that {η} is defined by a single nonzero
closure A
b
equation f ∈ A[1/l]
on the affine open subscheme
o
n
b
U := Spec(A[1/l])
= x ∈ ν0−1 (DProj(V•0 )+ (a)) : l(x) 6= 0 .
The hypothesis implies that µ∗ f generates the unit ideal on µ−1 (U ) =
Spec((W•′ )(a) [1/l]). Thus, there exist n1 , p ∈ N>1 , q ∈ N, and g ∈ Wn′ 1 such
that
g
(4.12)
f · p = lq (n1 = p deg a).
a
Let V•1 be the graded sub-k-algebra of W•′ generated by Wn′ 1 over k and
ν1 : X1 → Proj(V•1 ) a normalisation of Proj(V•1 ). Then X1 is also an integral
normal projective k-scheme with Rat(X1 ) = K ′ (see Step 3). Let V•0(n1 ) denote
the Veronese subalgebra of V•0 (see (4.6)). We consider the natural inclusion
homomorphism w• : V•0(n1 ) → V•1 and the commutative diagram
u
X0 o❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ X1
ν0
ν1
Proj(V•0 )
O
?
Spec(A)
∼
Proj(w )
/ Proj(V 0(n1 ) ) o❴ ❴ ❴ • ❴ Proj(V 1 )
O•
O •
?
DProj(V 0(n1 ) )+ (an1 ) o
?
Spec(B),
•
b denote the integral
where we set B := (V•1 )(an1 ) (see also Lemma 4.6). Let B
b
closure of B in K ′ . On the open subscheme u−1 (U ) = Spec(B[1/l]),
the
∗
equation u f is invertible since
f n1 ·
g n1
=1
an1 p ln1 q
HUAYI CHEN & HIDEAKI IKOMA
22
b
by (4.12) and gn1 /(an1 p ln1 q ) ∈ B[1/l].
However, on the other hand, the valuation ordη has a centre η1 on X1 ,
and OX1 ,η1 is a discrete valuation ring. Thus, by the valuative criterion of
properness [10, Corollaire II.7.3.6], η1 is contained in the maximal domain of
definition for u and is mapped to η ∈ U by u. Therefore, η1 ∈ ν1−1 (G(w• ));
hence η1 ∈ u−1 (U ) and u∗ f vanishes at η1 , which leads to a contradiction.
(1)
Step 5: By Claim 4.8 and Step 4, given any η ∈ X0 , there is a ξ ∈ P (1)
such that ordη is equivalent to ordξ |K ′ (see Notation and conventions 5).
Claim 4.10. — The set
n
o
(0)
(1)
/ X0 ∪ X0
Σ := ξ ∈ P (1) : cX0 (ordξ |K ′ ) ∈
is finite. In particular, we obtain a surjective map
(1)
(0)
(P (1) ∪ P (0) ) \ Σ → X0 ∪ X0 ,
ξ 7→ cX0 (ordξ |K ′ ).
Proof of Claim 4.10. — The inclusion K ′ ⊂ K yields a morphism π : U ′ → X0
from a nonempty open subscheme U ′ of P . By the theorem of generic flatness
[11, Théorème IV.6.9.1], there is a nonempty open subscheme U ′′ ⊂ X0 such
that π : π −1 (U ′′ ) → U ′′ is flat. If ξ ∈ π −1 (U ′′ )(1) , then
dim OX0 ,π(ξ) = dim OU ′ ,ξ − dimξ π −1 (π(ξ)) = 0 or 1
(1)
(0)
[12, Proposition III.9.5]. Thus π maps π −1 (U ′′ )(1) into X0 ∪ X0 .
Let eξ denote the ramification index of ordξ (see (1.4)) and set
X
ordξ (H)
(1)
′
: ξ ∈ P , ξ 7→ ξ , eξ 6= 0 {ξ ′ }.
(4.13)
D :=
min
e
ξ
(1)
ξ ′ ∈X0
By (4.8) and Claim
n
Wn′ = φ ∈ K ′ :
n
⊂ φ ∈ K′ :
n
= φ ∈ K′ :
4.10, we have
ordξ |K ′ (φ) > −n ordξ (H), ∀ ξ ∈ P (1)
o
o
ordξ |K ′ (φ) > −n ordξ (H), ∀ ξ ∈ P (1) \ Σ
o
(1)
eξ ordξ ′ (φ) > −n ordξ (H), ∀ ξ ′ ∈ X0
= H 0 (X0 , nD)
for every n ≫ 0.
Step 6: Finally, we consider the case where tr.degk K ′ = 1. In this case,
X is a regular projective curve over Spec k, which (as a set) is canonically in
bijection with D(K ′ /k). Moreover, we have a surjective map
P (1) → X (1) ∪ X (0) ,
ξ 7→ cX (ordξ |Rat(X) ).
ON SUBFINITENESS OF GRADED LINEAR SERIES
Hence, if we set
(4.14)
D :=
X
ξ ′ ∈X (1)
min
23
ordξ (H)
(1)
′
: ξ ∈ P , ξ 7→ ξ , eξ 6= 0 ξ ′ ,
eξ
then, by (4.8), we have Wn ∩ K ′ = H 0 (X, nD) for every sufficiently positive
integer n.
4.3. Rational fibrations associated with graded linear series. — In
the following, we give an alternative proof for Theorem 1.2 by using the
projective version of Zariski’s result (Theorem 1.4).
Corollary 4.11. — Let K/k be a finitely generated field extension and K ′ /k
a subextension of K/k. Let V• be a graded linear series of K ′ /k. If V• is
contained in a graded linear series W• of K/k and of finite type over k, then
V• is contained a graded linear series W•′ of K ′ /k and of finite type over k.
Proof. — We divide the proof into three steps.
Step 1: In this step, we make several reductions of the theorem. By the same
arguments as in the step 1 of Theorem 3.7, we can assume that V1 contains 1.
Claim 4.12. — By enlarging K if necessary, we can assume that W• is generated by W1 over k.
Proof of Claim 4.12. — Let f1 T d1 , . . . , fr T dr ∈ W>1 be homogeneous generators of W• over k. Let T1 , . . . , Tr be variables with deg Ti = 1 for every i. One
can find a homogeneous prime ideal p of W• [T1 , . . . , Tr ] such that p contains
I := (T1d1 − f1 T d1 , . . . , Trdr − fr T dr )
and such that p ∩ V• = {0}. In fact, let W•′ := W• [T1 , . . . , Tr ]/I and let a be
a homogeneous element of degree > 1. Since the morphism Spec (W•′ )(a) →
Spec (V• )(a) is dominant (Lemma 2.2), there exists a homogeneous prime
ideal p ∈ Proj(W•′ ) such that p ∩ V• = {0}. We set U• := W•′ /p. Then U• is a
graded linear series of k(U• )/k, W• → U• is injective, and U• is generated by
U1 = W1 + W0 T1 + · · · + W0 Tr .
In particular, we can assume that P := Proj(W• ) is a projective scheme over
^
k and that OP (1) := W
(1)• is an invertible sheaf on P .
b
Step 2: Let u : P → P be a normalisation and H the Cartier divisor defined
by the image of 1 via V1 → H 0 (Pb , u∗ OP (1)). We choose a very ample divisor
b such that H
b − H is effective and such that R(H)
b • is generated by R(H)
b 1T
H
b
over R(H)0 .
24
HUAYI CHEN & HIDEAKI IKOMA
Note that the graded k-algebra
c• := k ⊕
W
M
n>1
b n
H 0 (Pb, nH)T
is a graded linear series of K/k and of finite type over k (Lemma 3.5) and that
c• ) is isomorphic to Pb over k.
Proj(W
c• and K ′ /k, we can find an integral normal
Applying Theorem 1.4 to W
projective k-scheme X, an effective Q-divisor D on X, and an integer n0 > 1
b n ∩ K ′ ⊂ H 0 (X, nD) for
such that Rat(X) ⊂ K ′ and such that Vn ⊂ R(H)
every n with n > n0 .
b be a very ample divisor on X such that D
b − D is effective
Step 3: Let D
′
b
and such that R(D)• is finitely generated over k. Let W• be the graded linear
series generated by a basis of
M
Vn T n
n<n0
b • over k. Then Wn′ contains
over k and by finite number of generators of R(D)
′
Vn for every n > 0 and W• is finitely generated over k.
Theorem 1.4 is comparable with the existence theorem of Iitaka fibrations
for line bundles on normal projective varieties (see for example [15, Theorem
2.1.33]). As a consequence of Theorem 1.4, we can give an estimate of the
following type for graded linear series of subfinite type (see also [15, Corollary
2.1.38] and Theorem 5.2 infra).
Corollary 4.13. — Let K/k be a finitely generated field extension and V• a
graded linear series of K/k and of subfinite type. Let d be the Kodaira-Iitaka
dimension of V• . If d is nonnegative, then there exist an integral normal projective k-scheme X and Q-Cartier divisors D, D ′ on X such that the rational
function field of X is k-isomorphic to k(V• ), that both D and D ′ have KodairaIitaka dimension d, and that
H 0 (X, nD ′ ) ⊂ Vn ⊂ H 0 (X, nD) ⊂ k(V• )
for every sufficiently positive n with Vn 6= {0}.
Proof. — The existence of D results from the same arguments as in Corollary
4.11. Thus, it suffices to show the existence of D ′ having the prescribed
properties. By changing the grading of V• , we may assume that {n ∈ N :
Vn 6= {0}} generates Z as a Z-module. Choose any sufficiently positive integer
p0 such that k(Vp0 ) = k(V• ) (see Lemma 3.2). Let W• be the sub-k-algebra of
V• generated by Vp0 , and set
M
W•′ :=
Wp 0 n .
n∈N
ON SUBFINITENESS OF GRADED LINEAR SERIES
25
^
′
By Lemma 4.5, Wn′ =
Let P := Proj(W•′ ) and OP (1) := W
• (1).
0
H (P, OP (n)) ⊂ Vp0 n for every n ≫ 1. Let ν : Pb → P be a normalisation. Let p be any sufficiently positive integer divisible by p0 . Then one can
find an ample divisor A on Pb such that
H 0 (Pb, nA) = H 0 (P, ν∗ (OPb (nA))) ⊂ H 0 (P, OP (pn/p0 )) ⊂ Vpn
for every positive integer n (see [5, Démonstration de Proposition 3.6]).
Repeating the same arguments, one can choose an integral normal projective
k-scheme X, two big Cartier divisors A, A′ on X, and two coprime positive
integers p, p′ such that
H 0 (X, nA) ⊂ Vpn and H 0 (X, nA′ ) ⊂ Vp′ n
for any positive integer n. Moreover, one can choose an ample Q-Cartier divisor
D ′ on X and two coprime positive integers q, q ′ such that qq ′ D ′ is integral, that
q (resp. q ′ ) is divisible by p (resp. p′ ), and that
H 0 (X, qnD ′ ) ⊂ H 0 (X, (qn/p)A) ⊂ Vqn
and
H 0 (X, q ′ nD ′ ) ⊂ H 0 (X, (q ′ n/p′ )A) ⊂ Vq′ n
hold for every integer n ∈ N≥1 .
Since
H 0 (X, qnD ′ ) ⊗k H 0 (X, q ′ n′ D ′ ) → H 0 (X, (qn + q ′ n′ )D ′ )
is surjective for any sufficiently positive integers n, n′ (see for example [15,
Example 1.2.22], which is valid over fields of arbitrary characteristics), we
have H 0 (X, nD ′ ) ⊂ Vn for every sufficiently positive n (recall the arguments
in Lemma 3.2).
Theorem 4.14 (Fujita [8, Appendix]). — Let X be an integral normal projective k-scheme and D an effective Cartier divisor on X. If the Kodaira-Iitaka
dimension of D is 1, then the section ring R(D)• is finitely generated.
Proof. — Let K := Rat(X) and let C = D(k(R(D)• )/k) be the smooth
projective k-curve with rational function field k-isomorphic to K ′ := k(R(D)• ).
The inclusion K ′ ⊂ K defines a rational map X 99K C and, by taking a suitable
b → X, one obtains a flat morphism π : X
b → C. If we set
blow-up µ : X
X
ordξ (µ∗ D)
(1)
′
b
: ξ ∈ X , ξ 7→ ξ , eξ 6= 0 ξ ′ ,
E :=
min
e
ξ
(1)
′
ξ ∈C
b nµ∗ D) = H 0 (X, nD) for every positive integer n.
then H 0 (C, nE) = H 0 (X,
Hence the result is reduced to the classic case of curves.
HUAYI CHEN & HIDEAKI IKOMA
26
Remark 4.15. —
1. If X is a surface, Zariski [26] completely classified the
cases where R(D)• is finitely generated [26, Theorem 10.6 and Proposition 11.5]. Later, Fujita [8] generalised the case where the Kodaira-Iitaka
dimension is one to the form of Theorem 4.14 by using the Iitaka fibrations.
2. For a nef and big Cartier divisor D on X, R(D)• is finitely generated if
and only if D is semiample (see [15, Theorem 2.3.15]).
5. Applications
In this section, we apply the subfinite criterion (Theorem 1.2) to the study
of Fujita approximation for general subfinite graded linear series. Throughout
the section, we let k be a field and K/k be a finitely generated k-algebra.
Definition 5.1. — Let V• be a graded linear series of K/k and d be its
Kodaira-Iitaka dimension (see Definition 3.6). If d 6= −∞, we define the volume
of V• as
(5.1)
vol(V• ) := lim sup
n→+∞
dimk (Vn )
.
nd /d!
A priori this invariant takes value in [0, +∞]. We will see below that, if in
addition the graded linear series V• is of subfinite type, then its volume is
always a positive real number.
We say that a graded linear series V• satisfies the Fujita approximation
property if
vol(W• ) = vol(V• ),
sup
W• ⊂V•
W• of finite type
dim(W• )=dim(V• )
where W• runs over the set of all graded linear series of finite type which are
contained in V• and such that W• has the same Kodaira-Iitaka dimension as
V• .
The purpose of the section is to establish the following approximation result.
Theorem 5.2. — Any graded linear series V• of K/k which is of subfinite
type and has nonnegative Kodaira-Iitaka dimension d satisfies the Fujita approximation property. Moreover, one has
vol(V• ) =
dimk (Vn )
∈ (0, +∞),
nd /d!
n∈N(V• ), n→+∞
lim
where N(V• ) = {n ∈ N : Vn 6= {0}}.
ON SUBFINITENESS OF GRADED LINEAR SERIES
27
Proof. — By changing the grading we may assume without loss of generality
that Vn 6= {0} for sufficiently positive integer n. Let K ′ be the homogeneous
fraction field k(V• ). Note that K ′ /k is a subextension of K/k and hence is
finitely generated. Moreover, by Theorem 1.2, we obtain that V• viewed as
a graded linear series of K ′ /k is of subfinite type. Therefore, the assertions
follow from [6, Theorem 1.1] (by definition V• is birational if we consider it as
a graded linear series of K ′ ).
By combining the results of [5] and the subfiniteness result (Theorem 1.2),
we obtain the following upper bound for the Hilbert-Samuel function of general
graded linear series of subfinite type.
Theorem 5.3. — Let V• be a graded linear series of K/k and d its KodairaIitaka dimension. There then exists a function f : N → R+ such that
f (n) = vol(V• )
nd
+ O(nd−1 ),
d!
n → +∞
and
∀ n ∈ N,
dimk (Vn ) 6 f (n).
Remark 5.4. — The result [6, Theorem 1.1] actually provides more geometric information about the graded linear series of subfinite type. Let K/k be a
finitely generated transcendental field extension and let d be the transcendence
degree of K/k. We fix a flag
k = K0 ⊂ K1 ⊂ . . . ⊂ Kd = K
of subfields of K containing k such that each extension Ki /Ki−1 is transcendental and has transcendence degree 1. Let A(K/k) be the set of all graded
linear series of subfinite type V• of K/k such that k(V• ) = k. Then there has
been constructed in [6] a map ∆ from A(K/k) to the set of convex bodies in
Rd which satisfies the following conditions.
(a) If V• and V•′ are two graded linear series in A(K/k) such that V• ⊂ V•′ ,
then one has ∆(V• ) ⊂ ∆(V•′ ).
(b) If V• and W• are two graded linear series in A(K/k), then
∆(V• · W• ) ⊃ ∆(V• ) + ∆(W• ) := {x + y : x ∈ ∆(V• ), y ∈ ∆(W• )},
where V• · W• denotes the graded linear series whose n-th homogeneous
component is the k-vector space generated by {f g : f ∈ Vn , g ∈ Wn }.
(c) For any graded linear series V• in A(K/k), the volume of V• identifies with
the Lebesgue measure of ∆(V• ) multiplied by d!.
This allows us to construct the arithmetic analogue of Newton-Okounkov
bodies for general arithmetic graded linear series of subfinite type, using the
ideas of [1].
HUAYI CHEN & HIDEAKI IKOMA
28
In what follows, we assume that k is a number field. We denote by Mk
the set of all places of k. For each v ∈ Mk , let |.|v be an absolute value on k
which extends either the usual absolute value or certain p-adic absolute value
(so that |p|v = p−1 ) on Q.
As adelic vector bundle on Spec k, we refer to the data V = (V, (k.kv )v∈Mk )
of a finite dimensional vector space V over k and a family of norms k.kv over
V ⊗k kv such that there exists a basis (ei )ri=1 of V over k and a finite subset S
of Mk satisfying the following condition:
∀ v ∈ Mk \ S,
∀ (λ1 , . . . , λr ) ∈ kvr ,
kλ1 e1 + · · · + λr er kv =
max |λi |v .
i∈{1,...,r}
Given an adelic vector bundle V on Spec k, for any nonzero element s ∈ V , we
define the Arakelov degree of s as
X
d
deg(s)
:= −
[kv : Qv ] ln kskv .
v∈Mk
By the product formula
∀ a ∈ k× ,
X
[kv : Qv ] ln |a|v = 0
v∈Mk
we obtain that
d
d
∀ a ∈ k× , deg(as)
= deg(s).
Moreover, the Arakelov degree of V is defined as
X
−
ln kηkv,det ,
v∈Mk
where η is a nonzero element of det(V ), and
kηkv,det = inf{kx1 kv · · · kxr kv : η = x1 ∧ · · · ∧ xr }.
Again by the product formula we obtain that the definition does not depend
on the choice of η ∈ det(V ) \ {0}.
Let V be an adelic vector bundle of rank r on Spec k. For any t ∈ R, let
d
F t (V ) = Vectk ({s ∈ V \ {0} : deg(s)
> t}).
This is a decreasing R-filtration on V , called the R-filtration by minima. Note
that for any i ∈ {1, . . . , r}, the number
λi (V ) = sup{t ∈ R : rkk (F t (V )) > i}
coincides with the minus logarithmic version of the i-th minima in the sense
of Roy and Thunder. For any s ∈ V , we let
λ(s) := sup{t ∈ R : s ∈ F t (V )}.
In the following, we let K/k be a finitely generated field extension of
the number field k. Let V• be a graded linear series of subfinite type of
ON SUBFINITENESS OF GRADED LINEAR SERIES
29
K/k. For each n ∈ N, we equip Vn with a structure of adelic vector bundle (Vn , (k.kn,v )v∈Mk ) on Spec k such that, for any v ∈ Mk ,
(5.2)
∀ (n, m) ∈ N2 , ∀ (sn , sm ) ∈ Vn × Vm ,
ksn · sm kv 6 ksn kv · ksm kv .
We assume in addition that
λmax (V • ) := lim sup
n→+∞
λ1 (V n )
< +∞.
n
This condition implies that V• has a nonnegative Kodaira-Iitaka dimension.
For any t ∈ R, let
M
V•t :=
F nt (Vn ).
n∈N
It is a graded linear series of K/k. By definition one has Vnt = {0} if n ∈ N>1
and t > λmax (V • ).
Proposition 5.5. — For any t < λmax (V • ), one has k(V• ) = k(V•t ).
Proof. — Clearly one has k(V• ) ⊃ k(V•t ). It suffices to prove the converse
inclusion. Let n > 1 be an integer and f , g be nonzero elements in Vn . Since
t < λmax (V• ) there exist m ∈ N>1 and s ∈ Vm such that λ(s) > mt. Thus for
sufficiently positive integer ℓ one has λ(sℓ f ) > (ℓm+n)t and λ(sℓ g) > (ℓm+n)t.
t
Therefore {sℓ f, sℓ g} ⊂ Vℓm+n
, which implies f /g ∈ k(V•t ).
The above proposition allows us to consider V•t as a birational graded linear
series of k(V• )/k and to construct its Newton-Okounkov body as reminded in
Remark 5.4. We define the concave transform of V • as the function GV • on
∆(V• ) sending x ∈ ∆(V• ) to
sup{t < λmax (V • ) : x ∈ ∆(V•t )}.
By the condition (b) in Remark 5.4, the function GV • is concave.
The following result generalises [1, Theorem 2.8] to the case of subfinite
adelically normed graded linear series.
Theorem 5.6. L
— Let K/k be a finitely generated extension of a number field
k, and V • =
n∈N V n a graded linear series of subfinite type of K/k of
Kodaira-Iitaka dimension d > 0, equipped with structures of adelic vector
bundles on Spec k, which satisfy the submultiplicativity condition (5.2) and
the condition λmax (V • ) < +∞. Then the sequence of measures
rkk (Vn )
X
1
δλi (Vn )/n ,
rkk (Vn )
i=1
n ∈ N(V• ) = {m ∈ N : Vm 6= {0}}
30
HUAYI CHEN & HIDEAKI IKOMA
converges weakly to a Boreal probability measure on R, which is the image of
the uniform measure
1
1l
(x) dx
vol(∆(V• )) ∆(V• )
by the concave transform GV • .
Proof. — For any t < λmax (V • ), the graded linear series V•t has the same
homogeneous fraction field as V• (see Proposition 5.5). Hence we can construct
a decreasing family (∆(V•t ))t<λmax (V • ) of convex bodies contained in ∆(V• ), as
described in Remark 5.4. Moreover, if t1 and t2 are two real numbers which are
< λmax (V • ). Then by the same method as in [1, §1.3], we obtain the desired
result.
References
[1] S. Boucksom & H. Chen – “Okounkov bodies of filtered linear series”, Compositio Mathematica 147 (2011), no. 4, p. 1205–1229.
[2] N. Bourbaki – éléments de mathématique. Fascicule XXVIII. Algèbre commutative. Chapitre 3: Graduations, filtra- tions et topologies. Chapitre 4: Idéaux premiers
associés et décomposition primaire, Actualités Scientifiques et Industrielles, No. 1293,
Hermann, Paris, 1961.
[3]
, Éléments de mathématique, Hermann, Paris, 1964, Fasc. XXX. Algèbre
commutative. Chapitres 5 et 6.
[4] N. Bourbaki – Éléments de mathématique, Masson, Paris, 1981, Algèbre.
Chapitres 4 à 7.
[5] H. Chen – “Majorations explicites des fonctions de Hilbert-Samuel géométrique
et arithmétique”, Mathematische Zeitschrift 279 (2015), no. 1-2, p. 99–137.
[6]
,
“Okounkov
bodies:
an
approach
of
function
field
arithmetic”,
preprint
available
at
https://webusers.imj-prg.fr/~huayi.chen/Recherche/okounkov_cdf.pdf,
2017.
[7] S. D. Cutkosky – “Asymptotic multiplicities of graded families of ideals and
linear series”, Advances in Mathematics 264 (2014), p. 55–113.
[8] T. Fujita – “On L-dimension of coherent sheaves”, Journal of the Faculty of
Science. University of Tokyo. Section IA. Mathematics 28 (1981), no. 2, p. 215–236.
[9]
, “Approximating Zariski decomposition of big line bundles”, Kodai Mathematical Journal 17 (1994), no. 1, p. 1–3.
[10] A. Grothendieck & J. Dieudonné – “Éléments de géométrie algébrique. II.
Étude globale élémentaire de quelques classes de morphismes”, Institut des Hautes
Études Scientifiques. Publications Mathématiques 8 (1961), p. 222.
, “Éléments de géométrie algébrique. IV. Étude locale des schémas et des
[11]
morphismes de schémas. II”, Institut des Hautes Études Scientifiques. Publications
Mathématiques (1965), no. 24, p. 231.
ON SUBFINITENESS OF GRADED LINEAR SERIES
31
[12] R. Hartshorne – Algebraic geometry, Springer-Verlag, New York, 1977, Graduate Texts in Mathematics, No. 52.
[13] K. Kaveh & A. G. Khovanskii – “Newton-Okounkov bodies, semigroups of
integral points, graded algebras and intersection theory”, Annals of Mathematics.
Second Series 176 (2012), no. 2, p. 925–978.
[14] K. Kaveh & A. Khovanskii – “Algebraic equations and convex bodies”,
in Perspectives in analysis, geometry, and topology, Progr. Math., vol. 296,
Birkhäuser/Springer, New York, 2012, p. 263–282.
[15] R. Lazarsfeld – Positivity in algebraic geometry. I, Ergebnisse der Mathematik
und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics, vol. 48,
Springer-Verlag, Berlin, 2004, Classical setting: line bundles and linear series.
[16] R. Lazarsfeld & M. Mustaţă – “Convex bodies associated to linear series”,
Annales Scientifiques de l’École Normale Supérieure. Quatrième Série 42 (2009),
no. 5, p. 783–835.
[17] H. Matsumura – Commutative ring theory, second ed., Cambridge Studies
in Advanced Mathematics, vol. 8, Cambridge University Press, Cambridge, 1989,
Translated from the Japanese by M. Reid.
[18] D. Mumford – “Hilbert’s fourteenth problem — the finite generation of subrings
such as rings of invariants”, in Mathematical developments arising from Hilbert
problems (Proc. Sympos. Pure Math., Vol. XXVIII, Northern Illinois Univ., De
Kalb, Ill., 1974), Amer. Math. Soc., Providence, R. I., 1976, p. 431–444.
[19] M. Nagata – “Addition and corrections to my paper “A treatise on the 14th problem of Hilbert” ”, Mem. Coll. Sci. Univ. Kyoto. Ser. A. Math. 30 (1957),
p. 197–200.
, “On the 14-th problem of Hilbert”, American Journal of Mathematics
[20]
81 (1959), p. 766–772.
, “On the fourteenth problem of Hilbert”, in Proc. Internat. Congress
[21]
Math. 1958, Cambridge Univ. Press, New York, 1960, p. 459–462.
[22] A. Okounkov – “Brunn-Minkowski inequality for multiplicities”, Inventiones
Mathematicae 125 (1996), no. 3, p. 405–411.
, “Why would multiplicities be log-concave?”, in The orbit method in
[23]
geometry and physics (Marseille, 2000), Progr. Math., vol. 213, Birkhäuser Boston,
Boston, MA, 2003, p. 329–347.
[24] S. Takagi – “Fujita’s approximation theorem in positive characteristics”, Journal
of Mathematics of Kyoto University 47 (2007), no. 1, p. 179–202.
[25] O. Zariski – “Interprétations algébrico-géométriques du quatorzième problème
de Hilbert”, Bulletin des Sciences Mathématiques. 2e Série 78 (1954), p. 155–168.
[26] O. Zariski – “The theorem of Riemann-Roch for high multiples of an effective
divisor on an algebraic surface”, Annals of Mathematics. Second Series 76 (1962),
p. 560–615.
August 3, 2017
32
HUAYI CHEN & HIDEAKI IKOMA
Huayi Chen, Université Paris Diderot, Institut de Mathématiques de Jussieu - Paris
Rive Gauche, Bâtiment Sophie Germain, Boîte Courrier 7012, 75205 Paris Cedex
13, France. • Beijing International Center for Mathematical Research, Peking
University, Beijing, 100871, China. • E-mail : [email protected]
Url : webusers.imj-prg.fr/~huayi.chen
Hideaki Ikoma, Department of Mathematics, Kyoto University, Kyoto, 606-8502, Japan
E-mail : [email protected]
| 0math.AC
|
On the Achievability Region of Regenerating Codes for Multiple Erasures
arXiv:1802.00104v1 [cs.IT] 31 Jan 2018
Marwen Zorgui, Zhiying Wang
Center for Pervasive Communications and Computing (CPCC)
University of California, Irvine, USA
{mzorgui,zhiying}@uci.edu
Abstract—We study the problem of centralized exact repair of
multiple failures in distributed storage. We describe constructions
that achieve a new set of interior points under exact repair.
The constructions build upon the layered code construction by
Tian et al in [1], designed for exact repair of single failure.
We firstly improve upon the layered construction for general
system parameters. Then, we extend the improved construction
to support the repair of multiple failures, with varying number
of helpers. In particular, we prove the optimality of one point on
the functional repair tradeoff of multiple failures for some parameters. Finally, considering minimum bandwidth cooperative
repair (MBCR) codes as centralized repair codes, we determine
explicitly the best achievable region obtained by space-sharing
among all known points, including the MBCR point.
I. I NTRODUCTION
Driven by the growth of data-centric applications, efficient
data storage and retrieval has become of crucial importance for
several service providers. Distributed storage systems (DSS)
are currently widely employed for large-scale storage. DSS
provide scalable storage and high level of resiliency in the
face of server failures. To maintain the desired level of failure
tolerance, DSS utilize a replacement mechanism for out-ofaccess nodes, known also as the repair mechanism, that allows
to recover the content of inaccessible/failed nodes. The repair
process of a failed node is performed by downloading data
from accessible nodes (or a subset thereof) in the system and
recovering the lost data. Efficiency of a DSS is determined by
two parameters, namely, the overhead required for reliability
and the amount of data being transferred for a repair process.
The seminal work in [2] proposed a new class of erasure
codes, called regenerating codes, that optimally solve the
repair bandwidth problem. It is shown in [2] that one can
significantly reduce the amount of bandwidth required for
repair and the bandwidth decreases as each node stores more
information. Regenerating codes, as presented in [2], achieve
functional repair. In this case, the replacement nodes are
not required to be exact copies of the failed nodes, but the
repaired code should satisfy reliability constraints. However,
in practice, it is often more desirable to recover the exact same
information as the failed node, which is called exact repair.
Exact repair codes are easier to implement and maintain, and
thus are of more interest.
There has been a flurry of interest in designing exact repair
regenerating codes [3]–[11]. Moreover, there is a growing
literature focused on understanding the fundamental limits of
exact repair regenerating codes [12]–[15], as opposed to the
well-understood functional regenerating codes [2].
A. Multi-node recovery
In many practical scenarios, such as in large scale storage
systems, multiple failures are more frequent than a single
failure. Moreover, many systems apply a lazy repair strategy,
which seeks to limit the repair cost of erasure codes. Indeed, it
has been demonstrated that jointly repairing multiple failures
reduces the overall bandwidth compared to repairing each
failure individually [16]–[19]. We distinguish between two
ways of repairing multiple failures.
Cooperative regenerating codes: In this framework, each
replacement node first downloads information from d nodes
(helpers). Then, the replacement nodes exchange information
between themselves before regenerating the lost nodes. Of
interest to our work, we note that codes corresponding to
the extreme points on the cooperative tradeoff have been developed: minimum storage cooperative regenerating (MSCR)
codes [17], [20] and minimum bandwidth cooperative regeneration (MBCR) codes [21].
Centralized regenerating codes: Upon failure of e nodes,
the repair is carried out in a centralized way by contacting
any d helpers out of the n − e available nodes, d ≤ n − e,
and downloading β amount of information from each helper.
The content of any k out of n nodes in the system is
sufficient to reconstruct the entire data. Let α be the size
of each node, and F be the size of the entire data. A code
satisfying the centralized repair constraints is referred to as
an (F, n, k, d, e, α, β) code. We also say it is a code of the
(n, k, d, e) system. In our previous work [19], we characterized
the functional repair tradeoff for multi-node recovery. Let
q = ⌈ ke ⌉ − 1, t = k − qe. The normalized functional tradeoff
can then be written as follows
min(tᾱ, dβ̄) +
q−1
X
p=0
min(eᾱ, (d − t − pe)β̄) ≥ 1,
(1)
where ᾱ = α/F, β̄ = β/F . Inequality (1) gives q linear
bounds:
(t + pe)ᾱ +
q−1
X
i=p
(d − t − ie)β̄ ≥ 1,
p = 0, . . . , q − 1.
(2)
In this work, we are interested in designing centralized exact
repair regenerating codes for recovering multiple failures.
When e ≥ k, the tradeoff reduces to a single point, which
is trivially achievable [19]. We hereafter focus on the case
e < k.
In [18], it is argued that cooperative regenerative codes
can be used to construct centralized repair codes. The total
bandwidth in this case is obtained by taking into account the
bandwidth obtained from the helper nodes and disregarding the
communication between the replacement nodes. In particular,
MSCR codes achieve the same performance as centralized
minimum storage multi-node repair (MSMR) codes [11], [18].
Additionally, MBCR codes can be used as centralized repair
codes, which do not correspond to centralized minimum
bandwidth codes on the functional tradeoff [19]. These points
are given by
e
1
),
(ᾱM SM R , β̄M SM R ) = ( ,
k k(d − k + e)
2e
2d + e − 1
,
).
(ᾱM BCR , β̄M BCR ) = (
k(2d − k + e) k(2d − k + e)
(3)
(4)
Contributions of the paper: We improve upon the layered
construction presented in [1], which is concerned with single
node repair, to construct a family of regenerating codes that
is capable of repairing multiples failures. In particular, for the
(k + e, k, k, e) system, we first prove the optimality of a particular constructed point using the functional repair tradeoff;
combining the achievable points via our construction and also
the MBCR point, we then characterize the best achievable
region obtained by space-sharing between all known points.
The remainder of the paper is organized as follows. A
description of our first code construction is provided in Section
II. In Section III, we analyze the achievability region of
the (k + e, k, k, e) system. We describe our second code
construction in Section IV, before concluding in Section V.
Notation: we denote by [i] the set of integers {1, 2, . . . , i}
for i ≥ 1.
II. C ODE C ONSTRUCTION
Exact repair regenerating codes are characterized by parameters (F, n, k, d, e, α, β). We consider a distributed storage
system with n nodes storing F amount of information. The
data elements are distributed across the n storage nodes such
that each node can store up to α amount of information. We
use ᾱ = α/F, β̄ = β/F to denote the normalized storage size
and repair bandwidth, respectively. The system should satisfy
the following two properties:
• Reconstruction property: by connecting to any k ≤ n nodes
it should be sufficient to reconstruct the entire data.
• Repair property: upon failure of e nodes, a central node is
assumed to contact d helpers, k ≤ d ≤ n − e, and download β
amount of information from each of them. The exact content
of the failed nodes is determined by the central node. β is
called the repair bandwidth.
We first describe the code construction which is an improvement upon [1]. The construction is based on a collection of
subsets of [n], called a Steiner system. Information is first
encoded within each subset, and then distributed among the n
nodes. We recall the definition of Steiner systems.
Definition 1. A Steiner system S(t, r, n), t ≤ r ≤ n, is a
collection of subsets of size r, included in [n], such that any
subset of [n] of size t appears exactly once across all the
subsets.
Steiner systems do not exist for all design parameters. When
t = r, Steiner systems always exist, and the blocks in this
case are all r−combinations of the set [n]. The family of
(F, n, k, d, e, α, β) codes we describe below is parameterized
by t, m, r, for e ≤ m < r ≤ n, t ≤ r, where
n
Nr
t
F = N (r − m), N = r , α =
, k = n − m. (5)
n
t
Construction 1. Precoding step: We consider a Steiner system
(n)
S(t, r, n) and generate N = rt blocks such that each block is
(t)
indexed by a set J ∈ S(t, r, n). Block J corresponds to r − m
information symbols over an alphabet of size q, which is then
encoded using an MSMR code with length r and dimension
r−m over an alphabet of size q. The codeword symbols, called
the repair group J, is comprised of {cx,J : x ∈ J}. Moreover,
we assume that the MSMR code possesses the optimal repair
bandwidth (3) for any number of erasures l, 1 ≤ l ≤ m, and
any number of helpers d, r − m ≤ d ≤ r − l. The total number
of information symbols is F = N (r − m).
The code matrix: The code structure can be described by a
code matrix C, of size n × N . The rows of C are indexed by
integers in [n], corresponding to the different storage nodes,
and its columns are indexed by sets in S(t, r, n), arranged in
some arbitrary chosen order. We formally define C as
(
cx,J , if x ∈ J,
Cx,J =
(6)
−,
otherwise,
where ” − ” denotes an empty symbol. Node i ∈ [n] stores all
the non-empty symbols of row i in the code matrix C. It can
be checked that the storage per node is given by α = Nnr .
By abuse of notation, the terms block and repair group are
used interchangeably. The requirement on the alphabet size q is
dictated by the existence of an MSMR code with the required
property in (3). Such MSMR codes are known to exist [9].
Example 1. Consider a Steiner system S(t, r, n) = S(3, 4, 8).
So the number of blocks is N = 14 and each node number
appears α = rN
n = 7 times in the blocks. The 14 blocks are
given by
J1 = {1, 2, 4, 8}, J2 = {2, 3, 5, 8}, J3 = {3, 4, 6, 8}, J4 = {4, 5, 7, 8}
J5 = {1, 5, 6, 8}, J6 = {2, 6, 7, 8}, J7 = {1, 3, 7, 8}, J8 = {3, 5, 6, 7}
J9 = {1, 4, 6, 7}, J10 = {1, 2, 5, 7}, J11 = {1, 2, 3, 6}
J12 = {2, 3, 4, 7}, J13 = {1, 3, 4, 5}, J14 = {2, 4, 5, 6}.
The code matrix is given by (7) in Figure 1. Let m =
2, e = 2, d = n − e = 6. Then we can repair nodes 1 and
2 simultaneously, by downloading
• symbols c4,J1 , c8,J1 from nodes 4 and 8, respectively. These
help repair symbols c1,J1 and c2,J1 ,
• symbols c5,J10 , c7,J10 from nodes 5 and 7, respectively. These
help repair symbols c1,J10 and c2,J10 ,
• symbols c3,J11 , c6,J11 from nodes 3 and 6, respectively. These
help repair symbols c1,J11 and c2,J11 ,
• 12 symbol from each of the nodes 5, 6 and 8, to repair c1,J5 ,
• 21 symbol from each of the nodes 3, 7 and 8, to repair c1,J7 ,
• 21 symbol from each of the nodes 4, 6 and 7, to repair c1,J9 ,
• 21 symbol from each of the nodes 3, 4 and 5, to repair c1,J13 ,
• and similarly for node 2 to repair c2,J2 , c2,J6 , c2,J12 and
c2,J14 .
From the example above, we see that each repair group
J tolerates the failure of m nodes. Therefore, the code C
also tolerates the failure of up to any m nodes. Thus, it can
be checked that for Construction 1 from any k = n − m
nodes, we can recover the data, which is the reconstruction
parameter. Moreover, the code can recover from any m failures. Therefore, it is possible to repair simultaneously any
1 ≤ e ≤ m failures. The number of helpers is flexible, and
satisfies k ≤ d ≤ n − e. The repair bandwidth is given in
Propositions 1 and 2 for two different scenarios.
Proposition 1. Using Construction 1 with t = r, it is possible
to repair simultaneously any set of 1 ≤ e ≤ m nodes, using
n − m ≤ d ≤ n − e helpers, such that the contribution of each
helper, denoted by βe (d), is given by
βe (d) =
e min(n−d−e+s,r−1)
d − 1 n − d − e
X
X
s
e
. (8)
p
−
s
r
−
p
−
1
m
−
p+s
s
s=1
p=max(s,r−d)
Proof: In the repair procedure, any subset of missing
symbols belonging to the same repair group is repaired via
MSMR repair procedure, using all available helpers from the
same group among the chosen helper nodes. Fixing the set of
helper nodes, we argue that the repair is feasible. Indeed, let
H be the set of d helpers. For each repair group J, we denote
′
′
the set of remaining nodes in J as J . Using |H ∪ J | ≤ n − e
and d ≥ k = n − m, it follows that
′
′
be chosen from the remaining n − e − d nodes. Figure 2
summarizes the repair situation for given parameters s and
p. Summing over all repair contributions, and analyzing the
limit cases of p for a given s, (8) follows.
ĞͲƐ
1,J1
c2,J1
−
c4,J
1
C=
−
−
−
c8,J1
−
c2,J2
c3,J2
−
c5,J2
−
−
c8,J2
−
−
c3,J3
c4,J3
−
c6,J3
−
c8,J3
−
−
−
c4,J4
c5,J4
−
c7,J4
c8,J4
c1,J5
−
−
−
c5,J5
c6,J5
−
c8,J5
c1,J7
−
c3,J7
−
−
−
c7,J7
c8,J7
ŶͲĚͲĞͲƉнƐ
!"#$%&' ()*+,
functional tradeoff
space sharing
Construction 1
Layered code [1]
Exact-repair tradeoff
0.16
(9)
−
c2,J6
−
−
−
c6,J6
c7,J6
c8,J6
ƉͲƐ
(n,k,d,e)=(10,7,7,1)
≥ d + r − e − (n − e) = r + d − n
c
ĚͲƌнƉ
Remark 1. It can be seen that the repair procedure can
benefit from the MSMR repair property in the case n > k + 1.
In particular, the advantages of using MSMR codes in our
construction over maximum distance separable (MDS) codes
as in [1] are: 1) lower repair bandwidth, 2) symmetric repair
among helper nodes, which obviates the need for the expensive
procedure of duplicating the block design as in [1], and 3)
adaptability, meaning non-trivial repair strategies for multiple
erasures, 1 ≤ e ≤ m with the help of varying number of
helpers d, such that n − m ≤ d ≤ n − e. Figure 3 shows a
comparison between the performance of the layered code in
[1] and Construction 1, for an (n, k, d) = (10, 7, 7) system.
The MSR repair property clearly helps reduce the bandwidth.
|J ∩ H| = |H| + |J | − |H ∪ J |
Thus, for each repair group, we have enough information
across the set of helpers to recover the missing components.
We now analyze the contribution of a single helper h: h helps
in the simultaneous repair of s missing symbols of the same
repair group, such that 1 ≤ s ≤ e. For each size s, we count
all possible cases in which the repair can be done through
the help of r − p coded symbols among all the d helpers,
because the number of available coded symbols determines the
contribution of each helper, as dictated by the MSMR repair
bandwidth (3). It follows that, for the corresponding repair
group, r − p − 1 can be chosen from the set of d − 1 helpers
(helper h already belongs to the repair group by assumption),
while the remaining p − s elements of the repair group can
ƌͲƉ
Fig. 2: A repair situation associated to given parameters s and p.
′
≥ r − n + n − m = r − m.
Ɛ
Normalized bandwidth per helper
In total, we download 18 symbols. Each helper transmits 3
symbols.
0.14
0.12
0.1
0.08
0.06
0.04
0.14
0.16
0.18
0.2
0.22
0.24
0.26
Normalized storage per node
Fig. 3: Using the MSR repair property improves upon the
layered code repair performance.
The technique of using MSMR codes as building blocks for
outer code constructions has been used in the literature, for
instance in constructing codes with local regeneration [22],
[23].
−
−
c3,J8
−
c5,J8
c6,J8
c7,J8
−
c1,J9
−
−
c4,J9
−
c6,J9
c7,J9
−
c1,J10
c2,J10
−
−
c5,J10
−
c7,J10
−
c1,J11
c2,J11
c3,J11
−
−
c6,J11
−
−
−
c2,J12
c3,J12
c4,J12
−
−
c7,J12
−
Fig. 1: Code matrix for the system with parameters (t, r, n) = (3, 4, 8).
c1,J13
−
c3,J13
c4,J13
c5,J13
−
−
−
−
c2,J14
−
c4,J14
c5,J14
c6,J14
−
−
.
(7)
Remark 2. We argue that one can use a regenerating code
corresponding to an interior point instead of an MSMR
code as the inner code per repair group. Consider the case
(n, k, d, e) = (5, 4, 4, 1). Let r = t = 5, m = 1 in (5). The
code structure is given by
−
c1,J2 c1,J3 c1,J4 c1,J5
c2,J1
−
c2,J3 c2,J4 c2,J5
(10)
−
c3,J4 c3,J5
C = c3,J1 c3,J2
.
c4,J1 c4,J2 c4,J3
−
c4,J5
−
c5,J1 c5,J2 c5,J3 c5,J4
Thus, the code per column of C is of length r = 4 and dimension r − m = 3. We use the interior code: (ᾱ0 , β̄0 ) = ( 83 , 14 )
per repair group. Let F0 be the information size per column.
3
Thus, F = 5F0 and α = 3F2 0 . It follows that ᾱ = 10
. To
repair node 1, we download a total bandwidth of 3F0 . Thus,
3
3
3
. We obtain the achievable point (ᾱ, β̄) = ( 10
, 20
).
β̄ = 20
The same point is equally achievable using Construction 1
with (t, r, n, m, e) = (3, 3, 5, 1, 1) with an MSMR code as the
interior code. This point is optimal on the exact-repair tradeoff
of the (5, 4, 4, 1) system [10], [24], and is the optimal point
next to the minimum bandwidth regenerating point.
In Proposition 1, we considered Construction 1 with Steiner
systems such that t = r. We study next the use of a general
Steiner system for the specific (k + 2, k, k, 2) system.
Proposition 2. Construction 1 generates an (F, k +
2, k, k, 2, α, β) code such that
n−2
n−1
n
t−2
t−1
t
(11)
F = (r − 2) r , α = r−1 , β = r−2 ,
t
ᾱ =
t−1
t−2
r
r(r − 1)
, β̄ =
.
n(r − 2)
n(n − 1)(r − 2)
(12)
Proof: We consider a Steiner system S(t, r, n) and let
m = 2. From (5), we obtain F and α as in (11). To analyze
the repair bandwidth per helper, we distinguish two cases:
Case t = 2: If the helper node h shares a block with both failed
nodes, then, by design, h does not share any other block with
either of the failed nodes. Thus, h contributes a single symbol
(log2 q bits) that is useful for the repair of the missing symbols
of the shared repair group. Otherwise, h shares exclusively two
blocks with each of the failed nodes. In each of the shared
repair group, node h contributes 12 symbol ( 21 log2 q bits) to
help repair the corresponding missing symbol, by virtue of the
MSMR repair property (i.e., the missing symbol is repaired
with r − 1 helpers).
Case t ≥ 3: For a helper h, the number of blocks he shares
(n−3
t−3 )
. The number
with both failed nodes is given by λ3 , r−3
( t−3 )
of blocks node h shares exclusively with either of the failed
(n−2
(n−3
t−2 )
t−3 )
nodes is given by λ2 − λ3 , r−2
− r−3
. Therefore, the
( t−2 )
( t−3 )
contribution of each helper node is
2(λ2 − λ3 )
2λ3
+
= λ2 .
β=
(r − 2) − (r − 2) + 2 (r − 1) − (r − 2) + 1
The repair in Example 1 is an illustration of Proposition
2. Similar to Proposition 1, the repair bandwidth is identical
among the helper nodes, and independent of the choice of the
failed nodes and helpers.
Remark 3. We note here that ᾱ, β̄ do not depend on t by
(12). The advantage of using Steiner systems with smaller
t, whenever they exist, is that they induce smaller α and β,
for the same normalized parameters. Indeed, it can be shown
that α, as given by (11), is strictly increasing in t. Therefore,
to reduce the storage size per node, and therefore the repair
bandwidth, it is advantageous to use a Steiner System with the
smallest t, t ≤ r. Moreover, when e = 2, t = r, Proposition 1
and Proposition 2 give the same ᾱ, β̄.
III. A NALYSIS OF THE ACHIEVABILITY FOR AN
(n, k, d, e) = (k + e, k, k, e) SYSTEM
In this section, we analyze the achievable region for an
(n, k, d, e) = (k + e, k, k, e) system by means of Construction
1, using, for simplicity, a Steiner system with t = r.
Proposition 3. Construction 1 with t = r, m = e generates a
set of achievable points for an (F, k + e, k, k, e, α, β) system,
such that
F =
!
k+e
(r − e), α =
r
!
k+e−1
,β =
r−1
!
k+e−2
,
r−2
(13)
r(r − 1)
r
, β̄ =
,
(k + e)(r − e)
(k + e)(k + e − 1)(r − e)
e + 1 ≤ r ≤ k + e.
ᾱ =
(14)
Proof: When d = k, n = k + e, m = e, r is chosen such
that e + 1 ≤ r ≤ n, the general expression in (8) is given by
βe (k) =
e
X
t=max(1,r−k)
!
!
k−1
e−1
=
r−t−1
t−1
!
k+e−2
,
r−2
(15)
where the last equality follows from Vandermonde’s identity.
A. Optimality of one achievable point
Proposition 4. For the (k + e, k, k, e) system, the point
achieved in (13) for r = k + e − 1 is an optimal interior
point.
Proof: From (13) when r = k + e − 1, we achieve F =
(k + e)(k − 1), α = k + e − 1, β = k + e − 2. Thus,
(ᾱ, β̄) = (
k+e−1
k+e−2
,
).
(k + e)(k − 1) (k + e)(k − 1)
(16)
Substituting (16) in (2) and setting p = q − 1, we obtain
(t + qe − e)ᾱ + (d − t − qe + e)β̄ = (k − e)ᾱ + eβ̄ = 1.
Therefore, the above point lies on the functional repair lower
bound and hence is optimal. It lies on the first segment of the
bound near the MSMR point, and it is not the MSMR point
nor the MBCR point, as indicated by (3) and (4).
B. Optimal extension property
Proposition 5. Consider a (k +e, k, k, e) system and consider
the optimal point achieved by Construction 1 in Proposition 4,
one can extend the system to a (k + e + 1, k, k, e + 1) system,
operating at the optimal point of Proposition 4, by adding
another node to the system and increasing the storage per
node, while keeping the initial storage content.
Proof: Let αi , βi , Fi , for i = 1, 2, refer to the parameters
of the old and the new systems, respectively. Then, α2 − α1 =
1, β2 − β1 = 1, F2 − F1 = k − 1. Moreover, the number of
blocks N is increased by 1. Let k + e + 1 be the index of the
new node to be added. The new code is obtained by simply
adding another block, whose set is {1, . . . , k + e}, and adding
to the old sets the element (k + e + 1) to each of them, and
thus generating another coded symbol for the corresponding
repair group. A key requirement is to assume the use of an
MSMR code that can accommodate the addition of extra coded
symbols, when needed. This can be done by choosing the
number of nodes of the MSMR code to be as large as needed
(this may result in an increase in the underlying field size).
Each old node will store an extra symbol coming from the new
repair group, while the new node stores the newly generated
coded symbols from the old repair groups.
Normalized bandwidth per helper
From Proposition 4, Construction 1 gives us an optimal
point for any (k + e, k, k, e) system. Construction 1 also offers
the following property.
(n,k,d,e)=(17,14,14,3)
0.07
functional tradeoff
space sharing :MBCR and MSMR
Construction 1
space sharing MBCR and r=8
space sharing MBCR and r=9
0.065
0.06
0.055
0.05
r=8
0.045
r=6
0.04
0.035
0.03
0.025
0.07
0.08
0.09
0.1
0.11
0.12
0.13
0.14
0.15
0.16
Normalized storage per node
Fig. 4: Achievable points for an (n, k, d, e) = (17, 14, 14, 3)
system. The x-axis is the normalized storage per node ᾱ and
the y-axis is the normalized bandwidth β̄.
of Proposition 5, we can increase the fault tolerance of the
system by any desirable factor.
C. Acheivability region for the (k + e, k, k, e) system
In this subsection, we seek to determine the convex hull of
the known achievable points for the (k + e, k, k, e) system.
The convex hull, denoted by R, is the smallest convex set
Example 2. We illustrate the process of extending a (4, 3, 3, 1) containing all known achievable points, obtained by all convex
system to a (5, 3, 3, 2) system. Initially, each repair group is combinations (i.e., space-sharing) among the points achieved
by Construction 1, described in (14), and also the MBCR point
of size 3. The code blocks are given by
given by (4). The objective is therefore to determine which
J1 = {2, 3, 4}, J2 = {1, 3, 4}, J3 = {1, 2, 4}, J4 = {1, 2, 3}.
points are sufficient to describe R. We refer to these points as
corner points of R.
The code matrix is given by
Figure 4 presents the achievable points for an (17, 14, 14, 3)
−
c1,J2 c1,J3 c1,J4
system.
The achievable points of (13) are parameterized by r,
−
c2,J3 c2,J4
c
.
C1 = 2,J1
such that e + 1 ≤ r ≤ e + k. For each r, we denote the
−
c3,J4
c3,J1 c3,J2
corresponding point as (ᾱr , β̄r ). As r decreases, the storage
−
c4,J1 c4,J2 c4,J3
Adding node 5 to the system, we add another block J5 = αr increases. By abuse of notation, we refer to the point
{1, 2, 3, 4}, whose symbols will be distributed across the old (ᾱr , β̄r ) as point r. We state some guiding observations for
our subsequent analysis. First, one can eliminate some of the
nodes {1, 2, 3, 4}. The old blocks become
achievable points obtained by Construction 1. For instance,
J1 = {2, 3, 4, 5}, J2 = {1, 3, 4, 5}, J3 = {1, 2, 4, 5}, J4 = {1, 2, 3, 5}. point r = 5, with ᾱ = 0.1471, achieves a similar bandwidth
as the neighbor point r = 6, but at a larger storage size. Points
The new node 5 stores newly generated coded symbols of each
to the right of ᾱ = 0.1471, such that r < 5, can be also
of the old repair groups {J1 , . . . , J4 }. The new code matrix
immediately eliminated, because they can be outperformed
is given by
by space-sharing between the MBCR point and some interior
point. Interestingly, we observe that point r = 8 lies exactly
−
c1,J2 c1,J3 c1,J4 c1,J5
c2,J1
c
c
−
c
on the segment joining point r = 9 and the MBCR point. This
2,J5
2,J4
2,J3
means
that, while point r = 8 is not outperformed by spacec
−
c
c
c
.
C2 =
3,J5
3,J4
3,J2
3,J1
c4,J1 c4,J2 c4,J3
−
c4,J5
sharing, it is nonetheless not necessary for the description of
−
c5,J1 c5,J2 c5,J3 c5,J4
R, and thus it is not considered as a corner point. In the
following, we show that the observations from Figure 4 can
The above property is useful for systems for which the
be generalized and we explicitly determine the corner points
fault tolerance may be deemed insufficient. Therefore, one can
of R, depending on the system’s parameters e and k.
increase the fault tolerance of the system without sacrificing
the optimality on the exact repair tradeoff, or changing the Lemma 1. The achievable points in (13), with r < 2e, are
existing data. We note also that by a successive application not corner points in R.
Proof: From (13), it can be seen that ᾱ(r), seen as a
function of r, is decreasing. β̄(r) is a fractional function in r,
with a pole at r = e. For r > e, β̄(r) is convex in r. It can
be shown that it decreases and then increases monotonically.
Therefore, as ᾱ(r) is decreasing, the points of interest are
those for which β̄(r) increases. Moreover, by noticing that
β̄(2e) = β̄(2e − 1), it follows that points with r ≤ 2e − 1
do not contribute to the acheivability region (ᾱ, β̄), as these
points are outperformed by the point r = 2e in terms of both,
storage and bandwidth.
Lemma 1 implies that it is sufficient to consider the range
2e ≤ r ≤ k + e. We define the non-negative integer p such
that r = 2e + p. We now show that the achievable points
r = 2e, . . . , k + e can not be eliminated by space-sharing
between themselves, when not considering the MBCR point.
Lemma 2. The achievability region of the points (ᾱr , β̄r ), r =
e + 1, . . . , k + e has points with r ∈ {2e, . . . , k + e} as corner
points, when not considering the MBCR point.
Proof: By virtue of Lemma 1, points with e + 1 ≤ r < 2e
can be eliminated. We consider the segment joining the points
(ᾱr , β̄r ) and (ᾱr+1 , β̄r+1 ). The slope of the segment, denoted
sl(r), is given by
−r(r − 2e + 1)
β̄r+1 − β̄r
=
.
sl(r) =
ᾱr+1 − ᾱr
e(k + e − 1)
The slope sl(r) is strictly decreasing in r for r ≥ 2e.
This means, for any three consecutive points (ᾱr+2 , β̄r+2 ),
(ᾱr+1 , β̄r+1 ) and (ᾱr , β̄r ), the point (ᾱr+1 , β̄r+1 ) lies below
the segment joining the other two extreme points. Therefore,
space-sharing between (ᾱr+2 , β̄r+2 ) and (ᾱr , β̄r ) is suboptimal.
Now, we analyze the achievability region when adjoining
the MBCR point to the points in (13) with 2e ≤ r ≤ k + e.
Lemma 3. The MBCR point is a corner point for R .
2k+e−1
2
(k+e)k > ᾱ2e = (k+e) and
2(2e−1)
(k+e)(k+e−1) , along with Lemma 2
Proof: Noting that ᾱMBCR =
2e
(k+e)k
β̄MBCR =
< β̄2e =
concludes the result.
By Lemma 3, we only need to analyze whether spacesharing between the MBCR point and any other point r may
′
outperform some of the other achievable points r .
Lemma 4. If a point r, r ≥ 2e, is not outperformed by
space-sharing between the point r + 1 and the MBCR point,
then, all points r′ such that r′ ≥ r, are corner points of the
achievability region.
Proof: The assumption of the lemma implies that
the slope of the segment joining the points (ᾱr , β̄r ) and
(ᾱMBCR , β̄MBCR) is smaller than the slope of the segment between (ᾱr+1 , β̄r+1 ) and (ᾱMBCR , β̄MBCR ). As from Lemma 2,
the slope of the segment between (ᾱr , β̄r ) and (ᾱr+1 , β̄r+1 )
is decreasing in r, it follows that no point r′ ≥ r can be outperformed by space-sharing across any two other achievable
points, including the MBCR point.
Therefore, to determine the corner points of R, we need to
successively test for increasing values of p, such that 0 ≤ p ≤
k − e, whether the point r = 2e + p is outperformed by spacesharing of MBCR and point r + 1. Let p∗ denote the smallest
p such that r = 2e + p is not outperformed by space-sharing,
it follows by Lemma 4 the following achievability region.
Proposition 6. The achievability region R is given by the
corner points
R = {(ᾱr , β̄r ) : r ∈ {r : r = 2e + p and p∗ ≤ p ≤ k − e} ∪ {MBCR}},
(17)
where 1 ≤ p∗ ≤ k − e, and p∗ is given by
p∗ =
e − k − 2e2 + 1 +
2(e + k − 1)
√
∆
+ 1,
∆ = (2e2 − e + k − 1)2 + 8(k + e − 1)e(e − 1)(k − e − 1).
(18)
Proof: Consider r = 2e + p, 0 ≤ p ≤ k − e − 1. We
consider space-sharing between the MBCR point and the point
r + 1. We compute the normalized bandwidth, denoted by β¯′ r ,
achieved by the considered space-sharing, at the intermediate
point α = αr , and then determine whether β¯′ r > β̄r . Using
(14) and (4), we obtain after simplification
β¯′ r − β̄r
k(−2e2 + 2e + p2 + p) − p(−2e2 + e + 1) + 2e(e2 − 1) + p2 (e − 1)
(e + k)(e + p)(e + k − 1)(e2 + pe + k − p + kp − 1)
N1 (k)
(19)
,
D
2
2
(k + e − 1)p + p(2e + k − e − 1) + 2e(e − 1)(e + 1 − k)
=
(e + k)(e + p)(e + k − 1)(e2 + pe + k − p + kp − 1)
N2 (p)
.
(20)
,
D
=
We regard N1 as a function of k, for fixed e and p, and
N2 as a function of p, for fixed e and k. In this proof, we are
interested in analyzing N2 . We analyze N1 in a later proof.
Clearly D > 0. Thus, sign(β¯′ r − β̄r ) = sign(N2 (p)).
Therefore, it suffices to study the sign of N2 (p). We note that
β¯′ r − β̄r ≤ 0 implies that point r = 2e + p can be eliminated
by space-sharing and thus it is a not a corner point. N2 (p) is
a quadratic function in p. Let ∆ denote the discriminant of
N2 (p). It can be checked that
∆ = (2e2 − e + k − 1)2 + 8(k + e − 1)e(e − 1)(k − e − 1) > 0.
Thus, there exists p0,1 , p0,2 such that N2 (p0,1 ) = N2 (p0,2 ) =
0. As the leading coefficient of N2 (p) is positive, and N2 (0) =
−2e(e − 1)(k − e − 1) ≤ 0, it follows that one solution, say
p0,1 , is negative and the other solution p0,2 is non-negative.
That is, p0,1 < 0 and p0,2 ≥ 0. Then, it follows that ∀0 ≤
p ≤ p0,2 , N2 (p) ≤ 0, which implies that the set {p : p ≤ p0,2 }
can be eliminated. In particular, p = 0 is always eliminated.
Let p∗ = ⌊p0,2 ⌋ + 1, as in (18). Thus, p∗ outperforms spacesharing and so do all p ≥ p∗ . As N2 (k − e − 1) = (k − e)(k +
e − 1)(k − e − 1) ≥ 0, it follows that p0,2 ≤ k − e − 1, and
thus p∗ ≤ k − e.
Proposition 6 agrees with known particular cases. 1) When
e = 1, we have p∗ = 1 and the only eliminated point (p = 0)
coincides with the MBCR point, in agreement with [1]. 2)
The optimal point in Proposition 4 (p = k − e − 1) is not a
corner point for k = e + 1, because of p∗ = k − e > p and
Proposition 6. Indeed, the point with p = k − e − 1 lies exactly
on the segment joining the MBCR and the MSMR point. 3)
When k > e + 1, the optimal point in Proposition 4 is a corner
point, as β¯′ k+e−1 − β̄k+e−1 = (k−e)(k−e−1)
k(k+e)(k−1)2 > 0.
While Proposition 6 characterizes exactly R, it does not
give insight into when a particular point r = 2e + p is a corner
point or not. We focus on the analysis of the sign of N1 (k)
in (19). N1 (k) is linear in k. Depending on the sign of its the
leading coefficient −2e2 + 2e + p2 + p, there may exist an
integer kth such that when k ≥ kth space sharing enhances the
achievability region (i.e., N1 (k) ≤ 0 ) and does not enhance
it when k < kth . That is, a point with the same r may be a
corner point for some (k + e, k, k, e) systems and may be not
a corner point for other systems, with higher reconstruction
parameter k.
For example, when p = e − 1, we have N1 (k) = e(1 −
e)(k − 5e + 1). It follows that, for systems with k ≥ 5e − 1,
the point r = 2e + (e − 1) = 3e − 1 is outperformed by spacesharing. For systems with 3e − 1 ≤ k < 5e − 1, the point r is
a corner point.
The next proposition addresses the cases in which a particular point r = 2e+p is a corner point, using a similar argument
as the above example.
Proposition 7. Consider the achievable
point r = 2e + kp, for
j p
fixed (e, k), e > 1. Let pmax = 12 ( 8e(e − 1) − 1 − 1) and
e+1
(p+1
2 )+2( 2 )+ep
kth = (1 − e)
. Then, Table I specifies the
e
(p+1
2 )−2(2)
scenarios in which (ᾱr , β̄r ) is a corner point in R.
(ᾱr , β̄r )
k < kth (p)
k ≥ kth (p)
p ≤ pmax
X
✗
p > pmax
cannot be 0, for e > 0, ∀p ≥ 0. The maximum value of p
satisfying (23) is given by
1 p
( 8e(e − 1) − 1 − 1) .
pmax =
(24)
2
Thus, a point r = 2e + p, p > pmax is a corner point for
any (k + e, k, k, e) system such that r ≤ k + e. For each
0 ≤ p ≤ pmax , the point r = 2e + p is a corner point if and
only if sign(β¯′ r − β̄r ) = sign(N2 (k)) > 0. From (19), Let k0
be the solution to the linear equation N1 (k) = 0. Then, after
simplification, we have
(1 − e)(2e2 + 2ep + 2e + p2 + p)
−2e2 + 2e + p2 + p
p+1
+ 2 e+1 + ep
.
= (1 − e) 2 p+1 2 e
−2 2
2
k0 =
Proof: We examine N1 (k). First, we note that when
−2e2 +2e+p2+p > 0, the point r = 2e+p is a corner point for
all systems. Indeed, as N1 (e + 1) = 2ep(e + p) > 0, p > 0, we
have N1 (k) > 0, ∀k ≥ e + 1, p > 0. It follows that, for a fixed
(e, p), we need to determine the sign of −2e2 + 2e + p2 + p.
We have
(26)
As p ≤ pmax , we have −2e2 + 2e + p2 + p < 0, which also
implies that k0 > 0. As N1 (e + 1) = 2ep(e + p), we have
kth ≥ e + 1, with equality iff p = 0. It can be checked from
(26) that when p = e − 1, k0 = 5e − 1. For k ≥ k0 , point r is
not a corner point. As k is an integer and k0 is not necessarily
an integer, it follows that k ≥ k0 ⇐⇒ k ≥ ⌈k0 ⌉ , kth .
Using Proposition 7, Corollary 1 follows.
Corollary 1. For a (k + e, k, k, e) system with e ≥ 2, we have
•
X
TABLE I: Summary of cases for which (ᾱr , β̄r ) is a corner
point in R. The symbol X means (ᾱr , β̄r ) is a corner point
while the symbol ✗ denotes the other case.
(25)
•
•
p∗ in (18) can also be expressed as
p∗ = 1 + max{p : p ≤ pmax and k ≥ kth (p)}
(27)
p
1
= 1 + max p : p ≤
( 8e(e − 1) − 1 − 1)
2
&
'
p+1
+ 2 e+1
+ ep
2
2
and k ≥ (1 − e)
. (28)
p+1
− 2 2e
2
The number of corner points in R is given by nc , |{r :
2e + p∗ ≤ r ≤ k + e}| + 1 = k − e + 2 − p∗ .
As a function of k, p∗ levels out at k = kth (pmax ) and
its final value is given by 1 + pmax .
Example 3. We consider the setting of Figure 4: e = 3, k =
14. We obtain pmax = 2, p∗ = 3. This means the points r, for
6 ≤ r ≤ 2e + p∗ − 1 = 8 are not corner points in R and the
number of corner points is nc = 10. This clearly matches the
observations made in Figure 4.
− 2e2 + 2e + p2 + p < 0 ⇐⇒ p(p + 1) < 2e(e + 1),
(21)
p+1
e
⇐⇒
<2
,
2
2
r
⇐⇒ p <
2e2 − 2e −
(22)
1 1
1 p
− = ( 8e(e − 1) − 1 − 1).
4 2
2
(23)
We
p note that RHS of (23) can not be an integer, as otherwise
8e(e − 1) − 1 should be an odd integer, implying 8e(e −
1) − 1 ≡ 1 mod 4, which leads to a contradiction as 8e(e −
1) − 1 ≡ 3 mod 4. This also implies that the slope of N1 (k)
IV. C ODE
CONSTRUCTION
2
In this section, we present another family of codes improved
upon [1] that encapsulates Construction 1 as a special case.
Let G denote the N (r − m) × nα generator matrix after
vectorization of the code in (6), with t = r. Every node
corresponds to a set of α columns of G. Different from
Construction 1, we allow k ≤ n − m, hence we may feed
Fc , (r − m)N dependent symbols to the generator matrix.
Let T be kα columns of G corresponding to k out of the
n nodes. Let G|T be the submatrix of G consisting of the
p=max(1,r−(n−k))
(29)
The maximum amount of information that can be stored in the
system, F , is upper bounded by ρk,m,r , i.e., F ≤ ρk,m,r . For
instance, when m = n − k, it can be checked that ρk,m,r =
(r − m)N = Fc .
To generate the Fc dependent symbols, we add another layer
of inner code to Construction 1. Moreover, the information
symbols are assumed to be over Fκq , for the finite field Fq and
an appropriately chosen positive integer κ.
Construction 2. For an (n, k, d, e) system, similarly to Construction 1, the code construction is parameterized by m, r,
such that e ≤ m ≤ n − k and m + 1 ≤ r ≤ n (we
assume t = r). For each pair (r, m), let F be given by (29),
F
κ
α = n−r
r−1 . First, the F information symbols {vi }i=1 , vi ∈ Fq ,
are used to construct a linearized polynomial
f (x) =
F
X
vi xq
i−1
.
(30)
i=1
The linearized polynomial is then evaluated at Fc elements
of Fκq to obtain {f (θi ), 1 ≤ i ≤ Fc }, which when viewed
as vectors over Fq , are linearly independent. Finally, the
evaluation points {f (θi ), 1 ≤ i ≤ Fc } are fed to the encoder
in Construction 1.
Repair: The repair of e nodes is similar to Construction 1,
and the contribution of each helper is given by (8).
We note that the elements in Construction 1 are defined
over an alphabet of size q, while the evaluation points are
defined over Fκq . This difference can be resolved by viewing
{f (θi ), 1 ≤ i ≤ Fc } as vectors over Fq and applying
Construction 1 to each of their components. Similarly, the
repair is carried out component-wise. The linearized polynomial evaluations are an instance of rank-metric codes. In [1,
Proposition 5], it is shown that the use of rank-metric codes
guarantees the reconstruction property of the regenerating
code. Moreover, [1] shows that when κ ≥ Fc , the symbols
{f (θi ), 1 ≤ i ≤ Fc } can be made independent over Fq .
In fact, rank metric codes may be replaced by other linear
codes, as long as the reconstruction property is satisfied, so as
to reduce the field size [1]. Furthermore, we note that when
m = n − k, the use of rank-metric codes is not needed, and
the code obtained is simply the code in Construction 1.
Remark 4. Construction 2 generalizes the non-canonical
construction in [1], which is designed for repairing single
erasures. Moreover, the non-canonical construction in [1]
is based on MDS codes, rather than MSMR codes, and its
repair scheme is based on the naive repair scheme of MDS
codes. Finally, non-canonical codes in [1] set m = n − d,
while in Construction 2, m takes arbitrary values, such that
e ≤ m ≤ n − k.
Remark 5. The repair process in Construction 2 does not take
(n,k,d,e)=(19,13,14,3)
m=e
m=e+1
m=e+2
m=e+3
functional tradeoff
Space sharing: MSMR and MBCR
MBCR point
0.06
Normalized bandwidth per helper
columns of T . Then the rank of G|T , denoted by ρk,m,r , is
independent of the choice of the k nodes, and is given by [1]
min(k,r)
X
k
n−k
min(p, r − m).
ρk,m,r =
p
r−p
0.055
0.05
0.045
0.04
0.035
0.03
0.025
0.07
0.08
0.09
0.1
0.11
0.12
0.13
0.14
Normalized storage per node
Fig. 5: Achievable points using Construction
(n, k, d, e) = (19, 13, 14, 3) system. The x-axis
malized storage per node ᾱ and the y-axis is the
bandwidth β̄. When m = n − k = e + 3, the
coincides with Construction 1.
2 for an
is the nornormalized
blue curve
into account the dependency introduced by rank-metric codes
among the Fc = (r − m)N intermediate symbols. It may be
possible to reduce further the repair bandwidth by leveraging
such dependency.
By varying m and r in Construction 2, we obtain various achievability points. Construction 1 is a special case of
Construction 2, corresponding to m = n − k. In particular,
when k = d, n = k + e, Constructions 1 and 2 coincide as
m = n − k = e. For other parameters, simulation shows
that Construction 1 performs better closer to the MSMR point
while Construction 2 with m = e performs better closer to
the MBCR point. Figure 5 plots the achievable points by
Construction 2 for an (n, k, d, e) = (19, 13, 14, 3) system, for
various values of m, e ≤ m ≤ n − k.
V. C ONCLUSION
We studied the problem of centralized exact repair of
multiple failures in distributed storage. We first described a
construction that achieves a new set of interior points. In
particular, we proved the optimality of one point on the
functional centralized repair tadeoff. Moreover, considering
minimum bandwidth cooperative repair codes as centralized
repair codes, we determined explicitly the best achievable
region obtained by space-sharing among all known points,
for the (k + e, k, k, e) system. Finally, we described another
construction, that includes the first construction as a special
case, and that generates various achievable points for a general
(n, k, d, e) system. Future work includes investigating outer
bounds for the centralized exact repair problem.
R EFERENCES
[1] C. Tian, B. Sasidharan, V. Aggarwal, V. A. Vaishampayan, and P. V.
Kumar, “Layered exact-repair regenerating codes via embedded error
correction and block designs,” IEEE Trans. Inf. Theory, vol. 61, no. 4,
pp. 1933–1947, 2015.
[2] A. G. Dimakis, P. Godfrey, Y. Wu, M. J. Wainwright, and K. Ramchandran, “Network coding for distributed storage systems,” IEEE Trans. Inf.
Theory, vol. 56, no. 9, pp. 4539–4551, 2010.
[3] N. B. Shah, K. Rashmi, P. V. Kumar, and K. Ramchandran, “Interference
alignment in regenerating codes for distributed storage: Necessity and
code constructions,” IEEE Trans. Inf. Theory, vol. 58, no. 4, pp. 2134–
2158, 2012.
[4] C. Suh and K. Ramchandran, “Exact-repair MDS code construction
using interference alignment,” IEEE Trans. Inf. Theory, vol. 57, no. 3,
pp. 1425–1442, 2011.
[5] A. S. Rawat, O. O. Koyluoglu, and S. Vishwanath, “Progress on
high-rate msr codes: Enabling arbitrary number of helper nodes,” in
Information Theory and Applications Workshop (ITA), 2016.
[6] S. Goparaju, A. Fazeli, and A. Vardy, “Minimum storage regenerating
codes for all parameters,” IEEE Trans. Inf. Theory, vol. 63, no. 10, pp.
6318–6328, Oct 2017.
[7] V. R. Cadambe, S. A. Jafar, H. Maleki, K. Ramchandran, and C. Suh,
“Asymptotic interference alignment for optimal repair of MDS codes in
distributed storage,” IEEE Trans. Inf. Theory, vol. 59, no. 5, pp. 2974–
2987, May 2013.
[8] I. Tamo, Z. Wang, and J. Bruck, “Zigzag codes: MDS array codes with
optimal rebuilding,” IEEE Trans. Inf. Theory, vol. 59, no. 3, pp. 1597–
1616, March 2013.
[9] M. Ye and A. Barg, “Explicit constructions of high-rate MDS array
codes with optimal repair bandwidth,” IEEE Trans. Inf. Theory, vol. 63,
no. 4, pp. 2001–2014, April 2017.
[10] M. Elyasi and S. Mohajer, “Determinant coding: A novel framework
for exact-repair regenerating codes,” IEEE Trans. Inf. Theory, vol. 62,
no. 12, pp. 6683–6697, 2016.
[11] M. Zorgui and Z. Wang, “Centralized multi-node repair for minimum
storage regenerating codes,” in IEEE International Symposium on Information Theory (ISIT), June 2017, pp. 2213–2217.
[12] B. Sasidharan, K. Senthoor, and P. V. Kumar, “An improved outer bound
on the storage-repair-bandwidth tradeoff of exact-repair regenerating
codes,” in IEEE International Symposium on Information Theory (ISIT).
IEEE, 2014, pp. 2430–2434.
[13] I. M. Duursma, “Outer bounds for exact repair codes,” arXiv preprint
arXiv:1406.4852, 2014.
[14] B. Sasidharan, N. Prakash, M. N. Krishnan, M. Vajha, K. Senthoor, and
P. V. Kumar, “Outer bounds on the storage-repair bandwidth trade-off
of exact-repair regenerating codes,” International Journal of Information
and Coding Theory, vol. 3, no. 4, pp. 255–298, 2016.
[15] I. M. Duursma, “Shortened regenerating codes,” arXiv preprint
arXiv:1505.00178, 2015.
[16] A.-M. Kermarrec, N. Le Scouarnec, and G. Straub, “Repairing multiple
failures with coordinated and adaptive regenerating codes,” in International Symposium on Network Coding (NetCod). IEEE, 2011, pp. 1–6.
[17] K. W. Shum and Y. Hu, “Cooperative regenerating codes,” IEEE Trans.
Inf. Theory, vol. 59, no. 11, pp. 7229–7258, Nov 2013.
[18] A. S. Rawat, O. O. Koyluoglu, and S. Vishwanath, “Centralized repair
of multiple node failures with applications to communication efficient
secret sharing,” arXiv preprint arXiv:1603.04822, 2016.
[19] M. Zorgui and Z. Wang, “Centralized multi-node repair in distributed
storage,” in 2016 54th Annual Allerton Conference on Communication,
Control, and Computing (Allerton), Sept 2016, pp. 617–624.
[20] J. Li and B. Li, “Cooperative repair with minimum-storage regenerating
codes for distributed storage,” in INFOCOM, 2014 Proceedings IEEE.
IEEE, 2014, pp. 316–324.
[21] A. Wang and Z. Zhang, “Exact cooperative regenerating codes with
minimum-repair-bandwidth for distributed storage,” in 2013 Proceedings
IEEE INFOCOM, April 2013, pp. 400–404.
[22] G. M. Kamath, N. Prakash, V. Lalitha, and P. V. Kumar, “Codes with
local regeneration and erasure correction,” IEEE Trans. Inf. Theory,
vol. 60, no. 8, pp. 4637–4660, 2014.
[23] A. S. Rawat, N. Silberstein, O. O. Koyluoglu, and S. Vishwanath,
“Optimal locally repairable codes with local minimum storage regeneration via rank-metric codes,” in Information Theory and Applications
Workshop (ITA), 2013. IEEE, 2013, pp. 1–8.
[24] C. Tian, “A note on the rate region of exact-repair regenerating codes,”
arXiv preprint arXiv:1503.00011, 2015.
| 7cs.IT
|
arXiv:1509.01376v2 [math.GR] 17 Feb 2016
New topological methods to solve equations over groups
Anton Klyachko and Andreas Thom
Abstract. We show that the equation associated with a group word w ∈ G ∗ F2 can be
solved over a hyperlinear group G if its content – that is its augmentation in F2 – does
not lie in the second term of the lower central series of F2 . Moreover, if G is finite, then a
solution can be found in a finite extension of G. The method of proof extends techniques
developed by Gerstenhaber and Rothaus, and uses computations in p-local homotopy theory
and cohomology of compact Lie groups.
Contents
Introduction
1. The main results
2. The topology of SU(p) and PU(p)
3. Solvability of equations
Acknowledgments
References
1
2
5
11
15
15
Introduction
This paper is about the solvability of equations in groups. Let us start by briefly recalling
the analogous situation of polynomial equations with rational coefficients. Even though not
every non-constant polynomial p(t) ∈ Q[t] has a root in Q, there always exists a finite field
extension Q ⊂ K, such that p(t) = 0 can be solved in K, i.e., there exists α ∈ K with p(α) = 0.
Indeed, it is straightforward to construct some splitting field K with the desired property using
the machinery of commutative algebra. On the other side, it is also well-known that arguments
from algebraic topology (using notions of degree, winding number or the fundamental group)
can be used to show that every polynomial has a root in the topological field C and historically,
this was the first way to provide a field extension of Q in which p(t) = 0 can be solved – an
argument which essentially goes back to Gauss’ work from 1799. In the analogous situation,
when one wants to solve equations with coefficients in a group, the algebraic or combinatorial
approach fails to a large extend and the homotopy theoretic approach has been used by
Gerstenhaber-Rothaus [14] to obtain positive results – see the next section for definitions
and more precise statements. As a particular consequence, Gerstenhaber-Rothaus were able
to prove the Kervaire-Laudenbach conjecture for locally residually finite groups. First of all,
1
2
ANTON KLYACHKO AND ANDREAS THOM
we want to clarify a relationship between Connes’ Embedding Problem and the KervaireLaudenbach Conjecture that was observed in [36]. Moreover, we want to extend this study
to cover a larger class of groups and also a larger class of equations which can be handled
by methods from algebraic topology. Our methods involve a detailed study of the p-local
homotopy type of the simple Lie group PU(p) and the effect of word maps on the cohomology
ring with mod p coefficients.
Our first main result applies to any prime number p and any group word w ∈ SU(p) ∗ F2 .
If the augmentation of w in F2 does not lie in the second step of the exponent-p central series
[F2 , F2 ]p [F2 , [F2 , F2 ]], then the equation w(a, b) = 1 can be solved in SU(p). This implies
our second main result, which says that a two-variable equation with augmentation not in
[F2 , [F2 , F2 ]] can be solved over any hyperlinear group, see Section 1 for details. Moreover,
if the group G is finite, then a solution can be found in a finite extension of G. This covers
classes of singular equations, which were intractable by combinatorial methods or the topological methods developed by Gerstenhaber-Rothaus. Our main results are stated explicitly
as Theorems 1.5 and 1.6 in Section 1.2.
The paper is organized as follows. Section 1 collects various preliminaries and discusses
briefly the setup of group words and equations and the classes of hyperlinear and sofic groups.
Section 2 recalls some facts about the cohomology of SU(n) and PU(n), localization theory
of topological spaces, and computations of homotopy groups of spheres. This section is the
most technical part and also contains a review and extension of results of Kishimoto and Kono
[25]. Section 3 contains the proofs of our main results and discusses related low-dimensional
results and further directions.
1. The main results
1.1. Group words and equations. We denote by Fn the free group on generators
x1 , . . . , xn . For any group G, an element w in the free product G ∗ Fn determines a word map
w : G×n → G given by evaluation. We denote by ε : G ∗ Fn → Fn the natural augmentation
which sends G to the neutral element and call ε(w) the content of w. We call w ∈ G ∗ Fn a
group word in n variables with coefficients in G. Every group word w ∈ G ∗ Fn determines
an equation w(x1 , . . . , xn ) = 1 in n variables with coefficients in G in an obvious way. We say
that w ∈ G ∗ Fn can be solved over G if there exists an overgroup H ⊇ G and g1 , . . . , gn ∈ H
such that w(g1 , . . . , gn ) = 1, where 1 denotes the neutral element in H. Similarly, we say that
it can be solved in G if we can take H = G. We denote the normal subgroup in G ∗ Fn , which
is generated by the element w by hhwii. It is clear that an equation w ∈ G ∗ Fn can be solved
over G if and only if the natural homomorphism G → G ∗ Fn /hhwii is injective. Similarly, an
equation can be solved in G if and only if the natural homomorphism G → G ∗ Fn /hhwii is
split-injective, i.e., it has a left inverse.
The study of equations over groups dates back to the work of Bernhard Neumann [34].
There is an extensive literature about equations over groups, including [1,8–10,13,14,20–22,
24, 26, 27, 29, 30, 34]. See also Roman’kov’s recent survey about this topic [39].
It is well-known that not all equations with coefficients in G are solvable over G. For
example if G = ha, b | a2 , b3 i, then the equation w(x) = xax−1 b with variable x is not solvable
over G. Indeed, a and b cannot become conjugate in any overgroup of G. Another example
is G = Z/pZ = hai with the equation w(x) = xax−1 axa−1 x−1 a−2 . However, in both cases we
have ε(w) = 1 ∈ Fn . Indeed, the only known examples of equations which are not solvable
EQUATIONS OVER GROUPS
3
over some G are equations whose content is trivial. We call an equation w ∈ G ∗ Fn singular
if its content is trivial, and non-singular otherwise. This lets us put forward the following
conjecture:
Conjecture 1.1. Let G be a group and w ∈ G ∗ Fn be an equation in n variables with
coefficients in G. If w is non-singular, then it is solvable over G. In addition, if G is finite,
then a solution can be found in a finite extension.
The case n = 1 is the famous Kervaire-Laudenbach Conjecture. The one-variable case
was studied in classical work by Gerstenhaber-Rothaus from 1962, see [14]. They showed
that if G is finite, then every non-singular equation in one variable can be solved over G
(in fact in some finite extension of G). Their proof used computations in cohomology of the
compact Lie groups U (n). It is this proof, that inspired us to start this work. The work of
Gerstenhaber-Rothaus showed in fact that every non-singular equation in one variable with
coefficients in the unitary group U(n) can be solved already in U(n), for any n ∈ N. Their
strategy was to use homotopy theory to say that the associated word map w : U(n) → U(n)
has a non-vanishing degree (as a map of oriented manifolds) and thus must be surjective.
Any preimage of the neutral element provides a solution to the equation w. The key to the
computation of the degree is to observe that the degree depends only on the homotopy class
of w and thus – since U(n) is connected – does not change if w is replaced by ε(w). The
computation of the degree is now an easy consequence of classical computations of Hopf [19].
This property of solvability in a group is easily seen to pass to arbitrary Cartesian products
of groups and arbitrary quotients of groups. As a consequence, non-singular equations in
one variable with coefficients in G as above can Q
be solved over G if G is isomorphic to
a subgroup of a quotient of the infinite product n U(n) – an observation that is due to
Pestov [36]. Groups which admit such an embedding are called hyperlinear groups, see [36]
and Section 3.1 for more information on this class of groups – see also Remark 3.3. Thus,
the result of Gerstenhaber-Rothaus also holds for hyperlinear groups – in particular for all
amenable groups, or more generally, all sofic groups [36]. Connes’ Embedding Conjecture
predicts (among other things) that every countable group is hyperlinear and thus implies the
Kervaire-Laudenbach Conjecture – this was also observed by Pestov in [36].
Actually, Gerstenhaber-Rothaus [14] studied the more involved question whether m equations of the form w1 , . . . , wm ∈ G ∗ Fn in n variables can be solved simultaneously over G.
Their main result is that this is the case if G is finite (or more generally, locally residually finite) and the presentation two-complex X := Khx1 , . . . , xn | ε(w1 ), . . . , ε(wm )i satisfies
H2 (X, Z) = 0, i.e., the second homology of X with integral coefficients vanishes. Here, the
representation two-complex is the two-dimensional CW -complex associated with the presentation of a group – obtained by glueing m two-cells to a bouquet of n circles according to
the relations. (This amounts to a certain algebraic condition on the exponent sum matrix.)
A system of equations which satisfies this vanishing condition was called non-singular by
Jim Howie [20] – note that our terminology is not consistent with this, but there will be
no risk of confusion. Later, Howie [20] proved the same result for locally indicable groups
and conjectured it to hold for all groups – we call that Howie’s Conjecture. Again, Connes’
Embedding Conjecture implies Howie’s Conjecture – and more specifically, every hyperlinear
group satisfies Howie’s Conjecture.
Remark 1.2. Equations in one variable with at most three occurrences of the variable
are solvable by a result of Howie [21], which however also reduces this to the residually finite
4
ANTON KLYACHKO AND ANDREAS THOM
case and uses the results of Gerstenhaber-Rothaus. Similar results have been proved for
non-singular equations with four [10] and five [12] occurrences of variables.
Remark 1.3. Equations w ∈ G ∗ Fn with G torsion-free can be solved by more combinatorial methods. A systematic study of the torsion-free case was started by Levin [30]
who conjectured that equations in one variable with coefficients in a torsion-free group should
always be solvable if w is conjugate to an element in G. A result in this direction is due to the
first author who proved that this is indeed the case for one-variable equations with content
±1 ∈ Z, see [26]. Moreover, over any torsion-free group any several-variable equation whose
content is not a proper power (and not the neutral element) is solvable [28]. Again, due to
the absence of any counterexamples, it is conjectured that solvability in the torsion-free case
is true even if the content of the equation is trivial.
Remark 1.4. Existence of non-trivial solutions can be a subtle issue too, see for example
[40]. For G = Z/pZ = hai, the equation w(x) = axa−1 xax−1 a−1 x−2 can be solved in a finite
overgroup of G only with x = 1, even though non-trivial solutions exist in infinite extensions.
The mechanism behind these kind of examples was first discovered by Higman, see [18].
1.2. Statement of the main results. The main goal of this work is to provide examples
of non-singular equations in many variables which are solvable over every hyperlinear group,
where the condition on the equation only depends on its content. This should be compared for
example with results of Gersten [13], where the conditions on w depended on the unreduced
word obtained by deleting the coefficients from w. For simplicity, we concentrate on the
two-variable case. Our main result is:
Theorem 1.5. Let G be a hyperlinear group. An equation in two variables with coefficients
in G can be solved over G if its content does not lie in [F2 , [F2 , F2 ]]. Moreover, if G is finite,
then a solution can be found in a finite extension of G.
In order to prove our main result we have to refine the study of Gerstenhaber-Rothaus on
the effect of word maps on cohomology of compact Lie groups. Again, the strategy is to show
that such equations can be solved in SU(n) for sufficiently many n ∈ N. More specifically, we
prove:
Theorem 1.6. Let p be a prime number. Let w ∈ SU(p) ∗ F2 be a group word. If
ε(w) 6∈ [F2 , F2 ]p [F2 , [F2 , F2 ]],
then the equation w(a, b) = 1 can be solved in SU(p).
If ε(w) 6∈ [F2 , F2 ], then this theorem is a direct consequence of the work of Gerstenhaber
and Rothaus. However, if ε(w) ∈ [F2 , F2 ], then a new idea is needed. We show – under the
conditions on p which are mentioned above – that the induced word map w : PU(p)×PU(p) →
SU(p) is surjective, where SU(p) denotes the special unitary group and PU(p) its quotient by
the center. The strategy is to replace w by the much simpler and homotopic map induced
by ε(w) and study its effect on cohomology directly. This is done in Section 3.2 with the
necessary preparations from Section 2.
In general, the assumption on ε(w) cannot be omitted in the previous theorem. Indeed,
the second author showed in previous work:
Theorem 1.7 ([11]). For every k ∈ N and every neighborhood V ⊂ SU(k) of 1k ∈ SU(k),
there exists w ∈ F2 \{e}, such that the image of w : SU(k)×2 → SU(k) lies in V. In particular,
the equation w(a, b) = g is not solvable in SU(k) for g 6∈ V.
EQUATIONS OVER GROUPS
5
The construction that proves the preceding theorem yields words in F2 that lie deep in
the derived series, so that there is no contradiction with Theorem 1.6.
The surjectivity of word maps without coefficients is an interesting subject in itself.
Michael Larsen conjectured that for each non-trivial w ∈ F2 and n high enough, the associated word map w : PU(n) × PU(n) → PU(n) is surjective. This was shown (with some
divisibility restrictions on n) for words not in the second derived subgroup of F2 by Elkasapy
and the second author in [11]. In a similar direction, we believe that for n high enough –
or again, with some divisibility restrictions – the word map w should define a non-trivial
homotopy class and not even be homotopic to a non-surjective map.
In order to study words which lie deeper in the lower central series, we suspect that it
might be helpful to oberserve that the induced word map w : PU(p) × PU(p) → PU(p) does
not only lift to SU(p) – which is the simply connected cover of PU(p) – but lifts even to higher
connected covers of PU(p). Indeed, for example one can show that if w ∈ [F2 , [F2 , F2 ]], then
the associated word maps lifts to the complex analogue of the string group, see [45] for a
study of related groups.
2. The topology of SU(p) and PU(p)
In this section we collect some standard results from algebraic topology that will be used
in the proof of the main theorems. A classical result of Samelson says:
SU(2)
: SU(2)×2 → SU(2) is not
Theorem 2.1 (Samelson, see [41]). The commutator map c2
null-homotopic. In particular, since SU(2) is a sphere, any map homotopic to the commutator
map must be surjective.
This easily implies that for every group word w ∈ SU(2) ∗ F2 whose content is the commutator of the generators of F2 , the equation w(a, b) = 1 can be solved in SU(2). This already
has non-trivial consequences that (to the best of our knowledge) could not be proved using
combinatorial techniques. In order to treat SU(n) for higher n, we have to recall some aspects
of algebraic topology. Our methods in the proof of the main results follow closely ideas from
Hamanaka-Kishimoto-Kono [16] and Kishimoto-Kono [25].
2.1. The cohomology of SU(p) and PU(p). Let n be a positive integer. We denote
by SU(n) the special unitary group and by PU(n) the quotient of SU(n) by its center – the
projective unitary group. We denote the quotient map by π : SU(n) → PU(n) and the coset
of some u ∈ SU(n) in PU(n) by π(u) = ū. The cohomology rings of the simply-connected
classical Lie groups were computed by Borel in [3]. For example, it is well-known that as a
graded ring
H ∗ (SU(n), Z/pZ) = Λ∗Z/pZ (x2 , x3 , . . . , xn )
with |xi | = 2i − 1. Here, we denote by Λ∗k the exterior algebra over a field k on a certain
set of generators of particular degrees. The product map m : SU(n) × SU(n) → SU(n) and
the inversion turn H ∗ (SU(n), Z/pZ) into a Hopf algebra. However, the comultiplication turns
out to be trivial in this situation, i.e., ∆(xi ) = xi ⊗ 1 + 1 ⊗ xi for all 2 ≤ i ≤ n. We will
be mainly interested in the case n = p. The computation of the cohomology ring of PU(p)
is more involved than that of SU(p) and was also first studied by Borel in [4]. Later, the
comultiplication on the cohomology of PU(p) with Z/pZ-coefficients was computed in work of
Baum-Browder [2] and turns out to be not co-commutative. It is this lack of co-commutativity
which makes our approach work. Let us sumarize the situation:
6
ANTON KLYACHKO AND ANDREAS THOM
Theorem 2.2 (Baum-Browder, see [2]). Let p be an odd prime number. Then,
H ∗ (PU(p), Z/pZ) ∼
= (Z/pZ)[y]/(y p ) ⊗Z Λ∗Z/pZ (y1 , y2 , . . . , yp−1 )
with |y| = 2, |yi | = 2i − 1, π ∗ (yi ) = xi for 2 ≤ i ≤ p − 1, and π ∗ (y) = π ∗ (y1 ) = 0. Moreover,
the comultiplication takes the form
∆(y) = y ⊗ 1 + 1 ⊗ y,
and
∆(yi ) = yi ⊗ 1 + 1 ⊗ yi +
i−1
X
j−1
j=1
i−1
· yj ⊗ y i−j .
We denote by I(n) the kernel of the natural augmentation H ∗ (PU(p)n , Z/pZ) → Z/pZ.
We start out by recalling the effect of various natural word maps on the cohomology ring.
Lemma 2.3. Let n ∈ N and p be an odd prime number. Consider the map µn : PU(p) →
PU(p) given by µn (u) = un . The map induced by µn : PU(p) → PU(p) on cohomology satisfies
µ∗n (yi ) = n · yi
mod I(1)2
for all 1 ≤ i ≤ p − 1.
of the diagonal embedding PU(p) →
Qn Proof. The map µn arises as the composition Q
n
PU(p)
following
by
the
multiplication
map
m
:
i=1
i=1 PU(p) → PU(p). In cohomology,
(n)
∗
this induces first the the n-fold coproduct ∆ : H (PU(p), Z/pZ) → ⊗ni=1 H ∗ (PU(p), Z/pZ)
followed by the multiplication in cohomology ⊗ni=1 H ∗ (PU(p), Z/pZ) → H ∗ (PU(p), Z/pZ). An
easy verification shows that each generator is just multiplied by n, modulo sums of products
of at least two generators. This proves the claim.
It is also important for us to study the effect of the commutator map in cohomology. We
need the following result from the work of Hamanaka-Kishimoto-Kono [16], which is an easy
consequence of Theorem 2.2 above.
Lemma 2.4 (Proposition 6 in [16]). The commutator map
c : PU(p) × PU(p) → PU(p)
induces the cohomology map c∗ : H ∗ (PU(p), Z/pZ) → H ∗ (PU(p) × PU(p), Z/pZ) sending yi
to (i − 1)(yi−1 ⊗ y − y ⊗ yi−1 ) modulo the ideal I(2)3 . The elements y1 and y map to zero.
Note that the commutator map SU(n) × SU(n) ∋ (u, v) 7→ uvu−1 v −1 ∈ SU(n) induces a
well-defined map c : PU(n) × PU(n) → SU(n), which we will also call commutator map.
Our first aim is to show that the commutator map c : PU(p) × PU(p) → SU(p) is not
homotopic to a non-surjective map. We will show this by showing that the image of the
2
top-dimensional cohomology class x2 · · · xp ∈ H p −1 (SU(p), Z/pZ) does not vanish in the
2
group H p −1 (PU(p) × PU(p), Z/pZ). As it turns out, the study the images of x2 , . . . , xp−1 in
∗
H (PU(p) × PU(p), Z/pZ) is fairly straightforward, since these generators are the image of
generators y2 , . . . , yp−1 in the cohomology of PU(p). The study of the last generator xp is
considerably more complicated and we have to rely on some structure results on the p-local
homotopy type of PU(p). In fact, in the final argument we will not rely on Lemma 2.4, but
the required result is proved for all xi with 2 ≤ i ≤ p directly.
EQUATIONS OVER GROUPS
7
2.2. The lens spaces. Since S 2d−1 ⊂ Cp , there is a natural (Z/pZ)-action on S 2d−1
given by scalar multiplication with the complex number exp(2πi/p). In our considerations, we
only study the case d = p. We denote by L the lens space S 2p−1 /(Z/pZ) and let ρ : S 2p−1 → L
be the natural projection. The space L has a natural CW-structure with one cell in each
dimension, see [17, Example 2.43].
We denote by Lk := S 2k−1 ∪p D 2k the mod-p Moore space for 1 ≤ k ≤ p − 1. This
is, by definition, the space obtained by attaching D 2k to S 2k−1 along the attaching map
p : S 2k−1 → S 2k−1 , which is defined as (z1 , . . . , zk ) → (z1p , z2 , . . . , zk ). Note that the letter p is
over-used here, but this will not cause any confusion. The characteristic property of Lk (for
1 ≤ k ≤ p − 1) is that Hn (Lk , Z) = 0 unless n = 2k − 1 or n = 0, and H2k−1 (Lk , Z) = Z/pZ.
See [17, Chapter 2, Example 2.40] for more details. We set Lp := S 2p−1 .
For 1 ≤ k ≤ p − 1, we denote by qk : Lk → S 2k the so-called pinch map, which collapses
2k−1
S
(and hence the boundary of D 2k ) to a point. Note that the 2-skeleton of L is just L1 .
Indeed, L1 = S 1 ∪p D 2 , see [17, Example 2.43].
ι
π
There is a fibre bundle
SU(p − 1) → SU(p) → S 2p−1 , where the embedding ι sends a matrix
A to the matrix 10 A0 and the projection π sends a matrix to its first row. Similarly, the
ι
π
group PU(p) admits a fibre bundle SU(p − 1) → PU(p)
→ L, where the embedding ι sends a
matrix A to the class of matrices hexp(2πi/p)i 10 A0 , the projection π sends a matrix to its
first row (which is defined up to multiplication by exp(2πi/p)), and L = S 2p−1 /hexp(2πi/p)i
is the lens space.
2.3. Localization at a prime. We will freely use the concept of localization of topological spaces (simply connected or with abelian fundamental group) at a prime p, see the work
of Bousfield-Kan [7] or Mimura-Nishida-Toda [33] background and as general references. See
also [32] for a more recent presentation of this material.
Given a topological space X with abelian fundamental group, we denote by X(p) its plocalization which comes equipped with a natural map ι : X → X(p) . The p-localization can
be defined as a certain tower of spaces, and its defining properties are
πi (X(p) , ι(x)) = πi (X, x) ⊗Z Z(p) ,
∀x ∈ X, ∀i ≥ 1.
Here, Z(p) denotes the p-localization of Z, i.e., the ring of those fractions in Q, whose denominator is not divisible by p. For a continuous map f : X → Y (between topological spaces with
abelian fundamental group), we denote by f(p) : X(p) → Y(p) the induced map between the
p-localizations. We will freely use that ι : X → X(p) induces an isomorphism on cohomology
with coefficients in Z/pZ. A map is called a p-local homotopy equivalence if it is a homotopy
equivalence after p-localization.
If X is a double suspension, then [X, Y ] is an abelian group – here [X, Y ] denote as usual
the set of homotopy classes of maps from X to Y . We will use that if in addition X is also a
finite CW-complex, then the natural maps
[X, Y ] ⊗Z Z(p) → [X, Y(p) ] ← [X(p) , Y(p) ]
are isomorphisms, see for example [32, Chapter 6.6].
We will need the following computation of the homotopy groups of SU(n), which is due
to Bott [5].
8
ANTON KLYACHKO AND ANDREAS THOM
Theorem 2.5 (Bott).
0
Z
πk (SU(n)) =
0
Z/n!Z
k
k
k
k
= 0, 1, 2
= 2i − 1, 2 ≤ i ≤ n
.
= 2i, 2 ≤ i ≤ n − 1
= 2n
2k−1 → SU(n) be a generator of π
Let ǫQ
k: S
2k−1 (SU(n)) for 2 ≤ k ≤ n and consider the
n
map µ : k=2 S 2k−1 → SU(n) given by µ(x2 , . . . , xn ) := ǫ2 (x2 ) · · · ǫn (xn ), where we use multiplication in the group SU(n). The following theorem was first proved by Serre [42, Proposition
7] – even though without using the language of localization at the level of topological spaces.
Theorem 2.6 (Serre). Let p be a prime number. If p ≥ n, then the map
n
Y
µ:
S 2k−1 → SU(n)
k=2
is a p-local homotopy equivalence.
We will now concentrate on the case that n = p. For 2 ≤ i ≤ p, we denote by
2i−1
the composition of the homotopy inverse of µ(p) with the projection
λi : SU(p)(p) → S(p)
2i−1
. We also have the following computation of the p-local homotopy groups of odd
onto S(p)
spheres, which is also due to Serre [42, Proposition 11]:
0
0 ≤ k < 2i − 1
k = 2i − 1
Z(p)
2i−1
= 0
(1)
πk S(p)
2i − 1 < k < 2i + 2p − 4
Z/pZ k = 2i + 2p − 4
0
2i + 2p − 4 < k < 2i + 4p − 7.
2i−1
3 ).
) is equal to Σ2i−4 (α) for some generator α ∈ π2p (S(p)
Here, the generator in π2i+2p−4 (S(p)
Here, Σ(?) denotes as usual the suspension also on the level of maps. Hence, Theorem 2.6
together with the computation above implies that there is a more refined computation of the
homotopy groups of SU(p) localized at a prime p:
k = 0, 1, 2
0
Z
k = 2i − 1, 2 ≤ i ≤ p
(p)
πk SU(p)(p) = 0
k = 2i, 2 ≤ i ≤ p − 1
0
k = 2i + 1, p ≤ i < 2p − 1
Z/pZ k = 2i, p ≤ i < 2p − 1.
Note that this covers all dimensions up to 4p − 4, whereas Bott’s computation only gives
information up to dimension 2p – a fact that will be used later.
Now, the commutator map on SU(p) induces a secondary operation
h., .i : πi (SU(p)) × πj (SU(p)) → πi+j (SU(p)),
the so-called Samelson product, which was originally introduced in [41]. Bott already analyzed the Samelson products of the maps ǫk in [6]. He proved as a corollary to his main result
[6, Theorem 1] that the element hǫi , ǫp−i+1 i in π2p (SU(p)) = Z/p!Z is divisible by precisely
EQUATIONS OVER GROUPS
9
p!
(i−1)!(p−i)! ,
i.e., it is equal to (i−1)!(p−i)! times some generator of Z/p!Z. The maps ǫk induce
natural maps ǭk : S 2k−1 → PU(p) for 2 ≤ k ≤ p. Note, that we can also choose a natural
map ǭ1 : S 1 → PU(p) which yields a generator of π1 (PU(p)) = Z/pZ – and that Bott’s result
extends to the case i = 1. In the light of our computation of πk (SU(p)(p) ) from above, Bott’s
computation of the Samuelson products implies:
Theorem 2.7 (Bott). Let p be a prime number and 1 ≤ i < p. The element
hǭp , ǭi i ∈ π2p+2i−2 (SU(p)(p) ) = Z/pZ
does not vanish.
Proof. Indeed, Bott’s result from above yields that the image of the map
hǭp , ǭi i : S 2p+2i−2 → SU(p + i − 1)
in π2p+2i−2 (SU(p + i − 1)) = Z/(p + i − 1)!Z is (p − 1)!(i − 1)! times some generator, and hence
does not vanish modulo p. Since hǭp , ǭi i factors through SU(p), the assertion follows.
The non-vanishing of these Samuelson products modulo p will be the key to understand
the non-vanishing of certain cohomology classes after application of the commutator map.
2.4. The work of Kishimoto-Kono. In order to understand the effect of the commutator map on the cohomology of PU(p), we must now study the p-local homotopy type of
PU(p). We restate Proposition 2 from the work of Kishimoto-Kono [25].
Lemma 2.8. There exists a natural map η : L(p) → PU(p)(p) , such that the diagram
ǫp(p)
2p−1
S(p)
(2)
/ SU(p)
(p)
ρ(p)
η
L(p)
/ PU(p)(p)
commutes up to homotopy.
Using the notation introduced in Section 2.2, we are now ready to state and prove an
extension of Lemma 4 of Kishimoto-Kono [25].
Lemma 2.9. For 1 ≤ i ≤ p − 1, we have p-locally
2i+1
2i−1
→ S(p)
λi+1 ◦ c(p) ◦ (η|L1(p) ∧ ε̄i(p) ) = a · (q1 ∧ 1S 2i−1 )(p) : L1(p) ∧ S(p)
for some a ∈ Z×
(p) .
Proof. First of all, we know from [33, Proposition 9.6] that there exists a p-local splitting
(L ∧ S 1 )(p) =
p
_
(Lk ∧ S 1 )(p) .
k=1
Thus, (L ∧
S 2i−1 )(p)
=
Wp
k=1 (Lk
∧
S 2i−1 )(p)
and the map
2p−1
∧ S 2i−1} → L ∧ S 2i−1
ρ ∧ 1S 2i−1 : S
|
{z
S 2p+2i−2
10
ANTON KLYACHKO AND ANDREAS THOM
can be decomposed p-locally as
(3)
(ρ ∧ 1S 2i−1 )(p) =
p
_
fk ,
2p+2i−2
→ (Lk ∧ S 2i−1 )(p) ,
with fk : S(p)
1 ≤ k ≤ p.
k=1
S 2p−1
Since the map ρ :
→ L is a p-fold covering, we get fp is equal to multiplication by p.
p
Now, for 1 ≤ k ≤ p−1, the cofiber sequence S 2k−1 → S 2k−1 → Lk (coming from the definition
of Lk ) induces a long exact sequence
p
p
· · · → π2i+2k−1 (S 2i+1 ) → π2i+2k−1 (S 2i+1 ) → [Lk ∧ S 2i−1 , S 2i+1 ] → π2i+2k−2 (S 2i+1 ) → · · · .
From the computations of the p-local homotopy groups of spheres, we obtain for all 1 ≤ i ≤ p
that
k=1
Z(p)
2i−1
2i+1
[Lk ∧ S
,S
](p) = 0
2≤k ≤p−1
Z/pZ k = p.
Moreover, it follows that the group [Lk ∧S 2i−1 , S 2i+1 ](p) is generated by the map (q1 ∧1S 2i−1 )(p) .
Consider now the map
2i+1
2i−1
.
→ S(p)
λi+1 ◦ c(p) ◦ (η ∧ ǭi(p) ) : L(p) ∧ S(p)
We obtain from Equation (3) and the sentence after Equation (1) that
λi+1 ◦ c(p) ◦ (η ∧ ǭi(p) ) = ai · (q1 ∧ 1S 2i−1 )(p) ∨ bi · (Σ2i−2 (α))
(4)
for some ai , bi ∈ Z(p) and 1 ≤ i ≤ p − 1. Here, we consider Σ2i−2 (α) as map
1∧α(p)
2p
2p+2i−2
2i+1
2i−2
2i−2
2i−1
3
.
∧ S(p)
= S(p)
→ S(p)
∧ S(p)
= S(p)
= S(p)
Lp(p) ∧ S(p)
Now is the point, when we are going to use Theorem 2.7. Indeed, we have the following
identification of homotopy classes of maps from S 2p+2i−2 to S 2i+1 (note that π2p+2i−2 (S 2i+1 ) =
Z/pZ by Equation 1):
0
Th.2.7
6=
λi+1 ◦ hǭp , ǭi i(p)
=
λi+1 ◦ c(p) ◦ (ǭp(p) ∧ ǭi(p) )
(2)
=
λi+1 ◦ c(p) ◦ (η ∧ ǭi(p) ) ◦ (ρ(p) ∧ 1S 2i−1 )
ai · (q1 ∧ 1S 2i−1 )(p) ∨ bi · (Σ2i−2 (α)) ◦ (ρ(p) ∧ 1S 2i−1 )
=
ai · ((q1 ∧ 1S 2i−1 )(p) ◦ f1 ).
=
(4)
=
(3)
ai · ((q1 ∧ 1S 2i−1 )(p) ◦ f1 ) ∨ pbi · (Σ2i−2 (α))
Finally, this implies that (q1 ∧ 1S 2i−1 )(p) ◦ f1 is a non-zero multiple of the class
Σ2i−2 (α) ∈ π2p+2i−2 (S 2i+1 )(p) = Z/pZ
and we can conclude that ai ∈ Z×
(p) . This finishes the proof.
The preceding proof followed closely the work of Kishimoto-Kono [25] and we do not
claim any originality for this computation.
EQUATIONS OVER GROUPS
11
2.5. Applications to cohomology. We will now apply the computation from the previous section to study the effect of the commutator map on cohomology. Recall that
H ∗ (PU(p), Z/pZ) ∼
= (Z/pZ)[y]/(y p ) ⊗Z Λ∗ (y1 , y2 , . . . , yp−1 )
Z/pZ
with |y| = 2, |yi | = 2i − 1. We denote by Ji the ideal in the ring H ∗ (PU(p), Z/pZ) ⊗
H ∗ (PU(p), Z/pZ) which is generated by y 2 ⊗ 1, yj ⊗ 1, 1 ⊗ yk , 1 ⊗ y for 1 ≤ j, k ≤ p − 1 with
k 6= i.
Corollary 2.10. Let 2 ≤ i ≤ p and c : PU(p) × PU(p) → SU(p) the commutator map.
Then, the induced map
c∗ : H ∗ (SU(p), Z/pZ) → H ∗ (PU(p), Z/pZ) ⊗Z H ∗ (PU(p), Z/pZ)
satisfies
c∗ (xi ) = ai (y ⊗ yi−1 )
mod Ji−1
for some ai ∈ Z×
(p) .
Proof. Note that
2i−1
, Z/pZ = Λ∗ (y) ⊗Z Λ∗ (yi )
H ∗ L1(p) × S(p)
with |y| = 2 and |yi | = 2i − 1 such that the natural map
2i−1
→ PU(p) × PU(p)
η × ǭi(p) : L1(p) × S(p)
induces the natural homomorphism from
(Z/pZ)[y]/(y p ) ⊗Z Λ∗Z/pZ (y1 , y2 , . . . , yp−1 ) ⊗ (Z/pZ)[y]/(y p ) ⊗Z Λ∗Z/pZ (y1 , y2 , . . . , yp−1 )
to Λ∗ (y) ⊗Z Λ∗ (yi ) which sends y ⊗ 1 to y ⊗ 1, 1 ⊗ yi to 1 ⊗ yi and the other generators to
zero. The kernel of this homomorphism is precisely the ideal Ji . Now, Lemma 2.9 implies
that the composition
c
λi+1
2i+1
2i−1
→ PU(p)(p) ∧ PU(p)(p) → SU(p)(p) → S(p)
L1(p) ∧ S(p)
2i−1 →
is homotopic to ai (q1 ∧ 1S 2i−1 )(p) for some ai ∈ Z×
(p) . Since the map q1 ∧ 1S 2i−1 : L1 ∧ S
S 2i+1 sends the generator xi+1 of the cohomology of S 2i+1 to y ⊗ yi ∈ Λ∗ (y) ⊗Z Λ∗ (yi ), this
implies the claim.
3. Solvability of equations
3.1. Hyperlinear groups and related classes of groups. The unitary group U(n)
is equipped with a natural metric that arises from the normalized Frobenius norm, i.e.,
1/2
n
X
1
|uij − vij |2 .
d(u, v) = 1/2
n
i,j=1
Informally speaking, a group G is said to be hyperlinear if its multiplication table can be
modelled locally (that means on finite subsets of the group) by unitary matrices up to small
mistakes measured in the normalized Frobenius norm. More precisely:
Definition 3.1. A group G is called hyperlinear if for all finite subsets F ⊂ G and all
ε > 0, there exists n ∈ N and a map ϕ : G → U(n), such that
(1) d(ϕ(gh), ϕ(g)ϕ(h)) < ε for all g, h ∈ F , and
12
ANTON KLYACHKO AND ANDREAS THOM
(2) d(ϕ(g), 1n ) > 1, for all g ∈ F \ {e}.
There are variations on this definition but they are all equivalent. A detailed discussion
of the class of hyperlinear groups can be found in [36]. If in the above definition the unitary
groups with their metrics are replaced by symmetric groups Sym(n) with the normalized
Hamming metrics, then one obtains the definition of the class of sofic groups. This important
class of groups was introduced by Gromov [15] and Weiss [47] in order to study certain
problems in ergodic theory. Since the inclusion Sym(n) ⊂ U(n) is compatible enough with
the metrics, every sofic group is automatically hyperlinear – see [36] for more details. It is
not known if there are non-sofic groups.
We denote the set of prime numbers by P. We will need the following easy proposition.
Proposition
3.2. Let G be a countable hyperlinear group. Then G is a subgroup of a
Q
quotient of p∈P SU(p).
Proof. Let (gi )i∈N be an enumeration of G. Let nk be some integer such that the
definition of hyperlinearity is satisfied for the finite set {g1 , . . . , gk } with ε = 1/k. Let
ϕk : G → U(nk ) the the corresponding map. Without loss of generality, we may assume
that limk→∞ nk = ∞. Indeed, the natural diagonal embedding U(n) ⊂ U(nm) is isometric
with respect to the normalized Frobenius metric, so that we can replace nk by knk if necessary. Using the natural embedding U(n) ⊂ SU(n + 1), we may now assume without loss of
generality that the image of ϕk lies in SU(nk ). Similarly, replacing nk again by a suitable
number of the form mnk + 1, we may assume that nk is a prime number – using Dirichlet’s
theorem. Now, consider
Y
N := (up )p ∈
SU(p) | lim d(1nk , unk ) = 0 .
k→∞
p∈P
Q
Q
It is easy to see that N ⊂ p∈P SU(p) is a normal subgroup and ϕ = k∈N ϕk defines
Q
a injective homomorphism from G to the quotient of p∈P SU(p) by N . This proves the
claim.
Q
Q
Remark 3.3. It is also true that any subgroup of a quotient of n∈N U(n) (or p∈P SU(p)
for that matter) is hyperlinear. This follows from results in [44], based on ideas from work
of Nikolov-Segal, see [35]. It is not known if there are any groups that are not hyperlinear –
essentially all groups that are known to by hyperlinear, are also known to be sofic.
3.2. Proofs of the main results. We can now prove the main theorems, i.e., Theorem
1.5 and Theorem 1.6. Let us first study the commutator map again. It follows from Corollary
2.10 that
c∗ (x2 · · · xp ) = a · (y p−1 ⊗ y1 · · · yp−1 )
(5)
for some a ∈
Z×
(p) ,
where J is the ideal defined as
X
Y
Y ≥(2,2i−1)
J :=
(y ⊗ yi ) ·
Ji
,
E({1,...,p−1} i∈E
≥(2,2i−1)
mod J
i6∈E
where Ji
denotes the subspace of Ji , which is of bi-degree (2, 2i − 1) or more. Here,
we use the natural bi-grading of the tensor product H ∗ (PU(p), Z/pZ) ⊗Z H ∗ (PU(p), Z/pZ).
It is easy to see that y p−1 ⊗ y1 · · · yp−1 6∈ J . Indeed, assume that y p−1 ⊗ y1 · · · yp−1 ∈ J . The
EQUATIONS OVER GROUPS
13
element y p−1 ⊗ y1 · · · yp−1 is of bi-degree (2(p − 1), 1 + 3 + · · · + 2p − 1) so that all contributions
≥(2,2i−1)
of Ji
must of of minimal possible degree (2, 2i − 1). Note that y ⊗ yi 6∈ Ji by
construction. This implies that for any E ⊂ {1, . . . , p − 1}, a product of p − 1 factors of the
(2,2i−1)
form y ⊗ yi for i ∈ E and otherwise in Ji
cannot yield a summand y p−1 ⊗ y1 . . . yp−1
unless E = {1, . . . , p − 1}, but this is forbidden.
This proves as a first step that any continuous map which is homotopic to the commutator
map c : PU(p) × PU(p) → SU(p) must be surjective. Indeed, the previous computation shows
2
that its effect on the fundamental class x2 · · · xp ∈ H p −1 (SU(p), Z/pZ) is non-trivial on
cohomology – and this happens only if the map (and any map homotopic to it) is surjective.
We now attempt to extend this result to other words in F2 . Later, we will also allow to
vary p and will see that our approach works for all elements which do not lie in [F2 , [F2 , F2 ]].
(1)
First, we need some preparations. Note that F2 := [F2 , F2 ] is a free group with basis
{[xn1 , xm
2 ] | nm 6= 0}, see Proposition 4 in Chapter I, §1.3 of [43].
Proposition 3.4. Let w = [xn1 , xm
2 ]. Then
w∗ : H ∗ (SU(p), Z/pZ) → H ∗ (PU(p), Z/pZ) ⊗ H ∗ (PU(p), Z/pZ)
satisfies w∗ (xi ) = nm · ai · (y ⊗ yi−1 ) mod Ji−1 for some ai ∈ Z×
(p) independent of n, m. More
Qs
mk lk
nk
generally, if w = k=1 [x1 , x2 ] , then
!
s
X
∗
w (xi ) =
nk mk lk · ai · (y ⊗ yi−1 ) mod Ji−1 .
k=1
On the other side, if w ∈
(2)
F2 ,
the second derived subgroup, then we have w∗ = 0.
Proof. If w1 : PU(p) × PU(p) → SU(p) and w2 : PU(p) × PU(p) → SU(p) are word maps,
for w1 , w2 ∈ [F2 , F2 ], so that the associated word map can be factored through SU(p), then
w1 w2 : PU(p) × PU(p) → SU(p)
is equal to
mSU(p) ◦ (w1 × w2 ) ◦ ∆PU(p)×PU(p) : PU(p) × PU(p) → SU(p).
Hence, the effect on cohomology can be computed explicitly as follows:
xi 7→ xi ⊗ 1 + 1 ⊗ xi
7→ w1∗ (xi ) ⊗ 1 + 1 ⊗ w2∗ (xi )
7→ w1∗ (xi ) + w2∗ (xi ).
The results follow directly from Lemma 2.3 and Corollary 2.10. This finishes the proof of the
proposition.
Consider now the 3-dimensional Heisenberg group
1 a b
H3 (Z) := 0 1 c | a, b, c ∈ Z .
0 0 1
Sending x1 to the matrix with a = 1, b = c = 0 and x2 to the matrix with a = b = 0, c = 1,
Pl
Q
k lk
we get that w = lk=1 [xn1 k , xm
2 ] is sent to the matrix with a = c = 0 and b =
k=1 nk mk lk .
It is well known that H3 (Z) = F2 /[[F2 , F2 ], F2 ]. Coming back to the proof of Theorem 1.6,
14
ANTON KLYACHKO AND ANDREAS THOM
we see that we will succeed with our strategy if w can be mapped non-trivially to the central
quotient of H3 (Z) by pZ.
Proof of Theorem 1.6. The result is clear if ε(w) 6∈ [F2 , F2 ], so that we may assume
that ε(w) ∈ [F2 , F2 ]. Now, the assumption on ε(w) implies that it can be mapped nontrivially to the canonical central extension of Z2 by Z/pZ, which is just the central quotient
of H3 (Z)/pZ. This happens if any only if ε(w) 6∈ [F2 , F2 ]p [[F2 , F2 ], F2 ]. Hence, by our
main result above the induced word map w : PU(p) × PU(p) → SU(p) is is non-trivial on
2
the fundamental class x1 . . . xn ∈ H p −1 (SU(p), Z/pZ), and hence the word map must be
(1)
surjective. Since ε(w) 6∈ F2 , any lift to SU(p) of a preimage of the neutral element solves
the equation.
Proof of Theorem 1.5. This is a straightforward consequence of Theorem 1.6 and
Proposition 3.2. The claim about finite groups follows from Mal’cev’s theorem [31], stating
that finitely generated linear groups are residually finite.
3.3. Related results and low-dimensional cases. Let us finish by mentioning a few
low-dimensional results which go beyond the second step of the lower central series. So far, we
are unable to exploit the mechanisms behind these examples in order to get satisfactory results
for all hyperlinear groups. However, we would also like to mention some further directions
and possible extensions of the techniques used in this paper
Recall, we denote the commutator by [x, y] := xyx−1 y −1 . The iterated commutators
cn ∈ Fn = hx1 , . . . , xn i are defined by induction cn = [xn , cn−1 ] and c1 = x1 . The first result
that goes beyond the second step in the lower central series is the following result by Porter.
SU(2)
Theorem 3.5 (Porter, see [37]). The map c3
: SU(2)×3 → SU(2) is not null-homotopic.
In order to treat cn for n ≥ 4, we need to use some more sophisticated results from
algebraic topology related to homotopy nilpotence results. This was done by Rao in [38],
showing also that Spin(n) is not homotopy nilpotent for n ≥ 7.
PU(2)
Theorem 3.6 (Rao, [38]). The map cn
any n ∈ N.
: PU(2)×n → PU(2) is not null-homotopic for
As before, we obtain results concerning solvability of equations.
Corollary 3.7. Let G be any subgroup of SU(2) and let w ∈ G∗Fn be such that ε(w) = cn
for some n ≥ 2. Then, w(x1 , . . . , xn ) = 1 can be solved in some group containing G.
Proof. Since SU(2) = S 3 , every non-surjective map is null-homotopic. Thus, using the
same arguments as before, we can conclude that for every w ∈ SU(2) ∗ Fn with content cn ,
the induced word map w : SU(2)×n → SU(2) is surjective.
Finally, we want to mention some questions that appear naturally at this interface between
homotopy theory and the study of word maps. Given a topological group, it seems natural
to study the group of words modulo those which are null-homotopic. Let G be a compact Lie
group, set
Nn,G := {w ∈ Fn | w : Gn → G is homotopically trivial}
and define Hn,G := Fn /Nn,G .
Question 3.8. Can we compute H2,SU(n) ?
EQUATIONS OVER GROUPS
15
See [23, 49] for partial information about Hn,G in particular cases. In this direction, the
following result is implied by results on p.464 in [48].
Theorem 3.9 (Whitehead, see [48]). Let G be a connected and simply connected compact
Lie group. Then, HG is k-step nilpotent for some k ≤ 2 · dim(G).
Proof. We denote the degree of nilpotency of a group Γ by nil(Γ). Whitehead showed
that the homotopy set [X, G] is a group and
nil([X, G]) ≤ dim(X).
For X = G × G, we obtain nil([G × G, G]) ≤ dim(G × G) = 2 · dim(G). Now, the subgroup
generated by the coordinate projections is precisely HG . This proves the claim.
Let G be a topological group, e.g., a compact Lie group. We call w ∈ Fn homotopically
surjective with respect to G if every map in the homotopy class of w : G×n → G is surjective.
Question 3.10. Let w ∈ Fn \{1}. Is w : PU(n)×PU(n) → PU(n) homotopically surjective
for large n?
Acknowledgments
The work of the first author was supported by the Russian Foundation for Basic Research,
project no. 15-01-05823. He thanks TU Dresden for its hospitality during a visit in January
2015. He thanks this institution for its hospitality.
The second author wants to thank Tilman Bauer for some interesting comments. Part
of this paper was written, when he visited Institute Henri Poincaré for the trimester about
random walks and asymptotic geometry of groups in spring 2014. This research was supported
by ERC Starting Grant No. 277728 and the MPI-MIS in Leipzig.
We thank the unknown referee for many useful comments that improved the exposition.
References
[1] Dmitrii V. Baranov and Anton A. Klyachko, Efficient adjunction of square roots to groups, Sibirsk. Mat.
Zh. 53 (2012), no. 2, 250–257 (Russian, with Russian summary); English transl., Sib. Math. J. 53 (2012),
no. 2, 201–206. ↑2
[2] Paul F. Baum and William Browder, The cohomology of quotients of classical groups, Topology 3 (1965),
305–336. ↑5, 6
[3] Armand Borel, Sur la cohomologie des espaces fibrés principaux et des espaces homogènes de groupes de
Lie compacts, Ann. of Math. (2) 57 (1953), 115–207 (French). ↑5
[4]
, Sur l’homologie et la cohomologie des groupes de Lie compacts connexes, Amer. J. Math. 76
(1954), 273–342 (French). ↑5
[5] Raoul Bott, The space of loops on a Lie group, Michigan Math. J. 5 (1958), 35–61. ↑7
[6]
, A note on the Samelson product in the classical groups, Comment. Math. Helv. 34 (1960), 249–
256. ↑8
[7] Aldridge K. Bousfield and Daniel M. Kan, Homotopy limits, completions and localizations, Lecture Notes
in Mathematics, Vol. 304, Springer-Verlag, Berlin, 1972. ↑7
[8] Martin Edjvet and Ayre Juhász, Non-singular equations over groups I, Algebra Colloq. 18 (2011), no. 2,
221–240. ↑2
[9]
, Nonsingular equations over groups II, Comm. Algebra 38 (2010), no. 5, 1640–1657. ↑2
[10] Martin Edjvet and James Howie, The solution of length four equations over groups, Trans. Amer. Math.
Soc. 326 (1991), no. 1, 345–369. ↑2, 4
[11] Abdelrhman Elkasapy and Andreas Thom, About Gôto’s method showing the surjectivity of word maps,
arXiv:1207.5596, to appear in Indiana Univ. Math. J. ↑4, 5
16
ANTON KLYACHKO AND ANDREAS THOM
[12] Anastasia Evangelidou, The solution of length five equations over groups, Comm. Algebra 35 (2007), no. 6,
1914–1948. ↑4
[13] Steve M. Gersten, Reducible diagrams and equations over groups, Essays in group theory, Math. Sci. Res.
Inst. Publ., vol. 8, Springer, New York, 1987, pp. 15–73. ↑2, 4
[14] Murray Gerstenhaber and Oscar S. Rothaus, The solution of sets of equations in groups, Proc. Nat. Acad.
Sci. U.S.A. 48 (1962), 1531–1533. ↑1, 2, 3
[15] Misha Gromov, Endomorphisms of symbolic algebraic varieties, J. Eur. Math. Soc. (JEMS) 1 (1999),
no. 2, 109–197. ↑12
[16] Hiroaki Hamanaka, Daisuke Kishimoto, and Akira Kono, Self homotopy groups with large nilpotency
classes, Topology Appl. 153 (2006), no. 14, 2425–2429. ↑5, 6
[17] Allen Hatcher, Algebraic topology, Cambridge University Press, Cambridge, 2002. ↑7
[18] Graham Higman, A finitely generated infinite simple group, J. London Math. Soc. 26 (1951), 61–64. ↑4
[19] Heinz Hopf, Über den Rang geschlossener Liescher Gruppen, Comment. Math. Helv. 13 (1940), 119–143.
↑3
[20] James Howie, On pairs of 2-complexes and systems of equations over groups, J. Reine Angew. Math. 324
(1981), 165–174. ↑2, 3
[21]
, The solution of length three equations over groups, Proc. Edinburgh Math. Soc. (2) 26 (1983),
no. 1, 89–96. ↑2, 3
[22] Sergey V. Ivanov and Anton A. Klyachko, Solving equations of length at most six over torsion-free groups,
J. Group Theory 3 (2000), no. 3, 329–337. ↑2
[23] Ioan James and Emery Thomas, Which Lie groups are homotopy-abelian?, Proc. Nat. Acad. Sci. U.S.A.
45 (1959), 737–740. ↑15
[24] Arye Juhász, On the solvability of equations over groups, Comm. Algebra 32 (2004), no. 4, 1487–1496. ↑2
[25] Daisuke Kishimoto and Akira Kono, On a conjecture of Ōshima, Topology Appl. 156 (2009), no. 13,
2189–2192. ↑2, 5, 9, 10
[26] Anton A. Klyachko, A funny property of sphere and equations over groups, Comm. Algebra 21 (1993),
no. 7, 2555–2575. ↑2, 4
[27]
, Equations over groups, quasivarieties, and a residual property of a free group, J. Group Theory
2 (1999), no. 3, 319–327. ↑2
[28]
, How to generalize known results on equations over groups, Mat. Zametki 79 (2006), no. 3, 409–419
(Russian, with Russian summary); English transl., Math. Notes 79 (2006), no. 3-4, 377–386. ↑4
[29] Anton A. Klyachko and Anton V. Trofimov, The number of non-solutions of an equation in a group, J.
Group Theory 8 (2005), no. 6, 747–754. ↑2
[30] Frank Levin, Solutions of equations over groups, Bull. Amer. Math. Soc. 68 (1962), 603–604. ↑2, 4
[31] Anatoly Malcev, On isomorphic matrix representations of infinite groups, Rec. Math. [Mat. Sbornik] N.S.
8 (50) (1940), 405–422. ↑14
[32] Peter May and Kate Ponto, More concise algebraic topology, Chicago Lectures in Mathematics, University
of Chicago Press, Chicago, IL, 2012. Localization, completion, and model categories. ↑7
[33] Mamoru Mimura, Goro Nishida, and Hirosi Toda, Localization of CW-complexes and its applications, J.
Math. Soc. Japan 23 (1971), 593–624. ↑7, 9
[34] Bernhard H. Neumann, Adjunction of elements to groups, J. London Math. Soc. 18 (1943), 4–11. ↑2
[35] Nikolay Nikolov and Dan Segal, Generators and commutators in finite groups; abstract quotients of compact
groups, Invent. Math. 190 (2012), no. 3, 513–602. ↑12
[36] Vladimir G. Pestov, Hyperlinear and sofic groups: a brief guide, Bull. Symbolic Logic 14 (2008), no. 4,
449–480. ↑2, 3, 12
[37] Gerald J. Porter, Homotopical nilpotence of S 3 , Proc. Amer. Math. Soc. 15 (1964), 681–682. ↑14
[38] Vidhyanath K. Rao, Spin(n) is not homotopy nilpotent for n ≥ 7, Topology 32 (1993), no. 2, 239–249.
↑14
[39] Vitaliı̆ Roman’kov, Equations over groups, Groups Complex. Cryptol. 4 (2012), no. 2, 191–239. ↑2
[40] Oskar S. Rothaus, On the non-triviality of some group extensions given by generators and relations, Ann.
of Math. (2) 106 (1977), no. 3, 599–612. ↑4
[41] Hans Samelson, Groups and spaces of loops, Comment. Math. Helv. 28 (1954), 278–287. ↑5, 8
[42] Jean-Pierre Serre, Groupes d’homotopie et classes de groupes abéliens, Ann. of Math. (2) 58 (1953),
258–294 (French). ↑8
EQUATIONS OVER GROUPS
[43]
[44]
[45]
[46]
[47]
[48]
[49]
17
, Trees, Springer Monographs in Mathematics, Springer-Verlag, Berlin, 2003. Translated from the
French original by John Stillwell; Corrected 2nd printing of the 1980 English translation. ↑13
Abel Stolz and Andreas Thom, On the lattice of normal subgroups in ultraproducts of compact simple
groups, Proc. London Math. Soc. 108 (2014), 73–102. ↑12
Stephan Stolz and Peter Teichner, What is an elliptic object?, Topology, geometry and quantum field
theory, London Math. Soc. Lecture Note Ser., vol. 308, Cambridge Univ. Press, Cambridge, 2004, pp. 247–
343. ↑5
Andreas Thom, Convergent sequences in discrete groups, Canad. Math. Bull. 56 (2013), no. 2, 424–433.
↑4, 5
Benjamin Weiss, Sofic groups and dynamical systems, Sankhyā Ser. A 62 (2000), no. 3, 350–359. Ergodic
theory and harmonic analysis (Mumbai, 1999). ↑12
George W. Whitehead, Elements of homotopy theory, Graduate Texts in Mathematics, vol. 61, SpringerVerlag, New York, 1978. ↑15
Nobuaki Yagita, Homotopy nilpotency for simply connected Lie groups, Bull. London Math. Soc. 25 (1993),
no. 5, 481–486. ↑15
A.K., Moscow State University, Russia.
E-mail address: [email protected]
A.T., Institut für Geometrie, TU Dresden, Germany
E-mail address: [email protected]
| 4math.GR
|
Biconnectivity, st-numbering and other applications of DFS using O(n) bits I
Sankardeep Chakrabortya , Venkatesh Ramana , Srinivasa Rao Sattib
a The
Institute of Mathematical Sciences, HBNI, Chennai, India
b Seoul National University, Seoul, South Korea
arXiv:1606.08645v3 [cs.DS] 27 Jul 2017
Abstract
We consider space efficient implementations of some classical applications of DFS including the problem of
testing biconnectivity and 2-edge connectivity, finding cut vertices and cut edges, computing chain decomposition
and st-numbering of a given undirected graph G on n vertices and m edges. Classical algorithms for them typically
use DFS and some Ω(lg n) bits1 of information at each vertex. Building on a recent O(n)-bits implementation of
DFS due to Elmasry et al. (STACS 2015) we provide O(n)-bit implementations for all these applications of DFS.
Our algorithms take O(m lgc n lg lg n) time for some small constant c (where c ≤ 2). Central to our implementation
is a succinct representation of the DFS tree and a space efficient partitioning of the DFS tree into connected
subtrees, which maybe of independent interest for designing other space efficient graph algorithms.
1. Introduction
Space efficient algorithms are becoming increasingly important owing to their applications in the presence of
rapid growth of “big data” and the proliferation of specialized handheld devices and embedded systems that have
a limited supply of memory. Even if mobile devices and embedded systems are designed with large supply of
memory, it might be useful to restrict the number of write operations. For example, on flash memory, writing
is a costly operation in terms of speed, and it also reduces the reliability and longevity of the memory. Keeping
all these constraints in mind, it makes sense to consider algorithms that do not modify the input and use only a
limited amount of work space. One computational model that has been proposed in algorithmic literature to study
space efficient algorithms, is the read-only memory (ROM) model. In this article, we focus on space efficient
implementations of some fundamental graph algorithms in such settings without paying too much penalty on
time.
There is already a rich history of designing space efficient algorithms in the read-only memory model. The
complexity class L (also known as DLOGSPACE) is the class containing decision problems that can be solved by
a deterministic Turing machine using only logarithmic amount of work space for computation. There are several
important algorithmic results [29, 33, 34, 35] for this class, the most celebrated being Reingold’s method [62] for
checking st-reachability in an undirected graph, i.e., to determine if there is a path between two given vertices
s and t. NL is the non-deterministic analogue of L and it is known that the st-reachability problem for directed
graphs is NL-complete (with respect to log space reductions). Using Savitch’s algorithm [1], this problem can
be solved in nO(lg n) time using O(lg2 n) bits. Savitch’s algorithm is very space efficient but its running time is
superpolynomial. Among the deterministic algorithms running in polynomial time for directed st-reachability,
√ the
Θ( lg n)
most space efficient algorithm is due to Barnes et al. [9] who gave a slightly sublinear space (using n/2
bits) algorithm for this problem running in polynomial time. We know of no better polynomial time algorithm for
this problem with better space bound. Moreover, the space used by this algorithm matches a lower bound on space
for solving directed st-reachability on a restricted model of computation called Node Naming Jumping Automata
on Graphs (NNJAG) [24, 32]. This model was introduced especially for the study of directed st-reachability and
most of the known sublinear space algorithms for this problem
√ can be implemented on it. Thus, to design any
polynomial time ROM algorithm taking space less than n/2Θ( lg n) bits requires significantly new ideas. Recently
1 We
use lg to denote logarithm to the base 2.
of these results were announced in preliminary form in the proceedings of 27th International Symposium on Algorithms and
Computation (ISAAC 2016) LIPIcs, Volume 64, pages 22:1-22:13 [19].
Email addresses: [email protected] (Sankardeep Chakraborty), [email protected] (Venkatesh Raman),
[email protected] (Srinivasa Rao Satti)
I Some
Preprint submitted to Elsevier
July 28, 2017
there has been some improvement in the space bound for some special classes of graphs like planar and Hminor free graphs [4, 14]. Other than these fundamental graph theoretical problems, there have been some work
on designing space-efficient algorithms for the more classical selection and sorting problems [10, 57, 58], and
problems in computational geometry [2, 7, 8, 27] among others.
A drawback, however, for all these graph algorithms using small space i.e., sublinear bits, is that their running
time is often some polynomial of high degree. For example, to the best of our knowledge, the exact running time of
Reingold’s algorithm [62] for undirected s-t connectivity is not analysed, yet we know it admits a large polynomial
running time. This is not surprising as Tompa [69] showed that for directed st-reachability, if the number of
bits available is o(n) then some natural algorithmic approaches to the problem require super-polynomial time.
Motivated by these impossibility results from complexity theory and inspired by the practical applications of these
fundamental graph algorithms, recently there has been a surge of interest in improving the space complexity of
the fundamental graph algorithms without paying too much penalty in the running time i.e., reducing the working
space of the classical graph algorithms to O(n) bits with little or no penalty in running time. Generally most of the
classical linear time graph algorithms take O(n) words or equivalently O(n lg n) bits of space.
Starting with the paper of Asano et al. [3] who showed how one can implement DFS using O(n) bits, improving on the naive O(n lg n)-bit implementation, the recent series of papers [3, 5, 6, 16, 19, 18, 37] presented
space-efficient algorithms for a few other basic graph problems: namely BFS, maximum cardinality search, topological sort, connected components, minimum spanning tree, shortest path, recognition of outerplanar graph and
chordal graphs among others. We add to this small yet growing body of space-efficient algorithm design literature
by providing such algorithms for some classical graph problems that have been solved using DFS, namely the
problem of testing biconnectivity and 2-edge connectivity, finding cut vertices and cut edges, computing chain
decomposition and st-numbering among others.
1.1. Model of Computation
As is standard in the area of space-efficient graph algorithms [3, 6, 5, 37], we assume that the input graph is
given in a read-only memory (and so cannot be modified). If an algorithm must do some outputting, this is done
on a separate write-only memory. When something is written to this memory, the information cannot be read or
rewritten again. So the input is “read only” and the output is “write only”. In addition to the input and the output
media, a limited random-access workspace is available. The data on this workspace is manipulated at word level
as in the standard word RAM model, where the machine consists of words of size w = Ω(lg n) bits; and any
logical, arithmetic, and bitwise operations involving a constant number of words take a constant amount of time.
We count space in terms of the number of bits in the workspace used by the algorithms. Historically, this model is
called the register input model and it was introduced by Frederickson [43] while studying some problems related
to sorting and selection.
We assume that the input graphs G = (V, E) are represented using an adjacency array, i.e., G is represented by
an array of length |V| where the i-th entry stores a pointer to an array that stores all the neighbors of the i-th vertex.
For the directed graphs, we assume that the input representation has both in/out adjacency array for all the vertices
i.e., for directed graphs, every vertex v has access to two arrays, one array is for all the in-neighbors of v and the
other array is for all the out-neighbors of v. This representation which has now become somewhat standard was
also used in [5, 19, 37, 49] recently to design various other space efficient graph algorithms. We use n and m to
denote the number of vertices and the number of edges respectively, in the input graph G. Throughout the paper,
we assume that the input graph is a connected graph, and hence m ≥ n − 1.
1.2. Our results and organization of the paper
Asano et al. [3] showed that Depth First Search (DFS) in a directed or an undirected graph can be performed
in O(m lg n) time and O(n) bits of space. Elmasry et al. [37] improved the time to O(m lg lg n) still using O(n) bits
of space. We build upon these results to give space efficient implementations of several classical applications of
DFS.
First, as a warm up, we start with some simple applications of the space efficient DFS to show the following.
• An O(m lg n lg lg n) time and O(n) bits of space algorithm to compute the strongly connected components
of a directed graph in Section 3.1.
In addition, we also give
• an algorithm to output the vertices of a directed acyclic graph in a topologically sorted order in Section 3.2,
and
2
• an algorithm to find a sparse (with O(n) edges) spanning biconnected subgraph of an undirected biconnected
graph in Section 3.3
both using asymptotically the same time and space used for DFS, i.e., using O(n) bits and O(m lg lg n) time.
To develop fast and space efficient algorithms for other non-trivial graph problems which are also applications
of DFS, in Section 4, we develop and describe in detail a space efficient tree covering technique, and use this
in subsequent sections. This technique, roughly speaking, partitions the DFS tree into connected smaller sized
subtrees which can be stored using less space. Finally we solve the corresponding graph problem on these smaller
sized subtrees and merge the solutions across the subtrees to get an overall solution. All of these can be done using
less space and not paying too much penalty in the running time. Some of these ideas are borrowed from succinct
tree representation literature.
As the first application, we consider in Section 5.1, a space efficient implementation of chain decomposition of
an undirected graph. This is an important preprocessing routine for an algorithm to find cut vertices, biconnected
components, cut edges, and also to test 3-connectivity [63] among others. We provide an algorithm that takes
O(m lg2 n lg lg n) time using O(n) bits of space, improving on previous implementations that took Ω(n lg n) bits [64]
or Θ(m + n) bits [5] of space.
In Section 5.2, we give improved space efficient algorithms for testing whether a given undirected graph G
is biconnected, and if G is not biconnected, we also show how one can find all the cut vertices of G. For this,
we provide a space efficient implementation of Tarjan’s classical lowpoint algorithm [66]. Our algorithms take
O(m lg n lg lg n) time and O(n) bits of space. In Section 5.3, we provide a space efficient implementation for testing
2-edge connectivity of a given undirected graph G, and producing cut edges of G using O(m lg n lg lg n) time and
O(n) bits of space.
Given a biconnected graph, and two distinguished vertices s and t, st-numbering is a numbering of the vertices
of the graph so that s gets the smallest number, t gets the largest and every other vertex is adjacent both to a lowernumbered and to a higher-numbered vertex. Finding an st-numbering is an important preprocessing routine for a
planarity testing algorithm [38] among others. In Section 5.4, we give an algorithm to determine an st-numbering
of a biconnected graph that takes O(m lg2 n lg lg n) time using O(n) bits. This improves the earlier implementations
that take Ω(n lg n) bits [11, 31, 38, 68]. Using this as a subroutine, in Section 5.5, we provide improved space
effcient implementation for two-partitioning and two independent spanning tree problem among others. We direct
the readers to Section 2.1 where we provide all the necessary definitions.
1.3. Related Models
Several models of computation come close to read-only random-access model, the model we focus on this
paper, when it comes to design space-efficient graph algorithms. A single thread common to all of them is that
access to the input tape is restricted in some way. In the multi-pass streaming model [57] the input is kept in a
read-only sequentially-accessible media, and an algorithm tries to optimize on the number of passes it makes over
the input. In the semi-streaming model [42], the elements (or edges if the input is graph) are revealed one by
one and extra space allowed to the algorithm is O(n.polylg(n)) bits. Observe that, it is not possible to store the
whole graph if it is dense. The efficiency of an algorithm in this model is measured by the space it uses, the time it
requires to process each edge and the number of passes it makes over the stream. In the in-place model [12], one is
allowed a constant number of additional variables, but it is possible to rearrange (and sometimes even modify) the
input values. Chan et al. [20] introduced the restore model which is a more relaxed version of read-only memory
(and a restricted version of the in-place model), where the input is allowed to be modified, but at the end of the
computation, the input has to be restored to its original form. This has motivation, for example, in scenarios where
the input (in its original form) is required by some other application. Buhrman et al. [13, 54] introduced and
studied the catalytic-space model where a small amount (typically O(lg n) bits) of clean space is provided along
with additional auxiliary space, with the condition that the additional space is initially in an arbitrary, possibly
incompressible, state and must be returned to this state when the computation is finished. The input is assumed to
be given in ROM. They also provided implementations of some graph algorithms space efficiently.
2. Preliminaries
In this section, we list some preliminary results and graph theoretic definitions that will be used later in the
algorithms we develop. We also discuss briefly, at a very high level, the main technique that goes behind almost
all of our algorithms in this paper.
3
2.1. Graph theoretic terminology
Here we collect all the necessary graph theoretic definitions that will be used throughout the paper. A cut vertex
in an undirected graph G is a vertex v that when removed (along with its incident edges) from a graph creates more
components than previously in the graph. A (connected) graph with at least three vertices is biconnected (also
called 2-connected in the graph literature sometimes) if and only if it has no cut vertex. A biconnected component
is a maximal biconnected subgraph. These components are attached to each other at cut vertices. Similarly in an
undirected graph G, a bridge (or cut edge) is an edge that when removed (without removing the vertices) from a
graph creates more components than previously in the graph. A (connected) graph with at least two vertices is 2edge-connected (also called bridgeless sometimes) if and only if it has no bridge. A 2-edge connected component
is a maximal 2-edge connected subgraph.
Given a biconnected graph G, and two distinguished vertices s and t in V such that s , t, st-numbering is a
numbering of the vertices of the graph so that s gets the smallest number, t gets the largest and every other vertex
is adjacent both to a lower-numbered and to a higher-numbered vertex i.e., a numbering s = v1 , v2 , · · · , vn = t of
the vertices of G is called an st-numbering, if for all vertices v j , 1 < j < n, there exist 1 ≤ i < j < k ≤ n such
that {vi , v j }, {v j , vk } ∈ E. It is well-known that G is biconnected if and only if, for every edge {s, t} ∈ E, it has an
st-numbering. In the k-partitioning problem, we are given vertices a1 , · · · , ak of an undirected graph G and natural
numbers c1 , · · · , ck with c1 + · · · + ck = n, and we want to find a partition of V into sets V1 , · · · , Vk with ai ∈ Vi
and |Vi | = ci for every i such that every set Vi induces a connected graph in G. Given a graph G, we call a set of
k rooted spanning trees independent if they all have the same root vertex r and, for every vertex v , r, the paths
from v to r in all the k spanning trees are vertex-disjoint (except for their endpoints). A directed graph G is said to
be strongly connected if for every pair of vertices u and v in V, both u and v are reachable from each other. If G is
not strongly connected, it is possible to decompose G into its strongly connected components i.e. a maximal set
of vertices C ⊆ V such that for every pair of vertices u and v in C, both u and v are reachable from each other. A
topological sort or topological ordering of a directed acyclic graph is a linear ordering of its vertices such that for
every directed edge (u, v) ∈ E from vertex u to vertex v, u comes before v in the ordering. Let T be a depth-first
search tree of a connected undirected (or directed) graph G. For each vertex v of T , preorder number of v is the
number of vertices visited up to and including v during a preorder traversal of T . Similarly, postorder number of
v is the number of vertices visited up to and including v during a postorder traversal of T .
2.2. Tree cover and its space efficient construction
To implement our algorithms in O(n) bits, our main idea is to process the nodes of the DFS tree in batches
of O(n/ lg n) nodes, as we can only afford to store trees of size O(n/ lg n) explicitly with their labels. To do
this, we first use a tree-cover algorithm (that is used in succinct representations of trees) to partition the tree into
O(lg n) connected subtrees of size O(n/ lg n) each. We then solve the problem we are dealing with in these smaller
subtrees, and later merge them in a specific order to obtain the overall solution. In some cases, to obtain the overall
solution, we need to generate pairs of subtrees with explicit node labels, and then process the edges between them
in a specific order. We describe all the details of the tree cover approach in Section 4, and describe the algorithms
in Section 5.
2.3. Rank-Select
We use the following fundamental data structure on bitstrings in some of our algorithms. Given a bitvector B
of length n, the rank and select operations are defined as follows:
• ranka (i, B) = number of occurrences of a ∈ {0, 1} in B[1, i], for 1 ≤ i ≤ n;
• selecta (i, B) = position in B of the i-th occurrence of a ∈ {0, 1}.
The following theorem gives an efficient structure to support these operations.
Theorem 1 ([23, 56, 60]). Given a bitstring B of length n, one can construct a o(n)-bit auxiliary structure to
support rank and select operations in O(1) time. Also, such a structure can be constructed from the given bitstring
in O(n) time.
2.4. Related work on Space-efficient DFS
Recall that DFS starts exploring the given input graph G where each vertex is initially white meaning unexplored, becomes gray when DFS discovers for the first time and pushed on the stack, and is colored black when it
is finished i.e. its adjacency list has been checked completely and it leaves the stack. Recently Elmasry et al. [37]
showed the following tradeoff result for DFS,
4
Theorem 2 ([37]). For every function t : N → N such that t(n) can be computed within the resource bound of
this theorem (e.g., in O(n) time using O(n) bits), the vertices of a directed or undirected graph G can be visited in
lg n
) bits.
depth first order in O((m + n)t(n)) time with O(n + n lgt(n)
In particular, fixing t(n) = O(lg lg n), one can obtain a DFS implementation which runs in O(m lg lg n) time
using O(n) bits. We build on top of this DFS algorithm to provide space efficient implementation for various
applications of DFS in directed and undirected graphs in the rest of this paper.
3. Some simple applications of DFS using O(n) bits 2
Classical applications of DFS in directed graphs (see [25]) are to find strongly connected components of a
directed graph, and to do a topological sort of a directed acyclic graph among many others. Also, given an
undirected biconnected graph G, DFS is used as the main tool to produce a sparse spanning biconnected subgraph
of G. We show here that while topological sort and producing a sparse spanning biconnected subgraph of an
undirected biconnected graph can be solved using the same O(n) bits and O(m lg lg n) time (as for DFS), strongly
connected components of a directed graph can be obtained using O(n) bits and O(m lg n lg lg n) time.
3.1. Strongly Connected Components
There is a classical two pass algorithm (see [25] or [28]) for computing the Strongly Connected Components
(SCC) of a given directed graph G which works as follows. In the first step, it runs a DFS on GR , the reverse graph
of G. In the second pass, it runs the connected component algorithm using DFS in G but it processes the vertices
in the decreasing order of the finishing time from the first pass.
We can obtain GR by switching the role of in and out adjacency arrays present in the input representation. As
we can not remember the vertex ordering from the first pass due to space restriction, we process them in batches
of size n/ lg n in the reverse order i.e., we run a full DFS in GR to obtain and store the last n/ lg n vertices in an
array A as they are the ones which have the highest set of finishing numbers in decreasing order. I.e., we maintain
A as a queue of size n/ lg n and as and when a new element is finished, it is added to the queue and the element
with the earliest finish time at the other end of the queue is deleted. Now, we pick the vertices from A one by
one in the order from the queue with the latest finish time and start a fresh DFS in G to compute the connected
components and output all the vertices reachable as a SCC. The output vertices are marked in a bitmap so that we
don’t output them again. Once we are done with all the vertices in A, we restart the DFS from the beginning and
produce the next chunk of n/ lg n vertices by remembering the last vertex produced in the previous step and stop
as soon as we hit that boundary vertex. Then we repeat the connected component algorithm from this chunk of
vertices and continue this way. It is clear that the algorithm produces the SCCs correctly. As we are calling the
DFS algorithm O(lg n) times, total time taken by this algorithm is O(m lg lg n lg n) with O(n) bits of space. Hence,
we have the following,
Theorem 3. Given a directed graph G on n vertices and m edges, represented as in/out adjacency array, we can
output the strongly connected components of G in O(m lg n lg lg n) time and O(n) bits of space.
3.2. Topological Sort
The standard algorithm for computing topological sort [25] outputs the vertices of a DFS in reverse order. If
we can keep track of the DFS numbers, then reversing is an easy task. While working in space restricted setting
(with o(n lg n) bits), this is a challenge as we don’t have space to keep track of the DFS order. We can do as we did
in the strongly connected components algorithm in the last section, by storing and outputting vertices in batches
of n/ lg n resulting in an O(m lg n lg lg n) time algorithm.
Elmasry et al.[37] showed that, the vertices of a DAG G can be output in the order of a topological sort within
the time and space bounds of a DFS in G plus an additional O(n lg lg n) bits. As they also showed how to perform
DFS in O(m + n) time and O(n lg lg n) bits, overall their algorithm takes O(m + n) time and O(n lg lg n) bits to
compute a topological sorting of G. Their main idea is to maintain enough information about a DFS to resume it
in the middle and apply this repeatedly to reverse small chunks of its output, produced in reverse order, one by
one.
We observe that, instead of storing information to restart DFS and produce the reverse order, we simply work
with the reverse graph itself (which can be obtained from the input representation by switching the role of in and
2 The results of this section were announced in preliminary form in the proceedings of 22nd International Computing and Combinatorics
Conference (COCOON 2016), Springer LNCS volume 9797, pages 119-130 [5].
5
out adjacency arrays) and do a DFS in the reverse graph and output vertices as they are finished (or blackened)
i.e., in the increasing order of finishing time. To see the correctness of this procedure, note that the reverse graph
is also a DAG, and if (i, j) is an edge in the DAG G, then ( j, i) is an edge in the reverse graph and i will become
black before j while the algorithm performs DFS in the reverse graph. Hence, i will be placed before j in the
correct topological sorted order. Thus we have the following,
Theorem 4. Given a DAG G on n vertices and m edges, if the black vertices of the DFS of G can be output using
s(n) space and t(n) time, then its vertices can be output in topologically sorted order using O(s(n)) space and
O(t(n)) time assuming that the input representation has both the in and out adjacency array of the graph.
From Theorem 2 (setting t(n) = O(lg lg n)) and Theorem 4, we have the following.
Corollary 1. Given a DAG G on n vertices and m edges, its vertices can be output in topologically sorted order
using O(m lg lg n) time and O(n) bits.
Note that, we knew all along that DFS and topological sort take the same time, the main contribution of
Theorem 4 is that it shows they take the same space (improving on the result of [37] where they showed that
topological sort space = DFS space + O(n lg lg n) bits under the same time) when both the in/out adjacency arrays
are present in the input.
3.2.1. Topological Sort in Sublinear Space
We note the following theorem of Asano et al. [3].
Theorem 5. DFS on a DAG G can be performed in space O(
√n
2(
lg n)
) bits and in polynomial time.
While it should immediately follow from Theorem 4 that topological sort can also be performed using such
sublinear bits of space, there is one caveat. Asano et al.’s algorithm works assuming that the given DAG G has a
single source vertex. In particular, they determine whether a vertex is black by checking whether it is reachable
from the source without using the gray vertices (using the sublinear space reachability algorithm of [9]).
The algorithm can be easily extended to handle s many sources if we have some additional s log n bits. We
simply keep track of the indices of the sources from which DFS has been explored, and to determine whether a
vertex is black, we ask if it is reachable from an earlier source or from the current source without using the gray
vertices. Thus we have the following improved theorem.
Theorem 6. DFS on DAG G with s sources can be performed using s lg n + o(n) bits and polynomial time. In
particular, if s is o(n/ lg n), the overall space used is o(n) bits.
Thus from Theorem 4 and Theorem 6 we obtain the following,
Theorem 7. Topological Sort on a DAG G with s sinks can be performed using s lg n + o(n) bits and polynomial
time. In particular if s is o(n/ lg n), the overall space used is o(n) bits.
3.3. Finding a sparse biconnected subgraph of a biconnected graph
The problem of finding a k-connected spanning subgraph with the minimum number of edges of a k-connected
graph is known to be NP-hard for any k ≥ 2 [45]. But the complexity of the problem decreases drastically if all
we want is to produce a “sparse” k-connected spanning subgraph, i.e., one with O(n) edges. Nagamochi and
Ibaraki [61] gave a linear time algorithm which produces a k-connected spanning subgraph with at most kn − k(k+1)
2
edges. Later, Cheriyan et al. [21] gave another linear time algorithm for k = 2 and 3 that produced a 2-connected
spanning subgraph with at most 2n − 2 edges, and a 3-connected subgraph with at most 3n − 3 edges. Later,
Elmasry [36] gave an alternate linear time algorithm for producing a sparse spanning biconnected subgraph of a
given biconnected graph by performing a DFS with additional bookkeeping. In what follows, we provide a space
efficient implementation for it. In order to do that, we start by briefly describing Elmasry’s algorithm.
Let DFI(v) denote the index (integer) that represents the time at which the vertex v is first discovered from
the vertex u when performing a DFS i.e., u is the parent of v in the DFS tree. Let low(v) be the smallest DFI
value among the DFI values of vertices w such that (v, w) is a back edge. (Note that this quantity is different
from the “lowpoint” value used in Tarjan’s [66] classical biconnectivity algorithm.) Basically low(v) captures the
information regarding the deepest back edge going out of the vertex v. If v has no backedges, for convenience (the
reason will become clear in the following lemma), we adopt the convention that low(v) = DFI(parent(v)). The
edge (v, low(v)) is the deepest backedge out of v. Note that, it is actually the tree edge between v and its parent if
v does not have a backedge. The algorithm maintains all the edges of the DFS tree. In addition, for every vertex
6
in the graph, the algorithm maintains the DFI and the low values along with the back edge that realizes it. As the
root of the DFS tree does not have any back edge and, as the underlying graph is 2-connected, the root has only
one child v so that there is no back edge emanating from v as well. Thus we get at most n − 2 back edges along
with n − 1 tree edges, giving a subgraph with at most 2n − 3 edges. Elmasry [36] proved that the resulting graph is
indeed a spanning 2-connected subgraph of G. His algorithm takes O(m + n) time and O(n lg n) bits of space. We
improve the space bound, albeit with slight increase in time, by first proving a more general lemma as following,
va
vb
vc
vi
Figure 1: A part of the full DFS tree. The wiggling edges represent tree edges and the edges with arrow heads represent back edges. If
low(vi ) = va , we would come across vi in the adjacency array of va before encountering from the arrays of vb and vc . I.e., the back edge (va , vi )
will be processed before the other back edges (vb , vi ) and (vc , vi ) since we process the vertices (and the backedges incident to them) in their
DFS order.
Lemma 1. Given any undirected graph G with n vertices and m edges, we can compute and report the low(v)
values i.e., deepest back edge going out of v, for every vertex v, using O(n) bits of space and O(m lg lg n) time.
Proof. The aim is to output all the deepest back edges out of every vertex v in G as we perform the DFS. As
always, let {v1 , v2 , · · · , vn } be the vertices of the graph. We perform a DFS with the usual color array and other
relevant data structures (as required in Theorem 2 with t(n) = lg lg n) along with one more array of n bits, which
we call DBE (for Deepest Back Edge) array, which is initialized to all zero. DBE[i] is set to 1 if and only if the
algorithm has found and output the deepest back edge emanating from vertex vi . So whenever a white vertex vi
becomes gray (i.e., vi is visited for the first time), we scan vi ’s adjacency array to mark, for every white neighbor
v j , DBE[ j] to 1 if and only if it was 0 before. The correctness of this step follows from the fact that as we are
visiting vertices in DFS order, and if DBE[ j] is 0, then vertex v j is not adjacent to any of the vertices we have
visited so far, and as it is adjacent to vi , the deepest back edge emanating from v j is (vi , v j ). Hence we output
this edge and move on to the next neighbor and eventually with the next step of DFS until all the vertices are
exhausted. This completes the description of the algorithm. See Figure 1 for an illustration. Now to see how this
procedure produces all the deepest back edges out of every vertex, note that, at vertex vi , our algorithm reports all
the back edges e = (vi , v j ) where e is the deepest back edge from v j , and also all tree edges (vi , v j ) where v j has no
back edge. Observe that from our convention, in the second case, (vi , v j ) is the deepest back edge out of v j . This
concludes the proof of the lemma. As we performed just one DFS to produce all such edges, using Theorem 2,
the claimed running time and space bounds follow.
The way we will actually use Lemma 1 in our algorithms, is for finding and storing the low values for at most
n/ lg n vertices. So we state a corollary for that.
Corollary 2. Given any undirected graph G with n vertices and m edges and any set L of O(n/ lg n) vertices as
input, we can compute, report and store the low(v) values for every vertex v in L in the DFS tree T of G using O(n)
bits of space and O(m lg lg n) time.
Note that, Lemma 1 holds true for any undirected connected graph G. In what follows, we use Lemma 1 to
give a space efficient implementation of Elmasry’s algorithm when the input graph G is an undirected biconnected
graph. In particular, we show the following,
Theorem 8. Given an undirected biconnected graph G with n vertices and m edges, we can output the edges of a
sparse spanning biconnected subgraph of G using O(n) bits of space and O(m lg lg n) time.
7
Proof. When the underlying graph G is undirected biconnected graph, we know that Elmasry’s algorithm produces
a sparse spanning subgraph which is also biconnected. In order to implement that, given an undirected biconnected
graph G, we first run on G the algorithm of Lemma 1 which produces and reports all the deepest back edges out
of all the vertices v in G. Out of all those deepest back edges, note that, some are actually tree edges from our
convention. Hence, we don’t want to report them multiple time. More specifically, if a vertex v j has no back edge
going out of it, Lemma 1 outputs the edge (vi , v j ) as the deepest back edge out of v j , which is actually a tree edge
in the DFS tree T of G. In order to avoid reporting such edges more than once, we perform the following. During
the scanning of vi ’s adjacency array, we also check if any of its neighbor, other than its parent, is gray. If so, we
report the edge from vi to its parent. Note that if vi has a back edge to one of its ancestors (other than its parent),
then this step reports the tree edge from vi to its parent. Otherwise, vi didn’t have any back edge, and hence the
tree edge to its parent would have been output while DFS was exploring and outputting deepest back edges from
its parent; so we do not output the edge again. Note that, we can do this test along with the algorithm of Lemma 1
so that using just one DFS, we can produce all the tree edges and deepest back edges as required in Elmasry’s
algorithm. Thus using Theorem 2, we can output the edges of a sparse spanning biconnected subgraph of G using
O(n) bits of space and O(m lg lg n) time.
4. Tree Cover and Space Efficient Construction
Before moving on to handle other complex applications of DFS in undirected graphs, namely biconnectivity,
2-edge connectivity, st-numbering etc, in the this section we discuss the common methodology to attack all of
these problems. Once we set all our machinary here, in Section 4, we see afterwards how to use them almost in a
similar fashion to several problems. Central to all of our algorithms following this section is a decomposition of
the DFS tree. For this we use the well-known tree covering technique which was first proposed by Geary et al. [47]
in the context of succinct representation of rooted ordered trees. The high level idea is to decompose the tree into
subtrees called minitrees, and further decompose the mini-trees into yet smaller subtrees called microtrees. The
microtrees are tiny enough to be stored in a compact table. The root of a minitree can be shared by several other
minitrees. To represent the tree, we only have to represent the connections and links between the subtrees. Later He
et al. [50] extended this approach to produce a representation which supports several additional operations. Farzan
and Munro [40] modified the tree covering algorithm of [47] so that each minitree has at most one node, other than
the root of the minitree, that is connected to the root of another minitree. This simplifies the representation of the
tree, and guarantees that in each minitree, there exists at most one non-root node which is connected to (the root
of) another minitree. The tree decomposition method of Farzan and Munro [40] is summarized in the following
theorem:
Theorem 9 ([40]). For any parameter L ≥ 1, a rooted ordered tree with n nodes can be decomposed into Θ(n/L)
minitrees of size at most 2L which are pairwise disjoint aside from the minitree roots. Furthermore, aside from
edges stemming from the minitree root, there is at most one edge leaving a node of a minitree to its child in another
minitree. The decomposition can be performed in linear time.
See Figure 2 for an illustration. In our algorithms, we apply Theorem 9 with L = n/ lg n. For this parameter
L, since the number of minitrees is only O(lg n), we can represent the structure of the minitrees within the original
tree (i.e., how the minitrees are connected with each other) using O(lg2 n) bits. The decomposition algorithm
of [40] ensures that each minitree has at most one ‘child’ minitree (other than the minitrees that share its root)
in this structure. We use this property crucially in our algorithms. We refer to this as the minitree-structure. See
Figure 3(a) for the minitree structure of the tree decomposition shown in Figure 2.
Explicitly storing all the minitrees (using pointers) requires ω(n) bits overall. One way to represent them
efficiently using O(n) bits is to store each of them using any linear-bit encoding of a tree [? ]. But if we store these
minitrees separately, we loose the ability to compute the preorder or postorder numbers of the nodes in the entire
tree, which is needed in our algorithms. Hence, we encode the entire tree structure using a linear-bit encoding, and
store pointers into this encoding to represent the minitrees. We first encode the tree using the balanced parenthesis
(BP) representation [15, 59], summarized in the following theorem.3
Theorem 10 ([15]). Given a rooted ordered tree T on n nodes, it can be represented as a sequence of balanced
parentheses of length 2n. Given the preorder or postorder number of a node v in T , we can support subtree size
and various navigational queries (such as parent and i-th child) on v in O(1) time using an additional o(n) bits.
3 The representation of [59] does not support computing the i-th child of a node in constant time while the one in [15] can. When using
these representations to produce a tree cover, the representation of [59] is sufficient as we just need to compute the ‘next child’ as we traverse
the tree in post-order computing the subtree sizes of each subtree.
8
Figure 2: An illustration of Tree Covering technique with L = 5. The figure is reproduced from [40]. Each closed region formed by the
dotted lines represents a minitree. Note that each minitree has at most one ‘child’ minitree (other than the minitrees that share its root) in this
structure.
The following lemma by Farzan et al. [41, Lemma 2] (restated) shows that each minitree is split into a constant
number of consecutive chunks in the BP sequence. So we now represent each minitree by storing pointers to the
set of all chunks in the BP representation that together constitute the minitree.
Lemma 2. In the BP sequence of a tree, the bits corresponding to a mini-tree form a set of constant number of
substrings. Furthermore, these substrings concatenated together in order, form the BP sequence of the mini-tree.
Hence, one can store a representation of the minitrees by storing an O(lg2 n)-bit structure that stores pointers
to the starting positions of the chunks corresponding to each minitree in the BP sequence We refer to the representation obtained using this tree covering (TC) approach as the TC representation of the tree. See Figure 3 for
a complete example of a minitree structure along with the BP sequence of the tree of Figure 2. The following
lemma shows that we can construct the TC representation of the DFS tree of a given graph, using O(n) additional
bits.
Lemma 3. Given a graph G on n vertices and m edges, if there is an algorithm that takes t(n, m) time and s(n, m)
bits to perform DFS on G, then one can create the TC representation of the DFS tree in t(n, m) + O(n) time, using
s(n, m) + O(n) bits.
Proof. We first construct the balanced parenthesis (BP) representation of the DFS tree as follows. We start with
an empty sequence, BP, and append parentheses to it as we perform each step of the DFS algorithm. In particular,
whenever the DFS visits a vertex v for the first time, we append an open parenthesis to BP. Similarly when DFS
backtracks from v, we append a closing parenthesis. At the end of the DFS algorithm, as every vertex is assigned
a pair of parenthesis, length of BP is 2n bits. We just need to run the DFS algorithm once to construct this array,
hence the running time of this algorithm is asymptotically the same as the running time of the DFS algorithm.
We construct auxiliary structures to support various navigational operations on the DFS tree using the BP
sequence, as mentioned in Theorem 10. This takes o(n) time and space using the algorithm of [46]. We then use
the BP sequence along with the auxiliary structures to navigate the DFS tree in postorder, and simulate the tree
decomposition algorithm of Farzan and Munro [40] for constructing the TC representation of the DFS tree. If we
reconstruct the entire tree (with pointers), then the intermediate space would be Ω(n lg n) bits. Instead, we observe
that the tree decomposition algorithm of [40] never needs to keep more than O(L) temporary components (see [40]
for details) in addition to some of the permanent components. Each component (permanent or temporary) can be
stored by storing the root of the component together with its subtree size. Since L = n/ lg n, and the number of
9
a
b
d
M8
M3
k
M1
M2
M4
p
n
M4
o
r
q
x
M5
M6
M7
y
z
β
M7
(a)
(((())()(())())()((()((()(())()()())((())()()))()))()())
(b)
a b
M1
M2 b M3 d M4 o q
M5
M6 q r
M7
r o M4 d M8
a
(c)
Figure 3: (a) The minitree structure of the tree decomposition shown in Figure 2. (b) This array encodes the entire DFS tree using the balanced
parenthesis (BP) representation. (c) In this array, we demonstrate how the minitrees are split into a constant number of consecutive chunks in
the BP representation. Note that the bottom array can actually be encoded using O(lg2 n) bits, by storing, for each of the O(lg n) minitrees,
pointers to all the chunks in BP sequence indicating the starting and ending positions of the chunks corresponding to the minitrees.
permanent components is only O(lg n), the space required to store all the permanent and temporary components
at any point of time is bounded by O(n) bits. The construction algorithm takes O(n) time.
We use the following lemma in the description of our algorithms in the later sections.
Lemma 4. Let G be a graph, and T be its DFS tree. If there is an algorithm that takes t(n, m) time and s(n, m)
bits to perform DFS on G, then, using s(n, m) + O(n) bits, one can reconstruct any minitree given by its ranges in
the BP sequence of the TC representation of T , along with the labels of the corresponding nodes in the graph in
O(t(n, m)) time.
Proof. We first perform DFS to construct the BP representation of the DFS tree, T . We then construct the TC
representation of T , as described in Lemma 3. We now perform DFS algorithm again, keeping track of the
preorder number of the current node at each step. Whenever we visit a new node, we check its preorder number
to see if it falls within the ranges of the minitree that we want to reconstruct. (Note that, as mentioned above,
from [41, Lemma 2], the set of all preorder number of the nodes that belong to any minitree form a constant
number of ranges, since these nodes belong to a constant number of chunks in the BP sequence.) If it is within
one of the ranges corresponding to the minitree being constructed, then we add the node along with its label to the
minitree.
5. Applications of DFS using tree-covering technique
In this section, we provide O(n) bit implementations of various algorithmic graph problems that use DFS, by
using the tree covering technique developed in the previous section. At a higher level, we use the tree covering
10
technique to generate the minitrees one by one, and then partially solve the corresponding graph problem inside
that minitree before finally combining the solution across all the minitrees. The problems we consider include
algorithms to test biconnectivity, 2-edge connectivity and to output cut vertices, edges, and to find a chain decomposition and an st-numbering among others. To test for biconnectivity and related problems, the classical
algorithm due to Tarjan [66, 67] computes the so-called “low-point” values (which are defined in terms of a DFStree) for every vertex v, and checks some conditions based on these values. Brandes [11] and Gabow [44] gave
considerably simpler algorithms for testing biconnectivity and computing biconnected components by using some
path-generating rules instead of low-points; they call these algorithms path-based. An algorithm due to Schmidt
[64] is based on chain decomposition of graphs to determine biconnectivity (and/or 2-edge connected). All these
algorithms take O(m + n) time and O(n) words of space. Roughly these approaches compute DFS and process the
DFS tree in specific order maintaining some auxiliary information of the nodes. We start with a brief description
of chain decomposition and its application first before providing its space efficient implementation.
5.1. Chain decomposition
Schmidt [63] introduced a decomposition of the input graph that partitions the edge set of the graph into cycles
and paths, called chains, and used this to design an algorithm to find cut vertices and biconnected components [64]
and also to test 3-connectivity [63] among others. In what follows, we discuss briefly the decomposition algorithm,
and state his main result.
The algorithm first performs a depth first search on G. Let r be the root of the DFS tree T of G. DFS assigns an
index to every vertex v, namely, the time vertex v is discovered for the first time during DFS – call it the depth-firstindex of v (DFI(v)). Imagine that the back edges are directed away from r and the tree edges are directed towards
r. The algorithm decomposes the graph into a set of paths and cycles called chains as follows. See Figure 4 for an
example. First we mark all the vertices as unvisited. Then we visit every vertex starting at r in the increasing order
of DFI, and do the following. For every back edge e that originates at v, we traverse a directed cycle or a path.
This begins with v and the back edge e and proceeds along the tree towards the root and stops at the first visited
vertex or the root. During this step, we mark every encountered vertex as visited. This forms the first chain. Then
we proceed with the next back edge at v, if any, or move towards the next vertex in the increasing DFI order and
continue the process. Let D be the collection of all such cycles and paths. Notice that the cardinality of this set is
exactly the same as the number of back edges in the DFS tree as each back edge contributes to a cycle or a path.
Also, as initially every vertex is unvisited, the first chain would be a cycle as it would end in the starting vertex.
Using this, Schmidt proved the following theorem.
Theorem 11 ([64]). Let D be a chain decomposition of a connected graph G(V, E). Then G is 2-edge-connected if
and only if the chains in D partition E. Also, G is 2-vertex-connected if and only if δ(G) ≥ 2 (where δ(G) denotes
the minimum degree of G) and D1 is the only cycle in the set D where D1 is the first chain in the decomposition.
An edge e in G is bridge if and only if e is not contained in any chain in D. A vertex v in G is a cut vertex if and
only if v is the first vertex of a cycle in D \ D1 .
Now we are ready to describe an implementation of Schmidt’s chain decomposition algorithm using only O(n)
bits of space and in O(m lg2 n lg lg n) time using our partition of the DFS tree of Section 4. In the following
description, processing a back edge refers to the step of outputting the chain (directed path or cycle) containing
that edge and marking all the encountered vertices as visited. Processing a node refers to processing all the back
edges out of that node. The main idea of our implementation is to process all the back edges out of each node
in their preorder (as in Schmidt’s algorithm). To perform this efficiently (within the space limit of O(n) bits), we
process the nodes in chunks of size n/ lg n each (i.e., the first chunk of n/ lg n nodes in preorder are processed,
followed by the next chunk of n/ lg n nodes, and so on). But when processing the back edges out of a chunk C,
we process all the back edges that go from C to all the minitrees in their postorder, processing all the edges from
C to a minitree τ1 before processing any other back edges going out of C to a different minitree. This requires us
to go through all the edges out of each chunk at most O(lg n) times (once for each minitree). Thus the order in
which we process the back edges is different from the order in which we process them in Schmidt’s algorithm, but
we argue that this does not affect the correctness of the algorithm. In particular, we observe the following:
• Schmidt’s algorithm correctly produces a chain decomposition even if we process vertices to any order, as
long as we process a vertex v only after all its ancestors are also processed – for example, in level order
instead of preorder. This also implies that as long as we process the back edges coming to a vertex v (from
any of its descendants) only after we process all the back edges going to any of it’s ancestors from any of
v’s descendants, we can produce a chain decomposition correctly.
11
a
V1
b
V2
c
V3
e
d
i
f
V5
V4
V9
V6
j
g
V10
V7
h
V8
(b)
(a)
V1
V1
V2
V5
V6
V2
V4
V5
V9
V7
V3
V3
V3
V10
V8
(c)
Figure 4: Illustration of Chain Decomposition. (a) An input graph G. (b) A DFS traversal of G and the resulting edge-orientation along with
DFIs. (c) A chain decomposition D of G. The chains D2 and D3 are paths and rest of them are cycles. The edge (V5 , V6 ) is bridge as it is not
contained in any chain. V5 and V6 are cut vertices.
To process a back edge (u, v) between a chunk C and a minitree τ, where u belongs to C, v belongs to τ, and
u is an anscestor of v, we first output the edge (u, v), and then traverse the path from v to the root of τ, outputting
all the traversed edges and marking the nodes as visited. We then start another DFS to produce the minitree τ p
containing the parent p of the root of τ, and output the path from p to the root of τ p , and continue the process
untill we reach a vertex that has been marked as visited. Note that this process will terminate since u is marked
and is an ancestor of v. We maintain a bitvector of length n to keep track of the marked vertices, to perform this
efficiently. A crucial observation that we use in bounding the runtime is that once we produce a minitree τ p for
a particular pair (C, τ), we don’t need to produce it again, as the root of τ will be marked after the first time we
output it as part of a chain. Also, once we generate a chunk C and a minitree τ, we go through all the vertices of
C in preorder, and process all the edges that go between C and τ. We provide the pseudocode (see Algorithm 1)
below describing the high-level algorithm for outputting the chain decomposition.
Algorithm 1 Chain Decomposition
Let τ1 , τ2 , · · · , τO(lg n) be the minitrees in postorder and C1 , C2 , · · · , Clg n be the chunks of vertices in preorder
1: for i = 1 to lg n do
2:
for j = 1 to O(lg n) do
3:
for all back edges (u, v) with u ∈ Ci and v ∈ τ j do
4:
output the chain containing the edge (u, v)
5:
end for
6:
end for
7: end for
The time taken for the initial part, where we construct the DFS tree, decompose it into minitrees, and construct
the auxiliary structures, is O(m lg lg n), using Theorem 2 with t(n) = lg lg n. The running time of the rest of
the algorithm is dominated by the cost of processing the back edges. As outlined in Algorithm 1, we process
the back edges between every pair (Ci , τ j ), where Ci is the i-th chunk of n/ lg n nodes in preorder, and τ j is the
j-th minitree in postorder, for 1 ≤ i ≤ lg n and 1 ≤ j ≤ O(lg n). The outer loop of the algorithm generates
each chunk in preorder, and thus requires a signle DFS to produce all the chunks over the entire execution of the
algorithm. The inner loop goes through all the minitrees for each chunk. Since there are lg n chunks and O(lg n)
minitrees, and prodicing each minitree takes O(m lg lg n) time, the generation of all the chunk-minitree pairs takes
12
O(m lg lg n lg2 n) time.
For a particular pair (C, τ), we may need to generate many (O(lg n), in the worst-case) minitrees. But we
observe that, this happens for at most one back edge for a every pair (C, τ), since after processing the first such
back edge, the root of the minitree τ is marked and hence any chain that is output afterwards will stop before the
root of the minitree. Also, if a minitree τ` is generated when processing a pair (C, τ), then it will not be generated
when processing any other pair (C 0 , τ0 ) which is different from (C, τ) (since each minitree has at most one child
minitree). Thus the overall running time is dominated by generating all the pairs C, τ), which is O(m lg2 n lg lg n).
Thus, we obtain the following.
Theorem 12. Given an undirected graph G on n vertices and m edges, we can output a chain decomposition of G
in O(m lg2 n lg lg n) time using O(n) bits.
5.2. Testing biconnectivity and finding cut vertices
A naı̈ve algorithm to test for biconnectivity of a graph G = (V, E) is to check if (V \ {v}, E) is connected,
for each v ∈ V. Using the O(n) bits and O(m + n) time BFS algorithm [5] for checking connectivity, this gives
a simple O(n) bits algorithm running in time O(mn). Another approach is to use Theorem 12 combining with
criteria mentioned in Theorem 11 to test for biconnectivity and output cut vertices in O(m lg2 n lg lg n) time using
O(n) bits.
Here we show that using O(n) bits we can design an even faster algorithm running in O(m lg n lg lg n) time.
If G is not biconnected, then we also show how one can find all the cut-vertices of G within the same time and
space bounds. We implement the classical low-point algorithm of Tarjan [66]. Recall that, the algorithm performs
a DFS and computes for every vertex v, a value lowpoint[v] which is recursively defined as
lowpoint[v] = min{ DFI(v) ∪ {lowpoint[s]| s is a child of v}
∪ {DFI(w)|(v, w) is a back-edge} }
Tarjan proved that if a vertex v is not the root, then v is a cut vertex if and only if v has a child w such that
lowpoint[w] ≥ DFI(v). The root of a DFS tree is a cut vertex if and only if the root has more than one child.
Since the lowpoint values require Ω(n lg n) bits in the worst case, this poses the challenge of efficiently testing the
condition for biconnectivity with O(n) bits. To deal with this, as in the case of the chain decomposition algorithm,
we compute lowpoint values in O(lg n) batches using our tree covering algorithm. Cut vertices encountered in the
process, if at all, are stored in a separate bitmap. We show that each batch can be processed in O(m lg lg n) time
using DFS, resulting in an overall runtime of O(m lg n lg lg n).
5.2.1. Computing lowpoint and reporting cut vertices
We first obtain a TC representation of the DFS tree using the decomposition algorithm of Theorem 9 with
L = n/ lg n. We then process each minitree, in the postorder of the minitrees in the minitree structure. To process a
minitree, we compute the lowpoint values of each of the nodes in the minitree (except possibly the root) in overall
O(m) time. During the processing of any minitree, if we determine that a vertex is a cut vertex, we store this
information by marking the corresponding node in a seperate bit vector. Each minitree can be reconstructed in
O(m lg lg n) time using Lemma 4. The lowpoint value of a node is a function of the lowpoints of all its children.
However the root of a minitree may have children in other minitress. Hence for the root of the minitree, we
store the partial lowpoint value (till that point) which will be used to update its value when all its subtrees have
computed their lowpoint values (possibly in other minitrees).
Computing the lowpoint values in each of the minitrees is done in a two step process. In the first step, we
compute and store the low values of each node (which is the DFI value of the deepest back edge emanating from
that node) belonging to the minitree using Corollary 2. Note that the low values form one component of the values
among which we find the minimum, in the definition of lowpoint above, with a slight change. I.e., if a vertex v
has a backedge, then low(v) is nothing but min{DFI(w) : (v, w) is a back edge}. However, if v does not have a
backedge, by our convention low(v) has the DFI value of its parent which needs to be discounted in computing
lowpoint value of v. This is easily done if we also remember the DFI value of the parent of every node in the
minitree (using O(n) bits).
Once these low(v) values are computed and stored for all the vertices v belonging to a minitree, they are passed
on to the next step for computing lowpoint(v) values. More specifically, in the second step, we do another DFS
starting at the root of this minitree and compute the lowpoint values for all the vertices v belonging to a minitree
exactly in the same way as it is done in the classical Tarjan’s [66] algorithm using the explicitly stored low(v)
values. We provide the code snippet which actually shows how to compute lowpoint values recursively for a
minitree in Algorithm 2. Thus we obtain the following,
13
Lemma 5. Computing and storing the lowpoint(v) values for all the nodes v in a minitree can be performed in
O(m lg lg n) time, using O(n) bits.
Algorithm 2 DFS(v)
1: if low(v) = DFI(parent(v)) then
2: lowpoint(v) = DFI(v)
3: else lowpoint(v) = Min{DFI(v), low(v)}
4: for all y ∈ ad j(v) do
5:
if y is white then
6:
DFI(y) ← DFI(v) + 1
7:
DFS(y)
8:
if lowpoint(y) < lowpoint(v) then
9:
lowpoint(v) = lowpoint(y)
10:
end if
11:
end if
12: end for
To compute the effect of the roots of the minitrees on the lowpoint computation, we store various Θ(lg n)
bit information with each of the Θ(lg n) minitree roots including their partial/full lowpoint values, the rank of
its first/last child in its subtree. After we process one minitree, we generate the next minitree, in postorder, and
process it in a similar fashion and continue until we exhaust all the minitrees. As we can store the cut vertices in
a bitvector B of size n marking B[i] = 1 if and only if the vertex vi is a cut vertex, reporting them at the end of the
execution of the algorithm is a routine task. Clearly we have taken O(n) bits of space and the total running time is
O(m lg lg n lg n) as we run the DFS algorithm O(lg n) times overall. Thus we have the following
Theorem 13. Given an undirected graph G with n vertices and m edges, in O(m lg n lg lg n) time and O(n) bits of
space we can determine whether G is 2-vertex connected. If not, in the same amount of time and space, we can
compute and report all the cut vertices of G.
5.3. Testing 2-edge connectivity and finding bridges
The classical algorithm of Tarjan [67] to check if G is 2-edge connected takes O(m + n) time using O(n) words.
Schmidt’s algorithm [64] which is based on chain decomposition can also be implemented in linear time but with
O(m) words. The purpose of this section is to improve the space bound to O(n) bits, albeit with slightly increased
running time. For this, we use the following folklore characterization: a tree edge (v, w), where v is the parent of
w, is a bridge if and only if lowpoint[w] > DFI(v). That is to say, a tree edge (v, w) is a bridge if and only if the
vertex w and any of its descedants in the DFS tree cannot reach to vertex v or any of its ancestors. Thus if the
edge (v, w) is removed, the graph G becomes disconnected. Note that, since storing the lowpoint values requires
Ω(n lg n) bits, we cannot store all of them at once to check the criteria mentioned in the characterization, and this
poses the challenge of efficiently testing the condition for 2-edge connectivity with only O(n) bits. To perform this
test in a space efficient manner, we extend ideas similar to the ones developed in the previous section.
Similar to the biconnectivity algorithm, here also we first construct a TC representation of the DFS tree using
the decomposition algorithm of Theorem 9 with L = n/ lg n. We then process each minitree, in the postorder of the
minitrees in the minitree structure. To process a minitree, we compute the lowpoint values of each of the nodes in
the minitree (except possibly the root) in overall O(m) time. While processing these minitrees, if we come across
any bridge, we store it in a separate bitvector so that at the end of the execution of the algorithm we can report all
of them. Using Lemma 4, we know that each minitree can be reconstructed in O(m lg lg n) time, and also we store
for the root the partially computed lowpoint (till the point we are done processing minitrees). Now we compute
the lowpoint values for each of the vertices belonging to a minitree using Lemma 5.
Once we determine lowpoint values for all the vertices belonging to a minitree, we generate each minitree
along with the node labels, and easily test whether any tree edge is a bridge using the characterization mentioned
above. We also need to check this condition for edges that connect two minitrees; but this can also be done within
the same time and space bounds. We store this information using a bit vector B of length n − 1 such that B[i] = 1
if and only if the i-th edge in pre-order, of the DFS tree, is a bridge. Thus, by running another DFS, we can report
all the bridges of G. Clearly this procedure takes O(n) bits of space and the total running time is O(m lg lg n lg n)
as we run the DFS algorithm O(lg n) times overall. Hence we obtain the following.
14
Theorem 14. Given an undirected graph G with n vertices and m edges, in O(m lg n lg lg n) time and O(n) bits of
space we can determine whether G is 2-edge connected. If G is not 2-edge connected, then in the same amount of
time and space, we can compute and output all the bridges of G.
5.4. st-numbering
The st-ordering of vertices of an undirected graph is a fundamental tool for many graph algorithms, e.g., in
planarity testing and graph drawing. The first linear-time algorithm for st-ordering the vertices of a biconnected
graph is due to Even and Tarjan [38], and is further simplified by Ebert [31], Tarjan [68] and Brandes [11]. All
these algorithms, however, preprocess the graph using depth-first search, essentially to compute lowpoints which
in turn determine an (implicit) open ear decomposition. A second traversal is required to compute the actual stordering. All of these algorithms take O(n lg n) bits of space. We give an O(n) bits implementation of Tarjan’s [68]
algorithm.
We first describe the two pass classical algorithm of Tarjan without worrying about the space requirement. The
algorithm assumes, without loss of generality, that there exists an edge between the vertices s and t, otherwise it
adds the edge (s, t) before starting with the algorithm. Moreover, the algorithm starts a DFS from the vertex s and
the edge (s, t) is the first edge traversed in the DFS of G. Let p(v) be the parent of vertex v in the DFS tree. DFI(v)
and lowpoint(v) have the usual meaning as defined previously. The first pass is a depth first search during which
for every vertex v, p(v), DFI(v) and lowpoint(v) are computed and stored. The second pass constructs a list L,
which is initialized with [s, t], such that if the vertices are numbered in the order in which they occur in L, then we
obtain an st-ordering. In addition, we also have a sign array of n bits, intialized with sign[s]=-. The second pass is
a preorder traversal starting from the root s of the DFS tree and works as described in the following pseudocode
(Algorithm 3) below.
Algorithm 3 st-numbering
1: DFS(s) starts with the edge (s, t)
2: for all vertices v , s, t in preorder of DFS(s) do
3:
if sign(lowpoint(v)) == + then
4:
Insert v after p(v) in L
5:
sign(p(v)) = −
6:
end if
7:
if sign(lowpoint(v))==- then
8:
Insert v before p(v) in L
9:
sign(p(v))=+
10:
end if
11: end for
It is clear from the above pseudocode that the procedure runs in linear time using O(n lg n) bits of space for
storing elements in L. To make it space effcient, we use ideas similar to our biconnectivity algorithm. At a high
level, we generate the lowpoint values of the first n/ lg n vertices in depth first order and process them. Due to
space restriction, we cannot store the list L as in Tarjan’s algorithm; instead we use the BP sequence of the DFS
tree and augment it with some extra information to ‘encode’ the final st-ordering, as described below.
As in some of our earlier algorithms, this algorithm also runs in O(lg n) phases and in each phase it processes
n/ lg n vertices. In the first phase, to obtain the lowpoint values of the first n/ lg n vertices in depth first order, we
run as in our biconnectivity algorithm a procedure to store explicitly for these vertices their lowpoint values in an
array. Also during the execution of the biconnectivity algorithm, the BP sequence is generated and stored in the
BP array. We create two more arrays, of size n bits, that have one to one correspondence with the open parentheses
of the BP sequence. We can use rank/select operations (as defined Section 2.3) to map the position of a vertex
in these two arrays to the corresponding open parenthesis in the BP sequence. The first array, called Sign, is for
storing the sign for every vertex as in Tarjan’s algorithm. To simulate the effect of the list L, we create the second
array, called P, where we store the relative position, i.e., “before” or “after”, of every vertex with respect to its
parent. Namely, if u is the parent of v, and v comes before (after, respectively) u in the list L in Algorithm 3, then
we store P[v] = b (P[v] = a, respectively). One crucial observation is that, even though the list L is dynamic, the
relative position of the vertex v does not change with respect to the position of u, and is determined at the time of
insertion of v into the list L (new vertices may be added between u and v later). In what follows, we show how to
use the BP sequence, and the array P to emulate the effect of list L and produce the st-ordering.
We first describe how to reconstruct the list L using the BP sequence and the P array. The main observation we
use in the reconstruction of L is that a node v appears in L after all the nodes in its child subtrees whose roots are
15
marked with b in P, and also before all the nodes in its child subtrees whose roots are marked with a in P. Also, all
the nodes in a subtree appear “together” (consecutively) in the list L. Moreover, all the children marked b appear
in the increasing order of the DFI while all the children marked a appear in the decreasing order of the DFI. Thus
by looking at the P[v] values of all the children of a node u, and computing their subtree sizes, we can determine
the position in L of u among all the nodes in its subtree. Let us call a child v of u as after-child if v is marked a in
P. Similarly, if v is marked b in P, it is called before-child. Let T (v) denote the subtree rooted at the vertex v in
the DFS tree T of G and |T (v)| denotes the size of T (v). Let us also suppose that the vertex u has k + ` children,
out of which k children v1 , · · · , vk are before-children and the remaining ` children w1 , · · · , w` are after-children,
where DFI(v1 ) < DFI(v2 ) < · · · < DFI(vk ) and DFI(w1 ) < DFI(w2 ) < · · · < DFI(w` ). Then in L, all the vertices
from T (v1 ), T (v2 ), followed by till T (vk ) appear, followed by u and finally the vertices from T (w` ), T (w`−1 ) till
P
T (w1 ) appear. More specifically, u appears at the (S + 1)-th location where S = ki=1 |T (vi )|. With this approach,
we can reconstruct the list L, and hence output the st-numbers of all the nodes in linear time, if L can be stored
in memory - which requires O(n lg n) bits. Now to perform this step with O(n) bits, we repeat the whole process
of reconstruction lg n times, where in the i-th iteration, we reproduce sublist L[(i − 1)n/ lg n + 1, . . . , i.n/ lg n]
– by ignoring any node that falls outside this range – and reporting all these nodes with st-numbers in the range
[(i−1)n/ lg n+1, . . . , i.n/ lg n]. As each of these reconstruction takes O(m lg n lg lg n) time, we obtain the following,
Theorem 15. Given an undirected biconnected graph G on n vertices and m edges, and two distinct vertices s
and t, we can output an st-numbering of all the vertices of G in O(m lg2 n lg lg n) time, using O(n) bits of space.
5.5. Applications of st-numbering
In this section, we show that using the space efficient implementation of Theorem 15 for st-numbering, we
immediately obtain similar results for a few applications of st-numbering. We provide the details below.
5.5.1. Two-partitioning problem
In this problem, given vertices a1 , · · · , ak of a graph G and natural numbers c1 , · · · , ck with c1 + · · · + ck = n,
we want to find a partition of V into sets V1 , · · · , Vk with ai ∈ Vi and |Vi | = ci for every i such that every set Vi
induces a connected graph in G. This problem is called the k-partitioning problem. The problem is NP-hard even
when k = 2, G is bipartite and the condition ai ∈ Vi is relaxed [30]. But, Györi [48] and Lovasz [55] proved
that such a partition always exists if the input graph is k-connected and can be found in polynomial time in such
graphs. Let G be 2-connected. Then two-partitioning problem can be solved in the following manner [65]: Let
v1 := a1 and vn := a2 , compute an v1 vn -numbering v1 , v2 , · · · , vn and note that, from the property of st-numbering,
for any vertex vi (in particular for i = c1 ) the graphs induced by v1 , · · · , vi and by vi , · · · , vn are always connected
subgraph of G. Thus applying Theorem 15, we obtain the following:
Theorem 16. Given an undirected biconnected graph G, two distinct vertices a1 , a2 , and two natural numbers
c1 , c2 such that c1 + c2 = n, we can obtain a partition (V1 , V2 ) of the vertex set V of G in O(m lg2 n lg lg n) time,
using O(n) bits of space, such that a1 ∈ V1 and a2 ∈ V2 , |V1 | = c1 , |V2 | = c2 , and both V1 and V2 induce connected
subgraph on G.
5.5.2. Vertex-subset-two-partition problem
Wada and Kawaguchi [70] defined the following problem which they call the vertex-subset-k-partition problem. This is actually an extension of the k-partition problem defined in Section 5.5.1. The problem is defined as
follows:
Input:
1.
2.
3.
4.
An undirected graph G = (V, E) with n vertices and m edges;
a vertex subset V 0 (⊆ V) with n0 = |V 0 | ≥ k;
k distinct vertices ai (1 ≤ i ≤ k) ∈ V 0 , ai , a j (1 ≤ i < j ≤ k); and
P
k natural numbers n1 , n2 , · · · , nk such that ki=1 ni = n0 .
Output: a partition V1 ∪ V2 ∪ · · · ∪ Vk of the vertex set V and a partition V10 ∪ V20 ∪ · · · ∪ Vk0 of vertex set V 0 such
that for each i(1 ≤ i ≤ k)
1.
2.
3.
4.
ai ∈ Vi0 ;
|Vi0 | = ni ;
Vi0 ⊆ Vi and
each Vi induces a connected subgraph of G.
16
Note that this problem is an extension of the k-partition problem, since choosing V 0 = V corresponds to the
original k-partition problem. Wada and Kawaguchi [70] proved that vertex-subset-k-partition problem always
admits a solution if the input graph G is k-connected (for k ≥ 2). In particular, if G is 2-connected, using stordering, the vertex-subset-two-partitioning problem can be solved in the following manner [70]: suppose that
G, V 0 (⊆ V), a1 , a2 , n1 and n2 (n1 + n2 = n0 = |V 0 |) are the inputs. Let s = v1 := a1 and t = vn := a2 , compute
an st-numbering v1 , v2 , · · · , vn . From this st-numbering, note that, V now can be partitioned in two sets V1 and V2
such that |V1 ∩ V 0 | = n1 and |V2 ∩ V 0 | = n2 . From the property of st-numbering, we know that both V1 and V2
induce a connected subgraph of G. Moreover, a1 ∈ V1 and a2 ∈ V2 . Using Theorem 15 as a subroutine to compute
such an st-numbering of G, we obtain the following result.
Theorem 17. Given an undirected biconnected graph G, we can solve the vertex-subset-two-partitioning problem
in O(m lg2 n lg lg n) time, using O(n) bits of space.
5.5.3. Two independent spanning trees
Recall that k spanning trees of a graph G are independent if they all have the same root vertex r, and for every
vertex v , r, the paths from v to r in the k spanning trees are vertex-disjoint (except for their endpoints). Itai
and Rodeh [52] conjectured that every k-connected graph contains k independent spanning trees. Even though
the most general version of this conjecture has not been proved yet, this conjecture is shown to be true for k ≤ 4
[22, 26, 52, 71], and also for planar graphs [51]. In particular, if the given graph G is biconnected, we can generate
two independent spanning trees (let us call them S and T ) in the following manner [52].
Choose an arbitrary edge, say (s, t) in G. Let f be an st-numbering of G. To construct S , choose for every
vertex v , s an edge (u, v) such that f (u) < f (v), and for t choose an edge other than (s, t). To construct T , choose
the edge (s, t) and for every vertex v < {s, t} an edge (v, w), f (v) < f (w) . It is easy to prove that s is the root of
both S and T , and that S and T are independent spanning trees as, for every vertex v, the path from the root s to
v in S consists of vertices u with f (u) < f (v) but except the edge (s, t), whereas in T , along with the edge (s, t), it
consists of vertices w with f (v) < f (w). Using Theorem 15 to compute such an st-numbering of G, it is not hard
to produce S and T . Thus we obtain the following,
Theorem 18. Given an undirected biconnected graph G, we can report two independent spanning trees S and T
in O(m lg2 n lg lg n) time, using O(n) bits.
6. Concluding remarks and open problems
We have presented space efficient algorithms for a number of important applications of DFS. Obtaining linear
time algorithms for them while maintaining O(n) bits of space usage is both interesting and challenging open
problem. One of the main bottlenecks (with this approach) towards this is finding an O(n)-bit, O(m + n)-time
algorithm for DFS, which is also open, even though for BFS we know such implementations [5, 37]. Another
challenging open problem is to remove the poly-log terms in the running times of the algorithms described (e.g.,
the lg n term in the running time of 2-vertex and 2-edge connectivity algorithm, and the lg2 n term in the running
time of two independent spanning trees algorithm). These terms seem inherent in our tree covering approach. It
would be interesting to find other applications of our tree covering approach for space efficient algorithms. There
are also plenty of other applications of DFS, it would be interesting to study them from the point of view of space
efficiency. For example, planarity is one prime example where DFS has been used very crucially. So it is a natural
question that, can we test planarity of a given graph using O(n) bits?
References
References
[1] S. Arora and B. Barak. Computational Complexity - A Modern Approach. Cambridge University Press, 2009.
[2] T. Asano, K. Buchin, M. Buchin, M.Korman, W. Mulzer, G. Rote, and A. Schulz. Reprint of: Memory-constrained algorithms for simple
polygons. Comput. Geom., 47(3):469–479, 2014.
[3] T. Asano, T. Izumi, M. Kiyomi, M. Konagaya, H. Ono, Y. Otachi, P. Schweitzer, J. Tarui, and R. Uehara. Depth-first search using O(n)
bits. In 25th ISAAC, pages 553–564, 2014.
√
[4] T. Asano, D. G. Kirkpatrick, K. Nakagawa, and O. Watanabe. Õ( n)-space and polynomial-time algorithm for planar directed graph
reachability. In 39th MFCS LNCS 8634, pages 45–56, 2014.
[5] N. Banerjee, S. Chakraborty, and V. Raman. Improved space efficient algorithms for BFS, DFS and applications. In 22nd COCOON,
volume 9797, pages 119–130. Springer, LNCS, 2016.
[6] N. Banerjee, S. Chakraborty, V. Raman, S. Roy, and S. Saurabh. Time-space tradeoffs for dynamic programming in trees and bounded
treewidth graphs. In 21st COCOON, volume 9198, pages 349–360. springer, LNCS, 2015.
17
[7] L. Barba, M. Korman, S. Langerman, K. Sadakane, and R. I. Silveira. Space-time trade-offs for stack-based algorithms. Algorithmica,
72(4):1097–1129, 2015.
[8] L. Barba, M. Korman, S. Langerman, and R. I. Silveira. Computing a visibility polygon using few variables. Comput. Geom., 47(9):918–
926, 2014.
[9] G. Barnes, J. Buss, W. Ruzzo, and B. Schieber. A sublinear space, polynomial time algorithm for directed s-t connectivity. SIAM J.
Comput., 27(5):1273–1282, 1998.
[10] P. Beame. A General Sequential Time-Space Tradeoff for Finding Unique Elements. SIAM J. Comput., 20(2):270-277, 1991.
[11] U. Brandes. Eager st-ordering. In 10th ESA, pages 247–256, 2002.
[12] H. Brönnimann and T. M. Chan. Space-efficient algorithms for computing the convex hull of a simple polygonal line in linear time.
Comput. Geom., 34(2):75–82, 2006.
[13] H. Buhrman, R. Cleve, M. Koucký, B. Loff, and F. Speelman. Computing with a full memory: catalytic space. In Symposium on Theory
of Computing, STOC 2014, New York, NY, USA, May 31 - June 03, 2014, pages 857–866, 2014.
[14] D. Chakraborty, A. Pavan, R. Tewari, N. V. Vinodchandran, and L. Yang. New time-space upperbounds for directed reachability in
high-genus and h-minor-free graphs. In FSTTCS, pages 585–595, 2014.
[15] H. Lu and C. Yeh. Balanced parentheses strike back. ACM Trans. Algorithms, 4(3):28:1–28:13, July 2008.
[16] S. Chakraborty, S. Jo, and S. R. Satti. Improved space-efficient linear time algorithms for some classical graph problems. In 15th CTW,
2017.
[17] S. Chakraborty, V. Raman, and S. R. Satti. Biconnectivity, chain decomposition and st-numbering using O(n) bits. In 27th ISAAC,
volume 64 of LIPIcs, pages 22:1–22:13. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2016.
[18] S. Chakraborty and S. R. Satti. Space-efficient algorithms for maximum cardinality search, stack BFS, queue BFS and applications. In
23rd COCOON, 2017.
[19] S. Chakraborty, V. Raman, and S. R. Satti. Biconnectivity, chain decomposition and st-numbering using O(n) bits. In 27th ISAAC,
volume 64 of LIPIcs, pages 22:1–22:13. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2016.
[20] T. M. Chan, J. I. Munro, and V. Raman. Selection and sorting in the ”restore” model. In 25th-SODA, pages 995–1004, 2014.
[21] J. Cheriyan, M. Kao, and R. Thurimella. Scan-first search and sparse certificates: An improved parallel algorithms for k-vertex connectivity. SIAM J. Comput., 22(1):157–174, 1993.
[22] J. Cheriyan and S. N. Maheshwari. Finding nonseparating induced cycles and independent spanning trees in 3-connected graphs. J.
Algorithms, 9(4):507–537, 1988.
[23] D. R. Clark. Compact Pat Trees. PhD thesis. University of Waterloo, Canada, 1996.
[24] S. A. Cook and C. Rackoff. Space lower bounds for maze threadability on restricted machines. SIAM J. Comput., 9(3):636–652, 1980.
[25] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein. Introduction to Algorithms (3. ed.). MIT Press, 2009.
[26] S. Curran, O. Lee, and X. Yu. Finding four independent trees. SIAM J. Comput., 35(5):1023–1058, 2006.
[27] O. Darwish and A. Elmasry. Optimal time-space tradeoff for the 2d convex-hull problem. In 22th ESA, pages 284–295, 2014.
[28] S. Dasgupta, C. H. Papadimitriou, and U. V. Vazirani. Algorithms. McGraw-Hill, 2008.
[29] S. Datta, N. Limaye, P. Nimbhorkar, T. Thierauf, and F. Wagner. Planar graph isomorphism is in log-space. In 24th CCC, pages 203–214,
2009.
[30] M.E. Dyer and A.M. Frieze. On the complexity of partitioning graphs into connected subgraphs. Discrete Applied Mathematics,
10(2):139–153, 1985.
[31] J. Ebert. st-ordering the vertices of biconnected graphs. Computing, 30(1):19–33, 1983.
[32] J. Edmonds, C. K. Poon, and D. Achlioptas. Tight lower bounds for st-connectivity on the NNJAG model. SIAM J. Comput., 28(6):2257–
2284, 1999.
[33] M. Elberfeld, A. Jakoby, and T. Tantau. Logspace versions of the theorems of bodlaender and courcelle. In 51th FOCS, pages 143–152,
2010.
[34] M. Elberfeld and K. Kawarabayashi. Embedding and canonizing graphs of bounded genus in logspace. In Symposium on Theory of
Computing, STOC 2014, New York, NY, USA, May 31 - June 03, 2014, pages 383–392, 2014.
[35] M. Elberfeld and P. Schweitzer. Canonizing graphs of bounded tree width in logspace. In 33rd Symposium on Theoretical Aspects of
Computer Science, STACS 2016, February 17-20, 2016, Orléans, France, pages 32:1–32:14, 2016.
[36] A. Elmasry. Why depth-first search efficiently identifies two and three-connected graphs. In 21st ISAAC, pages 375–386, 2010.
[37] A. Elmasry, T. Hagerup, and F. Kammer. Space-efficient basic graph algorithms. In 32nd STACS, pages 288–301, 2015.
[38] S. Even and R. E. Tarjan. Computing an st-numbering. Theo. Comp. Sci., 2(3):339–344, 1976.
[39] S. Even and R. E. Tarjan. Corrigendum: Computing an st-numbering. TCS 2(1976):339-344. Theor. Comput. Sci., 4(1):123, 1977.
[40] A. Farzan and J. Ian Munro. Succinct representation of dynamic trees. Theor. Comput. Sci., 412(24):2668–2678, 2011.
[41] A. Farzan, R. Raman, and S. S. Rao. Universal succinct representations of trees? In 36th ICALP, pages 451–462, 2009.
[42] J. Feigenbaum, S. Kannan, A. McGregor, S. Suri, and J. Zhang. On graph problems in a semi-streaming model. Theor. Comput. Sci.,
348(2-3):207–216, 2005.
[43] G. N. Frederickson. Upper bounds for time-space trade-offs in sorting and selection. J. Comput. Syst. Sci., 34(1):19–26, 1987.
[44] H. N. Gabow. Path-based depth-first search for strong and biconnected components. Inf. Process. Lett., 74(3-4):107–114, 2000.
[45] M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to the Theory of NP-Completeness. W. H. Freeman, 1979.
[46] R. F. Geary, N. Rahman, R. Raman, and V. Raman. A simple optimal representation for balanced parentheses. Theor. Comput. Sci.,
368(3):231–246, 2006.
[47] R. F. Geary, R. Raman, and V. Raman. Succinct ordinal trees with level-ancestor queries. ACM Trans. Algorithms, 2(4):510–534, 2006.
[48] E. Győri. Partition conditions and vertex-connectivity of graphs. Combinatorica, 1(3):263–273, 1981.
[49] T. Hagerup and F. Kammer. Succinct choice dictionaries. CoRR, abs/1604.06058, 2016.
[50] M. He, J. I. Munro, and S. R. Satti. Succinct ordinal trees based on tree covering. ACM Trans. Algorithms, 8(4):42:1–42:32, 2012.
[51] A. Huck. Independent trees in planar graphs. Graphs and Combinatorics, 15(1):29–77, 1999.
[52] A. Itai and M. Rodeh. The multi-tree approach to reliability in distributed networks. Inf. Comput., 79(1):43–59, 1988.
[53] F. Kammer, D. Kratsch, and M. Laudahn. Space-efficient biconnected components and recognition of outerplanar graphs. In 41st MFCS,
2016.
[54] M. Koucký. Catalytic computation. Bulletin of the EATCS, 118, 2016.
[55] L. Lovasz. A homology theory for spanning tress of a graph. Acta Mathematica Hungarica, 30(3-4):241–251, 1977.
[56] J. I. Munro. Tables. In FSTTCS, pages 37–42, 1996.
[57] J. I. Munro and M. Paterson. Selection and sorting with limited storage. Theor. Comput. Sci., 12:315–323, 1980.
18
[58] J. I. Munro and V. Raman. Selection from Read-Only Memory and Sorting with Minimum Data Movement. Theor. Comput. Sci.,
165(2):311–323, 1996.
[59] J. I. Munro and V. Raman. Succinct representation of balanced parentheses and static trees. SIAM J. Comput., 31(3):762–776, 2001.
[60] J. I. Munro, V. Raman, and S. S. Rao. Space efficient suffix trees. J. Algorithms, 39(2):205–222, 2001.
[61] H. Nagamochi and T. Ibaraki. A linear-time algorithm for finding a sparse k-connected spanning subgraph of a k-connected graph.
Algorithmica, 7(5&6):583–596, 1992.
[62] O. Reingold. Undirected connectivity in log-space. J. ACM, 55(4), 2008.
[63] J. M. Schmidt. Structure and constructions of 3-connected graphs. PhD thesis, Free University of Berlin, 2010.
[64] J. M. Schmidt. A simple test on 2-vertex- and 2-edge-connectivity. Inf. Process. Lett., 113(7):241–244, 2013.
[65] J. M. Schmidt. The mondshein sequence. In 41st ICALP, pages 967–978, 2014.
[66] R. E. Tarjan. Depth-first search and linear graph algorithms. SIAM J. Comput., 1(2):146–160, 1972.
[67] R. E. Tarjan. A note on finding the bridges of a graph. Inf. Process. Lett., 2(6):160–161, 1974.
[68] R. E. Tarjan. Two streamlined depth-first search algorithms. Fund. Inf., 9:85–94, 1986.
[69] M. Tompa. Two familiar transitive closure algorithms which admit no polynomial time, sublinear space implementations. SIAM J.
Comput., 11(1):130–137, 1982.
[70] K. Wada and K. Kawaguchi. Efficient algorithms for tripartitioning triconnected graphs and 3-edge-connected graphs. In 19th WG,
pages 132–143, 1993.
[71] A. Zehavi and A. Itai. Three tree-paths. Journal of Graph Theory, 13(2):175–188, 1989.
19
| 8cs.DS
|
Let’s Dance: Learning From Online Dance Videos
Daniel Castro
Georgia Institute of Technology
Steven Hickson
Patsorn Sangkloy
[email protected]
patsorn [email protected]
arXiv:1801.07388v1 [cs.CV] 23 Jan 2018
[email protected]
Bhavishya Mittal
Sean Dai
James Hays
Irfan Essa
[email protected]
[email protected]
[email protected]
[email protected]
Abstract
In recent years, deep neural network approaches have
naturally extended to the video domain, in their simplest
case by aggregating per-frame classifications as a baseline
for action recognition. A majority of the work in this area
extends from the imaging domain, leading to visual-feature
heavy approaches on temporal data. To address this issue we introduce “Let’s Dance”, a 1000 video dataset (and
growing) comprised of 10 visually overlapping dance categories that require motion for their classification. We stress
the important of human motion as a key distinguisher in
our work given that, as we show in this work, visual information is not sufficient to classify motion-heavy categories.
We compare our datasets’ performance using imaging techniques with UCF-101 and demonstrate this inherent difficulty. We present a comparison of numerous state-of-theart techniques on our dataset using three different representations (video, optical flow and multi-person pose data)
in order to analyze these approaches. We discuss the motion
parameterization of each of them and their value in learning to categorize online dance videos. Lastly, we release
this dataset (and its three representations) for the research
community to use.
Figure 1. Each row contains frames from the class it represents.
This figure is best viewed digitally.
be applied to extended/long sequences and datasets lacking in such highly dynamic videos. Our goal is to determine which methods best represent motion as opposed to
methods that use a single (properly picked) frame [6] to
identify the activity, as we feel such approaches devalue the
necessity for video data. In this work we introduce a 1,000
video dataset and evaluate methods that focuses on highly
dynamic videos requiring motion analysis for classification.
We choose the domain of dance videos as (a) there is
large amount of dance videos available online and (b) the diversity of dynamics in these videos provides us with a challenging space of problems for highly dynamic video analysis. This enables us to conduct a focused study on the relevance of motion in dancing classification and the broader
value of motion in improving video classification.
The core challenge of this task is attaining an adequate
representation of human motion across a 10-second clip. In
order to highlight the trajectory of this work, we will evaluate the current approaches and demonstrate the value of
isolating motion for properly evaluating these approaches
and this dataset.
1. Introduction
Video is a rich medium with dynamic information that
can be used to determine, what is happening in a scene. In
this work, we consider highly dynamic video, video that
requires the parametrization of motion over extended sequences to identify the activity being performed. The main
challenge with highly dynamic video is that a single frame
cannot provide sufficient information to understand the action being performed. Therefore, multiple frames, leading
to an extended sequence of frames, need to be analyzed for
scene classification. One of the drawbacks of current action
classification research is both a lack of approaches that can
1
Many video classification techniques exist, either utilizing single frames, late fusion architectures, temporal (3D)
convolutional networks, or recurrent networks with long
short-term memory (LSTM). Current classification problems can often be identified by a single frame. We present
a more challenging problem wherein each class requires the
use of multiple frames to adequately classify each category.
Specifically, we propose the use of optical flow and
multi-agent pose estimation as motion representations
which augment traditional video classification approaches.
Comparing these approaches enables us to gain insights into
the inherent encoding of motion in neural networks that is
difficult to understand.
Our main contributions are: (1) An analysis of baseline
and state-of-the-art approaches in video classification, (2)
a general method for concurrently learning from multiple
motion parameterizations in video, and (3) A 1000 video
dataset of highly dynamic dance videos, contrasted with existing video datasets, to motivate further investigation and
understanding of motion parameterization in video classification.
rating temporal data with each video frame by extending
the convolution kernels from size m x m x 3 to m x m
x 3 x T , where T represents a temporal extent[11]. They
also point out one of the major challenges of using deep
learning for video classification – there are no large-scale
video datasets comparable to the quality and size of image
recognition datasets. Similarly, 3D convolutional kernels
that incorporate the spatial domain have been shown to be
successful for action classification in both security camera
and depth data recordings [9, 27]. Wang et al. use a similar two-stream late fusion approach [28], but they note that
without incorporating the learned features into an ensemble method with handcrafted features, these deep-learned
approaches still fail to outperform handcrafted approaches.
We combine these methods in our work by incorporating
preprocessed features (optical flow and multi-agent pose detection) with 3D convolutional kernels in order to integrate
the representation of motion into the network architecture.
Another common approach is to leverage the sequential
nature of a Long Short-Term Memory (LSTM) network–
a specific type of recurrent neural network with additional
gates to control the flow of information. LSTMs can process information over long term temporal sequences and
have been applied in video for various tasks such as caption generation [26] and learning video representations [23].
Similarly, the long-term recurrent convolutional networks
(LRCNs) proposed by Donahue et. al. introduce another
variation of an LSTM for this task. Despite their temporal
nature, these approaches have been less successful at encoding motion in comparison to two-stream networks [4] which
encode the spatial and temporal domain in concurrent architectures.
The most effective method for classifying motion in
video is still unclear. In the context of action recognition,
many of these approaches are learning features based on the
image’s context and not the inherent action. This is in part
because commonly used video datasets such as UCF-Sports
and more traditionally UCF-101 can generally be identified to moderately decent accuracy using single-frame approaches which do not encode motion parameters[11].
A specific method for encoding motion that has recently
gained traction in action recognition is the use of pose
detection over the temporal domain with neural networks
[3][14][24][30]. Detecting pose over this domain provides
us with the intrinsic motion of the subjects in the scene. As
highlighted earlier, an initial breakthrough was achieved by
Toshev et. al. [24] with state-of-the-art results in estimating
the pose of a single individual from an image. The importance of pose was further demonstrated in [3], incorporating pose features from a CNN into action recognition. This
work was extended over the next two years to attain jointspecific networks that work well with partial and occluded
poses [30]. It was then most recently implemented to detect
2. Related Work
In order to determine which competing state-of-the-art
approaches to examine, we first present a literature review
on video classification. While deep networks have been
shown to be very effective at classifying, localizing, and
segmenting images, it is still unclear how to properly extend
these methods to the video domain. There are a few common approaches, some of which are: (1) Applying proven
image classification deep network architectures to individual frames of a video; (2) Extending 2D convolutional operators to 3D convolutions acting on the time domain; and
(3) Preprocessing the video into images that encode motion,
such as optical flow, and running current image architectures on the processed frames.
A simple way to extend image-based neural networks to
video classification is to extract features from each individual frame of a video [31]. While this technique does
lead to some success if the network learns temporallyinvariant features, it is commonly only used as a baseline
approach to compare against networks that incorporate temporal data [9, 11]. One common variant is a two-stream late
fusion architecture with a still frame-based “spatial” network stream running in parallel alongside a “temporal” network performing classifications based on optical flow calculations [3, 6, 8, 20]. This network architecture significantly outperforms approaches based solely on individual
frame classification, suggesting that incorporating a temporal component is necessary. In our work we leverage the
benefit of a temporal network by incorporating it into the
design of our network architecture.
Karpathy et al. explore more direct methods of incorpo2
2.1.3
Atomic Visual Actions (AVA)
The AVA dataset [15] contains 80 atomic visual actions in
57,400 movie clips which are localized within the frame.
This work goes beyond simply understanding a simple action in a video clip to understanding the interaction, both between humans and with humans and objects. Although this
is somewhat less relevant to our work, it demonstrates the
need for understanding motion features in human interaction – specifically by localizing the action and its relevance
in a scene that may contain multiple subjects / objects.
Figure 2. Each of these examples represents a different class in our
dataset (they are types of ballroom dancing). Top Left: Waltz, Top
Right: Quickstep, Bottom Left: Foxtrot, Bottom Right: Tango.
3. Let’s Dance Dataset
Our main challenge in this work was determining a reliable way of testing how well a specific method can parameterize motion. We realized that available video datasets such
as UCF-101 [22] and UCF-Sports [18] tackled a known
classification problem, one that could be evaluated using
extensions of available image classification architectures.
With that in mind, we developed a new dataset that prioritizes motion as the key characteristic of the classification.
We assembled a 1,000 video dataset containing 10 dynamic
and visually overlapping dances. We chose the parent category of dancing because it has a variety of measurable features (i.e. rhythm, limb movement), and it is not represented
in the Sports-1M and UCF-101 datasets [11, 22]. The categories included for this dataset are:
multiple people within a single frame [2]. In our work we
will leverage our own implementation of multi-agent pose
detection to demonstrate the need for motion parameterization when classifying highly dynamic video.
2.1. Existing Datasets
There are a handful of relevant datasets that exist in
the research domain. We highlight some of the more relevant video datasets that are appropriate to our work. All
of these datasets demonstrate the growing need for understanding what type of motion features are relevant in classifying highly dynamic actions, which we explore in this
work.
2.1.1
• Ballet
• Flamenco
• Latin
• Square
• Tango
UCF-101
The UCF-101 dataset [22] contains approximately 13,000
clips and 101 action classes, totaling 27 hours of data. The
clip length varies largely from 1 second to 71 seconds depending on the activity at a resolution of 320x240. This was
one of the first datasets to tackle human actions in video.
However, as we will demonstrate in this work, most perframe (image-based) approaches still perform moderately
well on the dataset, illustrating the main question which we
seek to answer in this work – that being the representation
of motion as a classification feature.
2.1.2
• Break Dancing
• Foxtrot
• Quickstep
• Swing
• Waltz
The dataset contains 100 videos for each class. Each
video is 10-seconds long at 30 frames per second. The
videos themselves were taken from YouTube at a quality of
720p, and includes both dancing performances and plainclothes practicing. Examples of each class can be seen in
Figure 1.
We highlight that the dataset contains four different types
of ballroom dancing (quickstep, foxtrot, waltz, and tango)
as seen in Figure 2. The motivation behind picking these
dances is that their parent category is specifically the setting
in which the dance occurs (a ballroom). This satisfies our
main challenge of selecting classes that exemplify highly
dynamic video. On this note, we extract two different motion representations from our input data for use by the community; optical flow[5] and multi-person pose detection.
When attempting to detect pose, we found numerous
methods that focused on single-person pose detection. We
adapted these methods to multiple individuals (given that
dancing is generally a group activity, see Figure 3) through
Kinetics
The Kinetics dataset [12] contains 300,000 clips and 400 action classes, with a minimum of 400 videos per class. The
action classes are also loosely grouped in 32 parent classes
which further break down the dataset. This dataset was collected semi-automatically with curation through image classifiers and use of Amazon Mechanical Turk to determine
the action classes and if the video snippet was appropriate
to that class.
3
4.1. Frame-by-Frame
Using the architecture of a state-of-the-art convolutional
neural network for image classification, such as VGG [21],
a classification for the video can be achieved based on key
image frames from a video. A sample architecture based on
CaffeNet, a variation of AlexNet [13], is shown in Figure 4.
This approach does not explicitly encode motion in determining the video’s classification but rather categorizes each
frame and naively selects the majority label.
We do note that although there are numerous approaches
for aggregating a single class from multiple per-frame classifications, the network itself does not encode the temporal
domain.
4.2. Two-Stream Late Fusion
A common way of adding a temporal component to deep
networks is by separately performing a classification based
on spatial data (a single frame) and temporal data (i.e. optical flow). Merging these results produces an overall classification for the video, as shown in Figure 5.
This approach computes optical flow from two frames
(at time n and n − k where k is not necessarily 1) over the
period of the entire video. Each frame in this case can be
considered a single instance of motion that occurred in the
video. For dancing we envision this as a specific move in a
dance.
Figure 3. Distribution of number of people per frame using [17].
75% of frames had at least two people detected in the dataset. 56%
of the dataset has more than two people in the shot, which further
illustrates the added complexity of this task.
the use of a recent real-time person detector[17]. Similar
approaches can be seen in [7][16][2].
After detecting the bounding boxes for each person in
the scene we computed the pose for each individual using
[30]. Positive and negative examples of this methodology
can be seen in Figure 7.
4. Baseline Methods
Figure 5. Two-Stream Late Fusion Architecture (color key in 4).
This method incorporates motion (optical flow) into a traditional
CNN pipeline.
In order to better understand the need for motion parametrization in video, we have identified two
commonly-used architectures to establish as our baseline.
These are architectures which are commonly applied to
video architectures but only take a single-frame as input
(per architecture).
5. Proposed Approaches
In order to address the challenge of categorizing highly
dynamic videos we implement a number of methods which
explicitly encode motion. At the core of these approaches
is the notion of 3-dimensional kernels which process a series of video frames for classification. This enables us to
pass in very short video clips (16 frames or approx. 1/2 second) for the network to learn. The overall objective was to
incorporate motion in the learning pipeline of standard approaches and assess their performance. After testing these
approaches it was evident that combining numerous motion
parameterizations in a concurrent deep network architecture
would best represent the input video.
These approaches are extensions of very successful image classification techniques.
5.1. Temporal 3D CNN (RGB)
Figure 4. Frame-by-Frame Architecture: This is a traditional
CNN, commonly used in image recognition.
As stated, traditional convolutional neural networks can
be extended to video by using 3-dimensional kernels that
4
frames that are convolved in the temporal domain.
5.3. Three-Stream CNNs
We tested both single-frame and temporal approaches
for a three-steam convolutional network in order to directly compare the potential importance of embedding multiple frames into the learning pipeline in addition to providing multiple representations of your original input. We
highlight that these temporal convolutions are computing
2D convolutions over each of the input frames. Although
this increases the complexity of our model it still remains
significantly more tractable than computing 3D convolutions which require approximately twice the computational
power.
Figure 6. This pipeline displays a skeletal temporal CNN (3D Convolution) which processes the initial frames to obtain a multi person pose estimation from the input frames obtained by performing
a bounding box person detection from [17] which is then processed
by [30] for detecting the dancers’ pose.
5.3.1
Frame-by-Frame Architecture
For the frame-by-frame architecture, the first stack of our
network processes the spatial representation of our input
which is our RGB image. Our second stack processes
the optical flow representation which was computed from
frames n and n − 10 in order to accentuate particular motions from a given dance. Our third stack processes our
multi-person pose visualization explained in Figure 6. As
discussed earlier, this stack is essentially encoding the number of participants detected for a given dance frame and
their current pose.
Figure 7. Demonstration of outputs from our pose detection
pipeline. Top: Latin dancing positively classified. Bottom: Break
dancing being erroneously classified. The dancers’ left leg is accurate but his remaining limbs fail.
5.3.2
Temporal Architecture
convolve in the temporal domain. We focus on testing
this slow-fusion approach discussed in [9], which embeds
the high-level spatial and temporal information at the initial convolutional layers by propagating the information
through the network. One of the main setbacks of this proposed approach is the computational time it currently takes
to compute these methodologies. We discuss this further in
Section 7.
The temporal architecture utilizes the same three stacks but
processes chunks of 16 frames at a time in order to incorporate a temporal component into the loss of the network.
This enables us to learn motion parameters from the spatial, optical flow and multi-person pose representations. A
visualization of our pipeline can be seen in Figure 8 whose
convolutional and fully connected layers are based on the
standard AlexNet architecture[13].
5.2. Temporal 3D CNN (Skeletal)
6. Baseline Experiments
In this pipeline we compute a temporal CNN on multiperson pose information. We visualize the pipeline in Figure 6. This architecture demonstrates the importance of
leveraging context for particular videos. Dance videos inherently benefit from this representation given that there are
generally multiple people in the scene. Through the use of
a visualization of pose we are able to attain comparable results to single-frame CNN approaches. It is key to note that
this method does not use visual information from the original frame but solely visualized pose information as shown
in Figure 7. Similar to our optical flow approach, it is likely
that this method benefits heavily from encoding the number
of people in the frame in addition to the motion over the 16
We implement our proposed approaches with the goal of
determining which approach is most effective at highly dynamic video classification. Implementation details for each
approach are given below.
6.1. Dataset Splits
We extract individual frames from the Let’s Dance
dataset (1000 10-second videos at 30fps, resulting in
300000 frames), which we then randomly split per video
into 80% training, 10% testing and 10% validation (consistent across experiments). Optical flow and pose detection
was split in the same manner in order to consistently test
the approaches.
5
Figure 8. This visualizes the workflow for our three-stream temporal CNN which uses three convolutional stacks to process the spatial and
respective motion components. It aggregates the fc7 layers into one and outputs the dance classification for a 16 frame input.
6.2. Frame-by-Frame
that it is merely embedding the motion between two frames.
We will later demonstrate that larger frame chunks provide
significant improvements to this approach.
To perform a baseline video classification experiment,
we implemented the architecture shown in Figure 4 in
Tensorflow[1]. The weights for the network’s convolutional
layers are initialized to values from a network pre-trained
on the ILSVRC 2012 dataset [19]. Final video classification results can be determined by classifying each frame in
a video and voting to determine the video’s overall class.
For an initial comparison, we also tested the network
with optical flow imagery as the input.
Overall, we observed significant amounts of overfitting
in the original training accuracy which hints at the network learning too much about the appearance of the specific videos in the training set for each class. As we hypothesized, using image frames alone results in the network
learning features that do not generalize well to the dancing categories, since it has no way to observe the motion
inherent in the video. Testing accuracy peaks at 56.4%
over 10,000 iterations of fine-tuning the network. We compare these results to a similar framework introduced by [29]
which tested the frame-by-frame baseline on UCF-101, attaining an accuracy of 72.8%. This directly demonstrates
the possibility of solving the classification problem by carefully selecting the right frame versus understanding the underlying motion of the video.
We also ran the identical setup using optical flow estimation. Before training we pre-compute optical flow for
the entire dataset. We used Farneback’s method for calculating dense optical flow [5] to obtain a per-pixel estimate of
the horizontal and vertical components of motion and then
incorporate this into the same network architecture.
In this case we saw slightly worse performance at approximately 45% for testing. We do note that the overfitting
for optical flow images is subdued given that the per-frame
images no longer contain background information. Given
that a number of our dances occurred in similar or identical
settings, background information was a strong confounding
factor for the original images. The overall result for optical flow performs worse than training on RGB images given
6.3. Two-Stream Late Fusion
We implemented the two-stream late fusion architecture
shown in Figure 5 in Caffe[10]. The two-stream approach
follows intuitively from the previous subsection in which
we discuss the effects of both a frame-by-frame method on
images and on optical flow. Each individual stream uses the
CaffeNet architecture, with weights initialized to a network
pre-trained on the ILSVRC 2012 dataset [19]. We then finetune the network by training only the fully-connected layers
at the end of each stream, which are then concatenated and
passed through a final fully-connected layer which outputs
the respective classifications.
We note that each architecture in the two-stream method
is still using a single frame as input, and as such the network
is trained on a frame-by-frame basis. We chose to use the
CaffeNet architecture for each frame, initialized with the
ILSVRC 2012 weights, to be consistent with the baseline
frame-by-frame experiment described in the previous section. This allows us to perform a direct comparison between
the two-stream and frame-by-frame approaches, to determine the benefit of optical flow on this dataset. As with the
frame-by-frame approach, final video classifications can be
determined by classifying each individual frame and optical
flow image pair, followed by voting to determine an overall
class. It is interesting to note that the total per video classification accuracy of this method was 68.89% which is much
higher than the single frame-by-frame accuracy of 56.40%.
Although one may be compelled to argue that single-frame
motion is key to this classification, we refer back to Figure
3. This figure demonstrates that the frames throughout the
dataset also contain a tremendously varied number of participants. As we can further see in Figure 9, optical flow tends
to visually separate the dancers from the background, which
also explains the significant increase in the algorithm’s performance. In addition to the motion in a single frame pair,
6
two-stream comparison is comparing a two-stream accuracy
for UCF-101 that utilizes an SVM to combine its streams
whereas we concatenate the final layers of both convolutional streams into the fully connected output. As stated
earlier, this illustrates the core issue we encountered in looking for a highly dynamic dataset which further validates our
motivation to introduce the “Lets Dance” dataset to the research community.
Figure 9. An image of dancers performing ballet and their optical
flow estimation. As we can see, optical flow does a good job of
segmenting the subjects in the scene in addition to encoding their
motion.
7. Results & Discussion
In order to assess our temporal architectures we compare
with a number of state-of-the-art approaches that explicitly encode motion in order to determine their performance.
Overall it has become clear to us that we need to transition
from traditional per-frame CNN approaches when conducting video classification.
It is evident from Table 2 that methods which embed motion significantly outperform traditional methods and that
metrics to evaluate these approaches are necessary in order to better understand what each network architecture is
learning.
the foreground’s shape and representation is playing a key
role in the classification of the network.
The results demonstrate an improvement over each independent approach, with a classification accuracy of 64.69%
per-frame. This is a significant increase of 10% above the
imaging method and 20% over the optical flow method.
This increase was attained by combining the same architecture as the previous two methods, with the addition of a
single concatenation node to fuse the data at the end of the
network. It demonstrates that directly incorporating temporal data into a network can be immediately beneficial towards classifying video.
Leveraging the network to perform full video classification (rather than only per-frame classification), we tested
the trained network on our test set of videos, taking the class
with the largest number of per-frame votes as the final video
label. This resulted in a per-video classification accuracy of
66.14%. After further experimentation with the network architecture, we saw a significant improvement from computing a unique mean image to subtract from the optical flow,
which increased our accuracy to a final per-video classification rate of 68.89%.
The network performs well at classifying Ballet, Waltz,
Tango, Flamenco, and Foxtrot, with poor classification accuracy on Break and Swing dancing. Of particular interest is the network’s performance on Waltz, Tango and Foxtrot which occur in similar settings. As such, the network
shows that it’s capable of performing fine-grained classification within the Let’s Dance dataset.
Dataset
UCF-101 [22]
Let’s Dance
Frame-by-Frame
72.8%
56.4%
7.1. Temporal 3D CNN
In order to evaluate this approach we restructured our
data into 16-frame chunks that were needed as the input for
the 3D convolution. The network could be trained on the
3D features from 16-frame non-overlapping chunks of the
video. We fine-tuned from the network trained on UCF101
by [25]. This method yielded a per-video accuracy of
70.11%. This result was particularly impressive because it
demonstrated the inherent ability of a 3D convolution to extract motion features that are not explicitly computed. The
major drawback of this approach is its complexity. A 3D
convolution inherently takes significant computation for a
single-stream.
We were unable to perform multi-stream approaches using 3D convolutions due to this complexity. In order to
combat this we introduce more tractable approaches for
state-of-the-art graphics cards (Our current systems utilizes
Titan Z Pascal graphics cards) that achieve comparable performance by explicitly encoding motion into the network architecture. In addition to this we note that 3D convolutions
are limited to the initial input-size which in our case was 16
frames. This makes it difficult to encode more complex motions that last more than 1/2 second without sub sampling
frames which will invariably lead to a loss in detail. Most
temporal methods will invariably suffer from this limitation
given that variable inputs into a convolutional network has
not been fully explored.
Two-Stream
88.0%
68.89%
Table 1. Method Comparison of UCF-101 and Let’s Dance. UCF
Frame-by-Frame results obtained from [29], Two-Stream results
obtained from [20]
Lastly, we revisit our accuracy results with UCF-101, a
well-established activity recognition dataset. Table 1 illustrates high levels of accuracy on UCF-101 using the standard extensions of image classification techniques which
we discuss in this section. It is important to note that the
7.2. Skeletal Temporal 3D CNN
In order to embed human motion data, we incorporate
skeletal images into a temporal CNN. We visualize each
7
Approach
Frame-by Frame CNN
Two-Stream CNN
Temporal 3D CNN (RGB)
Temporal 3D CNN (Skeletal)
Three-Stream CNN
Temporal Three-Stream CNN
8. Conclusion and Future Work
Testing Accuracy
56.4%
68.89%
70.11%
57.14%
69.20%
71.60%
In this work we sought out to understand the effect of
motion on classifying videos. Recent work in the are has
demonstrated the relevance of these type of videos, most recently seen in [15] and [12]. The work we have conducted
demonstrates that traditional CNN approaches do not properly or intentionally encode motion in their methodology.
This fact is frequently overlooked by testing on videos that
do not inherently require motion. That was the primary motivator of this work. As we can see in Table 2, 3D convolution methods outperform more traditional approaches
by inherently encoding motion into their computation and
prediction. Similarly, two-stream methods that incorporate
optical flow can also leverage temporal features to significantly improve video classification.
Table 2. Comparison of numerous approaches and their testing accuracies on our dataset
pose into a single image which represents the pose for that
particular frame. We attained an accuracy of 57.14%. We
note that this accuracy still performs marginally better than
a frame-by-frame approach despite the fact that it does not
utilize the spatial (RGB) representation. Due to the computational complexity of running concurrent 3D convolutional
networks we propose a stacked 2D convolutional method
which allows us to combine multiple streams in a single
state-of-the-art graphics card.
This also opens up some potential for future work in incorporating optical flow and pose data. Hybrid approaches,
such as a three-stream temporal CNN, have the potential
to increase an algorithm’s understanding of the video. We
have also developed a more focused dataset that we believe
the research community will benefit from by intentionally
selecting highly dynamic actions in one specific class. We
tested a variety of traditional and more complex methods in
order to begin to understand the composition of our dataset
and its baseline performance. The Let’s Dance dataset will
continue to help us to assess adequate motion parameterization and hopefully assist in improving how we learn from
video data.
7.3. Frame-by-Frame Three-Stream CNN
Our Three-Stream Frame-by-Frame architecture utilizes
all three data modalities. We assess this as both a singleframe and as a stacked architecture in order to compare their
benefits and drawbacks. As shown in Table 2, this approach
attains an accuracy of 69.20%. This three-stream network
performs comparably to the two-stream fusion approach we
conducted as one of our baselines which indicates that there
is not a significant amount of information added from the
use of both skeletal and optical flow representations.
One of the biggest problems we ran into throughout this
research endeavor was determining the best classes to select for our dataset. Initially we had some intuition for
dancing and martial arts being adequate parent categories
but we quickly saw that martial arts represented a multiclass problem. Although dancing exhibits similar overlaps
the separation was much more evident when performing the
data collection. We also had to alternate between different
dances partly due to availability on YouTube and our own
understanding of these dances.
7.4. Temporal Three-Stream CNN
Logically, we extended our frame-by-frame approach
into the temporal domain by stacking the image input layers to produce a 16-frame chunk. This approach utilizes
the same input as the Temporal 3D CNN we implemented
at a much lower complexity for three streams. We saw this
method attain the best performance out of all of the methods
we evaluated, at an accuracy of 71.60%.
Looking at our most successful approaches, three-stream
methods and 3D convolution, we note that both achieve
very similar performance in per-video classification. However, the two methods are not equivalent in terms of computational resources. Beyond the increased workload and
restrictions inherent in appropriately formatting the data
for the temporal CNN, 3D convolution is much more
computationally-intensive at both training and testing time.
We observe that even though the temporal CNN was our
most successful approach, it may be sub-optimal when a
much simpler three-stream stacked convolutional network
approach is available.
One of the next steps we have considered in this work is
modifying the input data in order to blur out regions of the
video which are not in motion or considered background.
This would further enforce motion parameterization and
help us better understand how we can accomplish that to
improve more general video classification algorithms. This
could also be explored by independently classifying the
pose estimation which is significantly more challenging.
We seek to extend this work by continuing to develop human-motion representations that intentionally target these highly dynamic actions.
For more information, please visit: https://www.
cc.gatech.edu/cpl/projects/dance/
8
References
[15] C. Pantofaru, C. Sun, C. Gu, C. Schmid, D. Ross,
G. Toderici, J. Malik, R. Sukthankar, S. Vijayanarasimhan,
S. Ricco, et al. Ava: A video dataset of spatio-temporally
localized atomic visual actions. 2017. 3, 8
[16] L. Pishchulin, A. Jain, M. Andriluka, T. Thormählen, and
B. Schiele. Articulated people detection and pose estimation: Reshaping the future. In Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, pages
3178–3185. IEEE, 2012. 4
[17] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi. You
only look once: Unified, real-time object detection. arXiv
preprint arXiv:1506.02640, 2015. 4, 5
[18] M. D. Rodriguez, J. Ahmed, and M. Shah. Action mach
a spatio-temporal maximum average correlation height filter for action recognition. In Computer Vision and Pattern
Recognition, 2008. CVPR 2008. IEEE Conference on, pages
1–8. IEEE, 2008. 3
[19] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh,
S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein,
A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual
Recognition Challenge. International Journal of Computer
Vision (IJCV), 115(3):211–252, 2015. 6
[20] K. Simonyan and A. Zisserman. Two-stream convolutional
networks for action recognition in videos. In Z. Ghahramani,
M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger, editors, Advances in Neural Information Processing
Systems 27, pages 568–576. Curran Associates, Inc., 2014.
2, 7
[21] K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition. arXiv preprint
arXiv:1409.1556, 2014. 4
[22] K. Soomro, A. R. Zamir, and M. Shah. Ucf101: A dataset
of 101 human actions classes from videos in the wild. arXiv
preprint arXiv:1212.0402, 2012. 3, 7
[23] N. Srivastava, E. Mansimov, and R. Salakhutdinov. Unsupervised learning of video representations using lstms. CoRR,
abs/1502.04681, 2015. 2
[24] A. Toshev and C. Szegedy. Deeppose: Human pose estimation via deep neural networks. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition,
pages 1653–1660, 2014. 2
[25] D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri.
Learning spatiotemporal features with 3d convolutional networks. arXiv preprint arXiv:1412.0767, 2014. 7
[26] S. Venugopalan, M. Rohrbach, J. Donahue, R. Mooney,
T. Darrell, and K. Saenko. Sequence to sequence-video to
text. In Proceedings of the IEEE International Conference
on Computer Vision, pages 4534–4542, 2015. 2
[27] K. Wang, X. Wang, L. Lin, M. Wang, and W. Zuo. 3d human activity recognition with reconfigurable convolutional
neural networks. In Proceedings of the 22Nd ACM International Conference on Multimedia, MM ’14, pages 97–106,
New York, NY, USA, 2014. ACM. 2
[28] L. Wang, Y. Qiao, and X. Tang. Action recognition with
trajectory-pooled deep-convolutional descriptors. In The
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2015. 2
[1] M. Abadi and A. A. et. al. TensorFlow: Large-scale machine
learning on heterogeneous systems, 2015. Software available
from tensorflow.org. 6
[2] Z. Cao, T. Simon, S.-E. Wei, and Y. Sheikh. Realtime multiperson 2d pose estimation using part affinity fields. June
2017. 3, 4
[3] G. Chéron, I. Laptev, and C. Schmid. P-cnn: Pose-based cnn
features for action recognition. In Proceedings of the IEEE
International Conference on Computer Vision, pages 3218–
3226, 2015. 2
[4] J. Donahue, L. Anne Hendricks, S. Guadarrama,
M. Rohrbach, S. Venugopalan, K. Saenko, and T. Darrell. Long-term recurrent convolutional networks for visual
recognition and description. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition,
pages 2625–2634, 2015. 2
[5] G. Farnebäck. Two-frame motion estimation based on
polynomial expansion. In Image analysis, pages 363–370.
Springer, 2003. 3, 6
[6] C. Feichtenhofer, A. Pinz, and A. Zisserman. Convolutional two-stream network fusion for video action recognition. arXiv preprint arXiv:1604.06573, 2016. 1, 2
[7] G. Gkioxari, B. Hariharan, R. Girshick, and J. Malik. Using k-poselets for detecting people and localizing their keypoints. In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, pages 3582–3589, 2014. 4
[8] G. Gkioxari and J. Malik. Finding action tubes. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 759–768, 2015. 2
[9] S. Ji, W. Xu, M. Yang, and K. Yu. 3d convolutional neural networks for human action recognition. Pattern Analysis
and Machine Intelligence, IEEE Transactions on, 35(1):221–
231, Jan 2013. 2, 5
[10] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. In Proceedings of the 22nd ACM international conference on Multimedia, pages 675–678. ACM, 2014. 6
[11] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar,
and L. Fei-Fei. Large-scale video classification with convolutional neural networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2014. 2,
3
[12] W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijayanarasimhan, F. Viola, T. Green, T. Back, P. Natsev, et al.
The kinetics human action video dataset. arXiv preprint
arXiv:1705.06950, 2017. 3, 8
[13] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet
classification with deep convolutional neural networks. In
F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger,
editors, Advances in Neural Information Processing Systems
25, pages 1097–1105. Curran Associates, Inc., 2012. 4, 5
[14] J. Liu, A. Shahroudy, D. Xu, and G. Wang. Spatio-temporal
lstm with trust gates for 3d human action recognition. In
European Conference on Computer Vision, pages 816–833.
Springer, 2016. 2
9
[29] L. Wang, Y. Xiong, Z. Wang, and Y. Qiao. Towards good
practices for very deep two-stream convnets. arXiv preprint
arXiv:1507.02159, 2015. 6, 7
[30] S.-E. Wei, V. Ramakrishna, T. Kanade, and Y. Sheikh. Convolutional pose machines. arXiv preprint arXiv:1602.00134,
2016. 2, 4, 5
[31] W. Zou, S. Zhu, K. Yu, and A. Y. Ng. Deep learning of invariant features via simulated fixations in video. In F. Pereira,
C. J. C. Burges, L. Bottou, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 25,
pages 3203–3211. Curran Associates, Inc., 2012. 2
10
| 1cs.CV
|
arXiv:1803.03737v2 [cs.NE] 26 Mar 2018
Enhancing Evolutionary Optimization in Uncertain
Environments by Allocating Evaluations
via Multi-armed Bandit Algorithms
Xin Qiu
Risto Miikkulainen
Sentient Technologies, Inc
San Francisco, California
[email protected]
Sentient Technologies, Inc
San Francisco, California
The University of Texas at Austin
Austin, Texas
[email protected]
ABSTRACT
Optimization problems with uncertain fitness functions are common in the real world, and present unique challenges for evolutionary optimization approaches. Existing issues include excessively
expensive evaluation, lack of solution reliability, and incapability in
maintaining high overall fitness during optimization. Using conversion rate optimization as an example, this paper proposes a series of
new techniques for addressing these issues. The main innovation is
to augment evolutionary algorithms by allocating evaluation budget through multi-armed bandit algorithms. Experimental results
demonstrate that multi-armed bandit algorithms can be used to allocate evaluations efficiently, select the winning solution reliably and
increase overall fitness during exploration. The proposed methods
can be generalized to any optimization problems with noisy fitness
functions.
KEYWORDS
Evolutionary computation, multi-armed bandit algorithm, uncertain
environment
1
INTRODUCTION
In many real-world problems, fitness evaluations are noisy, and
such uncertainty makes evolutionary optimization difficult [14].
The true fitness of a solution can only be approximated by a large
number of samples [22], slowing down evaluations and making
results unreliable.
As a concrete example, consider the conversion rate optimization
(CRO) problem. CRO is an emerging field where web designs are
optimized to increase the percentage of converted visitors [25]. Conversion refers to any desired actions of website users, e.g., making
purchase, registering new account, or clicking on any desired links.
The true conversion rate of a web design can only be estimated via
a certain number of user interactions/visits. Recently, a new technology for CRO was developed by applying evolutionary algorithm
(EA) [20, 21]. Each website design is represented as a genome, and
will be evolved into better ones in terms of conversion rate.
Evolutionary CRO provides considerable advantages over traditional A/B or multivariant testing: the exploration in evolutionary
algorithm covers a large design space; evolution discovers and utilizes effective interactions among variables; the optimization of
website design is fully automated. Although impressive improvements over human design have been reported [20, 21], there are still
several open issues in evolutionary CRO. First, the evaluations for
candidate designs are expensive, and a fair amount of traffic may
be wasted on bad designs. Second, only weak statistical evidence
is necessary for winner selection among candidate designs, reducing reliability of the optimization outcome. Third, in some cases,
the target is to maintain a high overall conversion rate during the
optimization process instead of identifying a single best website
design after optimization; the current evolutionary CRO technique
can not accommodate to this demand.
To overcome these issues, a series of new mechanisms are proposed in this paper by augmenting evolutionary optimization with
multi-armed bandit (MAB) algorithms [2–4]. First, a new framework called MAB-EA is developed by utilizing MAB algorithms
for traffic allocation during fitness evaluation. The proposed framework aims at reducing the evaluation cost while maintaining the
optimization performance. Second, an enhanced variant of MAB-EA
is designed to select the winner reliably. The main idea is to include
an addtional verification phase, based on MAB algorithms, to the
end of the evolution process. Third, another variant of MAB-EA
is developed by introducing a new concept called asynchronous
statistics into MAB algorithms. The new variant is particularly well
suited for problems where overall fitness (average fitness over all
evaluations) during optimization needs to be maximized. Empirical
studies demonstrate that these techniques are effective in a simulated CRO domain. Although the studies in this work are based on
CRO domain, the principle ideas can be easily generalized to other
optimization problems with uncertain environments.
The remainder of this paper is organized as follows. Section
2 introduces the basic concepts of evolutionary CRO and MAB
algorithms. Section 3 explains the technical details of the new approaches and discusses the underlying rationales. Section 4 evaluates the proposed mechanisms via experiments. Section 5 provides
high-level discussions and suggests some future research diections.
2
BACKGROUND
This section provides a brief description of the basic concepts and
existing MAB algorithms that will be utilized in the proposed techniques.
2.1
Evolutionary Conversion Rate
Optimization
EA is a population-based metaheuristic inspired by natural evolution process. Each individual (genome) in the population represents
a single solution to the optimization problem, and these individuals will evolve through crossover, mutation and survival selection
iteratively. The most significant advantage of EAs is that they do
not make any assumption about the underlying landscape of the
optimization problems, thereby leading to exceptional ability in
finding good solutions to mathematically intractable problems [9].
Algorithm 1 UCB1 Algorithm
In evolutionary CRO [20, 21], each genome represents a web
interface design. The search space is pre-defined by the web designer. For each such space, the designer specifies the elements of
the interface and values that they can take. For instance in a landing
page, logo size, header image, button color, content order are such
elements, and they can each take on 2-4 values. Evolutionary CRO
searches for good designs in the space of possible combinations of
these values, which often number in millions. In each generation, all
the genomes will be evaluated with a fixed number of user interactions, and the conversion rates during evalution will be used as the
fitnesses for each genome. Fitness-proportionate selection is then
used to select parent genomes, and traditional genetic operations
such as crossover (recombination of the elements in two parent
genomes) and mutation (randomly changing one element in the
offspring genome) are performed to generate offspring candidates.
The same process will be repeated generation by generation until
the termination criterion is met, which usually means reaching a
fixed number of user interactions. In a typical winner-selection
application, the winning design is then selected among the best
candidates, with an estimate of its future performance. In campaignmode application, there is no winner but performance is measured
by the overall convergence rate throughout the entire experiment.
2.2
Require:
K: Total number of arms
1: for i = 1 to K do
2:
Pull arm i, and observe reward X i,0
3:
x̂ i = X i,0
4:
ni = 1
5: end for
6: for t = K + 1, K + 2, . . . do
p
7:
Pull arm i max := argmaxi x̂ i + 2 log (t)/ni , and observe
reward X i max,t
8:
9:
10:
x̂
×n
+X i max , t
x̂ i max = imax niimax+1
max
ni max = ni max + 1
end for
MAB problems due to its good theoretical guarantees [2]. The principle behind UCB is optimism in the face of uncertainty [15]. Generally, UCB constructs an optimistic guess on the potential reward
of each arm, and pulls the arm with the highest guess. Among the
UCB family of algorithms [11, 16, 18, 19], UCB1 [4] is a simple yet
efficient variant that can be directly applied to Bernoulli Bandits.
The optimistic guess in UCB1 is in the form of an upper confidence
bound derived from the Chernoff-Hoeffding inequality [7, 13].
Suppose X 1 , X 2 , . . . , X n are independent random variables on
Í
[0, 1] with expected values as µ 1 , µ 2 , . . . , µ n . For X = n1 ni=1 X i and
Í
µ = E(X ) = n1 ni=1 µ i , the Chernoff-Hoeffding inequality gives
an exponential upper bound on the probability that the value of X
deviates from its expectation for more than a:
Multi-armed Bandit Algorithms
This subsection explains the definition of multi-armed bandit problem, and introduces three representative multi-armed bandit algorithms used in this paper.
2
P(X > µ + a) ≤ e −2na .
2.2.1 Multi-armed Bandit Problem. In MAB problem, a slot machine with multiple arms is given, and the gambler has to decide
which arms to pull, how many times to pull each arm, and in which
order to pull them [28]. The most common is the stochastic MAB
problem, which is parameterized by the number of arms K, the
number of rounds n, and K fixed but unknown reward distributions
ν 1 , ν 2 , . . . , ν K associated with arm 1, arm 2, . . . , arm K, respectively.
For t = 1, 2, . . . , n, at round t, the agent (gambler) chooses an arm
It from the set of arms {1, 2, . . . , K } to pull, and observes a reward
sampled from ν It . Each reward sample is independent from the
past actions and observations. The CRO problem is a special case
(called Bernoulli bandit) of the general stochastic MAB problem:
the reward for each pull is either 1 or 0 (converted or not in CRO),
and for arm i the probability of success (reward = 1) is µ i .
An algorithm for the stochastic MAB problem must decide which
arm to pull at each round t, based on the outcomes of the previous
t − 1 pulls. In the classical MAB problem, the goal is to maximize
the cumulative sum of rewards over the n rounds [4, 24]. Since the
agent has no prior knowledge about the reward distributions, it
needs to explore the different arms, and at the same time, exploit
the seemingly most rewarding arms [3]. This goal aligns with the
campaign-mode application in CRO. For clarity of statement, we call
this type of problem the classical stochastic MAB problem. Another
target for stochastic MAB problem is to output a recommended arm
after a given number of pulls. The performance of the MAB algorithm is only evaluated by the average payoff of that recommended
arm. This goal aligns with the winner-selection application in CRO.
This is called pure exploration problem [5].
(1)
For the stochastic MAB problem, X i will be the reward of each pull
for a single arm and X will be the
p empirical average reward for that
arm after n pulls. Using a = 2 log (t)/n, where t represents the
total number of pulls so far for all the arms, P(X > µ + a) ≤ t −4 ,
which converges to zero very quickly as the total number of pulls
increases. This upper bound will then be used in UCB1 to balance
between exploration and exploitation. Algorithm 1 shows the basic
steps of UCB1 algorithm.
Two effects of the UCB1 algorithm needs to bepmentioned: first,
if one arm is never pulled, its corresponding x̂ i + 2 log (t)/ni will
grow at a rate that is faster than all the other pulled arms. This
means that an arm will never be permanently ruled out no matter
how poorly it performs. This effect encourages exploration. Second,
if one arm has been pulled as many times as the other arms but has a
significantly higher empirical mean reward x̂ i , it will be pulled more
subsequently. This effect encourages exploitation. A delicate tradeoff between exploration and exploitation is therefore the hallmark
of UCB1.
2.2.3 Thompson Sampling. Except for UCB, Thompson Sampling (TS) [27] is another good alternative MAB algorithm for the
classical stochastic MAB problem. The idea is to assume a simple prior distribution on the parameters of the reward distribution of every arm, and at each round, play an arm according to
its posterior probability of being the best arm [2]. The effectiveness of TS has been empirically demonstrated by several studies
[6, 12, 26], and the asymptotic optimality of TS has been theoretically proved for Bernoulli bandits [2, 17]. TS for Bernoulli bandits
utilizes beta distribution as priors, i.e., a family of continuous probability distributions on the interval [0, 1] parameterized by two
2.2.2 UCB Algorithm. Upper Confidence Bound (UCB) algorithm is arguably the most popular approach for solving classical
2
Algorithm 2 Thompson Sampling for Bernoulli Bandits
Algorithm 3 Successive Rejects Algorithm
Require:
K: Total number of arms
1: for i = 1 to K do
2:
Si = 0, Fi = 0
3: end for
4: for t = 1, 2, . . . do
5:
for i = 1 to K do
6:
Sample θ i,t from Beta(Si + 1, Fi + 1)
7:
end for
8:
Pull arm i max := argmaxi θ i,t , and observe reward X i max,t
9:
if X i max,t = 1 then
10:
Si = Si + 1
11:
else
12:
Fi = Fi + 1
13:
end if
14: end for
Require:
K: Total number of arms
n: Total number of pulls
Ensure:
Best arm Jn
1 ÍK 1 , n = 0
1: A1 = {1, . . . , K }, log(K) = 2 + i=2
0
i
2: for k = 1 to K − 1 do
3:
nk = ⌈ 1 Kn−K
+1−k ⌉
log(K )
4:
5:
6:
7:
8:
9:
3
positive shape parameters, denoted by α and β. The probability
density function (PDF) of the beta distribution Beta(α, β) is given
Γ(α +β )
by f (x; α, β) = Γ(α )Γ(β ) x α −1 (1 − x)β −1 , where Γ() is the gamma
function, and 0 ≤ x ≤ 1. The mean of Beta(α, β) is α α+β , and higher
α, β lead to tighter concentration of Beta(α, β) around the mean.
TS initially assumes each arm i to have prior reward distribution
Beta(1, 1), which is equivalent to uniform distribution on [0, 1]. At
round t, after having observed Si successes (reward = 1) and Fi
failures (reward = 0) in Si + Fi pulls for arm i, the reward distribution of arm i will be updated as Beta(Si + 1, Fi + 1). The algorithm
then samples from these updated reward distributions, and selects
the next arm to pull according to the sampled reward. Algorithm 2
describes the detailed procedure of TS.
1 n−K ⌉
log(K ) K −1
3.1
MAB-EA
In the original Evolutionary CRO framework [20, 21], the evaluation
of each candidate is performed in a static fashion: a fixed amount of
traffic will be allocated to each website design, and the fitness of each
candidate is measured by the conversion rate. In the case of CRO,
the evaluation is very expensive because it is consuming traffic on
a real website. Thus, the overall conversion rate during evolution
determines the real cost of the optimization process. One main
drawback of the original static evaluation is that a large amount
of traffic may be consumed by bad candidates. Thus, the goal is to
develop of a new framework that allocates traffic dynamically in a
more efficient way.
MAB algorithms are well suited for this role. More specifically,
each candidate web design is regarded as an arm, and each visit to
the website is equal to a pull. The reward of each visit to a single
web design is assumed to follow an unknown but fixed Bernoulli
distribution. The probability of getting reward 1 (the visited user is
successfully converted) is p and the probability of getting reward
0 (the visited user is not converted) is 1 − p, where p is the true
conversion rate of that web design. Given a fixed budget of traffic
(number of visits) for each generation, a Bernoulli MAB algorithm
will then be invoked to allocate traffic to the current candidates. The
fitness of each candidate is equivalent to its number of successful
conversions divided by its total visits (both numbers are counted
within the current generation). Based on these fitnesses, standard
EA operations such as parent selection, crossover, mutation and
survival selection will be conducted to generate the population for
next generation. Algorithm 4 depicts the procedure of the proposed
framework, namely MAB-EA.
Note that the goals of the MAB algorithm and the evaluation
phase in EA are inherently different: MAB algorithm only cares
about identifying the good arms efficiently, whereas EA aims at
estimating the fitnesses of all arms. In spite of this fact, MAB algorithm should not impair the optimization performance significantly,
and may even improve it. As shown in Algorithm 4, elite candidates
log(K )
times, and so on, except for the last two arms, which
are both pulled n K −1 = ⌈ 1 n−K
⌉ times, where log(K) = 21 +
log(K ) 2
ÍK 1
i=2 i . Therefore, the total number of pulls does not exceed the
budget n:
K −1
n 1 + n 2 + · · · + n K −1 + n K −1 ≤ K +
METHODOLOGY
This section describes the algorithmic details of the proposed approaches and mechanisms. A basic framework combining Evolutionary CRO technique with MAB algorithm is presented first, then
two enhanced variants are developed for tackling different use cases,
namely, Best Arm Identification and Campaign mode.
2.2.4 Successive Rejects Algorithm. Among many existing algorithms for solving the pure exploration problem [8, 10, 23], Successive Rejects (SR) algorithm stands out in that it is parameter-free
and independent of the scaling of the rewards [3]. The main task
for SR algorithm is to identify the best arm (the arm with truly best
mean reward) after a fixed number of pulls. Suppose we are given
K arms and n pulls. First the SR algorithm divides the n pulls into
K − 1 phases. At the end of each phase, the arm with the lowest
empirical mean reward will be discarded. During each phase, each
arm that has not been discarded yet will be pulled for equal number
of times. The only surviving arm after K − 1 phases, Jn , will be
recommended as the best arm.
The details of SR are described in Algorithm 3. As a result, one
arm is pulled n 1 = ⌈ 1 n−K
K ⌉ times, one arm is pulled n 2 =
⌈
for i ∈ Ak do
Pull arm i for nk − nk−1 rounds
end for
Ak +1 = Ak \argmini ∈Ak X̂ i,nk , where X̂ i,nk is the average
reward for arm i after nk pulls
end for
Let Jn be the unique element of AK
n−K 1 Õ
1
( +
) = n.
2
K
+
1−k
log(K)
k =1
(2)
The SR algorithm is essentially optimal because the regret (difference between the mean rewards of identified best arm and true best
arm) decreases exponentially at a rate which is, up to a logarithmic
factor, the best possible [3].
3
Algorithm 4 MAB-EA
Algorithm 5 Best Arm Identification Mode
Require:
K: Population size, G max : Maximum number of generations
T : Number of website visits for each generation
Ce : Percentage for elites, Cp : Percentage for parents
P1 : Initial population, Cm : Mutation probability
1: D = P 1 , D is the archive for storing evaluated candidates
2: for д = 1 to G max do
3:
Perform MAB algorithm on Pд with a traffic budget of T ,
record the number of conversions si and number of visits ni
within current generation for each candidate
4:
for i = 1 to K do
5:
Set fitness for candidate i as fi = si /ni
6:
end for
Create elite pool Eд as the best Ce percentile candidates in
7:
current generation
8:
Create parent pool Aд as the best Cp percentile candidates
in current generation
9:
Initialize offspring pool Oд as empty
10:
while Size of Eд + Oд is less than K do
11:
Perform fitness-proportionate selection on Aд to pick 2
parent candidates
Perform uniform crossover between the two parents to
12:
generate an offspring
13:
Perform mutation operation on the offspring, each element
of the offspring will have Cm probability to be randomly
altered
14:
if the offspring is not in D then
15:
Add the offspring to Oд
16:
Add the offspring to D
17:
end if
18:
end while
19:
Pд+1 = Eд + Oд
20: end for
Require:
Same control parameters as in Algorithm 4
Ke : Size for the elite pool
Te : Additional traffic for Best Arm Identification phase
1: D = P 1 , D is the archive for storing evaluated candidates
2: Initialize elite pool E as empty
3: for д = 1 to G max do
4:
Same as lines 3-6 in Algorithm 4
5:
Add the best Ce percentile candidates of current generation
to elite pool E
6:
while Size of E is larger than Ke do
7:
Remove the worst candidate from E
8:
end while
Create parent pool Aд as the best Cp percentile candidates
9:
in current generation
10:
Initialize offspring pool Oд as empty
11:
while Size of Oд is less than K do
12:
Same as lines 11-17 in Algorithm 4
13:
end while
14:
Pд+1 = Oд
15: end for
16: Perform pure exploration MAB algorithm on E with a traffic
budget of Te , return the identified best candidate
tests in Section 4.2 show that BAI mode can significantly improve
the reliability of identified best candidate without incurring any
additional cost.
One additional modification in BAI mode is the removal of elite
survival mechanism. No candidate is allowed to survive for more
than one generation, and all the candidates for next generation will
be totally new. The purpose for this modification is to further improve the explorative ability of the framework, considering the fact
that the evaluations are very expensive and only limited number of
generations (less than 10) are acceptable in real CRO cases. Since
the elite pool in BAI mode has already stored the outstanding candidates for every generation, the evolution can focus on exploring
more regions in search space.
play an important role in both parent selection and survival selection. Since MAB algorithms allocate more traffic to those promising
candidates, their fitnesses are actually more reliably estimated than
those of the least promising candidates. Selection mechanisms relying on good candidates are therefore further enhanced. Thus, the
proposed framework is expected to significantly increase the overall conversion rate during evolution without sacrificing the overall
optimization performance.
3.2
3.3
Campaign Mode with Asynchronous
Multi-armed Bandit Algorithm
The current evolutionary CRO technique is designed to identify
a good candidate at the end of optimization. However, in some
scenaria, the goal for CRO is to make the overall conversion rate
during optimization as high as possible. To fill this need, a Campaign
mode based on MAB-EA is developed by introducing a new concept
to existing MAB algorithms: asynchronous statistics.
The original MAB algorithms initialize the statistics (total reward, average reward, number of pulls, etc.) of all the arms as
0. In contrast, the MAB algorithms in Campaign mode run in an
asynchronous manner: All the candidates surviving from previous
generation preserve their statistics and use them to initialize the
MAB algorithm, which then updates them further, as usual. Taking
asynchronous TS as an example, each candidate has an Si and Fi ,
and these two numbers are updated over generations until the candidate fails to survive. The underlying rationale is that preservation
of statistics increases the survival probability of good candidates,
therefore Campaign mode focuses more on exploitation than exploration. Asynchronous MAB algorithms allocate more traffic to
Best Arm Identification Mode
One important goal for CRO is to identify a single best design that
can be delivered to website owner for long-term use. However,
during the normal evolutionary process, only weak statistical evidence is obtained. Although this evidence is enough to rank the
candidate designs for reproduction, it is not enough to identify the
true best candidate. To solve this problem, Best Arm Identification
(BAI) Mode (Algorithm 5) is developed based on the new MAB-EA
framework.
The BAI mode is applied after the evolution process has concluded. A MAB algorithm for pure exploration (e.g., SR algorithm)
will be performed on an elite pool, i.e., the collection of top candidates over all generations. A single winner will be returned after
the BAI phase. Although additional traffic is needed for running the
BAI phase, this cost can be compensated by extracting a small portion of traffic from each previous generation (e.g., 10%). Empirical
4
Algorithm 6 Campaign Mode with Asynchronous MAB Algorithm
a total of 8 elements in the website that needs to be optimized. The
elements have 5, 4, 2, 3, 4, 3, 3 and 4 choices, respectively. The
basic conversion rate for the website is 0.05, and the effect of each
element choice is within [−0.01, 0.01]. The mean conversion rate
for all possible designs is 0.04997, and the conversion rate for the
best possible design is 0.08494. This parametric setup is similar to
real world situations. For each simulated visit, a Bernoulli test with
success probability equal to the conversion rate of the web design
will be conducted. A successful trail corresponds to a successful
conversion, and givies a reward of 1. A failed trial returns a reward
of 0.
Require:
Same control parameters as in Algorithm 4 excluding Ce
1: Initialize the total number of conversions si and total number
of visits ni for each candidate as 0
2: for д = 1 to G max do
3:
Perform asynchronous MAB algorithm on Pд with a traffic
budget of T , update the total number of conversions si and
total number of visits ni for each candidate
4:
Same as lines 4-6 in Algorithm 4
Create parent pool Aд as the best Cp percentile candidates
5:
in current generation
6:
Remove the worst Cp percentile candidates from Pд
7:
Initialize offspring pool Oд as empty
8:
while Size of Pд + Oд is less than K do
9:
Same as lines 11-13 in Algorithm 4
10:
if the offspring is not in Pд + Oд then
11:
Initialize the total number of conversions si and total
number of visits ni for the offspring as 0
12:
Add the offspring to Oд
13:
end if
14:
end while
15:
Pд+1 = Pд + Oд
16: end for
4.2
the existing elites without reevaluating them from scratch, thus
improving overall conversion rate. Algorithm 6 summarizes the
structure of Campaign mode.
Except for asynchronous statistics in MAB algorithms, the Campaign mode differs from the original MAB-EA in two other aspects.
First, the duplication avoidance mechanism is weakened. Since exploration is not the first priority in Campaign mode, duplications
between different generations are allowed to encourage the revival
of underestimated candidates. Second, only the worst Cp percentile
candidates are replaced by the new offspring, which are generated
from the top Cp percentile candidates. By setting Cp as 20 or even
less, the portion of newly generated offspring is limited, and the
overall conversion rate is more stable. Moreover, because all the
offspring are generated based on the top candidates, the overall
quality of offspring tends to be better than purely random sampling.
Under these mechanisms, the crossover and mutation operations
can continue the exploration at a steady pace.
4
EMPIRICAL STUDY
This section evaluates the effectiveness of the proposed framework
and mechanisms via simulated experiments in the conversion rate
optimization domain. All the conclusions in this section are supported by t-test at a 5% significance level.
4.1
Overall Performance Evaluation
In order to evaluate the performance of the new framework, three
representative MAB algorithms (SR, TS and UCB1) are incorporated
into it, and an empirical comparison between these three variants
and the original evolutionary CRO algorithm is conducted. The
original algorithm is the same as Algorithm 4 except for traffic
allocation: instead of varying it based on a MAB algorithm, all
candidates evenly share the traffic budget. For convenience, the
original evolutionary CRO algorithm is named "Standard Method"
in the rest of the paper.
The traffic budget for each generation is fixed at 10,000, and the
maximun number of generations is set at 10 , conforming to cost
limitations in real-world CRO. The population size K is 20, and the
mutation probability Cm is 0.01. Different values of elite and parent
percentages, Ce and Cp , are tested to investigate the robustness of
the proposed framework. Two performance metrics are utilized:
one is the best conversion rate, i.e., the true conversion rate of
the best-performing candidate in each generation; the other is the
overall conversion rate for each generation, i.e., the total number of
conversions in one generation divided by the total number of visits
in that generation. Note that the overall conversion rate is different
from simply averaging the conversion rates of all the candidates,
because the traffic allocated to each candidate may be different.
Figure 1 shows the results based on 500 independent runs. From
Figure 1, it is clear that the proposed framework with TS and SR
significantly increases the overall conversion rate during evolution
without deteriorating the optimization performance. In fact, the
incorporation of MAB algorithms even improves the optimization
performance in terms of best conversion rate. Regarding the influence of Ce and Cp , larger values of these two parameters lead
to more explorative behaviors at a cost of overall conversion rate
at early stage. In real-world cases, the largest acceptable generation number is usually 5, so a reasonable choice for Ce and Cp
would be 20 or even less. Under these circumstances, the TS and
SR variants perform best both in terms of overall conversion rate
and best conversion rate. There are three explanations: first, the
MAB algorithm allocates more traffic to the promising candidates,
thereby increasing the overall conversion rate during evaluation;
second, since the top candidates receive more traffic from MAB
algorithm, the reliability of best performing candidate is enhanced;
third, under small Ce and Cp , the quality of offspring relies heavily
on the top candidates, and more reliable top candidates tend to
generate more reliable offspring. The overall quality of candidates
is therefore improved, and the overall conversion rate is further
increased in this way. Regarding UCB1, since the average reward
in
p the simulated CRO case is very low (e.g., 0.05), the second term
2 log (t)/ni in line 7 of Algorithm 1 plays a more important role
Experimental Setup
A CRO simulator was built to simulate the interactions between
users and website designs. There are multiple possible choices for
each element of a website, and each choice will increase or decrease the basic conversion rate of the website independently (that
is, interactions between elements are not modeled in these experiments: although they are important in real-world CRO, they do
not play a large role in traffic allocation). The effect of each choice
is predefined and kept fixed during the CRO process. For all the
experiments in this section, the following setup is used: There are
5
Figure 1: This figure shows the best conversion rate and overall conversion rate in each generation. The results are averaged over
500 independent runs for different Ce and Cp settings. TS and SR perform significantly better than Standard Method in terms
of both measures. The differences in best conversion rate are statistically significant after generation 1, and the differences in
overall conversion rate are statistically significant over all the generations.
Best Conversion Rate over Generations
in arm selection. This encourages evenly allocation of the traffic,
thereby leading to similar behaviors as in Standard Method.
Effectiveness of Best Arm Identification
Mode
0.070
Best Conversion Rate
4.3
0.075
This subsection demonstrates the effectiveness of BAI mode through
an experimental comparison with the Standard Method, MAB-EA,
and a state-of-the-art approach [22] where the average fitness
within a predefined neighborhood (in solution space) is used to evaluate candidates. In the experiments, the neighborhood approach
was further improved by considering all the previous candidates
when calculating neighbood fitnesses.
For a fair comparison, MAB-EA and neighborhood approach
have 11,000 visits per generation; BAI mode has 10,000 visits for
each generation and 10,000 additional visits in the BAI phase. The
other parameters are identical for all algorithms: Ce = 20, Cp = 20,
K = 20, G max = 15, Cm = 0.01. For BAI mode, Ke = 20, and
SR algorithm is used in BAI phase. For neighborhood approach,
neighborhood size is fixed at 5.
Figure 2 compares the best conversion rates of all the algorithms
over 500 independent runs. BAI mode consistently improves over
the Standard Method, MAB-EA, and neighborhood approach. It both
converges faster early on, and explores more efficiently later. After
generation 10, BAI mode significantly outperforms MAB-EA even
with less total traffic. The neighborhood approach’s performance
gradually improves with the collection of more candidates. However,
BAI mode is still more reliable than the neighborhood approach even
in later stages. Based on the experimental results, BAI mode allows
selecting a better winner, and estimates its future/true performance
0.065
UCB1
Successive Rejects
Thompson Sampling
Standard Method
UCB1+BAI
Successive Rejects+BAI
Thompson Sampling+BAI
Standard Method+BAI
Neighborhood Approach
0.060
0.055
0
2
4
6
8
10
Generation
12
14
16
Figure 2: Best conversion rate over generations. The methods with a BAI phase perform significantly better, i.e. they
allow identifying a candidate where true performance is significantly better than methods without a BAI phase. The
neighborhood approach is also better than non-BAI methods, but not as good as BAI. The results are averaged over 500
independent runs, and the performance differences between
BAI variants and non-BAI variants are statistically significant .
more accurately. It therefore provides important improvements in
practical applications.
6
Overall Conversion Rate over Generations
of them can be generalized to other optimization problems with
noisy fitness functions.
The main idea of MAB-EA is to utilize MAB algorithms to allocate the evaluation budget. Although poorly performing candidates
cannot receive sufficient traffic to accurate estimate of their true
fitnesses, the optimization performance does not deteriorate. This
is because the evolution pressure in EAs comes from the parent
selection and survival selection, and these two steps rely primarily
on those good candidates. That is, efficient detection of good candidates is more important than accurate evaluation of bad candidates.
Thus, MAB algorithms reduce evaluation cost without sacrificing
optimization performance.
The BAI mode is a significant improvement in settings where
result reliability is critical. The EA maintains an elite archive that
collects good candidates throughout the optimization process. After
evolution is finished, a pure exploration MAB algorithm (e.g., SR)
is performed on the elite archive to select the final winner. This
process amounts to a two-level winner selection, and reliability of
the optimization outcome is enhanced. The additional traffic for
the BAI phase is extracted from previous generations, thus no extra
cost is incurred.
For problems in which overall fitness during optimization matters, the new concepts in Campaign mode can be applied. Asynchronous MAB algorithms together with a high survival probability
and greedy offspring generation lead to a high yet stable overall
fitness during evolution. It is notable that only MAB algorithms
without sharing parameters among candidates are suitable for asynchronization, such as TS and SR but not UCB1.
From the experimental results in Section 4.3, neighborhood-based
fitness works better than original fitness if the candidate pool is big
enough. Therefore, one promising future direction is to combine
neighborhood-based fitness with MAB algorithms in later optimization stages. The reliability of winner candidate should be further
improved in this way. Another interesting future direction is to
introduce contextual bandit algorithms [1], in which the interrelations among variables are explicitly modeled. The model will then
be used by the MAB algorithms to allocate evaluation budget more
efficiently. Moreover, the model can be used in the crossover or
mutation operations to propagate promising variable combinations
more often, thus increasing overall performance and efficiency further. Third direction is the incorporation of asynchronous statistics
into BAI mode. Initializing elites in BAI phase with their statistics
in the main optimization phase may further increase the reliability
of best candidate.
Overall Conversion Rate
0.075
0.070
0.065
0.060
UCB1
Successive Rejects
Thompson Sampling
Standard Method
Asynchronous UCB1
Asynchronous Successive Rejects
Asynchronous Thompson Sampling
0.055
0.050
0
10
20
30
Generation
40
50
Figure 3: The overall conversion rate for entire optimization
process in Campaign Mode. The data point at generation д
shows the overall conversion rate until generation д. The
asynchronous versions of TS and SR perform significantly
better than other versions, leading to better conversion rate
over the entire campaign. The results are averaged over 500
independent runs, and the performance differences between
asynchronous versions and original versions are statistically
significant.
4.4
Effectiveness of Asynchronous MAB
Algorithm in Campaign Mode
The main difference between Campaign mode and MAB-EA is
the new asynchronous MAB algorithm. This section verifies the
effectiveness of asynchronous statistics in MAB algorithms via an
empirical comparison. In the experiments, SR, TS and UCB1 are
modified to run asynchronously and compared with their original
versions, as well as with the Standard Method. The same parameters
are used for all algorithms: Cp = 20, K = 20, Cm = 0.01, T = 10, 000.
Since Campaign mode generally run for longer period of time, G max
is set at 50.
Figure 3 compares the results over 500 independent runs. Asynchronous SR and asynchronous TS perform significantly better than
their original versions. For UCB1, the asynchronous version is better only in the early stages. This is because all the candidates in
UCB1 algorithm share the parameter t (total number of visits for all
the candidates). A candidate that has survived for a long period in
the asynchronous variant will lead to a very large t. A significant
bias towards less visited candidates will then be introduced during
the traffic allocation (line 7 of Algorithm 1), thereby wasting more
traffic in those unreliable candidates. The candidates in SR and TS
do not share any parameters, so the Campaign mode works properly
with their asynchronous versions, improving overall conversion
rate during optimization significantly.
5
6
CONCLUSION
This paper demonstrates how MAB algorithms can be used to make
EAs more effective in uncertain environments. First, the proposed
MAB-EA framework makes it possible to allocate the evaluation
budget more efficiently. Second, a more reliable two-level winner
selection mechanism is employed in BAI mode based on pure exploration MAB algorithms. Third, a Campaign mode with asynchronous MAB algorithms is designed to achieve high and stable overall
fitness during the entire campaign. All proposed mechanisms are
shown to be effective through experimental comparisons in the
simulated CRO domain. The ideas and conclusions in this work can
be generalized to other uncertain optimization problems as well.
DISCUSSION AND FUTURE WORK
The proposed mechanisms in this work solves three general issues
in uncertain optimization problems: how to allocate the evaluation
budget efficiently, how to make the optimization outcome reliable,
and how to maintain high overall fitness during evolution. Although
the new approaches are only demonstrated in the CRO domain, all
7
REFERENCES
[23] Volodymyr Mnih, Csaba Szepesvári, and Jean-Yves Audibert. 2008. Empirical
Bernstein Stopping. In Proceedings of the 25th International Conference on Machine
Learning (ICML ’08). ACM, New York, NY, USA, 672–679. https://doi.org/10.1145/
1390156.1390241
[24] Herbert Robbins. 1952. Some aspects of the sequential design of experiments.
Bull. Amer. Math. Soc. 58, 5 (09 1952), 527–535. https://projecteuclid.org:443/
euclid.bams/1183517370
[25] Khalid Salehd and Ayat Shukairy. 2011. Conversion Optimization: The Art and
Science of Converting Prospects to Customers. O’Reilly Media, Inc., Sebastopol,
CA.
[26] Steven L. Scott. 2010. A Modern Bayesian Look at the Multi-armed Bandit. Appl.
Stoch. Model. Bus. Ind. 26, 6 (Nov. 2010), 639–658. https://doi.org/10.1002/asmb.874
[27] William R. Thompson. 1933. On the Likelihood that One Unknown Probability
Exceeds Another in View of the Evidence of Two Samples. Biometrika 25, 3/4
(1933), 285–294. http://www.jstor.org/stable/2332286
[28] Richard Weber. 1992. On the Gittins Index for Multiarmed Bandits. The Annals of
Applied Probability 2, 4 (1992), 1024–1033. http://www.jstor.org/stable/2959678
[1] Alekh Agarwal, Daniel J. Hsu, Satyen Kale, John Langford, Lihong Li, and Robert E.
Schapire. 2014. Taming the Monster: A Fast and Simple Algorithm for Contextual
Bandits. CoRR abs/1402.0555 (2014). arXiv:1402.0555
[2] Shipra Agrawal and Navin Goyal. 2012. Analysis of Thompson Sampling for
the Multi-armed Bandit Problem. In COLT 2012 - The 25th Annual Conference on
Learning Theory, June 25-27, 2012, Edinburgh, Scotland. 39.1–39.26. http://www.
jmlr.org/proceedings/papers/v23/agrawal12/agrawal12.pdf
[3] Jean-Yves Audibert and Sébastien Bubeck. 2010. Best Arm Identification in
Multi-Armed Bandits. https://www.microsoft.com/en-us/research/publication/
best-arm-identification-multi-armed-bandits/
[4] Peter Auer, Nicolò Cesa-Bianchi, and Paul Fischer. 2002. Finite-time Analysis of
the Multiarmed Bandit Problem. Machine Learning 47, 2 (01 May 2002), 235–256.
https://doi.org/10.1023/A:1013689704352
[5] Sébastien Bubeck, Rémi Munos, and Gilles Stoltz. 2009. Pure Exploration in Multiarmed Bandits Problems. Springer Berlin Heidelberg, Berlin, Heidelberg, 23–37.
https://doi.org/10.1007/978-3-642-04414-4_7
[6] Olivier Chapelle and Lihong Li. 2011. An Empirical Evaluation of Thompson
Sampling. In Proceedings of the 24th International Conference on Neural Information
Processing Systems (NIPS’11). Curran Associates Inc., USA, 2249–2257. http://dl.
acm.org/citation.cfm?id=2986459.2986710
[7] Herman Chernoff. 1952. A Measure of Asymptotic Efficiency for Tests of a
Hypothesis Based on the sum of Observations. The Annals of Mathematical
Statistics 23, 4 (1952), 493–507. http://www.jstor.org/stable/2236576
[8] Carlos Domingo, Ricard Gavaldà, and Osamu Watanabe. 2002. Adaptive Sampling
Methods for Scaling Up Knowledge Discovery Algorithms. Data Mining and
Knowledge Discovery 6, 2 (01 Apr 2002), 131–152. https://doi.org/10.1023/A:
1014091514039
[9] Agoston E. Eiben and Jim Smith. 2015. From evolutionary computation to the
evolution of things. Nature 521 (27 05 2015), 476 EP –. http://dx.doi.org/10.1038/
nature14544
[10] Eyal Even-Dar, Shie Mannor, and Yishay Mansour. 2006. Action Elimination and
Stopping Conditions for the Multi-Armed Bandit and Reinforcement Learning
Problems. J. Mach. Learn. Res. 7 (Dec. 2006), 1079–1105. http://dl.acm.org/citation.
cfm?id=1248547.1248586
[11] Aurélien Garivier and Olivier Cappé. 2011. The KL-UCB Algorithm for Bounded
Stochastic Bandits and Beyond. (02 2011).
[12] OleâĂŘChristoffer Granmo. 2010. Solving twoâĂŘarmed Bernoulli bandit problems using a Bayesian learning automaton. International Journal of Intelligent Computing and Cybernetics 3, 2 (2010), 207–234. https://doi.org/10.1108/
17563781011049179 arXiv:https://doi.org/10.1108/17563781011049179
[13] Wassily Hoeffding. 1963. Probability Inequalities for Sums of Bounded Random
Variables. J. Amer. Statist. Assoc. 58, 301 (1963), 13–30. http://www.jstor.org/
stable/2282952
[14] Yaochu Jin and J. Branke. 2005. Evolutionary optimization in uncertain
environments-a survey. IEEE Transactions on Evolutionary Computation 9, 3
(June 2005), 303–317. https://doi.org/10.1109/TEVC.2005.846356
[15] Moto Kamiura and Kohei Sano. 2017. Optimism in the face of uncertainty supported by a statistically-designed multi-armed bandit algorithm. Biosystems 160,
Supplement C (2017), 25 – 32. https://doi.org/10.1016/j.biosystems.2017.08.004
[16] Emilie Kaufmann, Olivier Cappe, and Aurelien Garivier. 2012. On Bayesian
Upper Confidence Bounds for Bandit Problems. In Proceedings of the Fifteenth
International Conference on Artificial Intelligence and Statistics (Proceedings of
Machine Learning Research), Neil D. Lawrence and Mark Girolami (Eds.), Vol. 22.
PMLR, La Palma, Canary Islands, 592–600. http://proceedings.mlr.press/v22/
kaufmann12.html
[17] Emilie Kaufmann, Nathaniel Korda, and Rémi Munos. 2012. Thompson Sampling:
An Asymptotically Optimal Finite-time Analysis. In Proceedings of the 23rd International Conference on Algorithmic Learning Theory (ALT’12). Springer-Verlag,
Berlin, Heidelberg, 199–213. https://doi.org/10.1007/978-3-642-34106-9_18
[18] T.L Lai and Herbert Robbins. 1985. Asymptotically efficient adaptive allocation
rules. Advances in Applied Mathematics 6, 1 (1985), 4 – 22. https://doi.org/10.
1016/0196-8858(85)90002-8
[19] Odalric-Ambrym Maillard, Remi Munos, and Gilles Stoltz. 2011. A Finite-Time
Analysis of Multi-armed Bandits Problems with Kullback-Leibler Divergences.
19 (05 2011).
[20] Risto Miikkulainen, Neil Iscoe, Aaron Shagrin, Ron Cordell, Sam Nazari, Cory
Schoolland, Myles Brundage, Jonathan Epstein, Randy Dean, and Gurmeet Lamba.
2017. Conversion Rate Optimization Through Evolutionary Computation. In
Proceedings of the Genetic and Evolutionary Computation Conference (GECCO ’17).
ACM, New York, NY, USA, 1193–1199. https://doi.org/10.1145/3071178.3071312
[21] Risto Miikkulainen, Neil Iscoe, Aaron Shagrin, Ryan Rapp, Sam Nazari, Patrick McGrath, Cory Schoolland, Elyas Achkar, Myles Brundage, Jeremy Miller, Jonathan
Epstein, and Gurmeet Lamba. 2018. Sentient Ascend: AI-Based Massively Multivariate Conversion Rate Optimization. In Proceedings of the Thirtieth Innovative
Applications of Artificial Intelligence Conference. AAAI. http://nn.cs.utexas.edu/
?miikkulainen:iaai18
[22] Risto Miikkulainen, Hormoz Shahrzad, Nigel Duffy, and Phil Long. 2017. How
to Select a Winner in Evolutionary Optimization?. In Proceedings of the IEEE
Symposium Series in Computational Intelligence. IEEE. http://nn.cs.utexas.edu/
?miikkulainen:ssci17
8
| 9cs.NE
|
Lifting from the Deep: Convolutional 3D Pose Estimation from a Single Image
arXiv:1701.00295v4 [cs.CV] 11 Oct 2017
Denis Tome
University College London
[email protected]
Chris Russell
The Turing Institute and
The University of Edinburgh
[email protected]
Lourdes Agapito
University College London
[email protected]
http://visual.cs.ucl.ac.uk/pubs/liftingFromTheDeep
Abstract
We propose a unified formulation for the problem of
3D human pose estimation from a single raw RGB image
that reasons jointly about 2D joint estimation and 3D pose
reconstruction to improve both tasks. We take an integrated approach that fuses probabilistic knowledge of 3D
human pose with a multi-stage CNN architecture and uses
the knowledge of plausible 3D landmark locations to refine
the search for better 2D locations. The entire process is
trained end-to-end, is extremely efficient and obtains stateof-the-art results on Human3.6M outperforming previous
approaches both on 2D and 3D errors.
1. Introduction
Estimating the full 3D pose of a human from a single
RGB image is one of the most challenging problems in
computer vision. It involves tackling two inherently ambiguous tasks. First, the 2D location of the human joints, or
landmarks, must be found in the image, a problem plagued
with ambiguities due to the large variations in visual appearance caused by different camera viewpoints, external
and self occlusions or changes in clothing, body shape or
illumination. Next, lifting the coordinates of the 2D landmarks into 3D from a single image is still an ill-posed problem – the space of possible 3D poses consistent with the
2D landmark locations of a human, is infinite. Finding the
correct 3D pose that matches the image requires injecting
additional information usually in the form of 3D geometric
pose priors and temporal or structural constraints.
We propose a new joint approach to 2D landmark detection and full 3D pose estimation from a single RGB image that takes advantage of reasoning jointly about the estimation of 2D and 3D landmark locations to improve both
tasks. We propose a novel CNN architecture that learns to
combine the image appearance based predictions provided
by convolutional-pose-machine style 2D landmark detectors [44], with the geometric 3D skeletal information en-
coded in a novel pretrained model of 3D human pose.
Information captured by the 3D human pose model is
embedded in the CNN architecture as an additional layer
that lifts 2D landmark coordinates into 3D while imposing that they lie on the space of physically plausible poses.
The advantage of integrating the output proposed by the 2D
landmark location predictors – based purely on image appearance – with the 3D pose predicted by a probabilistic
model, is that the 2D landmark location estimates are improved by guaranteeing that they satisfy the anatomical 3D
constraints encapsulated in the human 3D pose model. In
this way, both tasks clearly benefit from each other.
A further advantage of our approach is that the 2D and
3D training data sources may be completely independent.
The deep architecture only needs that images are annotated
with 2D poses, not 3D poses. The human pose model is
trained independently and exclusively from 3D mocap data.
This decoupling between 2D and 3D training data presents
a huge advantage since we can augment the training sets
completely independently. For instance we can take advantage of extra 2D pose annotations without the need for 3D
ground truth or extend the 3D training data to further mocap
datasets without the need for synchronized 2D images.
Our contribution: In this work, we show how to integrate
a prelearned 3D human pose model directly within a novel
CNN architecture (illustrated in figure 1) for joint 2D landmark and 3D human pose estimation. In contrast to preexisting methods, we do not take a pipeline approach that
takes 2D landmarks as given. Instead, we show how such
a model can be used as part of the CNN architecture itself,
and how the architecture can learn to use physically plausible 3D reconstructions in its search for better 2D landmark
locations. Our method achieves state-of-the-art results on
the Human3.6M dataset both in terms of 2D and 3D errors.
2. Related Work
We first describe methods that assume that 2D joint locations are provided as input and focus on solving the 3D
Fusion
2D
fusion
predicted
belief maps
Fusion
Loss
Fusion
Loss
Loss
Final 3D pose
fused
belief maps
projected pose
belief maps
Input image
9
9
9
9
9
9
9
9
Output 2D pose
11 1
1
predicted
belief maps
Probabilistic 3D
pose model
predicted
belief maps
predicted
belief maps
projected pose
belief maps
Feature extraction
2D joint prediction
3D/2D
projection
projected pose
belief maps
3D pose
STAGE 1
3D lifting &
projection
Fusion
2D
fusion
fused
belief maps
2D Loss
STAGE 2
STAGE 6
Feature extraction
2D joint prediction
3D lifting &
projection
Fusion
2D Loss
Feature extraction
2D joint prediction
3D lifting &
projection
Probabilistic 3D
pose model
Fusion
2D Loss
Final 3D pose
Figure 1: The multistage deep architecture for 2D/3D human pose estimation. Each stage produces as output a set of belief maps for the
location of the 2D landmarks (one per landmark). The belief maps from each stage, as well as the image, are used as input to the next
stage. Internally, each stage learns to combine: (a) belief maps provided by convolutional 2D joint predictors, with (b) projected pose
belief maps, proposed by the probabilistic 3D pose model. The 3D pose layer is responsible for lifting 2D landmark coordinates into 3D
and projecting them onto the space of valid 3D poses. These two belief maps are then fused into a single set of output proposals for the 2D
landmark locations per stage. The accuracy of the 2D and 3D landmark locations increases progressively through the stages. The loss used
at each stage requires only 2D pose annotations, not 3D. The overall architecture is fully differentiable – including the new projected-pose
belief maps and 2D-fusion layers – and can be trained end-to-end using back-propagation. [Best viewed in color.]
lifting problem and follow with methods that learn to estimate the 3D pose directly from images.
3D pose from known 2D joint positions: A large body
of work has focused on recovering the 3D pose of people
given perfect 2D joint positions as input. Early approaches
[19, 34, 25, 6] took advantage of anatomical knowledge of
the human skeleton or joint angle limits to recover pose
from a single image. More recent methods [13, 28, 3] have
focused on learning a prior statistical model of the human
body directly from 3D mocap data.
Non-rigid structure from motion approaches (NRSfM)
also recover 3D articulated motion [8, 4, 14, 20] given
known 2D correspondences for the joints in every frame
of a monocular video. Their huge advantage, as unsupervised methods, is they do not need 3D training data, instead
they can learn a linear basis for the 3D poses purely from
2D data. Their main drawback is their need for significant
camera movement throughout the sequence to guarantee accurate 3D reconstruction. Recent work on NRSfM applied
to human pose estimation has focused on escaping these
limitations by the use of a linear model to represent shape
variations of the human body. For instance, [10] defined
a generative model based on the assumption that complex
shape variations can be decomposed into a mixture of primitive shape variations and achieve competitive results.
Representing human 3D pose as a linear combination of
a sparse set of 3D bases, pretrained using 3D mocap data,
has also proved a popular approach for articulated human
motion [28, 43, 49], while [49] propose a convex relaxation
to jointly estimate the coefficients of the sparse representation and the camera viewpoint [28] and [43] enforce limb
length constraints. Although these approaches can reconstruct 3D pose from a single image, their best results come
from imposing temporal smoothness on the reconstructions
of a video sequence.
Recently, Zhao et al. [47] achieved state-of-the-art results by training a simple neural network to recover 3D pose
from known 2D joint positions. Although the results on
perfect 2D input data are impressive, the inaccuracies in 2D
joint estimation are not modeled and the performance of this
approach combined with joint detectors is unknown.
3D pose from images: Most approaches to 3D pose inference directly from images fall into one of two categories: (i)
models that learn to regress the 3D pose directly from image
features and (ii) pipeline approaches where the 2D pose is
first estimated, typically using discriminatively trained part
models or joint predictors, and then lifted into 3D. While
regression based methods suffer from the need to annotate
all images with ground truth 3D poses – a technically complex and elaborate process – for pipeline approaches the
challenge is how to account for uncertainty in the measurements. Crucial to both types of approaches is the question
of how to incorporate the 3D dependencies between the different body joints or to leverage other useful 3D geometric
information in the inference process.
Many earlier works on human pose estimation from a
single image relied on discriminatively trained models to
learn a direct mapping from image features such as silhouettes, HOG or SIFT, to 3D human poses without passing
through 2D landmark estimation [1, 12, 11, 24, 32].
Recent direct approaches make use of deep learning [21,
22, 40, 41]. Regression-based approaches train an end-toend network to predict 3D joint locations directly from the
image [41, 21, 22, 48]. Li et al. [22] incorporate model joint
dependencies in the CNN via a max-margin formalism, others [48] impose kinematic constraints by embedding a dif-
ferentiable kinematic model into the deep learning architecture. Tekin et al. [35] propose a deep regression architecture
for structured prediction that combines traditional CNNs for
supervised learning with an auto-encoder that implicitly encodes 3D dependencies between body parts.
As CNNs have become more prevalent, 2D joint estimation [44] has become increasingly reliable and many recent
works have looked to exploit this using a pipeline approach.
Papers such as [9, 16, 40, 26] first estimate 2D landmarks
and later 3D spatial relationships are imposed between them
using structured learning or graphical models.
Simo-Serra et al. [33] were one of the first to propose
an approach that naturally copes with the noisy detections
inherent to off-the-shelf body part detectors by modeling
their uncertainty and propagating it through 3D shape space
while satisfying geometric and kinematic 3D constraints.
The work [31] also estimates the location of 2D joints before predicting 3D pose using appearance and the probable
3D pose of discovered parts using a non-parametric model.
Another recent example is Bogo et al. [7], who fit a detailed
statistical 3D body model [23] to 2D joint proposals.
Zhou et al. [50] tackles the problem of 3D pose estimation for a monocular image sequence integrating 2D, 3D
and temporal information to account for uncertainties in the
model and the measurements. Similar to our proposed approach, Zhou et al.’s method [50] does not need synchronized 2D-3D training data, i.e. it only needs 2D pose annotations to train the CNN joint regressor and a separate
3D mocap dataset to learn the 3D sparse basis. Unlike our
approach, it relies on temporal smoothness for its best performance, and performs poorly on a single image.
Finally, Wu et al. [45]’s 3D Interpreter Network, a recent
approach to estimate the skeletal structure of common objects (chairs, sofas, ...) bears similarities with our method.
Although our approaches share common ground in the decoupling of 3D and 2D training data and the use of projection from 3D to improve 2D predictions the network architectures are very different and, unlike us, they do not carry
out a quantitative evaluation on 3D human pose estimation.
3. Network Architecture
Figure 1 illustrates the main contribution of our approach, a new multi-stage CNN architecture that can be
trained end-to-end to estimate jointly 2D and 3D joint locations. Crucially it includes a novel layer, based on a probabilistic 3D model of human pose, responsible for lifting
2D poses into 3D and propagating 3D information about
the skeletal structure to the 2D convolutional layers. In this
way, the prediction of 2D pose benefits from the 3D information encoded. Section 4 describes the new probabilistic
3D model of human pose, trained on a dataset of 3D mocap data. Section 5 describes all the new components and
layers of the CNN architecture. Finally, Section 6 describes
experimental evaluation on the Human3.6M dataset where
we obtain state-of-the-art results. In addition we show qualitative results on images from the MPII and Leeds datasets.
4. Probabilistic 3D Model of Human Pose
One fundamental challenge in creating models of human
poses lies in the lack of access to 3D data of sufficient variety to characterize the space of human poses. To compensate for this lack of data we identify and eliminate confounding factors such as rotation in the ground plane, limb
length, and left-right symmetry that lead to conceptually
similar poses being unrecognized in the training data.
Simple preprocessing eliminates some factors. Size variance is addressed by normalizing the data such that the sum
of squared limb lengths on the human skeleton is one; while
left-right symmetry is exploited by flipping each pose in the
x-axis and re-annotating left as right and vice-versa.
4.1. Aligning 3D Human Poses in the Training Set
Allowing for rotational invariance in the ground-plane
is more challenging and requires integration with the data
model. We seek the optimal rotations for each pose such
that after rotating the poses they are closely approximated
by a low-rank compact Gaussian distribution.
We formulate this as a problem of optimization over a
set of variables. Given a set of N training 3D poses, each
represented as a (3 × L) matrix Pi of 3D landmark locations, where i ∈ {1, 2, .., N } and L is the number of human
joints/landmarks; we seek global estimates of an average
3D pose µ, a set of J orthonormal basis matrices1 e and
noise variance σ, alongside per sample rotations Ri and basis coefficients ai to minimize the following estimate
arg min
N
X
R,µ,a,e,σ i=1
||Pi − Ri (µ + ai · e) ||22
+
(1)
J
J
X
X
(ai,j · σj )2 + ln
σj2
j=1
j=1
Where ai · e = j ai,j ej is the tensor analog of a multiplication between a vector and a matrix, and || · ||22 is the
squared Frobenius norm of the matrix. Here the y-axis is
assumed to point up, and the rotation matrices Ri considered are ground plane rotations. With the large number of
3D pose samples considered (of the order of 1 million when
training on the Human3.6M dataset [15]), and the complex
inter-dependencies between samples for e and σ, the memory requirements mean that it is not possible to solve directly as a joint optimization over all variables using a nonlinear solver such as Ceres. Instead, we carefully initialize
P
1 When we say e is a set of orthonormal basis matrices we mean that
each matrix, if unwrapped into a vector, is of unit norm and orthogonal to
all other unwrapped matrices.
e)
f)
d)
h)
c)
b)
a)
g)
Figure 2: Visualization of the 3D training data after alignment
(see section 4.1) using 2D PCA. Notice how all poses have the
same orientation. Standing-up poses a), b), c) and d) are all close
to each other and far from sitting-down poses f) and h) which form
another clear cluster.
and alternate between performing closed-form PPCA [38]
to update µ, a, e, σ; and updating Ri using Ceres [2] to minimize the above error. As we do this, we steadily increase
the size of the basis from 1 through to its target size J. This
stops apparent deformations that could be resolved through
rotations from becoming locked into the basis at an early
stage, and empirically leads to lower cost solutions.
To initialize we use a variant of the Tomasi-Kanade [39]
algorithm to estimate the mean 3D pose µ. As the y component is not altered by planar rotations, we take as our estimate of the y component of µ, the mean of each point in
the y direction. For the x and z components, we interleave
the x and z components of each sample and concatenate
them into a large 2 N × L matrix M, and find the rank two
approximation of this such that M ≈ A · B. We then calculate  by replacing each adjacent pair of rows of A with
the closest orthonormal matrix of rank two, and take † M
as our estimate2 of the x and z components of µ.
The end result of this optimization is a compact lowrank approximation of the data in which all reconstructed
poses appear to have the same orientation (see Figure 2). In
the next section we extend the model to be described as a
multi-modal distribution to better capture the variations in
the space of 3D human poses.
4.2. A Multi-Modal Model of 3D Human Pose
Although the learned Gaussian model of section 4.1 can
be directly used to estimate the 3D (see Table 1), inspection of figure 2 shows that the data is not Gaussian distributed and is better described using a multiple modal distribution. In doing this, we are heavily inspired both by
approaches such as [27] which characterize the space of
2 A†
being the pseudo-inverse of A.
human poses as a mixture of PCA bases, and by related
works such as [42, 8] that represent poses as an interpolation between exemplars. These approaches are extremely
good at modeling tightly distributed poses (e.g. walking)
where samples in the testing data are likely to be close to
poses seen in training. This is emphatically not the case in
much of the Human3.6M dataset, which we use for evaluation. Zooming in on the edges of Figure 2 reveals many
isolated paths where motions occur once and are never revisited again.
Nonetheless, it is precisely these regions of low-density
that we are interested in modeling. As such, we seek a
coarse representation of the pose space that says something
about the regions of low density but also characterizes the
multi-modal nature of the pose space. We represent the data
as a mixture of probabilistic PCA models using few clusters, and trained using the EM-algorithm [38]. When using
a small number of clusters, it is important to initialize the
algorithm correctly, as accidentally initializing with multiple clusters about a single mode, can lead to poor density
estimates. To initialize we make use of a simple heuristic.
We first subsample the aligned poses (which we refer to
as P ), and then compute the Euclidean distance d among
pairs. We seek a set of k samples S such that the distance
between points and their nearest sample is minimized
X
arg min
min d(s, p)
(2)
S
p∈P
s∈S
We find S using greedy selection, holding our previous estimate of S constant, and iteratively selecting the next candidate s such that {s} ∪ S minimizes the above cost. A selection of 3D pose samples found using this procedure can
be seen in the rendered poses of Figure 2. In practice, we
stop proposing candidates when they occur too close to the
existing candidates, as shown by samples (a–d), and only
choose one candidate from the dominant mode.
Given these candidates for cluster centers, we assign
each aligned point to a cluster representing its nearest candidate and then run the EM algorithm of [38], building a
mixture of probabilistic PCA bases.
5. A New Convolutional Architecture for 2D
and 3D Pose Inference
Our 3D pose inference from a single RGB image makes
use of a multistage deep convolutional architecture, trained
end-to-end, that repeatedly fuses and refines 2D and 3D
poses, and a second module which takes the final predicted
2D landmarks and lifts them one last time into 3D space for
the final estimate (see Figure 1).
At its heart, the architecture is a novel refinement of the
Convolutional Pose Machine of Wei et al. [44], who reasoned exclusively in 2D, and proposed an architecture that
Stage 1
Stage 2
Stage 3
Stage 4
Stage 5
Stage 6
Stage 1
Stage 2
Stage 3
Stage 4
Stage 5
Stage 6
9.19 mm 7.30 mm 6.64 mm 3.34 mm 3.28 mm
3.10 mm
Figure 3: Results returned by different stages of the architecture. Top Left: Evolution of the 2D skeleton after projecting the 3D points
back into the 2D space; Bottom Left: Evolution of the beliefs for the landmark Left hand through the stages. Right: 3D skeleton with the
relative mean error per landmark in millimeters. Even with incorrect landmark locations, the model returns a physically plausible solution.
iteratively refined 2D pose estimations of landmarks using
a mixture of knowledge of the image and of the estimates
of landmark locations of the previous stage. We modify this
architecture by generating, at each stage, projected 3D pose
belief maps which are fused in a learned manner with the
standard maps. From an implementation point of view this
is done by introducing two distinct layers, the probabilistic
3D pose layer and the fusion layer (see Figure 1).
Figure 3 shows how the 2D uncertainty in the belief
maps is reduced at each stage of the architecture and how
the accuracy of the 3D poses increases with each stage.
5.1. Architecture of each stage
The sequential architecture consists of 6 stages. Each
stage consists of 4 distinct components (see Figure 1):
Predicting CNN-based belief-maps: we use a set of convolutional and pooling layers, equivalent to those used in
the original CPM architecture [44], that combine evidence
obtained from image learned features with the belief maps
obtained from the previous stage (t − 1) to predict an updated set of belief maps for the 2D human joint positions.
Lifting 2D belief-maps into 3D: the output of the CNNbased belief maps is taken as input to a new layer that uses
new pretrained probabilistic 3D human pose model to lift
the proposed 2D poses into 3D.
Projected 2D pose belief maps: The 3D pose estimated by
the previous layer is projected back onto the image plane
to produce a new set of projected pose belief maps. These
maps encapsulate 3D dependencies between the body parts.
2D Fusion layer: The final layer in each stage (described in
section 5.5) learns the weights to fuse the two sets of belief
maps into a single estimate passed to the next stage.
Final lifting: The belief maps produced as the output of the
final stage (t = 6) are then lifted into 3D to give the final
estimate for the pose (see Figure 1) using our algorithm to
lift 2D poses into 3D.
5.2. Predicting CNN-based belief-maps
Convolutional Pose Machines [44] can be understood as
an updating of the earlier work of Ramakrishna et al. [29] to
use a deep convolutional architecture. In both approaches,
at each stage t and for each landmark p, the algorithm returns dense per pixel belief maps bpt [u, v], which show how
confident it is that a joint center or landmark occurs in any
given pixel (u, v). For stages t ∈ {2, . . . , T } the belief
maps are a function of not just the information contained in
the image but also the information computed by the previous stage.
In the case of convolutional pose machines, and in our
work which uses the same architecture, a summary of the
convolution widths and architecture design is shown in Figure 1, with more details of training given in [44].
Both [29, 44] predict the locations of different landmarks
to those captured in the Human3.6M dataset. As such the
input and output layers in each stage of the architecture are
replaced with a larger set to account for the greater number
of landmarks. The new architecture is then initialized by
using the weights with those found in CPM’s model for all
preexisting layers, with the new layers randomly initialized.
After retraining, CPMs return per-pixel estimates of
landmark locations, while the techniques for 3D estimation
(described in the next section) make use of 2D locations.
To transform these belief maps into locations, we select the
most confident pixel as the location of each landmark
Yp = arg max bp [u, v]
(3)
(u,v)
5.3. Lifting 2D belief-maps into 3D
We follow [50] in assuming a weak perspective model,
and first describe the simplest case of estimating the 3D
pose of a single frame using a unimodal Gaussian 3D pose
model as described in section 4. This model is composed
of a mean shape µ, a set of basis matrices e and variances
σ 2 , and from this we can compute the most probable sample
Results from the Human3.6M dataset
1.5
1.5
1.0Z
0.5
0.0
1.0
0.5
0.4
-0.4
1.5
1.5
Z
1.0
1.0
0.5
0.5
2.0
1.5
Z
1.0 Z
0.5
0.0
0.0
-0.3
Y
0
0.1.0
-0.3
0.0
0.2
X
0.0 .3
-0
Y
X
0.2.0
0 .2
-0
0.4
0.2 .0
0 .3
-0
-0.3
00.1.0
Y
2.0
0.0
X
0.0
0.2
-0
0 .1
0.1.0
X
2.0
-0.4
Y
X
0.0
Y
Z
0.0
0.3
0.0 .2
-0
0.0 .3
-0
-0.1
00.1.0
1.5
1.0Z
0.5
0.0
0.4
X
Success cases from MPII and Leeds
Y
2.0
0.5
0.0
0.0
-2.2
Y
0.6
X
-0.6
X
0.0
0.0
-0.4
0.0 .3
-0
0.0
-0.2
0
0.1.0
0.8
0.5
0.4
Y
Z
1.0
0.5
0.0
-0.3
0
0.1.0
1.5
1.0 Z
Z
0.5
X
Failure cases from MPII and Leeds
1.5
1.5
1.0
Y
2.0
1.5
Z
1.0
0.5
0.0
1.0
0.4
-1.2
X
Y
0.7
0.0
-0.5
0.0
0.2
-0.5
-0.7
Y
0.0
X
-0.4
-0.5
0.0
0.0
0.0
0.2
Z
1.0
0.5
0.0
0.6
Y
Z
0.5
0.0
X
2.0
1.5
1.5
1.0
Figure 4: Left: Results from the Human3.6M dataset. The identified 2D landmark positions and 3D skeleton is shown for each pose taken
from different actions: Walking, Phoning, Greeting, Discussion, Sitting Down. Right: Results on images from the MPII [5] (columns 1 to
3) and Leeds [18] datasets (last column). The model was not trained on images as diverse as those contained in these datasets, however it
often retrieves correct 2D and 3D joint positions. The last row shows example cases where the method fails either in the identification of
2D or 3D landmarks.
2D Pose refinement in Human3.6M
Figure 5: Landmark refinement: Left: 2D predicted landmark positions; Right: improved predictions using the projected 3D pose.
from the model that could give rise to a projected image.
arg min ||Y − sΠER(µ + a · e)||22 + ||σ · a||22
(4)
R,a
Where Π is the orthographic projection matrix, E a known
external camera calibration matrix, and s the estimated perframe scale. Although, given R this problem is convex in
a and s together3 , for an unknown rotation matrix R the
problem is extremely non-convex – even if a is known – and
prone to sticking in local minima using gradient descent.
Local optima often lie far apart in pose space and a poor
optima leads to a significantly worse 3D reconstructions.
We take advantage of the matrix R’s restricted form that
allows it to be parameterized in terms of a single angle θ.
Rather than attempting to solve this optimization problem
3 To see this consider the trivial reparameterization where we solve for
sµ + b · e and then let a = b/s.
using local methods we quantize over the space of possible
rotations, and for each choice of rotation, we hold this fixed
and solve for s and a, before picking the minimum cost solution of any choice of R. With fixed choices of rotation the
terms ΠERµ and ΠERe can be precomputed and finding
the optimal a becomes a simple linear least square problem.
This process is highly efficient and by oversampling the
rotations and exhaustively checking in 10, 000 locations we
can guarantee that a solution extremely close to the global
optima is found. In practice, using 20 samples and refining
the rotations and basis coefficients of the best found solution
using a non-linear least squares solver obtains the same reconstruction, and we make use of the faster option of checking 80 locations and using the best found solution as our 3D
estimate. This puts us close to the global optima and has the
same average accuracy as finding the global optima. Moreover, it allows us to upgrade from sparse landmark locations
to 3D using a single Gaussian at around 3,000 frames a second using python code on a standard laptop.
To handle models consisting of a mixture of Gaussians,
we follow [27] and simply solve for each Gaussian independently and select the most probably solution.
5.4. Projecting 3D poses onto 2D belief maps
The projected pose model is interleaved throughout the
architecture (see Figure 1). The goal is to correct the beliefs
regarding landmark locations at each stage, by fusing extra
information about 3D physical plausibility. Given the solution R, s, and a from the previous component, we estimate
a physically plausible projected 3D pose as
Ŷ = sΠER(µ + a · e)
(5)
Directions
Discussion
Eating
Greeting
Phoning
Photo
Posing
Purchases
LinKDE [15]
Li et al. [22]
Tekin et al. [37]
Tekin et al. [35]
Tekin et al. [36]
Zhou et al. [50]
Sanzari et al. [31]
132.71
102.39
85.03
87.36
48.82
183.55
136.88
158.52
129.06
108.79
109.31
56.31
132.37
96.94
87.95
91.43
84.38
87.05
95.98
164.39
124.74
126.83
121.68
98.94
103.16
84.78
162.12
118.37
119.39
116.18
96.47
205.94
168.68
185.02
162.17
95.65
143.32
105.58
150.61
114.69
98.49
106.88
66.30
171.31
107.61
93.77
99.78
107.41
Ours - Single PPCA Model
Ours - Mixture PPCA Model
68.55
64.98
78.27
73.47
77.22
76.82
89.05
86.43
91.63
86.28
110.05
110.67
74.92
68.93
83.71
74.79
Sitting
Sitting Down Smoking Waiting Walk Dog Walking Walk Together
Average
LinKDE [15]
Li et al. [22]
Tekin et al. [37]
Tekin et al. [35]
Tekin et al. [36]
Zhou et al. [50]
Sanzari et al. [31]
151.57
136.15
73.76
124.52
116.89
243.03
205.65
170.4
199.23
129.63
162.14
118.21
85.08
107.42
97.84
170.69
146.66
116.91
118.09
65.94
177.13
132.17
128.11
130.53
113.72
114.23
130.46
96.60
69.97
65.86
65.75
62.08
79.39
92.58
127.88
77.21
94.83
97.70
102.21
162.14
125.28
100.08
113.01
93.15
Ours - Single PPCA Model
Ours - Mixture PPCA Model
115.94
110.19
185.72
173.91
88.25
84.95
88.73
85.78
92.37
86.26
76.48
71.36
77.95
73.14
92.96
88.39
Table 1: A comparison of the 3D pose estimation results of our approach on the Human3.6M dataset against competitors that follow
Protocol #1 for evaluation (3D errors are given in mm). We substantially outperform all other methods in terms of average error showing a
4.7mm average improvement over our closest competitor. Note that some approaches [37, 50] use video as input instead of a single frame.
which is then embedded in a belief map as
(
1 if(i, j) = Ŷp
p
b̂i,j =
0 otherwise.
(6)
and then convolved using Gaussian filters.
5.5. 2D Fusion of belief maps
The 2D belief maps predicted by the probabilistic 3D
pose model are fused with the CNN-based belief maps bp
according to the following equation
ftp = wt ∗ bpt + (1 − wt ) ∗ b̂pt
(7)
where wt ∈ [0, 1] is a weight trained as part of the end-toend learning. This set of fused belief maps ft is then passed
to the next stage and used as an input to guide the 2D reestimation of joint locations, instead of the belief maps bt
used by convolutional pose machines.
5.6. The Objective and Training
Following [44], the objective or cost function ct minimized at each stage is the the squared distance between
the generated fusion maps of the layer ftp , and ground-truth
belief maps bp∗ generated by Gaussian blurring the sparse
ground-truth locations of each landmark p
ct =
L+1
X
X
p=1 z∈Z
||ftp − bp∗ ||22
(8)
For end-to-end
training the total loss is the sum over all
P
layers t≤6 ct . The novel layers were implemented as
an extension of the published code of Convolutional Pose
Machines [44] inside the Caffe framework [17] as Python
layers, with weights updated using Stochastic Gradient Descent with momentum. Details of the novel gradient updates
used lifting estimates through 3d pose space are given in the
supplementary materials.
6. Experimental evaluation
Human3.6M dataset: The model was trained and tested
on the Human3.6M dataset consisting of 3.6 million accurate 3D human poses [15]. This is a video and mocap
dataset of 5 female and 6 male subjects, captured from 4 different viewpoints, that show them performing typical activities (talking on the phone, walking, greeting, eating, etc.).
2D Evaluation: Figure 5 shows how the 2D predictions are
improved by the projected pose model, reducing the overall mean error per landmark. The 2D error reduction using
our full approach over the estimates of [44] is comparable
in magnitude to the improvement due to the change of architecture moving from the work Zhou et al. [50] to the
state-of-the-art 2d architecture [44] (i.e. a reduction of 0.59
pixels vs. 0.81 pixels). See Table 2 for details.
3D Evaluation: Several evaluation protocols have been
followed by different authors to measure the performance
of their 3D pose estimation methods on the Human3.6M
dataset. Tables 1 and 2 show comparisons of the 3D pose
Evaluation of 3D error (mm)
Yasin et al. [46]
Rogez et al. [30]
Ours - Mixture PPCA Model
Evaluation of 3D error (mm)
Bogo et al. [7]
Ours - Mixture PPCA Model
Protocol #2
108.3
88.1
70.7
Protocol #3
82.3
79.6
Evaluation of 2D pixel error
Zhou et al. [50]
10.85
Trained CPM [44] architecture
Ours using 3D refinement
10.04
9.47
Table 2: Further evaluation on the Human3.6M dataset. Top two
tables compare our 3D pose estimation errors against competitors
on Protocols #2 or #3. Bottom table compares our 2D pose estimation error against competitors. Our approach, which lifts the 2D
landmark predictions into a plausible 3D model and then projects
them back into the image, substantially reduces the error. Note
that [50] use video as input and knowledge of the action label.
estimation with previous works, where we take care to evaluate using the appropriate protocol.
Protocol #1, the most standard evaluation protocol on
Human3.6M, was followed by [15, 22, 37, 35, 36, 50, 31].
The training set consists of 5 subjects (S1, S5, S6, S7, S8),
while the test set includes 2 subjects (S9, S11). The original frame rate of 50 FPS is down-sampled to 10 FPS and
the evaluation is on sequences coming from all 4 cameras
and all trials. The reported error metric is the 3D error i.e.
the Euclidean distance from the estimated 3D joints to the
ground truth, averaged over all 17 joints of the Human3.6M
skeletal model. Table 1 shows a comparison between our
approach and competing approaches using Protocol #1. Our
baseline method using a single unimodal probabilistic PCA
model outperforms almost every method in most action
types, with the exception of Sanzari et al. [31], which it
still outperforms on average across the entire dataset. The
mixture model improves on this again, offering a 4.76mm
improvement over Sanzari et al., our closest competitor.
Protocol #2, followed by [46, 30], selects 6 subjects (S1,
S5, S6, S7, S8 and S9) for training and subject S11 for
testing. The original video is down-sampled to every 64th
frame and evaluation is performed on sequences from all 4
cameras and all trials. The error metric reported in this case
is the 3D pose error equivalent to the per-joint 3D error up
to a similarity transformation (i.e. each estimated 3D pose
is aligned with the ground truth pose, on a per-frame basis,
using Procrustes analysis). The error is averaged over 14
joints. Table 2 shows a comparison between our approach
and other approaches that use Protocol #2. Although, our
model was trained using only the 5 subjects used for training in Protocol #1 (one fewer subject), it still outperforms
the other methods [30, 46].
Protocol #3, followed by [7], selects the same subjects
for training and testing as Protocol #1. However, evaluation is only on sequences captured from the frontal camera
(“cam 3”) from trial 1 and the original video is not subsampled. The error metric used in this case is the 3D pose
error as described in Protocol #2. The error is averaged
over a subset of 14 joints. Table 2 shows a comparison
between our approach and [7]. Our method outperforms
Bogo et al. [7] by almost 3mm on average, even though
Bogo et al. exploits a high-quality detailed statistical 3D
body model [23] trained on thousands of 3D body scans,
that captures both the variation of human body shape and
its deformation through pose.
MPII and Leeds datasets: The proposed approach
trained exclusively on the Human3.6M dataset can be used
to identify 2D and 3D landmarks of images contained in
different datasets. Figure 4 shows some qualitative results
on the MPII dataset [5] and on the Leeds dataset [18], including failure cases. Notice how the probabilistic 3D pose
model generates anatomically plausible poses even though
the 2D landmark estimations are not all correct. However,
as shown in bottom row, even small errors in 2D pose can
lead to drastically different 3D poses. These inaccuracies
could be mitigated without further 3D data by annotating
additional RGB images for training from different datasets.
7. Conclusion
We have presented a novel approach to human 3D pose
estimation from a single image that outperforms previous
solutions. We approach this as a problem of iterative refinement in which 3D proposals help refine and improve upon
the 2D estimates. Our approach shows the importance of
thinking in 3D even for 2D pose estimation within a single
image, with our method demonstrating better 2D accuracy
than [44], the 2D approach it is based upon. Our novel approach for upgrading from 2D to 3D is extremely efficient.
When using 3 models, as in Tables 1 and 2, the upgrade
for each stage in CPU-based Python code runs at approximately 1,000 frames a second, while a GPU-based realtime approach for Convolutional Pose Machines has been
announced. Integrating these systems to provide a reliable
real-time 3D pose estimator is a natural future direction, as
is integrating this work with a simpler 2D approach for realtime pose estimation on lower power devices.
Acknowledgments
This work was funded by the SecondHands project, from
the European Union’s Horizon 2020 Research and Innovation programme under grant agreement No 643950. Chris
Russell was partially supported by The Alan Turing Institute under EPSRC grant EP/N510129/1.
References
[1] A. Agarwal and B. Triggs. Recovering 3d human pose from
monocular images. IEEE transactions on pattern analysis
and machine intelligence, 28(1):44–58, 2006. 2
[2] S. Agarwal, K. Mierle, and Others. Ceres solver. http:
//ceres-solver.org. 4
[3] I. Akhter and M. J. Black. Pose-conditioned joint angle limits for 3d human pose reconstruction. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, pages 1446–1455, 2015. 2
[4] I. Akhter, Y. Sheikh, S. Khan, and T. Kanade. Trajectory
space: A dual representation for nonrigid structure from motion. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 2011. 2
[5] M. Andriluka, L. Pishchulin, P. Gehler, and B. Schiele. Human pose estimation: New benchmark and state of the art
analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 6, 8
[6] C. Barrón and I. A. Kakadiaris. Estimating anthropometry
and pose from a single uncalibrated image. Computer Vision
and Image Understanding, 81(3):269–284, 2001. 2
[7] F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero,
and M. J. Black. Keep it smpl: Automatic estimation of 3d
human pose and shape from a single image. In European
Conference on Computer Vision, pages 561–578. Springer,
2016. 3, 8
[8] C. Bregler, A. Hertzmann, and H. Biermann. Recovering
non-rigid 3d shape from image streams. In Computer Vision
and Pattern Recognition, 2000. Proceedings. IEEE Conference on, volume 2, pages 690–696. IEEE, 2000. 2, 4
[9] X. Chen and A. L. Yuille. Articulated pose estimation by a
graphical model with image dependent pairwise relations. In
Advances in Neural Information Processing Systems, pages
1736–1744, 2014. 3
[10] J. Cho, M. Lee, and S. Oh.
Complex non-rigid 3d
shape recovery using a procrustean normal distribution mixture model. International Journal of Computer Vision,
117(3):226–246, 2016. 2
[11] C. H. Ek, P. H. S. Torr, and N. D. Lawrence. Gaussian
process latent variable models for human pose estimation.
In A. Popescu-Belis, S. Renals, and H. Bourlard, editors,
MLMI, volume 4892 of Lecture Notes in Computer Science,
pages 132–143. Springer, 2007. 2
[12] A. Elgammal and C. Lee. Inferring 3d body pose from silhouettes using activity manifold learning. In CVPR, 2004.
2
[13] X. Fan, K. Zheng, Y. Zhou, and S. Wang. Pose locality constrained representation for 3d human pose reconstruction. In
European Conference on Computer Vision, pages 174–188.
Springer, 2014. 2
[14] P. Gotardo and A. Martinez. Computing smooth time trajectories for camera and deformable shape in structure from
motion with occlusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2011. 2
[15] C. Ionescu, D. Papava, V. Olaru, and C. Sminchisescu.
Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
transactions on pattern analysis and machine intelligence,
36(7):1325–1339, 2014. 3, 7, 8
A. Jain, J. Tompson, M. Andriluka, G. W. Taylor, and C. Bregler. Learning human pose estimation features with convolutional networks. arXiv preprint arXiv:1312.7302, 2013. 3
Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. arXiv preprint
arXiv:1408.5093, 2014. 7
S. Johnson and M. Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In Proceedings of the British Machine Vision Conference, 2010.
doi:10.5244/C.24.12. 6, 8
H.-J. Lee and Z. Chen. Determination of 3d human body
postures from a single view. Computer Vision, Graphics,
and Image Processing, 30(2):148–168, 1985. 2
M. Lee, J. Cho, and S. Oh. Procrustean normal distribution
for non-rigid structure from motion. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 2016. 2
S. Li and A. B. Chan. 3d human pose estimation from
monocular images with deep convolutional neural network.
In Asian Conference on Computer Vision, pages 332–347.
Springer, 2014. 2
S. Li, W. Zhang, and A. B. Chan. Maximum-margin structured learning with deep networks for 3d human pose estimation. In Proceedings of the IEEE International Conference
on Computer Vision, pages 2848–2856, 2015. 2, 7, 8
M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J.
Black. Smpl: A skinned multi-person linear model. ACM
Transactions on Graphics (TOG), 34(6):248, 2015. 3, 8
G. Mori and J. Malik. Recovering 3d human body configurations using shape contexts. PAMI, 2006. 2
V. Parameswaran and R. Chellappa. View independent human body pose estimation from a single perspective image.
In Computer Vision and Pattern Recognition, 2004. CVPR
2004. Proceedings of the 2004 IEEE Computer Society Conference on, volume 2, pages II–16. IEEE, 2004. 2
T. Pfister, J. Charles, and A. Zisserman. Flowing convnets
for human pose estimation in videos. In Proceedings of the
IEEE International Conference on Computer Vision, pages
1913–1921, 2015. 3
N. Pitelis, C. Russell, and L. Agapito. Learning a manifold as an atlas. In Computer Vision and Pattern Recognition (CVPR), 2013 IEEE Conference on, pages 1642–1649.
IEEE, 2013. 4, 6
V. Ramakrishna, T. Kanade, and Y. Sheikh. Reconstructing
3d human pose from 2d image landmarks. In European Conference on Computer Vision, pages 573–586. Springer, 2012.
2
V. Ramakrishna, D. Munoz, M. Hebert, J. A. Bagnell, and
Y. Sheikh. Pose machines: Articulated pose estimation via
inference machines. In European Conference on Computer
Vision, pages 33–47. Springer, 2014. 5
G. Rogez and C. Schmid. Mocap-guided data augmentation
for 3d pose estimation in the wild. In Advances in Neural
Information Processing Systems, pages 3108–3116, 2016. 8
[31] M. Sanzari, V. Ntouskos, and F. Pirri. Bayesian image based
3d pose estimation. In European Conference on Computer
Vision, pages 566–582. Springer, 2016. 3, 7, 8
[32] L. Sigal, R. Memisevic, and D. Fleet. Shared kernel information embedding for discriminative inference. In CVPR,
2009. 2
[33] E. Simo-Serra, A. Ramisa, G. Alenyà, C. Torras, and
F. Moreno-Noguer. Single image 3d human pose estimation from noisy observations. In Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, pages
2673–2680. IEEE, 2012. 3
[34] C. J. Taylor. Reconstruction of articulated objects from point
correspondences in a single uncalibrated image. In Computer Vision and Pattern Recognition, 2000. Proceedings.
IEEE Conference on, volume 1, pages 677–684. IEEE, 2000.
2
[35] B. Tekin, I. Katircioglu, M. Salzmann, V. Lepetit, and P. Fua.
Structured prediction of 3d human pose with deep neural networks. In British Machine Vision Conference (BMVC), 2016.
3, 7, 8
[36] B. Tekin, P. Márquez-Neila, M. Salzmann, and P. Fua. Fusing 2d uncertainty and 3d cues for monocular body pose estimation. arXiv preprint arXiv:1611.05708, 2016. 7, 8
[37] B. Tekin, X. Sun, X. Wang, V. Lepetit, and P. Fua. Predicting people’s 3d poses from short sequences. arXiv preprint
arXiv:1504.08200, 2015. 7, 8
[38] M. E. Tipping and C. M. Bishop. Probabilistic principal
component analysis. Journal of the Royal Statistical Society, Series B, 61:611–622, 1999. 4
[39] C. Tomasi and T. Kanade. Shape and motion from image
streams under orthography: a factorization method. International Journal of Computer Vision, 9(2):137–154, 1992.
4
[40] J. J. Tompson, A. Jain, Y. LeCun, and C. Bregler. Joint training of a convolutional network and a graphical model for
human pose estimation. In Advances in neural information
processing systems, pages 1799–1807, 2014. 2, 3
[41] A. Toshev and C. Szegedy. Deeppose: Human pose estimation via deep neural networks. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition,
pages 1653–1660, 2014. 2
[42] C. Wang, J. Flynn, Y. Wang, and A. L. Yuille. Representing data by a mixture of activated simplices. arXiv preprint
arXiv:1412.4102, 2014. 4
[43] C. Wang, Y. Wang, Z. Lin, , A. Yuille, and W. Gao. Robust
estimation of human poses from a single image. In Computer
Vision and Pattern Recognition (CVPR), 2014. 2
[44] S.-E. Wei, V. Ramakrishna, T. Kanade, and Y. Sheikh. Convolutional pose machines. arXiv preprint arXiv:1602.00134,
2016. 1, 3, 4, 5, 7, 8
[45] J. Wu, T. Xue, J. J. Lim, Y. Tian, J. B. Tenenbaum, A. Torralba, and W. T. Freeman. Single image 3d interpreter network. In European Conference on Computer Vision, pages
365–382. Springer, 2016. 3
[46] H. Yasin, U. Iqbal, B. Kruger, A. Weber, and J. Gall. A
dual-source approach for 3d pose estimation from a single
image. In IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), 2016. 8
[47] R. Zhao, Y. Wang, and A. Martinez. A simple, fast and
highly-accurate algorithm to recover 3d shape from 2d landmarks on a single image. arXiv preprint arXiv:1609.09058,
2016. 2
[48] X. Zhou, X. Sun, W. Zhang, S. Liang, and Y. Wei. Deep kinematic pose regression. arXiv preprint arXiv:1609.05317,
2016. 2
[49] X. Zhou, M. Zhu, S. Leonardos, and K. Daniilidis. Sparse
representation for 3d shape estimation: A convex relaxation
approach. arXiv preprint arXiv:1509.04309, 2015. 2
[50] X. Zhou, M. Zhu, S. Leonardos, K. Derpanis, and K. Daniilidis. Sparseness meets deepness: 3d human pose estimation
from monocular video. arXiv preprint arXiv:1511.09439,
2015. 3, 5, 7, 8
Supplementary Material
1. Computing derivatives for back-propagation through our lifted model
As discussed by the Convolution Pose Machine paper [1] recurrent like architectures such as ours have problems with
vanishing gradients and for effective training they require an additional loss function to be defined for each layer, that independently drives each individual layer to return correct predictions regardless of how this information is used in subsequent
layers.
Before we give the derivation of the gradients it should be emphasized that it is entirely possible to train the network
without using them – in fact similar results can be obtained by only using the 3D lifting for the forward pass, and not
back-propagating the lifting derivatives through the rest of the network. As the additional layers make use of custom Pythonbased derivatives rather than an efficient implementation, for computational reasons it might preferable to avoid this step.
Nonetheless for completeness we include the derivatives.
There are two reasons the gradients are unneeded: Our lifting 3D model we use makes its best predictions when the
2D predictions of the same layer are closest to ground truth, and this is a constraint naturally enforced by the objective of
equation (8) of the main paper. Further, as with Convolutional Pose Machines [1] our architecture suffers from problems with
vanishing gradients. To overcome this Wei et al. [1] defined an objective at each layer, which acted to locally strengthen the
gradients. However, a side effect of this multi-stage objective is that most of the effects of back-propagation happen locally
and gradients back-propagated from other layers have little effect on the learning. This makes subtle interactions between
layers less influential, and forces the learning process to concentrate on simply making accurate 2D predictions in each layer.
We first give the results for computing the gradients of sparse predicted locations Ŷ from Y (see section 5 of the main
paper), before discussing the gradients induced on the confidence maps by these sparse locations.
1.1. Landmark Gradients
In the interests of readability we neglect the use of indices to indicate stages, the reader should assume that all variables are
taken from the same stage. Similarly, when dealing with a mixture of Gaussians, as we are only interested in computing a subgradient, the reader should assume that the best model has already been selected in the forward pass and we are computing
gradients using only this model.
Recall (section 5 of main body of paper) that the mapping from the initial landmarks Y to the projected 3D proposals Ŷ
is given by
Ŷ = ΠR(µ + a · e)
(1)
where
R, a =
arg min
{R∗ ∈R,a∗ ∈RJ }
||Y − ΠR∗ (µ + a∗ · e)||22 + (σ · a∗ )2
(2)
where R is a discrete set of rotations we exhaustively minimize over, and J is the number of bases in e. Owing to the use
of discrete rotations, this mapping from Y to Ŷ is a piecewise smooth approximation of the smooth function defined over a
continious R, and sub-gradients can be induced by fixing R to its current state. Hence:
dŶ
= ΠRe
da
(3)
For the remainder of the section, and to compact notation we will write E for the matrix of size 2L × J (L the number of
landmark points and J being the number of bases in e ) formed by unwrapping tensor ΠRe. Similarly, we will unwrap the
2 × L matrices Y and Ŷ and write them as y and ŷ. We also write p for the vector representing the unwrapped set of 2D
landmark positions ΠRµ.
1
We will use [y, 0] for the vector formed by vector y followed by J zeros, and Ē for the matrix of size (2L + J) × J
formed by concatenating E with the matrix that has values σ along the diagonals and zero everywhere else. We can rewrite
equation (3) in its new notation as:
dŷ
=E
(4)
da
and given R, we can rewrite equation (2) as
a = arg min |[y, 0] − [p, 0] − a∗ Ē||22
(5)
a = [(y − p), 0]Ē †
(6)
da
= Ē †
d[y, 0]
(7)
dŷ da
dŷ
=
= E Ē t
dy
da dy
(8)
a∗∈RJ
or
with Ē † continuing to represent the pseudo-inverse of Ē. Hence
and
where Ē t is the truncation of Ē † .
1.2. Mapping belief gradients to coordinate transform
The coordinates of each predicted landmark Yˆp induce a Gaussian in the belief map b̂p . So a change in the x component
of Yˆp induces an update which is equivalent to a difference of Gaussians.
db̂p
G(Yˆp
≈
dYˆx
(x)
p
(x)
+ δx ) − G(Yˆp − δx )
2δx
(9)
and the same for the y component as well. For computational purposes we take δx as one pixel. As such, an induced gradient
on the projected belief map near the predicted location Ŷ b̂p induces an updating of Ŷ that is propagated through to Y using
the sub-gradients described in equation (8).
Updating B
such that
Writing B for the the set of all bp , and assuming Yp is not in the right location, i.e. given updates ∆B̂ on B̂
∆B̂.
dB̂ dŶ
6= 0,
dŶ dYp
any update of b in which we decrease the belief at bpYp and increase anywhere else is a valid sub-gradient. We choose as a
dŶ
sensible update a negative step at bp of magnitude m = k||∆B̂. ddŶB̂ dY
|| and a positive update for each element Y of of Bp
p
of the magnitude m · N (Y, σ 2 ) in the quadrant of a Gaussian of the same width used to generate b̂ (i.e. σ = 1 see section 5.6
dŶ
of main paper) and with the same direction as ∆B̂. ddŶB̂ dY
in each x and y coordinate.
p
References
[1] S.-E. Wei, V. Ramakrishna, T. Kanade, and Y. Sheikh. Convolutional pose machines. arXiv preprint arXiv:1602.00134, 2016. 1
2
| 1cs.CV
|
Neural Belief Tracker: Data-Driven Dialogue State Tracking
Nikola Mrkšić1 , Diarmuid Ó Séaghdha2
Tsung-Hsien Wen1 , Blaise Thomson2 , Steve Young1
1
University of Cambridge
2
Apple Inc.
{nm480, thw28, sjy}@cam.ac.uk
{doseaghdha, blaisethom}@apple.com
arXiv:1606.03777v2 [cs.CL] 21 Apr 2017
Abstract
One of the core components of modern spoken dialogue systems is the belief tracker,
which estimates the user’s goal at every
step of the dialogue. However, most current
approaches have difficulty scaling to larger,
more complex dialogue domains. This is
due to their dependency on either: a) Spoken Language Understanding models that
require large amounts of annotated training
data; or b) hand-crafted lexicons for capturing some of the linguistic variation in users’
language. We propose a novel Neural Belief Tracking (NBT) framework which overcomes these problems by building on recent advances in representation learning.
NBT models reason over pre-trained word
vectors, learning to compose them into distributed representations of user utterances
and dialogue context. Our evaluation on
two datasets shows that this approach surpasses past limitations, matching the performance of state-of-the-art models which
rely on hand-crafted semantic lexicons and
outperforming them when such lexicons
are not provided.
1
Introduction
Spoken dialogue systems (SDS) allow users to interact with computer applications through conversation. Task-based systems help users achieve goals
such as finding restaurants or booking flights. The
dialogue state tracking (DST) component of an
SDS serves to interpret user input and update the
belief state, which is the system’s internal representation of the state of the conversation (Young
et al., 2010). This is a probability distribution over
dialogue states used by the downstream dialogue
manager to decide which action the system should
User: I’m looking for a cheaper restaurant
inform(price=cheap)
System: Sure. What kind - and where?
User: Thai food, somewhere downtown
inform(price=cheap, food=Thai,
area=centre)
System: The House serves cheap Thai food
User: Where is it?
inform(price=cheap, food=Thai,
area=centre); request(address)
System: The House is at 106 Regent Street
Figure 1: Annotated dialogue states in a sample dialogue. Underlined words show rephrasings which
are typically handled using semantic dictionaries.
perform next (Su et al., 2016a,b); the system action
is then verbalised by the natural language generator
(Wen et al., 2015a,b; Dušek and Jurčı́ček, 2015).
The Dialogue State Tracking Challenge (DSTC)
series of shared tasks has provided a common evaluation framework accompanied by labelled datasets
(Williams et al., 2016). In this framework, the dialogue system is supported by a domain ontology
which describes the range of user intents the system can process. The ontology defines a collection
of slots and the values that each slot can take. The
system must track the search constraints expressed
by users (goals or informable slots) and questions
the users ask about search results (requests), taking into account each user utterance (input via a
speech recogniser) and the dialogue context (e.g.,
what the system just said). The example in Figure
1 shows the true state after each user utterance in
a three-turn conversation. As can be seen in this
example, DST models depend on identifying mentions of ontology items in user utterances. This
becomes a non-trivial task when confronted with
lexical variation, the dynamics of context and noisy
automated speech recognition (ASR) output.
F OOD =C HEAP : [affordable, budget, low-cost,
low-priced, inexpensive, cheaper, economic, ...]
R ATING =H IGH : [best, high-rated, highly rated,
top-rated, cool, chic, popular, trendy, ...]
A REA =C ENTRE : [center, downtown, central,
city centre, midtown, town centre, ...]
Figure 2: An example semantic dictionary with
rephrasings for three ontology values in a restaurant search domain.
Traditional statistical approaches use separate
Spoken Language Understanding (SLU) modules
to address lexical variability within a single dialogue turn. However, training such models requires
substantial amounts of domain-specific annotation.
Alternatively, turn-level SLU and cross-turn DST
can be coalesced into a single model to achieve
superior belief tracking performance, as shown by
Henderson et al. (2014d). Such coupled models
typically rely on manually constructed semantic
dictionaries to identify alternative mentions of ontology items that vary lexically or morphologically.
Figure 2 gives an example of such a dictionary
for three slot-value pairs. This approach, which
we term delexicalisation, is clearly not scalable
to larger, more complex dialogue domains. Importantly, the focus on English in DST research
understates the considerable challenges that morphology poses to systems based on exact matching
in morphologically richer languages such as Italian
or German (see Vulić et al. (2017)).
In this paper, we present two new models, collectively called the Neural Belief Tracker (NBT)
family. The proposed models couple SLU and DST,
efficiently learning to handle variation without requiring any hand-crafted resources. To do that,
NBT models move away from exact matching and
instead reason entirely over pre-trained word vectors. The vectors making up the user utterance and
preceding system output are first composed into intermediate representations. These representations
are then used to decide which of the ontologydefined intents have been expressed by the user
up to that point in the conversation.
To the best of our knowledge, NBT models are
the first to successfully use pre-trained word vector
spaces to improve the language understanding capability of belief tracking models. In evaluation on
two datasets, we show that: a) NBT models match
the performance of delexicalisation-based models
which make use of hand-crafted semantic lexicons;
and b) the NBT models significantly outperform
those models when such resources are not available. Consequently, we believe this work proposes
a framework better-suited to scaling belief tracking
models for deployment in real-world dialogue systems operating over sophisticated application domains where the creation of such domain-specific
lexicons would be infeasible.
2
Background
Models for probabilistic dialogue state tracking, or
belief tracking, were introduced as components of
spoken dialogue systems in order to better handle
noisy speech recognition and other sources of uncertainty in understanding a user’s goals (Bohus
and Rudnicky, 2006; Williams and Young, 2007;
Young et al., 2010). Modern dialogue management
policies can learn to use a tracker’s distribution
over intents to decide whether to execute an action
or request clarification from the user. As mentioned above, the DSTC shared tasks have spurred
research on this problem and established a standard
evaluation paradigm (Williams et al., 2013; Henderson et al., 2014b,a). In this setting, the task is
defined by an ontology that enumerates the goals a
user can specify and the attributes of entities that
the user can request information about. Many different belief tracking models have been proposed
in the literature, from generative (Thomson and
Young, 2010) and discriminative (Henderson et al.,
2014d) statistical models to rule-based systems
(Wang and Lemon, 2013). To motivate the work
presented here, we categorise prior research according to their reliance (or otherwise) on a separate
SLU module for interpreting user utterances:1
Separate SLU Traditional SDS pipelines use
Spoken Language Understanding (SLU) decoders
to detect slot-value pairs expressed in the Automatic Speech Recognition (ASR) output. The
downstream DST model then combines this information with the past dialogue context to update the belief state (Thomson and Young, 2010;
Wang and Lemon, 2013; Lee and Kim, 2016; Perez,
2016; Perez and Liu, 2017; Sun et al., 2016; Jang
et al., 2016; Shi et al., 2016; Dernoncourt et al.,
2016; Liu and Perez, 2017; Vodolán et al., 2017).
1
The best-performing models in DSTC2 all used both raw
ASR output and the output of (potentially more than one) SLU
decoders (Williams, 2014; Williams et al., 2016). This does
not mean that those models are immune to the drawbacks
identified here for the two model categories; in fact, they share
the drawbacks of both.
Figure 3: Architecture of the NBT Model. The implementation of the three representation learning
subcomponents can be modified, as long as these produce adequate vector representations which the
downstream model components can use to decide whether the current candidate slot-value pair was
expressed in the user utterance (taking into account the preceding system act).
In the DSTC challenges, some systems used the
output of template-based matching systems such
as Phoenix (Wang, 1994). However, more robust
and accurate statistical SLU systems are available.
Many discriminative approaches to spoken dialogue SLU train independent binary models that
decide whether each slot-value pair was expressed
in the user utterance. Given enough data, these
models can learn which lexical features are good
indicators for a given value and can capture elements of paraphrasing (Mairesse et al., 2009). This
line of work later shifted focus to robust handling of
rich ASR output (Henderson et al., 2012; Tur et al.,
2013). SLU has also been treated as a sequence
labelling problem, where each word in an utterance
is labelled according to its role in the user’s intent;
standard labelling models such as CRFs or Recurrent Neural Networks can then be used (Raymond
and Ricardi, 2007; Yao et al., 2014; Celikyilmaz
and Hakkani-Tur, 2015; Mesnil et al., 2015; Peng
et al., 2015; Zhang and Wang, 2016; Liu and Lane,
2016b; Vu et al., 2016; Liu and Lane, 2016a, i.a.).
Other approaches adopt a more complex modelling
structure inspired by semantic parsing (Saleh et al.,
2014; Vlachos and Clark, 2014). One drawback
shared by these methods is their resource requirements, either because they need to learn independent parameters for each slot and value or because
they need fine-grained manual annotation at the
word level. This hinders scaling to larger, more
realistic application domains.
Joint SLU/DST Research on belief tracking has
found it advantageous to reason about SLU and
DST jointly, taking ASR predictions as input and
generating belief states as output (Henderson et al.,
2014d; Sun et al., 2014; Zilka and Jurcicek, 2015;
Mrkšić et al., 2015). In DSTC2, systems which
used no external SLU module outperformed all systems that only used external SLU features. Joint
models typically rely on a strategy known as delexicalisation whereby slots and values mentioned in
the text are replaced with generic labels. Once
the dataset is transformed in this manner, one can
extract a collection of template-like n-gram features such as [want tagged-value food]. To perform belief tracking, the shared model iterates over
all slot-value pairs, extracting delexicalised feature
vectors and making a separate binary decision regarding each pair. Delexicalisation introduces a
hidden dependency that is rarely discussed: how
do we identify slot/value mentions in text? For
toy domains, one can manually construct semantic
dictionaries which list the potential rephrasings for
all slot values. As shown by Mrkšić et al. (2016),
the use of such dictionaries is essential for the performance of current delexicalisation-based models.
Again though, this will not scale to the rich variety
of user language or to general domains.
The primary motivation for the work presented
in this paper is to overcome the limitations that
affect previous belief tracking models. The
NBT model efficiently learns from the avail-
able data by: 1) leveraging semantic information from pre-trained word vectors to resolve lexical/morphological ambiguity; 2) maximising the
number of parameters shared across ontology values; and 3) having the flexibility to learn domainspecific paraphrasings and other kinds of variation
that make it infeasible to rely on exact matching
and delexicalisation as a robust strategy.
3
Neural Belief Tracker
The Neural Belief Tracker (NBT) is a model designed to detect the slot-value pairs that make up
the user’s goal at a given turn during the flow of
dialogue. Its input consists of the system dialogue
acts preceding the user input, the user utterance
itself, and a single candidate slot-value pair that
it needs to make a decision about. For instance,
the model might have to decide whether the goal
FOOD =I TALIAN has been expressed in ‘I’m looking for good pizza’. To perform belief tracking, the
NBT model iterates over all candidate slot-value
pairs (defined by the ontology), and decides which
ones have just been expressed by the user.
Figure 3 presents the flow of information in the
model. The first layer in the NBT hierarchy performs representation learning given the three model
inputs, producing vector representations for the
user utterance (r), the current candidate slot-value
pair (c) and the system dialogue acts (tq , ts , tv ).
Subsequently, the learned vector representations
interact through the context modelling and semantic decoding submodules to obtain the intermediate
interaction summary vectors dr , dc and d. These
are used as input to the final decision-making module which decides whether the user expressed the
intent represented by the candidate slot-value pair.
3.1
Representation Learning
For any given user utterance, system act(s) and candidate slot-value pair, the representation learning
submodules produce vector representations which
act as input for the downstream components of
the model. All representation learning subcomponents make use of pre-trained collections of word
vectors. As shown by Mrkšić et al. (2016), specialising word vectors to express semantic similarity
rather than relatedness is essential for improving
belief tracking performance. For this reason, we
use the semantically-specialised Paragram-SL999
word vectors (Wieting et al., 2015) throughout this
work. The NBT training procedure keeps these
vectors fixed: that way, at test time, unseen words
semantically related to familiar slot values (i.e. inexpensive to cheap) will be recognised purely by
their position in the original vector space (see also
Rocktäschel et al. (2016)). This means that the
NBT model parameters can be shared across all
values of the given slot, or even across all slots.
Let u represent a user utterance consisting of
ku words u1 , u2 , . . . , uku . Each word has an associated word vector u1 , . . . , uku . We propose two
model variants which differ in the method used to
produce vector representations of u: NBT-DNN
and NBT-CNN. Both act over the constituent ngrams of the utterance. Let vin be the concatenation
of the n word vectors starting at index i, so that:
vin = ui ⊕ . . . ⊕ ui+n−1
(1)
where ⊕ denotes vector concatenation. The simpler
of our two models, which we term NBT-DNN, is
shown in Figure 4. This model computes cumulative n-gram representation vectors r1 , r2 and r3 ,
which are the n-gram ‘summaries’ of the unigrams,
bigrams and trigrams in the user utterance:
rn =
kuX
−n+1
vin
(2)
i=1
Each of these vectors is then non-linearly mapped
to intermediate representations of the same size:
r0n = σ(Wns rn + bsn )
(3)
where the weight matrices and bias terms map the
cumulative n-grams to vectors of the same dimensionality and σ denotes the sigmoid activation function. We maintain a separate set of parameters for
each slot (indicated by superscript s). The three
vectors are then summed to obtain a single representation for the user utterance:
r = r01 + r02 + r03
(4)
The cumulative n-gram representations used by
this model are just unweighted sums of all word
vectors in the utterance. Ideally, the model should
learn to recognise which parts of the utterance are
more relevant for the subsequent classification task.
For instance, it could learn to ignore verbs or stop
words and pay more attention to adjectives and
nouns which are more likely to express slot values.
Figure 4: NBT-DNN M ODEL. Word vectors of n-grams (n = 1, 2, 3) are summed to obtain cumulative
n-grams, then passed through another hidden layer and summed to obtain the utterance representation r.
Figure 5: NBT-CNN Model. L convolutional filters of window sizes 1, 2, 3 are applied to word vectors
of the given utterance (L = 3 in the diagram, but L = 300 in the system). The convolutions are followed
by the ReLU activation function and max-pooling to produce summary n-gram representations. These are
summed to obtain the utterance representation r.
NBT-CNN Our second model draws inspiration
from successful applications of Convolutional Neural Networks (CNNs) for language understanding
(Collobert et al., 2011; Kalchbrenner et al., 2014;
Kim, 2014). These models typically apply a number of convolutional filters to n-grams in the input
sentence, followed by non-linear activation functions and max-pooling. Following this approach,
the NBT-CNN model applies L = 300 different filters for n-gram lengths of 1, 2 and 3 (Figure 5). Let Fns ∈ RL×nD denote the collection
of filters for each value of n, where D = 300 is
the word vector dimensionality. If vin denotes the
concatenation of n word vectors starting at index
i, let mn = [v1n ; v2n ; . . . ; vknu −n+1 ] be the list of
n-grams that convolutional filters of length n run
over. The three intermediate representations are
then given by:
Rn = Fns mn
(5)
Each column of the intermediate matrices Rn is
produced by a single convolutional filter of length
n. We obtain summary n-gram representations
by pushing these representations through a recti-
fied linear unit (ReLU) activation function (Nair
and Hinton, 2010) and max-pooling over time
(i.e. columns of the matrix) to get a single feature
for each of the L filters applied to the utterance:
r0n = maxpool (ReLU (Rn + bsn ))
(6)
where bsn is a bias term broadcast across all filters.
Finally, the three summary n-gram representations
are summed to obtain the final utterance representation vector r (as in Equation 4). The NBT-CNN
model is (by design) better suited to longer utterances, as its convolutional filters interact directly
with subsequences of the utterance, and not just
their noisy summaries given by the NBT-DNN’s
cumulative n-grams.
3.2
Semantic Decoding
The NBT diagram in Figure 3 shows that the utterance representation r and the candidate slotvalue pair representation c directly interact through
the semantic decoding module. This component decides whether the user explicitly expressed
an intent matching the current candidate pair
(i.e. without taking the dialogue context into account). Examples of such matches would be ‘I
want Thai food’ with food=Thai or more demanding ones such as ‘a pricey restaurant’ with
price=expensive. This is where the use of
high-quality pre-trained word vectors comes into
play: a delexicalisation-based model could deal
with the former example but would be helpless in
the latter case, unless a human expert had provided
a semantic dictionary listing all potential rephrasings for each value in the domain ontology.
Let the vector space representations of a candidate pair’s slot name and value be given by cs and
cv (with vectors of multi-word slot names/values
summed together). The NBT model learns to map
this tuple into a single vector c of the same dimensionality as the utterance representation r. These
two representations are then forced to interact in order to learn a similarity metric which discriminates
between interactions of utterances with slot-value
pairs that they either do or do not express:
c = σ Wcs (cs + cv ) + bsc
(7)
d=r⊗c
(8)
where ⊗ denotes element-wise vector multiplication. The dot product, which may seem like the
more intuitive similarity metric, would reduce the
rich set of features in d to a single scalar. The
element-wise multiplication allows the downstream
network to make better use of its parameters by
learning non-linear interactions between sets of
features in r and c.2
3.3
Context Modelling
This ‘decoder’ does not yet suffice to extract intents
from utterances in human-machine dialogue. To
understand some queries, the belief tracker must be
aware of context, i.e. the flow of dialogue leading
up to the latest user utterance. While all previous
system and user utterances are important, the most
relevant one is the last system utterance, in which
the dialogue system could have performed (among
others) one of the following two system acts:
1. System Request: The system asks the user
about the value of a specific slot Tq . If the
system utterance is: ‘what price range would
2
We also tried to concatenate r and c and pass that vector
to the downstream decision-making neural network. However,
this set-up led to very weak performance since our relatively
small datasets did not suffice for the network to learn to model
the interaction between the two feature vectors.
you like?’ and the user answers with any, the
model must infer the reference to price range,
and not to other slots such as area or food.
2. System Confirm: The system asks the user
to confirm whether a specific slot-value pair
(Ts , Tv ) is part of their desired constraints. For
example, if the user responds to ‘how about
Turkish food?’ with ‘yes’, the model must be
aware of the system act in order to correctly
update the belief state.
If we make the Markovian decision to only consider the last set of system acts, we can incorporate
context modelling into the NBT. Let tq and (ts , tv )
be the word vectors of the arguments for the system request and confirm acts (zero vectors if none).
The model computes the following measures of
similarity between the system acts, candidate pair
(cs , cv ) and utterance representation r:
mr = (cs · tq )r
mc = (cs · ts )(cv · tv )r
(9)
(10)
where · denotes dot product. The computed similarity terms act as gating mechanisms which only pass
the utterance representation through if the system
asked about the current candidate slot or slot-value
pair. This type of interaction is particularly useful
for the confirm system act: if the system asks the
user to confirm, the user is likely not to mention any
slot values, but to just respond affirmatively or negatively. This means that the model must consider
the three-way interaction between the utterance,
candidate slot-value pair and the slot value pair offered by the system. If (and only if) the latter two
are the same should the model consider the affirmative or negative polarity of the user utterance when
making the subsequent binary decision.
Binary Decision Maker The intermediate representations are passed through another hidden layer
and then combined. If φdim (x) = σ(W x + b) is a
layer which maps input vector x to a vector of size
dim, the input to the final binary softmax (which
represents the decision) is given by:
y = φ2 φ100 (d) + φ100 (mr ) + φ100 (mc )
4
Belief State Update Mechanism
In spoken dialogue systems, belief tracking models
operate over the output of automatic speech recognition (ASR). Despite improvements to speech
recognition, the need to make the most out of imperfect ASR will persist as dialogue systems are
used in increasingly noisy environments.
In this work, we define a simple rule-based
belief state update mechanism which can be
applied to ASR N -best lists. For dialogue turn t,
let syst−1 denote the preceding system output, and
let ht denote the list of N ASR hypotheses hti with
posterior probabilities pti . For any hypothesis hti ,
slot s and slot value v ∈ Vs , NBT models estimate
P(s, v | hti , syst−1 ), which is the (turn-level)
probability that (s, v) was expressed in the given
hypothesis. The predictions for N such hypotheses
are then combined as:
P(s, v | ht , syst−1 ) =
N
X
pti P s, v | hti , syst
i=1
This turn-level belief state estimate is then combined with the (cumulative) belief state up to time
(t − 1) to get the updated belief state estimate:
P(s, v | h1:t , sys1:t−1 ) = λ P s, v | ht , syst−1
+ (1 − λ) P s, v | h1:t−1 , sys1:t−2
where λ is the coefficient which determines the
relative weight of the turn-level and previous turns’
belief state estimates.3 For slot s, the set of its
detected values at turn t is then given by:
Vst = {v ∈ Vs | P s, v | h1:t , sys1:t−1 ≥ 0.5}
For informable (i.e. goal-tracking) slots, the value
in Vst with the highest probability is chosen as the
current goal (if Vst 6= {∅}). For requests, all slots
t are deemed to have been requested. As
in Vreq
requestable slots serve to model single-turn user
queries, they require no belief tracking across turns.
5
5.1
Experiments
Datasets
Two datasets were used for training and evaluation. Both consist of user conversations with taskoriented dialogue systems designed to help users
find suitable restaurants around Cambridge, UK.
The two corpora share the same domain ontology,
which contains three informable (i.e. goal-tracking)
slots: FOOD, AREA and PRICE. The users can specify values for these slots in order to find restaurants
3
This coefficient was tuned on the DSTC2 development
set. The best performance was achieved with λ = 0.55.
which best meet their criteria. Once the system suggests a restaurant, the users can ask about the values
of up to eight requestable slots (PHONE NUMBER ,
ADDRESS , etc.). The two datasets are:
1. DSTC2: We use the transcriptions, ASR hypotheses and turn-level semantic labels provided for the Dialogue State Tracking Challenge 2 (Henderson et al., 2014a). The official transcriptions contain various spelling
errors which we corrected manually; the
cleaned version of the dataset is available at mi.eng.cam.ac.uk/˜nm480/
dstc2-clean.zip. The training data contains 2207 dialogues and the test set consists
of 1117 dialogues. We train NBT models on
transcriptions but report belief tracking performance on test set ASR hypotheses provided
in the original challenge.
2. WOZ 2.0: Wen et al. (2017) performed a Wizard of Oz style experiment in which Amazon
Mechanical Turk users assumed the role of
the system or the user of a task-oriented dialogue system based on the DSTC2 ontology.
Users typed instead of using speech, which
means performance in the WOZ experiments
is more indicative of the model’s capacity for
semantic understanding than its robustness to
ASR errors. Whereas in the DSTC2 dialogues
users would quickly adapt to the system’s
(lack of) language understanding capability,
the WOZ experimental design gave them freedom to use more sophisticated language. We
expanded the original WOZ dataset from Wen
et al. (2017) using the same data collection
procedure, yielding a total of 1200 dialogues.
We divided these into 600 training, 200 validation and 400 test set dialogues. The WOZ
2.0 dataset is available at mi.eng.cam.ac.
uk/˜nm480/woz_2.0.zip.
Training Examples The two corpora are used to
create training data for two separate experiments.
For each dataset, we iterate over all train set utterances, generating one example for each of the slotvalue pairs in the ontology. An example consists
of a transcription, its context (i.e. list of preceding
system acts) and a candidate slot-value pair. The
binary label for each example indicates whether or
not its utterance and context express the example’s
candidate pair. For instance, ‘I would like Irish
food’ would generate a positive example for candidate pair FOOD =I RISH, and a negative example for
every other slot-value pair in the ontology.
but over delexicalised features akin to those
used by Henderson et al. (2014c).
2. The same baseline model supplemented with
a task-specific semantic dictionary (produced
by the baseline system creators). The two
dictionaries are available at mi.eng.cam.
ac.uk/˜nm480/sem-dict.zip. The
DSTC2 dictionary contains only three rephrasings. Nonetheless, the use of these rephrasings translates to substantial gains in DST performance (see Sect. 6.1). We believe this
result supports our claim that the vocabulary used by Mechanical Turkers in DSTC2
was constrained by the system’s inability to
cope with lexical variation and ASR noise.
The WOZ dictionary includes 38 rephrasings,
showing that the unconstrained language used
by Mechanical Turkers in the Wizard-of-Oz
setup requires more elaborate lexicons.
Evaluation We focus on two key evaluation metrics introduced in (Henderson et al., 2014a):
1. Goals (‘joint goal accuracy’): the proportion
of dialogue turns where all the user’s search
goal constraints were correctly identified;
2. Requests: similarly, the proportion of dialogue turns where user’s requests for information were identified correctly.
5.2
Models
We evaluate two NBT model variants: NBT-DNN
and NBT-CNN. To train the models, we use the
Adam optimizer (Kingma and Ba, 2015) with crossentropy loss, backpropagating through all the NBT
subcomponents while keeping the pre-trained word
vectors fixed (in order to allow the model to deal
with unseen words at test time). The model is
trained separately for each slot. Due to the high
class bias (most of the constructed examples are
negative), we incorporate a fixed number of positive examples in each mini-batch.4
Baseline Models For each of the two datasets,
we compare the NBT models to:
1. A baseline system that implements a wellknown competitive delexicalisation-based
model for that dataset. For DSTC2, the model
is that of Henderson et al. (2014c; 2014d).
This model is an n-gram based neural network model with recurrent connections between turns (but not inside utterances) which
replaces occurrences of slot names and values with generic delexicalised features. For
WOZ 2.0, we compare the NBT models to a
more sophisticated belief tracking model presented in (Wen et al., 2017). This model uses
an RNN for belief state updates and a CNN
for turn-level feature extraction. Unlike NBTCNN, their CNN operates not over vectors,
4
Model hyperparameters were tuned on the respective validation sets. For both datasets, the initial Adam learning rate
was set to 0.001, and 18 th of positive examples were included
in each mini-batch. The batch size did not affect performance:
it was set to 256 in all experiments. Gradient clipping (to
[−2.0, 2.0]) was used to handle exploding gradients. Dropout
(Srivastava et al., 2014) was used for regularisation (with 50%
dropout rate on all intermediate representations). Both NBT
models were implemented in TensorFlow (Abadi et al., 2015).
Both baseline models map exact matches
of ontology-defined intents (and their lexiconspecified rephrasings) to one-hot delexicalised ngram features. This means that pre-trained vectors
cannot be incorporated directly into these models.
6
6.1
Results
Belief Tracking Performance
Table 1 shows the performance of NBT models
trained and evaluated on DSTC2 and WOZ 2.0
datasets. The NBT models outperformed the baseline models in terms of both joint goal and request
accuracies. For goals, the gains are always statistically significant (paired t-test, p < 0.05). Moreover, there was no statistically significant variation
between the NBT and the lexicon-supplemented
models, showing that the NBT can handle semantic relations which otherwise had to be explicitly
encoded in semantic dictionaries.
While the NBT performs well across the board,
we can compare its performance on the two datasets
to understand its strengths. The improvement over
the baseline is greater on WOZ 2.0, which corroborates our intuition that the NBT’s ability to
learn linguistic variation is vital for this dataset
containing longer sentences, richer vocabulary and
no ASR errors. By comparison, the language of
the subjects in the DSTC2 dataset is less rich, and
compensating for ASR errors is the main hurdle:
given access to the DSTC2 test set transcriptions,
the NBT models’ goal accuracy rises to 0.96. This
DST Model
Delexicalisation-Based Model
Delexicalisation-Based Model + Semantic Dictionary
N EURAL B ELIEF T RACKER : NBT-DNN
N EURAL B ELIEF T RACKER : NBT-CNN
DSTC2
Goals Requests
69.1
95.7
72.9*
95.7
72.6*
96.4
73.4*
96.5
WOZ 2.0
Goals Requests
70.8
87.1
83.7*
87.6
84.4*
91.2*
84.2*
91.6*
Table 1: DSTC2 and WOZ 2.0 test set accuracies for: a) joint goals; and b) turn-level requests. The
asterisk indicates statistically significant improvement over the baseline trackers (paired t-test; p < 0.05).
indicates that future work should focus on better
ASR compensation if the model is to be deployed
in environments with challenging acoustics.
6.2
Word Vectors
(No Info.)
GloVe
Paragram-SL999
XAVIER
DSTC2
Goals Requests
64.2
81.2
69.0*
96.4*
73.4*
96.5*
WOZ 2.0
Goals Requests
81.2
90.7
80.1
91.4
84.2*
91.6
The Importance of Word Vector Spaces
The NBT models use the semantic relations embedded in the pre-trained word vectors to handle
semantic variation and produce high-quality intermediate representations. Table 2 shows the performance of NBT-CNN5 models making use of
three different word vector collections: 1) ‘random’
word vectors initialised using the XAVIER initialisation (Glorot and Bengio, 2010); 2) distributional
GloVe vectors (Pennington et al., 2014), trained
using co-occurrence information in large textual
corpora; and 3) semantically specialised ParagramSL999 vectors (Wieting et al., 2015), which are obtained by injecting semantic similarity constraints
from the Paraphrase Database (Ganitkevitch et al.,
2013) into the distributional GloVe vectors in order
to improve their semantic content.
The results in Table 2 show that the use of semantically specialised word vectors leads to considerable performance gains: Paragram-SL999 vectors
(significantly) outperformed GloVe and XAVIER
vectors for goal tracking on both datasets. The
gains are particularly robust for noisy DSTC2
data, where both collections of pre-trained vectors consistently outperformed random initialisation. The gains are weaker for the noise-free WOZ
2.0 dataset, which seems to be large (and clean)
enough for the NBT model to learn task-specific
rephrasings and compensate for the lack of semantic content in the word vectors. For this dataset,
GloVe vectors do not improve over the randomly
initialised ones. We believe this happens because
distributional models keep related, yet antonymous
words close together (e.g. north and south, expensive and inexpensive), offsetting the useful semantic content embedded in this vector spaces.
5
The NBT-DNN model showed the same trends. For
brevity, Table 2 presents only the NBT-CNN figures.
Table 2: DSTC2 and WOZ 2.0 test set performance
(joint goals and requests) of the NBT-CNN model
making use of three different word vector collections. The asterisk indicates statistically significant
improvement over the baseline XAVIER (random)
word vectors (paired t-test; p < 0.05).
7
Conclusion
In this paper, we have proposed a novel neural
belief tracking (NBT) framework designed to overcome current obstacles to deploying dialogue systems in real-world dialogue domains. The NBT
models offer the known advantages of coupling
Spoken Language Understanding and Dialogue
State Tracking, without relying on hand-crafted
semantic lexicons to achieve state-of-the-art performance. Our evaluation demonstrated these benefits:
the NBT models match the performance of models
which make use of such lexicons and vastly outperform them when these are not available. Finally, we
have shown that the performance of NBT models
improves with the semantic quality of the underlying word vectors. To the best of our knowledge,
we are the first to move past intrinsic evaluation
and show that semantic specialisation boosts performance in downstream tasks.
In future work, we intend to explore applications
of the NBT for multi-domain dialogue systems, as
well as in languages other than English that require
handling of complex morphological variation.
Acknowledgements
The authors would like to thank Ivan Vulić, Ulrich
Paquet, the Cambridge Dialogue Systems Group
and the anonymous ACL reviewers for their constructive feedback and helpful discussions.
References
Martı́n Abadi, Ashish Agarwal, Paul Barham, Eugene
Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin,
Sanjay Ghemawat, Ian Goodfellow, Andrew Harp,
Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal
Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh
Levenberg, Dan Mané, Rajat Monga, Sherry Moore,
Derek Murray, Chris Olah, Mike Schuster, Jonathon
Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar,
Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan,
Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. 2015. TensorFlow: Large-scale machine learning on heterogeneous systems.
Dan Bohus and Alex Rudnicky. 2006. A “k hypotheses + other” belief updating model. In Proceedings
of the AAAI Workshop on Statistical and Empirical
Methods in Spoken Dialogue Systems.
Asli Celikyilmaz and Dilek Hakkani-Tur. 2015. Convolutional Neural Network Based Semantic Tagging
with Entity Embeddings. In Proceedings of NIPS
Workshop on Machine Learning for Spoken Language Understanding and Interaction.
Ronan Collobert, Jason Weston, Leon Bottou, Michael
Karlen, Koray Kavukcuoglu, and Pavel Kuksa. 2011.
Natural language processing (almost) from scratch.
Journal of Machine Learning Research 12:2493–
2537.
Franck Dernoncourt, Ji Young Lee, Trung H. Bui, and
Hung H. Bui. 2016. Robust dialog state tracking for
large ontologies. In Proceedings of IWSDS.
Ondřej Dušek and Filip Jurčı́ček. 2015. Training a Natural Language Generator From Unaligned Data. In
Proceedings of ACL.
Juri Ganitkevitch, Benjamin Van Durme, and Chris
Callison-Burch. 2013.
PPDB: The Paraphrase
Database. In Proceedings of NAACL HLT.
Xavier Glorot and Yoshua Bengio. 2010. Understanding the difficulty of training deep feedforward neural
networks. In Proceedings of AISTATS.
Matthew Henderson, Milica Gašić, Blaise Thomson,
Pirros Tsiakoulis, Kai Yu, and Steve Young. 2012.
Discriminative Spoken Language Understanding Using Word Confusion Networks. In Spoken Language
Technology Workshop, 2012. IEEE.
Matthew Henderson, Blaise Thomson, and Jason D.
Wiliams. 2014a. The Second Dialog State Tracking
Challenge. In Proceedings of SIGDIAL.
Matthew Henderson, Blaise Thomson, and Jason D.
Wiliams. 2014b. The Third Dialog State Tracking
Challenge. In Proceedings of IEEE SLT.
Matthew Henderson, Blaise Thomson, and Steve
Young. 2014c. Robust Dialog State Tracking using
Delexicalised Recurrent Neural Networks and Unsupervised Adaptation. In Proceedings of IEEE SLT.
Matthew Henderson, Blaise Thomson, and Steve
Young. 2014d. Word-Based Dialog State Tracking
with Recurrent Neural Networks. In Proceedings of
SIGDIAL.
Youngsoo Jang, Jiyeon Ham, Byung-Jun Lee, Youngjae Chang, and Kee-Eung Kim. 2016. Neural dialog
state tracker for large ontologies by attention mechanism. In Proceedings of IEEE SLT.
Nal Kalchbrenner, Edward Grefenstette, and Phil Blunsom. 2014. A Convolutional Neural Network for
Modelling Sentences. In Proceedings of ACL.
Yoon Kim. 2014. Convolutional Neural Networks for
Sentence Classification. In Proceedings of EMNLP.
Diederik P. Kingma and Jimmy Ba. 2015. Adam: A
Method for Stochastic Optimization. In Proceedings of ICLR.
Byung-Jun Lee and Kee-Eung Kim. 2016. Dialog History Construction with Long-Short Term Memory
for Robust Generative Dialog State Tracking. Dialogue & Discourse 7(3):47–64.
Bing Liu and Ian Lane. 2016a. Attention-Based Recurrent Neural Network Models for Joint Intent Detection and Slot Filling. In Proceedings of Interspeech.
Bing Liu and Ian Lane. 2016b. Joint Online Spoken
Language Understanding and Language Modeling
with Recurrent Neural Networks. In Proceedings of
SIGDIAL.
Fei Liu and Julien Perez. 2017. Gated End-to-End
Memory Networks. In Proceedings of EACL.
F. Mairesse, M. Gasic, F. Jurcicek, S. Keizer, B. Thomson, K. Yu, and S. Young. 2009. Spoken Language
Understanding from Unaligned Data using Discriminative Classification Models. In Proceedings of
ICASSP.
Grégoire Mesnil, Yann Dauphin, Kaisheng Yao,
Yoshua Bengio, Li Deng, Dilek Hakkani-Tur, Xiaodong He, Larry Heck, Dong Yu, and Geoffrey
Zweig. 2015. Using recurrent neural networks
for slot filling in spoken language understanding.
IEEE/ACM Transactions on Audio, Speech, and Language Processing 23(3):530–539.
Nikola Mrkšić, Diarmuid Ó Séaghdha, Blaise Thomson, Milica Gašić, Lina Rojas-Barahona, Pei-Hao
Su, David Vandyke, Tsung-Hsien Wen, and Steve
Young. 2016. Counter-fitting Word Vectors to Linguistic Constraints. In Proceedings of HLT-NAACL.
Nikola Mrkšić, Diarmuid Ó Séaghdha, Blaise Thomson, Milica Gašić, Pei-Hao Su, David Vandyke,
Tsung-Hsien Wen, and Steve Young. 2015. Multidomain Dialog State Tracking using Recurrent Neural Networks. In Proceedings of ACL.
Vinod Nair and Geoffrey E. Hinton. 2010. Rectified
linear units improve restricted Boltzmann machines.
In Proceedings of ICML.
Kai Sun, Qizhe Xie, and Kai Yu. 2016. Recurrent Polynomial Network for Dialogue State Tracking. Dialogue & Discourse 7(3):65–88.
Baolin Peng, Kaisheng Yao, Li Jing, and Kam-Fai
Wong. 2015. Recurrent Neural Networks with External Memory for Language Understanding. In Proceedings of the National CCF Conference on Natural Language Processing and Chinese Computing.
Blaise Thomson and Steve Young. 2010. Bayesian update of dialogue state: A POMDP framework for
spoken dialogue systems. Computer Speech and
Language .
Jeffrey Pennington, Richard Socher, and Christopher
Manning. 2014. Glove: Global Vectors for Word
Representation. In Proceedings of EMNLP.
Gokhan Tur, Anoop Deoras, and Dilek Hakkani-Tur.
2013. Semantic Parsing Using Word Confusion Networks With Conditional Random Fields. In Proceedings of Interspeech.
Julien Perez. 2016. Spectral decomposition method of
dialog state tracking via collective matrix factorization. Dialogue & Discourse 7(3):34–46.
Andreas Vlachos and Stephen Clark. 2014. A new corpus and imitation learning framework for contextdependent semantic parsing. TACL 2:547–559.
Julien Perez and Fei Liu. 2017. Dialog state tracking, a
machine reading approach using Memory Network.
In Proceedings of EACL.
Miroslav Vodolán, Rudolf Kadlec, and Jan Kleindienst.
2017. Hybrid Dialog State Tracker with ASR Features. In Proceedings of EACL.
Christian Raymond and Giuseppe Ricardi. 2007. Generative and discriminative algorithms for spoken
language understanding. In Proceedings of Interspeech.
Ngoc Thang Vu, Pankaj Gupta, Heike Adel, and Hinrich Schütze. 2016. Bi-directional recurrent neural
network with ranking loss for spoken language understanding. In Proceedings of ICASSP.
Tim Rocktäschel, Edward Grefenstette, Karl Moritz
Hermann, Tomas Kocisky, and Phil Blunsom. 2016.
Reasoning about entailment with neural attention.
In ICLR.
Iman Saleh, Shafiq Joty, Lluı́s Màrquez, Alessandro
Moschitti, Preslav Nakov, Scott Cyphers, and Jim
Glass. 2014. A study of using syntactic and semantic structures for concept segmentation and labeling.
In Proceedings of COLING.
Hongjie Shi, Takashi Ushio, Mitsuru Endo, Katsuyoshi
Yamagami, and Noriaki Horii. 2016. Convolutional
Neural Networks for Multi-topic Dialog State Tracking. In Proceedings of IWSDS.
Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky,
Ilya Sutskever, and Ruslan Salakhutdinov. 2014.
Dropout: A Simple Way to Prevent Neural Networks
from Overfitting. Journal of Machine Learning Research .
Pei-Hao Su, Milica Gašić, Nikola Mrkšić, Lina RojasBarahona, Stefan Ultes, David Vandyke, TsungHsien Wen, and Steve Young. 2016a. Continuously
learning neural dialogue management. In arXiv
preprint: 1606.02689.
Pei-Hao Su, Milica Gašić, Nikola Mrkšić, Lina RojasBarahona, Stefan Ultes, David Vandyke, TsungHsien Wen, and Steve Young. 2016b. On-line active
reward learning for policy optimisation in spoken dialogue systems. In Proceedings of ACL.
Kai Sun, Lu Chen, Su Zhu, and Kai Yu. 2014. The
SJTU System for Dialog State Tracking Challenge
2. In Proceedings of SIGDIAL.
Ivan Vulić, Nikola Mrkšić, Roi Reichart, Diarmuid Ó
Séaghdha, Steve Young, and Anna Korhonen. 2017.
Morph-fitting: Fine-tuning word vector spaces with
simple language-specific rules. In Proceedings of
ACL.
Wayne Wang. 1994. Extracting Information From
Spontaneous Speech. In Proceedings of Interspeech.
Zhuoran Wang and Oliver Lemon. 2013. A Simple and
Generic Belief Tracking Mechanism for the Dialog
State Tracking Challenge: On the believability of observed information. In Proceedings of SIGDIAL.
Tsung-Hsien Wen, Milica Gašić, Dongho Kim, Nikola
Mrkšić, Pei-Hao Su, David Vandyke, and Steve
Young. 2015a. Stochastic Language Generation
in Dialogue using Recurrent Neural Networks with
Convolutional Sentence Reranking. In Proceedings
of SIGDIAL.
Tsung-Hsien Wen, Milica Gašić, Nikola Mrkšić,
Pei-Hao Su, David Vandyke, and Steve Young.
2015b. Semantically Conditioned LSTM-based Natural Language Generation for Spoken Dialogue Systems. In Proceedings of EMNLP.
Tsung-Hsien Wen, David Vandyke, Nikola Mrkšić,
Milica Gašić, Lina M. Rojas-Barahona, Pei-Hao Su,
Stefan Ultes, and Steve Young. 2017. A networkbased end-to-end trainable task-oriented dialogue
system. In Proceedings of EACL.
John Wieting, Mohit Bansal, Kevin Gimpel, and Karen
Livescu. 2015. From paraphrase database to compositional paraphrase model and back. TACL 3:345–
358.
Jason D. Williams. 2014. Web-style ranking and SLU
combination for dialog state tracking. In Proceedings of SIGDIAL.
Jason D. Williams, Antoine Raux, and Matthew Henderson. 2016. The Dialog State Tracking Challenge
series: A review. Dialogue & Discourse 7(3):4–33.
Jason D. Williams, Antoine Raux, Deepak Ramachandran, and Alan W. Black. 2013. The Dialogue State
Tracking Challenge. In Proceedings of SIGDIAL.
Jason D. Williams and Steve Young. 2007. Partially
observable markov decision processes for spoken
dialog systems. Computer Speech and Language
21:393–422.
Kaisheng Yao, Baolin Peng, Yu Zhang, Dong Yu, Geoffrey Zweig, and Yangyang Shi. 2014. Spoken language understanding using long short-term memory
neural networks. In Proceedings of ASRU.
Steve Young, Milica Gašić, Simon Keizer, François
Mairesse, Jost Schatzmann, Blaise Thomson, and
Kai Yu. 2010. The hidden information state model:
A practical framework for POMDP-based spoken dialogue management. Computer Speech and Language 24:150–174.
Xiaodong Zhang and Houfeng Wang. 2016. A Joint
Model of Intent Determination and Slot Filling for
Spoken Language Understanding. In Proceedings
of IJCAI.
Lukas Zilka and Filip Jurcicek. 2015. Incremental
LSTM-based dialog state tracker. In Proceedings of
ASRU.
| 2cs.AI
|
A Capacity Scaling Law for Artificial Neural Networks
Gerald Friedland∗, Mario Michael Krell†
[email protected], [email protected]
September 15, 2017
arXiv:1708.06019v2 [cs.NE] 18 Sep 2017
Abstract
observed in the self-assembly of matter or the
Ising model of ferromagnetism. As a result, it is
impossible to make an artificial neural network
that is sensitive to, but not disrupted by, new
information once a certain threshold is reached.
Likewise, too many input dimensions for too few
data points will always result in overfitting and
the network will only behave as associative memory.
Our theoretical derivation, backed up by repeatable empirical evidence, shows the scaling
of the capacity of a neural network based on two
critical points, which we call lossless-memory
(LM) dimension and MacKay (MK) dimension,
respectively. The LM dimension defines the
point of guaranteed operation as memory and
the MK dimension defines the point of guaranteed forgetting (generalization), even for very
high dimensional networks. Both points scale
strictly linear with the number of weights, exemplified as follows.
For a binary classifier, consisting of h threshold perceptrons in the hidden layer with a bias, k
weighted inputs, and one binary output perceptron (a configuration commonly referred to as
3-layer multilayer perceptron (MLP)), we show
that when we classify n points of input in chaotic
position (see Section 4), the following behavior
will be observed.
By assuming an ideal neural network with gating functions handling the worst case data, we
derive the calculation of two critical numbers
predicting the behavior of perceptron networks.
First, we derive the calculation of what we
call the lossless memory (LM) dimension. The
LM dimension is a generalization of the Vapnik–Chervonenkis (VC) dimension that avoids
structured data and therefore provides an upper bound for perfectly fitting any training data.
Second, we derive what we call the MacKay
(MK) dimension. This limit indicates necessary
forgetting, that is, the lower limit for most generalization uses of the network. Our derivations
are performed by embedding the ideal network
into Shannon’s communication model which allows to interpret the two points as capacities
measured in bits. We validate our upper bounds
with repeatable experiments using different network configurations, diverse implementations,
varying activation functions, and several learning algorithms. The bottom line is that the
two capacity points scale strictly linear with the
number of weights. Among other practical applications, our result allows network implementations with gating functions (e. g., sigmoid or
rectified linear units) to be evaluated against our
upper limit independent of a concrete task.
1
• The LM dimension is bounded by
DLM = h(k + 1) + (h + 1)
Introduction
(1)
for k > 1. Having n ≤ DLM samples implies perfect labeling in an ideal network.
In this paper, we show that feed-forward neural networks, just like individual perceptrons
or hopfield networks, can be best understood
as associative memory and one way to understand challenges like overfitting and forgetting
is to measure the capacity of that memory. Furthermore, a closer look at the error as a function of capacity then reveals that perceptrons go
through two phase transitions, as indicated earlier by Wolfgang Kinzel [23], similar to the ones
• The MK dimension is bounded by
DM K = 2 × DLM
(2)
for k > 1. Having n ≥ DM K samples implies a ≥ 50% chance that a labeling cannot
be learned even with perfect training and an
ideal network.
• For any connected perceptron network, the
maximum capacity C (MK dimension) of
each perceptron is C = 2k and scales linearly in k, the number of weights per perceptron (including bias).
∗ University of California, Berkeley and Lawrence Livermore National Lab
† International Computer Science Institute, Berkeley
Both authors contributed equally to this paper.
1
Sender
dataset,
labeling
Encoder
Channel
Learning dataset,
Method weights
Identity
Decoder
dataset,
weights
Receiver
Neural dataset,
Network labeling'
Figure 1: Shannon’s communication model applied to labeling in machine learning. A dataset
consisting of n sample points and the ground truth labeling of n bits are sent to the neural network.
The learning method transfers it to a parametrization (i. e., network weights). The dataset is kept
as constant information in the whole process. The identity channel forwards everything without
loss. In the decoding step, the network then uses the weights together with the dataset to try to
reproduce the original labeling.
• Using an embedding into Shannon’s communication model (Fig. 1), the capacity can
be measured in bits and connected to other
information-theoretic methods.
leading to a variety of claims about the capabilities and limits of each of these algorithms
(see for example [51]). Even though artificial
neural networks have been popular for decades,
The article is structured as follows: Section 2 understanding of the processes underlying them
starts with the discussion of related work on is usually based solely on empirical evidence in
bounds for artificial neural networks before we a particular application domain or task (see for
summarize the main parts of MacKay’s single example [32].
perceptron proof [30] in Section 3. Section 4
In fact, the perceptron was introduced in
then generalizes the result for arbitrary net1958 [38] and since then has been extended in
works, including a discussion on how the capacmany variants, including but not limited to as
ities can be interpreted in bits and some convedescribed in [12, 13, 27, 28]. The perceptron
nient consequences of that. We have validated
uses a k-dimensional input and generates the
the two bounds experimentally using standard
output by applying a linear function to the inimplementations of neural networks. These reput, followed by a gating function. The gatsults are presented in Section 5 together with
ing function is typically the identity function,
actual source code for reproducibility in Apthe sign function, a sigmoid function, or the
pendix A. Section 6 finally presents direct pracrectified linear unit (ReLU) [22, 33]. Motitical applications and discusses the limits of our
vated by brain research [15], perceptrons are
results before Section 7 concludes the paper with
stacked together to networks and usually trained
future work.
by chain rule (backpropagation) [39, 40]. Even
though perceptrons have been utilized for a long
time, its capacities have been rarely explored be2 Related Work
yond discussion of linear separability versus the
Understanding machine learning, as opposed to XOR function and general mentions of overfitusing it as a black box, requires insights into the ting, generalization and catastrophic forgetting.
training and testing data, the available hypothe- Moreover, overfitting and catastrophic forgetsis space of a chosen algorithm, the convergence ting have so far not been explained satisfactoand other properties of the optimization algo- rily.
rithm, and the effect of generalization and loss
Catastrophic forgetting [31, 35] describes the
terms in the optimization problem formulation.
effect that when the net is first trained on one
set of labels and then on another set of labels,
2.1 The Perceptron
it very quickly looses its capability to classify
One of the core questions that machine learning the first set of labels. Recently, an approach
theory focuses on is the complexity of the hy- was introduced to overcome catastrophic forpothesis space and what functions can be mod- getting by avoiding too large weight changes
eled. For artificial neural networks, this ques- for weights that are associated with the prevition has recently become relevant again as deep ously learned labels [24]. Our interpretation is
learning seems to outperform shallow learning. that this approach is valid as small, alternating
For deep learning, single perceptrons with a non- weight changes will give better results in case the
linear, continuous gating function are concate- total number of weight updates exceed the canated in a layered fashion. Techniques like con- pacity of the network. Hence, we suggest future
volutional filters, drop out, early stopping, reg- algorithms should avoid catastrophic forgetting
ularization, etc., are used to tune performance, by determining the capacity and avoid overflow.
2
2.2
VC Dimension
2.3
Other Measures
One measure that handles the properties of
given data is the Rademacher complexity [5].
For understanding the properties of large neural
networks, Zhang et al. [51] recently performed
randomization tests. They show that their observed networks can memorize the data as well
as the noise. This is proven by evaluating that
their neural networks perfectly learn with random labels or with random data. This shows
that the VC dimension of the analyzed networks
is above the size of the used dataset. But it is
not clear what the full capacity of the networks
is. This observation also gives a good reason for
why smaller size networks can outperform larger
networks even though they have a lower capacity. Their capacity is still large enough to memorize the labeling of the data. A more elaborate
extension of this evaluation has been provided
by Arpit et al. [2]. Our paper indicates the lower
limit for the size of the network.
A different approach using information theory comes from the group around Naftali
Tishby [45]. They use the information bottleneck principle to analyze deep learning. For each
layer, the previous layers are treated as an encoder that compresses the data X to some better
representation T which is then decoded to the labels Y by the consecutive layers. By calculating
the respective mutual information I(X, T ) and
I(T, Y ) for each layer they analyze networks and
their behavior during training or when changing
the amount of training data. A relevant statement to our work is then presented in [45]: “An
immediate consequence of the DPI (Data Processing Inequality) is that information about Y
that is lost in one layer cannot be recovered in
higher layers.” Thereby describing the learning
capabilities of one layer, one can go to the minimum over all layers to describe the overall capabilities of a network. His results seem to be
consistent with our experimental outcomes for
second and higher deeper layers (see Section 5).
However, we describe the learning capabilities
of neural networks using a different information
theoretic view, namely the interpretation of neurons as memory cells. Apart from speculations
in Section 6 that deeper hidden layers are able
to serve as error correcting codes for previous
layers, we assume an ideal network and do not
investigate concrete architectures.
One of the largest contribution to machine learning theory comes from Vladimir N. Vapnik and
Alexey Ya. Chervonenkis [47], including the
Vapnik-Chervonenkis (VC) dimension. The VC
dimension has been well known for decades [48].
It is defined as the largest natural number of
samples in a dataset that can be shattered by
a hypothesis space. This means that for a hypothesis space having VC dimension DV C , there
exists a dataset with DV C samples such that
for any binary labeling (2DV C possibilities) there
exists a perfect classifier f in the hypothesis
space, that is, f maps the samples perfectly to
the labels. Due to perfect memorizing, it holds
DV C = ∞ for 1-nearest neighbor. Tight bounds
have so far been computed for linear classifiers
(k + 1) as well as decision trees [3].
The definition of VC dimension comes with
two major drawbacks, however. First, it considers only the potential hypothesis space but not
other aspects like the optimization algorithm, or
loss and regularization function that effect the
choice of the hypothesis [2]. Second, it is sufficient to provide only one example of a dataset
to match the VC dimension. So given a more
complex structure of the hypothesis space, the
chosen data can take advantage of this structure. As a result, shatterability can be increased
by increasing the structure of the data. While
these aspects don’t matter much for simple algorithms, it is a major point for deep neural networks.
In [49], Vapnik et al. suggest to determine
the VC dimension empirically, but state in their
conclusion that the described approach does not
apply to neural networks as they are “beyond
theory". So far, the VC dimension has only been
approximated for neural networks. For example, Abu Mostafa argued loosely that the capacity must be bounded by N 2 with N being the
number of perceptrons [1]. Recently, [42] determined in their book that for a sigmoid activation function and a limited amount of bits for
the weights, the loose upper bound of the VC dimension is O(|E|) where E is the set of edges and
consequently |E| the number of nonzero weights.
Extensions of the boundaries have been derived
for example for recurrent neural networks [26]
and networks with piecewise polynomials [4] and
piecewise linear [21] gating functions. Another
article [25] describes a quadratic VC dimension
for a very special case. The authors use a regular
grid of n times n points in the two dimensional
space and tailor their multilayer perceptron directly to this structure to use only 3n gates and
8n weights.
2.4
Memory Capacity
We are aware of recent questioning of the approach of discussing the memory capacity of
neural networks [2, 51]. However, Occam’s razor [6] dictates to follow the path of least as3
Definition 3.1 (VC Dimension [47]). The VC
dimension DV C of a hypothesis space f is the
maximum integer D = DV C such that some
dataset of cardinality D can be shattered by f .
Shattered by f means that any arbitrary labeling can be represented by a hypothesis in f (or
as in our context learned by a perceptron). If
there is no maximum, it holds DV C = ∞.
sumptions and perceptrons were initially conceived as a "generalizing memory", as detailed
for example, in the early works of Bernard
Widrow [50]. This approach has also been suggested by [1] and later explained in depth by
Sir David J. C. MacKay [30]. In fact, initial
capacity derivations for linear separating functions have already been reported by Thomas
Cover [11]. Also, the Ising model of ferromagnetism, which is clearly a model used to explain
memory storage, has already been reported to
have similarities to perceptrons [17, 18] and also
the neurons in the retina [46].
MacKay is the first one to interpret a perceptron as an encoder in a Shannon communication
model ([30], Chapter 40). In our article, we use
a slightly modified version of the model depicted
in Fig. 1. MacKay’s use of the Shannon model
allows the measurement of the memory capacity of the perceptron in bits. Furthermore, it
allows the discussion of a perceptron’s capability without taking into account the number of
bits used to store the weights (64 bit doubles,
real-valued, etc.). He also points out that there
are two distinct transition points in the error
measurement. The first one is discontinuous and
happens at the VC dimension. For a single perceptron with offset, that point is DV C = k + 1,
when k is the dimensionality of the data. Below
this point, the error should be 0 given perfect
training, as the perceptron is able to generate
any arbitrary shattering of the hypothesis space.
In honor of him describing the perceptron in a
Shannon communication model, which allows for
the generalization outlined in this paper, we call
the second point the MacKay (MK) dimension.
The MK dimension describes the largest number
of samples such that at least 50 % of all possible
labelings can be separated by the binary classifier. The MK dimension of a single perceptron
with offset is DM K = 2(k + 1). For large k (not
really required), MacKay proves that there is a
sharp drop at this point in performance. Hence,
when the neuron or the neural network has to
learn much more than its capacity allows for,
catastrophic forgetting will occur. In the following section, we will summarize the proof and
then rely on it in the remainder of the paper.
3
Definition 3.2 (General Position [30]). “A set
of points {xn } in K-dimensional space are in general position if any subset of size ≤ K is linearly independent, and no K + 1 of them lie in
a (K − 1)-dimensional plane.”
3.2
MacKay interprets a perceptron as an encoder in
a Shannon communication model [43] (compatible to our interpretation in Fig. 1). The input of
the encoder are n points in general position and
a random labeling. The output of the encoder
are the weights of a perceptron. The decoder
receives the (perfectly learned) weights over a
lossless channel. The question is then: Given
the received set of weights and the knowledge of
the data, can the decoder reconstruct the original labels of the points? In other words, the
perceptron is interpreted as memory that stores
a labeling of n points relative to the data and the
question is how much information can be stored
by training a perceptron. In other words, we
ask about the memory capacity of a perceptron.
This communication definition not only has the
advantage that the mathematical framework of
information theory can be applied to machine
learning, it also allows to predict and measure
neural network capacity in the actual unit of information, bits (more on this in Section 4.4).
The functionality of a perceptron is typically
explained by the XOR example (i. e., showing
that a perceptron with 2 input variables, which
can have 4 states, can only model 14 of the 16
possible output functions). XOR and its negation cannot be linearly separated by a single
threshold function of two variables and a bias.
For an example of this explanation, see [36], section 3.2.2. MacKay effectively changes the computability question to a labeling question by asking: Given n points, how many of the 2n possible labelings in {0, 1}n can be learned by the
model without an error (rather than computing
binary functions of k variables). Just as done
by [11, 36], MacKay uses the relationship between the input dimensionality of the data k
and the number of inputs n to the perceptron,
which is denoted by a function T (n, k) that indicates the number of “distinct threshold func-
Capacity of a Perceptron
In this section, we summarize the proof appearing in [30], Chapter 40.
3.1
Summary of Proof and Interpretation
Definitions
The following definitions will be required.
4
n\k
1
2
3
4
5
6
7
8
tions” (separating hyperplanes) of n points in
general position in k dimensions. The original
function was derived by [41]. It can be calculated recursively as:
T (n, k) = T (n − 1, k) + T (n − 1, k − 1),
(3)
where T (n, 1) = T (1, k) = 2 or iteratively:
1
2
2
2
2
2
2
2
2
2
2
4
6
8
10
12
14
16
3
2
4
8
14
22
32
44
58
4
2
4
8
16
30
52
84
128
5
2
4
8
16
32
62
114
198
6
2
4
8
16
32
64
126
240
7
2
4
8
16
32
64
128
254
8
2
4
8
16
32
64
128
256
Table 1: Some values of the T (n, k) function indicating the number of distinct threshold funcT (n, k) = 2
(4)
tions on n points in general position in k dimenl=0
sions as defined by [30].
Table 1 shows the T (n, k) function for small
n and k. It turns out that T (n, k) has very inexactly 50 % and with more samples the “perforteresting properties.
mance” largely decreases and with less samples
Namely,
it largely increases because the summands in the
n
T (n, k) = 2 for k >= n.
(5) middle are the largest.
MacKay concludes that the capacity of a perThis allows to derive the VC dimension for the ceptron is therefore 2k as the error before that
case k = n where the number of possible binary point is small. We follow Kinzel’s physical interlabelings for n points is 2n . Since k = n and pretation [23] and understand that the percepT (n, n) = 2n , all possible labelings of the input tron error function undergoes two phase transican be realized.
tions: A first order transition at the VC dimenWhen k < n, the T (n, k) function follows a sion and a continuous one at the MK dimension.
calculation scheme based on the Pascal Trian- Based on this interpretation, we predict that the
gle [10], which means that the bit loss due to different phases will play a role on structuring
incomplete shattering is still highly predictable. and explaining machine learning algorithms. We
MacKay uses an error function based on the will therefore, throughout this paper, discuss the
cumulative distribution of the standard Gaus- two points separately.
sian to perform that prediction and approximate
the resulting distribution. More importantly,
he defines a second point, which we call MK 4
Capacity of a Perceptron
dimension. The MK dimension describes the
Network
largest number of samples such that typically
only about 50 % of all possible labelings can be
separated by the binary classifier. He proofs this The following further definitions will be useful.
point to be a n = 2k for large k and illustrates
that there is a sharp continuous drop in perfor- 4.1 Definitions
mance at this point. Since the sum of two independent normally distributed random variables For the remainder of this article, we will asis normal, with its mean being the sum of the sume a feed-forward network. The weights are
two means, and its variance being the sum of the assumed to be real-valued and each unit has a
two variances, it is only natural that we will see bias, which counts as a weight. Note that no
in the following section that the MacKay point further assumptions about the architecture are
is linearly additive in the best case.
required.
The quick analytic proof for what is
The definition of general position used in the
2k
shown in [30] is
that
T
(2k,
2k)
=
2
=
previous
section is typically used in linear al
P2k−1 2k − 1
gebra and is the most general case needed for
and that it is a symmetric
2 l=0
l
a perceptron that uses a hyperplane for linear
sum where the last summand is equal to the separation (see also Table 1 in [11]). For neural
first, etc. Hence, if we just take half of the networks, a stricter setting is required since neusummands, we get half of the values of the sum ral networks can implement arbitrary non-linear
which means T (2k, k) = 22k /2. This means that separations.
a perceptron with k inputs can handle half of all
possible labelings of 2k samples. Note that this Definition 4.1 (Chaotic Position). A set of
direct proof actually does not require large n and points {xn } in K-dimensional space is in chaotic
that the transition is sharp in the sense that it is position, if and only if from any subset of size
k−1
X
n−1
l
5
< n it is not possible to infer anything about the
positions of the remaining points.
needs to be lossless, regardless of individual neurons being lossy. In fact, it is tempting to call
Hp = log2 (T (n, k)) something like the entropy of
a perceptron as it indicates the expected number of bits of labeling a perceptron can assign
to a set of points. However, empirically it can
be easily observed that Hp is not additive as
it leads to an overestimation of the capabilities.
Analytically, it is also not clear how individual
hyperplanes are intersecting when perceptrons
are connected in various ways. More particularly, it seems hard to assume a completely disjoint labeling of points when perceptrons that
are able to implement the same functions are
stacked. Instead, our proof abstracts from a concrete network architecture and assumes the ideal
network, handling the worst case data. The two
critical points described in Section 3 were obtained by describing the single perceptron as a
memory cell. In the generalization, we interpret
an artificial neural network as an ideal combination of these memory cells.
We can now show that the critical points DLM
and DM K of any perceptron network scale linearly.
Note that chaotic position implies general position, which was only excluding linear inference.
Bear in mind that slightly distorted grid settings, as for example used in [25], are in general
position but not in chaotic position. Chaotic
position is equivalent of saying that no inference
is possible about the structure of the data and
the only thing a machine learner can do is memorize. The only distribution that satisfies this
constraint is the uniform distribution [19].
As explained in Section 2, it is possible to
achieve very high VC dimension by the choice of
very special datasets. This has not been an issue
yet for learning theory but from a practitioner
perspective, this has been criticized [44, 51, 2].
To avoid the reported problems and consistent
with our embedding into the Shannon communication model, we therefore propose a generalization of the VC dimension which we call lossless
memory dimension.
Definition 4.2 (Lossless Memory Dimension).
The lossless-memory dimension DLM is the
maximum integer number DLM such that for Theorem 4.1 (Capacity Scaling Law).
any dataset with cardinality n ≤ DLM and
points in chaotic position, all possible labelings of
l
l
X
X
this dataset can be represented with a function
C(Pj ) = C
Pj
(6)
in the hypothesis space, that is, learned even
j=1
j=1
with perfect training.
where Pj is an arbitrary perceptron with nj inNote that for a single perceptron DLM = DV C puts including a potential offset weight. The cabecause chaotic position implies general posi- pacity is either C = D
M K or C = DLM dependtion. As explained in Section 2, we will name ing on the targeted phase. P P denotes a neural
j
the corresponding point where loss is guaranteed network that combines the respective perceptrons
MacKay dimension.
perfectly and the data points are assumed to be
Definition 4.3 (MacKay Dimension). The in chaotic position.
MacKay dimension DM K is the maximum integer DM K such that for any dataset with cardinality n ≤ DM K and points in chaotic position at least 50% of all possible labelings of these
datasets can be represented with a function in
the hypothesis space, that is, learned with perfect training [30] (100% accuracy).
Proof. Each neuron with weights k (including
bias) is able to implement exactly T (n, k) different binary threshold functions over n sample
points. With the maximum number of binary
labelings of n points being 2n , it follows that
the perceptron is at DLE when T (n, k) = 2n .
It is then able to store n bits. Adding lossConsequently, a higher cardinality than DM K less memory cells with capacity n and m inimplies less than 50% of the labelings can be creases their capacity linearly in the amount
represented. We will show that for an ideal per- of bits. Lossy memory cells are bit-additive
as well. In fact, if perceptrons are combined
ceptron network the limit is exactly 50%.
in the best possible way (i. e., they label disjoint sets of sample points correctly), two per4.2 Capacity Scaling Law
ceptrons with k1 and k2 weights can maximally
k1
k2
We will now extend MacKay’s capacity proof label 2 ∗ 2 points. The transitivity of addifrom a single perceptron to neural networks tion implies that DLM is additive for any perit holds
!
of many perceptrons. Generalizing to a neu- ceptron or network. In other words,
l
l
P
P
ral network that is able to losslessly memorize that
DLM (Pj ) = DLM
Pj in the case
the input labels means the network as a whole
j=1
j=1
6
the number of sample points n and each neuron
is exactly responsible for its fraction of the labeling of the dataset. In other words, there are
no “free labels to borrow" from anywhere. since
T (n, k) is monotonically falling for an increasing nk , we can assume that there are no other
extreme points between DLM and DM K . Furthermore, we have to assume that capacity is
asymptotically approaching 0 for high values of
n
k . We can therefore assume that the phase transitions observed in a single perceptron generalize
to a network of perceptrons.
that perceptrons are either lossless, or lossy but
combined in the best possible way.
To proof the linear scalability of the MK dimension, we take a look at a ideal network at
DLM , where the number of weights k equals the
number of sample points n. It follows that each
neuron is exactly responsible for its fraction of
the labeling of the dataset thus T (n, k) = 2n . If
we increased n without increasing k and there
was a hard cut-off (e. g., conventional transistor memory), we would not be able guarantee
the assignment of any labels to new points as
we would have to erase the original configuration bit-by-bit for each new label. However, as
we discussed in Section 3, there is no hard cutoff as T (2k, k) = 12 T (2k, 2k). Assuming each
perceptron equally contributes to the labeling of
more incoming points as needed, doubling n for
a fixed k should result in each neuron being able
to memorize the labeling of half of all points.
Therefore DM k = 2DLM generalizes to an ideal
neural network.
In consequence, both DLM and!DM K scale
l
l
P
P
linearly, or
C(Pj ) = C
Pj for a perj=1
4.3
Consequences
In conclusion, in an ideal network, all we need
to do is add up the weights until the number
of input points is reached to guarantee lossless
memorization. Halving the number of weights,
results in the MK dimension. Generalizing to a
neural network that is able to losslessly memorize the input labels means the network as a
whole needs to be lossless, regardless of individual neurons being lossy. Note that capacities in
information theory are upper bounds, the measurable LM and MK dimension of a given concrete network structure is therefore most likely
lower. It can be assumed that two lossy memory
cells are more likely to complement each other
when a global learning function is used. The
learning function will therefore influence an empirically measured capacity, so does the gating
function. In the extreme case of using the identity as gating function in an MLP, the capacities
are limited to the first perceptron of the feed forward step, and we get again k + 1 and 2k + 2
as dimensions because the overall binary separation function is still linear no matter how the
perceptrons are combined.
The following formulas for the capacity of a
network might come in handy in practice. If we
count the bias weights of each perceptron as an
edge of a graph, the number of incoming edges
|E| and the number of weights P are identical
and the following holds:
j=1
ceptron Pj with nj weights and C = DM K or
C = DLM assuming an ideal network and points
in chaotic position.
Practically, Equation 6 is an inequality “≥”
when the data is not in chaotic position because
the network should be able to exploit redundancies.
Please allow us to provide the following intuition regarding the MK dimension. Each perceptron reduces the amount of possible labelings by
half but it is a joint reduction in the best case. It
is not a separate reduction for each perceptron
which would result in a reduction by 2p , with p
being the number of perceptrons. This can be
understood by taking another look at the the
T (n, k) function. It is composed of the sum of
two T (n, k) functions of n − 1: one that uses k
weights and one that uses k−1 weights. So when
n is increased, the new function space is composed of a space of functions that requires k parameters (let’s call them “difficult"), and a space
of functions that uses k −1 parameters (let’s call
them “easy"). We could therefore argue that a
network at MK dimension can only handle the
easier labelings, which turn out to be exactly
50 % of the functions modeled at DLM if we had
twice the amount of edges. We do not know how
a network of perceptrons behaves between DLM
and DM K and after DM K as we cannot assume
T (n, k) to hold for non-linear separation functions. However, we can again assume an ideal
network where the number of weights k equals
DM K = 2|E| .
(7)
As of the previous section, each perceptron
has MK dimension DM K = 2|E|. Hence, h unconnected hidden units would have capacity
DM K = h ∗ |E| and DM K = 2 ∗ DLM .
(8)
Connecting the hidden units with an output
unit results in two observations. First, the output unit is a perceptron with h inputs and a
threshold and has the task of encoding 2h possible labelings. The maximum capacity that this
unit can add is therefore DM K = 2|E| = 2h bits.
7
Input
Layer
Hidden
Layer
Output
Layer
1
Offset
k+1
k+1
k+1
k+1
k+1
2
Input #1
3
Input #2
4
Output
h(k+1)+h+1
1
Input #3
h
Input #k
Offset
Figure 2: Capacity calculation of a 3-Layer MLP with k inputs and h hidden nodes. Displayed are
the contributions of each weight for the joint LM dimension as explained in Section 4.
in Shannon’s communication model. The intuition is that a binary labeling of d random points
is a bit vector with dimension d. If the capacity
allows, the weights can label all the points as
targeted by a training method. In other words,
if the network has the capacity to store the labeling of the d points, there is no error and all
the labeling bits are restored at the output. As
shown in the previous sections, this is guaranteed at the LM dimension, which is the case for
d = k + 1 = C(P ) = DLM . The guaranteed
capacity is therefore d bits. For the MK dimension, the capacity is therefore d/2 bits.
As a consequence, the LM and MK dimensions
can be interpreted using information theoretic
measures, such as the KL Divergence which allows predicting the minimally expected error for
a given neural network, given some input data.
It is also worth noting that the KL divergence
is closely related to the cross entropy, which in
turn has recently been proposed as an optimization criterion for deep learning algorithms [20].
Second, since we treat a perceptron as a memory unit, we do not care that the output unit
looks at a restricted space compared to the input space, the capacity of the unit itself doesn’t
change.
Hence, for a 3-layer Multilayer Perceptron
(MLP) with h hidden units, k inputs, a single
output node, and respective offsets we get:
DM K = 2 (h(k + 1) + h + 1) .
(9)
The LM dimension DLM is consequently
DLM = |E|
(10)
DLM = (h(k + 1) + h + 1)
(11)
and respectively
for the MLP.
A graphical illustration of the scaling law can
be seen in Figure 2. The scaling law is also
strongly supported by our empirical results (see
Section 5).
One can generalize to more than binary classification as follows: If we assume a binary classification network at DLM , it means that any
binary labeling can be implemented with one
output unit. Adding a second output unit of
the same capacity as the first one must work for
a specific second binary labeling as well. Just
imagine the first output unit away. A multiclassifier MK dimension gives every new output
node a 50% chance that it can implement a particular function.
5
Capacity Measurements
This section describes how we evaluated LM and
MK dimension with empirical means. We observe that our theoretical capacities are indeed
upper limits.
5.1
Experimental Setup
The basic principle for our empirical evaluation
is to obtain samples from randomly generated
Measuring the capacities C of a neural network data and increase the number of input points to
in bits follows directly from MacKay’s defini- the network step-by-step to test if the network
tion of a perceptron being an encoder/decoder can learn all possible labelings for the LM di-
4.4
Unit of Capacity
8
nodes, we looked at [1, 2, 3, 4, 5, 6, 7, 8]. Our implementation does not consider the difficulties of
an MLP with class imbalance or redundancies.
Here, higher empirical dimensions due to oversampling might be achievable.
mension or half of the possible labelings for the
MK dimension.
Obviously, we expect our empirical measurements to be lower than the theoretical capacities. Practically, neither the ideal network nor
the perfect training alogrithm exists. Furthermore, for higher dimensions, we were only able
to sample from the hypothesis space and could
not test all labelings exhaustively. Therefore our
goal was to create the best conditions possible
and give the network the highest chance of reaching optimal capacity without violating the constraints of the theoretical framework. Thereby
some practical workarounds are required for
speedup and some limitations arise due to the
exponential increase of the search space.
We mainly used the MLP implementation in
scikit-learn [34] with L-BFGS [29] as optimizer.
Our code is provided in the Appendix A. To control the randomness and ensure consistent results, we seed the randomizers with the respective index of the repetition. In case the optimizer does not fit the training data, we repeat
its training up to 20 times. Our data was randomly generated by sampling from a normal distribution. We repeated evaluations with up to
20 different datasets if a labeling could not be
fitted in the case of the LM dimension or if 50%
of the labelings could not be fitted in the case
of the MK dimension. The processing time of
the latter is much higher for two reasons. First,
a larger amount of samples has to be analyzed
since at least 50% of all labelings have to be evaluated every time. Second, with more data the
convergence of the MLP takes more iterations.
For completeness, every labeling would have
to be tested. Due to symmetry in the class handling by the MLP, a minor speedup was achieved
by testing only labelings where the last sample
was labeled with a "0" and not a "1". This
was not possible for large LM dimensions. Testing more than 215 labelings was computationally too expensive for us. Hence, for more than
15 samples, we tested only a random selection
of 215 labelings. Due to this approximation, results might be above the true values for the given
structure. The processing effort of the MK dimension is even worse and required to have a
limit of 210 samples. Given more resources, one
could imagine a better approach where multiple
random samplings are tested and the median result for the MK dimension and the worst result
for the LM dimension is taken. We leave this as
future work.
The number of tested labels also limits the
possible dimensions of the MLP. We analyzed
input dimensions: [1, 2, 3, 4]. 1 did not provide reliable results. For the number of hidden
5.2
Tuning the Implementation
Apart from the aforementioned implementation,
we tested all other optimizers like “Adam” and
“SGD” as well as the Keras library [8]. In most
cases, the net was not able to fit the data in
contrast to using L-BFGS. Hence the measured
dimensions were very low. This could be interpreted as generalization capability of Adam and
SGD because the optimizer is avoiding overfitting. Note, that L-BFGS approximates the second order derivative which makes it more accurate but also computationally more expensive
and prone to get stuck in local minima.
We also tested different gating functions. Using the identity function, the network mostly behaved like a single perceptron as expected. For
tanh and logistic function, results looked similar
to the ReLU function but needed more repetitions and processing time.
As expected, the generation of the data had
a significant impact on the results. Originally,
we tested with uniformly sampled data. Changing it to sampling from a normal distribution
improved our results dramatically (i. e., the empirically measured upper bound came closer to
the theoretical). The number of different tested
datasets using the same distribution had only
a minor effect on when the empirical calculation reached its limit in LM or MK dimension.
The testing of more than one dataset was solely
to capture the randomness in the training algorithm and had no significant impact on the
empirical results.
5.3
One Hidden Layer
The results with our described setting for the
LM dimension are depicted in Figure 3.
Using just one hidden neuron behaved always
like a single perceptron with LM dimension k +
1 and MK dimension 2(k + 1). The predicted
linear relationship in h as well as in k for both
dimensions can also be observed.
The comparison between theoretical and empirical dimensions shows a similar linear behavior. For the larger LM dimensions, the differences get smaller but this is probably due to
sampling error. For the VC dimensions, it is
more important to test all labelings because a
single misclassification has an impact, whereas
for the MK dimension this effect is less severe.
9
Figure 3: Experimental results for LM dimension (left) and MK dimension (right). Displayed are
the functional dependency on k (top) and on h (bottom). The solid lines depict the theoretical
boundaries whereas the respective dotted lines display our empirical results. The black lines display
the number of samples where not all labelings are tested anymore but a random sample, which
makes the empirical results less reliable.
5.4
This could be improved in the future with more
processing power.
We observed that the empirical MK dimension
is extremely close to twice the empirical VC dimension. This is expected from the theoretical
derivations but considering the aforementioned
practical shortcuts, the clarity of this result increases our confidence in the validation experiments.
Two Hidden Layers
We also performed experiments with going
deeper than one layer and, as expected, there
was no more than linear increase in the capacity
of the network. In fact, in case of using small k
and h, the obtained results were better by just
one sample compared to the respective LM dimension with a one hidden layer architecture. In
most cases, we observed that the LM dimension
was actually far below the empirical values of a
respective network with one hidden layer.
The empirical values for the VC dimension
come quite close to hk + 1 for small numbers
which is 2h off from the optimal value. By
increasing the number of iterations and tested
6 Interpretation and Utility
datasets, we also detected three special cases
that are worth pointing out here.
A perceptron can be interpreted as either a unit
For an MLP with 2 hidden nodes and input di- that models a function of n input variables or,
mensions of 3, 4, or 5, we found a dataset exam- as we have done in this paper, as a memory unit
ple of 9, 11, or 13 samples respectively that could that tries to label n sample points with binary
be shattered. In those cases, we tested all label- digits. The choice of interpretation makes a difings. Those sample values are exactly one sam- ference. For example, interpreting the neuron
ple higher than h(k + 1) and therefore above the as a binary functional unit, one perceptron has
n
storage capabilities of the hidden layer. Hence, to be able to represent 22 different functions
the output neuron is making a significant contri- of n input variables with k + 1 weights. A task
bution to the resulting learning capabilities, as that it is not able to achieve as the number of
predicted by the memory capacity formulation functions grows too quickly (see also [36], Secin this paper.
tion 6.3). Interpreting the perceptron as mem10
ory unit, the k + 1 variables only have to be able
to model 2n binary labelings of n points. Intuitively, given that a hyperplane is already separating a space into two different regions each of
which is labeled with one bit, the difference between log2 T (n, k + 1) and log2 2n appears almost
insignificant. This has already been discussed
in [30] (Figure 40.9). However, this seemingly
insignificant difference allows for one thing that
a memory stick (which perfectly stores any 2n
labelings with n bits) is not able to do: generalization. Assume points form two clusters, the
promise is that very few hyperplanes can be used
to parametrize the difference between the two
clusters. So interpreting perceptrons as memory units does not exclude generalization but at
the same time allows us to understand their limits better. We therefore like to concur with the
early assumptions in [1] that understanding neural networks as memory is indeed beneficial.
application, in that the variation of the linear
separability of the input data and the learning
function for that will dominate the quality of the
prediction of the network.
Our result has at least three practical applications.
First, an overflow in capacity of all, some,
or even a single perceptron serves as one definite explanation for (potentially catastrophic)
forgetting.
Second, our experimental methodology, including the code in Appendix A, serves as a
benchmark for the evaluation of neural network implementations. Using points in general
position, one can test any learning algorithm
and network architecture against the theoretical
limit both for performance and efficiency (convergence rate).
Third, statistically speaking we should be able
to compare networks with each other by simple
However, it is important to understand that bit counting of edges and comparing it to the
the capacities, we derived for neural networks, total number of samples. For example, for a
are upper limits and that optimization meth- binary function of D variables, our analysis says
ods, target functions, feature extraction, prepro- that we will need a network structure with more
cessing, initialization methods, specialized ar- than 2D edges. For a 3-layer MLP this means,
D−1
chitectures, and other such work will still be that we need more than 2D+2 hidden neurons.
of practical value. Furthermore, as explained Having less than 2D−1 edges, we know from the
above, our proof is not constructive. Just like MK Dimension, that the chance of a labeling not
in Claude Shannon’s original work [43], we are being learned is even higher than 50%.
not discussing the structure of the encoder. We
are only quantifying maximum capabilities. Additionally, the main use of neural networks is 7
Conclusion
generalization, not memorization. Therefore it
might be acceptable to lose bits as long as they Using an information theoretic proof and emare the right ones.
pirical evidence, we show that the capacities of
While our work is network architecture agnostic and therefore implying that deep and flat networks have the same capacity, this should not be
interpreted as them having necessarily the same
performance. The first hidden layer has complete exposure to the input data and further
layers only have exposure to the informationreduced output of the previous layers. Then
again, deeper layers have the same capacity as
earlier layers (assuming the same amount of
weights) while looking at a smaller input space.
So deep layers have more free capacity. This
is countered by results [37] that neural networks
with a so-called shortcut (i. e., where all the hidden layers are directly connected to the input)
need overall less weights to learn the same labeling (still bounded by our capacity law). However, it is also easy to show that a deeper hidden layer can serve as an error correcting code to
the previous layers. For example, the output of
the previous hidden layer can be interpreted as
repetition code. Practically speaking, the concrete architecture will be most dependent on the
11
perceptrons in a neural network scale linearly
with the number of weights. In fact, for guaranteed perfect memorization accuracy, a neural
network can learn one bit per weight. The goal
can be stretched to 2 bits per weight without too
much error. After that, the error is guaranteed
to be very large. While our work is an extension
of the initial work by David MacKay, this article is the first to generalize the critical points
to multiple perceptrons and derive a concrete
scaling law. We show using repeatable experiments (see Appendix A) with a widely used open
source framework that linear scaling holds and
our theoretical bounds are actual upper bounds.
This holds for all threshold-like activation function tested, including sigmoid and ReLU.
Future work should be dedicated to the phase
between LM and MK dimension, which could be
physically interpreted as the “liquid phase”, because in that phase it is clear that the network
will make errors (i. e., lose bits compared to the
input) but the error is log-proportional to the
network complexity (i. e., the number of possi-
ble configurations) [14]. The use of neural networks is typically not that of memory but that
of abstraction (assumed as learning) and qualifying this information loss for concrete data would
therefore allow to reason about the generalization capabilities. For example, if only insignificant bits (noise) were lost as part of the error
of the network, it would back up the hypothesis
that one should learn in a way to avoid overfitting the data.
It is highly conceivable that weight values are
not uniformly distributed, especially for deeply
layered nets. This would mean that the capacity
of individual perceptrons within a network can
overflow as well. As a consequence, it will forget some previously learned information based
on learning more. That individual overflow can
cause interesting effects in the dynamics of the
overall system [16] and gives one possible cause
for why the theoretical upper limit is not always reached. Investigating methods to detect
such a local overflow and redirecting further input to other neurons could be interesting, especially when neural networks are trained in a
parallelized fashion.
More challenges include understanding the
deeper hidden layers as error correcting codes,
generalizing these results to networks with more
complex activation functions (such as in Radial
Basis Function1 networks [7]), and finding the
bounds for convolutional networks. Last but not
least, it remains to provide more constructive
proofs of the capabilities depending on the gating function and the network architecture. Unfortunately, in general, finding an explicit formula that shows if a certain function is implemented by a network or not is very hard. Even
just the case of binary functions would reduce
to the problem of boolean satisfiability [9].
Acknowledgements
This work was performed under the auspices
of the U.S. Department of Energy by Lawrence
Livermore National Laboratory under Contract
DE-AC52-07NA27344. It was also partially
supported by a Lawrence Livermore Laboratory Directed Research & Development grant
(17-ERD-096). IM release number LLNL-TR736950. Mario Michael Krell was supported by
the Federal Ministry of Education and Research
(BMBF, grant no. 01IM14006A) and by a fellowship within the FITweltweit program of the
German Academic Exchange Service (DAAD).
Any findings and conclusions are those of the
authors, and do not necessarily reflect the views
1 e. g.,
2
e−||x−w||
/(2σ 2 )
instead of max{0,
P
wi xi }
12
of the funders. We want to cordially thank Raúl
Rojas for in depth discussion on the chaining of
the T () function. We also want to thank Alfredo Metere, Jerome Feldman, Kannan Ramchandran, Jan Hendrik Metzen, Bhiksha Raj,
Naftali Tishby, Jaeyoung Choi, Friedrich Sommer and Andrew Feit for their insightful advise
and Barry Chen and Brenda Ng for their support.
References
[1] Y. Abu-Mostafa.
Information theory,
complexity and neural networks. IEEE
Communications Magazine, 27(11):25–28,
November 1989.
[2] D. Arpit, S. Jastrzȩbski, N. Ballas,
D. Krueger, E. Bengio, M. S. Kanwal,
T. Maharaj, A. Fischer, A. Courville,
Y. Bengio, and S. Lacoste-Julien. A Closer
Look at Memorization in Deep Networks.
jun 2017.
[3] O. Asian, O. T. Yildiz, and E. Alpaydin.
Calculating the VC-dimension of decision
trees. In 24th International Symposium on
Computer and Information Sciences, pages
193–198. IEEE, sep 2009.
[4] P. L. Bartlett, V. Maiorov, and R. Meir. Almost linear vc dimension bounds for piecewise polynomial networks. In Advances
in Neural Information Processing Systems,
pages 190–196, 1999.
[5] P. L. Bartlett and S. Mendelson.
Rademacher and Gaussian Complexities: Risk Bounds and Structural Results.
Journal of Machine Learning Research,
3:463–482, 2001.
[6] A. Blumer, A. Ehrenfeucht, D. Haussler,
and M. K. Warmuth. Occam’s razor. Information processing letters, 24(6):377–380,
1987.
[7] D. S. Broomhead and D. Lowe. Radial basis functions, multi-variable functional interpolation and adaptive networks. Technical report, Royal Signals and Radar Establishment Malvern (United Kingdom), 1988.
[8] F. Chollet et al. Keras. https://github.
com/fchollet/keras, 2015.
[9] S. A. Cook. The complexity of theoremproving procedures. In Proceedings of the
third annual ACM symposium on Theory of
computing, pages 151–158. ACM, 1971.
[10] J. L. Coolidge. The story of the binomial theorem. The American Mathematical
Monthly, 56(3):147–157, 1949.
[22] K. He, X. Zhang, S. Ren, and J. Sun.
Delving Deep into Rectifiers: Surpassing
Human-Level Performance on ImageNet
Classification. In 2015 IEEE International
[11] T. M. Cover. Geometrical and statistiConference on Computer Vision (ICCV),
cal properties of systems of linear inequalpages 1026–1034. IEEE, dec 2015.
ities with applications in pattern recognition. IEEE transactions on electronic com- [23] W. Kinzel. Phase transitions of neural
networks. Philosophical Magazine Part B,
puters, (3):326–334, 1965.
77(5):1455–1477, 1998.
[12] K. Crammer, O. Dekel, J. Keshet, [24] J. Kirkpatrick, R. Pascanu, N. RabiS. Shalev-Shwartz, and Y. Singer. Online
nowitz, J. Veness, G. Desjardins, A. A.
Passive-Aggressive Algorithms. Journal of
Rusu, K. Milan, J. Quan, T. RaMachine Learning Research, 7:551 – 585,
malho, A. Grabska-Barwinska, D. Hassabis,
2006.
C. Clopath, D. Kumaran, and R. Hadsell.
Overcoming catastrophic forgetting in neural networks. In Proceedings of the National
Academy of Sciences of the United States
of America, volume 114, pages 3521–3526.
National Academy of Sciences, mar 2017.
[13] O. Dekel, S. Shalev-Shwartz, and Y. Singer.
The Forgetron: A Kernel-Based Perceptron
on a Budget. SIAM Journal on Computing,
37(5):1342–1372, jan 2008.
[14] M. Dzugutov, E. Aurell, and A. Vulpi- [25] P. Koiran and E. D. Sontag.
Neural
ani.
Universal Relation between the
Networks with Quadratic VC Dimension.
Kolmogorov-Sinai Entropy and the TherJournal of Computer and System Sciences,
modynamical Entropy in Simple Liquids.
54(1):190–198, feb 1997.
Phys. Rev. Lett., 81:1762–1765, Aug 1998.
[26] P. Koiran and E. D. Sontag. VapnikChervonenkis dimension of recurrent neu[15] J. A. Feldman. Dynamic connections in
ral networks. Discrete Applied Mathematneural networks. Biological Cybernetics,
ics, 86(1):63–79, aug 1998.
46(1):27–39, dec 1982.
[27]
[16] G. Friedland and A. Metere. An Isomorphism between Lyapunov Exponents
and Shannon’s Channel Capacity. arXiv
preprint arXiv:1706.08638, June 2017
https://arxiv.org/abs/1706.08638.
[28]
[17] E. Gardner. Maximum storage capacity in
neural networks. EPL (Europhysics Letters), 4(4):481, 1987.
M. M. Krell. Generalizing, Decoding, and
Optimizing Support Vector Machine Classification. Phd thesis, University of Bremen,
Bremen, 2015.
M. M. Krell and H. Wöhrle. New one-class
classifiers based on the origin separation approach. Pattern Recognition Letters, 53:93–
99, feb 2015.
[29] D. C. Liu and J. Nocedal. On the limited memory BFGS method for large scale
optimization. Mathematical Programming,
45(1-3):503–528, aug 1989.
[18] E. Gardner. The space of interactions in
neural network models. Journal of physics
A: Mathematical and general, 21(1):257,
1988.
[30] D. J. C. MacKay. Information Theory, Inference, and Learning Algorithms. Cambridge University Press, New York, NY,
USA, 2003 http://www.inference.org.
uk/itila/book.html.
[19] J. W. Gibbs. The scientific papers of J.
Willard Gibbs, volume 1. Longmans, Green
and Company, 1906.
[20] I. Goodfellow, Y. Bengio, and A. Courville. [31] M. McCloskey and N. J. Cohen. CatasDeep Learning. MIT Press, 2016.
trophic Interference in Connectionist Networks: The Sequential Learning Problem.
[21] N. Harvey, C. Liaw, and A. MehraPsychology of Learning and Motivation,
bian. Nearly-tight VC-dimension bounds
24:109–165, 1989.
for piecewise linear neural networks. Proceedings of Machine Learning Research: [32] N. Morgan. Deep and Wide: Multiple LayConference on Learning Theory, 7-10 July
ers in Automatic Speech Recognition. IEEE
2017, Amsterdam, Netherlands, 65:1064–
Transactions on Audio, Speech, and Lan1068, mar 2017.
guage Processing, 20(1):7–13, Jan 2012.
13
[33] V. Nair and G. E. Hinton. Rectified Linear [44] R. Shwartz-Ziv and N. Tishby. Opening the
Units Improve Restricted Boltzmann MaBlack Box of Deep Neural Networks via Inchines. In Proceedings of the 27th Internaformation. arXiv preprint, 2017.
tional Conference on International Conference on Machine Learning, ICML’10, pages [45] N. Tishby and N. Zaslavsky. Deep learning
and the information bottleneck principle. In
807–814, USA, 2010. Omnipress.
2015 IEEE Information Theory Workshop
(ITW), pages 1–5, April 2015.
[34] F. Pedregosa, G. Varoquaux, A. Gramfort,
V. Michel, B. Thirion, O. Grisel, M. Blon[46] G. Tkacik, E. Schneidman, I. Berry,
del, P. Prettenhofer, R. Weiss, V. Dubourg,
J. Michael, and W. Bialek. Ising models
J. Vanderplas, A. Passos, D. Cournapeau,
for networks of real neurons. arXiv preprint
M. Brucher, M. Perrot, and É. Duchesq-bio/0611072, 2006.
nay. Scikit-learn: Machine Learning in
Python. Journal of Machine Learning Re- [47] V. N. Vapnik. The nature of statistical
search, 12:2825–2830, feb 2011.
learning theory. Springer, 2000.
[35] R. Ratcliff. Connectionist models of recog- [48] V. N. Vapnik and A. Y. Chervonenkis. On
nition memory: constraints imposed by
the Uniform Convergence of Relative Frelearning and forgetting functions. Psychoquencies of Events to Their Probabilities.
logical review, 97(2):285–308, apr 1990.
Theory of Probability & Its Applications,
16(2):264–280, jan 1971.
[36] R. Rojas. Neural networks: a systematic introduction. Springer-Verlag, 1996 https:// [49] V. N. Vapnik, E. Levin, and Y. L. Cun.
page.mi.fu-berlin.de/rojas/neural/.
Measuring the VC-Dimension of a Learning
Machine. Neural Computation, 6(5):851–
[37] R. Rojas.
Deepest Neural Networks.
876, sep 1994.
ArXiv e-prints arXiv:1707.02617, July
2017 http://adsabs.harvard.edu/abs/ [50] B. Widrow. Generalization and information storage in network of adaline’neurons’.
2017arXiv170702617R.
Self-organizing systems-1962, pages 435–
[38] F. Rosenblatt. The perceptron: a proba462, 1962.
bilistic model for information storage and
organization in the brain. Psychological re- [51] C. Zhang, S. Bengio, M. Hardt, B. Recht,
and O. Vinyals. Understanding deep learnview, 65(6):386–408, November 1958.
ing requires rethinking generalization. In
[39] D. E. Rumelhart, G. E. Hinton, and R. J.
International Conference on Learning RepWilliams. Parallel Distributed Processresentations (ICLR), 2017.
ing: Explorations in the Microstructure
of Cognition, Vol. 1. chapter Learning
Internal Representations by Error Propagation, pages 318–362. MIT Press, Cambridge, MA, USA, 1986.
[40] D. E. Rumelhart, G. E. Hinton, and R. J.
Williams. Neurocomputing: Foundations of
research. chapter Learning Representations
by Back-propagating Errors, pages 696–
699. MIT Press, Cambridge, MA, USA,
1988.
[41] L. Schläfli. Theorie der vielfachen Kontinuität. Birkhäuser, 1852.
[42] S. Shalev-Shwartz and S. Ben-David. Understanding Machine Learning: From Theory to Algorithms. Cambridge University
Press, New York, NY, USA, 2014.
[43] C. E. Shannon. The Bell System Technical
Journal. A mathematical theory of communication, 27:379–423, 1948.
14
A
Code
print ( ’n ’ , ’k ’ , ’h ’ , ’ correct ’ , ’ rate ’ )
for k in K: # input dimension
numpy . random . s e e d ( 0 )
for h in H: # number of hidden layers
numpy . random . s e e d ( 0 )
for n in range (N ) :
# dataset size
n += 1 # We start with one sample .
d a t a _ r e s = [ ] # Good results
# first label is fixed to be zero
l _ l e n = min ( n−1 ,max_l−1)
# 20 different random datasets
for r_data in range ( 2 0 ) :
numpy . random . s e e d ( r_data )
# normal distributed data
data = numpy . random . normal (
s i z e =[N, k ] )
numpy . random . s e e d ( 0 )
true_results = 0
for l a b e l _ i n t in range ( 2 ∗ ∗ l _ l e n ) :
if max_l < n :
label_int = \
random . r a n d i n t ( 0 , 2 ∗ ∗ ( n −1))
l a b e l s = [ int ( i ) for i in bin (
l a b e l _ i n t ∗2+2∗∗(N+2))[ −n : ] ]
d = data [ : n ]
converged = False
# repeated runs till converged
for r_mlp in range ( 2 0 ) :
c l f = MLPClassifier (
h i d d e n _ l a y e r _ s i z e s =(h , ) ,
random_state=r_mlp ,
a c t i v a t i o n=’ relu ’ ,
s o l v e r=’ lbfgs ’ , a l p h a =0)
c l f . f i t (d , l a b e l s )
p = c l f . predict (d)
if ( p == l a b e l s ) . all ( ) :
t r u e _ r e s u l t s += 1
c o n v e r g e d = True
break # short converged
if not c o n v e r g e d :
break # shortcut after miss
d a t a _ r e s . append ( t r u e _ r e s u l t s )
# All labelings correct ?
if t r u e _ r e s u l t s == 2∗∗ l _ l e n :
break
t r u e _ r e s u l t s = max ( d a t a _ r e s )
print ( n , k , h , t r u e _ r e s u l t s ,
true_results ∗1.0/2∗∗ l_len )
if t r u e _ r e s u l t s ∗ 1 . 0 / 2 ∗ ∗ l _ l e n < 0 . 9 5 :
break
# Variable definitions
N = 120 # Maximum number of samples
K = [ 1 , 2 , 3 , 4 ] # Analyzed dimensions
# Analyzed numbers of hidden layers
H = [1 ,2 ,3 ,4 ,5 ,6 ,7 ,8]
# Maximum number of samples ,
max_l = 10 # for random labelings
# Imports
import i t e r t o o l s
import numpy
import random
from s k l e a r n . neural_network \
import M L P C l a s s i f i e r
print ( " n " , " k " , " h " , " correct " , " rate " )
for k in K: # input dimension
numpy . random . s e e d ( 0 )
for h in H: # number of hidden layers
numpy . random . s e e d ( 0 )
for n in range (N ) :
# dataset size
n += 1 # We start with one sample
if n <= 2 ∗ ( h ) ∗ ( k−1)+k +1:
continue # shortcut
data_res = [ ]
# Good results
# first label is fixed to be zero
l _ l e n = min ( n−1 ,max_l−1)
# 20 different datasets
for r_data in range ( 2 0 ) :
numpy . random . s e e d ( r_data )
data = numpy . random . normal (
s i z e =[N, k ] )
numpy . random . s e e d ( 0 )
true_results = 0
for l a b e l _ i n t in range ( 2 ∗ ∗ l _ l e n ) :
index = label_int
if max_l < n :
label_int = \
random . r a n d i n t ( 0 , 2 ∗ ∗ ( n −1))
l a b e l s = [ int ( i ) for i in bin (
l a b e l _ i n t ∗2+2∗∗(N+2))[ −n : ] ]
d = data [ : n ]
converged = False
for r_mlp in range ( 2 0 ) :
c l f = MLPClassifier (
h i d d e n _ l a y e r _ s i z e s =(h , ) ,
random_state=r_mlp ,
a c t i v a t i o n=’ relu ’ ,
s o l v e r=" lbfgs " , a l p h a =0)
c l f . f i t (d , l a b e l s )
p = c l f . predict (d)
if ( p == l a b e l s ) . all ( ) :
t r u e _ r e s u l t s += 1
c o n v e r g e d = True
break # short converged
# 50% labelings correct ?
if t r u e _ r e s u l t s >=2∗∗( l _ l e n −1):
break # short success
if index−t r u e _ r e s u l t s > 2 ∗ ∗ (
l _ l e n −1):
break # short fail
if t r u e _ r e s u l t s >= 2 ∗ ∗ ( l _ l e n −1):
d a t a _ r e s . append ( t r u e _ r e s u l t s )
break # short success
d a t a _ r e s . append ( t r u e _ r e s u l t s )
t r u e _ r e s u l t s = max ( d a t a _ r e s )
print ( n , k , h , t r u e _ r e s u l t s ,
true_results ∗1.0/2∗∗ l_len )
if t r u e _ r e s u l t s ∗ 1 . 0 / 2 ∗ ∗ l _ l e n < 0 . 4 5 :
break
Figure 4: Python 2.7 code used for measuring the LM dimension with one hidden layer.
Figure 5: Python 2.7 code used for measuring the MK Dimension with one hidden layer.
# Variable definitions
N = 80 # Maximum number of samples
K = [ 1 , 2 , 3 , 4 ] # Analyzed dimensions
# Analyzed numbers of hidden layers
H = [1 ,2 ,3 ,4 ,5 ,6 ,7 ,8]
# Maximum number of samples ,
max_l = 15 # for random labelings
# Imports
import i t e r t o o l s
import numpy
import random
from s k l e a r n . neural_network \
import M L P C l a s s i f i e r
15
| 9cs.NE
|
ON FINITE COMPLETE REWRITING SYSTEMS, FINITE
DERIVATION TYPE, AND AUTOMATICITY FOR
HOMOGENEOUS MONOIDS
arXiv:1407.7428v2 [math.GR] 14 May 2017
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
Abstract. This paper investigates the class of finitely presented monoids
defined by homogeneous (length-preserving) relations from a computational perspective. The properties of admitting a finite complete rewriting system, having finite derivation type, being automatic, and being
biautomatic are investigated for this class of monoids. The first main
result shows that for any consistent combination of these properties and
their negations, there is a homogeneous monoid with exactly this combination of properties. We then introduce the new concept of abstract
Rees-commensurability (an analogue of the notion of abstract commensurability for groups) in order to extend this result to show that the
same statement holds even if one restricts attention to the class of nary homogeneous monoids (where every side of every relation has fixed
length n). We then introduce a new encoding technique that allows
us to extend the result partially to the class of n-ary multihomogenous
monoids.
1. Introduction
Numerous interesting algebras arise as semigroup algebras K[S], where
K is a field and S is a homogeneous semigroup (that is, a semigroup that
is defined by a presentation where all relations are length-preserving); examples include algebras yielding set-theoretic solutions to the Yang–Baxter
equation and quadratic algebras of skew type (see for example [1, 2, 3] and
[4, 5, 6]), algebras related to Young diagrams, representation theory and
algebraic combinatorics such as the plactic and Chinese algebras (see [7,
Ch. 5], [8, 9] and [10, 11, 12]), and algebras defined by permutation relations
(see [13, 4, 14]). In these examples, there are strong connections between
the structure of the algebra K[S] and that of the underlying semigroup S.
Further motivation for studying this class comes from other important semigroups in the literature that admit homogeneous presentations, such as the
hypoplactic monoid [15], shifted plactic monoid [16], monoids with the same
multihomogeneous growth as the plactic monoid [17], trace monoids [18],
The first author was supported by an FCT Ciência 2008 fellowship and later by an
Investigador FCT fellowship (IF/01622/2013/CP1161/CT0001).
The second author was partially supported by the EPSRC grant EP/N033353/1 ‘Special inverse monoids: subgroups, structure, geometry, rewriting systems and the word
problem’.
This work was partially supported by the Fundação para a Ciência e a Tecnologia (Portuguese Foundation for Science and Technology) through the project
UID/MAT/00297/2013 (Centro de Matemática e Aplicações).
1
2
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
divisibility monoids [19], queue monoids [20], and positive braid monoids
[21, 22].
When investigating a semigroup S defined by homogeneous relations, and
its associated semigroup algebra K[S], a useful first step is to find a good set
of normal forms (canonical representatives over the generating set) for the
elements of the monoid, and thus for elements of the algebra. (See the list of
open problems in [13, Section 3] for more on the importance of this problem
in the context of semigroups defined by permutation relations.) Specifically
we would like a set of normal forms that is a regular language, and we want
to be able to compute effectively with these normal forms. Two situations
where such a good set of normal forms does exist are for monoids that admit
presentations by finite complete rewriting systems (see [23]), and for monoids
and semigroups that are automatic (see [24, 25]). Each of these properties
also has implications for properties of the corresponding semigroup algebra.
Indeed, if the semigroup admits a finite complete rewriting system, then the
semigroup algebra admits a finite Gröbner–Shirshov basis (see [26] for an explanation of the connection between Gröbner–Shirshov bases and complete
rewriting systems), while the automaticity of the semigroup implies that the
algebra is an automaton algebra in the sense of Ufnarovskij; see [27] and [4,
Section 1].
Many of the examples of homogeneous semigroups mentioned above have
been shown to admit presentations by finite complete rewriting systems, and
have been shown to be biautomatic; see for example [28, 29, 30, 31, 32, 33].
It is natural to ask to what extent these results generalise to arbitrary homogeneous semigroups. One can ask: Does every homogeneous semigroup
admit a presentation by a finite complete rewriting system? Is every such
semigroup biautomatic? Within the class of homogeneous semigroups, what
is the relationship between admitting a finite complete rewriting system and
being biautomatic? (For general semigroups, these properties are independent; see [34]). The aim of this paper is to make a comprehensive investigation of these questions. In fact, we shall consider two different strengths of
automaticity, called automaticity and biautomaticity, and we shall also investigate the homotopical finiteness property of finite derivation type (FDT)
in the sense of Squier [35], which is a finiteness property that is satisfied by
monoids that admit presentations by finite complete rewriting systems (full
definitions of all of these concepts will be given in Section 2).
There are various degrees of homogeneity that one can impose on a semigroup presentation. We shall consider finite presentations hA | Ri which
are:
• homogeneous: relations are length-preserving;
• multihomogeneous: for each letter a in the alphabet A, and for every
relation u = v in R, the number of occurrences of the letter a in u
equals the number of occurrences of the letter a in v;
• n-ary homogeneous: there is a fixed global constant n such that for
every relation u = v in R the lengths of the words u and v are both
n;
• n-ary multihomogeneous: simultaneously n-ary homogeneous and
multihomogeneous.
ON HOMOGENEOUS MONOIDS
3
fcrs =⇒ fdt biauto =⇒ auto Example
Y
Y
Y
Y
N
N
N
N
N
N
Y
Y
Y
N
Y
N
N
Y
Y
See
Y
Plactic monoid
[28]
Y
N
Y
Y
Mfcrs
auto
Mfcrs
nonauto
Mfdt
biauto
Mnonfdt
biauto
fdt
Mfcrs
auto ∗ Mbiauto
fdt
Mbiauto ∗ Mfcrs
nonauto
nonfdt
∗
M
Mfcrs
biauto
auto
nonfdt
Mfcrs
nonauto ∗ Mbiauto
Example
Example
Example
Example
3.1
3.6
3.8
3.10
Section 4
Y
N
Y
Section 4
Y
N
N
Section 4
N
N
Y
Section 4
N
N
N
Table 1. Summary of examples of homogeneous monoids
exhibiting all consistent combinations of the properties fcrs,
fdt, biauto, and auto. Examples with the same combinations of properties also exist in the class of n-ary homogeneous monoids (see Section 5).
fcrs =⇒ fdt biauto =⇒ auto Exists? See
Y
Y
Y
Y
Y
[28]
Y
Y
N
N
Y
Y
Y
N
N
N
Y
Y
Y
N
Y
Y
Y
Y
?
Y
Theorem
Theorem
Question
Theorem
5.2
5.2
5.4
5.3
N
Y
N
Y
?
Question 5.4
N
Y
N
N
?
Question 5.4
N
N
N
Y
Y
Theorem 5.3
N
N
N
N
Y
Theorem 5.3
Table 2. Summary of the existence of examples of multihomogeneous monoids with consistent combinations of the
properties fcrs, fdt, biauto, and auto. Examples with
the same combinations of properties also exist in the class of
n-ary multihomogeneous monoids.
Of course, the most restricted class listed here is the class of n-ary multihomogeneous presentations.
For brevity, we introduce the following terminology for the four properties
we are interested in: a monoid is
• fcrs if it admits a presentation via a finite complete rewriting system
(with respect to some finite generating set);
• fdt if it has finite derivation type;
• biauto if it is biautomatic;
• auto if it is automatic.
We will also use the natural negated terms: non-fcrs, non-fdt, non-biauto,
and non-auto.
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
au
t
t
fd
s
o
bi
au
t
r
fc
o
4
Plactic
monoid
Mfdt
biauto
Mfcrs
auto
Mfcrs
nonauto
Mnonfdt
biauto
∗
∗
∗
∗
Figure 1. The semilattice showing the relationship between
examples. By taking the free product of two examples, one
obtains a new monoid whose properties are given by taking
the logical conjunction (that is, the ‘and’ operation) of corresponding properties of the original example monoids. This
corresponds to the meet operation in this semilattice.
We are interested in which combinations of these properties a homogeneous monoid can have. Since in general fcrs implies fdt, and biauto
implies auto, not all combinations will be possible. We refer to any combination of properties that satisfies these restrictions, and that does not contain a property and its negation, as consistent. Our first main result shows
that any consistent combination is possible within the class of homogeneous
monoids. We show this by constructing examples of homogeneous monoids
with each consistent combination of properties. We adopt the following
naming scheme: in the example monoid Mab , the superscript a will be one of
fcrs, fdt, or nonfdt, indicating that the monoid is respectively fcrs (and
thus also fdt), fdt but not fcrs, or non-fdt (and thus also non-fcrs);
while the subscript b will be one of biauto, auto, or nonauto, indicating
that the monoid is respectively biauto (and thus also auto), auto but not
biauto, or non-auto (and thus also non-biauto). In Section 3, we presents
fdt
fcrs
nonfdt
the fundamental examples Mfcrs
auto , Mbiauto , Mnonauto , and Mbiauto . Section
4 contains general results about the behaviour of the various properties under free products of monoids, which we then use to construct the remaining
examples. These results are summarised in Table 1 and the relationship
between the various examples is illustrated in Figure 1.
In Sections 6 and 7 we introduce new concepts and prove new results,
in order to study the combinations of properties can occur in even more
restricted classes. We first introduce and investigate the notion of abstract Rees-commensurability (an analogue of abstract commensurability for
groups [36, §§ iv.27ff.]), which allows us to show that every consistent combination can arise within the class of n-ary homogeneous monoids. (Thus
Table 1 and Figure 1 could also describe the situation for n-ary homogeneous monoids.) We then develop a new encoding technique that embeds
ON HOMOGENEOUS MONOIDS
5
a homogeneous monoid into a 2-generated multihomogeneous monoid. This
encoding technique allows us to obtain most of the consistent combinations
of properties in the class of multihomogeneous or n-ary multihomogenous
monoids. Specifically, it allows us to construct n-ary multihomogeneous
monoids with any possile combination of the properties fcrs, biauto, and
auto, or any combination of the properties fdt, biauto, and auto. However, it does not allow us to construct examples to separate the properties
fcrs and fdt within the class of multihomogeneous or n-ary multihomogeneous monoids. Table 2 summarises the known consistent combinations of
properties in the class of multihomogeneous monoids. Using the results for
abstract Rees commensurability, the same table also describes the situation
for the class of n-ary multihomogeneous monoids.
2. Preliminaries
The subsection below on derivation graphs, homotopy bases and finite
derivation type is self-contained, but it can be complemented with [37, 38].
There is an alternative formulation of the same concepts in terms of strict
monoidal categories/groupoids and higher-dimensional variations of them,
and homotopical algebra in higher categories [39, 40]. However, our approach, using Squier complexes, is the same one used in papers by Otto
[41, 42], Wang [43], Pride and the second and third authors [44], and the
third author [45]; we will require methods and results from these papers in
Sections 4, 6, and 7.
For further information on automatic semigroups, see [25]. We assume
familiarity with basic notions of automata and regular languages (see, for
example, [46]) and transducers and rational relations (see, for example, [47]),
although we will recall some key results that we use frequently. For background on string rewriting systems we refer the reader to [48, 23].
2.1. Words, rewriting systems, and presentations. We denote the
empty word (over any alphabet) by ε. For an alphabet A, we denote by
A∗ the set of all words over A. When A is a generating set for a monoid M ,
every element of A∗ can be interpreted either as a word or as an element of
M . For words u, v ∈ A∗ , we write u = v to indicate that u and v are equal
as words and u =M v to denote that u and v represent the same element of
the monoid M . The length of u ∈ A∗ is denoted |u|, and, for any a ∈ A, the
number of symbols a in u is denoted |u|a . We denote by urev the reversal of
a word u; that is, if u = a1 · · · an−1 an then urev = an an−1 · · · a1 , with ai ∈ A.
If R is a relation on A∗ , then R# denotes the smallest monoid congruence
generated by R.
We use standard terminology and notation from the theory of string
rewriting systems; see [23] or [48] for background reading.
If M is a monoid, a presentation of M is a pair hA | Ri such that M is
isomorphic to the quotient A∗ /R# , in which case, the elements of R are
called the defining relations. We write [u]M for the set of words in A∗ equal
to u in M . The presentation hA | Ri is homogeneous (respectively, multihomogeneous) if for every (u, v) ∈ R and a ∈ A, we have |u| = |v| (respectively,
|u|a = |v|a ). That is, in a homogeneous presentation, defining relations preserve length; in a multihomogenous presentation, defining relations preserve
6
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
the number of each symbol. A monoid is homogeneous (respectively, multihomogeneous) if it admits a homogeneous (respectively, multihomogeneous)
presentation. Note that homogeneous and multihomogeneous presentations
are not required to be finite presentations.
A string rewriting system, or simply a rewriting system, is a pair (A, R),
where A is a finite alphabet and R is a set of pairs (ℓ, r), usually written
ℓ → r, known as rewriting rules or simply rules, drawn from A∗ × A∗ .
The single reduction relation →R is defined as follows: u →R v (where
u, v ∈ A∗ ) if there exists a rewriting rule (ℓ, r) ∈ R and words x, y ∈ A∗
such that u = xℓy and v = xry. That is, u →R v if one can obtain v
from u by substituting the word r for a subword ℓ of u, where ℓ → r is
a rewriting rule. The reduction relation →∗R is the reflexive and transitive
closure of →R . The subscript R is omitted when it is clear from context.
The process of replacing a subword ℓ by a word r, where ℓ → r is a rule, is
called reduction by application of the rule ℓ → r; the iteration of this process
is also called reduction. A word w ∈ A∗ is reducible if it contains a subword
ℓ that forms the left-hand side of a rewriting rule in R; it is otherwise called
irreducible.
The rewriting system (A, R) is finite if both A and R are finite. The
rewriting system (A, R) is noetherian if there is no infinite sequence u1 , u2 , . . .
of words from A∗ such that ui → ui+1 for all i ∈ N. That is, (A, R) is
noetherian if any process of reduction must eventually terminate with an
irreducible word. The rewriting system (A, R) is confluent if, for any words
u, u′ , u′′ ∈ A∗ with u →∗ u′ and u →∗ u′′ , the pair u′ and u′′ resolves, that
is, there exists a word v ∈ A∗ such that u′ →∗ v and u′′ →∗ v. It is well
known that a noetherian system is confluent if and only if all critical pairs
resolve, where critical pairs are obtained by considering overlaps of left-hand
sides of the rewrite rules in R; see [23] for more details. A rewriting system
that is both confluent and noetherian is complete. If a monoid admits a
presentation with respect to some generating set A that forms a finite complete rewriting system R, the monoid is fcrs. In that case, the irreducible
elements form a set of unique normal forms, over A, for the elements of the
monoid.
The Thue congruence ↔∗R is the equivalence relation generated by →R .
The elements of the monoid presented by hA | Ri are the ↔∗R -equivalence
classes. The relations ↔∗R and R# coincide.
Let M be a homogeneous monoid. Let hA | Ri be a homogeneous presentation of M . Without lost of generality, we can assume that R has no
trivial relations of the form a = a′ , for letters a, a′ in A. Since none of the
generators represented by A can be non-trivially decomposed, the alphabet
A represents a unique minimal generating set for M , and any generating
set must contain this minimal generating set. Any two words over A representing the same element of M must be of the same length. So there is
a well-defined function λ : M → N where xλ is defined to be the length
of any word over A representing x. (Here and elsewhere we shall write the
function symbol on the right.) It is easy to see that λ is a homomorphism.
Following [49, Definition 2.1 in §4] the function λ is called a grading of M ,
and so homogeneous monoids are graded monoids.
ON HOMOGENEOUS MONOIDS
7
2.2. Derivation graphs, homotopy bases, and finite derivation type.
Associated with any monoid presentation hA | Ri is a 2-complex D, called
the Squier complex, whose 1-skeleton has vertex set A∗ and edges corresponding to applications of relations from R, and that has 2-cells adjoined
for each instance of “non-overlapping” applications of relations from R (see
below for a formal definition of non-overlapping relations). The free monoid
A∗ acts on D in a natural way via left and right multiplication. A collection
of closed paths in D is called a homotopy base if the complex obtained by
adjoining cells for each of these paths, and those that they generate under
the action of the free monoid on the Squier complex, has trivial fundamental
groups. A monoid defined by a presentation is said to have finite derivation type (or fdt for short) if the corresponding Squier complex admits a
finite homotopy base. It was shown by Squier [35] that the property fdt is
independent of the choice of a finite presentation, so we may speak of fdt
monoids. The original motivation for studying this notion is Squier’s result
[35] which says that if a monoid admits a presentation by a finite complete
rewriting system then the monoid must have finite derivation type. The
study of these concepts is motivated further by the fact that the fundamental groups of connected components of Squier complexes, which are called
diagram groups, have turned out to be a very interesting class of groups;
see [50]. In recent important work [51], acyclic polygraphs have been used
to define a higher-dimensional homotopical finiteness condition for higher
categories. In particular, this work gives rise to a definition of fdtn that
extends the notion of finite derivation type to arbitrary dimensions.
In more detail, with any monoid presentation P = hA | Ri we associate a
graph (in the sense of Serre [52]) as follows. The derivation graph of P is an
infinite graph Γ = Γ(P) = (V, E, ι, τ,−1 ) with vertex set V = A∗ , and edge
set E consisting of the collection of 4-tuples
{(w1 , r, ǫ, w2 ) : w1 , w2 ∈ A∗ , r ∈ R, and ǫ ∈ {+1, −1}}.
The functions ι, τ : E → V associate with each edge E = (w1 , r, ǫ, w2 )
(with r = (r+1 , r−1 ) ∈ R) its initial and terminal vertices ιE = w1 rǫ w2 and
τ E = w1 r−ǫ w2 , respectively. The mapping −1 : E → E associates with each
edge E = (w1 , r, ǫ, w2 ) an inverse edge E−1 = (w1 , r, −ǫ, w2 ).
A non-empty path in Γ is a sequence of edges P = E1 ◦E2 ◦. . .◦En , written
in the diagrammatic order, where τ Ei = ιEi+1 for i = 1, . . . , n − 1. Here P
is a path from ιE1 to τ En and we extend the mappings ι and τ to paths by
defining ιP = ιE1 and τ P = τ En . The inverse of a path P = E1 ◦ E2 ◦ . . . ◦ En
−1
−1
is the path P−1 = E−1
n ◦ En−1 ◦ . . . ◦ E1 , which is a path from τ P to ιP.
A closed path is a path P satisfying ιP = τ P. For two paths P and Q with
τ P = ιQ the composition P ◦ Q is defined.
We denote the set of paths in Γ by P (Γ), where for each vertex w ∈ V
we include a path 1w with no edges, called the empty path at w. The free
monoid A∗ acts on both sides of the set of edges E of Γ by
x · E · y = (xw1 , r, ǫ, w2 y)
where E = (w1 , r, ǫ, w2 ) and x, y ∈ A∗ . This extends naturally to a two-sided
action of A∗ on P (Γ) where for a path P = E1 ◦ E2 ◦ . . . ◦ En we define
x · P · y = (x · E1 · y) ◦ (x · E2 · y) ◦ . . . ◦ (x · En · y).
8
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
ιE1 ιE2
E1 · ιE2
ιE1 · E2
τ E1 ιE2
ιE1 τ E2
τ E1 · E2
E1 · τ E2
τ E1 τ E2
Figure 2. Disjoint derivations in Γ.
If P and Q are paths such that ιP = ιQ and τ P = τ Q then we say that
P and Q are parallel, and write P k Q. We use k to denote the subset of
P (Γ) × P (Γ) of all pairs of parallel paths.
An equivalence relation ∼ on P (Γ) is called a homotopy relation if it is
contained in k and satisfies the following four conditions.
(1) If E1 and E2 are edges of Γ, then
(E1 · ιE2 ) ◦ (τ E1 · E2 ) ∼ (ιE1 · E2 ) ◦ (E1 · τ E2 ).
(2) For any P, Q ∈ P (Γ) and x, y ∈ A∗
P ∼ Q implies x · P · y ∼ x · Q · y.
(3) For any P, Q, R, S ∈ P (Γ) with τ R = ιP = ιQ and ιS = τ P = τ Q
P ∼ Q implies R ◦ P ◦ S ∼ R ◦ Q ◦ S.
(4) If P ∈ P (Γ) then PP−1 ∼ 1ιP , where 1ιP denotes the empty path at
the vertex ιP.
The idea behind condition 1 is the following. Suppose that a word w has
two disjoint occurrences of rewriting rules in the sense that w = αrǫ βα′ rǫ′ ′ β ′
where α, β, α′ , β ′ ∈ A∗ , r, r ′ ∈ R and ǫ, ǫ′ ∈ {−1, +1}. Let E1 = (α, r, ǫ, β)
and E2 = (α′ , r ′ , ǫ′ , β ′ ). Then the paths
P = (E1 · ιE2 ) ◦ (τ E1 · E2 ),
P′ = (ιE1 · E2 ) ◦ (E1 · τ E2 )
give two different ways of rewriting the word w = αrǫ βα′ rǫ′ ′ β ′ to the word
′ β ′ , where in P we first apply the left-hand relation and
w = αr−ǫ βα′ r−ǫ
′
then the right-hand, while in P′ the relations are applied in the opposite
order; see Figure 2. We want to regard these two paths as being essentially
the same, and this is achieved by condition 1. Equivalent paths under this
condition are said to be homotopic by disjoint derivations. This relation is
also often refereed to as the exchange relation or the interchange law in the
literature; see [40].
For a subset C of k, the homotopy relation ∼C generated by C is the
smallest (with respect to inclusion) homotopy relation containing C. The
homotopy relation generated by the empty set ∅ is denoted by ∼0 . If ∼C
coincides with k, then C is called a homotopy base for Γ. The presentation
hA | Ri is said to have finite derivation type if the derivation graph Γ of
hA | Ri admits a finite homotopy base. A finitely presented monoid M is
said to have finite derivation type, or to be fdt, if some (and hence any by
[35, Theorem 4.3]) finite presentation for M has finite derivation type.
ON HOMOGENEOUS MONOIDS
9
It is not difficult to see that a subset C of k is a homotopy base of Γ if
and only if the set
{(P ◦ Q−1 , 1ιP ) : (P, Q) ∈ C}
is a homotopy base for Γ. Thus we say that a set D of closed paths is a
homotopy base if the corresponding set {(P, 1ιP ) : P ∈ D} is a homotopy
base.
2.3. Rational relations. For references purposes, we briefly recall here
some basic definitions and results regarding rational relations; we only consider relations of the form R ⊆ A∗ ×B ∗ . The set of rational relations between
A∗ and B ∗ is the smallest subset of PA∗ × B ∗ that contains the empty set
∅, all singleton sets {(u, v)}, and is closed under the operations of union,
product, and the Kleene star
∞
[
∗
X i.
X 7→ X =
i=0
Note that the set of rational relations is also closed under the Kleene plus
operation X 7→ X + = X ∗ X.
Proposition 2.1 ([47, Examples 5.1 & 5.5]). For any regular language L ⊆
A∗ , the relation
{ (u, u) : u ∈ L } ⊆ A∗ × A∗
is rational.
Proposition 2.2. Let K ⊆ A∗ and L ⊆ B ∗ be regular languages. If R ⊆
A∗ × B ∗ is a rational relation, then R ∩ (K × L) is a rational relation. In
particular, K × L is a rational relation.
Proof. Let T be a transducer recognizing R and let M and N be finite
automata recognizing K and L respectively. Adapt T to simulate M and
N on the inputs from its first and second tapes, respectively, and to accept
only if the simulated copies of M and N are in accept states. This adapted
transducer recognizes R ∩ (K × L).
2.4. Automaticity and biautomaticity.
Definition 2.3. Let A be an alphabet and let $ be a new symbol not in A.
Define the mapping $ : A∗ × A∗ → ((A ∪ {$}) × (A ∪ {$}))∗ by
(u1 · · · um , v1 · · · vn ) 7→
if m = n,
(u1 , v1 ) · · · (um , vn )
(u1 , v1 ) · · · (un , vn )(un+1 , $) · · · (um , $) if m > n,
(u1 , v1 ) · · · (um , vm )($, vm+1 ) · · · ($, vn ) if m < n,
and the mapping
$
: A∗ × A∗ → ((A ∪ {$}) × (A ∪ {$}))∗ by
(u1 · · · um , v1 · · · vn ) 7→
if m = n,
(u1 , v1 ) · · · (um , vn )
(u1 , $) · · · (um−n , $)(um−n+1 , v1 ) · · · (um , vn ) if m > n,
($, v1 ) · · · ($, vn−m )(u1 , vn−m+1 ) · · · (um , vn ) if m < n,
where ui , vi ∈ A.
10
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
Definition 2.4. Let M be a finitely generated monoid. Let A be a finite set
of generators for M and let L ⊆ A∗ be a regular language such that every
element of M has at least one representative in L. For each a ∈ A ∪ {ε},
define the relations
La = {(u, v) : u, v ∈ L, ua =M v}
aL
= {(u, v) : u, v ∈ L, au =M v}.
The pair (A, L) is an automatic structure for M if L$a is a regular language
over (A ∪ {$}) × (A ∪ {$}) for all a ∈ A ∪ {ε}. A monoid M is automatic,
or auto, if it admits an automatic structure with respect to some finite
generating set.
The pair (A, L) is a biautomatic structure for M if L$a , a L$ , $ La , and $a L
are regular languages over (A ∪ {$}) × (A ∪ {$}) for all a ∈ A ∪ {ε}. A
monoid M is biautomatic, or biauto, if it admits a biautomatic structure
with respect to some finite generating set. [Note that biauto implies auto.]
Hoffmann & Thomas have made a careful study of biautomaticity for
semigroups [53]. They distinguish four notions of biautomaticity for semigroups that require at least one of L$a and $ La and at least one of a L$ and $a L
to be regular. These notions are all equivalent for groups and more generally for cancellative semigroups [53, Theorem 1] but distinct for semigroups
[53, Remark 1 & § 4]. biauto clearly implies all four Hoffmann–Thomas
notions of biautomaticity. However, we shall shortly prove that, within the
class of homogeneous monoids, any of the Hoffmann–Thomas notions of
biautomaticity implies biauto (see Proposition 2.7).
In proving that R$ or $ R is regular, where R is a relation on A∗ , a useful
strategy is to prove that R is a rational relation (that is, a relation recognized
by a finite transducer [47, Theorem 6.1]) and then apply the following result,
which is a combination of [54, Corollary 2.5] and [53, Proposition 4]:
Proposition 2.5. If R ⊆ A∗ ×A∗ is rational relation and there is a constant
k such that |u| − |v| ≤ k for all (u, v) ∈ R, then R$ and $ R are regular.
Now we shall prove some results on automaticity and biautomaticity for
the class of homogeneous monoids.
Unlike the situation for groups, both automaticity and biautomaticity for
monoids and semigroups are dependent on the choice of generating set [25,
Example 4.5]. However, for monoids, biautomaticity and automaticity are
independent of the choice of semigroup generating sets [55, Theorem 1.1].
In our particular case of homogeneous monoids, we do have independence
of the choice of generating set:
Proposition 2.6. Let M be a homogeneous monoid that is auto (respectively, biauto). Then for any finite generating set C of M there is a language K over C such that (C, K) is an automatic (respectively, biautomatic)
structure for M .
Proof. We first consider the case for auto. Suppose (B, L) is an automatic
structure for M .
Notice that both the alphabet B and the alphabet C must contain a
subalphabet representing the unique minimal generating set of M . Without
ON HOMOGENEOUS MONOIDS
11
loss of generality, assume that they both contain the alphabet A representing
this minimal generating set. For each b ∈ B, let wb ∈ A∗ be such that
wb =M b. Let Q ⊆ B ∗ × A∗ be the relation
∗
(b, w) : b ∈ B .
Since Q is simply the subset of B ∗ × A∗ obtained by taking the Kleene star
of the finite set of elements of the form (b, wb ), it is by definition a rational
relation. Let
K = L ◦ Q = v ∈ A∗ : (∃u ∈ L) (u, v) ∈ Q .
Let a ∈ A ∪ {ε}. Then
(u, v) ∈ Ka ⇐⇒ u ∈ K ∧ v ∈ K ∧ ua =M v
⇐⇒ (∃u′ , v ′ ∈ L)((u′ , u) ∈ Q ∧ (v ′ , v) ∈ Q ∧ u′ a =M v ′ )
⇐⇒ (∃u′ , v ′ ∈ L)((u′ , u) ∈ Q ∧ (v ′ , v) ∈ Q ∧ (u′ , v ′ ) ∈ La )
⇐⇒ (u, v) ∈ Q−1 ◦ La ◦ Q.
Since a composition of rational relations is rational [47, Theorem 4.4], it
follows that Ka = Q−1 ◦ La ◦ Q is a rational relation. Furthermore
(u, v) ∈ Ka =⇒ ua =M v
=⇒ (ua)λ = vλ
=⇒ uλ + aλ = vλ
=⇒ |u| + 1 = |v|;
thus Ka$ is a regular language by Proposition 2.5.
For c ∈ C − A, let u = u1 · · · um ∈ A∗ be such that u = c; then Kc$ =
(Ku1 ◦ Ku2 ◦ · · · ◦ Kum )$ is regular by [25, Proposition 2.3] and similarly $ Kc
is regular. Hence (C, K) is an automatic structure for M .
For biauto, assume (B, L) is a biautomatic structure for M and follow
the above reasoning to show that each of the languages Ka$ , a K $ , $ Ka , and
$ K are regular.
a
Proposition 2.7. Let M be a homogeneous monoid, let B be a finite generating set for M , and let L be a regular language over B such that every
element of M has at least one representative in L and such that, for each
b ∈ B ∪ {ε}, at least one of L$b and $ Lb and at least one of b L$ and $b L is
regular. Then M is biauto.
Proof. Suppose L$b and b L$ are regular; the other cases are similar.
As in the proof of Proposition 2.6, the alphabet B must contain the unique
minimal generating set A of M . Construct the relation Q ⊆ B ∗ × A∗ as in
the proof of Proposition 2.6. Let K = L ◦ Q.
Let a ∈ A ⊆ B. Then at least one of L$a and $ La and at least one of
$
$
a L and a L is regular. In particular, La and a L are rational relations. So
−1
Ka = Q ◦ La ◦ Q and a K = Q−1 ◦ a L ◦ Q are rational relations. If (u, v)
is in Ka or a K, then |u| + 1 = |v|. Hence Ka$ , $ Ka , a K $ , and $a K are all
regular by Proposition 2.5. Since a ∈ A ∪ {ε} was arbitrary, this proves that
(A, K) is a biautomatic structure for M .
12
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
Despite the positive results obtained so far, note that auto does not
imply biauto in the class of homogeneous monoids, as we shall see below
in Example 3.1.
3. Fundamental examples
3.1. An fcrs, fdt, non-biauto, auto homogeneous monoid. In this
subsection we present a homogeneous monoid that is fcrs and thus fdt,
is auto, but is not biauto. By considering the reversal semigroup of this
example we will get a homogeneous monoid that admits a finite complete
rewriting system but is not automatic.
Example 3.1. Let Mfcrs
auto be the monoid defined by the presentation hA | Ri,
where A = {a, b, c} and R consists of the rewriting rules
cxyz → cxcz,
cbca → cacb.
(x, y, z ∈ {a, b})
Proposition 3.2. The monoid Mfcrs
auto is fcrs.
Proof. The rewriting system (A, R) is noetherian because every rewriting
rule either decreases the number of non-c symbols, or it stays the same and
decreases the number of symbols b to the left of symbols a. To see that it is
confluent, notice that the only overlaps are those between the left-hand side
of cbca → cacb and the left-hand side of a rule of the form cayz → cacz,
where y, z ∈ {a, b}. However, they resolve since we have
(
cacbyz → cacbcz
cbcayz →
, for any y, z ∈ {a, b}.
cbcacz → cacbcz
Therefore (A, R) is confluent.
Proposition 3.3. The monoid
Mfcrs
auto
is auto, but non-biauto.
Proof. Let L be the language of normal forms of (A, R). Since (A, R) is
finite, L is regular [23, Lemma 2.1.3]. Let u ∈ L. Consider the following
cases separately:
(1) uc must also be in normal form, since no left-hand side of a rewriting
rule ends in c. Hence
Lc = { (u, u) : u ∈ L }(ε, c),
and so Lc is rational by Proposition 2.1.
(2) If ub is not in normal form, then ub must end with the left-hand side
of a rewriting rule. Hence u = u′ cxy for some x, y ∈ {a, b}, then
ub = u′ cxyb → u′ cxcb. This word u′ cxcb is in normal form since
u′ cx (which is a prefix of u) is in normal form and no rewriting rule
has left-hand side cxcb for any x ∈ {a, b}. Thus
Lb = { (u, u)(ε, b) : u ∈ L, ub ∈ L }
∪ { (u′ cxy, u′ cxcb) : x, y ∈ {a, b}, u′ cxy ∈ L }
= { (u, u)(ε, b) : u ∈ L, ub ∈ L }
∪ { (u′ , u′ ) : u′ ∈ L }{ (cxy, cxcb) : x, y ∈ {a, b} } ∩ (L × L) ,
and so Lb is rational by Proposition 2.2.
ON HOMOGENEOUS MONOIDS
13
(3) If ua is not in normal form, then ua must end with the left-hand
side of a rewriting rule and so either u = u′ cbc or u = u′ cxy for some
x, y ∈ {a, b}.
(a) If u = u′′ (cb)α y, with y ∈ {a, b, c} and where α ≥ 1 is maximal,
then ua = u′′ (cb)α ya →∗ u′′ ca(cb)α , since ua = u′′ (cb)α ya →
u′′ (cb)α ca when y ∈ {a, b}, and then cbca → cacb. Now, u′′ ca(cb)α
is in normal form since u′′ and ca(ab)α are in normal form and
the only left-hand side of a rewriting rule of that ends in ca is
cbca, and α is maximal.
(b) If u = u′ cay, with y ∈ {a, b}, then ua = u′ caya → u′ caca and
this word is in normal form since u′ ca is in normal form.
Therefore
La = { (u, ua) : u ∈ L, ua ∈ L }
∪ { (u′′ (cb)α y, u′′ ca(cb)α ) : α ∈ N, y ∈ {a, b, c}, u′′ (cb)α y ∈ L, u′′ ∈
/ A∗ cb }
∪ { (u′ cay, u′ caca) : y ∈ {a, b}, u′ cay ∈ L }
= { (u, u)) : u ∈ L }(ε, a) ∩ (L × L)
∪ { (u′′ u′′ ) : u′′ ∈ L \ (A∗ cb) }(ε, ca)(cb, cb)+ { (y, ε) : y ∈ {a, b, c} } ∩ L × L
∪ { (u′ , u′ ) : u′ ∈ L }{ (cay, caca) : y ∈ {a, b} } ∩ L × L ,
and so La is a union of relations, each of which is rational by Proposition
2.1 and 2.2.
Note also that Lε = {(u, u) : u ∈ L} is rational. Hence Lx is a rational
relation for any x ∈ A ∪ {ε}. Moreover, if (u, v) lies in one of these relations,
then |u| − |v| ≤ 1 and so L$x is regular for all x ∈ A ∪ {ε} by Propositions
2.5 and 2.2. Hence Mfcrs
auto is auto.
Suppose, with the aim of obtaining a contradiction, that Mfcrs
auto is biauto.
Then by Proposition 2.6 it admits a biautomatic structure (A, L). Thus L is
$
a regular language mapping onto Mfcrs
auto and c L is regular. This contradicts
Lemma 3.4 below.
Lemma 3.4. There is no regular language L ⊆ A∗ such that L maps onto
$
Mfcrs
auto and c L is regular.
Proof. Suppose, with the aim of obtaining a contradiction, that such a language L exists. Then (c L ◦ c L−1 )$ is regular. Let n be an even number
exceeding the number of states in an automaton recognizing (c L ◦ c L−1 )$ .
Observe that
−1
= {(u, v) ∈ L : cu =M cv}.
cL ◦ cL
Notice that an bn+1 is not represented by any other word over A and similarly
bn an b is not represented by any other word over A. So an bn+1 , bn an b ∈ L.
Furthermore,
can bn+1 →∗ (ca)n/2 (cb)(n/2)+1
and
cbn an b →∗ (cb)n/2 (ca)n/2 cb
→∗ (ca)n/2 (cb)(n/2)+1
14
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
and so (an bn+1 , bn an b) ∈ c L ◦ c L−1 . Since n exceeds the number of states in
an automaton recognizing (c L ◦ c L−1 )$ , we can apply the pumping lemma to
a segment of the word that lies within the first n letters of (an bn+1 , bn an b)$
(that is, to a subword of the form (ak , bk )$ for some k ≥ 1) to see that
(an+ik bn+1 , bn+ik an b) ∈ c L ◦ c L−1 for some k ≥ 1 and for all i ∈ N. Hence,
by definition of the relation c L, we have can+ik bn+1 =M cbn+ik an b. But
can+2k bn+1 →∗ (ca)n/2+k (cb)(n/2)+1
and
cbn+2k an b →∗ (cb)n/2+k (ca)n/2 cb
→∗ (ca)n/2 (cb)(n/2)+k+1 ,
and so the normal forms of can+ik bn+1 and cbn+ik an b are unequal; this
contradicts the previous equality. Thus M is not biauto.
3.2. An fcrs, fdt, non-biauto, non-auto homogeneous monoid.
Definition 3.5. Let S be a monoid defined by a presentation hA | Ri. Denote by M rev the monoid defined by the presentation hA | Rrev i, where
Rrev = {(lrev , r rev ) : (l, r) ∈ R}, that is called the reversal monoid of M .
[Note that (M rev )rev ≃ M .]
fcrs rev , which is defined by the presenExample 3.6. Let Mfcrs
nonauto = (Mauto )
tation hA | Rrev i, where hA | Ri is the presentation defining Example 3.1.
rev i we can argue as in the proof
Since Mfcrs
nonauto is presented by hA | R
of Proposition 3.2 that the rewriting system is noetherian and that the
overlaps, that result in critical pairs, resolve in a similar way. Thus Mfcrs
nonauto
is also fcrs and thus fdt.
Proposition 3.7. Mfcrs
nonauto is non-auto.
Proof. Suppose, with the aim of obtaining a contradiction, that Mfcrs
nonauto
$
is auto. Let (A, L) be an automatic structure for Mfcrs
nonauto . Then La is
fcrs
regular for all a ∈ A ∪ {ε}. Since Mnonauto is homogeneous, if (u, v) ∈ La ,
then |u| − |v| ≤ 1 and so $ La is regular by [56, Corollary 4.2]. Notice
$
that ($ La )rev = a (Lrev ) . Hence Lrev is a regular language mapping onto
$
rev
rev
≃ Mfcrs
Mfcrs
such that a (Lrev ) is regular. Since Mfcrs
nonauto
auto ,
nonauto
fcrs
this contradicts Lemma 3.4 and so Mnonauto is indeed non-auto.
3.3. A non-fcrs, fdt, biauto, auto homogeneous monoid. The following homogeneous monoid was introduced by Katsura and Kobayashi [57,
Example 3], who showed that it is non-fcrs, but is fdt. We shall prove
that it is biauto and thus auto.
Example 3.8. Let A = {a, bi , ci , di : i = 1, 2, 3} and let R consist of the
rewriting rules
(3.1)
bi a → abi
i ∈ {1, 2, 3},
(3.2)
cj bj → c1 b1
j ∈ {2, 3},
(3.3)
bj dj → b1 d1
j ∈ {2, 3}.
ON HOMOGENEOUS MONOIDS
15
fdt
Let Mfdt
biauto = hA | Ri. Then Mbiauto is fdt [57, § 4] but is non-fcrs [57,
Proposition 3].
Proposition 3.9. The monoid Mfdt
biauto of Example 3.8 is biauto and thus
auto.
Proof. Let S consist of the following rewriting rules:
(3.4)
bi a → abi
(3.5)
cj ak bj → c1 ak b1
k
i ∈ {1, 2, 3},
k
(3.6)
c1 a b1 dj → cj a b1 d1
(3.7)
bj dj → b1 d1
j ∈ {2, 3}, k ∈ N ∪ {0},
j ∈ {2, 3}, k ∈ N ∪ {0},
j ∈ {2, 3}.
Notice that every rule in S is a consequence of those in R. Indeed, using
rules in R, we have
cj ak bj ↔ cj ak−1 bj a ↔ . . . ↔ cj bj ak ↔ c1 b1 ak ↔ c1 ab1 ak−1 ↔ . . . ↔ c1 ak b1 ;
and
c1 ak b1 dj ↔ . . . ↔ c1 b1 ak dj ↔ cj bj ak dj ↔ . . . ↔ cj ak bj dj ↔ cj ak b1 d1 .
Consider any ordering < of A satisfying b1 < bj < a < d1 < dj (for
j ∈ {2, 3}). By [48, Lemma 2.4.3] the right-to-left length-plus-lexicographic
order induced by < is noetherian. Moreover, any rewriting using a rule in S
decreases a word with respect to this ordering. Thus the rewriting system
(A, S) is noetherian.
To see that (A, S) is confluent, notice that there are two possible overlaps
of left-hand sides of rewriting rules: an overlap of (3.4) and (3.5), and an
overlap of (3.5) and (3.7). However, critical pairs resolve, since
(
cj ak+1 bj → c1 ak+1 b1
k
cj a bj a →
c1 ak b1 a → c1 ak+1 b1
and
(
cj ak b1 d1
cj ak bj dj →
c1 ak b1 dj → cj ak b1 d1
.
Let L be the language of S-irreducible words of A∗ . That is,
L = A∗ − A∗ {b1 a, b2 a, b3 a} ∪ {c2 , c3 }a∗ {b2 , b3 }
∪ c1 a∗ b1 {d2 , d3 } ∪ {b2 d2 , b3 d3 } A∗ ;
thus L is regular. To prove that (A, L) is an automatic structure for Mfdt
biauto,
we first show that Lx and x L are rational relations for all x ∈ A ∪ {ε}. Since
L is a cross-section for Mfdt
biauto , the relations Lε and ε L are both equal to
the equality relation and hence are trivially rational.
So let x ∈ A and w ∈ L. Suppose first that x ∈ {b2 , b3 }, then there may be
a left-hand side of a rewriting rule of type (3.5) at the rightmost end of the
word wx. In this case, w must be of the form w′ cj ak for some w′ ∈ L (since
a prefix of a irreducible word is irreducible), j ∈ {2, 3} and k ∈ N ∪ {0};
applying the rewriting rule yields w′ c1 ak b1 , which is in normal form since
16
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
no left-hand side of a rewrite rule contains c1 except for (3.6), which clearly
cannot be applied. So when x ∈ {b2 , b3 }, at most one application of a rewrite
rule at the rightmost end turns wx into a normal form word. Hence
Lbj = (w, wx) : w ∈ L \ Lcj a∗
∪ (w′ cj ak , w′ c1 ak b1 ) : w′ ∈ L, k ∈ N ∪ {0}
= (w, w) : w ∈ L \ Lcj a∗ (ε, x)
∪ (w′ , w′ ) : w′ ∈ L (cj , c1 )(a, a)∗ (ε, b1 ).
Hence each Lbj is a rational relation by Proposition 2.2.
Suppose now that x ∈ {d2 , d3 }. Reasoning similar to the previous paragraph shows that wx is in normal form, or one application of a rewrite rule
of type (3.6) or (3.7) turns it into normal form. (Note that an application
of a rule of type (3.7) might be followed by one of type (3.5), but these can
be replaced by one of type (3.6).) Hence
Ldj = (w, wx) : w ∈ L \ (Lc1 a∗ b1 ∪ Ld2 ∪ Ld3 )
∪ (w′ c1 ak b1 , w′ cj ak b1 d1 ) : w′ ∈ L, k ∈ N ∪ {0}
∪ (w′ bj , w′ b1 d1 ) : w′ ∈ L \ Lc1 a∗
= (w, w) : w ∈ L \ (Lc1 a∗ b1 ∪ Ld2 ∪ Ld3 ) (ε, x)
∪ (w′ , w′ ) : w′ ∈ L (c1 , cj )(a, a)∗ (b1 , b1 )(ε, d1 )
∪ (w′ , w′ ) : w′ ∈ L \ Lc1 a∗ (bj , b1 )(ε, d1 ).
Hence each Ldj is a rational relation by Proposition 2.2.
Suppose that x = a. Then the only rewriting rules that can apply to
/
wx are a sequence of rule of type (3.4), rewriting w′ ybi1 · · · bik a (where y ∈
{b1 , b2 , b3 }) to w′ yabi1 · · · bik . This resulting word is in normal form, since
the only way a rewriting rule could apply was if y = ci1 , but this means the
word w would contain ci1 bi1 , which contradicts w ∈ L. Hence
La = (w′ ybj1 · · · bjk , w′ yabj1 · · · bjk ) : w′ y ∈ L, bj1 , . . . , bjk ∈ {1, 2, 3}, k ∈ N ∪ {0}
= (w′ y, w′ y) : w′ y ∈ L (ε, a){ (bj , bj ) : j ∈ {1, 2, 3} }∗ .
Finally, if x ∈ {b1 , c1 , c2 , c3 , d1 }, then wx is already in normal form: hence,
in this case,
Lx = { (w, w)(ε, x) : w ∈ L }.
In each case, Lx is a rational relation. Since Mfdt
biauto is homogeneous, if
(u, v) ∈ Lx for x ∈ A, then |v| = |ua| = |u| + 1. Furthermore, if (u, v) ∈ Lε ,
then |u| = |v|. Hence L$x and $ Lx are regular for all x ∈ A ∪ {ε} by
Proposition 2.5.
Similar reasoning shows that x L is a rational relation: if x ∈ {b1 , b2 , b3 }
and w ∈ L, then rewriting xw to normal form can consist of a sequence
of applications of rules of type (3.4) followed possibly by one of type (3.3);
for all other x, at most one rewriting rule is required. Proposition 2.5 then
applies to show that x L$ and $x L are regular.
Hence (A, L) is a biautomatic structure for Mfdt
biauto .
ON HOMOGENEOUS MONOIDS
17
3.4. A non-fcrs, non-fdt, biauto, auto homogeneous monoid. In
this section we give an example of a homogeneous monoid that is non-fdt
and thus non-fcrs, but which is biauto and thus auto.
Example 3.10. Let A = {a, b} and let R be the rewriting system on A∪{c}
consisting of the three rules:
Ka : ac → ca
Kb : bc → cb
C
: cab → cbb.
Let P be the presentation hA ∪ {c} | Ri, and let Mnonfdt
biauto be the monoid
presented by P.
Theorem 3.11. The monoid Mnonfdt
biauto :
∗
+
∗
∗
(1) has A ∪ c b a as set of unique normal forms (that is, it is a set,
over the generating set A, in one-to-one correspondence with the
elements of Mnonfdt
biauto );
(2) is biauto and thus auto;
(3) is non-fdt and thus non-fcrs.
Part 1 of Theorem 3.11 will follow from Lemma 3.13 below, and part 2 is
proved in Lemma 3.14. Then, the rest of the subsection will be devoted to
proving that Mnonfdt
biauto is non-fdt, thus establishing part 3.
Remark 3.12. The methods we use here to prove that Example 3.10 is notfdt are similar to those used in the proof of [44, Theorem 1]. In particular
we will use the notion of critical peaks, and resolution of critical peaks, in
our proof. We refer the reader to [44, Section 2] for the definitions of these
concepts, and their connection with complete rewriting systems and fdt.
Let us begin by fixing some of the notation. We start by adding to P
infinitely many rules of the form
C u : cuab → cubb (u ∈ A∗ )
and denote by R′ the set of all these rules. Notice first that C ǫ is precisely
the rule C defined above and that, for any word u ∈ A∗ the words cuab
and cubb represent the same element of the monoid Mnonfdt
biauto , since in the
word cuab we can use relations of the form Kx to pass the letter c through
the word u from left to right, then replace cab by cbb using the relation
C, and finally move the c back through u again from right to left using
the relations Kx . It follows that the presentations P = hA ∪ {c} | Ri and
P = hA ∪ {c} | R ∪ R′ i are equivalent presentations, in the sense that two
words u, v ∈ (A ∪ {c})∗ are equivalent modulo the relations R if and only if
they are equivalent modulo the relations R ∪ R′ . In particular, the monoid
Mnonfdt
biauto is also defined by the infinite presentation
P = A ∪ {c} | R ∪ R′ .
Lemma 3.13. The infinite presentation P is a complete presentation of
Mnonfdt
biauto . The set of irreducible words with respect to this complete rewriting
system is
A∗ ∪ c+ b∗ a∗ = A∗ ∪ { ci bj cl : i ∈ N, j, k ∈ N ∪ {0} }.
18
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
Proof. The fact that P is a presentation for Mnonfdt
biauto follows from the comments made before the statement of the lemma. By considering the (left-toright) length-plus-lexicographic ordering on {a, b, c}∗ induced by a > b > c
one sees that the rewriting system P is noetherian.
The set of irreducible words under this rewriting system is the set A∗ ∪
c+ b∗ a∗ . Indeed, if a irreducible word contains a symbol c, it cannot be to
the left of a symbol from A, otherwise we could apply a relation Ka or Kb .
Moreover, if the word also contains a symbol b, then all symbols a must
be to the right of the rightmost symbol b, since otherwise we could use a
relation of the form C u .
Finally, to prove that P is confluent it suffices to consider all possible
overlaps between left-hand sides of the rewriting rules Kx (x ∈ A) and C u
(u ∈ A∗ ), showing that all critical peaks arising from these overlaps resolve
(see [44, Section 2]). There are three different ways in which these rewrite
rules can overlap, giving rise to three types of critical peaks, all of which
can be resolved; see Figure 3. This proves that P is confluent and thus
completes the proof of the lemma.
Lemma 3.14. The monoid Mnonfdt
biauto is biauto.
Proof. Let L = A∗ ∪ c+ b∗ a∗ . We will prove that (A∪ {c}, L) is a biautomatic
structure for Mnonfdt
biauto . By the previous lemma, L is a regular language such
that every element of Mnonfdt
biauto has a unique representative in L. Hence
Lε = { (w, w) : w ∈ L };
thus Lε is a rational relation.
Now let u ∈ L. Regardless of whether u ∈ A∗ or u ∈ c+ b∗ a∗ , the word ua
also lies in L. Hence
La = { (u, ua) : u ∈ L } = { (u, u) : u ∈ L }(ε, a)
is a rational relation by Proposition 2.1.
If u ∈ A∗ , then ub also lies in L. On the other hand, if u = ci bj ak , then
ub = ci bj ak b →∗ ci bj+k+1 via a sequence of applications of rules C u . Hence
Lb = { (u, ub) : u ∈ A∗ } ∪ { (ci bj ak , ci bj+k+1 ) : i ≥ 1, j, k ≥ 0 }
= { (u, u) : u ∈ A∗ }(ε, b) ∪ (c, c)+ (b, b)∗ (a, b)∗ (ε, a)
is a rational relation by Proposition 2.1.
If u = ak , then uc = ak c →∗ cak using a sequence of applications of rules
Ka . If u ∈ A∗ and u contains at least one symbol b, then u = u′ bak for
′
some u′ ∈ A∗ and k ∈ N ∪ {0} and so uc = u′ bak c →∗ cu′ bak →∗ cb|u |+1 ak
by a sequences of applications of rules of Ka and Kb and a sequence of
applications of rules C u . Hence
Lc = { (ak , cak ) : k ≥ 0 }
∪ { (ubak , cb|u|+1 ak ) : u ∈ A∗ , k ≥ 0 }
∪ { (ci bj ak , ci+1 bj ak ) : i ≥ 1, j, k ≥ 0 }
= (ε, c)(a, a)∗
∪ (ε, c){(a, b), (b, b)}∗ (b, b)(a, a)∗
∪ (ε, c)(c, c)+ (b, b)∗ (a, a)∗
ON HOMOGENEOUS MONOIDS
19
is a rational relation.
Similar reasoning shows that
aL
= { (u, au) : u ∈ A∗ }
∪ { (ci ak , ci ak+1 ) : i ≥ 1, k ≥ 0 }
∪ { (ci bj ak , ci bj+1 ak ) : i, j ≥ 1, k ≥ 0 }
= (ε, a){(a, a), (b, b)}∗
∪ (c, c)+ (a, a)∗ (ε, a)
∪ (c, c)+ (b, b)+ (ε, b)(a, a)∗ ;
bL
= { (u, bu) : u ∈ A∗ }
∪ { (ci bj ak , ci bj+1 ak ) : i ≥ 1, j, k ≥ 0 }
= (ε, b){(a, a), (b, b)}∗
∪ (c, c)+ (ε, b)(b, b)+ (aa, a)∗
cL
= { (ak , cak ) : k ≥ 0 };
∪ { (ubak , cb|u|+1 ak ) : u ∈ A∗ , k ≥ 0 }
∪ { (ci bj ak , ci+1 bj ak ) : i ≥ 1, j, k ≥ 0 }
= (ε, c)(a, a)∗
∪ (ε, c){(a, b), (b, b)} ∗ (b, b)(a, a)∗
∪ (ε, c)(c, c)+ (b, b)∗ (a, a)∗ ;
thus a L, b L, and c L are all rational.
Since (u, v) ∈ x L ∪ Lx for any x ∈ A ∪ ε implies |u| − |v| ≤ 1 , Proposition
2.5 shows that their images under $ and $ are regular. Hence (A, L) is a
biautomatic structure for Mnonfdt
biauto .
Let Γ denote the derivation graph of P, and Γ the derivation graph of P.
Let ΓZ denote the connected components of Γ with vertex set the set of all
words in A ∪ {c} with at least two occurrences of the letter c. Likewise let
ΓZ be the connected component of Γ with the same vertex set as ΓZ .
There are three infinite families (CT 1), (CT 2) and (CT 3) of closed paths
in Γ, as displayed in Figure 3, that correspond to resolutions of all critical
peaks. Each such closed path we obtain we shall call a critical circuit. Let us
denote by C the critical circuits of the form (CT 1) and (CT 3) and denote by
Z the critical circuits of the form (CT 2). Observe that the critical circuits
in Z are in ΓZ since the words labelling vertices in (CT 2) all contain two
occurrences of the letter c. The set of critical circuits C ∪ Z forms an infinite
homotopy base for Γ (see [44, Lemma 3]).
We now want to use the infinite homotopy base C ∪ Z for Γ to obtain an
infinite homotopy base for Γ. In order to do this we need to take the critical
circuits (CT 1)–(CT 3) and transform them into circuits in the derivation
graph Γ by replacing each occurrence of an edge C u by a corresponding
path in Γ.
As mentioned above when proving that P and P are equivalent presentations, the edges C u can be realized in Γ by paths Cu which are defined
20
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
cuabc
xcuab
Kx · uab
cxuab
x · Cu
(CT 1)
xcubb
Kx · ubb
C xu
cua · Kb
Cu · c
cxubb
cubbc
(CT 2)
cuacb
cub · Kb
cu · Ka · b
cubcb
cucab
cu · C
cu · Kb · b
cucbb
cuabvab
C uabv
C u · vab
cubbvab
(CT 3)
cuabvbb
C ubbv
C u · vbb
cubbvbb
Figure 3. Resolutions of the critical peaks in the derivation
graph Γ of the presentation P. Here x, y ∈ A and u, v ∈ A∗ .
inductively as follows: we first set Cǫ to be the rule C, and then for u = xu′ ,
with x ∈ A and u′ ∈ A∗ we set Cu to be the path
(3.8)
Cu :
K −1 ·u′ ab
x·C
K ·u′ bb
x
x
cxu′ ab −−−
−−−→ xcu′ ab −−−u→ xcu′ bb −−−
−−→ cxu′ bb.
′
So, Cu is the path in Γ from cuab to cubb given by commuting c through u
using the relations Kx , applying the relation C to transform ucab into ucbb,
and then commuting c back through u again using the relations Kx , ending
at the vertex cubb.
Now let us define a mapping ϕ from the set of paths P (Γ) in Γ to the set
of paths P (Γ) in Γ. Let ϕ : P (Γ) → P (Γ) be the map given by (α·C u ·β)ϕ =
α · Cu · β, for all α, β ∈ (A ∪ {c})∗ and u ∈ A∗ , and defined to be the identity
on every other edge of Γ. Let C = (C)ϕ and Z = (Z)ϕ. Since C ∪ Z forms a
homotopy base for Γ it follows that C ∪ Z is an infinite homotopy base for
Γ (see [35, Corollary 3.7]).
Observe that the infinite homotopy base C ∪ Z for Γ is nothing more than
the set of circuits in Γ obtained by taking the set of circuits (CT 1)–(CT 3)
and replacing each occurrence of the edge C u by the path Cu defined in (3.8).
Let us denote this corresponding set of circuits C ∪ Z in Γ by (CT 1)–(CT 3).
The monoid Mnonfdt
biauto is presented by the finite presentation P and Γ is
the derivation graph of P which has an infinite homotopy base C ∪ Z.
Lemma 3.15 ([35, Discussion before Definition 3.3]). If Mnonfdt
biauto were fdt
then there would be a finite subset C0 ∪ Z0 of C ∪ Z which would be a finite
homotopy base for Γ.
Sketch. Since D is a finite homotopy base, each path in D is homotopic to an
empty path using finitely many paths from C ∪ Z, and thus the finite subset
of C ∪ Z consisting of all paths arising this way is a homotopy base.
ON HOMOGENEOUS MONOIDS
21
Our aim now is to show that this leads to a contradiction, and thus
conclude that Mnonfdt
biauto is not fdt. In order to do this we shall now define a
mapping from the set of paths P (Γ) in Γ into the integral monoid ring ZM.
Define Φ : P (Γ) → ZM to be the unique map which extends the mapping:
• (α · Ka · β)Φ = α, (α · Kb · β)Φ = α, and
• (α · C · β)Φ = α,
nonfdt
where α, β ∈ (A ∪ {c})∗ and α ∈ Mnonfdt
biauto denotes the element of Mbiauto
represented by the word α, to paths in such a way that
(P ◦ Q)Φ = (P)Φ + (Q)Φ and (P−1 )Φ = −(P)Φ.
The following basic properties of Φ are then easily verified for all paths
P, Q ∈ P (Γ) and words α, β ∈ (A ∪ {c})∗ :
(1) (α · P · β)Φ = α · (P)Φ
(2) (P ◦ P−1 )Φ = 0
(3) ([P, Q])Φ = 0 where
[P, Q] = (P · ιQ) ◦ (τ P · Q) ◦ (P−1 · τ Q) ◦ (ιP · Q−1 ).
(4) If P ∼0 Q then (P)Φ = (Q)Φ.
Here, property 4 follows from properties 1, 2 and 3. Note that property 4
implies that Φ induces a well-defined map on the homotopy classes of paths
of Γ.
In what follows we shall often omit bars from the top of words in the
images under Φ and simply write words from (A ∪ {c})∗ with the obvious
intended meaning. Recall that C is the set of critical circuits of the form
(CT 1) and (CT 3), and that C = (C)ϕ is the corresponding set of circuits in
Γ. Let F = A∗ denote the free monoid on the alphabet A.
Lemma 3.16. If Mnonfdt
biauto is fdt then the submodule h(C)ΦiZF , of the left
ZF -module ZF , generated by (C)Φ is a finitely generated left ZF -module.
Proof. Assume that Mnonfdt
biauto is fdt and therefore P is fdt. Since C ∪ Z is
a homotopy base for its derivation graph Γ, by Lemma 3.15 there are finite
subsets C0 ⊆ C and Z0 ⊆ Z such that C0 ∪ Z0 is a finite homotopy base for Γ.
Let P ∈ C be arbitrary. We claim that (P)Φ ∈ h(C0 )ΦiZF . Once established,
this will prove the lemma, since (C0 )Φ is a finite subset of h(C)ΦiZF .
By [44, Lemma 2], since P is a closed path in Γ and C0 ∪ Z0 is a homotopy
base for Γ, we can write
(3.9)
−1
P ∼0 P−1
1 ◦ (α1 · Q1 · β1 ) ◦ P1 ◦ · · · ◦ Pn ◦ (αn · Qn · βn ) ◦ Pn ,
where each Pi ∈ P (Γ), αi , βi ∈ (A ∪ {c})∗ and Qi ∈ (C0 ∪ Z0 )±1 .
Since the vertices of P have exactly one c, and all the relations in the
presentation P involve the letter c, it follows that αi , βi ∈ A∗ and Qi ∈ C0 ,
for all i ∈ {1, . . . , n}. Applying Φ gives
(P)Φ = α1 (Q1 )Φ + · · · + αn (Qn )Φ ∈ h(C0 )ΦiZF
as claimed.
To complete our proof, it remains to compute the subset (C)Φ of ZF and
then prove that the submodule h(C)ΦiZF of ZF is not finitely generated as
a left ZF -module. Recall that C is the set of critical circuits of the form
22
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
(CT 1) and (CT 3), and that C = (C)ϕ. So C is the set of closed paths (CT 1)
and (CT 3) obtained by applying the mapping ϕ to the closed paths (CT 1)
and (CT 3), that is, obtained by taking each occurrence of C u and replacing
it by the path Cu .
From equation (3.8) we have, for any word u ∈ A∗ , if u = xu′ , with x ∈ A,
the equality (Cu )Φ = −1M +x·(Cu′ )Φ+1M = x·(Cu′ )Φ, and thus we deduce
that (Cu )Φ = u. Using this fact, the result of computing a critical circuit in
C under the map Φ is given by:
• (P)Φ = 0ZM , for a critical circuit P of the family (CT 1);
• (P)Φ = u(ab − bb)v, for a critical circuit P of the family (CT 3).
The next lemma completes the proof of Theorem 3.11.
Lemma 3.17. The submodule h(C)ΦiZF of ZF , where
(C)Φ = {u(ab − bb)v | u, v ∈ A∗ } ∪ {0ZM },
is not finitely generated as a left ZF -module, and therefore Mnonfdt
biauto is not
fdt.
Proof. Suppose, with the aim of obtaining a contradiction, that h(C)ΦiZF is
finitely generated as a left ZF -module. Then there exists a finite subset X
of (C)Φ such that hXiZF ⊇ (C)Φ. Let n ∈ N be the maximal length of a
word v where u(ab − bb)v ∈ X. We shall show that (ab − bb)an+1 belongs to
(C)Φ but not to hXiZF . Suppose that in ZF we have
n+1
(ab − bb)a
=
k
X
αi (ab − bb)vi ,
i=1
where αi , vi ∈ A∗ and |vi | ≤ n, for i = 1, . . . , k. Then for some j ∈ {1, . . . , k}
we have either bban+1 = αj abvj or bban+1 = αj bbvj in the free monoid
F = {a, b}∗ , which clearly contradicts the fact that |vi | ≤ n. We conclude
that (C)Φ is not finitely generated as a left ZF -module, and it then follows
from Lemma 3.16 that Mnonfdt
biauto is not fdt.
4. Free products of homogeneous monoids
In Section 3, we gave four examples of homogeneous monoids that possess
certain combinations of the properties fcrs, fdt, biauto, and auto. In
this section, we use free products to construct examples with the remaining
consistent combinations of these properties. Note that if monoids M1 and
M2 have homogeneous presentations hA1 | R1 i and hA2 | R2 i, then their free
product M1 ∗ M2 is defined by the presentation hA1 ∪ A2 | R1 ∪ R2 i and is
thus also homogeneous.
First we consider the interaction of the free product with biauto and
auto. It is known that the free product of two monoids is auto if and
only if each of the monoids is auto ([25, Theorem 6.2], [55, Theorem 1.2]).
While it would be possible to extend this result to biautomaticity for general
monoids, this generalization does not appear in the literature. This paper
required the biautomaticity result only for homogeneous monoids, and the
proofs are simpler in this case:
ON HOMOGENEOUS MONOIDS
23
Proposition 4.1. Let M1 and M2 be homogeneous monoids. Then M1 ∗
M2 is auto (respectively, biauto) if and only if M1 and M2 are auto
(respectively, biauto).
Proof. Let hA1 | R1 i and hA2 | R2 i be homogeneous presentations for M1
and M2 , respectively.
Suppose M1 ∗M2 is auto; the proof for biauto is similar. By Proposition
2.6, M1 ∗ M2 admits an automatic structure (A1 ∪ A2 , L). Since M1 is
homogeneous, there is no non-empty word over A1 representing the identity
of M1 ; hence every word in (A1 ∪ A2 )∗ representing an element of M2 must
lie in A∗2 . Thus (A2 , L ∩ A∗2 ) is an automatic structure for M2 ; similarly,
(A1 , L ∩ A∗1 ) is an automatic structure for M1 .
On the other hand, suppose M1 and M2 are auto; again, the proof for
biauto is similar. Then there are automatic structures (A1 , L1 ) and (A2 , L2 )
for M1 and M2 respectively. By [25, Corollary 5.5], assume without loss of
generality that every element of M1 and M2 has a unique representative in
L1 and L2 , respectively. Let
(4.1) L = {ε}
∪
(4.2)
∪
(4.3)
∪
(4.4)
∪
(4.5)
+
+
+
+
L1 ∩ (A+
×
A
)
L
∩
(A
×
A
)
2
1
1
2
2
∗
+
+
+
+
L1 ∩ (A+
×
A
)
L
∩
(A
×
A
)
L1 ∩ (A+
2
1
1
2
2
1 × A1 )
+
+
+
+
L2 ∩ (A+
×
A
)
L
∩
(A
×
A
)
1
2
2
1
1
∗
+
+
+
+
L2 ∩ (A+
×
A
)
L
∩
(A
×
A
)
L2 ∩ (A+
1
2
2
1
1
2 × A2 ) ;
note that L is a disjoint union of the languages (4.1)–(4.5). (Note that the
language L is not equal to (L1 L2 )∗ : every word in L is a product of words
strictly alternating between non-empty words from L1 and words from L2 .)
Since every element of M1 ∗ M2 is represented by a unique element of L,
Lε = { (u, u) : u ∈ L }
is a rational relation.
Now let a ∈ A1 , and let va be the unique word in L1 that is equal to a
(note that va will either be a itself or another generator that is equal to a).
Suppose that u ∈ L lies in (4.2) or (4.5). Then uva is the unique word in L
that is equal to ua. On the other hand, if u ∈ L lies in (4.3) or (4.4), then
it is of the form u′ s, where s ∈ L2 and s is the maximal suffix of u lying
in A∗1 . In this case, the unique word in L equal to ua is u′ t, where t is the
unique word in L1 equal to sa: that is, where (s, t) ∈ (L1 )a . Hence
La = {(ε, va )}
+
+
+
∪ (L1 )ε ∩ (A+
1 × A1 ) (L2 )ε ∩ (A2 × A2 ) (ε, va )
∗
+
+
+
+
∪ (L1 )ε ∩ (A+
×
A
)
(L
)
∩
(A
×
A
)
(L1 )a ∩ (A+
2 ε
1
1
2
2
1 × A1 )
∗
+
+
+
+
+
+
∪ (L2 )ε ∩ (A+
×
A
)
(L
)
∩
(A
×
A
)
(L
)
∩
(A
×
A
)
(L1 )a ∩ (A+
1 ε
2 ε
2
2
1
1
2
2
1 × A1 )
∗
+
+
+
+
+
+
∪ (L2 )ε ∩ (A+
×
A
)
(L
)
∩
(A
×
A
)
(L2 )ε ∩ (A+
1 ε
2
2
1
1
2 × A2 ) (L1 )a ∩ (A1 × A1 ) .
24
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
Thus La is a rational relation.
Since M1 ∗ M2 is homogeneous, if (u, v) ∈ Lx for x ∈ A, then |v| = |ua| =
|u| + 1. Furthermore, if (u, v) ∈ Lε , then |u| = |v|. Hence L$x and $ Lx are
regular for all x ∈ A ∪ {ε} by Proposition 2.5. Thus M1 ∗ M2 is automatic.
(This reasoning is essentially [25, Proof of Theorem 6.2], but simplified
because we consider only homogeneous semigroups.)
Now we consider the interaction of free product with fcrs within the
class of homogeneous monoids.
Theorem 4.2 ([58, Theorem D]). Let M1 and M2 be monoids. Suppose
that M2 has no non-trivial left- or right-invertible elements. If M1 ∗ M2 is
fcrs, then M1 and M2 are fcrs.
Proposition 4.3. Let M1 and M2 be homogeneous monoids. Then M1 ∗M2
is fcrs if and only if M1 and M2 are fcrs.
Proof. Suppose M1 ∗M2 is fcrs. Since M1 and M2 are homogeneous, neither
contains any non-trivial left- or right-invertible elements. So M1 and M2 are
both fcrs by Theorem 4.2.
The converse part follows by [42, Proposition 4.3].
Finally, we recall the following result on the interaction of the free product
and fdt:
Theorem 4.4 ([41, 42]). Let M1 and M2 be monoids. Then M1 ∗ M2 is
fdt if and only if M1 and M2 are fdt.
Now, the examples in Section 3 suffice to construct the remaining examples in Figure 1 by taking their free products. Regarding Figure 1 as a
semilattice, examples in Section 3 correspond to the elements of this semilattice that have no non-trivial decomposition. Proposition 4.3, Theorem 4.4,
and Proposition 4.1 together show that by taking a free product, one obtains
a new monoid whose properties are given by taking the logical conjunction
(that is, the ‘and’ operation) of the corresponding properties. That is, given
two monoids Mab and Mcd , their free product Mab ∗ Mcd will have the weaker
of the two properties a and b (which lie in {fcrs, fdt, nonfdt}) and the
weaker of the two properties c and d (which lie in {biauto, auto, nonauto}).
Thus, to summarize:
fdt
• Mfcrs
auto ∗ Mbiauto is non-fcrs, fdt, non-biauto, auto;
fdt
• Mbiauto ∗ Mfcrs
nonauto is non-fcrs, fdt, non-biauto, non-auto;
nonfdt is non-fcrs, non-fdt, non-biauto, auto;
∗
M
• Mfcrs
biauto
auto
nonfdt
• Mfcrs
nonauto ∗ Mbiauto is non-fcrs, non-fdt, non-biauto, non-auto.
Theorem 4.5. For each consistent combination of the properties fcrs, fdt,
biauto, auto, and their negations, there exists a homogeneous monoid with
exactly that combination of properties.
5. From homogeneous to n-ary multihomogeneous monoids
Thus far we have proved that for every consistent combination of the
properties fcrs, fdt, biauto, auto, and their negations, there exists a
homogeneous monoid with exactly those properties. In the remainder of
ON HOMOGENEOUS MONOIDS
25
the paper, we show that such monoids exist in the more restricted class of
n-ary homogenous monoids, and show that monoids with some consistent
combinations of these properties exist in the even more restricted class of
n-ary multihomogeneous monoids. The current section describes the overall
strategy and results; the following two sections then develop the necessary
concepts and techniques. First, we introduce and investigate the theory of
abstractly Rees-commensuable semigroups in Section 6, ultimately proving
Corollary 6.5, which implies that from each homogeneous monoid listed
in Table 1, we can obtain an n-ary homogeneous monoid with the same
combination of properties. Thus we have the following analogy of Theorem
4.5 for n-ary homogeneous monoids:
Theorem 5.1. For each consistent combination of the properties fcrs,
fdt, biauto, auto, and their negations, there exists an n-ary homogeneous monoid with exactly that combination of properties.
Since we have examples of n-ary homogeneous monoids with all consistent combinations of the properties fcrs, fdt, biauto, auto, and their
negations, the next step is to extend these examples to n-ary multihomogeneous monoids. With that aim, in Section 7 we define and investigate
an embedding of a (n-ary) homogeneous monoid into a 2-generated (n-ary)
multihomogeneous monoid. As stated below in Corollary 7.13, passing to
and from the (n-ary) multihomogeneous monoid preserves fdt, auto, and
biauto. Furthermore, passing to the (n-ary) multihomogeneous monoid
preserves fcrs. It is unknown whether fcrs is preserved passing back to
the original (n-ary) homogeneous monoid, or, equivalently, whether nonfcrs is preserved on passing to the (n-ary) multihomogeneous monoid. Applying this embedding technique to the list of n-ary homogeneous monoids
we discussed above, we get the following results:
Theorem 5.2. For each consistent combination of the properties fcrs, biauto, auto, and their negations, there exists an n-ary multihomogeneous
monoid with exactly that combination of properties.
Theorem 5.3. For each consistent combination of the properties fdt, biauto, auto, and their negations, there exists an n-ary multihomogeneous
monoid with exactly that combination of properties.
To obtain an analogue of Theorem 5.1 for n-ary multihomogeneous monoids
it would be sufficient to find a multihomogeneous monoid that is non-fcrs,
fdt, and biauto. Indeed, in that case, combining Theorems 5.2 and 5.3
with the results of Section 4 and noting that a free product of multihomogeneous monoids is again multihomogeneous, we would get, for any consistent
combination of biauto, auto, and their negations, an example of an fdt,
non-fcrs multihomogeneous monoid with exactly that combination of properties. Therefore, from Corollary 6.5 we would get examples of n-ary multihomogeneous monoids with exactly the same discussed properties. Joining
these examples with the examples from Theorems 5.2 and 5.3 we would get
the intended result. Thus we have the following question:
Question 5.4. Does there exist a multihomogeneous monoid that is fdt,
but non-fcrs?
26
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
6. Abstractly Rees-commensurable semigroups
We introduce a new definition which is inspired by the notions of abstractly
commensurable groups [36, §§ iv.27ff.] and Rees index for semigroups [59].
A subsemigroup T of a given semigroup S has finite Rees index if S \ T is
finite. In that case the semigroup S is said to be a small extension of T , and
T a large subsemigroup of S. The main interest is that large subsemigroups
and small extensions of a given semigroup share many important properties
of that semigroup (see [60] for a survey).
Definition 6.1. Two semigroups S1 and S2 are said to be abstractly Reescommensurable if there are finite Rees index subsemigroups Ti ⊆ Si (for
i = 1, 2) with T1 ∼
= T2 .
Is is easy to verify that abstract Rees-commensurability is an equivalence
relation on semigroups.
This notion can be naturally extended to ideals.
Definition 6.2. Two semigroups S1 and S2 are said to be abstract Reesideal-commensurable if there are finite Rees index ideals Ui ⊆ Si (for i = 1, 2)
with U1 ∼
= U2 .
The idea behind these notions is that abstract Rees-ideal-commensurable
semigroups share many important properties, such as fcrs, fdt, biauto,
and auto.
Proposition 6.3. fcrs, fdt, biauto, and auto are preserved under abstract Rees-ideal-commensurability:
• It is known that fcrs is inherited by small extensions [43,
Theorem 1] and by large subsemigroups [61, Theorem 1.1]. Although
the result for small extensions was stated in the context of monoids,
it can be naturally extended to semigroups. These two results imply that fcrs is preserved under abstract Rees(-ideal)-commensurability.
• It is known that fdt is inherited by small extensions [43, Theorem 2]
of monoids, and by large semigroup ideals [45, Theorem 1]. We
recall that the notion of finite derivation type was first introduced
for monoids, but it was naturally extended to the semigroup case
[62, Section 2].
The result on small extensions can be easily adapted for the semigroup case. Indeed, let T be a semigroup and consider the monoid
T 1 obtained from T by adding an identity. If T is fdt the derivation graph of T 1 can be obtained from the derivation graph of T
by adding an extra connected component with a single vertex corresponding to the empty word. Thus T 1 is fdt. Now, if S is a small
extension of the semigroup T , it turns out that the monoid S 1 is a
small extension of the monoid T 1 . Therefore, by [43, Theorem 2] the
monoid S 1 is fdt. But S is a large ideal of S 1 , and by [45, Theorem
1] we conclude that S is fdt.
The two results on small extensions and large ideals of semigroups
show that fdt is preserved under abstract Rees-ideal-commensurability.
Proof.
ON HOMOGENEOUS MONOIDS
27
• By [63, Theorem 1.1] and its natural analogue for biauto, both
auto and biauto are inherited by small extensions and by large
subsemigroups, and therefore auto and biauto are preserved under
abstract Rees(-ideal)-commensurability.
The preceding result is important because, in the case of (multi)homogeneous
and n-ary (multi)homogeneous monoids, the following result holds:
Proposition 6.4. Every finitely presented (multi)homogeneous monoid is
Rees-ideal-commensurable to an n-ary (multi)homogeneous monoid, where
n can be chosen arbitrarily as long as it is greater than or equal to the length
of the longest relation in R.
Proof. Let hA | Ri be a finite homogeneous presentation of a monoid M .
Let n be chosen arbitrarily, provided it is greater than or equal to the
maximum length of a relation in R. We are going to construct an nary (multi)homogeneous monoid M ′ and show it is abstract Rees-idealcommensurable to M by finding isomorphic ideals I and I ′ of M and M ′
respectively.
Let I = {[u]M ∈ M : |u| ≥ n}. Note that I is an ideal
S of finite Rees index
in M , since its complement in M is the finite set 0≤i<n {[a1 ]M · · · [ai ]M :
a1 , . . . , ai ∈ A}.
Let R′ = {(uℓv, urv) : (ℓ, r) ∈ R, u, v ∈ A∗ , |uℓv| = n}. Consider the
n-ary (multi)homogeneous presentation hA | R′ i, and let M ′ be the monoid
defined by it.
The set I ′ = {[u]M ′ ∈ M ′ : |u| ≥ n} is an ideal of M ′ . Moreover, I ′ has
finite Rees index in M ′ .
Since R′ is contained in the Thue congruence generated by R, we can
define a map ϕ : I ′ → I, with ([u]M ′ ) ϕ = [u]M . Given [u]M in I, thus with
u ∈ A∗ and |u| ≥ n, we have [u]M ′ ∈ I ′ by definition of I ′ . Therefore ϕ
is surjective. It is also injective since, for any [u]M ′ , [v]M ′ ∈ I ′ such that
([u]M ′ ) ϕ = ([v]M ′ ) ϕ, we get [u]M = [v]M , with |u| = |v| ≥ n, and therefore
[u]M ′ = [v]M ′ . It is routine to check that ϕ is a homomorphism. Thus I and
I ′ are isomorphic finite Rees index ideals of M and M ′ . So M and M ′ are
abstract Rees-ideal-commensurable.
Corollary 6.5. Let P be a set of properties preserved under abstract Reesideal-commensurability. Then there exists a (multi)homogeneous monoid
satisfying every property in P if and only if there exists an n-ary (multi)homogeneous monoid satisfying every property in P .
Note that the set of properties P can contain ‘negative’ properties like
‘not finitely generated’. As an immediate consequence, we obtain Theorem
5.1
7. From n-ary homogeneous to n-ary multihomogeneous monoids
We now develop the embedding technique that allows us to construct
multihomogeneous examples. The technique embeds a homogeneous monoid
into a multihomogenous monoid that shares several of the properties under
28
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
study. If the original monoid is n-ary homogeneous, then the monoid it is embedded into is n-ary multihomogenous, so we use the formulations ‘(n-ary)
homogeneous’ and ‘(n-ary) multihomogeneous’ to indicate that statements
apply to both the general and n-ary cases.
Let us start by fixing some notation that will be maintained throughout
this section. Let M be a finitely generated (n-ary) homogeneous monoid,
and let hA | Ri be a (n-ary) homogeneous presentation defining M . Recall
that A = {a1 , . . . , am } is minimal, so that any other generating set of M
contains A (see discussion at the end of Subsection 2.1).
We define a homomorphism
φ : A∗ → {x, y}∗ ,
ai 7→ x2 y i xy m+1−i .
Denote by N the finitely generated monoid presented by hx, y | Rφi, where
Rφ denotes the set {(ℓφ, rφ) : (ℓ, r) ∈ R}.
Proposition 7.1. The monoid N , defined by the presentation hx, y | Rφi,
is (n(m + 4)-ary) multihomogeneous.
Proof. The presentation hA | Ri is homogeneous hence, for any (u, v) ∈ R,
we have |u| = |v|. Since ai φ contains 3 symbols x and m + 1 symbols y for
all i ∈ {1, . . . , m}, it follows that |uφ|x = 3|u| = 3|v| = |vφ|x and |uφ|y =
(m + 1)|u| = (m + 1)|v| = |vφ|y . Hence hx, y | Rφi is multihomogeneous.
If the presentation hA | Ri is n-ary homogeneous, then also for any relation (u, v) ∈ R, we have |u| = |v| = n. But in that case we get |uφ| =
|uφ|x + |uφ|y = 3|u| + (m + 1)|u| = 3n + (m + 1)n = n(m + 4), and similarly
for |vφ|.
A subset C of {x, y}∗ is a code if it is a set of free generators for the
submonoid of {x, y}∗ generated by C [64, Subsection 7.2]. That is the case
for the set Aφ = {x2 y i xy m+1−i : i = 1, . . . , m}. Furthermore, Aφ is what is
called a prefix code, since no word of Aφ is a proper prefix of another word
in Aφ [7, Chapter 6].
By [7, Proposition 6.1.3], since φ induces a bijection from A to the code
Aφ, we conclude that φ is injective, and for that reason φ is called a coding
morphism for Aφ.
Proposition 7.2. The monoid M presented by hA | Ri embeds into the
monoid N presented by hx, y | Rφi via the map [u]M 7→ [uφ]N , and the
words over {x, y} representing elements of (the image of ) M are precisely
the words in (A∗ )φ.
Proof. Consider the natural projection ρN of {x, y}∗ onto N , whose kernel
is generated by Rφ. Since φ is injective, the kernel of the composition
φ ◦ ρN : A∗ → N is the congruence generated by R.
The natural projection of ρM of A∗ onto M has the same kernel has the
map φ ◦ ρN . By [64, Theorem 1.5.2] there is a monomorphism φ : M → N
such that (M )φ = (A∗ )φ ◦ ρN and ([u]M )φ = (u)φ ◦ ρN = [uφ]N .
We shall now turn our attention to investigating the relationship between
fdt holding in M and fdt holding in N . For this we shall prove some
results which relate the Squier graphs of M and N .
ON HOMOGENEOUS MONOIDS
29
Let ∆ denote the subgraph of Γ(hx, y | Rφi) induced on the set of vertices
(Aφ)∗ .
Let us extend the mapping φ : A∗ → {x, y}∗ to a mapping from the
derivation graph Γ(hA | Ri) to the derivation graph Γ(hx, y | Rφi) mapping
an edge E = (w1 , (ℓ, r), ǫ, w2 ) to an edge Eφ = (w1 φ, (ℓφ, rφ), ǫ, w2 φ). The
mapping φ extends to a mapping between paths by putting
(E1 E2 . . . Ek )φ = (E1 φ)(E2 φ) . . . (Ek φ),
for any edges E1 , E2 , . . . , Ek of Γ(hA | Ri).
Lemma 7.3. The mapping φ : Γ(hA | Ri) → Γ(hx, y | Rφi) has the following
properties.
(1) φ maps A∗ bijectively to (Aφ)∗ ;
(2) For every edge E in Γ(hA | Ri) we have
ι(Eφ) = (ιE)φ,
and
τ (Eφ) = (τ E)φ;
(3) For every edge E in Γ(hA | Ri) and every pair of words u, v ∈ A∗ we
have
(u · E · v)φ = uφ · Eφ · vφ;
(4) φ maps the edge-set of Γ(hA | Ri) bijectively to the edge-set of ∆.
Proof. (i) It follows from the fact that φ is an injective homomorphism from
A∗ into {x, y}∗ , and thus φ maps bijectively A∗ to its image (A∗ )φ = (Aφ)∗ .
(ii) & (iii) The identities follow from the definition of the extended map φ
on edges and from the fact that φ is a homomorphism.
(iv) The injectivity of the edge-set of Γ(hA | Ri) to the edge-set of ∆, follows
from the injectivity of φ. Now, consider an edge F = (z1 , (ℓφ, rφ), ǫ, z2 ) of
∆, and suppose without lost of generality, that ǫ = +1.
The word ιF = z1 (ℓφ)z2 decomposes as a concatenation of words in Aφ;
that is, words of the form x2 y j xy m+1−j for various j ∈ {1, . . . , m}. Since
subwords x2 only occur at the start of such words in Aφ, the x2 at the start
of ℓφ lies at the start of a word from Aφ in the decomposition of ιF. Since
all words in Aφ have the same length, ℓφ must also finish at the end of some
word from Aφ in the decomposition of ιF. Hence z1 and z2 are (possibly
empty) concatenations of words in Aφ, and so z1 = w1 φ and z2 = w2 φ for
some w1 , w2 ∈ A∗ . Thus F = Eφ, for the edge E = (w1 , (ℓ, r), +1, w2 ) of
Γ(hA | Ri).
Let ψ : (Aφ)∗ → A∗ be the right inverse of the injective mapping φ. Note
that (w1 w2 )ψ = (w1 ψ)(w2 ψ), for any elements w1 , w2 ∈ (Aφ)∗ . By part
(iv) of the above lemma we can extend ψ to a mapping of edges by simply
setting (Eφ)ψ = E and using the fact that every edge of ∆ has the form Eφ
for a unique edge E from Γ(hA | Ri).
Lemma 7.4. The mapping ψ : ∆ → Γ(hA | Ri) has the following properties.
(1) ψ maps (Aφ)∗ bijectively to A∗ ;
(2) For every edge E in ∆ we have
ι(Eψ) = (ιE)ψ,
and
τ (Eψ) = (τ E)ψ;
30
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
(3) For every edge E in ∆ and every pair of words u, v ∈ {x, y}∗ we have
that u · E · v is an edge in ∆ if and only if u, v ∈ (Aφ)∗ in which case
(u · E · v)ψ = uψ · Eψ · vψ.
(4) ψ maps the edge-set of ∆ bijectively to the edge-set of Γ(hA | Ri).
Proof. (i) Follows immediatly from the definition of ψ.
(ii) From Lemma 7.3 each edge of ∆ has the form Eφ = (w1 φ, (ℓφ, rφ), ǫ, w2 φ),
with w1 , w2 ∈ A∗ , (ℓ, r) ∈ R and ǫ = ±1. The image of the edge under ψ
is E. Since φ is an homomorphism and from the definition of ψ we get
((w1 φ)(ℓφ)(w2 φ))ψ = w1 ℓw2 and ((w1 φ)(rφ)(w2 φ))ψ = w1 rw2 as required.
(iii) Suppose that E is an edge in ∆ and that for some u, v ∈ {x, y}∗ we
have u · E · v an edge in ∆. So the vertex u(ιE)v of E is a word in (Aφ)∗ .
Arguing as in Lemma 7.3 (iv) we can conclude that u, v ∈ (Aφ)∗ .
The converse part of the equivalence follows trivially.
If E in ∆ and u, v ∈ (Aφ)∗ , there exists E′ in Γ(hA | Ri) and u′ , v ′ ∈ A∗ ,
such that (u · E · v) = u′ φ · Eφ · v ′ φ. From the definition of ψ we get
(u · E · v)ψ = uψ · Eψ · vψ.
(iv) It follows from the definition of ψ and Lemma 7.3.
From the above lemmas, we can regard, up to the given encoding, the
graph Γ(hA | Ri) as a subgraph of Γ(hx, y | Rφi), identifying Γ(hA | Ri) with
∆.
Lemma 7.5. Let C be a set of closed paths in Γ(hA | Ri). Then Cφ is a
set of closed paths in ∆. Moreover, for every closed path P in Γ(hA | Ri), if
P ∼C 1ιP in Γ(hA | Ri) then Pφ ∼Cφ 1(ιP)φ in Γ(hx, y | Rφi).
Proof. The result follows by [35, Theorem 3.6].
At this point, our strategy starts to emerge. The identification between ∆
with the derivation graph Γ(hA | Ri), will allow us to get a finite homotopy
base of Γ(hA | Ri), from a finite homotopy bases of Γ(hx, y | Rφi).
Lemma 7.6. Let D be a set of closed paths in ∆. Then Dψ is a set of closed
paths in Γ(hA | Ri). Moreover, for every closed path P in ∆, if P ∼D 1ιP in
Γ(hx, y | Rφi) then Pψ ∼Dψ 1(ιP)ψ in Γ(hA | Ri).
Proof. The result follows from [39, Lemma 1], where the author uses strict
monoidal categories to prove an analogue of [35, Theorem 3.6].
Indeed, in that context, both Γ(hA | Ri) and ∆ are strict monoidal categories (notice that ∆ is not a derivation graph, and so we can not refer to
[35, Theorem 3.6] to prove the lemma). The mapping ψ is a functor that
preserves the multiplicative structure, by Lemma 7.4. The homotopy relation ∼Dψ is what is called a 2-congruence on ∆, satisfying Pψ ∼Dψ 1(ιP)ψ , for
any closed path P in D. From [39, Lemma 1], it follows that Pψ ∼Dψ 1(ιP)ψ ,
for any closed path P of ∆ such that P ∼D 1ιP .
Next, we describe the remaining connected componnets of Γ(hx, y | Rφi).
Lemma 7.7. The derivation graph Γ(hx, y | Rφi) has the following properties:
ON HOMOGENEOUS MONOIDS
31
(1) If a vertex z has maximal factors (with respect to length) u and
v in (Aφ)∗ , then either z = z0 uz2 = z0 vz2 , or z = z0 uz1 vz2 , or
z = z0 vz1 uz2 , with z0 , z1 , z2 ∈
/ (Aφ)+ and z1 non-empty;
(2) Any vertex has a unique decomposition of the form
w0 u1 w1 · · · wk−1 uk wk ,
where each ui is a maximal factor in (Aφ)∗ , and each wi has no
factor in (Aφ)∗ ;
(3) Any edge has the form z1 · (Eφ) · z2 , for some edge E in Γ(hA | Ri),
where ιEφ and τ Eφ are maximal factors in (Aφ)∗ of the corresponding initial and terminal vertices of Eφ, and z1 , z2 are not in (Aφ)+ .
Proof. (i) Let u and v be maximal factors in (Aφ)∗ of z. Suppose u and v
overlap in the word z. Without lost of generality, suppose that z = u0 uu1 =
v0 vv1 with |u0 | ≤ |v0 | < |u0 u|. Since v ∈ (Aφ)∗ is a sequence of words of the
form x2 y j xy m+1−j for various j ∈ {1, . . . , m}, and subwords x2 only occur
at the start of such words in Aφ, the x2 at the start of v lies at the start
of a word from Aφ in the decomposition of u. By the maximality of v and
since z = u0 uu1 = v0 vv1 we get u0 = v0 . Since all words in Aφ have the
same length, u and v must also finish at the end of some word from Aφ. By
the maximality of u and v we get u = v.
(ii) The result follows from (i).
(iii) From the definition of Γ(hx, y | Rφi), an edge has for initial and terminal
vertices words t and s in {x, y}∗ where one is obtained from the other by
a single application of a defining relation in Rφ, that is, s = w0 (ℓφ)w1 and
t = w0 (rφ)w1 , with (ℓ, r) ∈ R. Notice that ℓφ, rφ ∈ (Aφ)∗ . So, we can
consider maximal factors p and q in (Aφ)∗ having as factors ℓφ and rφ,
respectively, in such a way that
s = z1 (pφ)z2 ,
t = z1 (qφ)z2 .
The words p and q are the initial and terminal vertices of an edge E in
Γ(hA | Ri). The result follows as in the statement.
Lemma 7.8. A length-two path in Γ(hx, y | Rφi) has one of the forms:
(1) (z0 · E1 φ · z2 ) ◦ (z0 · E2 φ · z2 ), and E1 ◦ E2 is a path in Γ(hA | Ri); or
(2) (z0 · E1 φ · z1 (ιE2 φ)z2 ) ◦ (z0 (τ E1 φ)z1 · E2 φ · z2 ); or
(3) (z0 (ιE2 φ)z1 · E1 φ · z2 ) ◦ (z0 · E2 φ · z1 (τ E1 φ)z2 ),
for z0 , z1 , z2 ∈
/ (Aφ)+ and z1 non-empty, and E1 , E2 edges in Γ(hA | Ri).
Proof. By Lemma 7.7 (iii), the path of consecutive edges has the form (w1 ·
(E1 φ) · w2 ) ◦ (z1 · (E2 φ) · z2 ). In the conditions of that lemma the words τ E1 φ
and ιE2 φ are maximal factors in (Aφ)∗ of the same word w1 (τ E1 φ)w2 =
z1 (ιE2 φ)z2 . From Lemma 7.7 (i) we have three possible cases. Each of the
cases corresponds to one the possible cases (i), (ii) and (iii) of the statement.
Lemma 7.9. Let P be a non-empty (closed) path in Γ(hx, y | Rφi). Then
(7.1)
P ∼0 P1 ◦ P2 ◦ · · · ◦ Pk ,
32
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
for some k ∈ N, and where each Pi has the form
w0 (τ Q1 φ)w1 · · · (τ Qi−1 φ)wi−1 · (Qi φ) · wi (ιQi+1 φ) · · · wk−1 (ιQk φ)wk ,
for some (closed) path Qi in Γ(hA | Ri), with w0 , . . . , wk ∈
/ (Aφ)∗ and w1 , . . . , wk−1
non-empty.
Proof. The key observation to prove the lemma is to note that in Lemma
7.8 the paths on items (ii) and (iii) are ∼0 -homotopic. Indeed, this is a
consequence of disjoint derivations.
By Lemma 7.7 (ii) the initial vertex of P can be factorized in the form
w0 u1 w1 · · · wk−1 uk wk ,
where each ui is a maximal factor in (Aφ)∗ and each wi has no factor in
(Aφ)∗ . Notice, also by the same lemma, that any edge in P has a factorization of the form z1 · (Eφ) · z2 , for some edge E in Γ(hA | Ri), where ιEφ and
τ Eφ are maximal factors in (Aφ)∗ of the corresponding initial and terminal
vertices of Eφ.
Since wi ’s have no factor in (Aφ)∗ , no relation from Rφ is going to be
applied to them, and hence they are fixed in this path. Also, for each
i ∈ {1, . . . , k} we can identify Bi as the set of edges in P where the relation
is applied to a word between wi−1 and wi .
From Lemma 7.8 any two consecutive edges Ei ◦ Ej , with Ei ∈ Bi and
Ej ∈ Bj , and i > j, we get Ei ◦ Ej ∼0 Ej ◦ Ei by disjoint derivations.
Consequently, taking i = 1 we can group together all edges in B1 by
finding a ∼0 -homotopic path in which all those edges are at the beginning
of the path. Proceeding in this way with the remaining edges we get the
intended result.
The following result can be viewed as a generalization of Lemma 2.3 in
[65].
Proposition 7.10. The monoid M is fdt if and only if the monoid N is
fdt.
Proof. Suppose that M is fdt and let C be a finite homotopy base (of closed
paths) for Γ(hA | Ri). We shall see that the finite set Cφ = {Pφ : P ∈ C} of
closed paths is a homotopy base for Γ(hx, y | Rφi) .
Let P be a non-empty closed path in Γ(hx, y | Rφi). By Lemma 7.9 there
are closed paths Qi , for i = 1, . . . , k, in Γ(hA | Ri) such that
P ∼0 P1 ◦ P2 ◦ · · · ◦ Pk ,
and each Pi = wi · Qi φ · wi′ , for some wi , wi′ ∈ {x, y}∗ .
Since C is a homotopy base, for each i = 1, . . . , k, we have Qi ∼C 1ι(Qi ) .
Thus Pi ∼Cφ 1ι(Pi φ) by Lemma 7.5, for all i = 1, . . . , k, which in turn implies that P ∼Cφ 1ι(Pφ) . Consequently, Cφ is a finite homotopy base for
Γ(hx, y | Rφi), and so N is fdt.
Conversely, suppose that N is fdt, and let E be a finite homotopy base
(of closed paths) for Γ(hx, y | Rφi). Again by Lemma 7.9, for each P in E,
there are closed paths Q1 , . . . , Qk in Γ(hA | Ri), such that
P ∼0 P1 ◦ P2 ◦ · · · ◦ Pk ,
ON HOMOGENEOUS MONOIDS
33
and Pi = wi · Qi φ · wi′ , for some wi , wi′ ∈ {x, y}∗ . Denote by C the finite set
of all Qi ’s, for all P ∈ E, and let Cφ denote the set {Pφ : P ∈ C}. From the
definition of Cφ, the paths Pi satisfy Pi ∼Cφ 1ι(Pi ) , and so each closed path P
in E satisfies P ∼Cφ 1ι(P) . Thus, the set Cφ generates the homotopy relation
∼E , and therefore also Cφ is a finite homotopy base of Γ(hx, y | Rφi). Note
that Cφ is a set of closed paths in ∆ and that Cφψ = C.
To conclude the proof, let P be a closed path in Γ(hA | Ri). Since Cφ is
a homotopy base of Γ(hx, y | Rφi), we have Pφ ∼Cφ 1ι(Pφ) . By Lemma 7.6,
we get Pφψ ∼Cφψ 1ι(Pφψ) . Since Pφψ = P we get P ∼C 1ι(P) as required.
Therefore, C is a finite homotopy base of Γ(hA | Ri), and so M is fdt.
Proposition 7.11. The monoid M is auto (respectively, biauto) if and
only if the monoid N is auto (respectively, biauto).
Proof. We will prove the result for biauto; the result for auto follows by
considering multiplication only on one side.
Suppose that N is biauto. By Proposition 2.6, there is a biautomatic
structure ({x, y}, K) for N . By Proposition 7.2, words over {x, y} representing elements of the image under φ of M are precisely those in (Aφ)∗ . So
K ∩ (Aφ)∗ must map onto the image of M .
The map φ is a rational relation, since
φ = { (ai , x2 y i xy m+1i ) : ai ∈ A }∗ .
Thus its converse φ−1 is also a rational relation. Let L = Kφ−1 ; note that
L ⊆ A∗ . Since φ−1 is a rational relation and K is regular, it follows that L is
also regular. Since K ∩ (Aφ)∗ must map onto the image of M , the language
L maps onto M . For any a ∈ A ∪ {ε},
(u, v) ∈ La ⇐⇒ u ∈ L ∧ v ∈ L ∧ ua =M v
⇐⇒ (∃u′ , v ′ ∈ K) (u, u′ ) ∈ φ ∧ (v, v ′ ) ∈ φ ∧ u′ (aφ) =N v ′
⇐⇒ (∃u′ , v ′ ∈ K) (u, u′ ) ∈ φ ∧ (v, v ′ ) ∈ φ ∧ (u′ , v ′ ) ∈ Kaφ
⇐⇒ (u, v) ∈ φ ◦ Kaφ ◦ φ−1 ;
thus La = φ ◦ Kaφ ◦ φ−1 . Since ({x, y}, K) is an automatic structure for
N , each relation Kaφ is rational. Since φ and φ−1 are rational relations,
it follows that La is a rational relation. Since M is homogeneous, (u, v) ∈
La =⇒ |u| − |v| ≤ 1, and so L$a and $ La are regular by Proposition 2.5.
Symmetrical reasoning shows that a L$ and $a L are regular. Hence (A, L) is
a biautomatic structure for M .
Now suppose that M is biauto. By Proposition 2.6, there is a biautomatic structure (A, L) for M . The aim is to construct a biautomatic
structure for N .
Let J = {x, y}+ −{x, y}∗ A+ φ{x, y}∗ ; so J consists of all non-empty words
over {x, y}∗ that do not contain subwords equal to elements M φ. Notice in
particular that ε ∈
/ J and that J is closed under taking non-empty subwords.
Note further that J is regular. For future use, let ∆ = { (z, z) : z ∈ J };
note that ∆ is a rational relation by Proposition 2.1.
34
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
Consider a word w ∈ {x, y}+ . By Lemma 7.7, w can be uniquely factored
into an alternating product of words from J and words from A+ φ. That is,
(7.2)
w = z0 u1 z1 · · · zk−1 uk zk , w
where each ui lies in A+ φ and each zi lies in J, except that z0 and zk may
also be ε. Note that each ui is equal to some word in Lφ−{ε}. The idea is to
build an automatic structure where the language of representatives consists
of alternating products of words from J and words from Lφ. That is, the
language will consist of words (7.2) where each ui lies in Lφ − {ε}.
More formally, let
∗
K = (J ∪ {ε}) (Lφ ∩ {x, y}+ )J (J ∪ {ε});
note that K is regular since J is regular and φ is a homomorphism, and
that every element of N is equal to some word in K by the reasoning in the
previous paragraph.
Now let w ∈ K and factor w as (7.2) where each ui lies in Lφ − {ε} and
each zi lies in J, except that z0 and zk may also be ε.
Consider right-multiplication of w by a generator x. Since zk x cannot
contain a subword from A+ φ (since zk contains no such subword, and such
subwords do not end with symbols x), the words in L to which wx is equal
are precisely words of the form z0 u′1 z1 · · · zk−1 u′k zk x, where each u′i is any
word in Lφ that is equal to ui . (Recally that defining relations in Rφ only
apply to words in A+ φ; thus the subwords zi are fixed.) Hence
∗
Kx = (∆ ∪ {(ε, ε)}) (φ−1 ◦ Lε ◦ φ)∆ Lε φ(ε, x) ∪ (ε, x)
is a rational relation.
Now consider right-multiplication of w by a generator y. The zk y may end
with a subword x2 y i xy m−i+1 ∈ Aφ, and thus it is necessary to distinguish
three cases:
• zk = tx2 y i xy m−i , where t 6= ε. In this case, the factorization of wy
into an alternating product of subwords from J and A+ φ is
w = z0 u1 z1 · · · zk−1 uk tx2 y i xy m−i ;
note that t ∈ J since J is closed under taking subwords. The
words in L to which wy is equal are precisely words of the form
z0 u′1 z1 · · · zk−1 u′k zk v, where each u′i is any word in Lφ that is equal
to ui , and v is any word in Lφ that is equal to x2 y i xy m−i = ai φ. Let
Hi = { p ∈ L : p = ai }; then Hi is regular by [25, Proposition 3.1].
(In fact, since M is homogeneous, H can contain only generators
from A and so must be finite.) Let
Ky(1) = (∆ ∪ m{(ε, ε)}) (φ−1 ◦ Lε ◦ φ)∆
m
∗ [
i=1
{x2 y i xy m−i } × Hi φ .
Since Hi φ is regular, {x2 y i xy m−i } × Hi φ is a rational relation by
(1)
Proposition 2.2 and so Ky is a rational relation since it is a concatenation of rational relations.
(1)
Then Ky is a rational relation that describes right-multiplication
by a generator y in this case.
ON HOMOGENEOUS MONOIDS
35
• zk = x2 y i xy m−i . In this case, the factorization of wy into an alternating product of subwords from J and A+ φ is
w = z0 u1 z1 · · · zk−1 uk x2 y i xy m−i ;
note that the last factor is uk x2 y i xy m−i ∈ A+ φ. The words in L to
which wy is equal are precisely words of the form z0 u′1 z1 · · · zk−1 v,
where each u′i is any word in Lφ that is equal to ui , and v is any
word in Lφ that is equal to uk x2 y i xy m−i . That is, v can be any word
such that (uk , v) ∈ Lai φ. Let
∗
Ky(2) = (∆ ∪ {(ε, ε)}) (φ−1 ◦ Lε ◦ φ)∆
·
n
[
i=1
(2)
(φ−1 ◦ Lai ◦ φ)(x2 y i xy m−i , ε) .
Then Ky is a rational relation that describes right-multiplication
by a generator y in this case.
• zk does not end with a suffix of the form x2 y i xy m−i . The words in L
to which wy is equal are precisely words of the form z0 u′1 z1 · · · zk−1 u′k zk y,
where each u′i is any word in Lφ that is equal to ui . Let
∗
Ky(3) = (∆ ∪ {(ε, ε)}) (φ−1 ◦ Lε ◦ φ)∆ (φ−1 ◦ Lε ◦ φ)
· (v, v) : v ∈ J − {x, y}
(2)
m
[
x2 y i xy m−i (ε, y).
i=1
Then Ky is a rational relation that describes right-multiplication
by a generator y in this case.
(1)
(2)
(3)
Thus Ky = Ky ∪ Ky ∪ Ky is a rational relation.
Since N is homogeneous, (u, v) ∈ Kt =⇒ |u| − |v| ≤ 1, and so Kt$ and
$ K are regular by Proposition 2.5.
t
Similar reasoning shows that t K $ and $t K. Hence ({x, y}, K) is a biautomatic structure for N .
Proposition 7.12. If hB | Qi is a finite presentation for M , and so A ⊆ B,
then
P = hx, y, B | Q, (aφ, a) (∀a ∈ A)i
is a finite presentation defining N . Moreover, the presentation hB | Qi is
complete if and only if the presentation P is complete. Thus, if M is fcrs
then N is fcrs.
Proof. Using Tietze transformations we obtain from the presentation hx, y | Rφi
a new presentation for N as follows: for each a ∈ A, insert a generator a
and a relation (aφ, a), thus obtaining a Tietze equivalent presentation
hx, y, A | Rφ, (aφ, a) (∀a ∈ A)i .
Since φ is a homomorphism, identifying each symbol a with the word aφ,
performing these substitutions on the words from Rφ, we obtain another
Tietze equivalent presentation defining the same monoid:
hx, y, A | R, (aφ, a) (∀a ∈ A)i .
36
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
As it is possible to obtain from the presentation hA | Ri the presentation
hB | Qi using finitely many Tietze transformations, we can obtain from the
presentation hx, y, A | R, (aφ, a) (∀a ∈ A)i the presentation P by using the
same Tietze transformations.
Suppose that hB | Qi is also complete. Observe that Q relates words from
the alphabet B, and that a relation from the set E = {(aφ, a) : a ∈ A} has
left-hand side in {x, y}∗ and right-hand side in A. Thus, if w →Q w′ →E w′′ ,
we can find w such that w →E w →Q w′′ . Indeed, since w′ →E w′′ the word
w′ has a factor from the alphabet {x, y}∗ . Since w →Q w′ the word w′ is
obtained from w by changing some factor in B ∗ . Thus, the word w also
contains the left-hand side of the relation from E applied to w′ . This means
that w has both left-hand sides of the relations being applied, and those lefthand sides do not overlap. So, we can alternatively apply first the relation
from E, obtaining a word w, and then apply the relation from Q, obtaining
the word w′′ .
Hence, →E quasi-commutes over →Q , that is, →Q ◦→E ⊆ →E ◦→∗Q∪E . By
[66, Theorem 1], the rewriting system ({x, y} ∪ A, Q ∪ E) is terminating if,
and only if, both Q and E are terminating. By assumption Q is terminating
and from where E is also terminating by length-reduction.
As before, since Q relates words from the alphabet B, and that relations
from E have left-hand side in {x, y}∗ and right-hand side in A, we can
also deduce that whenever w′ ←∗Q w →∗E w′′ , there exists w such that
w′ →∗E w ←∗Q w′′ . Therefore, the relations →∗E and ←∗Q commute, that is,
←∗Q ◦ →∗E ⊆ →∗E ◦ ←∗Q . By [48, Lemma 2.7.10], if E and Q are confluent and
→∗E and ←∗Q commute, then Q ∪ E is also confluent. Since by assumption Q
is confluent it remains to show that E is confluent.
Observe that two left-hand sides ai φ and aj φ of rules in E can overlap if
and only if ai φ = aj φ. Since φ is injective we get ai = aj . Therefore, E has
no critical pairs and thus is also confluent.
Conversely, suppose that P is a complete presentation. Since hB | Qi is
contained in P, we deduce that hB | Qi is terminating. Confluence also holds
from the fact that in P all critical pairs are resolved, in particular, those
arising from relations in Q. Hence, each resolution associated to relations
from Q, can only involve relations from Q, since left-hand sides of rules in
E belong to {x, y}∗ . Therefore, hB | Qi is complete.
It is unknown if the property fcrs is preserved from N to M . If this were
true, we would have an example satisfying the conditions of Question 5.4.
Combining Propositions 7.10, 7.11, and 7.12, we conclude the following:
Corollary 7.13. For each of the properties fcrs, (non-)fdt, (non-)auto
and (non-)biauto, the (n-ary) homogeneous monoid M has that property
if and only if the (n-ary) multihomogeneous monoid N has that property.
References
[1] P. Etingof, T. Schedler, A. Soloviev, Set-theoretical solutions to the
quantum Yang-Baxter equation, Duke Math. J. 100 (2) (1999) 169–209.
doi:10.1215/S0012-7094-99-10007-X.
[2] E. Jespers, J. Okniński, Noetherian semigroup algebras, Vol. 7 of Algebras and Applications, Springer, Dordrecht, 2007.
ON HOMOGENEOUS MONOIDS
37
[3] F. Cedó, E. Jespers, J. Okniński, Retractability of set theoretic solutions of the Yang-Baxter equation, Adv. Math. 224 (6) (2010) 2472–2484.
doi:10.1016/j.aim.2010.02.001.
[4] F. Cedó, E. Jespers, J. Okniński, Algebras and groups defined by permutation relations of alternating type, J. Algebra 324 (6) (2010) 1290–1313.
doi:10.1016/j.jalgebra.2010.07.014.
[5] T. Gateva-Ivanova, E. Jespers, J. Okniński, Quadratic algebras of skew
type and the underlying monoids, J. Algebra 270 (2) (2003) 635–659.
doi:10.1016/j.jalgebra.2003.06.005.
[6] E. Jespers, J. Okniński, Quadratic algebras of skew type, in: Algebras, rings
and their representations, World Sci. Publ., Hackensack, NJ, 2006, pp. 93–112.
doi:10.1142/9789812774552\_0009.
[7] M. Lothaire, Algebraic Combinatorics on Words, no. 90 in Encyclopedia of Mathematics and its Applications, Cambridge University Press, 2002.
[8] F. Cedó, J. Okniński, Plactic algebras, J. Algebra 274 (1) (2004) 97–117.
doi:10.1016/j.jalgebra.2003.12.004.
[9] A. Lascoux, M.-P. Schützenberger, Le monoı̈de plaxique, in: Noncommutative
structures in algebra and geometric combinatorics, no. 109 in Quaderni de ”La
Ricerca Scientifica”, CNR, Rome, 1981, pp. 129–156.
URL http://igm.univ-mlv.fr/~ berstel/Mps/Travaux/A/1981-1PlaxiqueNaples.pdf
[10] J. Cassaigne, M. Espie, D. Krob, J.-C. Novelli, F. Hivert, The Chinese Monoid, Int.
J. Alg. Comput. 11 (3) (2001) 301–334. doi:10.1142/S0218196701000425.
[11] J. Jaszuńska, J. Okniński, Structure of Chinese algebras, J. Algebra 346 (1) (2011)
31–81. doi:10.1016/j.jalgebra.2011.08.020.
[12] F. Cedó, J. Okniński, Minimal spectrum and the radical of chinese algebras, Algebr.
Represent. Theor. 16 (4) (2013) 905–930. doi:10.1007/s10468-012-9339-1.
[13] F. Cedó, E. Jespers, J. Okniński, Finitely presented algebras and groups defined by
permutation relations, Journal of Pure and Applied Algebra 214 (7) (2010) 1095–
1102. doi:10.1016/j.jpaa.2009.09.015.
[14] F. Cedó, E. Jespers, G. Klein, Finitely presented monoids and algebras defined by
permutation relations of abelian type, Journal of Pure and Applied Algebra 216 (5)
(2012) 1033–1039. doi:10.1016/j.jpaa.2011.10.021.
[15] J.-C. Novelli, On the hypoplactic monoid, Discrete Mathematics 217 (1-3) (2000)
315–336. doi:10.1016/S0012-365X(99)00270-8.
[16] L. Serrano, The shifted plactic monoid, Mathematische Zeitschrift 266 (2) (2009)
363–392. doi:10.1007/s00209-009-0573-0.
[17] G. Duchamp, D. Krob, Plactic-growth-like monoids, in: M. Ito, H. Jürgensen (Eds.),
Words, Languages and Combinatorics, II, World Scientific, River Edge, NJ, 1994, p.
124–142.
[18] V. Diekert, Y. Métivier, Partial commutation and traces, in: Handbook of formal
languages, Vol. 3, Springer, Berlin, 1997, pp. 457–533.
[19] D. Kuske, Divisibility monoids: presentation, word problem, and rational languages,
in: Fundamentals of computation theory (Riga, 2001), Vol. 2138 of Lecture Notes in
Comput. Sci., Springer, Berlin, 2001, pp. 227–239. doi:10.1007/3-540-44669-9_23.
[20] M. Huschenbett, D. Kuske, G. Zetzsche, The monoid of queue actions, in: E. CsuhajVarjú, M. Dietzfelbinger, Z. Ésik (Eds.), Mathematical Foundations of Computer
Science 2014, no. 8634 in Lecture Notes in Comput. Sci., Springer, pp. 340–351.
doi:10.1007/978-3-662-44522-8\_29.
[21] J. Birman, K. H. Ko, S. J. Lee, A new approach to the word and conjugacy problems in
the braid groups, Adv. Math. 139 (2) (1998) 322–353. doi:10.1006/aima.1998.1761.
[22] P. Dehornoy, L. Paris, Gaussian groups and garside groups, two generalisations of
artin groups, Proceedings of the London Mathematical Society 79 (3) (1999) 569–
604. doi:10.1112/S0024611599012071.
[23] R. V. Book, F. Otto, String Rewriting Systems, Texts and Monographs in Computer
Science, Springer, 1993.
[24] D. B. Epstein, J. W. Cannon, D. F. Holt, S. V. Levy, M. S. Paterson, W. P. Thurston,
Word Processing in Groups, Jones & Bartlett, Boston, MA, 1992.
38
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
[25] C. M. Campbell, E. F. Robertson, N. Ruškuc, R. M. Thomas, Automatic semigroups, Theoret. Comput. Sci. 250 (1-2) (2001) 365–391.
doi:10.1016/S0304-3975(99)00151-6.
[26] A. Heyworth, Rewriting as a special case of non-commutative Gröbner basis theory,
in: M. Atkinson, N. Gilbert, J. Howie, S. Linton, E. Robertson (Eds.), Computational and Geometric Aspects of Modern Algebra, no. 275 in London Mathematical Society Lecture Note Series, Cambridge University Press, 2000, pp. 101–105.
doi:10.1017/CBO9780511600609.009.
[27] V. A. Ufnarovskij, Combinatorial and asymptotic methods in algebra, in: Algebra, VI, Vol. 57 of Encyclopaedia Math. Sci., Springer, Berlin, 1995, pp. 1–196.
doi:10.1007/978-3-662-06292-0\_1.
[28] A. J. Cain, R. D. Gray, A. Malheiro, Finite Gröbner–Shirshov bases for Plactic algebras and biautomatic structures for Plactic monoids, J. Algebra 423 (2015) 37–53.
arXiv:1205.4885, doi:10.1016/j.jalgebra.2014.09.037.
[29] A. J. Cain, R. D. Gray, A. Malheiro, Rewriting systems and biautomatic structures
for Chinese, hypoplactic, and sylvester monoids, Internat. J. Algebra Comput. 25 (12). arXiv:1310.6572, doi:10.1142/S0218196715400044.
[30] F. Cedó, J. Okniński, Gröbner bases for quadratic algebras of skew type,
Proceedings of the Edinburgh Mathematical Society 55 (2) (2012) 387–401.
doi:10.1017/S0013091511000447.
[31] E. Güzel Karpuz, Complete rewriting system for the Chinese monoid, Appl. Math.
Sci (Ruse) 4 (22) (2010) 1081–1087.
URL http://www.m-hikari.com/ams/ams-2010/ams-21-24-2010/karpuzAMS21-24-2010-2.pdf
[32] L. Kubat,
J. Okniński,
Gröbner-Shirshov bases for plactic algebras, Algebra Colloquium 21 (04) (2014) 591–596. arXiv:1010.3338v1,
doi:10.1142/S1005386714000534.
[33] Y. Chen, J. Qiu, Gröbner–Shirshov basis for the Chinese monoid, Journal of Algebra and Its Applications 07 (05) 623–628. arXiv:0804.0972v2,
doi:10.1142/S0219498808003028.
[34] F. Otto, A. Sattler-Klein, K. Madlener, Automatic monoids versus monoids with
finite convergent presentations, in: T. Nipkow (Ed.), Rewriting Techniques and Applications, no. 1379 in Lecture Notes in Comput. Sci., Springer, 1998, pp. 32–46.
doi:10.1007/BFb0052359.
[35] C. C. Squier, F. Otto, Y. Kobayashi, A finiteness condition for rewriting systems,
Theoret. Comput. Sci. 131 (2) (1994) 271–294. doi:10.1016/0304-3975(94)90175-9.
[36] P. de la Harpe, Topics in Geometric Group Theory, University of Chicago Press,
2000.
[37] F. Otto, Y. Kobayashi, Properties of monoids that are presented by finite convergent
string-rewriting systems: a survey, in: D.-Z. Du, K.-I. Ko (Eds.), Advances in algorithms, languages, and complexity, Kluwer Academic Publishers, Dordrecht, 1997, p.
225–266. doi:10.1007/978-1-4613-3394-4\_12.
[38] Y. Kobayashi, F. Otto, On homotopical and homological finiteness conditions for
finitely presented monoids, International Journal of Algebra and Computation 11 (03)
(2001) 391–403. doi:10.1142/S0218196701000577.
[39] Y. Lafont, A new finiteness condition for monoids presented by complete rewriting
systems (after Craig C. Squier), Journal of Pure and Applied Algebra 98 (3) (1995)
229 – 244. doi:10.1016/0022-4049(94)00043-I.
[40] Y. Guiraud, P. Malbos, Higher-dimensional categories with finite derivation type,
Theory Appl. Categ. 22 (2009) No. 18, 420–478.
[41] F. Otto, Modular properties of monoids and string-rewriting systems., Nehaniv,
Chrystopher L. (ed.) et al., Algebraic engineering. Proceedings of the international
workshop on formal languages and computer systems, Kyoto, Japan, March 18–21,
1997 and proceedings of the first international conference on semigroups and algebraic
engineering held in Aizu, Japan, March 24–28, 1997. Singapore: World Scientific. 538554 (1999). (1999).
URL http://www.theory.informatik.uni-kassel.de/techreps/TR-97-6.ps
ON HOMOGENEOUS MONOIDS
39
[42] F. Otto, On properties of monoids that are modular for free products and for certain free products with amalgama
URL http://www.theory.informatik.uni-kassel.de/techreps/TR-97-6.ps
[43] J. Wang, Finite Complete Rewriting Systems and Finite Derivation Type
for Small Extensions of Monoids, J. Algebra 204 (2) (1998) 493–503.
doi:10.1006/jabr.1997.7388.
[44] R. Gray, A. Malheiro, S. Pride, On properties not inherited by monoids from their
Schützenberger groups, Information and Computation 209 (7) (2011) 1120–1134.
doi:10.1016/j.ic.2011.03.004.
[45] A. Malheiro, Finite derivation type for large ideals, Semigroup Forum 78 (3) (2008)
450–485. doi:10.1007/s00233-008-9109-x.
[46] J. E. Hopcroft, J. D. Ullman, Introduction to Automata Theory, Languages, and
Computation, 1st Edition, Addison–Wesley, Reading, MA, 1979.
[47] J. Berstel, Transductions and Context-Free Languages, no. 38 in Leitfäden der Angewandten Mathematik und Mechanik, B. G. Teubner, Stuttgart, 1979.
[48] F. Baader, T. Nipkow, Term Rewriting and All That, Cambridge University Press,
1999.
[49] M. Droste, W. Kuich, H. Vogler, Handbook of Weighted Automata, 1st Edition,
Springer Publishing Company, Incorporated, 2009.
[50] V. Guba, M. Sapir, Diagram groups, Mem. Amer. Math. Soc. 130 (620) (1997)
viii+117. doi:10.1090/memo/0620.
[51] Y. Guiraud,
P. Malbos,
Higher-dimensional normalisation
strategies
for acyclicity, Advances in Mathematics 231 (3-4) (2012) 2294–2351.
doi:10.1016/j.aim.2012.05.010.
[52] J.-P. Serre, Trees, Springer-Verlag, 1980.
[53] M. Hoffmann, R. M. Thomas, Biautomatic semigroups, in: M. Liskiewicz, R. Reischuk
(Eds.), Fundamentals of Computation Theory, no. 3623 in Lecture Notes in Computer
Sciene, Springer, 2005, pp. 56–67. doi:10.1007/11537311\_6.
[54] C. Frougny, J. Sakarovitch, Synchronized rational relations of finite and infinite words,
Theoret. Comput. Sci. 108 (1) (1993) 45–82. doi:10.1016/0304-3975(93)90230-Q.
[55] A. J. Duncan, E. F. Robertson, N. Ruškuc, Automatic monoids and change of generators, Mathematical Proceedings of the Cambridge Philosophical Society 127 (3)
(1999) 403–409. doi:10.1017/S0305004199003722.
[56] M. Hoffmann, R. M. Thomas, Notions of automaticity in semigroups, Semigroup
Forum 66 (3) (2002) 337–367. doi:10.1007/s002330010161.
[57] M. Katsura, Y. Kobayashi, Constructing finitely presented monoids which have
no finite complete presentation, Semigroup Forum 54 (1) (1997) 292–302.
doi:10.1007/BF02676612.
[58] S. J. Pride, J. Wang, Rewriting systems, finiteness conditions, and associated functions, in: Algorithmic problems in groups and semigroups (Lincoln, NE, 1998), Trends
Math., Birkhäuser Boston, Boston, MA, 2000, pp. 195–216.
[59] A. Jura, Determining ideals of a given finite index in a finitely presented semigroup,
Demonstratio Math. 11 (3) (1978) 813–827.
[60] A. J. Cain, V. Maltcev, For a few elements more: A survey of finite Rees index, Tech.
rep. (7 2013). arXiv:1307.8259.
[61] K. Wong, P. Wong, On finite complete rewriting systems and large subsemigroups,
J. Algebra 345 (1) (2011) 242–256. doi:10.1016/j.jalgebra.2011.08.022.
[62] A. Malheiro, Finite derivation type for Rees matrix semigroups, Theoret. Comput.
Sci. 355 (3) (2006) 274–290. doi:10.1016/j.tcs.2005.12.011.
[63] M. Hoffmann, R. M. Thomas, N. Ruškuc, Automatic semigroups with subsemigroups
of finite ress index, International Journal of Algebra and Computation 12 (03) (2002)
463–476. doi:10.1142/S0218196702000833.
[64] J. M. Howie, Fundamentals of Semigroup Theory, no. 12 in London Mathematical
Society Monographs: New Series, Clarendon Press, Oxford University Press, New
York, 1995.
[65] Y. Kobayashi, Finite Homotopy Bases of One-Relator Monoids, J. Algebra 229 (2)
(2000) 547–569. doi:10.1006/jabr.1999.8251.
40
ALAN J. CAIN, ROBERT D. GRAY, AND ANTÓNIO MALHEIRO
[66] L. Bachmair, N. Dershowitz, Commutation, transformation, and termination, in:
J. H. Siekmann (Ed.), 8th International Conference on Automated Deduction, Vol.
230 of Lecture Notes in Computer Science, Springer Berlin Heidelberg, 1986, pp.
5–20. doi:10.1007/3-540-16780-3\_76.
Centro de Matemática e Aplicações, Faculdade de Ciências e Tecnologia,
Universidade Nova de Lisboa, 2829–516 Caparica, Portugal
E-mail address: [email protected]
School of Mathematics, University of East Anglia, Norwich NR4 7TJ,
United Kingdom
E-mail address: [email protected]
Departamento de Matemática & Centro de Matemática e Aplicações, Faculdade de Ciências e Tecnologia, Universidade Nova de Lisboa, 2829–516 Caparica, Portugal
E-mail address: [email protected]
| 4math.GR
|
A Proposed Algorithm for Minimum Vertex Cover Problem and its Testing
Gang Hu
Email: [email protected]
Abstract
The paper presents an algorithm for minimum vertex cover problem, which is an NP-Complete
problem. The algorithm computes a minimum vertex cover of each input simple graph. Tested by the
attached MATLAB programs, Stage 1 of the algorithm is applicable to, i.e., yields a proved minimum
vertex cover for, about 99.99% of the tested 610,000 graphs of order 16 and 99.67% of the tested 1,200
graphs of order 32, and Stage 2 of the algorithm is applicable to all of the above tested graphs. All of
the tested graphs are randomly generated graphs of random "edge density" or in other words, random
probability of each edge. It is proved that Stage 1 and Stage 2 of the algorithm run in O(n5+logn) and
O(n3(5+logn)/2) time respectively, where n is the order of input graph. Because there is no theoretical
proof yet that Stage 2 is applicable to all graphs, further stages of the algorithm are proposed, which are
in a general form that is consistent with Stages 1 and 2.
1.
INTRODUCTION
If the algorithm is to be classified as one algorithm design technique, then transform-and-conquer
strategy maybe most suitable for it [1].
Suppose we need to find a minimum vertex cover of simple graph G, then the first part of the
algorithm is to generate an auxiliary simple graph H, which satisfies the following four conditions:
A. L(V(H)) ⊇ L(V(G)), of which L(X) denotes X’s label set, i.e., the set of X’s all vertex’s labels;
B. Different vertices of H may share a same label. For any edge of H, however, the two endpoints
have different labels;
C. Of each component P of H, all maximal-clique’s vertex-sets have a same cardinal number, and
thus this number is called "grade of component P" and denoted by g(P);
D. For any nonempty subset T of L(V(G)), there exists a maximal clique Q in H such that L(V(Q))
T.
Remark 1.1. H is generated by Steps 1 to 3 in Section 2, and it is not any other graph which
satisfies the above conditions.
The remaining part of the algorithm is to find out a maximal clique of H whose label set can be
proved to be the label set of a minimum vertex cover of graph G.
Let N[x] = N(x)∪{x}, of which N(x) is the neighbor set of vertex x. Suppose a subgraph of H
which satisfies both of the following conditions.
Condition 1. For any vertex v of the subgraph, there exists a vertex cover C of G such that
L(N[v]) L(C).
Condition 2. For any edge uv of the subgraph, there exists a vertex cover C of G such that
L(N[uv]) L(C), where N[uv] N[u]∩N[v].
Then for each component of H in ascending order of its grade, which is defined in Condition C,
find out a maximal subgraph S of it which satisfies Conditions 1 and 2. Iterate this computation for the
next component until S is nonempty. Then find out a minimal-order subgraph I of S which contains a
nonempty subgraph that satisfies Conditions 1 and 2. If I is a clique, then it is proved by Claim 2.10
that L(V(I)) is the label set of a minimum vertex cover of graph G. If I is not a clique, then go to Stage 2
of the algorithm, which is introduced in Section 5
In Section 3, Stage 1 of the algorithm is proved to run in O(n5+logn) time.
The algorithm was tested by MATLAB programs, and it was found that I is not a clique for only 44
of the tested 711,200 graphs. Then by Claim 2.10, Stage 1 of the algorithm cannot yield a minimum
vertex cover for only about 0.0062% of all tested graphs. The test results are detailed in Section 4.
For the graphs to which Stage 1 of the algorithm is not applicable, a stronger version of Condition
2 is introduced. With this change, the algorithm reaches Stage 2 for the graphs to which Stage 1 is not
1
applicable, and it was found that, with O(n3(5+logn)/2) time, Stage 2 is applicable to, i.e., yields a
minimum vertex cover of, each of the 44 tested graphs. Because Stage 2 is same as Stage 1 except that
Condition 2 is replaced by a stronger version, so Stage 2 actually works for all of the 711,200 tested
graphs.
Furthermore, if there exist graphs to which Stage 2 of the algorithm is not applicable, those graphs
will go to further stages of the algorithm which are expressed in a general form that is consistent with
Stages 1 and 2. Those stages are introduced in Section 6.
2.
STEPS OF THE ALGORITHM
Firstly, graph H is constructed by Steps 1 to 3.
Step 1. Suppose L(V(G)) {1, 2, …, n}, and k log2ni.e., k is the smallest integer which
satisfies n 2k. Let L(V(H)) {1, 2, 3, …,2k}.
Then define Z, which is a family of sets, as follows.
Definition 2.1 Let L(V(H)) be the "first" member of Z. Then partition L(V(H)) into two
equal-sized disjoint sets {1, 2, …, 2k1} and {2k11, 2k12, …,2k}, whose elements are consecutive
numbers, and let the two sets become members of Z. For each of the two "new" members, if its cardinal
number is larger than 1, then continue the partition and "member-assignment" process as described
above. Keep on the process until the cardinal number of each "new" member is 1.
■
Z can also be generated in a reversed way, which is shown in the attached MATLAB program.
For example, if k 3, then Z {{1}, {2}, {3}, {4}, {5}, {6}, {7}, {8}, {1, 2}, {3, 4}, {5, 6}, {7,
8}, {1, 2, 3, 4}, {5, 6, 7, 8}, {1, 2, …, 8}}.
Step 2. l ∈ L(V(G)), generate one and only one first-grade component of H whose label is l,
and denote this component by P({l}).
Remark 2.2 The grade of H’s component is defined in Condition C of Section 1, it is obvious that
the set of all first-grade components of H are actually the set of H’s all isolated vertices.
■
Step 3. Other components of H are generated in sequence by their grades. For example, all of the
second-grade components are to be constructed before the construction of any higher-grade component.
Components of any grade larger than 1 are defined in a general form as follows. But firstly, another
definition is needed.
Definition 2.3. Suppose A is a member of Z, which is defined by Definition 2.1, and b is a
positive integer. Let U ∪P, of which P is H’s component which satisfies: L(V(P)) ⊆ A and g(P) b.
Then we call U a same-grade union of A and b, and denote it by U(A, b). (See Fig. 2.4.)
1
1
2
5
3
5
6
7
6
2
3
4
4
U({1,2,3,4}, 2)
7
8
8
U({5,6,7,8}, 3)
Figure 2.4. Examples of same-grade unions
Each component of U(A, b), where b is larger than 1, is defined as follows.
Definition 2.5. If and only if P is a join of two same-grade unions U(A1, b1) and U(A2, b2), i.e., P
U(A1, b1) U(A2, b2), of which A1, A2, b1 and b2 satisfy the following two conditions:
2.5.1 A1∩A2 and A1∪A2 A, of which A ∈ Z;
2.5.2 b1 and b2 are positive integers, and b1 b2 b,
then let P be one of H’s components denoted by P(A1, b1, A2, b2) or P(A, b1, b2), of which b1
corresponds to A’s "first half" containing smaller numbers.
After and only after for the same b, P of each A1, A2, b1 and b2 which satisfy conditions 2.5.1 and
2
2.5.2 is generated, we start to construct H’s components of grade b+1. When the highest-grade
component of H, which is a 2k-clique, is generated, the construction of H is complete, then we call H a
minimum-covering-computation graph.
■
Then we have
Claim 2.6. For any nonempty subset T of L(V(G)) and any member A of Z which satisfies T A,
there exists a maximal clique Q in U(A, |T|) such that L(V(Q)) T.
Proof. We use induction on |T|. The claim obviously holds when |T| 1.
For the induction step, let T be a subset of L(V(G)) with |T| 2, then by Definition 2.1, we can
suppose Am to be the only minimal member of Z which includes T. Because |T| 2, we can suppose that
Am A1∪A2 of which A1 and A2 are also Z’s members and A1∩A2 , then let T1 A1∩T and T2 A2
∩T. Since Am is the minimal member of Z which includes T, it is obvious that T1,T2 . By the
induction hypothesis, there exists a maximal clique Q1 in U(A1, |T1|) such that L(V(Q1)) T1, and there
exists a maximal clique Q2 in U(A2, |T2|) such that L(V(Q2)) T2. Then by Definition 2.5, there exists a
maximal clique Q in P(A1, |T1|, A2, |T2|) U(A1, |T1|) U(A2, |T2|) such that L(V(Q)) L(V(Q1)∪V(Q2))
T1∪T2 T.
P(A1, |T1|, A2, |T2|) is a component of U(Am, |T|). Because Am is the only minimal member of Z
which includes T, so for any member A of Z which includes T, A includes Am and P(A1, |T1|, A2, |T2|) is
also a component of U(A, |T|), thus there exists a maximal clique Q in U(A, |T|) such that L(V(Q)) T.
■
We need a claim as follows to understand the next step.
Claim 2.7. For any subgraph D of H, there is only one maximal subgraph of D which satisfies
Conditions 1 and 2 in Section 1.
Proof. Assume S1 and S2 are two maximal subgraphs of D which satisfy Conditions 1 and 2, then
S1 ⊈ S2 and S2 ⊈ S1. Let S S1∪S2, then S⊃S1 and S⊃S2. However, it is obvious that S also satisfies
Conditions 1 and 2, which is contradictory to the assumption.
■
Step 4. For each component of H in ascending order of its grade, which is defined in Condition
C of Section 1, by deleting a minimal subgraph of the component, get a maximal subgraph S of it
which satisfies Conditions 1 and 2. Suppose the grade of the current component is b. If S is empty,
iterate this computation for the next component, whose grade is also b or otherwise b+1 when all
components of grade b have been computed. If S is nonempty, record S for the next step.
■
Then we have a claim as follows.
Claim 2.8. For the recorded S of Step 4, the size of its maximum clique is no larger than the
minimum size of graph G’s vertex cover.
Proof. Suppose C is a minimum vertex cover of G, then by Claim 2.6, there exists a maximal
clique Q in U(L(V(H)), |C|) such that L(V(Q)) L(C). Thus Q satisfies Conditions 1 and 2, so there
exists component of U(L(V(H)), |C|) which has nonempty subgraph satisfying Conditions 1 and 2.
Because each component P of H is computed in ascending order of its grade and by Claim 2.7 S is the
only maximal subgraph of P which satisfies Conditions 1 and 2, so for the recorded nonempty S of Step
4, the size of its maximum clique is no larger than |C|.
■
Step 5. For the recorded S of Step 4, by deleting a maximal subset of V(S), get a minimal-order
subgraph I of S which contains a nonempty subgraph SS that satisfies Conditions 1 and 2. Record this I
for the next step.
Remark 2.9. By Claim 2.7, we can tell whether I contains a nonempty subgraph that satisfies
Conditions 1 and 2 by finding out the only one maximal subgraph of I which satisfies the two
conditions.
■
Then for I recorded by Step 5, we have a claim as follows.
Claim 2.10. If I is a clique, then L(V(I)) is the label set of a minimum vertex cover of graph G.
Proof. By Step 5, I contains a nonempty subgraph that satisfies Conditions 1 and 2, so by
Condition 1 or 2, we can tell that there exists a vertex cover C of G such that L(V(I)) L(C).
Suppose b is the size of a maximum clique of the recorded S in Step 4. If I is a clique, then |V(I)|
b because I is a subgraph of S. By Claim 2.8, b and so |V(I)| are no larger than the minimum size of
graph G’s vertex cover. Thus L(V(I)) L(C) is the label set of a minimum vertex cover of G.
■
3
Thus we have
Step 6. Check whether the recorded I of Step 5 is a clique. If I is a clique, then output L(V(I)). If
I is not a clique, then go to the next stage of the algorithm.
Remark 2.11. Sections 5 and 6 explain "the next stage of the algorithm".
■
3.
Claim 3.1
EFFICIENCY OF THE ALGORITHM
For L(V(H)) {1, 2, 3, …,2k}(k 1),
| V H |
k
i 1
2
i
2 .
Proof.
When k 1, H has two first-grade components and one second-grade component, and so
V(H) 4 (2k2).
We use induction on k and assume the claim holds for any k when 1 k h.
Now for k h 1, L(V(H)) {1, 2, 3, …,2h1}. By Steps 1 to 3 of the algorithm, we can suppose
two isomorphic minimum-covering-computation graph H1 and H2 which satisfy: L(V(H1)) {1, 2,
3, …,2h} and L(V(H2)) {2h 1, 2h 2,…,2h1}. Then by the induction hypothesis,
| V H 1 |
h
i 1
2
2 .
i
Because H2 is isomorphic to H1, V(H2) V(H1). Suppose P is a component of H1 or H2, then by
Steps 1 to 3 of the algorithm, we can tell that P is also a component of H. Thus we have
Corollary 3.1.1. Let X ∪P, of which P is H’s component and P is also H1’s or H2’s component,
h
then
| V X | 2
2i 2 .
i 1
Now the other vertex-amount we need to count is of H’s components of which each is neither H1’s
nor H2’s component. Suppose P* is such a kind of component, then by Definition 2.1 and 2.5, we can
tell that L(V(P*)) ⊈ L(V(H1)), L(V(P*)) ⊈ L(V(H2)) and so L(V(P*)) L(V(H)). Thus by Definition 2.5,
P* U(A1, b1) U(A2, b2) U(L(V(H1)), b1) U(L(V(H2)), b2).
In the above formula, for each U(L(V(H1)), b1), b2 of U(L(V(H2)), b2) can be any integer from 1 to
|L(V(H2))| 2h, which means that each U(L(V(H1)), b1) was copied for 2h times to generate different
components of H. Because
2h
H1 b 1U L V H1 , b1 ,
1
so H1 was copied for 2h times to construct H. And there is a same conclusion for H2. Combine this
conclusion with Corollary 3.1.1, then we have
V(H) V(X) + 2h (V(H1) + V(H2))
h
h
h
2 2 2 ( 2 2 2 2)
i1
i1
i1
h 1
2 2 .
i 1
i
2
h
i
i
i
Thus the induction succeeds and the claim is proved.
■
Claim 3.2. Suppose graph G of order n is the input and T1(n) is the running time of Stage 1 of
the algorithm, then T1(n) O(n5+logn).
Proof. By Step 1, k is the smallest integer which satisfies n 2k, so 2k1 n. Then by Claim 3.1,
| V H |
k
i 1
2
i
2
k
i 1
2 2
i 1
k ( k 3) / 2
2n
( log 2 n 4 ) / 2
.
Thus the asymptotic upper bounds for |V(H)| and |E(H)| can be n(5+logn)/2 and n5+logn, respectively, so
the construction of H runs in O(n5+logn) time.
Then in Step 4, the running time for computing subgraphs of H’s components by Conditions 1 and
2 is still no more than O(n5+logn), and this is also the case for Step 5. Therefore Stage 1 of the algorithm
takes no more than O(n5+logn) time.
■
4
4.
TESTING OF THE ALGORITHM
The algorithm was tested by the two attached programs, "Generating_k_and_H" and "Testing_G",
on MATLAB (R2016a, trial use). The first file generates and saves k and H for the second file. The
second file generates random graphs of random "edge density", i.e., random probability of each edge,
for testing of the algorithm. "Edge density" in the program approximately equals to the ratio of |E(G)|
to |E(K)|, of which K is a 2k-order complete graph, and it is generated from a normal distribution with
mean 0.5 and standard deviation 0.2,0.3,0.4,0.5 or 0.6 for the testing. The reasons for the introduction
of this distribution are that there exist the most graphs, no matter labelled or unlabeled, when "edge
density" equals to 0.5, and there exists a same amount of graphs of "edge density" being d as that of
"edge density" being 1 d.
An ordinary personal computer is used for the testing. Testing parameters and results are listed as
follows.
k, |V(G)|
3, 8
4, 16
4, 16
4, 16
4, 16
4, 16
5, 32
5, 32
5, 32
Number of
tested
graphs
100,000
50,000
220,000
150,000
130,000
60,000
400
400
400
Table 4.1. Testing parameters and results
Standard
Approximate
No. of graphs to
deviation of
running time
which Stage 1 is
"edge density" (for Testing_G.m)
not applicable
0.3
16 minutes
0
0.2
10 hours
2
0.3
43 hours
19
0.4
29 hours
8
0.5
25 hours
9
0.6
12 hours
2
0.3
58 hours
1
0.4
58 hours
2
0.5
58 hours
1
No. of graphs to
which Stage 2 is
not applicable
0
0
0
0
0
0
0
0
0
Remark 4.2. Stage 2 is explained in Section 5.
Remark 4.3. When k 5, the running time for the computer used for testing is so long that it is
impossible to test a large number of graphs in a short period. There is no test data for k 5 for the same
reason.
k, |V(G)|
4, 16
5, 32
Table 4.4. Recorded "edge densities" of graphs to which Stage 1 is not applicable
Recorded "edge densities" of graphs to which Stage 1 is not applicable
0.2410, 0.2599, 0.2973, 0.3170, 0.3173, 0.3204, 0.3323, 0.3359, 0.3533, 0.3551, 0.3642,
0.3773, 0.3774, 0.3942, 0.4089, 0.4232, 0.4436, 0.4475, 0.4495, 0.4500, 0.4509, 0.4539,
0.4821, 0.4919, 0.4938, 0.5032, 0.5318, 0.5337, 0.5343, 0.5356, 0.6200, 0.6476
0.1479, 0.1889, 0.2127, 0.2868
By Table 4.1, when k 3, Stage 1 is applicable to 100% of tested random graphs (i.e. Step 5 of
Stage 1 yields a clique and, by Claim 2.10, Step 6 yields the label set of a minimum vertex cover of
each tested graphs for k 3); when k 4, Stage 1 is applicable to about 99.99% of tested random
graphs; and when k 5, Stage 1 is applicable to about 99.67% of tested random graphs.
By Table 4.4, the recorded "edge densities" of tested graphs to which Stage 1 is not applicable
scatter.
Except for k 3, the numbers of tested graphs are much less than the numbers of unlabeled graphs
of respective orders. However, because of the large number of tested graphs, and because the
possibility of duplicated graphs to which the algorithm is applicable is same as that of duplicated
graphs to which the algorithm is not applicable, it is reasonable to believe that the "applicability ratios"
of tested graphs are close to the "applicability ratios" of all unlabeled (or labelled) graphs of respective
orders.
5
5.
STAGE 2 OF THE ALGORITHM
For those graphs to which Stage 1 of the algorithm is not applicable, a stronger version of
Condition 2 is introduced as follows.
Condition 2+. For any edge uv of the subgraph, there exists a vertex cover C of G such that
L(N[uv]) L(C), where
N[uv] {w ∈ N[u]∩N[v] : There exists a vertex cover C* of G such that L(N[u]∩N[v]∩N[w])
L(C*)}.
■
With this change, the algorithm goes to Stage 2 for those graphs to which Stage 1 is not applicable.
Steps 1 to 3 are for construction of H and apparently they do not need to run again. Besides, for those
components of H which have been computed in Step 4 of Stage 1 and concluded that they do not have
nonempty subgraph satisfying Conditions 1 and 2, they do not need to be computed again in Stage 2
because Condition 2+ is stronger than Condition 2. Thus Stage 2 starts at Step 4 for the component of H
which has nonempty subgraph satisfying Conditions 1 and 2, and the following computation is same as
Stage 1 except that Condition 2 is substituted by Condition 2+.
Stage 2 of the algorithm has been implemented in the attached MATLAB program. It was found
that Stage 2 yielded cliques for all of the 44 tested graphs for which Stage 1 did not. Because Claims
2.8 and 2.10 still hold for Stage 2, so Stage 2 is applicable to all of the 44 tested graphs. Then since
Condition 2+ is stronger than Condition 2, we can tell that Stage 2 actually works for all of the 711,200
tested graphs.
By the definition of N[uv] in Condition 2+, 3-cliques of S in Step 4 and SS in Step 5 require
computation, thus the running time for the algorithm that reaches Stage 2 as its final stage is
O(n3(5+logn)/2).
However, the definition of N[uv] in Condition 2+ looks "strange", so it requires an expression in a
more general way as shown in the following section.
6.
FURTHER STAGES OF THE ALGORITHM
Because it has not been theoretically proved that Stage 2 of the algorithm is applicable to all
graphs, it is worthwhile to conceive further stages of the algorithm which shall be in a general form that
is consistent with Stages 1 and 2. Then the concept of hyperedge of hypergraph is needed.
A hypergraph consists of a collection of vertices and a collection of hyperedges. If the vertex set is
V, then the hyperedges are subsets of V. Then for Stage t of the algorithm, where t is a positive integer,
we give the following five rules for the algorithm:
1. Suppose Es(H) is the set of all hyperedges of size s in H, then let
Es(H) = {e : e is the vertex set of an s-clique in H and H is generated by Steps 1 to 3}, s
t+1;
Es(H) = , s t+1.
2. For any hyperedge e of size no larger than t, let N[e] {x : {x}∪e is a subset of a hyperedge
of H}.
3. For any hyperedge e of size t+1, let N[e] {x : x ∈ N[{v}], v ∈ e}.
4. By deleting hyperedges in H, subgraph S in Step 4 and subgraph SS in Step 5 shall satisfy
Condition X. For any hyperedge e of the subgraph, there exists a vertex cover C of G such that
L(N[e]) L(C).
5. If Stage t does not yield clique at Step 5, start Stage t+1 at Step 4 for the component of H
where Stage t ends.
Then it is not difficult to show that the above rules are consistent with Stages 1 and 2 of the
algorithm.
Claims 2.8 and 2.10 can be easily proved to hold for any stage of the algorithm, so Stage t yields a
minimum vertex cover of graph G at Step 6 if it yields a clique at Step 5.
Because in Stage t, hyperedges of size t+1, i.e., vertex sets of (t+1)-cliques, of S in Step 4 and SS
in Step 5 require computation, the algorithm runs in O(n(t+1) (5+logn)/2) time if the "maximum" stage to
6
which it reaches is Stage t.
7.
CONCLUSIONS
Conclusions for this algorithm are summarized as follows.
A. For minimum vertex cover problem of an order-n graph, the algorithm runs in O(n(t+1) (5+logn)/2)
time if the "maximum" stage to which it reaches is Stage t, where t is a positive integer.
Therefore, if the "maximum" stage to which the algorithm reaches for a graph is Stage 1 or 2,
then the algorithm runs in O(n5+logn) or O(n3(5+logn)/2) time for the graph respectively.
B. Stage 1 of the algorithm is applicable to, i.e., yields a proved minimum vertex cover for, each
of more than 99% of tested graphs of order no larger than 32, while Stage 2 of the algorithm
works for 100% of tested graphs of order no larger than 32. And it is reasonable to believe that
the above ratios are close to the "real applicability ratios" for this algorithm to all unlabeled
(or labelled) graphs of order no larger than 32.
C. The "applicability ratio" of Stage 1 for all tested order-8 graphs, order-16 graphs and order-32
graphs is 100%, 99.99% and 99.67% respectively, so it is unlikely that the "applicability ratio"
of Stage 1 decreases sharply when the order of graph increases. Besides, Stage 2 works for
each of the tested graphs and we still have stages beyond Stage 2. Therefore it is an efficient
algorithm which is already applicable to practical use.
D. As summarized above, the performance of this algorithm is extraordinary for NP-Complete
problems. Although there is no systematic and theoretical explanation for it yet, it is at least an
important finding which is valuable for further research. However, like many findings or
conjectures in mathematics, its theoretical explanation may take many years even decades to
be found. Thus it is decided to make the algorithm public so that the explanation and
improvement of it become possible.
E. Further tests for graphs of order 64 or more are necessary and valuable. However, they require
computer of large capacity which is not available for the author.
Reference
[1] Levitin, A.: Introduction to the Design and Analysis of Algorithms. Pearson Education, Boston
(2007)
Appendix 1.
Generating_k_and_H.m
% This MATLAB file is the first program to test the algorithm in the
manuspcript and shall run before the running of file "Testing_G".
function Generating_k_and_H
k=input('When the order of each randomly generated graph is 2^k, k equals
to:');
if (k<=0)||(k<ceil(k))
error('k must be positive interger')
end
if k==6
disp('Note: This program requires large memory when k is larger than
5. The size of the swap file is recommended to be set to 100GB-200GB for
k=6.')
7
disp('And the running time for k=6 on an ordinary personal computer
is about 2 to 3 hours.')
end
if k>=7
disp('Note: For k>6, this program requires large memory which may
exceed the capacity of the computer,')
disp('and the running time on an ordinary personal computer is very
long.')
end
if k>7
error('For k>7, all of the uint8 class of matrix in this program shall
be changed to other class.')
end
Z=Generating_Z(k);
H=Generating_H(k,Z);
disp('Saving file "k_and_H". Please wait.') % When k is larger than 5,
the following saving of file 'k_and_H' may take a long time.
save('k_and_H','k','H','-v7.3')
end
% Step 1
function Z=Generating_Z(k)
% Cell-array Z represents the family of sets by Definition 2.1 of the
algorithm, while its non-empty cells are arrays which represent the members
of
% family Z. Z's cells of size 1 are first generated as follows.
Z=cell(k+1,2^k);
for c=1:2^k
M=c;
Z{1,c}=M;
end
% Z's cells of larger size are generated as follows. Non-empty cells of
the rth row of cell-array Z represent family Z's member of size 2^(r-1).
for r=2:k+1
for c=1:2^(k+1-r)
Z{r,c}=union(Z{r-1,2*c-1}, Z{r-1,2*c});
end
end
end
% Steps 2&3
function H=Generating_H(k,Z)
% In this program, component P of graph H is expressed by its adjacency
matrix where diagonal numbers are the labels of corresponding vertices
8
% instead of zeros. For first-grade component P(Z{1,d2}), H{1,d2,1,1} is
assigned to be its adjacency matrix. However, for higher-grade component
% P(Z{d1,d2},d3,d4) where d1>1, H{d1,d2,d3,d4} is assigned to be its
adjacency matrix.
H=cell(k+1,2^k,2^(k-1),2^(k-1));
% k+1=size(Z,1). 2^k=size(Z,2).
% When component P's grade is larger than 2^(k-1), P is the join of
same-grade unions U(Z{k,1},d3) and U(Z{k,2},d4). Because
|Z{k,1}|=|Z{k,2}|=2^(k-1),
% so both d3 and d4 are no larger than 2^(k-1) for any H{d1,d2,d3,d4}.
U=cell(k+1,2^k,2^(k-1));
% Same-grade union U(Z{d1,d2},D3) is expressed by its adjacency matrix
U{d1,d2,D3} where diagonal numbers are the labels of corresponding
vertices
% instead of zeros.
% Matrices for the first-grade components and first-grade unions are first
generated as follows.
for d2=1:2^k
H{1,d2,1,1}=uint8(d2); % The uint8 class is used to save memory and
time. However, it shall be changed to other class when k is larger than
7.
U=Generating_U(U,k,H{1,d2,1,1},1,d2,1);
end
% Matrices for higher-grade components and higher-grade unions are
generated as follows.
for b=2:2^k
disp(['Proceeding to b = ',num2str(b),' of ',num2str(2^k)]) % This
command is to show the progress of the running program when k is large.
for d3=max(1,b-2^(k-1)):min(b-1,2^(k-1))
% b=d3+d4. b3 and b4 are positive integers and both maximum values
of d3 and d4 for H{d1,d2,d3,d4} are 2^(k-1).
d4=b-d3;
for d1=1:k
if (size(Z{d1,1},2)<d3)||(size(Z{d1,1},2)<d4)
% By Definition 2.5, for component P(Z{d1+1,*},d3,d4) and
nonempty H{d1+1,*,d3,d4}, d3 and d4 are not larger than size(Z{d1,1},2).
continue
end
for d2=1:2:2^k-1
if isempty(Z{d1,d2+1})==0
% Z{d1,d2} and Z{d1,d2+1} represent A1 and A2 in
Definition 2.5, respectively.
sx=size(U{d1,d2,d3},1); % Matrix U{d1,d2,d3} expresses
same-grade union U(Z{d1,d2},d3)
9
sy=size(U{d1,d2+1,d4},1); % Matrix U{d1,d2+1,d4}
expresses same-grade union U(Z{d1,d2+1},d4)
H{d1+1,(d2+1)/2,d3,d4}=uint8([U{d1,d2,d3},ones(sx,sy);ones(sy,sx),U{d
1,d2+1,d4}]);
% H{d1+1,(d2+1)/2,d3,d4} is the adjacency matrix of
graph H's component P(Z{d1+1,(d2+1)/2},d3,d4), which is the join of
% U(Z{d1,d2},d3) and U(Z{d1,d2+1},d4).
% Z{d1+1,(d2+1)/2}=union(Z{d1,d2},Z{d1,d2+1}).
if d1<k
% When d1=k, H{d1+1,*,*,*} expresses component of
same-grade union U(Z{k+1,1},*). Because |Z{k+1,1}| = 2^k, i.e., Z{k+1,1}
% contains all labels, so U(Z{k+1,1},*) will not be
used to generate any component of H and thus they are not generated here.
U=Generating_U(U,k,H{d1+1,(d2+1)/2,d3,d4},d1+1,(d2+1)/2,b);
end
else % which indicates Z{d1,d2+1} is empty, then go to the
next d1.
break
end
end
end
end
end
end
function U=Generating_U(U,k,P,a1,a2,b)
% This function is to put component P(Z{a1,a2},b1,b2) into each same-grade
union U(Z{d1,d2},b) of which b=b1+b2 and Z{a1,a2} is a subset of Z{d1,d2}.
d1=a1;
d2=a2;
while d1<=k
% As explained in lines 80 and 81, U(Z{k+1,*},*) will not be used to
generate any cell of H and thus they are not generated here.
sizeU=size(U{d1,d2,b},1);
sizeP=size(P,1);
U{d1,d2,b}=uint8([U{d1,d2,b},zeros(sizeU,sizeP);zeros(sizeP,sizeU),P]
);
% The above command puts component P(Z{a1,a2},b1,b2) into same-grade
union U(Z{d1,d2},b).
d1=d1+1; % It is obvious that for any d1, there exists at most one
Z{d1,d2} which includes Z{a1,a2}.
10
d2=ceil(d2/2); % If Z{d1,d2} includes Z{a1,a2}, then
Z{d1+1,ceil(d2/2)} includes Z{a1,a2}.
end
end
Appendix 2.
Testing_G.m
% This MATLAB file is the second program to test the algorithm in the
manuspcript. Another file named "Generating_k_and_H.m" needs to run
first.
function Testing_G
% By generating m random graphs, of which the order is 2^k, the program
finds out how many of the tested random graphs do not yield clique in Step
5
% and thus are not the graphs to which the algorithm is applicable. (If
a clique is yielded in Step 5, then by Claim 2.10, its label set is the
label
% set of a minimum vertex cover of the tested graph.)
m=input('How many random graphs are to be generated and tested for the
algorithm?');
if (m<=0) || (m<ceil(m))
error('m must be positive integer')
end
disp('Loading file "k_and_H". Please wait.') % When k is larger than 5,
the following loading of file 'k_and_H' may take a long time.
load('k_and_H');
% File "k_and_H.mat" is generated by
"Generating_k_and_H.m".
if k==5
disp('Note: When k=5, the running time for an ordinary personal
computer to compute a graph is about 9 minutes.')
end
if k>5
disp('Note: When k>5, the running time for an ordinary personal
computer to compute a graph is very long.')
end
n1=0; % n1 will be the number of tested graphs to which Stage 1 of the
algorithm is not applicable.
n2=0; % n2 will be the number of tested graphs to which Stage 2 is not
applicable.
d1=[]; % d1 will record the approximate edge density of each graph to which
Stage 1 is not applicable.
d2=[]; % d2 will record the approximate edge density of each graph to which
Stage 2 is not applicable.
11
for h=1:m
disp(['Computing the ',num2str(h),'th random graph'])
density=2;
while density<=0 || density>1
density=0.5+0.3*randn;
end
% "density" will approximately equals to the ratio of |E(G)| to |E(K)|,
of which K is a 2^k-order complete graph, and it is generated from a normal
% distribution with mean 0.5 and standard deviation 0.2,0.3,0.4,0.5
or 0.6.
G=Generating_G(k,density);
if G==zeros(size(G))
disp(['The ',num2str(h),'th random graph has no edge.'])
continue
end
stage=1;
bstart=2;
[S1,bstart]=Step_4(stage,bstart,k,G,H);
I1=Step_5(stage,S1,G);
result1=Step_6(I1);
if result1==1 % which indicates I1 is a clique.
disp(['Stage 1 of the algorithm is applicable to the
',num2str(h),'th random graph.']);
else
disp(['Stage 1 of the algorithm is NOT applicable to the
',num2str(h),'th random graph.']);
n1=n1+1;
d1(length(d1)+1)=density;
stage=2;
[S2,~]=Step_4(stage,bstart,k,G,H);
I2=Step_5(stage,S2,G);
result2=Step_6(I2);
if result2==1 % which indicates I2 is a clique.
disp(['Stage 2 of the algorithm is applicable to the
',num2str(h),'th random graph.']);
else
disp(['Stage 2 of the algorithm is NOT applicable to the
',num2str(h),'th random graph.']);
n2=n2+1;
d2(length(d2)+1)=density;
end
end
disp(['Currently, Stage 1 of the algorithm is NOT applicable to
',num2str(n1),' of ',num2str(h),' computed random graphs.'])
12
disp(['Currently, Stage 2 of the algorithm is NOT applicable to
',num2str(n2),' of ',num2str(h),' computed random graphs.'])
end
d1
disp('d1 records the approximate edge density of each graph to which Stage
1 is not applicable.')
d2
disp('d2 records the approximate edge density of each graph to which Stage
2 is not applicable.')
end
function G=Generating_G(k,density)
G=sprandsym(2^k,density,0.1);
% sprandsym(n,density) is a symmetric random, n-by-n, sparse matrix with
approximately density*n*n nonzeros. However, when density equals to 1,
the
% function still generates some zeros. For better performance,
sprandsym(n,density,rc) is used here and it works much better, although
ocassionly still
% generates zeros, for density=1.
G(logical(eye(2^k)))=0;
G=spones(G);
end
function [S,bstart]=Step_4(stage,bstart,k,G,H)
% For each component P of graph H in ascending order of its grade, by
deleting a minimal subgraph of P, get P's maximal subgraph S which satisfies
% Conditions 1 and 2 (or 2+ for Stage 2). Iterate this computation for
the next component until P's S is nonempty.
% In this program, component P of graph H is expressed by its adjacency
matrix, where diagonal numbers are the labels of corresponding vertices
% instead of zeros. For first-grade component P(Z{1,d2}), H{1,d2,1,1} is
assigned to be its adjacency matrix. However, for higher-grade component
% P(Z{d1,d2},d3,d4) where d1>1, H{d1,d2,d3,d4} is assigned to be its
adjacency matrix.
if stage==1
for d2=1:2^k % This "for loop" is only for graph H's first-grade
components. Stage 2 never starts from any first-grade component because
a
% nonempty subgraph of a first-grade component is always a clique.
P=H{1,d2,1,1};
S=Maximal_subgraph_satisfying_conditions_1and2_or_1and2plus(stage,P,G
);
13
if isempty(S)==0 % which indicates P(Z{1,d2}) has a nonempty
subgraph S, which must be P(Z{1,d2}) itself, that satisfies Conditions
1 and 2.
return
end
end
end
for b=bstart:2^k % This "for loop" is for graph H's components of grade
larger than 1. Thus bstart=2 for Stage 1, while Stage 2 starts at the grade
% of H's component which has nonempty subgraph satisfying Conditions
1 and 2.
for d3=max(1,b-2^(k-1)):min(b-1,2^(k-1))
% b=d3+d4. d3 and d4 must be positive integers and both maximum
values of d3 and d4 for H{d1,d2,d3,d4} are 2^(k-1).
d4=b-d3;
for d1=2:k+1
for d2=1:2^k
P=H{d1,d2,d3,d4};
if isempty(P)==0
% which indicates this cell represents
a nonempty component P(Z{d1,d2},d3,d4).
S=Maximal_subgraph_satisfying_conditions_1and2_or_1and2plus(stage,P,G
);
if isempty(S)==0 % which indicates P(Z{d1,d2},d3,d4)
has a nonempty subgraph S which satisfies conditions 1 & 2(or 2+ for Stage
2).
bstart=b; % This command records the grade of H's
component at which the possible Stage 2 will start.
return
end
end
end
end
end
end
end
function I=Step_5(stage,S,G)
% For the recorded S of Step 4, by deleting a maximal subset of V(S), get
a minimal-order subgraph I of S which contains a nonempty subgraph SS that
% satisfies Conditions 1 and 2(or 2+ for Stage 2). Record this I for the
next step.
v=1;
while v<=size(S,1)
14
% The vth vertex of S is to be checked whether it can be deleted, i.e.,
whether S still has a nonempty subgraph which satisfies Conditions 1 and
2
% (or 2+ for Stage 2) after the vth vertex of S is deleted.
I=S;
I(v,:)=[];
I(:,v)=[];
SS=Maximal_subgraph_satisfying_conditions_1and2_or_1and2plus(stage,I,
G);
if isempty(SS)==0 % which indicates that I still has a nonempty
subgraph which satisfies Conditions 1 and 2(or 2+ for Stage 2).
S=SS; % If this command were S=I instead of S=SS, then the next
round of function
"Maximal_subgraph_satisfying_conditions_1and2_or_1and2plus"
% would delete again the edges of E(I)-E(SS).
% For any i which satisfies 1 <= i < v, it is obvious that the ith
vertex of the last S remains not deletable and so it is still the ith vertex
% of the current S.
else % which indicates the vth vertex of S shall not be deleted.
v=v+1;
end
end
I=S;
end
function result=Step_6(I)
% This function checks whether I represents a clique.
if size(I,1)==1
result=1;
return
end
for i=1:size(I,1)-1 % Because of symmetry of matrix I, only the "upper
half" of the matrix is checked.
for j=i+1: size(I,1)
if I(i,j)==0
result=0;
return;
end
end
end
result=1;
end
15
function
M=Maximal_subgraph_satisfying_conditions_1and2_or_1and2plus(stage,D,G
)
% For the graph defined by adjacency matrix D, this function finds out
its maximal subgraph which satisfies Conditions 1 and 2 (or 2+ for Stage
2).
if isempty(D)==1
M=D;
return
end
x=1;
while x==1
DD=A_round_of_edge_deleting_by_Condition_2or2plus(stage,D,G);
if DD==D % which indicates all of the edges of the graph expressed
by D satisfy Condition 2 (or 2+ for Stage 2).
break
end
% At this point, DD~=D, which indicates certain edges of the graph
expressed by D were deleted in the last round of edge-deleting and D becomes
DD.
% However,the graph expressed by DD may still have edges which do not
satisfy Condition 2 (or 2+ for Stage 2), so the next round of edge-deleting
% is necessary until DD=D.
D=DD;
end
% At this point, all edges of the graph expressed by D satisfy Condition
2 (or 2+ for Stage 2). Then it is obvious that for any vertex which has
% incident edge, Condition 1 is satisfied. Thus the following part of this
function is to delete the isolated vertices which do not satisfy Condition
1.
i=1;
while i<=size(D,1) % The ith vertex of D will be checked whether it is
an islated vertex and whether it satisfies Condition 1.
A=D(i,:); % Thus A(i)=D(i,i) is the label of the ith vertex of D.
A(i)=0;
if A==zeros(size(A))
% which indicates the ith vertex of D has no incident edge, then
we need to check whether the label of this vertex is also the label of
a G's
% vertex cover of size 1 as follows.
GG=G;
GG(D(i,i),:)=0; % The label of the xth vertex of G is x.
GG(:,D(i,i))=0;
% D(i,i) is the label of the ith vertex of D. Suppose j=D(i,i).
16
If j is the label of a G's vertex cover of size 1, then for any G(r,c)~=0,
% r=j or c=j. (Note that all of the diagonal numbers of matrix G
are zeros.)
if GG==zeros(size(G))
% which indicates the label of the ith vertex of D is the label
of a G's vertex cover of size 1, then this vertex satisfies Condition 1
% and shall not be deleted.
i=i+1;
else % Then this vertex does not satisfy Condition 1 and shall
be deleted.
D(i,:)=[];
D(:,i)=[];
end
else % which indicates the ith vertex of D has incident edge and shall
not be deleted.
i=i+1;
end
end
M=D;
end
function DD=A_round_of_edge_deleting_by_Condition_2or2plus(stage,D,G)
% For each round of edge-deleting, all edges which do not satisfy Condition
2 (or 2+ for Stage 2) of the input graph expressed by D are deleted.
% However, the deletion may cause some more edges which no longer satisfy
Condition 2 (or 2+ for Stage 2), and these edges may not be all deleted
in
% this round of edge-deleting.
i=2;
while i<=size(D,1)
for j=1:i-1 % Because matrix D is symmetrical, only the part under
its diagonal is checked.
if D(i,j)==1
dia=diag(D); % dia(1,x) is the label of the xth vertex of D.
Ne2=D(i,:).*D(j,:);
% ".*" is element-by-element multiplication, so Ne2(1,x) is
non-zero only if both D(i,x) and D(j,x) are non-zero. Therefore, Ne2(1,x)
is
% non-zero only if x=i, x=j, or the xth vertex is adjacent to
both the ith and the jth vertices, so Ne2=N[ij] for Stage 1.
if stage==2 % then Condition 2+ shall be satisfied instead of
Condition 2.
for k=1:size(D,1)
if k~=i && k~=j && Ne2(1,k)==1 % which indicates the
17
kth vertex is adjacent to both the ith and the jth vertices.
Ne3=Ne2.*D(k,:); % Similar as explained for Ne2,
Ne3(1,x) is non-zero only if x=i, x=j, x=k, or the xth vertex is adjacent
to
% all of the ith, jth and kth vertices.
LNe3=dia(logical(Ne3)); % LNe3 is the label set of
Ne3.
C=Does_the_label_set_cover_E_of_G(LNe3,G);
if C==0 % which indicates the 3-clique of vertices
i,j and k does not satisfy the definition of N[uv] in Condition 2+, so
% vertex k shall not be considered to be an element
of N[ij] for Stage 2.
Ne2(1,k)=0;
end
end
end
% At this point, Ne2=N[ij] for Stage 2.
end
LNe2=dia(logical(Ne2));
% LNe2 is the label set of Ne2=N[ij].
C=Does_the_label_set_cover_E_of_G(LNe2,G);
if C==0
% which indicates the edge corresponding to D(i,j)
does not satisfy Condition 2 (or 2+ for Stage 2).
D(i,j)=0;
D(j,i)=0;
end
end
end
i=i+1;
end
DD=D;
end
function C=Does_the_label_set_cover_E_of_G(LN,G)
% This function checks whether LN is a label set of a vertex cover of G.
G(LN,:)=0;
G(:,LN)=0;
if G==zeros(size(G))
C=1;
else
C=0;
end
end
18
| 8cs.DS
|
Working Paper
Computing and Testing Pareto Optimal Committees
arXiv:1803.06644v1 [cs.GT] 18 Mar 2018
Haris Aziz · Jérôme Lang · Jérôme Monnot
Abstract Selecting a set of alternatives based on the preferences of agents is an important problem in committee selection and beyond. Among the various criteria put
forth for desirability of a committee, Pareto optimality is a minimal and important
requirement. As asking agents to specify their preferences over exponentially many
subsets of alternatives is practically infeasible, we assume that each agent specifies
a weak order on single alternatives, from which a preference relation over subsets is
derived using some preference extension. We consider five prominent extensions (responsive, downward lexicographic, upward lexicographic, best, and worst). For each
of them, we consider the corresponding Pareto optimality notion, and we study the
complexity of computing and verifying Pareto optimal outcomes. We also consider
strategic issues: for four of the set extensions, we present a linear-time, Pareto optimal
and strategyproof algorithm that even works for weak preferences.
Keywords committee selection · multiwinner voting · Pareto optimality · algorithms
and complexity · set extensions.
JEL Classification: C70 · D61 · D71
1 Introduction
Pareto optimality is a central concept in economics and has been termed the “single
most important tool of normative economic analysis” (Moulin, 2003). An outcome
H. Aziz
Data61, CSIRO and UNSW, Sydney 2052 , Australia
Tel.: +61-2-8306 0490
Fax: +61-2-8306 0405
E-mail: [email protected]
J. Lang and J. Monnot
LAMSADE, Université Paris-Dauphine
Paris, France
E-mail: {lang, jerome.monnot}@lamsade.dauphine.fr
2
Haris Aziz et al.
is Pareto optimal if there does not exist another outcome that all agents like at least
as much and at least one agent strictly prefers. Although Pareto optimality has been
considered extensively in single-winner voting and other social choice settings such
as fair division or hedonic games, it has received only little attention in multiwinner
voting, in which the outcomes are sets of alternatives. Multiwinner voting applies
to selecting a set of plans or a committee, hiring team members, movie recommendations, and more. For convenience, we use the terminology “committee” even if
our results have an impact far beyond committee elections (Faliszewski et al., 2017;
Aziz et al., 2017).
In single-winner voting setting, agents express preferences over alternatives and
a single alternative is selected. Pareto optimality in this context is straightforward to
define, achieve, and verify. In multiwinner voting, a well-known difficulty is that it
is unrealistic to assume that agents will report preferences over all possible committees, since there is an exponential number of them. For this reason, most approaches
assume that they only report a small part of their preferences, and that some extension principle is used to induce a preference over all possible subsets from this ‘small
input’ over single alternatives (Barberà et al., 2004). Such preference extensions are
also widely used in other social choice settings such as fair division or matching. The
most two widely used choices of ‘small inputs’ in multiwinner voting are rankings
(linear orders) over alternatives and sets of approved alternatives. In this paper we
make a choice that generalizes both of them: agents report weak orders over single
alternatives. Then we consider five prominent preference extension principles: the
responsive extension, where a set of alternatives S is at least as preferred as a set of
alternatives T if S is obtained from T by repeated replacements of an alternative by
another alternative which is at least as preferred; the optimistic, or ‘best’ (respectively
pessimistic, or ‘worst’) extension, which orders subsets of alternatives according to
their most (respectively, least) preferred element; the downward lexicographic extension, a lexicographic refinement of the optimistic extension, and the upward lexicographic extension, a lexicographic refinement of the pessimistic (worst) extension.
The responsive extension (Barberà et al., 2004; Roth and Sotomayor, 1990) can
be seen as the ordinal counterpart of additivity. The downward lexicographic extension has been considered in various papers (Bossert, 1995; Lang et al., 2012;
Klamler et al., 2012). The ‘best’ set extension has been considered in a number of
approaches such as full proportional representation (Chamberlin and Courant, 1983;
Monroe, 1995) and other committee voting settings Elkind et al. (2015). The ‘worst’
set extension, also used by Klamler et al. (2012) and Skowron et al. (2015b), captures
settings where the impact of a bad alternative in the selection overwhelms the benefits of good alternatives: for instance, when the decision about a crucial issue will
be made by one of the members of the committee but the agent ignores which one
and is risk-averse; or the case of a parent’s preferences over a set of movies to be
watched by a child. The ‘best’ and ‘worst’ set extensions have been used in coalition
formation (Aziz and Savani, 2016; Cechlárová, 2008).
Although set extensions have been implicitly or explicitly considered in multiwinner voting, most of the computational work has dealt with specific voting rules
(see the related work section). Instead, we concentrate on Pareto optimality, consider
the computation and verification of Pareto optimal committees, as well as the exis-
Computing and Testing Pareto Optimal Committees
3
Computation
Verification
Responsive (RS)
in PIC (Th. 5)
coNP-C (Th. 2), W[2]-hard
in P—dich. prefs and tw ≤ 2 (Th. 4)
Downward Lexicographic (DL)
in PIC (Th. 8)
coNP-C (Th. 1), W[2]-hard
Set Extension
IC
Upward Lexicographic (UL)
in P
Best (B)
NP-hard (Th. 7)
in P for strict prefs
(Th. 11)
coNP-C (Th. 2), W[2]-hard
coNP-C, W[2]-hard (Th. 6)
Worst (W)
in PIC (Th. 10)
in P (Th. 9)
Table 1 Complexity of computing and verifying Pareto optimal committees. PIC (coined by Christos Papadimitriou in a seminar at Simons Institute in 2015) indicates a class of problems in which agents provide
the input and the problems admit a strategyproof and polynomial-time algorithm.
tence of a polynomial-time and strategyproof algorithm that returns Pareto optimal
outcomes.
Contributions We consider Pareto optimality with respect to the five aforementioned
preference set extensions. We present various connections between the Pareto optimality notions. For each of the notions, we undertake a detailed study of complexity
of computing and verifying Pareto optimal outcomes. Table 1 summarizes the complexity results.
An important take-home message of the results is that testing Pareto optimality or
obtaining Pareto improvements over status-quo committees is computationally hard
even though computing some Pareto optimal committee is easy. For responsive and
downward lexicographic extensions we give a complete characterization of the complexity of testing Pareto optimality when preferences are dichotomous and the size of
top equivalence class is two: unless P = NP, Pareto optimality can be tested in polynomial time if and only if the size of the first equivalence classes is at most two. For the
‘best’ extension, we show that even computing a Pareto optimal outcome is NP-hard.
Another interesting contrast with the responsive set extension is that even when preferences are dichotomous and the size of top equivalence class is two, testing Pareto
optimality is coNP-complete. In contrast to the other extensions, for the ‘worst’ extension, both problems of computing and verifying Pareto optimal outcomes admit
polynomial-time algorithms.
We also consider the requirement of strategyproofness on top of Pareto optimality. We show that there exist linear-time Pareto optimal and strategyproof algorithms
for committee voting even for weak preferences for four of the five set extensions.
The algorithms can be considered as careful adaptations of serial dictatorship for
committee voting.
2 Related Work
A first related stream of work involves studying specific committee elections rules
from a computational point of view (generally with little or no focus on Pareto opti-
4
Haris Aziz et al.
mality). Our focus on determining whether a committee is Pareto optimal or on finding a Pareto optimal committee, is in some sense orthogonal to the study of committee
election rules. The simplest (and most widely used) rules for electing a committee,
called best-k rules, compute a score for each alternative based on the ranks, and the
alternatives with the best k scores are elected Elkind et al. (2014); Faliszewski et al.
(2016). Scoring-based extension principles have also been used by Darmann (2013).
Note that the output of a best-k rule is obviously Pareto-optimal for the preferences
induced by this scoring function, but not necessarily with respect to other set extensions.
Klamler et al. Klamler et al. (2012) compute optimal committees under a weight
constraint for a single agent (therefore optimality is equivalent to Pareto optimality),
using several preference extensions including ‘worst’, ‘best’, and downward lexicographic.
The ‘best’ (B) extension principle has been used in a number of papers on committee elections by full proportional representation, starting with
(Chamberlin and Courant, 1983) and studied from a computational point of view
in a long series of papers (e.g., (Procaccia et al., 2008; Lu and Boutilier, 2011;
Betzler et al., 2013; Skowron et al., 2015a; Elkind and Ismaili, 2015). These rules
obviously output Pareto optimal committees for B, but not necessarily for other extensions.
Some of the set extensions considered in this paper have corresponding analogues
when extending preferences over alternatives to preferences over ‘lotteries over alternatives.’ In particular, the RS set extension corresponds to SD (stochastic dominance)
lottery extension. Also the DL and UL set extensions considered in this paper correspond to DL and UL lottery extensions considered in works in probabilistic social
choice (Brandl, 2013; Aziz et al., 2013b; Cho, 2016).
Some works are based on the Hamming extension. Each agent specifies his ideal
committee and he prefers committees with less Hamming distance from the ideal
committee. The Hamming distance notion can be used to define specific rules such as
minimax approval voting (Brams et al., 2007), which selects the committee minimizing the maximum Hamming distance for the agents. Although the output of minimax
approval voting is not always Pareto-optimal for the Hamming extension, there are
good Pareto-optimal approximations of it Caragiannis et al. (2010). Note that for dichotomous preferences, the Hamming extension coincides with the responsive and
the downward lexicographic extensions, therefore our computational results for responsive set extension for dichotomous preferences also hold for the Hamming and
downward lexicographic extensions.
A second line of work concerns understanding the classes of rules that result
in Pareto optimal outcomes. Most works along this line bear on a different type of
committee elections, called designated-seat voting, where candidates must declare
the seat they contest (Benoı̂t and Kornhauser, 2010).1 Results about the existence or
non-existence of Pareto optimal rules have been presented Özkal-Sanver and Sanver
(2006); Benoı̂t and Kornhauser (2010); Cuhadaroǧlu and Lainé (2012).
1 If there are exactly two candidates per seat, then designated voting is equivalent to multiple referenda,
where a decision has to be taken on each of a series of yes-no issues.
Computing and Testing Pareto Optimal Committees
5
3 Setup
We consider a set of agents N = {1, . . . , n}, a set of alternatives A = {a1 , . . . , am } and
a preference profile %= (%1 , . . . , %n ) such that each %i is a complete and transitive
relation over A. We write a %i b to denote that agent i values a at least as much as b
and use ≻i for the strict part of %i , i.e., a ≻i b iff a %i b but not b %i a. Finally, ∼i
denotes i’s indifference relation, i.e., a ∼i b iff both a %i b and b %i a.
The relation %i results in equivalence classes Ei1 , Ei2 , . . . , Eiki for some ki such that
′
a ≻i a′ if a ∈ Eil and a′ ∈ Eil for some l < l′ . We will use these equivalence classes
to represent the preference relation of an agent as a preference list i : Ei1 , Ei2 , . . . , Eiki .
For example, we will denote the preferences a ∼i b ≻i c by the list i : {a, b}, {c}. An
agent i’s preferences are strict if the size of each equivalence class is 1. An agent i’s
preferences are dichotomous if he partitions the alternatives into just two equivalence
classes, i.e., ki = 2. Let T opwidth(%) be the maximum size of the most preferred
equivalence class, i.e., T opwidth(%) = maxi≤n |Ei1 |. For any S ⊆ A, we will denote
by max%i (S ) and min%i (S ) the alternatives in S that are maximally and minimally
preferred by i respectively. Thus, if q and r are respectively the smallest and the
q
q
largest indices such that Ei ∩ S , ∅ and Eir ∩ S , ∅, then max%i (S ) = Ei ∩ S and
min%i (S ) = Eir ∩ S . For k ≤ m, let S k (A) = {W ⊆ A : |W| = k}.
4 Set Extensions and Pareto Optimality
Set Extensions Set extensions are used for reasoning about the preferences of an
agent over sets of alternatives given their preferences over single alternatives. For
fixed-size committee voting, the responsive extension is very natural and has been applied in various matching settings as well (Barberà et al., 2004; Roth and Sotomayor,
1990). For all V, W ∈ S k (A), we say that W %RS
i V if and only if there is an injection f
from V to W such that for each a ∈ V, agent i weakly prefers f (a) to a, i.e. f (a) %i a.
We define the ‘best’ set extension and the ‘worst’ set extension which are denoted
B and W respectively. For all W, V ∈ S k (A), W %B
i V if and only if w %i v for
w ∈ max%i (W) and v ∈ max%i (V). On the other side, W %W
i V if and only if w %i v
for w ∈ min %i (W) and v ∈ min%i (V).
In the downward lexicographic (DL) extension, an agent prefers a committee that
selects more alternatives from his most preferred equivalence class, in case of equality, the one with more alternatives for the second most preferred equivalence class,
and so on. Formally, W ≻iDL V iff for the smallest (if any) l with |W ∩ Eil | , |V ∩ Eil |
we have |W ∩ Eil | > |V ∩ Eil |.
In the upward lexicographic (UL) extension, an agent prefers a committee that
selects less alternatives from his least preferred equivalence class, in case of equality,
the one with less alternatives for the second least preferred equivalence class, and so
on. Formally, W ≻UL
V iff for the largest (if any) l with |W ∩ Eil | , |V ∩ Eil | we have
i
l
l
|W ∩ Ei | < |V ∩ Ei |.
Remark 1 Consider an agent i with preferences %i over A. Let S , T ⊂ W such that
|S | = |T | = k. Then,
6
Haris Aziz et al.
–
–
–
–
S
S
S
S
%RS
i
%RS
i
≻RS
i
≻RS
i
T
T
T
T
=⇒
=⇒
=⇒
=⇒
S
S
S
S
%iDL
%UL
i
≻iDL
≻UL
i
T =⇒ S %B
i T
T =⇒ S %W
i T
T
T
The relations follow from the definitions.
Efficiency based on Set Extensions With each set extension E, we can define Pareto
optimality with respect to E. A committee W ∈ S k (A) is Pareto optimal with respect
to E, or simply E-efficient, if there exists no committee W ′ ∈ S k (A) such that W ′ %E
i
W for all i ∈ N and W ′ ≻E
i W for some i ∈ N. Note that for each of our set extensions,
E-efficiency coincides with standard Pareto optimality when k = 1. An outcome is
a Pareto improvement over another if each agent weakly improves and at least one
agent strictly improves.
Example 1 Consider the preference profile:
1 : a, b, c, d
2 : d, c, b, a
Suppose k = 2. Then,
–
–
–
–
–
The unique B-efficient committee is {a, d}
The unique W-efficient committee is {b, c}.
The DL-efficient committees are {a, d}, {a, b}, and {d, c}.
The UL-efficient committees are {b, c}, {a, b}, and {d, c}.
The RS-efficient committees are {a, d}, {b, c}, {a, b}, and {d, c},
Remark 2 Consider a committee S .
– If S is DL-efficient, then S is RS -efficient
– If S is UL-efficient, then S is RS -efficient
The argument is as follows. Suppose S is not RS -efficient, then there exists some
RS
other outcome T such that T %RS
i S for all i ∈ N and T ≻i S for some i ∈ N. In that
DL
DL
case T %i S for all i ∈ N and T ≻i S for some i ∈ N. Also T %UL
S for all i ∈ N
i
S
for
some
i
∈
N.
Hence
S
is
neither
DL-efficient
not
UL-efficient.
and T ≻UL
i
Remark 3 There always exists a B-efficient committee that is also DL-efficient: DL
Pareto improvements over a B-efficient does not harm any agent with respect to the
B relation.
Remark 4 There always exists a W-efficient committee that is also UL-efficient: UL
Pareto improvements over a W-efficient does not harm any agent with respect to the
W relation.
In Figure 1, we illustrate the relations between the different efficiency notions.
Later on in the paper we will present an algorithm that returns a committee that is
UL-efficient and DL-efficient, and hence RS -efficient.
We also make the following general observation.
Computing and Testing Pareto Optimal Committees
7
B-efficiency
W-efficiency
DL-efficiency
U L-efficiency
RS -efficiency
Fig. 1 Relations between the five notions of efficiency. An arrow from E1 -efficiency to E2 -efficiency means
that E1 -efficiency implies E1 -efficiency; a dashed line means there always exists a committee that is both
E1 - and E2 -efficient; absence of arrow or line means that the sets of E1 - and E2 -efficient committees can
be disjoint.
Lemma 1 If there is a polynomial-time algorithm to compute a Pareto improvement
over a committee, then there exists a polynomial-time algorithm to compute an Eefficient committee under set extensions E ∈ {RS , DL, UL, W, B}.
Proof. Here, we start from any committee and we recursively apply Pareto improvement until we reach a Pareto optimal committee. For the ‘best’ and ‘worst’ extensions, there can be at most mn Pareto improvements because for one agent there can
be at most m improvements. Since an RS -improvement implies an DL-improvement,
let us bound the number of Pareto-improvements with respect to DL. In each Paretoimprovement, for the agent who strictly improves, the most preferred equivalence
class that has different number of alternative in the outcome increases by at least one.
Therefore the most preferred equivalence class can be the improving class in at most
m of the Pareto improvements. Similarly, the number of Pareto improvements in the
subsequent less preferred equivalence class improves in a Pareto improvement can
be at most m of the Pareto improvements. Therefore the total number of DL Paretoimprovements is bounded by m2 n. A similar argument holds for UL as well.
⊔
⊓
We end this section by observing that, under any of the set extensions we consider,
a set of Pareto optimal alternatives may be Pareto dominated. Consider the following
example.
Example 2
1 : a, c, b, d
3 : b, c, a, d
2 : a, d, b, c
4 : b, d, a, c
The set {c, d} consists of Pareto optimal alternatives but is Pareto dominated by {a, b}
under any of our set extensions.
8
Haris Aziz et al.
5 Responsive Set Extension
There is a trivial way to achieve Pareto optimality under the responsive set extension
by taking any decreasing scoring vector consistent with the ordinal preferences, finding the total score of each alternative and returning the set of k alternatives with the
maximum scores. For instance, on Example 2, the outcome of the rule that outputs
the alternatives with the best k Borda scores is {a, b}.
Theorem 1 A Pareto optimal committee under the responsive set extension committee can be computed in linear time.
In many situations, one may already have a status-quo committee and one may
want to find a Pareto improvement over it. This problem of testing Pareto optimality
and finding a Pareto improvement under the responsive set extension turns out to be a
much harder task. Note that if there exists a polynomial-time algorithm to compute a
Pareto improvement, then it means that testing Pareto optimality is also polynomialtime solvable.
Theorem 2 Checking whether a committee is Pareto optimal under the responsive
set extension is coNP-complete, even for dichotomous preferences and T opwidth(%
) ≥ 3, or for strict preferences.
Proof. We only present the case where T opwidth(%) = 3. The reduction is from the
NP-complete problem vertex cover Garey and Johnson (1979). Given a simple graph
G = (V, E), the minimum vertex cover problem consists in finding a subset C ⊆ V of
minimum size such that every edge e ∈ E is incident to some node of C. Its decision
version vertex cover takes as input a simple graph G = (V, E) and an integer k and
problem is deciding if there exists a vertex cover C ⊆ V of G with |C| ≤ k.
Let h(V, E), ki be an instance of vertex cover, with [x, y] being one arbitrary edge
in E. We build the following instance of Pareto optimality under RS :
• N = ∪e∈E Ne ∪ {a}, where for each edge e ∈ E, Ne is a set of k agents, and a is a
special agent.
• A = V ∪ D, where D = {d1 , . . . , dk }.
• For each e = [u, v] ∈ E, the preferences of agent ei , for i = 1, . . . , k, and of agent
a, are
ei : {u, v, di }, (D − di ) ∪ (V \ {u, v})
a : {x, y}, D ∪ (V \ {x, y})
The reduction is clearly done within polynomial time and preferences are dichotomous. We can check easily that committee D (of size k) is not Pareto optimal under
RS if and only if there exists a vertex cover of G of size at most k.
For strict preferences, in the previous reduction we replace {u, v, di }, . . . by
{u}, {v}, {di }, . . . in the preferences of ei . It is easy to see that the proof is similar. ⊓
⊔
Using a similar reduction from the Hitting Set problem, we can also
prove Theorem 3 that concerns a parametrized complexity intractability result Downey and Fellows (2013). Hitting Set is defined as follows: given a ground
set X of elements, and a collection C = {C1 , . . . , Cℓ } of subsets of X, does there exist
a H ⊂ X such that |H| ≤ k and H ∩ C , ∅ for all C ∈ C?
Computing and Testing Pareto Optimal Committees
9
Theorem 3 Checking whether a committee is Pareto optimal under the responsive
set extension is W[2]-complete under parameter k, even for dichotomous preferences.
For dichotomous preferences we present a complete characterization of the complexity according to the T opwidth(%) parameter. If T opwidth(%) = 1, then in any
Pareto improvement over committee D, any alternative in D that is most preferred
by some agent needs to be kept selected, and therefore the problem of checking RS efficiency is easy. If T opwidth(%) ≥ 3, from Theorem 2, the problem is hard. Remains
the case T opwidth(%) = 2.
Theorem 4 For dichotomous preferences, a Pareto improvement over a committee
with respect to the responsive set extension can be computed in polynomial time when
T opwidth(%) ≤ 2.
Proof. Consider a preference profile %= (%1 , . . . , %n ) where each %i is dichotomous
and verifies T opwidth(%) = 2, and let D ∈ S k (A). For each i ∈ N, let (Ei1 , Ei2 ) be the
partition associated with %i .
First, if for all i ∈ N, Ei1 ⊆ D, then D is obviously RS -efficient. Assume it is not
the case, that is, (1) for some i ∈ N, Ei1 \ D , ∅. Let
– N ′ = {i ∈ N : Ei1 ∩ D = Ei1 }, W ′ = ∪i∈N ′ Ei1 (by construction, W ′ ⊆ D), and
k′ = |W ′ |.
– N ′′ = {i ∈ N \ N ′ : Ei1 ∩ (D \ W ′ ) , ∅} and A′′ = ∪i∈N ′′ Ei1 .
Now, we build a graph G = (V, E) with V = {v1 , . . . , vr } isomorphic to A′′ , and
[v p , vq ] ∈ E iff Ei1 = {a p , aq } for some i ∈ N ′′ : each edge of G corresponds to the top
two alternatives of some agent, provided one of them is in D \ W ′ . Let τ(G) be the
size of an optimal vertex cover of G.
We first claim that there is a Pareto improvement over D if and only if one of
follows two conditions is satisfied:
(i) τ(G) < k − k′ , or
(ii) τ(G) = k − k′ , and there is an optimal vertex cover of G containing either at least
an element of Ei1 for some i < N ′ ∪ N ′′ , or two elements of Ei1 for some i ∈ N ′′ .
We first show that (i) and (ii) are sufficient. If (i) holds then take a committee
corresponding to a minimum vertex cover of G, add to it the k′ alternatives of W ′ , and
add (k − k′ ) − τ(G) alternatives, with at least one in ∪i (Ei1 \ D); this is possible because
of (1). If (ii) holds, then take a committee corresponding to a minimum vertex cover
of G, and add to it the k′ alternatives of W ′ . In both cases, the obtained committee
contains Ei1 for all i ∈ N ′ , contains at least one element of Ei1 for all i ∈ N ′ , and
contains either two elements of Ei1 for some i ∈ N ′′ , or an element of Ei1 for some
i < N ∪ N ′′ . Therefore it is a Pareto-improvement over D.
Now, we show that (i) and (ii) are necessary. Let W ∈ S k (A) be a Pareto improvement of D containing a maximum number of alternatives from D. We have the
following two properties: W ′ ⊆ W and W \ W ′ is a vertex cover of G. W ′ ⊆ W holds,
since otherwise there would be an i ∈ N ′ such that W ′ %RS
W does not hold. For
i
similar reasons, C ′ = (W \ W ′ ) ∩ A′′ is a vertex cover of G. If |(W \ W ′ ) ∩ A′′ | < τ(G),
then by adding to it any set of D \ C ′ of size k − k′ − τ(G) we obtain a set of size k
10
Haris Aziz et al.
which constitutes a Pareto improvement of D because now, Ei1 ⊆ W for some i ∈ N ′′ .
If |(W \ W ′ ) ∩ A′′ | = τ(G), then (W \ W ′ ) ∩ A′′ = W \ W ′ and necessarily either
Ei1 ∩ C , ∅ for some i < (N ′ ∪ N ′′ ) or Ei1 ⊆ C for some i ∈ N ′′ .
It remains to be shown that (i) and (ii) can be checked in polynomial time. (i)
can be done in polynomial-time because G is bipartite: indeed, by construction, G
is two-colorable with color sets A′′ ∩ D and A′′ \ D, and by König’s theorem, for
bipartite graphs, the problem of finding the minimum vertex cover is equivalent to
computing a maximum matching, hence solvable in polynomial time. As for (ii), if
τ(G) = k − k′ , we have to check whether for some optimal vertex cover C of G, either
(ii.1) Ei1 ∩ C , ∅ holds for some i < (N ′ ∪ N ′′ ), or (ii.2) Ei1 ⊆ C for some i ∈ N ′′ .
In order to check (ii.1), for each i < (N ′ ∪ N ′′ ) such that there exists x ∈ Ei1 ∩ A′′ ,
we transform G into a new bipartite graph G{x} where we add a new vertex x′ and an
edge [x, x′ ]. In order to check (ii.2), for each i < N ′′ , let Ei1 = {x, y}; we transform
G into a new bipartite graph G{x,y} where we add two new vertices x′ and y′ , and two
edges [x, x′ ] and [y, y′ ]. Finally, we test if τ(G) = τ(G{x} ) or if τ(G) = τ(G{x,y} ) for one
of these graphs, because all optimal vertex covers of G{x} (respectively G{x,y} ) must
contain x (respectively {x, y}).
⊔
⊓
Example 3 We illustrate the algorithm in the proof of Theorem 4. Let k = 2 and
consider the dichotomous profile, where we specify only the top equivalence class of
each agent:
1 : {a, c}
4 : {d, e}
2 : {b, c}
5 : {e, f }
3 : {b, d}
Let D = {a, b}. We have N ′ = W ′ = ∅, k′ = 0, D \ W ′ = {a, b}, N ′′ = {1, 2, 3},
and A′′ = {a, b, c, d}. We construct the graph G = (V, E): V = {va , vb , vc , vd } and
E = {{va , vc }, {vb , vc }, {vb , vd }}. We have τ(G) = 2 = k − k′ . Now we consider the four
graphs G{d} , resulting from the addition to G of a new vertex vd′ and edge [vd , vd′ ], and
G{a,c} , G{b,c} and G{b,d} : G{a,c} results from the addition to G of two new vertices va′ , vc′
and edges [va , va′ ] and [vc , vc′ ], etc. Two of these graphs have an optimal cover of size
2: G{d} , with optimal cover {vc , vd }, and G{b,c} , with optimal cover {vb , vc }. Therefore,
{c, d} and {b, c} are RS -Pareto-improvements over {a, b}, and {a, b} is not RS -efficient.
va
vc
G
vb
va
vd
vc
vb
vd
G {d}
vb
va
vd′
vc′
vb′
vd
vc
G {b,c}
Fig. 2 Graphs corresponding to Example 3
Note that finding an algorithm that computes a Pareto improvement over a committee can be used to decide whether a given a committee D of size k, is Pareto
optimal under the responsive set extension.
Computing and Testing Pareto Optimal Committees
11
Pareto optimality and Strategyproofness We now try to achieve both RS -efficiency
and strategyproofness simultaneously. A mechanism f is strategyproof if reporting
truthful preferences is a dominant strategy with respect to the responsive set extension: f (%) %RS
f (%′i , %−i ) for all preference profiles % and (%′i , %−i ). Note that defining
i
strategyproofness in this way with respect to the RS extension is stronger than defining it for any of the other four extensions considered in this paper. Nonetheless, we
will present some positive results with respect to strategyproofness.
A naive way of achieving RS -efficiency and Pareto optimality is to enumerate
the list of possible winning sets and implement serial dictatorship over the possible
outcomes as is done in voting (Aziz et al., 2013b). However, the number of possible
outcomes is exponential and responsive preferences result in a partial order over the
possible winning sets and not a complete and transitive order. This problem is solved
by Algorithm 1 which can be viewed as a computationally efficient serial dictatorship.
Algorithm 1 Committee Voting Serial Dictatorship
Input: (N, A, %, k, permutation π of N)
Output: W ∈ S k (A).
1
2
3
4
5
6
7
8
9
10
11
12
13
14
L (last set to be refined) ←− A
r (number of alternatives yet to be fixed) ←− k; W ←− ∅
i′ (index of the permutation π) ←− 1
while r , 0 or i′ , n do
S
S
j
j
Agent i = π(i′ ) selects first t equivalence classes such that | tj=1 Ei ∩ L| ≥ r and | t−1
j=1 Ei ∩ L| < r.
S
St−1 j
j
W ←− W ∪ ( t−1
j=1 Ei ∩ L) (we say agent i fixes the alternatives in
j=1 Ei );
S
S
j
j
r ←− | tj=1 Ei ∩ L| − | t−1
E
∩
L|
j=1 i
L ←− Eit ; ri′ ←− r
′
Increment i by one
end while
if r > 0 then
pick any r alternatives from L and add them to W
end if
return W
Theorem 5 There exists a linear-time and strategyproof algorithm that returns a
committee that is Pareto optimal under the responsive set extension.
Proof. Consider Algorithm 1. We show that at each stage i′ , agent π(i′ ), implicitly
refines the set of feasible committees to the maximal set of most preferred outcomes
from the set by providing additional constraints. This is true for the base case i′ = 1.
Now assume it holds from 1 to i′ . Note that L contains all those alternatives that are
strictly less preferred by agents in {π(1), . . . , π(i′ )} than the ones they respectively
fixed. Moreover, each agent in {1, . . . , π(i′ )} is indifferent between the alternatives in
S
j
L. As for π(i′ + 1), he fixes the best | t−1
j=1 E π(i′ +1) ∩ L| alternatives in L where t is the
St−1 j
S
j
t
value such that |( tj=1 Eπ(i
′ +1) ) ∩ L| ≥ ri′ and |
j=1 E π(i′ +1) ∩ L| < ri′ . For E π(i′ +1) , the
St−1 j
St−1 j
agent only requires that ri′ +1 = |( j=1 Eπ(i′ +1) )∩L|−|( j=1 Eπ(i′ +1) )∩L| alternatives are
t
selected from his equivalence class Eπ(i
′ +1) which is ensured by the definition of the
12
Haris Aziz et al.
algorithm. It follows from the argument that the returned set is Pareto optimal under
the responsive set extension. For strategyproofness, when an agent π(i′ ) turn comes,
it only has a choice over fixing the alternatives in L and requiring ri′ alternatives from
t
his equivalence class Eπ(i
′ ) . In this case the algorithm already chooses one of the best
possible committees for the agent.
⊔
⊓
Note that for k = 1, the algorithm is equivalent to serial dictatorship as formalized
by Aziz et al. (2013a). Note that a committee that is Pareto optimal under the responsive set extension may not be a result of serial dictatorship. This holds even for k = 1
and the basic voting setting.
The problem with the serial dictatorship algorithm formalized is that it overly
favours the agent that is the first in the permutation. One way to limit his power
is to let him choose only ⌈k/n⌉ alternatives. We note that this attempt at having a
fairer extension of serial dictatorship comes at an expense because strategyproofness
is compromised. Consider the profile in which 1 has preferences a, b, c and 2 has
preferences a, c, b. For k = 2, and permutation 12, the outcome is {a, c}. But if agent
1 reports b, a, c, then the outcome is {a, b}.
6 ‘Best’ Set Extension
Next, we consider Pareto optimality with respect to B, which has been used for defining many rules (see Section 2).
Theorem 6 Unless P=NP, there is no polynomial-time algorithm to compute a
Pareto improvement over a committee with respect to B, even for dichotomous preferences and T opwidth(%) = 2.
Proof. We show that if it is not the case, then we can solve polynomially the vertex cover decision problem. Consider an instance of vertex cover given by a simple
graph G = (V, E) with V = {v1 , . . . , vq } and E = {e1 , . . . , er }, and an integer k. Assume
the existence of a polynomial-time algorithm Algo that computes a Pareto improvement over a committee with respect to B when T opwidth(%) ≤ 2: given a profile %
and a set of k alternatives W, Algo(%, W) returns, in time polynomial in | % |, Yes if
W is Pareto optimal with respect to B, and otherwise returns a k-set U of alternatives
which Pareto dominates W. We will now prove by applying at most n times Algo
with different inputs that we can decide in polynomial if G has a vertex cover C ⊆ V
of G with |C| ≤ k. We construct the following profile P:
• The set of agents is N = {1, . . . , q + r − k}, where agent i ≤ r corresponds to edge
ei ∈ E.
• The set of 2q − k alternatives is A = V ∪ D where D = {d1 , . . . , dq−k }.
• Let ei = [u, v] ∈ E be an edge of G; the preferences of agent i for i = 1, . . . , r are:
i : {u, v}, D ∪ (V \ {u, v}).
The preferences of the last set of q − k agents {r + 1, . . . , q + r − k} are given by:
for i = 1, . . . , q − k,
r + i : di , V ∪ (D \ {di }).
Computing and Testing Pareto Optimal Committees
13
The reduction is clearly done within polynomial time and the set of preferences
given by % are dichotomous.
Consider the following inductive procedure: W0 = V and for i ≥ 1, Wi = Algo
(%, Wi−1 ) if Wi−1 is not Pareto optimal with respect to B, otherwise we return Wi−1 .
Let W = Wq−k be the solution output after q − k calls. Because Algo is polynomial,
the whole procedure is polynomial.
We claim that G has a vertex cover of size k iff D ⊆ W. We will first prove by
induction that at each step i, Wi \ D is a vertex cover of G. For the initial step, it
is valid because V is a vertex cover of G. Assume that it is true for i < q − k and
let us prove that Wi+1 \ D is a vertex cover of G. If it is not the case, some edge
e j = [u, v] ∈ E is not covered. By assumption, e j is covered by Wi \ D. This implies
B
Wi ≻B
j Wi+1 , which is a contradiction. Hence, Wi+1 % j Wi . From this hypothesis, we
deduce D \ Wi+1 %B
j D \ Wi for j = n + 1, . . . , 2n − k with a strict preference for
some agent. Equivalently, D \ Wi ⊂ D \ Wi+1 . In conclusion, after q − k recursive calls,
|W \ D| ≤ k if and only if D ⊆ W.
⊔
⊓
Theorem 7 Computing a B-efficient committee is NP-hard, even for dichotomous
preferences.
Proof. We give a reduction from Hitting Set. Let N = {1, . . . , ℓ}, A = X and for each
i ∈ N, i’s dichotomous preferences are i : Ci , (X \ Ci ). If there exists a polynomialtime algorithm to compute a B-efficient committee, it will return a committee in
which each agent gets a most preferred alternative if such a committee exists. But
such a committee corresponds to a hitting set of size k.
⊔
⊓
7 Downward Lexicographic Set Extension
We point out that for dichotomous preferences, the responsive set extension coincides with the downward lexicographic set extension. Hence we get a corollary of
our results for responsive preferences:
Corollary 1 Checking whether a committee is DL-efficient is coNP-complete, even
for dichotomous preferences and T opwidth(%) ≥ 3.
Note that Algorithm 1 returns a DL-efficient committee. The reason is that each
agent in her turn refines the set of possible outcomes to her most preferred subset of
outcomes. Each committee is the refined set is at least as preferred with respect to RS
(and hence with respect to DL) to all committees in the set of possible outcomes.
Theorem 8 There exists a linear-time and strategyproof algorithm that returns a DLefficient committee.
8 ‘Worst’ Set Extension
In contrast to all the other set extensions considered in the paper, Pareto optimality
with respect to the ‘worst’ set extension can be checked in polynomial time.
14
Haris Aziz et al.
Theorem 9 There exists a polynomial-time algorithm that checks whether a committee is W-efficient and computes a Pareto improvement over it if possible.
Proof. Let W ∈ S k (A). For each i ∈ N, let Eiti be the least preferred equivalence class
such that Eiti ∩ W , ∅. We want to check whether there is a k-set D of alternatives
in which at least some agent i ∈ N gets a strictly better outcome and all the other
agents get at least as preferred an outcome. We check this as follows. For i ∈ N,
Si
S
Sk j
let Bi = A \ (( kℓ=t
Eiℓ ) ∪ j∈N\{i} ℓ=t
E ℓj )). We check whether |Bi | ≥ k or not.
i
j +1
If |Bi| ≥ k, we know that there exists a subset of Bi , that is strictly more preferred
by i ∈ N and at least as preferred by each agent. The reason is that Bi contains a
more preferred worst alternative for agent i than D and contains at least as preferred
worst alternative for other agents j than D. If |Bi| < k, then this means that a Pareto
improvement with i strictly improving is only possible if the size of the winning set
is less than k which is not feasible.
⊔
⊓
We now consider strategyproofness together with W-efficiency. We first note that
Algorithm 1 may not return a W-efficient outcome. However, we construct a suitable strategyproof and W-efficient by formalising an appropriate serial dictatorship
algorithm for the worst set extension.
Theorem 10 There exists a linear-time and strategyproof algorithm that returns a
W-efficient committee.
Proof. Consider the agents in a permutation π. The set of alternatives A′ is initialized
to A. We reduce the set A′ while ensuring that it of size at least k. The next agent i
in the permutation comes and deletes the maximum number of least preferred equivalence classes from his preferences and the corresponding alternatives in A′ while
ensuring that |A′ | ≥ k. Each successive agent in the permutation gets a most preferred
outcome while ensuring that agents before him in the permutation get at least as preferred an outcome as before. Thus the algorithm is strategyproof and Pareto optimal
with respect to the ‘worst’ set extension.
⊔
⊓
9 Upward Lexicographic Set Extension
We point out that for dichotomous preferences, the responsive set extension coincides
with the upward lexicographic set extension. Hence we get a corollary of our results
for responsive preferences:
Corollary 2 Checking whether a committee is UL-efficient is coNP-complete, even
for dichotomous preferences and T opwidth(%) ≥ 3.
Note that Algorithm 1 returns a UL-efficient committee. The reason is that each
agent in her turn refines the set of possible outcomes to her most preferred subset of
outcomes. Each committee is the refined set is at least as preferred with respect to RS
(and hence with respect to UL) to all committees in the set of possible outcomes.
Theorem 11 There exists a linear-time and strategyproof algorithm that returns a
UL-efficient committee.
Computing and Testing Pareto Optimal Committees
15
10 Conclusions
We considered Pareto optimality in multi-winner voting with respect to a number of
prominent set extensions. We presented results on the relations between the notions
as well as complexity of computing and verifying Pareto optimal outcomes. Further
directions of future work include considering Pareto optimality with respect to other
set extensions (Brandt and Brill, 2011). Another direction is consider the compatibility of Pareto optimality concepts with other axioms. Finally, we remark that our serial
dictatorship algorithm can be used to define a multiwinner generalization of random
serial dictatorship, which is worth investigating and raises interesting computational
problems.
Acknowledgments
This is the extended version of the IJCAI conference paper (Aziz et al., 2016). The
authors thank Felix Brandt for useful pointers and comments. They also thanks the
reviewers and attendees of IJCAI 2016 and COMSOC 2016 for useful comments.
Jérôme Lang and Jérôme Monnot thank the ANR project CoCoRICo-CoDec.
References
H. Aziz and R. Savani. Hedonic games. In F. Brandt, V. Conitzer, U. Endriss, J. Lang,
and A. D. Procaccia, editors, Handbook of Computational Social Choice, chapter 15. Cambridge University Press, 2016.
H. Aziz, F. Brandt, and M. Brill. The computational complexity of random serial
dictatorship. Economics Letters, 121(3):341–345, 2013a.
H. Aziz, F. Brandt, and M. Brill. On the tradeoff between economic efficiency and
strategyproofness in randomized social choice. In Proceedings of the 12th International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS),
pages 455–462. IFAAMAS, 2013b.
H. Aziz, J. Lang, and J. Monnot. Computing Pareto Optimal Committees. In Proceedings of the 25th International Joint Conference on Artificial Intelligence (IJCAI),
pages 60–66, 2016.
H. Aziz, F. Brandt, E. Elkind, and P. Skowron. Computational social choice: The first
ten years and beyond. In B. Steffen and G. Woeginger, editors, Computer Science
Today, volume 10000 of Lecture Notes in Computer Science (LNCS). SpringerVerlag, 2017. Forthcoming.
S. Barberà, W. Bossert, and P. K. Pattanaik. Ranking sets of objects. In S. Barberà, P. J. Hammond, and C. Seidl, editors, Handbook of Utility Theory, volume II,
chapter 17, pages 893–977. Kluwer Academic Publishers, 2004.
J.-P. Benoı̂t and L. Kornhauser. Only a dictatorship is efficient. Games and Economic
Behavior, 70(2):261–270, 2010.
N. Betzler, A. Slinko, and J. Uhlmann. On the computation of fully proportional
representation. JAIR, 47:475–519, 2013.
16
Haris Aziz et al.
W. Bossert. Preference extension rules for ranking sets of alternatives with a fixed
cardinality. Theory and Decision, 39:301—317, 1995.
S. Brams, D. Kilgour, and R. Sanver. A minimax procedure for electing committees.
Public Choice, 3-4(132):401–420, 2007.
F. Brandl. Efficiency and incentives in randomized social choice. Master’s thesis,
Technische Universität München, 2013.
F. Brandt and M. Brill. Necessary and sufficient conditions for the strategyproofness
of irresolute social choice functions. In Proceedings of the 13th Conference on
Theoretical Aspects of Rationality and Knowledge (TARK), pages 136–142. ACM
Press, 2011.
I. Caragiannis, D. Kalaitzis, and E. Markakis. Approximation algorithms and mechanism design for minimax approval voting. In Proceedings of the 24th AAAI Conference on Artificial Intelligence (AAAI), pages 737–742, 2010.
K. Cechlárová. Stable partition problem. In Encyclopedia of Algorithms, pages 885–
888. Springer, 2008.
J. R. Chamberlin and P. N. Courant. Representative deliberations and representative decisions: proportional representation and the Borda rule. American Political
Science Review, 77(3):718–733, 1983.
W. J. Cho. Incentive properties for ordinal mechanisms. Games and Economic Behavior, 95:168–177, 2016.
T. Cuhadaroǧlu and J. Lainé. Pareto efficiency in multiple referendum. Theory and
Decision, 72(4):525–536, 2012.
A. Darmann. How hard is it to tell which is a condorcet committee? Mathematical Social Sciences, 66(3):282–292, 2013. doi: 10.1016/j.mathsocsci.2013.06.004. URL
http://dx.doi.org/10.1016/j.mathsocsci.2013.06.004.
R. G. Downey and M. R. Fellows. Fundamentals of Parameterized Complexity. Texts
in Computer Science. Springer, 2013.
E. Elkind and A. Ismaili. OWA-Based Extensions of the Chamberlin-Courant Rule.
In Proceedings of the 4th International Conference on Algorithmic Decision Theory (ADT), pages 486–502. Springer-Verlag, 2015.
E. Elkind, P. Faliszewski, P. Skowron, and A. Slinko. Properties of multiwinner
voting rules. In Proceedings of the 13th International Conference on Autonomous
Agents and Multi-Agent Systems (AAMAS), pages 53–60, 2014.
E. Elkind, J. Lang, and A. Saffidine. Condorcet winning sets. Social Choice and
Welfare, 44(3):493–517, 2015.
P. Faliszewski, P. Skowron, A. Slinko, and N. Talmon. Multiwinner analogues of the
plurality rule: Axiomatic and algorithmic views. In Proceedings of the 28th AAAI
Conference on Artificial Intelligence (AAAI), 2016.
P. Faliszewski, P. Skowron, A. Slinko, and N. Talmon. Multiwinner voting: A new
challenge for social choice theory. In U. Endriss, editor, Trends in Computational
Social Choice, chapter 2. 2017. Forthcoming.
M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to the Theory
of NP-Completeness. W. H. Freeman, 1979.
C. Klamler, U. Pferschy, and S. Ruzika. Committee selection under weight constraints. Mathematical Social Sciences, 64(1):48–56, 2012.
Computing and Testing Pareto Optimal Committees
17
J. Lang, J. Mengin, and L. Xia. Aggregating conditionally lexicographic preferences
on multi-issue domains. In Principles and Practice of Constraint Programming,
pages 973–987, 2012.
T. Lu and C. Boutilier. Budgeted social choice: From consensus to personalized
decision making. In Proceedings of the 22nd International Joint Conference on
Artificial Intelligence (IJCAI), pages 280–286. AAAI Press, 2011.
B. L. Monroe. Fully proportional representation. The American Political Science
Review, 89(4):925–940, 1995.
H. Moulin. Fair Division and Collective Welfare. The MIT Press, 2003.
İ. Özkal-Sanver and R. Sanver. Ensuring Pareto-optimality by referendum voting.
Social Choice and Welfare, 27:211–219, 2006.
A. D. Procaccia, J. S. Rosenschein, and A. Zohar. On the complexity of achieving
proportional representation. Social Choice and Welfare, 30:353–362, 2008.
A. E. Roth and M. A. O. Sotomayor. Two-Sided Matching: A Study in Game Theoretic
Modelling and Analysis. Cambridge University Press, 1990.
P. Skowron, P. Faliszewski, and A. Slinko. Achieving fully proportional representation: Approximability results. Artif. Intell., 222:67–103, 2015a.
P. K. Skowron, P. Faliszewski, and J. Lang. Finding a collective set of items: From
proportional multirepresentation to group recommendation. In Proceedings of the
23rd International Joint Conference on Artificial Intelligence (IJCAI), pages 2131–
2137. AAAI Press, 2015b.
| 8cs.DS
|
Evolutionary Design: Philosophy, Theory, and Application Tactics
1
V.V. Kryssanov1, H. Tamaki2, S. Kitamura2
College of Information Science and Engineering, Ritsumeikan University, Kusatsu, Japan
2
Faculty of Engineering, Kobe University, Kobe, Japan
Abstract
Although it has contributed to remarkable improvements in some specific areas, attempts to develop a
universal design theory are generally characterized by failure. This paper sketches arguments for a new
approach to engineering design based on Semiotics – the science about signs. The approach is to combine
different design theories over all the product life cycle stages into one coherent and traceable framework.
Besides, it is to bring together the designer’s and user’s understandings of the notion of ‘good product’.
Building on the insight from natural sciences that complex systems always exhibit a self-organizing meaninginfluential hierarchical dynamics, objective laws controlling product development are found through an
examination of design as a semiosis process. These laws are then applied to support evolutionary design of
products. An experiment validating some of the theoretical findings is outlined, and concluding remarks are
given.
Keywords:
Design, Lifecycle, Semiotics
1 INTRODUCTION
Historically, design tasks have been verifying theories,
challenging their value in a number of domains varied
from quite ‘objectivistic’ but highly abstract, such as
philosophy and mathematics, to rather situated but
pragmatic, such as economics and management, and
further to essentially subjective and loosely structured,
such as art and sociology. It is then not surprising that in
the present era of technocratic civilization, design
receives significant research attention, having been
studied first as a craft and later – as a field of engineering
that promises to eventually grow into an ‘independent’
discipline developed in line with the traditions of classical
science. In an effort to contribute to the development of
this discipline, both academic researchers and
practitioners, although seeing the problem from fairly
different perspectives, have been trying to understand
and justify if and how one or another theory can be useful
for solving design tasks.
Adherents of mathematical approaches habitually assess
the usefulness of a theory from the standpoint of its
logical consistency and tractability: only those theories are
considered good, which allow for the generation of true
theorems. When it comes to product design, such
theorems can be reduced to statements of the following
form: ‘a design is right if π,’ where π is a formally
determinable quality or property (see, for example,
reference [1] and, to some extent, reference [2]). The role
of the theories is seen to build an apparatus (terminology,
axiomatic basis, inference methods, etc.) for the design
science, while their main target is to achieve better control
over the design process. However, there are many
obstacles preventing the broad introduction of these
theories into practice. The most fundamental obstacle is
the ontological contradiction between the notions of
‘objectively’ good (i.e. right) design and always
subjectively good product. This contradiction makes the
development of a ‘general’ design theory hardly feasible
unless a universal notion of ‘good’ product is formally
clarified.
Proponents of a different viewpoint claim that theories are
good if they allow for understanding design processes
and products better. Having an obvious analytical flavor,
these theories are expected to incorporate a large amount
of domain-specific knowledge to assist the designer in his
or her decision-making (e.g. [3, 4]). It is assumed that the
quality of a product solely depends on knowledge
possessed by the designer, and the problem of designing
good products is, in the main, a problem of professional
expertise. A weak point here is that possessing even
‘ideal’ knowledge is neither necessary nor sufficient
condition for the creation of a commercially successful
product, while the history witnesses that individuals with
little qualification and experience but well developed
creative abilities invented many great products.
In the spirit of recent advances in psychology,
ethnomethodology, anthropology, and cognitive sciences,
some put forward suggestions, highlighting a need for
observationally and analytically studying gifted people,
who seem to be able to design, devise, and develop
without any systematic guidance from the outside (see
reference [5]). It is believed that research of this kind
would ultimately lead one to discovering a set of thinking
techniques and methodologies that can be taught to
designers to make them designing better products (e.g.
[6]). An obvious danger, however, is that apart from the
usual difficulties with the application of any cognitive
model, many factors affecting designing will necessarily
be missed in the resultant techniques because of the
complexity and instability of the ‘network’ topology of
social and psychological relationships: the process of
product concept creation is not limited by activities at the
designer’s side, and it is generally not homogeneous in
terms of space- as well as time-situated interactions [7].
The latter could question the applicability and usefulness
of the obtained techniques at all.
It is now well recognized that there are multiple
perspectives on the building of the design discipline, each
of which contains part of the truth, but none of which
contains the entire solution [8]. Leaving alone marginal
and unconstructive claims like ‘to design, we should only
learn from nature’ or ‘design is art and never be science,’
which nonetheless reflect opinions of some professionals,
a principal question then arises: whether it is possible to
bridge all the different viewpoints and attitudes toward
design within one scientific framework, and if so – what
would be the basis for that?
It is our strong belief that Semiotics – a science about
signs, sign systems, and signification – is exceptionally
well suited to link the mathematical, social, and cognitive
theories together to provide insights about design as a
single discipline and encourage the design community to
develop new methods and tools to aid designers and
lighten the routine burden but improve the quality and
‘goodness’ of products. To justify this supposition, we will
discuss a semiotic approach to the design study and
illustrate it with an example in the following sections. This
paper continues the investigation, which the authors
started with the Semiotic Theory of Creativity for
engineering design [9] that later gave rise to the
development of a more general Semiotic Theory of
Evolutionary Design [10].
2 SEMIOTICS AND THE 3-LEVEL PARADIGM
Perhaps the most significant thing missing from the
traditional design theories is the failure to appreciate how
products are conceived by people as opposed to simply
perceived by people. Indeed, a product, as we
comprehend it in everyday life, is always more than just
an artifact – a physical entity to which human must adapt,
while all the product life-cycle processes have a complex,
meaning-influential dynamics with regard to the product,
its concept, and use-environments. People, through their
activities and practice, originate an ‘objective’ need for the
creation of the product by subjectively assigning its
intended meaning and, at a later time but also
subjectively, evaluate the product’s value in an
environment, i.e. its actual or emergent meaning, yet
having many different views and opinions while doing so.
Semiotics studies this essentially meaning-making
process, where the product among other phenomena of
the environment is construed as a sign, which needs to
be interpreted to allow for the use of the product
throughout its life cycle.
By Peirce, Semiotics deals with three subjects [11]: the
representamen – the sign itself, the object – that which
is signified, and the interpretant – the meaning that
follows semantically from the process of interpretation of
the representamen. The main postulate of Peircean
Semiotics is that no representamen directly points us to
an object, and a sign has a meaning only for a system of
interpretance that is, by its very nature, a system of other
signs – a sign system. Therefore, a representamen is a
sign of an object only for some sign system and not
necessarily for all sign systems: the same representamen
may signify a different object, or the same object may be
signified by a different representamen, etc. – the variety
of possible combinations is not limited by one-to-one
relations.
Modern biology, sociology, and physics teach us that the
richness of the complexity of many natural systems
derives from a strategy to organize smaller units into
larger ones, which are in turn to be arranged into still
larger ones, and so on [12]. As we shall further see, all
the meaning-making processes – the semiosis
processes – reveal a similar hierarchical organization,
and sing systems generally have a multi-level structure
subject to these processes.
Let us assume that in a multi-level sign system, signs on
level N are dynamically composed of signs of level N-1 so
that only those of all the possible combinations of that
lower-level signs occur, which are allowed by boundary
conditions set at level N+1; signs at level N-1 are then
constitutive for level N signs, and signs at level N+1 are
constraining for level N. The dynamics of semiosis can
be described in terms of the interactions among the
adjacent levels of such a sign system [13].
Most naturally for the designer, semiosis is a material
process, which necessarily involves a translation or rerepresentation – re-interpretation – of information from
one meaning-making level to another. Signs on level N
can be representamina of the objects and phenomena at
level N-1 for processes and structures at level N+1, which
form a system of interpretance. The lowest-level signs
(e.g.
physical
objects,
phenomena,
behavioral
dispositions, emotions, and the like) are perceived or
realized through their distinctions and get a
representation at the ‘intermediary’ level of the designer’s
cognition in respect to the interpretive laws of the highest,
experiential and environmentally (culturally, socially,
technically, economically, etc.) induced level, which
accommodates interpretants and assigns a meaning to
the representamina. The design process is the process of
introducing a sign at level N with a meaning for both level
N-1 (as an entity materially grounded there) and level N+1
(as an entity having a material relevance at this level).
Generally, the sign is allowed to have many possible
meanings, depending on contextual constraints from
higher interpretive levels.
People perceive a completed product through its
distinctions that need to be interpreted and may acquire a
meaning for a sign system. Although human perception is
relatively uniform and consistent, it is natural that the
meaning assigned to a representamen can vary
significantly, depending on the subjective dynamics of
perception (level N) and interpretation (level N+1) as well
as on relations between the adjacent semiotic levels.
With a potentially infinite hierarchy of interpretative levels,
where signs on level N+1 can in their turn be constitutive,
i.e. be objects for a higher level and so on, a product is
perceived and conceived in the same (ensemble of) sign
system(s) that makes up a language (here understood in
a very general sense and not limited to handling verbal
constructions). A product can be considered as a ‘text’
written in this language, which has a syntax constraining
the product organization, semantics defining the
meaning(s) of the product, and pragmatics reflecting
various use-effects (e.g. physiological, psychological, or
social) associated with the product. The design science
may be seen as a science about the evolution of the
language – a science that studies fundamental laws of the
semiosis processes, which govern the product life cycle.
3
ELEMENTS OF THE THEORY
3.1 Industrial Semiosis
The concept of Industrial Semiosis categorizes the
product life-cycle processes along three semiotic levels of
meaning emergence [14]: 1) the ontogenic level that
deals with the life history data and future expectations
about a single occurrence of a product; 2) the typogenic
level that holds the processes related to a product type or
generation; and 3) the phylogenic level that embraces
the meaning-affecting processes common to all of the
past and current types and occurrences of a product. The
three levels naturally differ by the characteristic durational
times of the grouped semiosis processes: as one moves
from the lowest, ontogenic level to the higher levels, the
objects become larger and more complicated and have
slower dynamics in both original interpretation and
meaning change.
A product (as concept) starts its development with initially
coinciding onto-, typo-, and phylogenesis processes but
distinct and pre-existing semiotic levels of interpretation.
The concept is evolved, and typogenesis works to reorganize the relationships between the onto- and
phylogenesis processes, as the variety of objects involved
in product development increases. Product types and
their interactions mediate – filter and buffer – between the
levels above and below: not all variety of distinctions
remains available for re-organization as phylos, nor every
lowest-level object have a material relevance there. The
phylogenic level is buffered against variations at the
ontogenic level by the stabilizing mediations at the
typogenic level. (Note that all the three levels of product
definition are to mediate between more global levels of
product environments.)
The dynamics of the interactions between the semiotic
levels can well be described in terms of the basic
processes of variation and selection [15]. In complex
system evolution, variation stands for the generation of a
variety of simultaneously present, distinct entities
(synchronic variety), or of subsequent, distinct states of
the same entity (diachronic variety). Variation makes
variety increase and produces more distinctions.
Selection means, in essence, the elimination of certain
distinct entities and/or states, and it reduces the number
of remaining entities and/or states.
From a semiotic point of view, the variety of a product
intended to operate in an environment is determined by
the devised product structure (i.e. the relations
established between product parts – its synchronic
variety) and the possible relations between the product
and the anticipated environment (i.e. the product feasible
states – its potential diachronic variety), which together
aggregate the product possible configurations. The
variety is defined on the ontogenic level that includes
elements for description of both the structure and
environment. The ontogenesis is driven by variation that
goes through different configurations of the product and
eventually discovers (by distinction selection at every
stage of the product life cycle) configurations, which are
stable on one or another time-scale. A constraint on the
configurations is then imposed, resulting in the selective
retention – emergence of a new meaning for a (not
necessarily new) sign – at the typogenic level. The latter
decreases the variety but specializes the ontogenic level
so that only those distinctions ultimately remain, which fit
to the environment (i.e. only dynamically stable relation
patterns are preserved). Analogously but at a slower timescale, the typogenesis results in the emergence of a new
meaning on the phylogenic level that consecutively
specializes the lower levels. Thus, the main semiotic
principle of product development is such that the
dynamics of the meaning-making processes always seeks
to decrease the number of possible relations between the
product and its environment and hence, the semiosis of
product life cycle is naturally simplified. At the same time,
however, the ‘natural’ dynamics is such that augments the
evolutive potential of the product concept for increasing
its organizational richness: the emergence of new signs
(that may lead to the emergence of new levels of
interpretation) requires a new kind of information and new
descriptive categories must be given to deal with the still
same product.
3.2 On Formalization
Among many possible approaches to formalization of
design semiosis, we have chosen Algebraic Semiotics
(see reference [16]) for its expressiveness, clarity, and
freshness. Algebraic Semiotics deals with signs as
members of sign systems defined to be algebraic theories
with extra structure, and semiosis processes are specified
through semiotic morphisms to be a kind of mapping of
algebraic theories. Instead of defining properties of a sign
system by reference to its members, Algebraic Semiotics,
unlike other approaches but Category Theory, does so by
reference to its external relationships with other sign
systems.
A sign system is represented as a five-tuple
Ξ = 〈S, V, C, R, A〉, where S is a sort-set for signs in the
system, V is a sort-set for data, C is a set of operations
called ‘constructors’ that are used to create signs from
other signs, R is a set of relations defined on the system
signs, and A is a set of axioms that constrain the possible
signs. S and C are partially ordered: by subsort and by
level, respectively; in turn, constructors are partially
ordered by priority within each level.
A semiotic morphism м: Ξ → Ξ′ is a translation that
consists of partial functions, which map sorts,
constructors, predicates and functions of a sign system Ξ
to sorts, constructors, predicates and functions of a sign
system Ξ′ and retain some of the structure of Ξ: the
mapping of sorts to sorts preserves arguments and result
sorts of constructors and predicates as well as the
subsort ordering, and it does not change data sorts.
3.3 Semiosis Laws
In this section, we will put forward a bold conjecture that
(the dynamics of) all the life-cycle meaning-making
processes can be described in terms of basic semiotic
components – algebraic constructions of the following
form:
Pn(мn: fn [Ξn] → Ξn+1),
(1)
where Ξn is a sign system corresponding to a
representation of a (design) problem at time t1, Ξn+1 is a
sign system corresponding to a representation of the
problem at time t2, t2 > t1, fn is a composition of semiotic
morphisms that specifies the interaction of variation and
selection under the condition of information closure, which
requires no external elements be added to the current
sign system; мn is a semiotic morphism, and Pn is the
probability associated with мn, ΣPn = 1, n=1,…,M, where
M is the number of the meaningful transformations of the
resultant sign system after fn. There is a partial ranking –
importance ordering – on the constraints of A in every Ξn,
such that lower ranked constraints can be violated in
order for higher ranked constraints to be satisfied. The
morphisms of fn preserve the ranking.
The Semiotic Theory of Self-Organizing Systems
postulates that in the scale hierarchy of dynamical
organization, a new level emerges if and only if a new
level in the hierarchy of semiotic interpretance
emerges [13]. As the development of a new product
always and naturally causes the emergence of a new
meaning, the above-cited Principle of Emergence directly
leads us to the formulation of the first law of life-cycle
semiosis as follows:
I. The semiosis of a product life cycle is represented by a
sequence of basic semiotic components, such that at
least one of the components is well defined in the
sense that not all of its morphisms of м and f are
isomorphisms, and at least one м in the sequence is
not level-preserving in the sense that it does not
preserve the original partial ordering on levels.
For the present (i.e. for an on-going process), there
exists a probability distribution over the possible мn for
every component in the sequence. For the past (i.e.
retrospectively), each of the distributions collapses to a
single mapping with Pn=1, while the sequence of basic
semiotic components is degenerated to a sequence of
functions. For the future, the life-cycle meaning-making
process can be considered in a very general probabilistic
sense only (e.g. in terms of probability distributions that
are characteristic of a specific domain, social group,
design approach, or the like).
It seems logical to assume that the successful (perhaps,
in any sense) introduction of a product to the market
effects the introduction and settlement of the
corresponding meanings at the onto-, typo-, and
phylogenic semiotic levels. Let us denote the number of
relations between the product and its environment as ε.
We can now formulate the second law of life-cycle
semiosis as follows:
II. A component Pn(мn: fn [Ξn] → Ξn+1) represents a
successful life-cycle semiosis process if the morphism
мn is natural in the sense that εn > εn+1.
Although the above laws have been formulated with
sufficient precision, it is recommended to apply them
(alike Algebraic Semiotics in general) in an informal way,
calling for details only in boundary and difficult situations.
The main purpose of these as well as other not-yetformulated laws of life-cycle semiosis is to guide the
examination of the product development and usage
processes, no matter which design theory or even
paradigm is employed at the lower, applied level.
4 EVOLUTIONARY DESIGN
Evolutionary Design is a relatively new paradigm that
encompasses the recently popular design approaches,
such as Sustainable Design, Design for X, Green Design,
and the like, which explicitly recognize the social,
evolutionary, and error-prone nature of product
development and postulate the tentative character of
design solutions, making them dependent on the
dynamics of product use-environments. In the following
paragraphs, we will give a semiotic interpretation for this
paradigm and show how the laws of life-cycle semiosis
could be applied to support evolutionary design.
4.1 Semiosis of Evolutionary Design
Every design is based on some expectations, which
explicitly or implicitly determine the product intended
meaning and which are realized as design requirements –
once conceived relation patterns. These expectations
always fit particular environmental conditions, which often
become obsolete before the product reaches into the
market place. A universal and obvious solution to this
problem is to increase as much as possible the
synchronic variety of the product by contriving appropriate
decisions in design (the notorious idea of re-configurable
products). Indeed, the more elaborate the structure of the
product (or its concept), the larger the number of
environmental situations in which it can maintain.
Different product configurations can fit (or be adapted to)
different situations and, therefore, in the case of dynamic
environments, design evolution should increase the
synchronic variety, making the product more complex to
adequately react to the environmental changes.
Although the latter statement does not contradict the lifecycle semiosis laws and is, perhaps, true in general, this
does not mean that the ‘best’ product must always be the
most complex one, i.e. be the product with the maximal
synchronic variety. Due to many reasons – economical
(costs), technical (reliability), ecological (energy and
material consumption, pollution), social and ergonomic
(safety, convenience and easiness in production and
operation), etc., the best is the product with the simplest
possible structure for the given functionality, i.e. with the
least possible (for the given environment) synchronic
variety. In this sense, the ‘goodness’ or, better say,
adequacy of the product depends on the characteristics of
product environment in relation to the implemented
design expectations, i.e. it depends on how well the
intended meaning matches the meanings emerging
through onto-, typo-, and phylogenesis (if any).
Design expectations can roughly be classified into two
categories [17]: a) functional – the expectations about
operation of the product and its functional parameters,
and b) environmental – the expectations about the
product-environment interaction. In the product life cycle,
the distinction dynamics is driven by violations of design
expectations. The dynamics of ontogenesis processes,
where relation patterns are originally detected to be
further interpreted and accepted or rejected for an action,
is subject to psychological and physiological laws, such
as, for instance, the well-known law of Weber-Fechner.
Having been differentiated by the time-scale of the
corresponding meaning-making processes, violations of
functional expectations control the product typogenesis,
while violations of environmental expectations influence
both the typo- and phylogenesis semiosis processes.
Resolving the product intended meaning mismatch is a
critical task in design and life cycle engineering that
requires the development of the appropriate information
technologies and tools. Below, we will outline an agentbased technology that was developed to detect violations
of design expectations and support, in this way,
evolutionary design of high-tech products, as well as to
assess the successfulness of the life cycle of an individual
product, product type, or product family on the whole
(also see reference [18]).
4.2 Computer-aided Evolutionary Design
The main idea of the developed technology of
evolutionary design support is to allow for evaluation and
change of the most important design expectations using
programmable mobile agents, called expectation agents,
which utilize design requirements represented explicitly to
monitor product functionality, usage, and operational
environment. An expectation agent typically consists of a
static hardware unit (including transducers, processing
and preprocessing blocks, etc.) integrated with the
product and a mobile software part, which allows the
agent to demonstrate a certain level of autonomy,
intelligence, and proactivity in respect to the product and
act on behalf of the designer or manufacturer. The agent
can execute various control procedures, transfer the
registered data via a communication line, assist the
product user, collect feedback directly from the user, and
update its own code. Data obtained with the agent can be
analyzed and used to detect when the usage and
environmental patterns shift thereby necessitating
optimization of the synchronic variety of the product by
adjusting its configuration. To provide for the evolutionary
support of the life cycle semiosis processes, an effective
information infrastructure connecting products supplied
with agents, service centers, manufacturers, and
designers is to be developed. Figure 1 gives an example
of the agent networking that would be arranged using, for
instance, the existing infrastructure of the World-Wide
Web.
There are three distinct but overlapping layers of
networking driven by the onto-, typo-, and phylogenesis
meaning-making processes. At the ontogenic layer,
expectation agents monitor individual products and their
actual environments. The agents communicate with each
other as well as with the design, manufacturing,
maintenance, and other involved parties. They create
product life histories and try for optimization of technical,
PHYLOGENESIS
Designer’s
cluster
Environmental
cluster
Communication
node
Environmental
cluster
product
product product product
product
product
product
product
ONTOGENESIS
Figure 1: Expectation agent networking.
technologic, and usage processes associated with the
products. Both the product operation and the expectation
agent functionality can be changed by updating the
agent’s program code. The ontogenic layer comprises
single products installed in dynamic environments. An
essential feature of the layer is that no order is imposed
from the outside on the agents’ communication, but the
agents are to be evolutionary grouped by the
environmental characteristics into specific clusters. The
emergence of a cluster at the typogenic layer implies the
emergence of a new meaning – a product type – at the
corresponding level of semiotic interpretance. In an ideal
case, the number of active clusters indicates the number
of product types requisite for the given environment
covered by the agent network.
The typogenic layer provides for systematization of the
design information flows generated by the agents.
Information and data are sorted and stored, depending on
the product type (version or generation). A majority of
actions and processes of this layer are defined on a
population of products grouped into an environmental
cluster. Of course, environmental clusters may naturally
be grouped into still larger units of characteristic
dynamics, e.g. by different manufacturers. Such upperscale grouping would initiate the emergence of a new
meaning – product family – at the phylogenic level of
interpretance.
The phylogenic layer links all the design, production,
maintenance, utilization, etc. processes required for or
associated with the realization of a distinct product or
technology concept. Information flows of this layer are
relatively stable and depend on the global (e.g. cultural
and ecological) rather than specific (e.g. technical and
economic) factors. The first law of life-cycle semiosis
regulates the main processes of phylogenesis and makes
the agent networking really possible.
Figure 2: Refrigerator with an installed expectation agent.
N of interactions
Service
center
Time-scale (from Fast to Slow)
T
Y Manufacturer’s
P
cluster
O
G
E
N
E
S
I
S Environmental
cluster
Adaptation
Normal functioning
Breakdown
Time
Figure 3: The dynamics of product-environment
interaction.
Figure 2 presents a product – refrigerator – with an
embedded expectation agent that is part of the
experimental setup built through our study. The setup is
to compare the behavior of virtual products (i.e. product
intended meanings reconstructed with virtual objects) and
real products by means of multi-media, haptic devices,
and remote sensors (see reference [10] that gives a
detailed account of the experiment). An analysis of
empirical data collected with the setup showed us that the
dynamics of product-environment interactions could serve
as an indicator of the ‘successful’ (for the given
environment) operation of the product.
Figure 3 depicts the characteristic change of the number
of product-environment interactions, such as product part
movements and working mode switchings, registered by
the expectation agent (high-frequency fluctuations have
been filtered out of the data set by a pre-processing block
of the agent). Although it is admitted that a larger scale
empirical study is required to prove the efficiency of the
evolutionary design support based on the expectation
agent networking, the obtained results principally
confirmed our theoretical conjectures and demonstrated
the technological feasibility of the approach (also, see
reference [17] that discusses the case of violations of
design expectations as well as possible strategies to react
to detected violations in the light of the Evolutionary
Design Theory).
5 CONCLUSIONS
We would like to conclude this paper by the following
remarks.
First, the idea of semiotic interpretation of the design
process is not new; moreover, it has become almost a
‘fashion’ for the design community in the last few years,
and there are many publications on this subject. One
principal difference of our work is that we have not limited
the investigation by a semiotic analysis and classification
of design objects and signs representing them but,
instead, focused on the processes responsible for the
development of these objects and their meanings.
Second, we see the rôle of the semiotic approach to
design as not to discriminate true from false, correct from
incorrect, bad from good, etc., but to provide designers
with a new perspective on design theories and
techniques, to better understand how this or that process
goes on, and what are the factors, subjective and
objective, affecting it. Further elaboration of the semiotic
approach will, we believe, shed light upon many nonobvious consequences and causes of the application of a
particular design theory or technique.
[8]
Finally, we come clean that Semiotics alone cannot
account for all the results and insights brought out by all
the design and life cycle theories. Rather, Semiotics can
help us merge all the stages of product development
together within a uniform and universal scientific
framework. The latter could be seen as the ultimate goal
of our research.
[13]
6 ACKNOWLEDGMENTS
The authors would like to acknowledge the financial
support of the Japan Society for the Promotion of Science
(project No 96P00702) that made this research possible.
We are greatly indebted to Dr. I. Goncharenko for his key
contribution of the experimental part of the study.
7 REFERENCES
[1] Suh, N.P., 1999, Applications of Axiomatic Design,
In: H. Kals and F. van Houten (eds), Integration of
Process Knowledge into Design Support Systems,
pp.1-46, Kluwer Academic Publishers
[2] Yoshikawa, H., 1981, General Design Theory and a
CAD System, In: T. Sata and E.A. Warman (eds),
Man-Machine Communication in CAD/CAM, pp. 3558, North-Holland
[3] Klein, R., 1998, A Knowledge Level Theory of
Design and Engineering, In: G. Jacucci, G.J. Olling,
K. Preiss, M.J. Wozny (eds), Globalization of
Manufacturing in the Digital Communication Era of
the 21st Century: Innovation, Agility, and the Virtual
Enterprise, pp. 271-286, Kluwer Academic
Publishers
[4] Hubka, V., Eder, W.E., 1996, Design Science:
Introduction to the needs, scope and organization of
Engineering Design knowledge, Springer-Verlag,
London
[5] Eder, W.E. (ed.), 1996, WDK 24. Engineering
Design and Creativity, The proceedings of the
Workshop EDC, 1995, Pilsen, Czech Republic,
Heurista
[6] Ullman, D.G., Herling, D., D’Ambrosio, B., 1997,
What to do next: Using the problem status to
determine the course of action, Research in
Engineering Design, 9/4:214-227.
[7] Latour, B., 1996, On interobjectivity, Mind, Culture,
and Activity, 3/4:228-245.
[9]
[10]
[11]
[12]
[14]
[15]
[16]
[17]
[18]
Cavallucci, D., Lutz, P., 2000, Intuitive Design
Method (IDM): A New Framework for Design Method
Integration, In: M. Shpitalni (ed.), Proceedings of the
2000 International CIRP Design Seminar, pp. 447452, Haifa, Israel
Kryssanov V.V., Tamaki, H., Kitamura, S., 2001,
Understanding
Design
Fundamentals:
How
Synthesis and Analysis Drive Creativity, Resulting in
Emergence, Artificial Intelligence in Engineering,
15/4:329-342.
Kryssanov, V.V., Goossenaerts, J., Goncharenko, I.,
Tamaki, H., 2000, A Semiotic theory of evolutionary
design: concepts and an illustration, In: F. Kimura
(ed.), Proceedings of the 7th CIRP International
Seminar on Life Cycle Engineering, pp. 68-75,
Tokyo
Hartshorne, C., Weiss, P. (eds), 1960, Collected
papers of Charles Sanders Peirce, 6 volumes, The
Belknap Press of Harvard University Press,
Cambridge, MA
Salthe, S., 1993, Development and Evolution, MIT
Press, Cambridge, MA
Lemke, J.L., 2000, Opening Up Closure: Semiotics
Across Scales, In: J. Chandler and G. van de Vijver
(eds), Closure: Emergent Organizations and their
Dynamics, Annals of the New York Academy of
Science, 901:100-111.
Goossenaerts, J.B.M., 2000, Industrial Semiosis:
Founding the deployment of the Ubiquitous
Information Infrastructure, Computers in Industry,
43:189-201.
Heylighen, F., 1995, (Meta)systems as Constraints
on Variation: a classification and natural history of
metasystem transitions, World Futures: the Journal
of General Evolution, 45:59-85.
Goguen, J.A., 1999, An introduction to algebraic
semiotics, with applications to user interface design,
In: C. Nehaniv (ed.), Computation for Metaphor,
Analogy and Agents, Springer Lecture Notes in
Artificial Intelligence, 1562:242-291.
Goncharenko, I., Kryssanov, V.V., Tamaki, H., 1999,
An Agent-based Approach for Collecting and
Utilizing Design Information Throughout the Product
Life Cycle, In: J.M. Fuertes (ed.), Proceedings of the
7th IEEE International Conference on Emerging
Technologies and Factory Automation ETFA’99, pp.
175-182, UPC, Barcelona, Spain
Kryssanov, V.V., Tamaki, H., Ueda, K., 2001, An
internet-enabled technology to support evolutionary
design, Journal of Engineering Manufacture,
215/B5:647-655.
| 5cs.CE
|
Igor O. Zavadskyi (Taras Shevchenko National University of Kyiv)
[email protected]
A family of fast exact pattern matching algorithms
Abstract. A family of comparison-based exact pattern matching algorithms is described. They utilize
multi-dimensional arrays in order to process more than one adjacent text window in each iteration of the
search cycle. This approach leads to a lower average time complexity by the cost of space. The algorithms of
this family perform well for short patterns and middle size alphabets. In such case the shift of the window by
several pattern lengths at once is quite probable, which is the main factor of algorithm success. Our
algorithms outperform the Boyer-Moore-Horspool algorithm, either in the original version or with Sunday’s
“Quick search” modification, in a wide area of pattern length / alphabet size plane. In some subareas the
proposed algorithms are the fastest among all known exact pattern matching algorithms. Namely, they
perform best when alphabet size is about 30–40 and pattern length is about 4–10. Such parameters are typical
for search in natural language text databases.
Key words: pattern matching, Boyer-Moore, fast search, text search, multi-window.
1 Introduction
Pattern matching is one of the most fundamental
techniques used in computer science. The most
common pattern matching problem is formulated as
finding all the exact occurrences of a given
substring in a larger body of text. Through the
entire presentation we use the following notation:
T – input text
P – pattern to be searched
n – length of the input text
m – length of the pattern
Σ – alphabet of input text and pattern
|Σ| – size of alphabet Σ
|ΣP| – number of different symbols in the pattern
This problem has been systematically studied
since the beginning of seventies. The number of
algorithms more efficient than the simplest straight
forward search (SF) has been discovered. The most
famous of them are the Knuth-Morris-Pratt
algorithm [1], which improves the worst-case time
complexity of the SF from O(nm) to O(n+m) and
Boyer-Moore algorithm [2], which significantly
outperforms the SF in the average.
Almost all known pattern matching algorithms
include the preprocessing stage, when some
preliminary values are obtained basing on the
pattern, and the main search cycle, when the text
body is scanned. In most cases the algorithm
efficiency strongly depends not only on the pattern
and text lengths, but also on the alphabet size.
Since, as a rule, the dependence of the main search
cycle on the text length is linear and time
complexity of the preprocessing stage is negligibly
small, it is worthwhile to compare the algorithm
efficiency on the (|Σ|,m)-plane.
Our research concerns the left up area, where m
is small, |Σ| is large. In this area two modifications
of BM algorithm, namely Boyer-Moore-Horspool
algorithm (BMH) [3] and Sunday’s “Quick Search”
(QS) [4] were considered the best for a decades.
However, in 2000s a number of more efficient
exact pattern matching algorithms were invented.
According to [5] they are FJS for m<8 and |Σ|>32,
TVSBS for m=2 and 8≤|Σ|≤32, EBOM for 4≤m≤16
and 8≤|Σ|≤32, SBNDM for 8≤m≤16 and |Σ|>32 and
FSBNDM for m>16 and |Σ|>8. The algorithms
mentioned cover all three known types of patternmatching algorithms: FJS and TVSBS are
comparison-based; EBOM is automata based,
while SBNDM and FSBNDM algorithms utilize
the bit-parallel operations. We propose the new
comparison-based
algorithms.
Almost
all
algorithms of this type, including FJS, TVSBS and
our new algorithms, exploit the idea of badcharacter shift, which originates from BM search.
It is to compare the last character of the search
window and the last character of the pattern and, if
they do not match, shift the window as long as
possible. The BMH algorithm is based on this idea
only. We develop a generalization of BMH
algorithm, which allows performing several badcharacter shifts at each iteration.
The search cycle of the BMH algorithm is
shown in Fig. 1. Since we analyze the algorithms in
operation level, we try to remove the unnecessary
subtractions denoting n–m by nm and m–1 by m1;
these values can be calculated in a preprocessing
stage. The bad character shift is performed in the
row 7 and its length is equal to D[T[pos+m–1]],
where pos is the current position of the search
window and D is the shift array calculated in a
preprocessing stage. If the ratio |ΣP|/|Σ| is small
enough, the symbol T[pos+m–1] most likely does
not occur in the pattern and the length of this shift
is maximum, i.e. it stands m. These maximum
length shifts are the main factor responsible for the
efficiency of BMH in the left up area of the (|Σ|,m)plane. And we are interested in the most left upper
subarea, where the ratio |ΣP|/|Σ| is particularly
small. In this case one can assume that probably
not only character T[pos+m–1] does not belong to
the pattern, but the characters T[pos+2m–1],
T[pos+3m–1] etc. as well. This means that the
search window can be shifted by several window
lengths at once, or, in other words, several adjacent
search windows can be processed at iteration of the
search cycle. This is the main idea of the multiwindow search algorithms.
1. while pos<nm
2.
j ← m1;
3.
while j>0 AND T[pos+j]=P[j]
4.
j ← j–1;
5.
if j=0
6.
output pos;
7.
pos ← pos+D[T[pos+m1]];
Fig. 1. The main search cycle of the BoyerMoore-Horspool algorithm
Of course, at least k symbols of input text must
be read and processed in each substring of the
length km in order not to miss the possible pattern
occurrence. Thus, at least k readings of input text
characters should be done for each window of the
length km – just the same number as at k iterations
of the single-window algorithm like BMH or QS.
However, we can reduce the number of other
operations using the k-dimensional arrays. Such
arrays occupy rather more memory than the pattern
shift arrays for single-window search algorithms
and their filling takes more preprocessing time.
Nevertheless, as will be shown below, these space
overheads are not that big comparing to memory
size of modern computers, while time overheads
are more than covered in the main search cycle,
resulting in a significant gain in the total for the
wide range of alphabet size / pattern length
combinations.
The idea of using two text windows while
searching all occurrences of the pattern is not new.
It was utilized in Two-Sliding-Windows (TSW)
algorithm [6]. However, these windows were
supposed to be processed in parallel and thus TSW
algorithm is suitable for parallel processor
structures only. Also the idea of two-dimensional
search array was already implemented in a number
of algorithms, for instance, in Berry-Ravindran
algorithm [7], TVSBS and EBOM. However, it
was always proposed to use two adjacent
characters of a text as indices. This significantly
increases the probability of the maximum length
shift if it is low for single-character check but
otherwise leads to superfluous density of the
checks. In other words, if even single-character
check causes the maximum shift with high
probability, there is no need to check two adjacent
characters to shift the text window by m or m+1
positions. In this case it may be better to perform
double-check of the characters T[pos] and
T[pos+m–1], which could shift the text window
by 2m positions at once.
2 Double window algorithm
Let us discuss how to process two adjacent
search windows of the length m, which could be
considered as one window of the double length 2m.
We try to reduce the total number of computing
operations required to process the substring of the
length 2m. Let us examine the main cycle of BMH
algorithm shown in Fig. 1. Two reads from shift
table D in row 7 in two iterations can be replaced
by one if we use two-dimensional shift table
D2|Σ|x|Σ| defined as follows.
D2[i][j] is the leftmost possible position of the
first character of the pattern under assumption that
T[m–1]=i and T[2m–1]=j. All shifts defined by the
table D2 can be divided in 4 types shown in Fig. 2.
a) Neither i nor j belong to pattern P. Then P
can be safely shifted by 2m positions forward.
b) Character i doesn’t belong to pattern P but
j belongs. In this case P can be safely shifted by
more than m–1 symbols but less than 2m. Namely,
the rightmost occurrence of j in P should be
aligned with T[2m–1].
c)
Character i belongs to P and P[m–1]≠i.
Then P can be safely shifted forward by less than m
symbols. Namely, the rightmost occurrence of i in
P should be aligned with T[m–1].
d)
P[m–1]=i. This is the case when the
pattern can be matched in the current position. One
should check if T[0]…T[m–1] coincides with the
pattern before proceed forward.
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
Fig. 2. Pattern shifts in Double window algorithm
while pos<nm
r ← D2[T[pos]][T[pos+m]];
if r=0
j ← 0;
while j<m AND T[pos-m1+j]=P[j]
j ← j+1;
if k=m
output pos-m1;
pos ← pos+D[T[pos]];
else
pos ← pos+r;
Fig. 3. The Double window algorithm main search cycle
Let us calculate the number of operations in
BMH and Double window (DW) algorithms
required for the shift of the text window by 2m
characters forward in the case of the maximum
possible shift. This is the most probable case when
pattern length is very small compared to alphabet
size. In this case only rows 1, 2, 3 and 7 in two
iterations of BMH algorithm and only rows 1, 2, 3
and 10 in one iteration of Double window
algorithm are executed. Note that getting the
element of one-dimensional array like D[x] is
equivalent to *(D+x) in C notation, which requires
one addition and two readings from memory, while
getting the element of two-dimensional |Σ|x|Σ|
array like D2[x][y] is equivalent to *(D2+|Σ|*x+y),
which requires two additions, one multiplication
and three readings from memory (|Σ| is a constant).
The calculations are shown in Table 1. The
number of operations in Double window algorithm
is more than twice less compared to BMH. Also, it
should be noted that memory reads generally take
longer time than operations that could be
performed using processor registers only, e.g.
arithmetic operations. If someone implements the
discussed algorithms in assembler language, the
values of variables pos, m1, m, j etc. can be stored
in processor registers, while the arrays T, D and D2
cannot. In such implementation 8 memory reads
are performed in two iterations of BMH (2 in row 3
and 2 in row 7 in each iteration) and only 3
memory reads needed in one iteration of Double
window algorithm, they are T[pos], T[pos+m] and
D2[T[pos]][T[pos+m]] in row 3. Thus, in the case
of efficient assembler implementation, the Double
window algorithm main cycle outperforms the
BMH main cycle even more than in the case of
implementation in the high-level programming
language, if the longest possible shifts are
performed (Figure 2a).
Table 1. The operational complexity of BoyerMoore-Horspool and Double window algorithms
Operation Boyer-MooreDouble window
Horspool
compari- 4 – (rows 1 and
2 (rows 1 and 3)
son
3)x2
assign4 – (rows 2 and
2 (rows 2 and 10)
ment
7)x2
memory
28 – (2 in row 1; 10 (2 in row 1; 6
reads
1 in row 2; 6 in
in row 2; 2 in row
row 3; 5 in row
10)
7)x2
additions 14 – (3 in row 3; 6 (5 in row 2 and
4 in row 7)x2
1 in row 10)
multipli–
1 (row 2)
cation
Total
50
21
One can observe that even one iteration of
Double window algorithm main cycle requires
fewer operations than one iteration of BMH main
cycle in the case when the condition r=0 is not met.
Therefore, in the cases shown in figures 2b and 2c
the Double window algorithm main cycle still
executes faster than BMH main cycle. Note that in
the case (c) the equality D2[i][j]=D[i] holds, i.e. the
shift length in Double window algorithm is just the
same as in BMH.
Of course, the advantage of DW main cycle
over BMH main cycle in the case (b) is lower than
in the case (a) and in the case (c) is lower than in
the case (b). While the ratio |Σ|P/|Σ| increases, the
balance between cases (a), (b) and (c) moves to (b)
and (c) and then to (c) only. If |Σ|P/|Σ| is close to 1,
the case (c) occurs almost always and
outperformance of DW main cycle over BMH
main cycle is small.
The case shown in Figure 2d for random text
and pattern occurs with the probability 1/|Σ|
regardless of |ΣP| value. In this case the internal
cycle of DW in rows 5 and 6 executes and the
internal cycle of BMH in rows 3 and 4 has more
than one iteration. Each iteration of DW internal
cycle requires more time than that one of BMH,
because the comparison T[pos–m1+j]=P[j] requires
4 additions/subtractions and 6 readings from
memory, while the comparison T[pos+j]=P[j]
consists of only 3 additions and 5 readings from
memory.
As a result, the main cycle of the DW algorithm
is essentially faster than the main cycle of BMH
algorithm when the following conditions are met:
(1) the alphabet is large enough to make the
case (d) not frequent;
(2) the ratio |ΣP|/|Σ| is small enough to make
the case (c) not frequent.
The simulation shows that only when alphabet
size is up to 4, the condition (1) violation forces the
DW algorithm main cycle to run slower than BMH
main cycle on random pattern and text. Any
alphabet of size greater than 4 could be considered
as “large enough”. The violation of condition (2)
forces the DW algorithm main cycle to run
approximately at the same speed as the BMH main
cycle. However, for the wide range of pattern
length / alphabet size combinations the DW is
essentially faster than BMH. This range covers m
in the range 20–30 and |Σ|>4.
3 Multi-window extension
Let us consider the possibility of processing
more than 2 adjacent text windows in one iteration.
The modification of the DW algorithm is simple:
the N-dimensional array DN should be used instead
of D2. It is defined as follows.
DN[i1]…[iN] is the leftmost possible position of
the beginning of the pattern under assumption that
T[km–1]=ik, k=1,..,N.
In Figure 3 only row 2 should be changed in a
following way:
r ← DN[T[pos]][T[pos+2m]]…[T[pos+Nm]];
Thus we obtain the Triple Window (TW),
Quadruple Window (QW) and other Multi-window
algorithms.
Using C notation this assignment can be
rewritten as r=*(DN+bN–1pos+…+b1(pos+(N–1)m)
+pos+Nm), where bk=|Σ|k. Of course, in order to
reduce the number of multiplications, the values
2m,…,Nm can be calculated in the preprocessing
stage, while the values bk could be considered as
constants. However, every next dimension adds
two additions, one multiplication and two memory
reads to the calculation routine. This overhead is
covered by longer shifts until the value of |ΣP|/|Σ| is
small enough, but in the case N≥3 the
preprocessing begins to play an important role,
since its time and space complexity grows
exponentially depending on N. The preprocessing
stage is discussed in the next section.
4 Preprocessing
On the preprocessing stage of the N-window
algorithm the values km, k=1,…,N are calculated
and arrays D containing |Σ| elements and DN with
|Σ|n elements are filled with the values. Filling the
array D is just the same as in the BMH algorithm
and runs in O(|Σ|). Obtaining values km is O(N)
time, which is ignorable small for realistic values
of N. Filling the array DN takes up almost all the
time. The following procedure completes this task.
1. Assign the value Nm to all elements.
2. Replace the values D[i1]…[iN], where
iN∈P, with Nm–qN–1, where qN is the rightmost
position of the character iN in the pattern P.
…
N+1.
Replace the values D[i1]…[iN],
where i1∈P, with m–q1–1, where q1 is the rightmost
position of the character i1 in the pattern P.
The first step takes O(|Σ|N) time, while each
other step – O(m|Σ|N–1) time. The overall time
complexity
of
preprocessing
stage
is
O(|Σ|N+Nm|Σ|N–1).
Using the special functions that copy memory
blocks, like memcpy from memory.h C library,
one can build the implementation that is faster in
times than the conventional method given above.
The space complexity of Multi-window
algorithms is, of course, strongly greater than that
one for BMH/QS. However, memory requirements
of D2 array even for relatively large alphabet
containing 256 symbols are only 64 Kb, which is
absolutely admissible for present-day computers
and programs.
The TW algorithm, as computational
experiment shows, is efficient only for b values
around 32. Then the size of D3 (32x32x32) array is
only 32 Kb. The QW algorithm could be efficient
for b=16 at maximum and the size of the respective
D4 array is 64 Kb.
5 Unrolling cycle
The algorithms based on bad character rule,
such as BMH or Multi-window, could be
accelerated using the unrolling cycle technique. It
consist of applying “blind” shifts, without checking
the end of file, until shift value is not positive. In
order not to miss the end of the file, the text is
appended by the fictitious pattern.
This technique could be applied to Multiwindow algorithms in a following way:
- Row 1 should be replaced with the endless
cycle “while 1”.
After row 3 we should check the condition
pos≥nm and break the cycle if it is met.
This allows us to check the end of the file only
when the shift array element is equal to zero and
speed up the multi-window algorithms by 2–4%.
6 Computational experiment
We implement the unrolled versions of Double
window, Triple-window (TW), Quadruple-window
(QW) and a number of other known algorithms in
C language, use the Microsoft Visual studio
compiler to build the executables and run them on
the Athlon II X2 245 processor, 2.9 GHz, Windows
7 platform. Preprocessing stage of the TW and QW
algorithms is implemented using the fast memory
fill functions. The text containing 10 MB
characters is randomly built and the patterns as
well. The distribution of characters is uniform.
The results for alphabet of size 32 and different
pattern lengths are presented in Table 2. The total
running time of 1000 runs is shown. The QW
results are not shown, since they are significantly
worse than DW and TW results for alphabet size
32. The QW algorithm could be efficient for
smaller alphabets and very short patterns.
Table 2. The total running time of pattern matching algorithms, 10 MB of text × 1000, in seconds
BMH
m=2
m=3
m=4
m=5
m=6
m=7
m=8
m=9
m=10
m=11
m=12
45,326
33,320
24,552
19,919
16,458
15,049
14,279
12,170
11,097
10,541
9,874
QS
31,678
25,596
20,366
17,358
14,583
13,724
13,213
11,327
10,411
9,788
9,432
DW
34,030
24,429
18,836
15,558
13,313
11,957
11,704
10,181
9,363
9,088
8,522
TW
27,229
21,396
15,812
13,630
11,623
10,706
10,913
9,316
8,990
8,547
8,198
7 Conclusions
As is seen, the Triple Window algorithm is
superior over all known algorithms for pattern
lengths from 4 to 9 and alphabet size 32. Both DW
and TW algorithms outperform classical
comparison-based algorithms, such as BMH and
QS for all considered values of m. The TVSBS,
FJS and SBNDM are better for very short patterns,
EBOM
90,742
52,538
33,695
25,020
19,524
17,003
15,945
13,299
11,667
10,673
9,949
FJS
26,169
22,890
19,420
16,397
13,926
12,880
12,550
10,846
9,826
9,175
8,859
TVSBS
24,186
21,109
18,182
16,025
13,824
13,002
12,680
11,212
10,218
9,454
9,071
SBNDM
FSBNDM
25,689
20,034
16,071
14,127
12,328
11,855
11,725
10,391
10,076
9,401
9,271
36,091
26,177
19,477
16,273
13,250
11,930
11,451
9,624
8,963
8,103
7,726
while for pattern length 10 and more the FSBNDM
algorithm becomes superior.
The technique used in Multi-window algorithm
family could be applied in order to accelerate the
algorithms based on comparisons of adjacent
characters, which is a future research direction.
References
1. KNUTH D. E., MORRIS J. H., and
PRATT V. R. (1977) Fast pattern matching in
strings, SIAM J. Comput. 6 (1), pp. 323–350.
2. BOYER R. S. and MOORE J. S. (1977) A
fast string searching algorithm, Commun. ACM
20, 10, pp. 762–772.
3. HORSPOOL R. N. (1980) Practical fast
search in strings, Softw.Pract. Exp. 10, 6, 501–506.
4. SUNDAY D. M. (1990) A very fast
substring search algorithm, Commun. ACM 33, 8,
pp. 132–142.
5. FARO S., LECROQ T. The Exact Online
String Matching Problem: a Review of the Most
Recent Results (2013), ACM Computing Surveys
(CSUR) Surveys Homepage archive. Volume 45
Issue 2, Article No. 13.
6. HUDAIB
A.,
AL-KHALID
R.,
SULEIMAN D., ITRIQ M., AL-ANANI A. (2008),
A fast pattern matching algorithm with two sliding
windows (TSW), J. Comput. Sci. 4, 5, pp. 393-401.
7. BERRY T., RAVINDRAN S. (1999) A fast
string matching algorithm and experimental
results, In Proceedings of the Prague Stringology
Club Workshop ’99, pp. 16–28.
| 8cs.DS
|
SECOND-ORDER MATRIX CONCENTRATION INEQUALITIES
arXiv:1504.05919v2 [math.PR] 3 Aug 2016
JOEL A. TROPP
A BSTRACT. Matrix concentration inequalities give bounds for the spectral-norm deviation of a random matrix from its
expected value. These results have a weak dimensional dependence that is sometimes, but not always, necessary. This paper identifies one of the sources of the dimensional term and exploits this insight to develop sharper matrix concentration
inequalities. In particular, this analysis delivers two refinements of the matrix Khintchine inequality that use information
beyond the matrix variance to reduce or eliminate the dimensional dependence.
1. M OTIVATION
Matrix concentration inequalities provide spectral information about a random matrix that depends smoothly
on many independent random variables. In recent years, these results have become a dominant tool in applied
random matrix theory. There are several reasons for the success of this approach.
• Flexibility. Matrix concentration applies to a wide range of random matrix models. In particular, we can
obtain bounds for the spectral norm of a sum of independent random matrices in terms of the properties
of the summands.
• Ease of Use. For many applications, matrix concentration tools require only a small amount of matrix
analysis. No expertise in random matrix theory is required to invoke the results.
• Power. For a large class of examples, including independent sums, matrix concentration bounds are provably close to optimal.
See the monograph [Tro15] for an overview of this theory and a comprehensive bibliography.
The matrix concentration inequalities in the literature are suboptimal for certain examples because of a weak
dependence on the dimension of the random matrix. Removing this dimensional term is difficult because there are
many situations where it is necessary. The purpose of this paper is to identify one of the sources of the dimensional
factor. Using this insight, we will develop some new matrix concentration inequalities that are qualitatively better
than the current generation of results, although they sacrifice some of our desiderata. Ultimately, we hope that this
line of research will lead to general tools for applied random matrix theory that are flexible, easy to use, and that
give sharp results in most cases.
2. T HE M ATRIX K HINTCHINE I NEQUALITY
To set the stage, we present and discuss the primordial matrix concentration result, the matrix Khintchine inequality, which describes the behavior of a special random matrix model, called a matrix Gaussian series. This
result already exhibits the key features of more sophisticated matrix concentration inequalities, and it can be used
to derive concentration bounds for more general models. As such, the matrix Khintchine inequality serves as a
natural starting point for deeper investigations.
2.1. Matrix Gaussian Series. In this work, we focus on an important class of random matrices that has a lot of
modeling power but still supports an interesting theory.
Definition 2.1 (Matrix Gaussian Series). Consider fixed Hermitian matrices H1 , . . . , Hn with common dimension d,
and let {γ1 , . . . , γn } be an independent family of standard normal random variables. Construct the random matrix
Xn
X := i=1 γi Hi .
(2.1)
Date: 13 March 2015. Revised 21 April 2015 and 3 August 2016.
2010 Mathematics Subject Classification. Primary: 60B20. Secondary: 60F10, 60G50, 60G42.
Key words and phrases. Concentration inequality, moment inequality, random matrix.
Email: [email protected] . Tel: 626.395.5957.
1
2
J. A. TROPP
We refer to a random matrix with this form as a matrix Gaussian series with Hermitian coefficients or, for brevity,
an Hermitian matrix Gaussian series.
Matrix Gaussian series enjoy a surprising amount of modeling power. It is easy to see that we can express
any random Hermitian matrix with jointly Gaussian entries in the form (2.1). More generally, we can use matrix
Gaussian series to analyze a sum of independent, zero-mean, random, Hermitian matrices Y 1 , . . . , Y n . Indeed, for
any norm |||·||| on matrices,
ii
h h ¯¯¯X
¯¯¯Xn
¯¯¯ p
¯¯¯ ¯
n
(2.2)
E ¯¯¯ i=1 Y i ¯¯¯ ≤ 2π · E E ¯¯¯ i=1 γi Y i ¯¯¯ ¯ Y 1 , . . . , Y n .
The process of passing from an independent sum to a conditional Gaussian series is called symmetrization. See [LT91,
Lem. 6.3 and Eqn. (4.8)] for details about this calculation. Furthermore, some techniques for Gaussian series can
be adapted to study independent sums directly without the artifice of symmetrization.
Note that our restriction to Hermitian matrices is not really a limitation. We can also analyze a rectangular matrix Z with jointly Gaussian entries by working with the Hermitian dilation of Z , sometimes known as the Jordan–
Wielandt matrix. See [Tro15, Sec. 2.1.16] for more information on this approach.
2.2. The Matrix Variance. Many matrix concentration inequalities are expressed most naturally in terms of a matrix extension of the variance.
Definition 2.2 (Matrix Variance). Let X be a random Hermitian matrix. The matrix variance is the deterministic
matrix
Var(X ) := E X 2 − (E X )2 .
We use the convention that the power binds before the expectation.
P
In particular, consider a matrix Gaussian series X := ni=1 γi Hi . It is easy to verify that
Xn
Xn
Var(X ) = E X 2 = i,j =1 E[γi γ j ] · Hi H j = i=1 Hi2 .
We see that the matrix variance has a clean expression in terms of the coefficients of the Gaussian series, so it is
easy to compute in practice.
2.3. The Matrix Khintchine Inequality. The matrix Khintchine inequality is a fundamental fact about the behavior of matrix Gaussian series. The first version of this result was established by Lust-Piquard [LP86], and the constants were refined in the papers [Pis98, Buc01]. The version here is adapted from [MJC+ 14, Sec. 7.1].
P
Proposition 2.3 (Matrix Khintchine). Consider an Hermitian matrix Gaussian series X := ni=1 γi Hi , as in (2.1).
Introduce the matrix standard deviation parameter
°¡X
°
°
¢1/2 °
°
° n
for q ≥ 1.
(2.3)
σq (X ) := °Var(X )1/2 °q = ° i=1 Hi2
°
q
Then, for each integer p ≥ 1,
σ2p (X )
≤
¡
2p ¢1/(2p)
E kX k2p
The symbol k·kq denotes the Schatten q-norm.
≤
p
2p − 1 · σ2p (X ).
(2.4)
The lower bound in (2.4) is simply Jensen’s inequality. Section 7 contains a short proof of the upper bound.
The matrix Khintchine inequality also yields an estimate for the spectral norm of a matrix Gaussian series. This
type of result is often more useful in practice.
P
Corollary 2.4 (Matrix Khintchine: Spectral Norm). Consider an Hermitian matrix Gaussian series X := ni=1 γi Hi
with dimension d, as in (2.1). Introduce the matrix standard deviation parameter
°1/2
°Xn
σ(X ) := kVar(X )k1/2 = ° i=1 Hi2 ° .
Then
q
1
e (1 + 2log d) · σ(X ).
p · σ(X ) ≤ E kX k ≤
2
The symbol k·k denotes the spectral norm, also known as the ℓ2 operator norm.
(2.5)
SECOND-ORDER MATRIX CONCENTRATION INEQUALITIES
3
Proof Sketch. For the upper bound, observe that
´
³
p
p
°
°
2p 1/(2p)
≤ 2p − 1 · °Var(X )1/2 °2p ≤ d 1/(2p) 2p − 1 · kVar(X )k1/2 .
E kX k ≤ E kX k2p
Indeed, the spectral norm is bounded above by the Schatten 2p-norm, and we can apply Lyapunov’s inequality
to increase the order of the moment from one to 2p. Invoke Proposition 2.3, and bound the trace in terms of the
spectral norm again. Finally, set p = ⌈log d⌉, and simplify the constants.
For the lower bound, note that
°2 ´1/2
¢1/2
1 ³ °Xn
1 ¡
1
E kX k ≥ p E ° i=1 γi Hi °
= p E kX 2 k
≥ p kVar(X )k1/2 .
2
2
2
The first relation follows from the optimal Khintchine–Kahane inequality [LO94]; the last is Jensen’s.
2.4. Two Examples. The bound (2.5) shows that the matrix standard deviation controls the expected norm of a
matrix Gaussian series up to a factor that is logarithmic in the dimension of the random matrix. One may wonder
whether the lower branch or the upper branch of (2.5) gives the more accurate result. In fact, natural examples
demonstrate that both extremes of behavior occur.
For an integer d ≥ 1, define
γ1
γ2
γ3
.
(2.6)
X diag := X diag (d) :=
..
.
γd
That is, X diag is a d × d diagonal matrix whose entries {γi : 1 ≤ i ≤ d} are independent standard normal variables.
Second, define
γ11 γ12 . . . γ1d
γ
21 γ22 . . . γ2d
1
∗
(2.7)
X goe := X goe (d) := p (G + G ) where G := G(d) := .
..
..
..
.
2d
..
.
.
γd 1 γd 2 . . . γd d
The symbol ∗ denotes conjugate transposition. Up to scaling, the d × d random matrix X goe is the Hermitian
part of a matrix G whose entries {γi j : 1 ≤ i , j ≤ d} are independent standard normal variables. The sequence
{X goe (d) : d = 1, 2, 3, . . . } is called the Gaussian orthogonal ensemble (GOE).
To apply the matrix Khintchine inequality, we represent each matrix as an Hermitian Gaussian series:
X diag =
Xd
i=1 γi
1 X
and X goe = p
1≤i,j ≤d γi j (Ei j + E j i ).
2d
Eii
We have written Ei j for the d × d matrix with a one in the (i , j ) position and zeros elsewhere. Respectively, the
matrix variances satisfy
¡
¢
Var(X diag ) = I and Var(X goe ) = 1 + d −1 · I.
The bound (2.5) delivers
1
p
2
≤
E kX k
/
q
2e log d
for X = X diag or X = X goe .
(2.8)
The relations
/ and ≈ suppress lower-order terms. In each case, the ratio between the lower and upper bound has
p
order log d. The matrix Khintchine inequality does not provide more precise information.
On the other hand, for these examples, detailed spectral information is available:
q
(2.9)
E kX goe k ≈ 2 and E kX diag k ≈ 2log d.
See [Tao12, Sec. 2.3] for a proof of the result on the GOE matrix; the bound for the diagonal matrix depends on the
familiar calculation of the expected maximum of d independent standard normal random variables. We see that
the norm of the GOE matrix is close to the lower bound provided by (2.5), while the norm of the diagonal matrix is
close to the upper bound.
4
J. A. TROPP
2.5. A Question. Corollary 2.4 shows that the matrix variance controls the expected norm of a matrix Gaussian
series. On the other hand, the two examples in the previous section demonstrate that we need more information
than the variance to determine the norm up to a constant factor. Therefore, we must ask...
Are there parameters that allow us to calculate the norm of a matrix Gaussian series more
precisely than the matrix variance?
This paper provides the first affirmative answer to this question.
3. B EYOND THE M ATRIX K HINTCHINE I NEQUALITY
This section presents new results that improve on the matrix Khintchine inequality, Proposition 2.3. First, we
motivate the type of parameters that arise when we try to refine this result. Then we define a quantity, called the
matrix alignment parameter, that describes how the coefficients in the matrix Gaussian series interact with each
other. In Section 3.3, we use the alignment parameter to state a new bound that provides a uniform improvement
over the matrix Khintchine inequality. Further refinements are possible if we consider random matrices with highly
symmetric distributions, so we introduce the class of strongly isotropic random matrices in Section 3.5. Section 3.6
contains a matrix Khintchine inequality for matrix Gaussian series that are strongly isotropic. This bound is good
enough to compute the norm of a large GOE matrix exactly. Finally, in Sections 3.7 and 3.8, we discuss extensions
and related work.
3.1. Prospects. What kind of parameters might allow us to refine Proposition 2.3? The result is already an identity
for p = 1. For inspiration, let us work out what happens when p = 2:
¡Xn
¢4 Xn
£
¤
E kX k44 = E tr
= i,j ,k,ℓ=1 E[γi γ j γk γℓ ] · tr Hi H j Hk Hℓ
i=1 γi H i
´2
³X
³X
´
n
n
2
+
tr
= 2tr
=: 2tr Var(X )2 + tr ∆.
H
H
H
H
H
i
j
i
j
i,j =1
i,j =1 i
We use the convention that powers bind before the trace. The product of Gaussian variables has expectation zero
unless the indices are paired. In the last expression, the first term comes from the cases where i = j and k = ℓ or
where i = ℓ and j = k; the second term comes from the case where i = k and j = ℓ. Once again, the matrix variance Var(X ) emerges, but we have a new second-order term ∆ that arises from the summands where the indices
alternate: (i , j , i , j ).1
In a sense, the matrix ∆ reflects the extent to which the coefficient matrices are aligned. When the family {Hi }
commutes, the matrix ∆ = Var(X )2 , so the second-order term provides no new information. More generally, when2p
ever the coefficients commute, the quantity (E kX k2p )1/(2p) can be expressed in terms of the matrix variance and
the number p, and the matrix Khintchine inequality, Proposition 2.3, gives an estimate of the correct order. In other
words, commuting coefficients are the worst possible circumstance. Most previous work on matrix concentration
implicitly uses this worst-case model in the analysis.
To achieve better results, we need to account for how the coefficient matrices Hi interact with each other. The
calculation above suggests that the matrix ∆ might contain the information we need. Heuristically, when the
coefficients fail to commute, the matrix ∆ should be small. As we will see, this idea is fruitful, but we need a
parameter more discerning than ∆.
Let us summarize this discussion in the following observation:
To improve on the matrix Khintchine inequality, we must quantify the extent to which the
coefficient matrices commute.
Our work builds on this intuition to establish new matrix concentration inequalities.
3.2. The Matrix Alignment Parameter. In this section, we introduce a new parameter for a matrix Gaussian series
that describes how much the coefficients commute with each other. In later sections, we will present extensions of
the matrix Khintchine inequality that rely on this parameter.
Definition 3.1 (Matrix Alignment Parameter). Let H1 , . . . , Hn be Hermitian matrices with dimension d. For each
p ≥ 1, the matrix alignment parameter of this sequence is the quantity
°¯
¯1/4 °
°¯Xn
¯ °
°
and w := w ∞ .
(3.1)
w p := max °¯ i,j =1 Hi Q 1 H j Q 2 Hi Q 3 H j ¯ °
°
Qℓ
p
1A related observation animates the theory of free probability, which gives a fine description of certain large random matrices. The key fact
about centered, free random variables Y and Z is that crossing moments, such as ϕ(Y Z Y Z ), must vanish [NS06].
SECOND-ORDER MATRIX CONCENTRATION INEQUALITIES
5
The maximum takes place over a triple (Q 1 ,Q 2 ,Q 3 ) of unitary matrices with dimension d. The matrix absolute
value is defined as |B | := (B ∗ B )1/2 .
Roughly, the matrix alignment parameter (3.1) describes how well the matrices H1 , . . . , Hn can be aligned with each
other under the worst choices of coordinates.
The quantity (3.1) appears mysterious, so it is worth a few paragraphs to clarify its meaning. First, let us compare
the alignment parameter with the matrix standard deviation parameter (2.3) that appears in the matrix Khintchine
inequality.
Proposition 3.2 (Standard Deviation versus Alignment). Let H1 , . . . , Hn be Hermitian matrices. Define the standard
deviation and alignment parameters
°¯
°¡X
¯1/4 °
°¯Xn
¢1/2 °
° n
°
¯ °
°
σp := ° i=1 Hi
and w p := max °¯ i,j =1 Hi Q 1 H j Q 2 Hi Q 3 H j ¯ °
for p ≥ 1.
°
°
p
Q
p
ℓ
Then
w p ≤ σp
for all p ≥ 4.
The proof of Proposition 3.2 appears in Section 8.9.
Next, let us return to the examples in the introduction. In Section 4, we provide detailed calculations of the
standard deviation and alignment parameters. For the diagonal Gaussian series X diag defined in (2.6), we have
σ(X diag ) = 1 and w(X diag ) = 1.
For the GOE matrix X goe defined in (2.7),
σ(X goe ) = 1 + d −1
and w(X goe ) ≤ (4d)−1/4 .
(3.2)
The matrix alignment parameter can tell the two examples apart, while the matrix standard deviation cannot!
Remark 3.3 (Notation for Alignment). Here and elsewhere, we abuse notation by writing w p (X ) and w(X ) for the
P
alignment parameter of a matrix Gaussian series X := ni=1 γi Hi , even though w is a function of the coefficient
matrices Hi in the representation of the series.
Remark 3.4 (Are the Unitaries Necessary?). At this stage, it may seem capricious to include the unitary matrices in
the definition (3.1). In fact, the example in Section 4.3 demonstrates that the alignment parameter would lose its
value if we were to remove the unitary matrices. On the other hand, there are situations where the unitary matrices
are not completely arbitrary, as discussed in Section 8.1.
3.3. A Second-Order Matrix Khintchine Inequality. The first major result of this paper is an improvement on
the matrix Khintchine inequality. This theorem uses the second-order information in the alignment parameter to
obtain better bounds.
P
Theorem 3.5 (Second-Order Matrix Khintchine). Consider an Hermitian matrix Gaussian series X := ni=1 γi Hi , as
in (2.1). Define the matrix standard deviation and matrix alignment parameters
°¯
¯1/4 °
°¡X
°¯Xn
¢ °
¯ °
° n
2 1/2 °
°
for q ≥ 1.
and w q (X ) := max °¯ i,j =1 Hi Q 1 H j Q 2 Hi Q 3 H j ¯ °
σq (X ) := ° i=1 Hi
°
°
q
Q
q
ℓ
The maximum takes place over a triple (Q 1 ,Q 2 ,Q 3 ) of unitary matrices. Then, for each integer p ≥ 3,
p
p
¡
2p ¢1/(2p)
E kX k2p
≤ 3 4 2p − 5 · σ2p (X ) + 2p − 4 · w 2p (X ).
(3.3)
The symbol k·kq denotes the Schatten q-norm.
The proof of Theorem 3.5 appears in Section 8.
We can also derive bounds for the spectral norm of a matrix Gaussian series.
Corollary 3.6 (Second-Order Matrix Khintchine: Spectral Norm). Consider an Hermitian matrix Gaussian series
P
X := ni=1 γi Hi with dimension d ≥ 8, as in (2.1). Define the matrix standard deviation and matrix alignment
parameters
°1/4
°X
°1/2
°Xn
°
° n
and w(X ) := max °
σ(X ) := °
H 2°
Hi Q 1 H j Q 2 Hi Q 3 H j ° .
i=1
i
Qℓ
i,j =1
The maximum ranges over a triple (Q 1 ,Q 2 ,Q 3 ) of unitary matrices. Then
q
q
E kX k ≤ 3σ(X ) 4 2e log d + w(X ) 2e log d.
(3.4)
6
J. A. TROPP
The symbol k·k denotes the spectral norm.
The result follows from Theorem 3.5 by setting p = ⌈log d⌉. The potential gain in (3.4) over (2.5) comes from the
reduction of the power on the first logarithm from one-half to one-quarter.
3.4. Matrix Khintchine versus Second-Order Matrix Khintchine. Let us make some comparisons between Proposition 2.3 and Theorem 3.5. First, recall that the alignment parameter is dominated by the standard deviation
parameter: w q (X ) ≤ σq (X ) for q ≥ 4 because of Proposition 3.2. Therefore, the bound (3.3) implies that
p
¡
¡ p
¢
2p ¢1/(2p)
E kX k2p
≤ 3 4 2p − 1 + 2p − 1 · σ2p (X ) for p = 3, 4, 5, . . . .
This is very close to the prediction from Proposition 2.3, so Theorem 3.5 is never significantly worse.
On the other hand, there are situations where Theorem 3.5 gives qualitatively better results. In particular, for
the GOE matrix X goe (d), the bound (3.3) and the calculation (3.2) yield
p
q
2e log d
for d ≥ 8.
E kX goe (d)k ≤ 3 4 2e log d + p
4
4d
This estimate beats our first attempt in (2.8), but it still falls short of the correct estimate E kX goe k ≈ 2.
3.5. Strongly Isotropic Random Matrices. As we have seen, Theorem 3.5 offers a qualitative improvement over
the matrix Khintchine inequality, Proposition 2.3. Nevertheless, the new result still lacks the power to determine
the norm of the GOE matrix correctly. We can obtain more satisfactory results by specializing our attention to a
class of random matrices with highly symmetric distributions.
Definition 3.7 (Strong Isotropy). Let X be a random Hermitian matrix. We say that X is strongly isotropic when
¡
¢
¯ X p · I for p = 0, 1, 2, . . . .
E X p = E tr
¯ denotes the normalized trace: tr
¯ A := d −1 tr A when A has dimension d.
The symbol tr
The easiest way to check that a random matrix is strongly isotropic is to exploit symmetry properties of the
distribution. We offer one of many possible results in this direction [CT14, Lem. 7.1].
Proposition 3.8 (Strong Isotropy: Sufficient Condition). Let X be a random Hermitian matrix. Suppose that the
distribution of X is invariant under signed permutation:
X ∼ Π∗ X Π for every signed permutation Π.
Then X is strongly isotropic. The symbol ∼ refers to equality of distribution. A signed permutation is a square matrix
that has precisely one nonzero entry in each row or column, this entry taking the values ±1.
Proof. Suppose that Π is a signed permutation, drawn uniformly at random. For p = 0, 1, 2, 3, . . . ,
¯ ¤¤
£
¤
£ £
£¡
¢ ¤ ¡
¢
¯ X p · I = E tr
¯ X p ·I
E X p = E (Π∗ X Π)p = E E Π∗ X p Π ¯ X = E tr
The first relation uses invariance under signed permutation, and the second relies on the fact that signed permutations are unitary. Averaging a fixed matrix over signed permutations yields the identity times the normalized trace
of the matrix.
Proposition 3.8 applies to many types of random matrices. In particular, the diagonal Gaussian matrix X diag and
the GOE matrix X goe are both strongly isotropic because of this result. Other types of distributional symmetry can
also lead to strong isotropy.
Remark 3.9 (Group Orbits). Here is a more general class of matrix Gaussian series where we can verify strong
isotropy using abstract arguments. Let G be a unitary representation of a finite group, and let A be a fixed Hermitian matrix with the same dimension. Consider the random Hermitian matrix
X
X :=
γU ·U AU ∗
U ∈G
where {γU : U ∈ G } is an independent family of standard normal variables. Since G acts on itself by permutation,
U XU∗ ∼ X
for each U ∈ G .
This observation allows us to perform averaging arguments like the one in Proposition 3.8.
SECOND-ORDER MATRIX CONCENTRATION INEQUALITIES
7
There are several ways to apply this property to argue that X is strongly isotropic. For example, it suffices that
G ′ := {M : MU = U M for all U ∈ G } = {zI : z ∈ C}.
It is also sufficient that {U a : U ∈ G } forms a (complete) tight frame for every vector a; see the paper [VW08] for
some situations where this condition holds.
Remark 3.10 (Spherical Designs). A spherical t -design is a collection {ui : i = 1, . . . , N } of points on the unit sphere
Sd −1 in Rd with the property that
Z
N
1 X
ϕ(ui )
ϕ(u) dµ(u) =
N i=1
Sd−1
where ϕ is an arbitrary algebraic polynomial in d variables with degree t and dµ is the Haar measure on the sphere
Sd −1 . See the paper [BRV13] for existence results and background references.
Given a spherical t -design, consider the random matrix
X :=
N
X
i=1
γi ui ui∗ .
where {γi : i = 1, . . . , N } is an independent family of standard normal variables. By construction, this random matrix
has the property that
¯ X p ) · I for p = 0, 1, 2, . . . , ⌊t /2⌋.
E X p = (E tr
This variant of the strong isotropy property is sufficient for many purposes, provided that t ≈ log d.
3.6. A Second-Order Khintchine Inequality under Strong Isotropy. The second major result of this paper is a
second-order matrix Khintchine inequality that is valid for matrix Gaussian series with the strong isotropy property.
Like Theorem 3.5, this result uses the alignment parameter to control the norm of the random matrix.
Theorem 3.11 (Second-Order Khintchine under Strong Isotropy). Consider an Hermitian matrix Gaussian series
P
X := ni=1 γi Hi with dimension d, as in (2.1), and assume that X is strongly isotropic. Introduce the matrix standard
deviation and matrix alignment parameters:
°1/4
°X
°1/2
°Xn
°
° n
and w(X ) := max ° i,j =1 Hi Q 1 H j Q 2 Hi Q 3 H j ° .
σ(X ) := ° i=1 Hi2 °
Qℓ
The maximum ranges over a triple (Q 1 ,Q 2 ,Q 3 ) of unitary matrices. Then, for each integer p ≥ 1,
¡
£
¤
2p ¢1/(2p)
E kX k2p
≤
2σ(X ) + 21/4 p 5/4 w(X ) · d 1/(2p) .
The symbol k·k refers to the spectral norm, while k·kq is the Schatten q-norm.
The proof of this result appears in Section 9, where we also establish a lower bound.
Theorem 3.11 shows that the moments of the random matrix X are controlled by the standard deviation σ(X )
whenever p 5/4 w(X ) ≪ σ(X ). If we take p = ⌈log d⌉, the Schatten 2p-norm is essentially the same as the spectral
norm, and the dimensional factor on the right-hand side is negligible. Therefore,
w(X ) log5/4 d ≪ σ(X ) implies
E kX k / 2σ(X ).
In the presence of strong isotropy, the spectral norm of a matrix Gaussian series is comparable with the standard
deviation σ(X ) whenever the alignment parameter w(X ) is relatively small!
In particular, we can apply this result to the GOE matrix X goe because of Proposition 3.8. The calculation (3.2)
of the standard deviation and alignment parameters ensures that E kX goe k / 2. As we observed in (2.9), this bound
is sharp. For this example, we can even take p ≈ d 1/5 , which leads to very good probability bounds via Markov’s
inequality. Furthermore, a more detailed version of Theorem 3.11, appearing in Section 9, is precise enough to
show that the semicircle law is the limiting spectral distribution of the GOE.
On the other hand, the dependence on the exponent p in Theorem 3.11 is suboptimal. This point is evident
when we consider the diagonal Gaussian matrix X diag (d). Indeed, Theorem 3.11 only implies the bound
E kX diag k ≤ const · log5/4 d.
As we observed in (2.9), the power on the logarithm should be one-half.
8
J. A. TROPP
3.7. Discussion. This paper opens a new chapter in the theory of matrix concentration and noncommutative moment inequalities. Our main technical contribution is to demonstrate that the matrix Khintchine inequality, Proposition 2.3, is not the last word on the behavior of a matrix Gaussian series. Indeed, we have shown that the matrix
variance does not contain sufficient information to determine the expected norm of a matrix Gaussian series. We
have also identified another quantity, the matrix alignment parameter, that allows us to obtain better bounds for
every matrix Gaussian series. Furthermore, in the presence of more extensive distributional information, it is even
possible to obtain numerically sharp bounds for the norm of certain matrix Gaussian series.
There are a number of ways to extend the ideas and results in this paper:
Higher-Order Alignment: If we consider alignment parameters involving 2k coefficient matrices, it is possible to
improve the term p 1/4 σ2p in Theorem 3.5 to p 1/(2k)σ2p . See Section 8.1 for some additional details.
Other Matrix Series: We can use exchangeable pairs techniques [MJC+ 14] to study matrix series of the form X :=
Pn
i=1 ξi H i where {ξi } is an independent family of scalar random variables. This approach is potentially
quite interesting when the ξi are Bernoulli (that is, 0–1) random variables.
Independent Sums: We can use conditioning and symmetrization, as in (2.2), to apply Theorem 3.5 to a sum of
independent random matrices. See [CGT12, App.] for an example of this type of argument.
Rectangular Matrices: The techniques here also give results for rectangular random matrices by way of the Hermitian dilation [Tro15, Sec. 2.1.16]. In this setting, a different notion of strong isotropy becomes relevant;
see Section 9.1.
We have not elaborated on these ideas because there is also evidence that alignment parameters will not lead to
final results on matrix concentration.
3.8. Related Work. There are very few techniques in the literature on random matrices that satisfy all three of
our three requirements: flexibility, ease of use, and power. In particular, for many practical applications, it is
important to be able to work with an arbitrary sum of independent random matrices. We have chosen to study
matrix Gaussian series because they are the simplest instance of this model, and they may lead to further insights
about the general problem.
Most classical work in random matrix theory concerns very special classes of random matrices; the books [BS10,
Tao12] provide an overview of some of the main lines of research in this field. There are some specific subareas
of random matrix theory that address more general models. The monograph [NS06] gives an introduction to free
probability. The book chapter [Ver12] describes a collection of methods from Banach space geometry. The monograph [Tro15] covers the theory of matrix concentration. The last three works have a wide scope of applicability,
but none of them provides the ultimate description of the behavior of a sum of independent random matrices.
There is one specific strand of research that we would like to draw out because it is very close in spirit to this
paper. Recently, Bandeira & van Handel [BV14] and van Handel [vH15] have studied the behavior of a real symmetric Gaussian matrix whose entries are independent and centered but have inhomogeneous variances (the
independent-entry model). A d × d random matrix from this class can be written as
X indep :=
Xd
i,j =1 a i j γi j
· (Ei j + E j i ) for ai j ∈ R.
As usual, {γi j } is an independent family of standard normal random variables, and we assume that ai j = a j i without loss of generality.
To situate this model in the context of our work, observe that matrix Gaussian series are significantly more
general than the independent-entry model. The strongly isotropic model is incomparable with the independententry model. To see why, recall that strongly isotropic matrices can have dependent entries. At the same time,
p
E X indep is diagonal for each integer p ≥ 0, but it need not be a scalar matrix.
For the independent-entry model, Bandeira & van Handel [BV14] established the following (sharp) bound:
q
E kX indep k / 2σ(X indep ) + const · maxi j |ai j | · log d
(3.5)
The maximum entry maxi j |ai j | plays the same role in this formula as the matrix alignment parameter plays in this
paper. The paper [BV14] leans heavily on the independence assumption, so it is not clear whether the ideas extend
to a more general setting.
SECOND-ORDER MATRIX CONCENTRATION INEQUALITIES
9
To compare the result (3.5) with the work here, we can compute the matrix alignment parameter for the independententry model using a difficult extension of the calculation in Section 4.2. This effort yields
³
´1/4
X
w(X indep ) ≈ maxi j |ai j |4
.
We see that the matrix alignment parameter is somewhat larger than the maximum entry maxi j |ai j |. Thus, for the
independent model, Theorem 3.5 gives us a better result than the classical Khintchine inequality, Proposition 2.3,
but it is somewhat weaker than (3.5). Theorem 3.11 would give a result close to the bound (3.5), but it does not
always apply because the independent-entry model need not be strongly isotropic.
The independent-entry model is not adequate to reach results with the same power and scope as the current
generation of matrix concentration bounds [Tro15]. Nevertheless, the estimate (3.5) strongly suggests that there
are better ways of summarizing the interactions of the coefficients in an Hermitian matrix Gaussian series X :=
Pn
i=1 γi H i than the alignment parameter w(X ). One possibility is the weak variance parameter:
¡Xn
¢1/2
|〈u, Hi v〉|2
σ⋆ (X ) := sup
.
i=1
kuk=kv k=1
For the independent-entry model, this quantity reduces to const · maxi j |ai j |. The idea of considering σ⋆ (X ) is
motivated by the discussion in [Tro12, Sec. 4], as well as the work in [BV14, vH15]. Unfortunately, at this stage,
it is not clear whether there are any parameters that allow us to obtain a simple description of the behavior of a
Gaussian matrix in the absence of burdensome independence or isotropy assumptions. This is a frontier for future
work.
4. C OMPUTATION OF THE M ATRIX A LIGNMENT PARAMETERS
In this section, we show how to compute the matrix alignment parameter for the two random matrices in the
introduction, the diagonal Gaussian matrix and the GOE matrix. Afterward, we show by example that neither
Theorem 3.5 nor Theorem 3.11 can hold if we remove the unitary factors from the matrix alignment parameter.
4.1. A Diagonal Gaussian Matrix. The diagonal Gaussian matrix takes the form
Xd
X diag := i=1 γi Eii .
2
The matrix variance Var(X diag ) = E X diag
= I. It follows that the matrix standard deviation parameters, defined
in (2.3), satisfy
°
°
σp (X diag ) = °Var(X diag )1/2 ° = d 1/p for 1 ≤ p ≤ ∞.
p
We will show that the matrix alignment parameters, defined in (3.1), satisfy
w p (X diag ) = d 1/p
for 4 ≤ p ≤ ∞.
Thus, for this example, the second-order matrix Khintchine inequalities, Theorem 3.5 and Theorem 3.11, do not
improve over the matrix Khintchine inequality, Proposition 2.3. This outcome is natural, given that the classical
result is essentially optimal in this case.
Let us evaluate the matrix alignment parameter. For a triple (Q, S,U ) of unitary matrices, form the sum
Xd
Xd
W (Q, S,U ) := i,j =1 Eii QE j j SEii U E j j = i,j =1 q i j s j i ui j · Ei j = Q ⊙ S t ⊙U .
We have written ⊙ for the Schur (i.e., componentwise) product, and t is the transpose operation. When Q = S =
U = I, the sum collapses: W (I, I, I) = I. Therefore,
°
°
w p (X diag ) = max °|W (Q, S,U )|1/4 °p ≥ kIkp = d 1/p for p ≥ 1.
Q,S,U
But Proposition 3.2 shows that
w p (X diag ) ≤ σp (X diag ) = d 1/p
for each p ≥ 4.
Therefore, w p (X diag ) = σp (X diag ) = d 1/p for p ≥ 4. The result for p = ∞ follows when we take limits.
Remark 4.1 (Commutativity). A similar calculation is valid whenever the family {Hi } of coefficient matrices in the
matrix Gaussian series (2.1) commutes.
10
J. A. TROPP
4.2. A GOE Matrix. The GOE matrix takes the form
1 Xd
X goe := p
i,j =1 γi j (Ei j + E j i ).
2d
An easy calculation shows that the matrix variance satisfies
1 Xd
2
Var(X goe ) = E X goe
=
(Ei j + E j i )2 = (1 + d −1 ) · I.
2d i,j =1
Therefore, the matrix standard deviation parameters, defined in (2.3), equal
p
°
°
σp (X goe ) = °Var(X goe )1/2 °p = 1 + d −1 · d 1/p for 1 ≤ p ≤ ∞.
We will demonstrate that the matrix alignment parameters, defined in (3.1), satisfy
¡
¢1/4 1/p
w p (X goe ) ≤ d −1 + 3d −2
·d
for 4 ≤ p ≤ ∞.
When d is large, the matrix alignment parameters are much smaller than the matrix standard deviation parameters.
As a consequence, the second-order matrix Khintchine inequalities deliver a substantial gain over the classical
matrix Khintchine inequality.
Let us compute the matrix alignment parameter. For a triple (Q, S,U ) of unitary matrices, introduce the (unnormalized) sum
Xd
W (Q, S,U ) := i 1 ,i 2 ,j 1 ,j 2 =1 (Ei 1 i 2 + Ei 2 i 1 )Q(E j 1 j 2 + E j 2 j 1 )S(Ei 1 i 2 + Ei 2 i 1 )U (E j 1 j 2 + E j 2 j 1 ).
It is not hard to evaluate this sum if we take care. First, distribute terms:
h¡
Xd
¢
W (Q, S,U ) = i 1 ,i 2 ,j 1 ,j 2 =1 q i 2 j 1 s j 2 i 1 ui 2 j 2 + q i 2 j 1 s j 2 i 2 ui 1 j 2 + q i 2 j 2 s j 1 i 1 ui 2 j 2 + q i 2 j 2 s j 1 i 2 ui 1 j 2 · Ei 1 j 1
¡
¢
+ q i 2 j 1 s j 2 i 1 ui 2 j 1 + q i 2 j 1 s j 2 i 2 ui 1 j 1 + q i 2 j 2 s j 1 i 1 ui 2 j 1 + q i 2 j 2 s j 1 i 2 ui 1 j 1 · Ei 1 j 2
¡
¢
+ q i 1 j 1 s j 2 i 1 ui 2 j 2 + q i 1 j 1 s j 2 i 2 ui 1 j 2 + q i 1 j 2 s j 1 i 1 ui 2 j 2 + q i 1 j 2 s j 1 i 2 ui 1 j 2 · Ei 2 j 1
i
¡
¢
+ q i 1 j 1 s j 2 i 1 ui 2 j 1 + q i 1 j 1 s j 2 i 2 ui 1 j 1 + q i 1 j 2 s j 1 i 1 ui 2 j 1 + q i 1 j 2 s j 1 i 2 ui 1 j 1 · Ei 2 j 2 .
In each line, we can sum through the two free indices to identify four matrix products. For example, in the first line,
we can sum on i 2 and j 2 . This step yields
Xd
¡
¢
W (Q, S,U ) = i 1 ,j 1 =1 S tU tQ +U SQ + tr(Q tU ) · S t +U Q t S t i 1 j 1 · Ei 1 j 1
Xd
¡
¢
+ i 1 ,j 2 =1 tr(Q tU ) · S t +U Q t S t + S tU tQ +U SQ i 1 j 2 · Ei 1 j 2
Xd
¡
¢
+ i 2 ,j 1 =1 U SQ + S tU tQ +U Q t S t + tr(QU ) · S t i 2 j 1 · Ei 2 j 1
i
Xd
¡
¢
+ i 2 ,j 2 =1 U Q t S t + tr(QU ) · S t +U SQ + S tU tQ i 2 j 2 · Ei 2 j 2 .
Sum through the remaining indices to reach
¡
¢
W (Q, S,U ) = S tU tQ +U SQ + tr(Q tU ) · S t +U Q t S t
¡
¢
+ tr(Q tU ) · S t +U Q t S t + S tU tQ +U SQ
¡
¢
+ U SQ + S tU tQ +U Q t S t + tr(QU ) · S t
¡
¢
+ U Q t S t + tr(QU ) · S t +U SQ + S tU t Q .
Twelve of the sixteen terms are unitary matrices, and the remaining four are scaled unitary matrices. Furthermore,
each trace is bounded in magnitude by d, the worst case being Q = U = I. Applying the definition of the Schatten
norm, the triangle inequality, and unitary invariance, we find that
°
°
¢
¡
°|W (Q, S,U )|1/4 ° = kW (Q, S,U )k1/4 ≤ (4d + 12) · kIkp/4 1/4 = (4d + 12)1/4 · d 1/p for p ≥ 4.
p/4
p
To compute w p (X goe ), we must reintroduce the scaling (2d)−1/2 , which gives the advertised result:
¡
¢1/4 1/p
w p (X goe ) ≤ (2d)−1/2 · (4d + 12)1/4 · d 1/p = d −1 + 3d −2
·d .
To obtain the bound for p = ∞, we simply take limits.
SECOND-ORDER MATRIX CONCENTRATION INEQUALITIES
11
P
4.3. The Unitaries are Necessary. Suppose that X := ni=1 γi Hi is an Hermitian matrix Gaussian series with dimension d, and let σ(X ) be the matrix standard deviation (2.3). Consider the alternative alignment parameter
°1/4
°X
°
° n
δ(X ) := ° i,j =1 Hi H j Hi H j ° .
This quantity is suggested by the discussion in Section 3.1. Consider a general estimate of the form
E kX k
≤
f (d) · σ(X ) + g (d) · δ(X ).
(4.1)
p
We will demonstrate that, for every choice of the function g , there is a lower bound f (d) ≥ const · log d. From
this claim, we deduce that it is impossible to improve over the classical Khintchine inequality by using the secondorder quantity δ(X ). Therefore, the unitary matrices in the alignment parameter w(X ) play a critical role. Most of
this argument was developed by Afonso Bandeira; we are grateful to him for allowing us to include it.
Introduce the Pauli spin matrices
¸
¸
·
¸
·
·
0 i
0 1
1 0
.
H3 :=
H2 :=
H1 :=
−i 0
1 0
0 −1
These matrices are Hermitian and unitary, so Hi2 = I for i = 1, 2, 3. Furthermore, they satisfy the relations (Hi H j )2 =
p
p
−I when i 6= j . Next, define H0 := α I, where α := 2 3 − 3. Calculate that
X3
X4
X3
X3
X3
4
i,j =1 H i H j H i H j
i,j =0 H i H j H i H j =
i=0 H i +
j =1 H 0 H j H 0 H j +
i=1 H i H 0 H i H 0 +
2
2
i6= j
= (α + 3) I + 6αI − 6I = (α + 6α − 3) I = 0.
Indeed, α is a positive root of the quadratic.
Consider the two-dimensional Gaussian series Y generated by the matrices H0 , . . . , H3 :
X3
Y := i=0 γi Hi .
As usual, {γi } is an independent family of standard normal variables. For the series Y , we have already shown that
the alternative alignment parameter δ(Y ) = 0. Let us compute the variance and standard deviation:
X3
p
Var(Y ) = i=0 Hi2 = (α + 3) I = 2 3 I and σ(Y ) = kVar(Y )k1/2 = 121/4 .
Expanding the random matrix Y in coordinates, we also find that
¸
·p
α γ0 + γ1
γ2 + iγ3
p
.
Y=
αγ0 − γ1
γ2 − iγ3
p
Therefore, the top-left entry (Y )11 is a centered normal random variable with variance 1 + α = 2( 3 − 1).
To obtain the counterexample to the bound (4.1), fix an integer d ≥ 1. Let Y 1 , . . . , Y d be independent copies of
the two-dimensional Gaussian series Y , and construct the 2d-dimensional matrix Gaussian series
Xd
Xd X3
X spin := Y 1 ⊕ · · · ⊕ Y d = j =1 E j j ⊗ Y j ∼ j =1 i=0 γi j (E j j ⊗ Hi ).
We have written ⊕ for direct sum and ⊗ for the Kronecker product; the matrices E j j are the diagonal units with
dimension d × d; and {γi j } is an independent family of standard normal variables.
Extending the calculations above, we find that σ(X spin ) = 121/4 and δ(X spin ) = 0. Meanwhile, the norm of X spin
is bounded below by the absolute value of each of its diagonal entries. In particular,
¢1/2 q
¡ p
· log d.
E kX spin k ≥ E max j |(Y j )11 | ≥ const · 2( 3 − 1)
We
p have used the fact that the expected maximum of d independent standard normal variables is proportional to
log d. Assuming that (4.1) is valid, we can sequence these estimates to obtain
q
const · log d ≤ f (d) · σ(X spin ) + g (d) · δ(X spin ) = 121/4 · f (d).
p
Therefore, the function f (d) must grow at least as fast as log d . We conclude that a bound of the form (4.1) can
never improve over the classical matrix Khintchine inequality.
12
J. A. TROPP
5. N OTATION & B ACKGROUND
Before we enter into the body of the paper, let us set some additional notation and state a few background
results. First, Md denotes the complex linear space of d ×d matrices with complex entries. We write Hd for the reallinear subspace of Md that consists of Hermitian matrices. The symbol ∗ represents conjugate transposition. We
write 0 for the zero matrix and I for the identity. The matrix Ei j has a one in the (i , j ) position and zeros elsewhere.
The dimensions of these matrices are typically determined by context.
For an Hermitian matrix A, we define the integer powers A p for p = 0, 1, 2, 3, . . . in the usual way by iterated multiplication. For a positive-semidefinite matrix P, we can also define complex powers P z by raising each eigenvalue
of P to the power z while maintaining the eigenvectors. In particular, P 1/2 is the unique positive-semidefinite
square root of P. The matrix absolute value is defined for a general matrix B by the rule |B | := (B ∗ B )1/2 . Note that
|P| = P when P is positive semidefinite.
The trace and normalized trace of a matrix are given by
tr B :=
Xd
i=1 b ii
and
¯ B :=
tr
1 Xd
b ii
d i=1
for B ∈ Md .
We use the convention that a power binds before the trace to avoid unnecessary parentheses; powers also bind
before expectation. The Schatten p-norm is defined for an arbitrary matrix B via the rule
¢1/p
¡
kB kp := tr |B |p
for p ≥ 1.
The Schatten ∞-norm k·k∞ coincides with the spectral norm k·k. This work uses both trace powers and Schatten
norms, depending on which one is conceptually clearer. We require some Hölder inequalities involving the trace
and the Schatten norms. For matrices A, B ∈ Md and ̺ ≥ 1,
¢1/̺ ¡
¡
′ ¢1/̺′
|tr(AB )| ≤ tr |A|̺
where ̺′ := ̺/(̺ − 1).
(5.1)
· tr |B |̺
Furthermore,
kA ∗ B k2̺ ≤ kA ∗ Ak̺ · kB ∗ B k̺ .
(5.2)
These results are drawn from [Bha97, Chap. IV].
6. T HE T RACE M OMENTS OF A M ATRIX G AUSSIAN S ERIES
For each major result in this paper, the starting point is a formula for the trace moments of a matrix Gaussian
series.
P
Lemma 6.1 (Trace Moment Identity). Let X := ni=1 γi Hi be an Hermitian matrix Gaussian series, as in (2.1). For
each integer p ≥ 1, we have the identity
X2p−2 Xn
£
¤
q
2p−2−q
.
(6.1)
E tr X 2p = q=0
i=1 E tr H i X H i X
The easy proof of Lemma 6.1 appears in the next two subsections.
Integration by parts is not foreign in the study of Gaussian random matrices; for example, see [AGZ10, Sec. 2.4.1]
or [Kem13, Sec. 9]. The exchangeable pairs method for establishing matrix concentration is also based on an
elementary, but conceptually challenging, analog of integration by parts [MJC+ 14, Lem. 2.4]. Aside from these
works, we are not aware of any application of related techniques to prove results on matrix concentration.
6.1. Preliminaries. To obtain Lemma 6.1, the main auxiliary tool is the classical integration by parts formula for a
function of a standard normal vector [NP12, Lem. 1.1.1]. In the form required here, the result can be derived with
basic calculus.
Fact 6.2 (Gaussian Integration by Parts). Let γ ∈ Rn be a vector with independent standard normal entries, and let
f : Rn → R be a function whose derivative is absolutely integrable with respect to the standard normal measure. Then
Xn
£
¤ Xn
£
¤
i=1 E γi · f (γ) =
i=1 E (∂i f )(γ) .
The symbol ∂i denotes differentiation with respect to the i th coordinate.
We also use a well-known formula for the derivative of a matrix power [Bha97, Sec. X.4].
SECOND-ORDER MATRIX CONCENTRATION INEQUALITIES
Fact 6.3 (Derivative of a Matrix Power). Let A : R → Md be a differentiable function. For each integer ̺ ≥ 1,
¢ X̺−1
d
d ¡
A(u) · A(u)̺−1−k .
A(u)̺ = k=0 A(u)̺ ·
du
du
In particular,
·
¸
d
d
tr A(u)̺ = ̺ · tr A(u)̺−1 ·
A(u) .
du
du
13
(6.2)
The symbol · refers to ordinary matrix multiplication.
6.2. Proof of Lemma 6.1. Let us treat the random matrix X as a matrix-valued function of the standard normal
vector γ := (γ1 , . . . , γn ). That is,
Xn
X = X (γ) = i=1 γi Hi .
Write X = X · X 2p−1 and distribute the sum in the first factor:
£
£
¤¤
¢ 2p−1 ¤ Xn
£¡Xn
= i=1 E γi · tr Hi X 2p−1
E tr X 2p = E tr
i=1 γi H i X
The Gaussian integration by parts formula, Fact 6.2, implies that
Xn
£
¡
¢¤
E tr X 2p = i=1 E tr Hi · ∂i X 2p−1 .
Since ∂i X = Hi , the derivative formula (6.2) yields
i X
h
Xn
X2p−2
£
¤
2p−2 Xn
q
2p−2−q
.
E tr X 2p = i=1 tr Hi · q=0 X q Hi X 2p−2−q = q=0
i=1 tr H i X H i X
This completes the proof of the formula (6.1).
7. A S HORT P ROOF OF THE M ATRIX K HINTCHINE I NEQUALITY
Historically, proofs of the matrix Khintchine inequality have been rather complicated, but the result is actually
an immediate consequence of Lemma 6.1. We will present this argument in detail because it has not appeared in
the literature. Furthermore, the approach serves as a template for the more sophisticated theorems that are the
main contributions of this paper. Let us restate Proposition 2.3 in the form that we will establish it.
P
Proposition 7.1 (Matrix Khintchine). Let X := ni=1 γi Hi be an Hermitian matrix Gaussian series, as in (2.1). Define
the matrix variance and standard deviation parameters
Xn
¡
¢1/(2q)
V := Var(X ) = i=1 Hi2 and σ2q := trV q
for each q ≥ 1.
(7.1)
Then, for each integer p ≥ 1,
¡
E tr X 2p
¢1/(2p)
≤
p
2p − 1 · σ2p .
The short proof of Proposition 7.1 appears in the next two sections. The approach parallels the exchangeable pairs
method that has been used to establish the matrix Khintchine inequality for Rademacher series [MJC+ 14, Cor. 7.3].
Here, we replace exchangeable pairs with the conceptually simpler argument based on Gaussian integration by
parts. To reach the statement of Proposition 2.3, we simply rewrite the trace in terms of a Schatten norm.
Remark 7.2 (Noninteger Moments). Our proof of Proposition 7.1 can be adapted to obtain moment bounds for all
p ≥ 2. See [MJC+ 14, Cor. 7.3] for a closely related argument.
7.1. Preliminaries. The main idea in the proof is to simplify the trace moment identity (6.1) with an elementary
matrix inequality. Anticipating subsequent arguments, we state the inequality in greater generality than we need
right now.
Proposition 7.3. Suppose that H and A are Hermitian matrices of the same size. Let q and r be integers that satisfy
0 ≤ q ≤ r . For each real number s in the range 0 ≤ s ≤ min{q, r − q},
¤
£
¤
£
tr H A q H A r −q ≤ tr H |A|s H |A|r −s .
The proof of Proposition 7.3 depends on a numerical fact. For nonnegative numbers α and β, the function
θ 7→ αθ β1−θ + α1−θ βθ is convex on the interval [0, 1], and it achieves its minimum at θ = 12 . Therefore,
©
ª
′
′
′
′
αθ β1−θ + α1−θ βθ ≤ αθ β1−θ + α1−θ βθ when 0 ≤ θ′ ≤ min θ, 1 − θ .
(7.2)
We need to lift this scalar inequality to matrices.
14
J. A. TROPP
P
Proof. Without loss of generality, we may change coordinates so that A is diagonal: A = i ai Eii . Expanding both
copies of A,
¡ q r −q
£
¤
£
¤ X
£
¤ X
r −q q ¢
q r −q
tr H A q H A r −q = i,j ai a j · tr H Eii H E j j = i,j 21 ai a j + ai a j · tr H Eii H E j j .
After we take absolute values, the inequality (7.2) implies that
q
r −q
r −q
q
a j ≤ |ai |s |a j |r −s + |ai |r −s |a j |s .
£
¤
The remaining trace is nonnegative: tr H Eii H E j j = |hi j |2 , where hi j are the components of the matrix H . As a
consequence,
£
¤ X
¡
¢ £
¤
¤
£
tr H A q H A r −q ≤ i,j 12 |ai |s |a j |r −s + |ai |r −s |a j |s · tr H Eii H E j j = tr H |A|s H |A|r −s .
ai a j
+ ai
To reach the last identity, we reversed our steps to reassemble the sum into a trace.
7.2. Proof of the Matrix Khintchine Inequality. We may now establish Proposition 7.1. Let us introduce notation
for the quantity of interest:
E 2p := E tr X 2p .
Use the integration by parts result, Lemma 6.1, to rewrite the trace moment:
X2p−2 Xn
£
¤
q
2p−2−q
.
E 2p = q=0
i=1 E tr H i X H i X
For each choice of q, apply the matrix inequality from Proposition 7.3 with r = 2p − 2 and s = 0 to reach
Xn
£
¤
£
¤
E 2p ≤ (2p − 1) i=1 E tr Hi2 X 2(p−1) = (2p − 1) · E tr V X 2(p−1)
We have identified the matrix variance V defined in (7.1).
Next, let us identify a copy of E on the right-hand side and solve the resulting algebraic inequality. To that end,
invoke Hölder’s inequality (5.1) for the trace with ̺ = p and ̺′ = p/(p − 1):
¡
¢1/p ¡
¢(p−1)/p
E 2p ≤ (2p − 1) · trV p
· E tr X 2p
¡
¢(p−1)/p
≤ (2p − 1) · σ22p · E tr X 2p
= (2p − 1) · σ22p · E 2(p−1) .
We have identified the quantity σ2p from (7.1). The second inequality is Lyapunov’s. Since the unknown E is
nonnegative, we can solve the polynomial inequality to reach
p
E ≤ 2p − 1 · σ2p .
This is the required result.
8. A S ECOND -O RDER M ATRIX K HINTCHINE I NEQUALITY
In this Section, we prove Theorem 3.5, the second-order matrix Khintchine inequality. Let us restate the result
in the form that we will establish it.
P
Theorem 8.1 (Second-Order Matrix Khintchine). Let X = ni=1 γi Hi be an Hermitian matrix Gaussian series, as
in (2.1). Define the matrix variance and standard deviation parameter
Xn
¡
¢1/(2p)
V := Var(X ) = i=1 Hi2 and σ2p := trV p
for p ≥ 1.
(8.1)
Define the matrix alignment parameter
µ ¯
¯p/2 ¶1/(2p)
¯
¯Xn
w 2p := max tr ¯ i,j =1 Hi Q 1 H j Q 2 Hi Q 3 H j ¯
Qℓ
for p ≥ 1
where the maximum ranges over a triple (Q 1 ,Q 2 ,Q 3 ) of unitary matrices. Then, for each integer p ≥ 3,
p
p
¡
¢1/(2p)
E tr X 2p
≤ 3 4 2p − 5 · σ2p + 2p − 4 · w 2p .
(8.2)
(8.3)
The proof of Theorem 8.1 will occupy us for the rest of the section. To reach the statement in the introduction, we
rewrite traces in terms of Schatten norms. We also provide the proof of Proposition 3.2 in Section 8.9.
SECOND-ORDER MATRIX CONCENTRATION INEQUALITIES
15
8.1. Discussion. Before we establish Theorem 8.1, let us spend a moment to discuss the proof of this result. Theorem 8.1 is based on the same pattern of argument as the matrix Khintchine inequality, Proposition 7.1. This time,
we apply Proposition 7.3 more surgically to control the terms in the trace moment identity from Lemma 6.1. The
most significant new observation is that we can use complex interpolation to reorganize the products of matrices
that arise during the calculation.
We can refine this argument in several ways. First, if we apply complex interpolation with more care, it is possible to define the matrix alignment parameter (8.2) as a maximum over the set
©
ª
Q 1 ,Q 2 ,Q 3 are commuting unitaries and Q ℓ = I for some ℓ .
Given that commuting matrices are simultaneously diagonalizable, this improvement might make it easier to
bound the matrix alignment parameters.
Second, it is quite clear from the proof that we can proceed beyond the second-order terms. For example, for
an integer p ≥ 3, we can obtain results in terms of the third-order quantities
µ ¯
¯p/3 ¶1/(2p)
¯Xn
¯
w 2p,1 := max tr ¯ i,j ,k=1 Hi Q 1 H j Q 2 Hk Q 3 Hi Q 4 H j Q 5 Hk ¯
Qℓ
µ ¯
¯p/3 ¶1/(2p)
¯Xn
¯
w 2p,2 := max tr ¯ i,j ,k=1 Hi Q 1 H j Q 2 Hk Q 3 Hi Q 4 Hk Q 5 H j ¯
.
Qℓ
The ordering of indices is (i , j , k, i , j , k) and (i , j , k, i , k, j ), respectively. This refinement allows us reduce the order
of coefficient on the standard deviation term σ2p in (8.3) to p 1/6 . Unfortunately, we must also compute both
alignment parameters w 2p,1 and w 2p,2 , instead of just w 2p . This observation shows why it is unproductive to press
forward with this approach. Indeed, the number of orderings of indices grows super-exponentially as we consider
longer products, which is an awful prospect for applications.
8.2. Preliminaries. In the proof of Theorem 8.1, we will use two interpolation results to reorganize products of
matrices. The first one is a type of matrix Hölder inequality [LP86, Cor. 1]. Here is a version of the result specialized
to our setting.
Fact 8.2 (Lust-Piquard). Consider a finite sequence (A 1 , . . . , A n ) of Hermitian matrices with the same dimension,
and let B be a positive-semidefinite matrix of the same dimension. For each number ̺ ≥ 2,
³ ¡X
¢1/̺
¢ ´1/̺ ¡
¢̺/2 ´2/̺ ³ ¡Xn
n
2 ̺
· tr B ̺
.
tr
A
B
A
≤
tr
A
i
i
i=1
i=1 i
See [PX97, Lem. 1.1] for a proof based on the Hadamard Three-Lines Theorem [Gar07, Prop. 9.1.1].
The second result is a more complicated interpolation for a multilinear function whose arguments are powers
of random matrices.
Proposition 8.3 (Multilinear Interpolation). Suppose that F : (Md )k → C is a multilinear function. Fix nonnegaP
tive integers α1 , . . . , αk with ki=1 αi = α. Let Y i ∈ Hd be random matrices, not necessarily independent, for which
E kY i kα < ∞. Then
¯ ¡
¯ ¡ α1
¢¯
¢¯
¯E F Y , . . . , Y αk ¯ ≤ max E max ¯F Q 1 , . . . , Q i−1 , Q i Y α , Q i+1 , . . . , Q k ¯ .
1
k
i=1,...,k
Qℓ
i
In this expression, each Q ℓ is a (random) unitary matrix that commutes with Y ℓ .
As with Fact 8.2, the proof of Proposition 8.3 depends on the Hadamard Three-Lines Theorem [Gar07, Prop. 9.1.1].
The argument is standard but somewhat involved, so we postpone the details to Appendix A.
8.3. The Overture. Let us commence with the proof of Theorem 8.1. The initial steps are similar with the argument
that leads to the matrix Khintchine inequality, Proposition 7.1. Introduce notation for the quantity of interest:
X2p−2 Xn
£
¤
q
2p−2−q
.
(8.4)
E 2p := E tr X 2p = q=0
i=1 E tr H i X H i X
The identity follows from the integration by parts result, Lemma 6.1.
This time,
© we make finerªestimates for the summands in (8.4). Apply Proposition 7.3 with s = 0 to the terms
where q ∈ 0, 1, 2p − 3, 2p − 2 . For the remaining 2p − 5 values of the exponent q, apply Proposition 7.3 with s = 1.
We reach the bound
Xn
Xn
£
¤
£
¤
E 2p ≤ 4 i=1 E tr Hi2 X 2p−2 + (2p − 5) i=1 E tr Hi X 2 Hi X 2p−4 .
(8.5)
We can take advantage of the fact that the Hi are interleaved with the powers X r of the random matrix in the
second term.
16
J. A. TROPP
8.4. The First Term. To treat the first term on the right-hand side of (8.5), simply repeat the arguments from Section 7.2 to obtain a bound in terms of the quantity E . We have
Xn
i=1
£
¤
£
¤ ¡
¢1/p ¡
¢2(p−1)
E tr Hi2 X 2p−2 = E tr V X 2p−2 ≤ trV p
· E tr X 2p
= σ22p · E 2(p−1) .
(8.6)
The quantities V and σ2p are defined in (8.1), and we have identified a copy of E .
8.5. Integration by Parts, Again. To continue, we want to break down the matrix X 2 that appears in the second
term on the right-hand side of (8.5). To do so, we perform another Gaussian integration by parts. Write X 2 =
Pn
j =1 γ j H j X , and invoke Fact 6.2 to obtain
Xn
i=1
£
¤ Xn
£
£
¤¤
E tr Hi X 2 Hi X 2p−4 = i,j =1 E γ j · tr Hi H j X Hi X 2p−4
³X
´i
h
Xn
£
¤ Xn
2p−5 r
2p−5−r
.
= i,j =1 E tr Hi H 2j Hi X 2p−4 + i,j =1 E tr Hi H j X Hi
r =0 X H j X
(8.7)
This result follows from the product rule and the formula (6.2) for the derivative of a power. We will bound the first
term on the right-hand side of (8.7) in terms of the standard deviation parameter σ2p , and the second term will
lead to the matrix alignment parameter w 2p .
8.6. Finding the Standard Deviation Parameter. Let us address the first term on the right-hand side of (8.7). First,
draw the sum back into the trace and identify the matrix variance V , defined in (8.1):
Xn
£
¤
£¡Xn
¢ 2(p−2) ¤
2
2p−4
= E tr
.
i,j =1 E tr H i H j H i X
i=1 H i V H i X
To isolate the random matrix X , apply Hölder’s inequality (5.1) with exponents ̺ = p/2 and ̺′ = p/(p − 2), and
follow up with Lyapunov’s inequality. Thus,
E tr
¢ 2(p−2) ¤ ³ ¡Xn
¢(p−2)/p
¢p/2 ´2/p ¡
H
V
H
X
≤
tr
H
V
H
· E tr X 2p
.
i
i
i
i
i=1
i=1
£¡Xn
The Lust-Piquard inequality, Fact 8.2, with ̺ = p implies that
³ ¡X
¢1/p ¡
¢2/p
¢ ´1/p ¡
¢p/2 ´2/p ³ ¡Xn
n
2 p
· trV p
= trV p
= σ42p .
≤ tr
tr
i=1 H i
i=1 H i V H i
Once again, we identified V and σ2p from (8.1). Combine the last three displays to arrive at
Xn
i,j =1 E tr
£
¤
¡
¢(p−2)/p
Hi H 2j Hi X 2p−4 ≤ σ42p · E tr X 2p
= σ42p · E 2(p−2) .
(8.8)
We have identified another copy of E .
8.7. Finding the Matrix Alignment Parameter. It remains to study the second term on the right-hand side of (8.7).
Rearranging the sums, we write this object as
X2p−5 Xn
r =0
i,j =1 E tr
£
¤
Hi H j X Hi X r H j X 2p−5−r .
We can apply the interpolation result, Proposition 8.3, to consolidate the powers of the random matrix X . Consider
the multilinear function
Xn
£
¤
F (A 1 , A 2 , A 3 ) := i,j =1 tr Hi H j A 1 Hi A 2 H j A 3 .
Since X is a matrix Gaussian series, it has moments of all orders. Therefore, for each index r ,
¯ £
¤¯
¯E tr Hi H j X Hi X r H j X 2p−5−r ¯
¾
½
¯ ¡
¯ ¡
¯ ¡
¢¯
¢¯
¢¯
≤ max E max ¯F Q 1 X 2p−4 ,Q 2 ,Q 3 ¯ , E max ¯F Q 1 ,Q 2 X 2p−4 ,Q 3 ¯ , E max ¯F Q 1 ,Q 2 ,Q 3 X 2p−4 ¯ .
Qℓ
Qℓ
Qℓ
SECOND-ORDER MATRIX CONCENTRATION INEQUALITIES
17
All three terms in the maximum admit the same bound, so we may as well consider the third one:
¯X
¯
¯
£
¤¯¯
¯ n
E max ¯F (Q 1 ,Q 2 ,Q 3 X 2p−4 )¯ = E max ¯ i,j =1 tr Hi H j Q 1 Hi Q 2 H j Q 3 X 2(p−2) ¯
Qℓ
Qℓ
¯ h³X
i¯
´
¯
n
2(p−2) ¯
X
Q
H
H
Q
H
Q
H
= E max ¯tr
¯
3
1
2
i
j
i
j
i,j =1
Qℓ
#
"µ
¯X
¯p/2 ¶2/p ¡
¢
¯
¯ n
2p (p−2)/p
· tr X
≤ E max tr ¯ i,j =1 Hi H j Q 1 Hi Q 2 H j ¯
Qℓ
µ ¯
¯p/2 ¶2/p ¡
¢(p−2)/p
¯
¯Xn
≤ max tr ¯ i,j =1 Hi H j Q 1 Hi Q 2 H j ¯
· E tr X 2p
Qℓ
4
≤ w 2p
· E 2(p−2) .
The first step is the definition of F . To reach the second line, we use the fact that Q 3 commutes with X , then we
cycle the trace. The third line is Hölder’s inequality (5.1) with ̺ = p/2 and ̺′ = p/(p − 2), and we have used the
left unitary invariance of the matrix absolute value to delete Q 3 . Next, take the maximum over all unitary matrices,
and apply Lyapunov’s inequality to draw the expectation into the term involving X . Finally, identify the quantity E
4
and note that the maximum is bounded by the alignment parameter w 2p
, defined in (8.2). Similar calculations are
valid for the other two terms, whence
¯ £
¤¯
¯E tr Hi H j X Hi X r H j X 2p−5−r ¯ ≤ w 4 · E 2(p−2) .
2p
Since there are 2p − 4 possible choices of r , we determine that
X2p−5 Xn
¤
£
r
2p−5−r
4
· E 2(p−2) .
≤ (2p − 4) · w 2p
i,j =1 E tr H i H j X H i X H j X
r =0
(8.9)
The main part of the argument is finished.
8.8. Putting the Pieces Together. To conclude, we merge the bounds we have obtained and solve the resulting
inequality for the quantity E . Combine (8.5), (8.6), (8.7), (8.8), and (8.9) to reach
i
h
4
· E 2(p−2) .
E 2p ≤ 4σ22p · E 2(p−1) + (2p − 5) · σ42p + (2p − 4) · w 2p
Clearing factors of E , we reach the inequality
i
h
4
.
E 4 ≤ 4σ22p · E 2 + (2p − 5) · σ42p + (2p − 4) · w 2p
If α andp
β are nonnegative numbers, each nonnegative solution to the quadratic inequality t 2 ≤ αt +β must satisfy
t ≤ α + β. It follows that
i1/2
h
p
4
E 2 ≤ 4σ22p + 2p − 5 · σ42p + (2p − 4) · w 2p
.
Take the square root, and invoke subadditivity of the square root (twice) to reach
p
p
¢
¡
E ≤ 2 + 4 2p − 5 · σ2p + 4 (2p − 5)(2p − 4) · w 2p .
Finally, we simplify the numerical constants to arrive at (8.3).
8.9. Comparison of Standard Deviation and Alignment Parameters. Our last task in this section is to establish
Proposition 3.2, which states that the alignment parameter w 2p never exceeds the standard deviation σ2p . The
easiest way to obtain this result is to use block matrices and inequalities for the Schatten norm.
Fix an integer p ≥ 2, and fix a triple (Q 1 ,Q 2 ,Q 3 ) of unitary matrices. Consider the quantity
µ ¯
¯p/2 ¶2/p
¯Xn
¯
S := tr ¯ i,j =1 Hi Q 1 H j Q 2 Hi Q 3 H j ¯
.
To establish Proposition 3.2, it suffices to show that S ≤ σ42p . Using block matrices and converting the trace into a
Schatten norm, we can write
°
∗
°
°
°
..
..
°
°
.
.
°
°
° ∗
°
S = °Q 2 H j Q 1∗ Hi Hi Q 3 H j ° .
°
°
..
..
°
°
°
°
.
.
p/2
18
J. A. TROPP
The entries of the block column matrices are indexed by pairs (i , j ), arranged in lexicographic order. Invoke the
Cauchy–Schwarz inequality (5.2) for Schatten norms with ̺ = p/2:
°
°
°
∗
°
∗
°
°
°
°
..
..
..
..
°
°
°
°
.
.
.
.
°
°
°
°
°
°
°
∗
° ∗
2
∗
∗
S ≤ °Q 2 H j Q 1 Hi Q 2 H j Q 1 Hi ° × °Hi Q 3 H j Hi Q 3 H j ° .
°
°
°
°
..
..
..
..
°
°
°
°
°
°
°
°
.
.
.
.
p/2
p/2
Write each product of two block matrices as a sum:
°
°X
°
° n
S 2 ≤ ° i,j =1 Hi Q 1∗ H 2j Q 1 Hi °
p/2
°
°X
°
° n
× ° i,j =1 H j Q 3∗ Hi2Q 3 H j °
p/2
.
The two factors have the same form, so it suffices to bound the first one. Indeed,
°
°X
°
° °
°
°Xn
°Xn
°
° n
= ° i=1 Hi Q 1∗V Q 1 Hi °p/2 ≤ ° i=1 Hi2 °p · °Q 1∗V Q 1 °p = kV k2p = σ42p .
° i,j =1 Hi Q 1∗ H 2j Q 1 Hi °
p/2
We have identified the matrix variance V , defined in (8.1), and then we applied the Lust-Piquard inequality, Fact 8.2,
with ̺ = p. We identified V again, invoked unitary invariance of the Schatten norm, and then we recognized the
quantity σ2p from (8.1). In summary, we have established that S 2 ≤ σ82p . This is what we needed to show.
9. S ECOND -O RDER M ATRIX K HINTCHINE UNDER S TRONG I SOTROPY
In this section, we prove an extension of Theorem 3.11 that gives both lower and upper bounds for the trace
moments of a strongly isotropic matrix Gaussian series.
P
Theorem 9.1 (Second-Order Matrix Khintchine under Strong Isotropy). Let X := ni=1 γi Hi be an Hermitian matrix
Gaussian series, as in (2.1). Assume that X has the strong isotropy property
¯ X p ) · I for each p = 0, 1, 2, . . . .
E X p = (E tr
Define the matrix standard deviation parameter and matrix alignment parameter
°1/4
°X
°1/2
°Xn
°
° n
and w := max ° i,j =1 Hi Q 1 H j Q 2 Hi Q 3 H j ° .
σ := ° i=1 Hi2 °
Qℓ
(9.1)
(9.2)
The maximum ranges over a triple (Q 1 ,Q 2 ,Q 3 ) of unitary matrices. Then, for each integer p ≥ 1,
£
¤1/(2p)
¡
¢
1/(2p)
¯ X 2p 1/(2p) ≤ Cat1/(2p)
Catp
· σ · 1 − (pw/σ)4 +
≤
E tr
· σ + 21/4 p 5/4 · w.
p
The lower bound also requires that p 7/4 w ≤ 0.7σ. We have written Catp for the pth Catalan number, the function
¯ is the normalized trace.
[a]+ := max{a, 0}, and tr
The proof of this result appears below, starting in Section 9.3. To reach the statement of Theorem 3.11 in the
introduction, we rewrite normalized traces in terms of Schatten norms. Fact 9.2 (below) states that the Catalan
numbers satisfy the bound Catp ≤ 4p for each p = 1, 2, 3, . . . , which gives an explicit numerical form for the upper
bound.
9.1. Discussion. Before we establish Theorem 9.1, let us comment on the proof and the meaning of this result. The
most important observation is that the estimate is extremely accurate, at least for some examples. In particular, for
the GOE matrix X goe defined in (2.7), we showed in Section 4.2 that the standard deviation parameter σ ≈ 1 while
the alignment parameter w ≈ d −1/4 . Therefore, Theorem 9.1 implies that
2p
¯ X goe ≈ Catp
E tr
when p ≪ d 1/7 .
This estimate is sufficient to prove that the limiting spectral distribution of the GOE is the semicircle law. See [Tao12,
Sec. 2.3] for details about how to derive the law from the trace moments. Furthermore, Markov’s inequality implies
that the norm kX goe k ≈ 2 with high probability.
The proof of Theorem 9.1 has a lot in common with the arguments leading up to Proposition 7.1 and Theorem 8.1. The main innovation is that we can use the strong isotropy to imitate a moment identity that would hold
in free probability. This idea allows us to remove the dependence on p from the standard deviation term.
Although it may seem that the proof requires the matrix X to be a Gaussian series, there are analogous techniques, based on the theory of exchangeable pairs [MJC+ 14], that allow us to deal with other types of random
matrix series. This observation has the potential to lead to universality laws. It is also clear from the argument that
we could prove related results with an approximate form of strong isotropy.
SECOND-ORDER MATRIX CONCENTRATION INEQUALITIES
In addition, it is possible to extend these ideas to a rectangular matrix Gaussian series Z :=
In this case, we consider the Hermitian dilation:
¸
·
0
Z
.
H (Z ) = ∗
Z
0
19
Pn
i=1 γi S i
∈ Cd1 ×d2 .
The correct analog of strong isotropy is that
E H (Z )2p =
·
¯ Z ∗ )p ) · I
(E tr(Z
0
0
¯ ∗ Z )p ) · I
(E tr(Z
¸
for p = 0, 1, 2, . . . .
This observation allows us to obtain sharp bounds for the trace moments of rectangular Gaussian matrices. In this
fashion, we can even show that the limiting spectral density of a sequence of rectangular Gaussian matrices is the
Marcenko–Pastur distribution, provided that the aspect ratio of the sequence is held constant.
Finally, we remark that similar arguments can be applied to obtain algebraic relations for the Stieltjes transform of the matrix X . This approach may lead more directly to limit laws for sequences of random matrices with
increasing dimension. See [AGZ10, Sec. 2.4.1] or [Kem13, Sec. 9] for an argument of this species.
9.2. Preliminaries. Aside from the results we have collected so far, the proof of Theorem 9.1 requires a few additional ingredients. First, we state some of the basic and well-known properties of the Catalan numbers.
Fact 9.2 (Catalan Numbers). The pth Catalan number is defined by the formula
à !
1 2p
Catp :=
for p = 0, 1, 2, . . . .
p +1 p
(9.3)
In particular, p 7→ Catp is nondecreasing, and Catp ≤ 4p for each p. The Catalan numbers satisfy the recursion
Cat0 = 1 and Catp+1 =
Xp
q=0 Catq Catp−q .
(9.4)
The next result is a covariance identity for a product of centered functions of a Gaussian vector [NP12, Thm. 2.9.1].
It can be regarded as a refinement of the Poincaré inequality, which provides a bound for the variance of a centered
function of a Gaussian vector.
Fact 9.3 (Gaussian Covariance Identity). Let γ, γ′ ∈ Rn be independent standard normal vectors. Let f , g : Rn → C
be functions whose derivatives are square integrable with respect to the standard normal measure, and assume that
E f (γ) = E g (γ) = 0. Then
£
¤
E f (γ) · g (γ) =
Z∞
0
p
£
¤
dt Xn
−t
E
(∂
f
)(γ)
·
(∂
g
)(γ
)
where
γ
:=
e
γ
+
1 − e−2t γ′ .
t
t
j
j
j
=1
et
The symbol ∂ j refers to differentiation with respect to the j th coordinate.
The usual statement of this result involves the Ornstein–Uhlenbeck semigroup, but we have given a more elementary formulation.
Finally, we need a bound for the solution to a certain type of polynomial inequality. This estimate is related to
Fujiwara’s inequality [Mar66, Sec. 27]. We include a proof sketch since we could not locate the precise statement
in the literature.
Proposition 9.4 (Polynomial Inequalities). Consider an integer k ≥ 3, and fix positive numbers α and β. Then
u k ≤ α + βu k−2
implies u ≤ α1/k + β1/2
Proof Sketch. Consider the polynomial ϕ : u 7→ u k − βu k−2 − α. The Descartes Rule of Signs implies that ϕ has
exactly one positive root, say u+ . Furthermore, ϕ(u) ≤ 0 for a positive number u if and only if u ≤ u+ because
ϕ(0) < 0. By direct calculation, one may verify that u⋆ = α1/k + β1/2 satisfies ϕ(u⋆ ) > 0, which means that u+ < u⋆ .
We conclude that ϕ(u) ≤ 0 implies u ≤ u+ < u⋆ .
20
J. A. TROPP
9.3. The Normalized Trace Moments. Let us commence the proof of Theorem 9.1. First, we introduce notation
for the normalized trace moments of the matrix Gaussian series:
¯ Xp
µp := E tr
for p = 0, 1, 2, . . . .
It is clear that µ0 = 1. Since X is a symmetric random variable, the odd trace moments are zero:
µ2p+1 = 0 for p = 0, 1, 2, . . . .
It remains to calculate the even trace moments.
We can obtain the second moment from a simple argument:
Xn
¡
¢
2
2
¯ 2 · I = µ2 · I.
i=1 H i = E X = E tr X
(9.5)
The first identity follows from a direct calculation using the definition (2.1) of the matrix Gaussian series. The
second identity is the strong isotropy hypothesis (9.1), and the last relation is the definition of µ2 . Take the spectral
norm of (9.5) to see that
µ2 = σ2 .
(9.6)
We have identified the standard deviation parameter, defined in (9.2).
9.4. Representation of Higher-Order Moments. The major challenge is to compute the rest of the even moments.
As usual, the first step is to invoke Gaussian integration by parts. For each integer p ≥ 1, Lemma 6.1 implies that
X2p Xn
£
¤
¯ Hi X q Hi X 2p−q .
µ2(p+1) = q=0 i=1 E tr
We are considering µ2(p+1) instead of µ2p because it makes the argument cleaner. To analyze this expression, we
will examine each index q separately and subject each one to the same treatment.
Fix an index 0 ≤ q ≤ 2p. First, we center both X q and X 2p−q by adding and subtracting their expectations:
Xn
£
¤ X
£ ¡
¢ ¡
¢¤
¯ Hi X q Hi X 2p−q = n tr
¯ Hi E X q Hi E X 2p−q
E tr
i=1
i=1
Xn
£ ¡
¢ ¡
¢¤
¯ Hi X q − E X q Hi X 2p−q − E X 2p−q .
+ i=1 E tr
The cross-terms vanish because each one has zero mean. It is productive to think of the first sum on the right-hand
side as an approximation to the left-hand side, while the second sum is a perturbation.
Let us focus on the first sum on the right-hand side of the last display. We can use the strong isotropy hypothesis (9.1) to simplify this expression:
Xn
£ ¡
¢ ¡
¢¤ Xn
£ ¡¡
¢ ¢ ¡¡
¢ ¢¤
q
¯
¯ Hi E tr
¯ X q · I Hi E tr
¯ X 2p−q · I
Hi E X 2p−q = i=1 tr
i=1 tr H i E X
¡ Xn
¢¡
¢¡
¢
2
¯
¯ X q E tr
¯ X 2p−q
= tr
E tr
i=1 H i
= σ2 · µq µ2p−q .
The last identity follows from (9.5) and (9.6). As a side note, our motivation here is to imitate the moment identity
that would hold if X and Hi were free from each other, in the sense of free probability.
Finally, we combine the last three displays to reach
Xp
X2p−1 Xn
£ ¡ q
¢ ¡ 2p−q
¢¤
q
¯
Hi X
− E X 2p−q .
(9.7)
µ2(p+1) = σ2 · q=0 µ2q µ2(p−q) + q=1
i=1 E tr H i X − E X
Observe that we have modified the indexing of both sums. This step depends on the facts that µq = 0 for odd q and
that X 0 = I.
9.5. The Perturbation Term. The next step in the argument is to bound the perturbation term in (9.7) in terms of
the alignment parameter w, defined in (9.2). We will use the Gaussian covariance identity, Fact 9.3.
To that end, let us explain how to write each summand in the perturbation term as a covariance. Let H be a real,
diagonal matrix: H = diag(h1 , . . . , hd ). Expanding the normalized trace, using α, β for coordinate indices, we find
that
£ ¡
¢ ¡
¢¤ 1 Xd
£¡
¢ ¡
¢ ¤
¯ H X q − E X q H X 2p−q − E X 2p−q =
E tr
hα hβ · E X q − E X q αβ X 2p−q − E X 2p−q βα .
d α,β=1
To apply the Gaussian covariance identity to the expectation, we introduce a parameterized family {X t : t ≥ 0} of
random matrices where
p
Xn ¡
¢
X t := k=1 e−t γk + 1 − e−2t γ′k · Hk where γ′ is an independent copy of γ.
SECOND-ORDER MATRIX CONCENTRATION INEQUALITIES
21
Observe that X and X t have the same distribution, although they are dependent. Fact 9.3 and Fact 6.3 deliver
£¡
¢ ¡
¢ ¤
E X q − E X q αβ X 2p−q − E X 2p−q βα
Z∞
h³ X
´ ³X
´ i
dt Xn
q−1 r
2p−q−1 s
2p−q−1−s
q−1−r
=
E
X
H
X
X
H
X
j
j
t
t
j =1
r =0
s=0
et
αβ
βα
0
Xq−1 X2p−q−1 Z∞ dt Xn
£¡ r
¢ ¡ s
2p−q−1−s ¢ ¤
q−1−r
= r =0 s=0
βα .
αβ X t H j X t
j =1 E X H j X
et
0
Combining these formulas and expressing the result in terms of the normalized trace again, we find that
£ ¡
¢ ¡
¢¤
¯ H X q − E X q H X 2p−q − E X 2p−q
E tr
Xq−1 X2p−q−1 Z∞ dt Xn
¢¤
£ ¡
¢ ¡
¯ H X r H j X q−1−r H X ts H j X t2p−q−1−s .
= r =0 s=0
E tr
j
=1
t
e
0
In fact, this expression is valid for any Hermitian matrix H because of the unitary invariance of the trace. Summing
the last identity over H = Hi , we reach
Xn
£ ¡ q
¢ ¡ 2p−q
¢¤
q
¯
− E X 2p−q
i=1 E tr H i X − E X H i X
Xq−1 X2p−q−1 Z∞ dt Xn
£
2p−q−1−s ¤
r
q−1−r
¯
. (9.8)
Hi X ts H j X t
= r =0 s=0
i,j =1 E tr H i X H j X
et
0
At this point, the alignment parameter w starts to become visible.
9.6. Finding the Matrix Alignment Parameter. Our next goal is to control the expression (9.8) in terms of the
alignment parameter w. To do so, we use the interpolation result, Proposition 8.3, to bound the sum over (i , j ). For
each choice of indices (q, r, s), we obtain the estimate
¯X
¯X
n
¤¯
£
¤¯
£
¯ n
¯ Hi Q 1 X 2p−2 H j Q 2 Hi Q 3 H j Q 4 ¯¯ ,
¯ Hi X r H j X q−1−r Hi X ts H j X t2p−q−1−s ¯¯ ≤ max E max ¯¯ n
tr
¯ i,j =1 E tr
i,j
=1
Qℓ
¯X
£
¤¯
¯ n
¯ Hi Q 1 H j Q 2 X 2p−2 Hi Q 3 H j Q 4 ¯¯ ,
E max ¯ i,j =1 tr
Qℓ
¯X
¤¯
£
¯ n
¯ Hi Q 1 H j Q 2 Hi Q 3 X t2p−2 H j Q 4 ¯¯ ,
E max ¯ i,j =1 tr
Qℓ
¯X
¤¯ o
£
¯ n
¯ Hi Q 1 H j Q 2 Hi Q 3 H j Q 4 X t2p−2 ¯¯ .
E max ¯ i,j =1 tr
Qℓ
Each random unitary matrix Q ℓ commutes with the corresponding random matrix X or X t .
As in Section 8.7, we can bound each term in the maximum in the same fashion. For example, consider the
fourth term:
¯X
h°X
° ¡
¢i
¤¯
£
¯ n
° ¯ 2p−2
¯ Hi Q 1 H j Q 2 Hi Q 3 H j Q 4 X t2p−2 ¯¯ ≤ E max ° n
E max ¯ i,j =1 tr
i,j =1 H i Q 1 H j Q 2 H i Q 3 H j · tr X t
Qℓ
Qℓ
°Xn
° ¡
¢
¯ X t2p−2
≤ max ° i,j =1 Hi Q 1 H j Q 2 Hi Q 3 H j ° · E tr
Qℓ
= w 4 · µ2(p−1) .
The first step is Hölder’s inequality for the trace, and the second step is Hölder’s inequality for the expectation. In
the last line, we recall that X t has the same distribution as X to identify µ2(p−1) . Finally, we recognize the matrix
alignment parameter w, defined in (9.2).
In summary, we have shown that
¯X
¤¯
£
¯ n
¯ Hi X r H j X q−1−r Hi X ts H j X t2p−q−1−s ¯¯ ≤ w 4 · µ2(p−1) .
¯ i,j =1 E tr
Introduce this bound into (9.8) to arrive at
¯Xn
£ ¡ q
¢ ¡ 2p−q
¢¤¯
q
¯
¯
Hi X
− E X 2p−q ¯ ≤ q(2p − q) · w 4 · µ2(p−1) ≤ p 2 · w 4 · µ2(p−1) .
i=1 E tr H i X − E X
We have used the numerical inequality u(a − u) ≤ a 2 /4, valid for u ∈ R. Finally, we sum this expression over the
index q to conclude that
¯X
£ ¡
¢ ¡
¢¤¯
¯ 2p−1 Xn
¯ Hi X q − E X q Hi X 2p−q − E X 2p−q ¯¯ ≤ 2p 3 · w 4 · µ2(p−1) .
E
tr
(9.9)
¯ q=0
i=1
This is the required bound for the perturbation term in (9.7).
22
J. A. TROPP
9.7. A Recursion for the Trace Moments. In view of (9.7) and (9.9), we have shown that
Xp
µ2(p+1) = σ2 · q=0 µ2q µ2(p−q) ± 2p 3 · w 4 · µ2(p−1) for p = 1, 2, 3, . . . .
(9.10)
We have written ± to indicate that the expression contains both a lower bound and an upper bound for the normalized trace moment µ2(p+1) .
In the next two sections, we will solve this recursion to obtain explicit bounds on the trace moments. First, we
obtain the upper bound
1/(2p)
µ2p
1/(2p)
≤ Catp
· σ + 21/4 p 5/4 · w
for p = 1, 2, 3, . . . .
(9.11)
7/4
This result gives us a Khintchine-type inequality. Afterward, assuming p w ≤ 0.7σ, we establish the lower bound
£
¤1/(2p)
1/(2p)
1/(2p)
µ2p
≥ Catp
· σ · 1 − (pw/σ)4 +
for p = 1, 2, 3, . . . .
(9.12)
Together these estimates yield the statement of Theorem 9.1.
9.8. Solving the Recursion: Upper Bound. We begin with the proof of the upper bound (9.11). The first step in
the argument is to remove the lag term µ2(p−1) from the recursion (9.10) using moment comparison. Fix an integer
p ≥ 1. Observe that
¡
¢
¡
¢
¯ X 2(p−1) ≤ E tr
¯ X 2(p+1) (p−1)/(p+1) ≤ E tr
¯ X 2(p+1) (p−1)/(p+1) = µ(p−1)/(p+1) .
µ2(p−1) = E tr
2(p+1)
¯ A q )1/q is increasing for any positive-semidefinite matrix A, while the
The first inequality holds because q 7→ (tr
second inequality is Lyapunov’s. Introduce this estimate into the recursion (9.10) to obtain
Xp
(p−1)/(p+1)
.
µ2(p+1) ≤ σ2 · q=0 µ2q µ2(p−q) + 2p 3 w 4 · µ2(p+1)
This is a polynomial inequality of the form u p+1 ≤ α+ βu p−1 , so Proposition 9.4 ensures that u ≤ α1/(p+1) + β1/2 . In
other words,
´1/(p+1) p
³
Xp
1/(p+1)
+ 2 p 3/2 · w 2 for p = 1, 2, 3, . . . .
µ2(p+1) ≤ σ2 · q=0 µ2q µ2(p−q)
(9.13)
Using this formula, we will apply induction to prove that
p
1/p
1/p
µ2p ≤ Catp · σ2 + 2 p 5/2 · w 2
when p = 1, 2, 3, . . . .
(9.14)
The stated result (9.11) follows from (9.14) once we take the square root and invoke subadditivity.
Let us commence the induction. The formula (9.14) holds for p = 1 because µ2 = σ2 , as noted in (9.6). Assuming
that the bound (9.14) holds for each integer p in the range 1, 2, 3, . . . , r , we will verify that the same bound is also
valid for p = r + 1. For any integer q in the range 1 ≤ q ≤ r , the bound (9.14) implies that
Ã
!
Ãp
p 5/2 2 !q
2q w
2 q 7/2 w 2
2q
2q
µ2q ≤ Catq · σ · 1 +
.
(9.15)
≤ Catq · σ · exp
1/q
σ2
Cat σ2
q
−1/q
Using the definition (9.3) of the Catalan numbers, one may verify that q 7→ q 5/2 Catq
is increasing, so
Ã
!q
p
2(r + 1)5/2 w 2
for q = 0, 1, 2, . . . , r .
µ2q ≤ Catq · σ2q · 1 +
+1) 2
Cat1/(r
σ
r +1
The case q = 0 follows by inspection. Now, the latter bound and the recursion (9.4) for Catalan numbers together
imply that
Ã
!r
p
Xr
2 (r + 1)5/2 w 2
2
2(r +1)
.
σ · q=0 µ2q µ2(r −q) ≤ Catr +1 · σ
· 1+
+1) 2
Cat1/(r
σ
r +1
Take the r + 1 root to determine that
Ã
´1/(r +1)
³
Xr
1/(r +1)
2
2
≤ Catr +1 · σ · 1 +
σ · q=0 µ2q µ2(r −q)
!
p
r
2 (r + 1)5/2 w 2
,
·
+1) 2
r +1
Cat1/(r
σ
r +1
p
p
+1)
= Cat1/(r
· σ2 + 2 (r + 1)5/2 w 2 − 2 (r + 1)3/2 w 2 .
r +1
We have used the numerical inequality (1 + x)α ≤ 1 + αx, valid for 0 ≤ α ≤ 1 and x ≥ 0. Combine this estimate with
the recursive bound (9.13) for p = r to obtain
p
+1)
1/(r +1)
µ1/(r
· σ2 + 2(r + 1)5/2 w 2 .
2(r +1) ≤ Catr +1
SECOND-ORDER MATRIX CONCENTRATION INEQUALITIES
23
We see that (9.14) holds for p = r + 1, and the induction may proceed.
9.9. Solving the Recursion: Lower Bound. We turn to the proof of the lower bound (9.12). Assuming that p 7/4 w ≤
0.7σ, we will use induction to show that
£
¤
(9.16)
µ2p ≥ σ2p · Catp · 1 − (pw/σ)4 + for p = 0, 1, 2, 3, . . . .
The result (9.12) follows once we take the (2p)th root.
To begin the induction, recall that µ0 = 1, so the formula (9.16) is valid for p = 0. Suppose now that (9.16) is
valid for each integer p in the range 0, 1, 2, . . . , r . We will verify the formula for p = r + 1. The lower branch of the
recursion (9.10) states that
Xr
µ2(r +1) ≥ σ2 · q=0 µ2q µ2(r −q) − 2r 3 w 4 µ2(r −1) .
The induction hypothesis (9.16) yields
Xr
Xr
£
¤
σ2 · q=0 µ2q µ2(r −q) ≥ σ2(p+1) · q=0 Catq Catr −q · 1 − (q w/σ)4 − ((r − q)w/σ)4
≥ Catr +1 · σ2(p+1) · (r w/σ)4 .
We have used the fact that q 7→ q 4 + (r − q)4 achieves its maximum value on [0, r ] at one of the endpoints because
of convexity. We also applied the recursive formula (9.4) for the Catalan numbers. The bound (9.15) implies that
s
r 7/4 w
log 2
2(r −1)
≤
µ2(r −1) ≤ 2Catr −1 σ
when
p ≈ 0.7.
σ
2
Therefore,
2r 3 w 4 µ2(r −1) ≤ 4Catr −1 · σ2(r −1) · r 3 w 4 ≤ 4Catr +1 · σ2(r +1) · r 3 (w/σ)4 .
The second inequality holds because the Catalan numbers are nondecreasing. Combine the last three displays to
arrive at
£
¤
£
¤
µ2(r +1) ≥ Catr +1 · σ2(r +1) · 1 − (r 4 + 4r 3 )(w/σ)4 ≥ Catr +1 · σ2(r +1) · 1 − (r + 1)4 (w/σ)4 .
We have verified the formula (9.16) for p = r + 1, which completes the proof.
A PPENDIX A. I NTERPOLATION R ESULTS
In this appendix, we establish Proposition 8.3, the interpolation inequality for a multilinear function of a random matrix, whose proof appears below in Appendix A.4.
A.1. Multivariate Complex Interpolation. The interpolation result we use in the body of the paper is a consequence of a more general theorem on interpolation for a function of several complex variables.
Proposition A.1 (Multivariate Complex Interpolation). Let k be a natural number. For a positive number α, define
the simplicial prism
n
o
Xk
∆k (α) := (c 1 , . . . , c k ) ∈ Ck : Re c i ≥ 0 for each i and i=1 Re c i ≤ α .
Consider a bounded, continuous function G : ∆k (α) → C. For each pair (i , j ) of distinct indices and each c ∈ ∆k (α),
assume that G has the analytic section property:
z 7→ G(c 1 , . . . , c i + z, . . . , c j − z, . . . , c k ) is analytic on − Re c i < Re z < Re c j .
P
Then, for each z ∈ ∆k (α) with β := ni=1 Re zi > 0,
|G(z1 , . . . , zk )| ≤
³Y
k
i=1
¯Re z ´1/β
¯
.
suptℓ ∈R ¯G(it1 , . . . , iti−1 , β + iti , iti+1 , . . . , itk )¯ i
(A.1)
(A.2)
We establish Proposition A.1 in the next two sections. The argument relies on the same principles that support
standard univariate complex interpolation. Although it seems likely that a result of this form already appears in
the literature, we were not able to locate a reference.
24
J. A. TROPP
A.2. Preliminaries. Proposition A.1 depends on Hadamard’s Three-Lines Theorem [Gar07, Prop. 9.1.1].
Proposition A.2 (Three-Lines Theorem). Consider the vertical strip ∆1 (1) in the complex plane:
∆1 (1) := {z ∈ C : 0 ≤ Re z ≤ 1}.
Consider a bounded, continuous function f : ∆1 (1) → C that is analytic in the interior of ∆1 (1). For each θ ∈ [0, 1],
sup | f (θ + it )| ≤ sup | f (1 + it )|θ · sup | f (it )|1−θ .
t ∈R
t ∈R
t ∈R
As we will see, this result delivers the k = 2 case of Proposition A.1.
A.3. Proof of Proposition A.1. The proof of the multivariate interpolation result, Proposition A.1, follows by induction on the number k of arguments.
Let us begin with the base cases. When the function G has one argument only, the inequality (A.2) is obviously
true. Next, consider a bivariate function G 2 : ∆2 (α) → C that is bounded and continuous and has the analytic
section property (A.1). Fix a point z ∈ ∆2 (α) with β := Re z1 + Re z2 > 0. Define the bounded, continuous function
f (y) := G 2 (βy + i Im z1 , β(1 − y) + i Im z2 ) for y ∈ ∆1 (1).
The assumption (A.1) implies that f is analytic on 0 < Re y < 1. Select θ = Re z1 /β, which gives 1 − θ = Re z2 /β. An
application of the Three-Lines Theorem, Proposition A.2, implies that
¯
¯
¯
¯θ
¯
¯1−θ
|G 2 (z1 , z2 )| = ¯ f (θ)¯ ≤ supt ∈R ¯ f (1 + it )¯ · sup t ∈R ¯ f (it )¯
.
Introducing the definition of f and simplifying,
¯Re z /β
¯
|G 2 (z1 , z2 )| ≤ supt ∈R ¯G 2 (β(1 + it ) + i Im z1 , −βit + i Im z2 )¯ 1
¯
¯Re z /β
× supt ∈R ¯G 2 (βit + i Im z1 , β(1 − it ) + i Im z2 )¯ 2
³
¯
¯Re z
¯
¯Re z ´1/β
≤ sups 1 ,s 2 ∈R ¯G 2 (β + is1 , is2 )¯ 1 · sups 1 ,s 2 ∈R ¯G 2 (is1 , β + is2 )¯ 2
.
(A.3)
This is the k = 2 case of Proposition A.1.
Fix a positive integer k, and suppose that we have established the inequality (A.2) for functions with k − 1 arguments. In other words, assume that G k−1 : ∆k−1 (α′ ) → C is bounded and continuous, and it has the analytic section
Pk−1
property (A.1). Then, for z ∈ ∆k−1 (α′ ) with β′ := i=1
Re zi > 0,
|G k−1(z1 , . . . , zk−1 )| ≤
³Y
k−1
i=1 supt ℓ ∈R
´1/β′
¯
¯
¯G k−1 (it1 , . . . , β′ + iti , . . . , itk−1 )¯Re zi
.
(A.4)
We need to extend this result to functions with k variables.
Consider a bounded, continuous function G k : ∆k (α) → C with the analytic section property (A.1). Fix a complex
P
Pk−1
vector z ∈ ∆k (α) with β := ki=1 Re zi > 0. Define the number β′ := i=1
Re zi . When β′ = 0, the formula (A.2) is
trivial for G = G k because Re z1 = · · · = Re zk−1 = 0. Therefore, we may assume that β′ > 0. Introduce the function
G k−1 (y 1 , . . . , y k−1 ) := G k (y 1 , . . . , y k−1 , zk ) for y ∈ ∆k−1 (β′ ).
One may verify that G k−1 inherits boundedness, continuity, and analytic sections from G k . Therefore, the induction hypothesis (A.4) gives
|G k (z1 , . . . , zk−1, zk )| ≤
³Y
k−1
i=1 supt ℓ ∈R
´1/β′
¯
¯
¯G k (it1 , . . . , β′ + iti , . . . , itk−1 , zk )¯Re zi
.
(A.5)
For each fixed choice of the index i and of the numbers t1 , . . . , tk−1 ∈ R, consider the function
G 2 (y i , y k ) := G k (it1 , . . . , y i , . . . , itk−1 , y k ) for (y i , y k ) ∈ ∆2 (α).
Since β′ + Re zk = β, the bivariate case (A.3) provides that
¯
¯
¯
¯β′ /β
¯G k (it1 , . . . , β + iti , . . . , itk−1 , zk )¯ ≤ sup
¯
¯
s i ,s k ∈R G k (it 1 , . . . , β + is i , . . . , it k−1 , is k )
¯
¯1−β′ /β
× sups i ,s k ∈R ¯G k (it1 , . . . , isi , . . . , itk−1 , β + isk )¯
.
(A.6)
SECOND-ORDER MATRIX CONCENTRATION INEQUALITIES
25
Combine the bounds (A.5) and (A.6) to reach
³Y
¯
¯Re zi ´1/β
k−1
¯
¯
|G k (z1 , . . . , zk−1 , zk )| ≤
i=1 supt ℓ ∈R G k (it 1 , . . . , β + it i , . . . , it k )
×
³Y
k−1
suptℓ ∈R
i=1
´(β−β′ )/(β′ β)
¯
¯
¯G k (it1 , . . . , iti , . . . , β + itk )¯Re zi
Pk−1
Since β′ = i=1
Re zi and β − β′ = Re zk , we see that the second product has the same form as the i = k term of the
first product. Thus,
³Y
¯Re zi ´1/β
¯
k
¯
¯
|G k (z1 , . . . , zk )| ≤
.
i=1 sup t ℓ ∈R G k (it 1 , . . . , β + it i , . . . , it k )
This step completes the induction. We have established Proposition A.1.
A.4. Interpolation for a Multilinear Function of Random Matrices. We are now prepared to establish the interpolation result, Proposition 8.3, for a multilinear function of random matrices. We will actually establish a somewhat
more precise version, which we state here.
Proposition A.3 (Refined Multilinear Interpolation). Suppose that F : (Md )k → C is a multilinear function. Fix
P
nonnegative integers α1 , . . . , αk with ki=1 αk = α. Let Y i ∈ Hd be random Hermitian matrices, not necessarily independent, for which E kY i kα < ∞. Then
¶
µ
Yk h
¯ ¡
¯ ¡ α1
¢¯ iαi 1/α
¢¯
α
¯
¯
¯E F Y , . . . , Y αk ¯ ≤
.
1
i=1 E max F Q 1 , . . . , Q i−1 , Q i Y i , Q i+1 , . . . , Q k
k
Qℓ
Each Q ℓ is a unitary matrix that commutes with Y ℓ .
We establish this result in the next section.
Observe that Proposition A.3 immediately implies Proposition 8.3, the interpolation result that we use in the
body of the paper. Indeed, we recognize the large parenthesis on the right-hand side as a geometric mean, and we
bound the geometric mean by the maximum of its components.
A.5. Proof of Proposition A.3. By a perturbative argument, we may assume that each matrix Y i is almost surely
nonsingular. Indeed, for a parameter ε > 0, we can replace each Y i by the modified matrix Yei := Y i + εγi I, where
{γi } is an independent family of standard normal variables. After completing the argument, we can draw ε down
to zero to obtain the inequality for the original random matrices.
The first step in the argument is to perform a polar factorization of each random Hermitian matrix: Y i = Ui P i
where Ui is unitary, P i is almost surely positive definite, and the two factors commute with Y i for each index i . For
α
clarity of argument, we introduce the unitary matrices S i = Ui i . With this notation,
¯ ¡ α1
¢¯ ¯ ¡
¢¯
¯E F Y , . . . , Y αk ¯ = ¯E F S 1 P α1 , . . . , S k P αk ¯
(A.7)
1
1
k
k
We will perform interpolation only on the positive-definite matrices.
Next, we introduce a complex-valued function by replacing the powers αi with complex variables:
¡
z ¢
z
G : (z1 , . . . , zk ) 7→ E F S 1 P 1 1 , . . . , S k P k k
for z ∈ ∆k (α).
The set ∆k (α) is the simplicial prism defined in the statement of Proposition A.1.
Claim A.4. The function G is bounded, continuous, and has analytic sections.
These are the properties required to apply the interpolation result, Proposition A.1.
Let us assume that Claim A.4 holds so that we can complete the proof. The relation (A.7) and Proposition A.1
imply that
¯
¯ ¯
¯
¯E F (Y α1 , . . . , Y αk )¯ = ¯E F (S 1 P α1 , . . . , S k P αk )¯
1
1
k
k
¯ ´1/α
¯ ¡
³Y
k
¯
it k ¢¯αi
α+it i
it 1
.
,
.
.
.
,
S
P
,
.
.
.
S
P
F
S
P
≤
sup
¯
¯E
1
i
k
t
∈
R
1
i=1
i
k
ℓ
it
Fix an index i in the product. Introduce the unitary matrix Q ii (ti ) = S i P i i Ui−α , where Ui is the polar factor of Y i .
It follows that
¡
¢¡
¢
α+it
it
S i P i i = S i P i i Ui−α Uiα P iα = Q ii (ti )Y iα
26
J. A. TROPP
it
Similarly, for each ℓ 6= i , we can define Q iℓ (tℓ ) = S ℓ P ℓ ℓ . Therefore,
!1/α
Ã
Yk
¯ ¡
¯ ¡ α1
¢¯αi
¢¯
α
¯
¯
¯E F Y , . . . , Y αk ¯ ≤
1
i=1 sup E F Q i1 (t 1 ), . . . , Q ii (t i )Y i , . . . , Q ik (t k )
k
≤
µ
t ℓ ∈R
Yk
i=1
h
¯ ¡
¢¯ iαi
E max ¯F Q 1 , . . . , Q i Y iα , . . . , Q k ¯
Qℓ
¶1/α
.
By construction Q iℓ (tℓ ) commutes with Y ℓ for each index ℓ. In the second line, we apply Jensen’s inequality. Then
we relax the supremum to include all unitary matrices Q ℓ that commute with the corresponding Y ℓ . We can replace
the supremum with a maximum since the unitary group is compact and the function F is continuous. This is what
we needed to show.
Finally, we must verify Claim A.4. Each multilinear function F : (Md )k → C is bounded and continuous:
Yk
|F (A 1 , . . . , A k )| ≤ Const · i=1 kA i k .
P
Fix a point z ∈ ∆k (α), and let β := ki=1 Re zi . Applying this observation to the function G,
hY
i
¯
k
z ¯
z
|G(z1 , . . . , zk )| = ¯E F (S 1 P 1z1 , . . . , S k P k k )¯ ≤ Const · E
kS i P i i k
i=1
·
¸β
hY
i
1 Xk
k
Re z i
kY i k
kY i k .
= Const · E
≤ Const · E
i=1
β i=1
The first estimate follows from Jensen’s inequality and the bound on the multilinear function F . The second inequality depends on the unitary invariance of the spectral norm, the identity kP z k = kPkRe z , and the polar decomposition Y i = Ui P i . The last bound is the inequality between the geometric and arithmetic mean. Since
E kY i kα < ∞, and β ≤ α, we conclude that G is bounded. Since F is continuous, an application of the Dominated
Convergence Theorem shows that G is a continuous function as well.
The proof that G has analytic sections is similar. Fix a vector c ∈ ∆k (α). Since F is multilinear, it is easy to check
that the map
¡
c j −z
c ¢
c +z
c
z 7→ F S 1 P 11 , . . . , S i P i i , . . . , S j P j
, . . . , S k P1k
is analytic on − Re c i < z < Re c j
for any fixed choice of S ℓ and P ℓ and any pair (i , j ) of distinct indices. Together, the Morera Theorem and the
Fubini–Tonelli Theorem allow us to conclude that
¡
c j −z
c ¢
c +z
c
z 7→ E F S 1 P 11 , . . . , S i P i i , . . . , S j P j
, . . . , S k P1k
also is analytic. Therefore, the analytic section property (A.1) is in force. Claim A.4 is established.
A CKNOWLEDGMENTS
Afonso Bandeira is responsible for the argument in Section 4.3, and Ramon van Handel has offered critical comments. Parts of this research were completed at Mathematisches Forschungsinstitut Oberwolfach (MFO) and at
Instituto Nacional de Matemática Pura e Aplicada (IMPA) in Rio de Janeiro. The author gratefully acknowledges
support from ONR award N00014-11-1002, a Sloan Research Fellowship, and the Gordon & Betty Moore Foundation.
R EFERENCES
[AGZ10]
[BV14]
[Bha97]
[BRV13]
[BS10]
[Buc01]
[CGT12]
[CT14]
G. W. Anderson, A. Guionnet, and O. Zeitouni. An introduction to random matrices, volume 118 of Cambridge Studies in Advanced
Mathematics. Cambridge University Press, Cambridge, 2010.
A. Bandeira and R. van Handel. Sharp nonasymptotic bounds on the norm of random matrices with independent entries. Available
at http://www.arXiv.org/abs/1408.6185 , Aug. 2014.
R. Bhatia. Matrix analysis, volume 169 of Graduate Texts in Mathematics. Springer-Verlag, New York, 1997.
A. Bondarenko, D. Radchenko, and M. Viazovska. Optimal asymptotic bounds for spherical designs. Ann. of Math. (2), 178(2):443–
452, 2013.
Z. Bai and J. W. Silverstein. Spectral analysis of large dimensional random matrices. Springer Series in Statistics. Springer, New York,
second edition, 2010.
A. Buchholz. Operator Khintchine inequality in non-commutative probability. Math. Ann., 319:1–16, 2001.
R. Y. Chen, A. Gittens, and J. A. Tropp. The masked sample covariance estimator: an analysis using matrix concentration inequalities.
Inf. Inference, 1(1):2–20, 2012.
R. Y. Chen and J. A. Tropp. Subadditivity of matrix ϕ-entropy and concentration of random matrices. Electron. J. Probab., 19:no. 27,
30, 2014.
SECOND-ORDER MATRIX CONCENTRATION INEQUALITIES
[Gar07]
[Kem13]
27
D. J. H. Garling. Inequalities: a journey into linear analysis. Cambridge University Press, Cambridge, 2007.
T. Kemp. Math 247a: Introduction to random matrix theory. Available at http://www.math.ucsd.edu/~tkemp/247A/247A.Notes.pdf ,
Nov. 2013.
[LO94]
R. Latała and K. Oleszkiewicz. On the best constant in the Khinchin-Kahane inequality. Studia Math., 109(1):101–104, 1994.
[LP86]
F. Lust-Piquard. Inégalités de Khintchine dans C p (1 < p < ∞). C. R. Math. Acad. Sci. Paris, 303(7):289–292, 1986.
[LT91]
M. Ledoux and M. Talagrand. Probability in Banach Spaces: Isoperimetry and Processes. Springer, Berlin, 1991.
[Mar66] M. Marden. Geometry of polynomials. Second edition. Mathematical Surveys, No. 3. American Mathematical Society, Providence,
R.I., 1966.
[MJC+ 14] L. Mackey, M. I. Jordan, R. Y. Chen, B. Farrell, and J. A. Tropp. Matrix concentration inequalities via the method of exchangeable pairs.
Ann. Probab., 42(3):906–945, 2014.
[NP12]
I. Nourdin and G. Peccati. Normal approximations with Malliavin calculus, volume 192 of Cambridge Tracts in Mathematics. Cambridge University Press, Cambridge, 2012. From Stein’s method to universality.
[NS06]
A. Nica and R. Speicher. Lectures on the combinatorics of free probability, volume 335 of London Mathematical Society Lecture Note
Series. Cambridge University Press, Cambridge, 2006.
[Pis98]
G. Pisier. Non-commutative vector valued L p -spaces and completely p-summing maps. Astérisque, (247):vi+131, 1998.
[PX97]
G. Pisier and Q. Xu. Non-commutative martingale inequalities. Comm. Math. Phys., 189(3):667–698, 1997.
[Tao12] T. Tao. Topics in random matrix theory, volume 132 of Graduate Studies in Mathematics. American Mathematical Society, Providence,
RI, 2012.
[Tro12]
J. A. Tropp. User-friendly tail bounds for sums of random matrices. Found. Comput. Math., 12(4):389–434, 2012.
[Tro15]
J. A. Tropp. An Introduction to Matrix Concentration Inequalities. Foundations and Trends in Machine Learning. Now, 2015. To appear. Available at http://www.arXiv.org/abs/1501.01571 .
[vH15]
R. van Handel. On the spectral norm of inhomogeneous random matrices. Available at http://www.arXiv.org/abs/1502.05003 ,
Feb. 2015.
[Ver12]
R. Vershynin. Introduction to the non-asymptotic analysis of random matrices. In Compressed sensing, pages 210–268. Cambridge
Univ. Press, Cambridge, 2012.
[VW08]
R. Vale and S. Waldron. Tight frames generated by finite nonabelian groups. Numer. Algorithms, 48(1-3):11–27, 2008.
| 10math.ST
|
Deep Transform: Cocktail Party Source Separation
via Complex Convolution in a Deep Neural Network
Andrew J.R. Simpson #1
#
Centre for Vision, Speech and Signal Processing, University of Surrey
Guildford, UK
1
[email protected]
Abstract—Convolutional deep neural networks (DNN) are
state of the art in many engineering problems but have not yet
addressed the issue of how to deal with complex spectrograms.
Here, we use circular statistics to provide a convenient
probabilistic estimate of spectrogram phase in a complex
convolutional DNN. In a typical cocktail party source separation
scenario, we trained a convolutional DNN to re-synthesize the
complex spectrograms of two source speech signals given a
complex spectrogram of the monaural mixture – a discriminative
deep transform (DT). We then used this complex convolutional
DT to obtain probabilistic estimates of the magnitude and phase
components of the source spectrograms. Our separation results
are on a par with equivalent binary-mask based non-complex
separation approaches.
Index terms—Deep learning, supervised learning, complex
convolution, deep transform.
I. INTRODUCTION
Convolutional deep neural networks (DNN) [1]-[3] are
capable of exploiting geometric assumptions about data
structure in order to share network weights. When the
convolutional DNN is applied in a sliding window fashion,
predictions for a given datapoint may be made for multiple
alternate (windowed) contexts and from this distribution of
predictions a probabilistic estimate may be obtained. In
computer vision problems [1]–[3], the datapoints representing
pixel intensities are positive real numbers. Therefore, if the
convolutional DNN is used to make predictions about
intensity for a given pixel, parametric statistics may be used to
obtain a probabilistic estimate of pixel intensity [4] that
summarizes the predictions made in the various different
contexts of the sliding window (containing the pixel in
question).
While the audio spectrogram provides an intuitive visual
counterpart to the 2D image of the computer vision problem,
the equivalent application of convolutional DNNs to computer
audition is less straight forward. In particular, the equivalence
of the 2D spectrogram and the 2D image holds only for
interpretations of the complex spectrogram that are limited to
the magnitude component [5]-[7]. The approach of ignoring
the phase component of the complex spectrogram does not
imply serious limitations for classification problems (where
phase may not be critical) but for audio synthesis phase is
critical. Furthermore, if the convolutional DNN is applied to
the phase component of the spectrogram it is not appropriate
to compute probabilistic estimates of phase (from the
overlapping windowed predictions) using parametric statistics.
Here, we use circular statistics to obtain probabilistic
estimates of phase computed using a convolutional DNN. We
illustrate our approach in the context of a typical cocktail
party source separation problem featuring complex
spectrograms.
We trained a complex convolutional DNN to separate and
re-synthesize speech in a two-speaker cocktail party scenario
[7]-[8]. The DNN was used as a complex convolutional deep
transform (DT) and trained to separate the speech of two
concurrent speakers. At the input layer, the DT DNN was
provided the complex spectrogram. At the output layer, the
DT DNN was trained to re-synthesize the respective complex
spectrograms of the two speakers. We then used the trained
complex convolutional DT to make probabilsitic predictions
about new concurrent speech mixtures of the same speakers.
Using objective source separation quality metrics, we
analyzed the separation quality. Our results are on a par with
equivalent convolutional-DNN probabilsitic binary-mask
source separation techniques [7] but offer slightly better sound
quality.
II. METHOD
We consider a typical simulated cocktail party listening
scenario featuring a male voice and a female voice speaking
concurrently. The speakers were each separately recorded (in
mono) reading from a story. The two speech signals were then
equalized in intensity and linearly summed (superposed) to
produce a competing voice scenario.
The speech signals were decimated to a sample rate of 4
kHz and transformed into spectrograms using the short-time
Fourier transform (STFT) with window size of 128 samples,
overlap interval of 1 sample and a Hanning window. This
provided complex spectrograms with 65 frequency bins. The
spectrograms computed from the first 2 minutes of the speech
signals were used as training data. A subsequent 10 seconds of
data were held back for later use in testing the separation of
the model.
Fig. 1. Probabilistic cocktail party source separation via complex
convolutional deep transform. The upper pair of spectrograms plot
a ~2-second excerpt from the original (test) speech audio. The single
(central) spectrogram plots the linear mixture of the two speech
signals. The lower pairs of spectrograms plot the respective source
signals separated and re-synthesized using complex CDT both with
and without output gain adaptation.
For training data, the mixture and component speech
spectrograms were cut up into windows (i.e., in time) of 20
samples. The windows overlaped at intervals of 10 samples.
Thus, for every 20-sample window, for training the model
there was mixture spectrogram matrix of size 65x20 samples
and a corresponding pair of male/female source spectrograms.
This gave approximately 50,000 training examples. For the
testing stage, 10 seconds of speech spectrogram was used at
overlap intervals of 1 sample, giving approximately 40,000
test frames (which would ultimately be applied in an
overlaping convolutional output stage). Prior to windowing,
the complex spectrogram data was separated into magnitude
and phase spectrograms respectively, and each spectrogram
was normalized to unit scale (i.e., both the magnitude and
phase data were mapped to the range [0,1]).
We used a feed-forward DNN of size 2600x2600x5200
units (65 x 20 x 2 = 2600, and 2600 x 2 = 5200). Each
spectrogram (magnitude and phase) window of size 65 x 20
was unpacked into a vector of length 1300, giving two
respective vectors of the same size. The DNN was configured
such that the input layer was a vector concatenated from the
mixture magnitude spectrogram (1300 samples) followed by
the mixture phase spectrogram (1300 samples), giving a total
vector of length 2600 (1300 x 2). The output layer was trained
to synthesize a vector featuring the sequential concatenation
of the magnitude and phase spectrograms of the respective
male and female speech component signals. This meant that
the DNN was trained to re-synthesize the respective
components into the respective concatenated locations in the
output layer (vector). The DNN employed the biased-sigmoid
activation function [9] throughout with zero bias for the
output layer. The DNN was trained using 500 full iterations of
stochastic gradient descent (SGD). Each iteration of SGD
featured a full sweep of the training data. Dropout was not
used in training.
DT Probabilistic Re-Synthesis. In the testing stage the
model was used as a feed-forward signal processing device
and the output layer activations were taken as synthetic output.
For the test data, there was an overlap interval of 1 sample.
This means that the test data described the speech spectrogram
in terms of a sliding window and the output of the model was
in the same sliding window format (i.e., it was a convolutional
model). Each frame of input mixture spectrogram was passed
through the model to produce the respective predictions of the
respective magnitude and phase spectrograms for the male and
female voice respectively.
An SGD trained autoencoder type DNN can feature
neurons in the output layer with some degree of invariant or
persistent activity [4]. To account for this activity, as in [4],
we included an ‘output gain adaptation’ stage, where the mean
activation across all test frames was subtracted from the
individual activations for each frame (i.e., there was no time
constant). These ‘adapted’ output layer predictions were then
accumulated as the sliding window, W, (size 20) moved by 1sample steps. Thus, 20 separate predictions were obtained for
each column of the output spectrograms. For each of the
male/female speakers, and for each of the respective
magnitude and phase spectrograms, this gave an output
distribution of spectrograms contained within a 3D matrix
indexed using time (t), frequency (f) and window index (w).
From the 3D magnitude matrices (M), the running average
magnitude spectrogram ( ) was calculated as;
=
,
1
, ,
(1)
From the 3D phase spectrogram matrices ( ) an equivalent
circular mean phase angle was computed as follows. The
phase spectrograms predicted in the output layer were first
remapped from the range [0,1] to the range [0, 2π] (not shown
here for convenience) and then transformed into a matrix of
unit vectors (R) in the 2D plane using the following elementwise matrix operation (subscript indices t,f,w are dropped for
convenience);
=
cos(
(
2 )
2 )
=
, ,
(2)
The circular sum ( ), over window size (W = 20), is computed
in an element-wise matrix operation;
,
(3)
Then, in a further element-wise matrix operation, the circular
mean angle matrix ( ̅ ) is computed by taking the fourquadrant inverse tangent (indices from here onwards are
dropped for convenience);
̅ = "#" 2( )
(4)
The respective estimated magnitude ( ) and phase ( ̅ )
spectrograms were then recombined into a complex
spectrogram ( $% ) with the following element-wise matrix
operation;
$% =
exp( ̅ )
(5)
The estimated complex spectrogram ( $% ) was then
subjected to an inverse STFT using an overlap-and-add
procedure. Separation quality for the resulting separated audio
(with respect to the original time domain audio signals) was
measured using the BSS-EVAL toolbox [10] and is quantified
in terms of signal-to-distortion ratio (SDR), signal-tointerference ratio (SIR) and signal-to-artefact ratio (SAR).
Separation quality was assessed after each iteration of SGD
training in order to evaluate the trajectory of performance
(with training) in each measure.
III. RESULTS
Fig. 1 plots spectrograms illustrating the stages of mixture
and separation for a brief excerpt (~2 seconds) from the test
data. The model had been trained for 500 iterations (i.e., N =
500). The top spectrograms plot the original male and female
speech audio. The single (central) spectrogram plots the linear
mixture, illustrating a large degree of overlap in this feature
space. Next (downwards) are plotted the spectrograms for the
complex convolutional DT probabilistic re-synthesized audio
where no output gain adaptation was employed. Finally, at the
bottom are plotted spectrograms representing the respective
complex convolutional DT probabilistic re-synthesized audio
featuring output gain adaptation.
Both sets of output spectrograms illustrate features which
closely resemble those of the original signals. By visual
inspection, the output gain adaptation results in less noise and
somewhat better definition of the features. Both output
spectrograms appear to have captured most of the bandwidth
of the original signals but the fricative (noise) components of
the originals have not been replicated very faithfully in either
case. Informal listening revealed that the respective separated
output audio was of good quality but was noisy (as Fig. 1
illustrates) in the case of the model that did not feature output
gain adaptation (this model was not tested further).
Fig. 2 plots the mean objective separation quality measures
(SDR, SIR, SAR), computed over the entire 10-second test
data (with respect to the original audio) and averaged across
the male and female voice signals, as a function of training
iteration (N) between the range of 1 and 500. All three
functions appear to be nearing convergence after around 300
iterations and all three functions are more or less monotonic.
Peak separation quality (around 500 iterations), averaged
across the two voices, is SDR: 8.1, SIR: 17.2, SAR: 8.9 dB.
This compares well with the equivalent DNN binary mask
approach reported previously (for identical data and
conditions), which achieved slightly worse artefact
performance (for an equivalent SIR of 17.2 – see Fig. 2 of [7]);
SDR: 7.5, SAR: 8.2 dB. This minor advantage is presumably
due to the ability of the present approach to employ phase
directly, where in the convolutional DNN binary mask
approach phase is taken from the mixture spectrogram. The
present results also compare reasonably well with the ideal
binary mask computed for the same test data (i.e., using a
mask computed from the source spectrograms), which
achieves SDR: 11.3, SIR: 21.6, SAR: 11.7. The present results
are also much better than the time-domain convolutional DT
probabilistic re-synthesis approach reported previously [11].
To some degree, it may be that the inclusion of both phase and
magnitude in the present approach provides some small
advantage in terms of sampling that may be broadly
equivalent to over-sampling [12].
We also note that, similar to the binary-mask based
convolutional DNN approach reported previously [7], the
present performance appears superior to earlier non-negative
matrix factorization (NMF) based approaches [5], [6], which
featured small scale DNN within the NMF pipeline. However,
the results of the NMF-based methods are not directly
comparable with the present results.
some insight into the neural processing of the auditory system
and its associated perceptual illusions and function.
ACKNOWLEDGMENT
AJRS was supported by grant EP/L027119/1 from the UK
Engineering and Physical Sciences Research Council
(EPSRC).
REFERENCES
Fig. 2. Complex convolutional DT: Separation quality as a
function of training iterations (N). Mean signal-to-interference
ratio (SIR), signal-to-distortion ratio (SDR) and signal-to-artefact
ratio (SAR) computed from the audio separated using complex
convolutional DT as a function of training iteration (N). The
measures were computed from the 10-second test audio using the
BSS-EVAL toolkit [10] and averaged across the two voices.
IV. DISCUSSION AND CONCLUSION
We have introduced a complex convolutional DT approach
to cocktail party source separation using spectrograms. Our
probabilistic approach features both parametric statistical
estimation of spectrogram magnitude and circular statistical
estimation of phase. The convolutional DNN was trained on
two minutes of the speech of two speakers and tested on 10
seconds of new speech from the same speakers. Separation
quality is similar to binary mask based convolutional DNN
aproaches but features slightly improved artefact performance.
Although the DNN employed here is of only 3 layers, if we
consider the degree of abstraction already provided by the
STFT and inverse STFT (giving an effective ‘depth’ of 5
layers of demodulation [9] and synthesis [4]) then it is not
surprising that the approach works as well as it does [7].
Furthermore, the ability of the model to operate with full
phase information, whilst retaining the 2D topographic
projection of the STFT, appears an advantage. It may also be
that performance in the present model is further enhanced by
the use of the Hanning window, which acts (similarly to oversampling) to mitigate aliasing that has been suggested to
affect DNN learning and performance [12]. More generally,
the circular statistical process of probabilistic synthesis
described here may be useful for more general probabilistic
synthesis at the level of the complex spectrogram.
At a more general level, if our model is interpreted as an
auditory model, the featured output gain adaptation appears
similar to that observed in the early auditory system [13]-[18].
Auditory gain adaptation is temporal and occurs on various
timescales up to tens of seconds and even minutes [13]-[18].
In principle, the output gain adaptation of the present model
may be interpreted as featuring a rectangular temporal
integration window of length 10 seconds (the entire test data).
Hence, the present model may be interpreted as demonstrating
that neuronal output gain adaptation may be useful in terms of
synthesis noise reduction. Hence, our findings may provide
[1] LeCun Y, Bengio Y (1995) “Convolutional networks for images,
speech, and time series”, The handbook of brain theory and neural
networks 3361: 310.
[2] Lawrence S, Giles CL, Tsoi AC, Back AD (1997) “Face Recognition:
A Convolutional Neural Network Approach”, Neural Networks, IEEE
Transactions on, 8: 98–113.
[3] Krizhevsky A, Sutskever I, Hinton GE (2012) “Imagenet classification
with deep convolutional neural networks”, In Advances in neural
information processing systems, pp. 1097-1105.
[4] Simpson AJR (2015) “Deep Transform: Error Correction via
Probabilistic Re-Synthesis”, arxiv.org abs/1502.04617.
[5] Grais E, Sen M, Erdogan H (2014) “Deep neural networks for single
channel source separation”, in Acoustics, Speech and Signal
Processing (ICASSP), 2014 IEEE Int. Conf. on, pp. 3734-3738.
[6] Huang PS, Kim M, Hasegawa-Johnson M, Smaragdis P, “Deep
learning for monaural speech separation”, in Acoustics, Speech and
Signal Processing (ICASSP), 2014 IEEE Int. Conf. on., pp. 1562-1566.
[7] Simpson AJR (2015) “Probabilistic Binary-Mask Cocktail-Party
Source Separation in a Convolutional Deep Neural Network”, arxiv.org
abs/1503.06962
[8] McDermott JH (2009) “The cocktail party problem”, Curr. Biol. 19:
R1024-R1027.
[9] Simpson AJR (2015) “Abstract Learning via Demodulation in a Deep
Neural Network”, arxiv.org abs/1502.04042
[10] Vincent E, Gribonval R, Févotte C (2006) “Performance measurement
in blind audio source separation”, IEEE Trans. on Audio, Speech and
Language Processing, 14:1462-1469.
[11] Simpson AJR (2015) “Deep Transform: Time-Domain Cocktail Party
Source Separation via Probabilistic Re-Synthesis”, arxiv.org
abs/1502.06046.
[12] Simpson AJR (2015) “Over-Sampling in a Deep Neural Network”,
arxiv.org abs/1502.03648.
[13] Dean I, Harper NS, McAlpine D. (2005) “Neural population coding of
sound level adapts to stimulus statistics”, Nat. Neurosci. 8, 1684 –
1689 (2005).
[14] Watkins PV, Barbour DL. (2008) “Specialized neuronal adaptation for
preserving input sensitivity”, Nat. Neurosci. 11, 1259-1261.
[15] Rabinowitz NC, Willmore DB, Schnupp JWH, King AJ (2011)
“Contrast gain control in auditory cortex”, Neuron. 70, 1178-1191.
[16] Ulanovsky N, Las L, Nelken I. (2003) “Processing of low-probability
sounds by cortical neurons”, Nat. Neurosci. 6, 391-398.
[17] Ulanovsky N, Las L, Farkas D, Nelken I. (2004) “Multiple time scales
of adaptation in auditory cortical neurons”, J. Neurosci. 24, 10440 –
10453.
[18] Simpson AJR, Harper NS, Reiss JD, McAlpine D (2014) “Selective
Adaptation to “Oddball” Sounds by the Human Auditory System”, J
Neurosci 34:1963-1969.
| 9cs.NE
|
Half-integral linkages in highly connected directed graphs∗
arXiv:1611.01004v1 [math.CO] 3 Nov 2016
Katherine Edwards†
Irene Muzi‡
Paul Wollan§
November 4, 2016
Abstract
We study the half-integral k-Directed Disjoint Paths Problem ( 12 kDDPP) in highly strongly connected
digraphs. The integral kDDPP is NP-complete even when restricted to instances where k = 2, and the
input graph is L-strongly connected, for any L ≥ 1. We show that when the integrality condition is relaxed
to allow each vertex to be used in two paths, the problem becomes efficiently solvable in highly connected
digraphs (even with k as part of the input). Specifically, we show that there is an absolute constant c such
that for each k ≥ 2 there exists L(k) such that 21 kDDPP is solvable in time O(|V (G)|c ) for a L(k)-strongly
connected directed graph G. As the function L(k) grows rather quickly, we also show that 12 kDDPP is
solvable in time O(|V (G)| f (k) ) in (36k3 + 2k)-strongly connected directed graphs. We also show that for
each ε < 1 deciding half-integral feasibility of kDDPP instances is NP-complete when k is given as part
of the input, even when restricted to graphs with strong connectivity εk.
1
Introduction
Let k ≥ 1 be a positive integer. An instance of a directed k-linkage problem is an ordered tuple (G, S, T ) where
G is a directed graph and S = (s1 , . . . , sk ) and T = (t1 , . . . ,tk ) are each ordered sets of k distinct vertices in G.
The instance is integrally feasible if there exist paths P1 , . . . , Pk such that Pi is a directed path from si to ti for
1 ≤ i ≤ k and the paths Pi are pairwise vertex disjoint. The paths P1 , . . . , Pk will be referred to as an integral
solution to the linkage problem.
The k-Directed Disjoint Paths Problem (kDDPP) takes as input an instance of a directed k-linkage problem.
If the problem is integrally feasible, we output an integral solution and otherwise, return that the problem is
not feasible. The kDDPP is notoriously difficult. The problem was shown to be NP-complete even under the
restriction that k = 2 by Fortune, Hopcroft and Wyllie [3].
In an attempt to make the kDDPP more tractable, Thomassen asked if the problem would be easier if we
assume the graph is highly connected. Define a separation in a directed graph G as a pair (A, B) with
A, B ⊆ V (G) such that A ∪ B = V (G) and where there does not exist an edge (u, v) with u ∈ A \ B and
v ∈ B \ A. The order of the separation (A, B) is |A ∩ B|. The separation is trivial if A ⊆ B or B ⊆ A. The graph
∗ Supported by the European Research Council under the European Unions Seventh Framework Programme (FP7/20072013)/ERC Grant Agreement no. 279558
† [email protected]
‡ Department of Computer Science, University of Rome, “La Sapienza”, Rome, Italy.
§ Department of Computer Science, University of Rome, “La Sapienza”, Rome, Italy. email: [email protected].
1
G is strongly k-connected if |V (G)| ≥ k + 1 and there does not exist a nontrivial separation of order at most
k − 1. Let k ≥ 1 and define a directed graph G to be integrally k-linked if every linkage problem (G, S, T ) is
integrally feasible. Thomassen conjectured [15] that there exists a function f such that every f (k)-strongly
connected digraph G is integrally k-linked. He later answered his own conjecture in the negative [16],
showing that no such function f (k) exists. Moreover, he also showed [16] for all L ≥ 1, the 2DDPP is
NP-complete even when restricted to problem instances where the graph is L-strongly connected.
In this article, we relax the kDDPP problem by requiring that a potential solution not use any vertex more
than twice. Define a directed k-linkage problem (G, S, T ) to be half-integrally feasible if S = (s1 , . . . , sk ), and
T = (t1 , . . . ,tk ) and there exist paths P1 , . . . , Pk such that:
• for all 1 ≤ i ≤ k, Pi is a directed path from si to ti , and
• for every vertex v ∈ V (G), v is contained in at most two distinct paths Pi .
The paths P1 , . . . , Pk form a half-integral solution.
The main result of this article is that the 21 kDDPP is polynomial time solvable (even with k as part of the
input) when the graph is sufficiently highly connected. Define a graph G to be half-integrally k-linked if
every k disjoint paths problem (G, S, T ) is half-integrally feasible.
Theorem 1. For all integers k ≥ 1, there exists a value L(k) such that every strongly L(k)-connected graph
is half-integrally k-linked. Moreover, there exists an absolute constant c such that given an instance (G, S, T )
of the 21 kDDPP where G is L(k)-connected, we can find a solution in time O(|V (G)|c ).
The assumption that G is highly connected in Theorem 1 cannot be omitted under the usual complexity
assumptions.
Theorem 2. For all ε < 1, it is NP-complete to determine whether a given kDDPP instance (G, S, T ) halfintegrally feasible, even under the assumption that G is εk-strongly connected.
The value for L(k) in Theorem 1 grows extremely quickly. However, when we fix k, we can still efficiently
solve the 12 kDDPP with a significantly weaker bound on the connectivity than that given in Theorem 1.
Theorem 3. There exists a function f satisfying the following. Let k ≥ 1 be a positive integer. Given a
k-linkage problem (G, S, T ) such that G is (36k3 + 2k)-strongly connected, we can determine if the problem
is half-integrally feasible and if so, output a half-integral solution, in time O(|V (G)| f (k) ).
Given that the kDDPP is NP-complete even in the case k = 2, previous work on the problem has focused on
various relaxations of the problem. Schrijver [13] showed that for fixed k, the kDDPP is polynomial time
solvable when the input graph is assumed to be planar. Later, Cygan et al [1] improved this result, showing
that the kDDPP is fixed parameter tractable with the assumption that the input graph is planar. In their recent
series of articles [8, 7, 9] leading to the breakthrough showing the grid theorem holds for directed graphs,
Kawarabayashi and Kreutzer and Kawarabayashi et al showed the following relaxation of the kDDPP can
be efficiently resolved for fixed k. They showed that there exists a polynomial algorithm which, given an
instance (G, S = (s1 , . . . , sk ), T = (t1 , . . . ,tk )) of the kDDPP, does one of the following:
• find directed paths Pi , 1 ≤ i ≤ k, such that Pi links si to ti and for every vertex v of G, v is in at most
four distinct Pi , or
2
• determine that no integral solution to (G, S, T ) exists.
In terms of hardness results, Slivkins [14] showed that the kDDPP is W [1]-complete even when restricted to
acyclic graphs. Kawarabayashi et al [7] announced that the proof of Slivkins result can be extended to show
that the 12 kDDPP is also W [1]-complete.
There are two primary steps in the proof of Theorem 1. First, we show that any highly connected graph
contains a large structure which we can use to connect up the appropriate pairs of vertices. The exact
structure we use is a bramble of depth two. A bramble is a set of pairwise touching, connected (strongly
connected) subgraphs; they are widely studied certificates of large tree-width both in directed and undirected
graphs. See Sections 2 and 3 for the exact definitions and further details. The existence of such a bramble of
depth two follows immediately from Kawarabayashi and Kreutzer’s proof of the grid theorem [5]; however,
the algorithm given in [5] only runs in polynomial time for fixed size of the bramble. We show in Section 4
that from appropriate assumptions which will hold both in the proof of Theorem 1 and Theorem 3, we are
able to find a large bramble of depth two in time O(nc ) for a graph on n vertices and some absolute constant
c.
The second main step in the proof of Theorem 1 is to show how we can use such a bramble of depth two to
find the desired solution to a given instance of the 12 kDDPP. Define a linkage to be a set of pairwise disjoint
paths. We show in Section 5 that given an instance (G, S, T ) and a large bramble B of depth two, we can find
a smaller, sub-bramble B 0 ⊆ B along with a linkage P of order k such that every element of P is a path
from an element of S to a distinct subgraph in B 0 . Moreover, the linkage P is internally disjoint from B 0 .
At the same time, we find a linkage Q from distinct subgraphs of B 0 to the vertices T . Thus, by linking the
appropriate endpoints of Q and P in the bramble B 0 , we are able to find the desired solution to (G, S, T ).
The fact that the bramble B 0 has depth two ensures that the solution we find uses each vertex at most twice.
This result is given as Theorem 15; the statement and proof are presented in Section 5.
Linking to a well-behaved structure (the bramble of depth two in the instance above) is a common technique
in disjoint path and cycle problems in undirected graphs. See [6, 12] for examples. The main contribution of
Theorem 15 is to extend the technique to directed graphs, and in particular, simultaneously find the linkage
from S to B 0 and the linkage Q from B 0 to T . This is made significantly more difficult in the directed case
by the directional nature of separations in directed graphs and the fact that there is no easy way to control
how the separations between S and B 0 and those between B 0 and T cross.
The proofs of Theorems 1 and 3 are given in Section 6 and the construction showing NP-completeness in
Theorem 2 is given in Section 7.
2
Directed tree-width
An arborescence is a directed graph R such that R has a vertex r0 , called the root of R, with the property that
for every vertex r ∈ V (R) there is a unique directed path from r0 to r. Thus every arborescence arises from
a tree by selecting a root and directing all edges away from the root. If r, r0 ∈ V (R) we write r0 > r if r0 6= r
and there exists a directed path in R from r to r0 . If (u, v) ∈ E(R) and r ∈ V (R), we write r > (u, v) if r > v or
r = v. Let G be a directed graph and Z ⊆ V (G). A set S ⊆ V (G) \ Z is Z-normal if there is no directed walk
in G − Z with the first and last vertex in S which also contains a vertex of V (G) \ (S ∪ Z). Note that every
Z-normal set is a union of strongly connected components of G − Z.
Let G be a directed graph. A tree decomposition of G is a triple (R, β , γ), where R is an arborescence,
3
β : V (R) → 2V (G) and γ : E(R) → 2V (G) are functions such that:
1. {β (r) : r ∈ V (R)} is a partition of V (G) into non-empty sets and
2. if e ∈ E(R), then {β (r) : r ∈ V (R), r > e} is γ(e)-normal.
The sets β (r) are called the bags of the decomposition and the sets γ(e) are called the guards of the decomposition. For any r ∈ V (R), we define Γ(r) := β (r) ∪ {γ(e) : e incident to r}. The width of (R, β , γ) is the
smallest integer w such that |Γ(r)| ≤ w + 1 for all r ∈ V (R). The directed tree-width of G is the minimum
width of a tree decomposition of G.
Johnson, Robertson, Seymour, and Thomas showed that if we assume k and w are fixed positive integers,
then we can efficiently resolve the kDDPP when restricted to directed graphs of tree-width at most w [4].
Theorem 4 ([4], Theorem 4.8). For all t ≥ 1, there exists a function f satisfying the following. Let k ≥ 1, and
let (G, S, T ) be an k-linkage problem such that the directed tree-width of G is at most t. Then we determine
if (G, S, T ) is integrally feasible and if so, output an integral solution, in time O(|V (G)| f (k) ).
A simple construction shows that the same result holds to efficiently resolve k-linkage problems halfintegrally when k and the tree-width of the graph are fixed. We first define the following operation. To
double a vertex v in a directed graph G, we create a new vertex v0 and add the edges (u, v0 ) for all edges
(u, v) ∈ E(G), the edges (v0 , u) for all edges (v, u) ∈ E(G) and the edges (v, v0 ) and (v0 , v).
Corollary 5. For all t ≥ 1, there exists a function f satisfying the following. Let k ≥ 1, and let (G, S, T ) be
an instance of a k-linkage problem such that the directed tree-width of G is at most t. Given in input (G, S, T )
and a directed tree-decomposition of G of width at most t, we can determine if the problem is half-integrally
feasible and if so, output a half-integral solution, in time O(|V (G)| f (k) ).
Proof. Fix w ≥ 1 to be a positive integer. Let (G, S = (s1 , . . . , sk ), T = (t1 , . . . ,tk )) be an instance of a klinkage problem where G has tree-width at most w. Let G0 be the directed graph obtained by doubling every
vertex v ∈ V (G). Define the k-linkage problem (G0 , S∗ = (s∗1 , . . . , s∗k ), T ∗ = (t1∗ , . . . ,tk∗ )) by letting s∗i = si
and ti∗ = ti0 for 1 ≤ i ≤ k. Thus, (G, S, T ) is half-integrally feasible if and only if (G0 , S∗ , T ∗ ) is integrally
feasible. Moreover, any integral solution to (G0 , S∗ , T ∗ ) can be easily converted to a half-integral solution for
the original problem (G, S, T ).
Let (R, β , γ) be a tree decomposition of G of width w. Observe that (R, β 0 , γ 0 ) defined by β 0 (r) = {{v, v0 } :
v ∈ β (r)} and γ 0 (r) = {{v, v0 } : v ∈ γ(r)} yields a tree decomposition of G0 of width at most 2w. Thus,
by Theorem 4, we can determine if (G0 , S∗ = (s∗1 , . . . , s∗k ), T ∗ = (t1∗ , . . . ,tk∗ )) is integrally feasible and find an
solution when it is, in polynomial time assuming k and w are fixed, proving the claim.
3
Certificates for large directed tree-width
A bramble in a directed graph G is a set B of strongly connected subgraphs B ⊆ G such that if B, B0 ∈ B,
then V (B) ∩V (B0 ) 6= 0/ or there exists edges e, e0 ∈ E(G) such that e links B to B0 and e0 links B0 to B. A cover
of B is a set X ⊆ V (G) such that V (B) ∩ X 6= 0/ for all B ∈ B. The order of a bramble is the minimum size of
a cover of B. The bramble number, denoted bn(G), is the maximum order of a bramble in G. The elements
of a bramble are called bags, and the size of a bramble, denoted |B|, is the number of bags it contains.
4
The bramble number of a directed graph gives a good approximation of the tree-width, as seen by the following theorem of [11] as formulated by [9].
Theorem 6 ([11],[9]). There exist constants c, c0 such that for all directed graphs G, it holds that
bn(G) ≤ c · tw(G) ≤ c0 · bn(G).
Johnson, Robertson, Seymour, and Thomas showed one can efficiently either find a large bramble in a directed graph or explicitly find a directed tree-decomposition. Note that the result is not stated algorithmically,
but that the algorithm follows from the construction in the proof. Additionally, they looked at an alternate
certificate of large tree-width, namely havens, but a haven of order 2t immediately gives a bramble of order
t by the definitions.
Theorem 7 ([4], 3.3). There exist constants c1 , c2 such that for all t and directed graphs G, we can algorithmically find in time |V (G)|c1 either a bramble in G of order t or a tree-decomposition of G of order at most
c2t. Moreover, if we find the bramble, it has at most |V (G)|2t elements.
A long open question of Johnson, Robertson, Seymour, and Thomas [4] was whether sufficiently large treewidth in a directed graph would force the presence of a large directed grid minor. Let r ≥ 2 be a positive
integer. The directed r-grid Jr is the graph defined as follows. Let C1 , . . . ,Cr be directed cycles of length
2r. Let the vertices of Ci be labeled vi1 , . . . , vi2r for 1 ≤ i ≤ r. For 1 ≤ i ≤ 2r, i odd, let Pi be the directed
1
path v1i , v2i , . . . , vri . For 1 ≤ i ≤ 2r, i even, let Pi be the directed path vri , vr−1
i , . . . , vi . The directed grid
Sr
S2r
Jr = 1 Ci ∪ 1 Pi .
In a major recent breakthrough, Kreutzer and Kawarabayashi have confirmed the conjecture of Johnson et
al.
Theorem 8 ([9]). There is a function f : N → N such that given any directed graph and any fixed constant
k, in polynomial time, we can obtain either
1. a cylindrical grid of order k as a butterfly minor, or
2. a directed tree decomposition of width at most f (k).
For our purposes, we will use brambles when attempting to solve the 21 kDDPP. However, in order to ensure
that the paths we find don’t use any vertex more than twice, we require the bramble to have depth two. Define
the depth of a bramble B = {B1 , . . . , Bt } in a directed graph G to be the maxv∈V (G) |{i : v ∈ V (Bi )}|; in other
words, a bramble has depth at most k for some positive integer k if no vertex is contained in more than k
distinct subgraphs in the bramble. Note that if B has depth k and size t, then it has order at least dt/ke.
Lemma 9. For all t ≥ 2, the directed t-grid contains a model of a bramble B of size t and depth two.
Proof. Let the cycles C1 , . . . ,Ck , paths P1 , . . . , P2k , and vertex labels vij , 1 ≤ i ≤ 2k, 1 ≤ j ≤ k, be as in the
definition of the cylindrical grid. For every l, 1 ≤ l ≤ k, and for every i, 1 ≤ i ≤ 2k, let Pi (l) be the subpath of
Pi with endpoints v1i and vli . For 1 ≤ i ≤ k − 1, let Ci0 be the (unique) cycle in Ci ∪C1 ∪ P2i−1 (i) ∪ P2i (i) which
contains all the vertices v1j , 1 ≤ j ≤ 2k. Let Ck0 = Ck . The cycles C10 , . . . ,Ck0 form a bramble of depth two and
size k, as desired.
5
4
Finding a bramble of depth two
In this section, we show that given what we call a sufficiently large well-linked set of vertices in a directed
graph, we are able to efficiently find a large bramble of depth two. The argument in many ways follows
Diestel et al’s proof of Robertson and Seymour’s grid theorem (see [2] for the proof) for undirected graphs.
We begin with a collection of disjoint linkages and show that in each of the linkages we can find a sublinkage
which are pairwise disjoint. We will need two classic results in graph theory, namely Ramsey and Menger’s
theorems.
Theorem 10 (Menger’s theorem [10]). Let G be a directed graph and S, T subsets of V (G). The maximum
number of vertex-disjoint S − T paths equals the minimum order of a separation which separates S from T .
Moreover, there exists an algorithm to find a maximum set of vertex-disjoint S − T paths and a minimum
order separation in time O(|V (G)|2 ).
Theorem 11 (Ramsey’s theorem). Let r,t be positive integers. For every (improper) two coloring of the
edges of the undirected clique K2r+t by red and blue, there exists either a subgraph Kr with every edge is
colored red or a subgraph Kt with every edge colored blue. Moreover, the desired Kr or Kt subgraph can be
found in time O(kc ) for some absolute constant c.
We first give two preparatory lemmas before presenting the main result of this section.
Lemma 12. Let G be a digraph on n vertices and let t, k ≥ 2 be positive integers. Let T = (10k)22(t+k) .
Let X = {x1 , x2 , . . . , xT } and Y = {y1 , y2 , . . . , yT } be two disjoint sets of vertices of order T . Let P =
{P1 , P2 , . . . , PT } be a linkage from X to Y and Q = {Q1 , Q2 , . . . , QT } a linkage from Y to X, each of size
T . Assume each Pi has endpoints xi and yi and assume that there exists a permutation π of [T ] such that Qi
has endpoints yi and xπ(i) . Then one of the following holds:
1. there exist B1 , . . . , Bt subgraphs of G forming a bramble of size t and depth two, or
2. there exists a subset J ⊆ [T ] with |J| = k such that the subgraphs Pj ∪ Q j ∪ Pπ( j) ∪ Qπ( j) ∪ Pπ(π( j)) are
pairwise disjoint for all j ∈ J.
Moreover, given G, P, and Q in input, we can find either B1 , . . . , Bt or J in time O(nc ) for some absolute
constant c.
Proof. Define an auxiliary undirected bipartite graph H with vertex set x1 , x2 , . . . , xT , y1 , y2 , . . . , yT and edges
xi yi and xπ(i) yi for all 1 ≤ i ≤ T . Thus, H is the union of two perfect matchings and each component is either
S
S
→
−
a cycle or a single edge. For any induced subgraph C of H, let C = i:xi yi ∈E(C) Pi ∪ i:xπ(i) yi ∈E(C) Qi . Note
→
−
that if C is a connected component of H, then C is a strongly connected subgraph of G.
Assume, as a case, that there are at least T 0 = 2k+t distinct components C1 , . . . ,CT 0 . By Theorem 11, there
exists I ⊆ [T 0 ] such that one of the following holds:
→
−
• |I| = t and the subgraphs Ci pairwise intersect for i ∈ I, or
→
−
• |I| = k and the subgraphs Ci are pairwise disjoint for i ∈ I.
6
Moreover, we can find I in time O(nc ).
→
−
In the first case, we claim that {Ci : i ∈ I} form a bramble of depth two; clearly by construction, the sets
→
−
{Ci : i ∈ I} form a bramble. To see that it has depth two, observe that Q and P are both linkages and given
→
−
the fact that for any vertex v such that v is contained in V (Ci ), it must be the case that v is contained in some
→
−
→
−
→
−
P ∈ P or Q ∈ Q which is a subpath of Ci . As the set elements of P and Q forming Ci and C j are disjoint
→
−
→
−
for all i 6= j, at most two Ci and C j can intersect in the vertex v. Thus the bramble is of depth two.
→
−
In the second case, for every i ∈ I, we fix j = j(i) such that Pj is contained in Ci . It follows that J = { j(i) :
i ∈ I} satisfies outcome 2 in the statement of the lemma.
We conclude that there are at most 2k+t distinct components, and thus, there exists a component C of size at
least (10k) · 2k+t . Let P1 , . . . , P2k+t be disjoint paths in C, each of length at least 10k.
Assume that there exists an index i, 1 ≤ i ≤ 2k+t such that for all edges e and f in E(Pi ) at distance at least
→
−
−
six, →
e ∩ f = 0.
/ Fix pairwise disjoint subpaths P10 , . . . , Pk0 of P, such that
−
• for 1 ≤ i ≤ k, the path Pi0 has length five and contains exactly three edges e such that →
e ∈ P;
• for 1 ≤ i < i0 ≤ k, Pi0 and Pi00 are at distance at least five.
By our assumption on Pi , the fact that each Pi0 starts and ends with an edge corresponding to a path in P,
→
− →
−
and the fact that the elements of P are pairwise disjoint, it follows that for all 1 ≤ i < i0 ≤ k, Pi0 ∩ Pi00 = 0,
/ as
in the outcome 2 of the lemma.
We conclude that for all 1 ≤ i ≤ 2k+t , the subgraph Pi has two edges at distance at least six such that their
corresponding paths in P ∪ Q intersect. It follows that there exists a five edge path Pi0 of Pi containing three
→
−
−
edges e such that →
e ∈ P and a strongly connected subgraph Hi of G containing Pi0 . By the same Ramsey
argument as above applied to the subgraphs H1 , . . . , H2k+t , we see that one of the desired outcomes holds for
G.
A subset X ⊆ V (G) of vertices of a directed graph G is well-linked if for any pair of subsets U1 ,U2 ⊆ X with
|U1 | = |U2 |, there exists a directed U1 to U2 linkage of order |U1 |.
Lemma 13. Let X be a well-linked set in a directed graph G on n vertices. Let k,t ≥ 2 be positive integers.
Let X1 , X2 ,Y1 ,Y2 ⊆ X be pairwise disjoint subsets of X. Let P be a linkage of order 2k 2T from X1 to X2 where
T = (10t)24t . Let R be a linkage of order 2k 2T from Y1 to Y2 . Then one of the following holds:
1. there exist P 0 ⊂ P and R 0 ⊂ R with |P 0 | = |R 0 | = k such that for all P ∈ P 0 and R ∈ R 0 , it holds
that P ∩ R = 0,
/ or
2. there exists B1 , . . . , Bt forming a bramble of size t and depth two.
Moreover, we can find the linkages satisfying outcome 1 or the bramble in 2 in time O(nc ) for some absolute
constant c.
Proof. By Theorem 11, we may assume that there exist P 0 ⊆ P and R 0 ⊆ R, each of order T such that
every element of P 0 intersects every element of R 0 . By the definition of well-linked set, there exists a
7
linkage Q X from X2 ∩V (P 0 ) to X1 ∩V (P 0 ) and similarly, a linkage QY from Y2 ∩V (R 0 ) to Y1 ∩V (R 0 ). By
Theorem 10, we can find the linkages Q X and QY in time O(n2 ). Label the elements of R 0 as R1 , . . . , RT
and the elements of P 0 as P1 , . . . , PT . Let πX and πY be two permutations of [T ] such that we can label the
elements of Q X and QY as QX1 , . . . , QXT and QY1 , . . . , QYT such that QXi has a common endpoint with Pi in X2
and a common endpoint with PπX (i) in X1 . Similarly, QYi has a common endpoint with Ri in Y2 and a common
endpoint with RπY (i) in Y1 .
Apply Lemma 12 to the linkages P1 , . . . , PT and QX1 , . . . , QXT . We may assume that we get outcome 2 in
the lemma. Without loss of generality, we may assume that πX (i) = i + 1 for 1 ≤ i ≤ 3k − 1 and that
the subgraphs P3 j+1 ∪ QX3 j+1 ∪ P3 j+2 ∪ QX3 j+2 ∪ P3 j+3 are pairwise disjoint for 0 ≤ j ≤ k − 1. Similarly, by
applying Lemma 12 to R1 , . . . , RT and QY1 , . . . , QYT that πY (i) = i + 1 for 1 ≤ i ≤ 3k − 1 and that the subgraphs
R3 j+1 ∪ QY3 j+1 ∪ R3 j+2 ∪ QY3 j+2 ∪ R3 j+3 are pairwise disjoint for 0 ≤ j ≤ k − 1.
Since for all 0 ≤ j ≤ k − 1, the paths P3 j+3 intersects R3 j+1 and R3 j+3 intersects P3 j+1 , we conclude that the
subgraph
P3 j+1 ∪ Q3 j+1 ∪ P3 j+2 ∪ Q3 j+2 ∪ P3 j+3 ∪ R3 j+1 ∪ QY3 j+1 ∪ R3 j+2 ∪ QY3 j+2 ∪ R3 j+3
contains a strongly connected subgraph B j which contains both P3 j+2 and R3 j+2 . Since every Pi intersects
every Ri0 , we have that {B0 , . . . , Bk−1 } forms a bramble of size k and depth two, as required.
We now show the main result of the section which is that given a sufficiently large well-linked set, we can
efficiently find a large bramble of depth two.
Theorem 14. There exists a function f which satisfies the following. Let G be a directed graph on n vertices
and t ≥ 1 a positive integer. Let P be a directed path and X ⊆ V (P) a well-linked set with |X| ≥ f (t). Then
G contains a bramble B = B1 , . . . , Bt of depth two. Moreover, given G, P, and X in input, we can find B in
time O(nc ) for some absolute constant c.
4t
Proof. For a given k ≥ 1, let ft (k) = 2k 2(10t)2 . We will use the notation ftl (k) for the function ft iterated
l times, beginning with input k, i.e. ft ( ft ( ft (. . . ( ft (k)) . . . ))). Let T = 22t . Fix pairwise disjoint subsets
Xiin , Xiout for 1 ≤ i ≤ T 2 in X such that there exist subpaths P1 , . . . , PT 2 of P which satisfy the following:
• for all 1 ≤ i < j ≤ T 2 , Pi ∩ Pj = 0;
/
• for all i, Xiin ∪ Xiout ⊆ V (Pi ) and traversing the directed path Pi , the vertices of Xiin occur before the
vertices of Xiout .
Moreover, we pick Xiin , Xiout , 1 ≤ i ≤ T 2 such that
8
|Xiin |, |Xiout | ≥ ftT (2).
8
By assuming that the function f in the statement of the theorem satisfies f (t) ≥ T 4 ftT (2), we see that such
Xiin , Xiout exist.
8
For all 1 ≤ i ≤ T 2 , 1 ≤ j ≤ T 2 , i 6= j, fix a directed linkage Q(i, j) from Xiout to X jin of order ftT (1). By
Theorem 10, we can find such linkages Q(i, j) in time O(T 4 n2 ). Fix a enumeration of A := {((i, j), (i0 , j0 )) :
1 ≤ i, i0 , j, j0 ≤ T 2 , i 6= j, i0 6= j0 } and let α = |A| = T 8 − 2T 6 + T 4 . For 0 ≤ l ≤ α, we define linkages
8
Q(i, j)l ⊆ Q(i, j) of order ftT −l (1) as follows. Let Q(i, j)0 = Q(i, j). For l = 1, . . . , α, let ((i, j), (i0 , j0 )) be
the l th pair in A. Apply Lemma 13 to the linkages Q(i, j)l−1 and Q(i0 , j0 )l−1 . We may assume that there exist
8
8
disjoint sublinkages of Q(i, j)l−1 and Q(i0 , j0 )l−1 , each of size ftT −l (1). Call them Q(i, j)l and Q(i0 , j0 )l ,
respectively. For (i00 , j00 ) distinct from (i, j) and (i0 , j0 ), fix Q(i00 , j00 )l to be an arbitrary subset of Q(i00 , j00 )l−1
8
of order ftT −l (1). We fix Q(i, j) to be an element of the linkage Q(i, j)α . By construction, the paths Q(i, j)
and Q(i0 , j0 ) are disjoint if (i, j) 6= (i0 , j0 ).
Define strongly connected subgraphs Ci and Ri for 1 ≤ i ≤ T as follows. Fix i, 1 ≤ i ≤ T . The subgraph
S −1
Ci = Tj=0
Q(i + jT, i + ( j + 1)T ) along with the subpath of Pi+ jT linking endpoints of Q(i + ( j − 1)T, i + jT )
and Q(i + jT, i + ( j + 1)T ) for 0 ≤ j ≤ T − 1 where the values are taken modulo T 2 . Similarly, we define
S −1
Ri = Tj=1
Q(iT + j, iT + j + 1) ∪ Q(iT + T, iT + 1) along with the subpaths of PiT + j linking the endpoints
of Q(iT + j − 1, iT + j) and Q(iT + j, iT + j + 1) for 2 ≤ j ≤ T − 1 along with the analogous subpaths of
PiT +1 and PiT +T . If we think of the paths Pi laid out in a T × T grid, the subgraphs C j are the natural strongly
connected graphs formed by following the Q(i0 , j0 ) through a column of the grid, and the R j are the strongly
connected graphs formed by the rows.
Every vertex in one of the subgraphs Ci is either in a path Q(i0 , j0 ) or Pl 0 ; moreover, the subset of Pl 0 contained
in Ci are disjoint from those contained in Ci00 for i0 6= i00 . Thus, it is possible that two Ci and Ci0 intersect in a
vertex, but not for three distinct Ci , Ci0 and Ci00 to all intersect in a common vertex.
By Theorem 11, there exists I ⊆ [T ] of size t such that for i, i0 ∈ I, either Ci intersect or they are pairwise
disjoint. By the above observation, if they are pairwise intersecting, then {Ci : i ∈ I} forms a bramble of size
t and depth two. Thus, we may assume that Ci ∩Ci0 = 0/ for all distinct i, i0 ∈ I. Similarly, there exists J ⊆ [T ],
|J| = t such that for all distinct j, j0 ∈ J, R j ∩ R j0 = 0.
/ Without loss of generality, assume that I = J = [t]. The
set {Ci ∪ Ri : i ∈ [t]} then forms a bramble of size t and depth two, completing the proof.
5
Linking in a bramble of depth two
The main result of this section is the following which shows that if we have a sufficiently large bramble of
depth two, we can use it to efficiently resolve a given instance of the 12 kDDPP under a modest assumption
on the connectivity of the graph.
Theorem 15. For all k ≥ 1, there exists a positive integer t such that if G is a (36k3 + 2k)-strongly connected
directed graph, and G contains a bramble B of depth two and size t, then for every k-linkage problem
instance (G, S, T ) is half-integrally feasible. Moreover, given (G, S, T ) and the bags of B, we can find a
solution in time O(k4 n2 ).
We begin with some notation. Recall that the doubling of a vertex in a directed graph was defined in Section
2. To contract a set of vertices U inducing a strongly connected subgraph of G is to delete U and create a
new vertex v, then add edges (w, v) for all edges (w, u) ∈ E(G) with u ∈ U, w ∈
/ U and edges (v, w) for all
edges (v, u) ∈ E(G) with u ∈ U, w ∈
/ U.
Let B be a depth two bramble in a directed graph G and B1 ⊆ B. Define the graph G(B1 ; B) as follows:
First, let G0 be the graph obtained from G by doubling every vertex belonging to two bags of B and to
at least one bag of B1 . For each such vertex v, denote its double by v0 . Let B 0 be the collection of |B1 |
subsets of V (G0 ) obtained from B1 by replacing each vertex v belonging to a bag of B with v0 in exactly
one of the bags it belongs to. Thus, the elements of B 0 are pairwise disjoint and each induces a strongly
connected subgraph of G0 , so B 0 is a depth 1 bramble in G0 . Let G(B1 ; B) be the graph obtained from G0
by contracting each element of B 0 . Denote by KB1 the set of contracted vertices in G(B1 ; B); note that
the vertices of KB1 form a bidirected clique. Observe that every double of a vertex of G0 gets contracted, so
9
V (G(B1 ; B)) \ KB1 ⊆ V (G). For a vertex v ∈ KB1 , we write im(v) for the bag of B1 corresponding to the
vertices contracted to v. We stress that each im(v) is a bag of B1 ; in particular im(v) ⊆ V (G).
Let S, T be disjoint subsets of the vertices of a directed graph G. A separation (A, B) separates S from T if
S ⊆ A and T ⊆ B. The separation (A, B) properly separates S from T if S \ B and T \ A are both nonempty.
For a positive integer α, we say S is α-connected to T if every separation separating S from T has order at
least α.
Let G be a directed graph, and B 0 ⊆ B be brambles of depth two. Let X ⊆ V (G(B 0 ; B)) \ KB0 . We say a
X − KB0 or KB0 − X linkage P1 , . . . , P|X| is B-minimal if none of the paths contains internally a vertex in KB0
or in im(v) for some v ∈ KB \ ∪i Pi .
We now give a quick outline of how the proof will proceed. Let us denote S = (s1 , . . . , sk ) and T = (t1 , . . . ,tk ).
Our approach to proving half-integral feasability is in two steps. We find three sets of paths, one set of k
paths linking S to the bramble B, another set linking B to T , and a third linking the appropriate ends of
paths in the first two sets to each other inside of B. To get the first two sets of paths, we take advantage of the
high connectivity of the graph. Linking half-integrally inside of the bramble is easy, and its structure allows
us to link any pairs of vertices we like half-integrally. We need the union of the three sets of paths to form
a half-integral solution, so we will choose the first and second sets each to be (almost) vertex-disjoint, and
to intersect the bramble B in a very limited way. The third set of paths will be half-integral and completely
contained in B.
The underlying idea behind our approach to finding the first two sets of paths is to contract each bag of the
bramble (after doubling vertices in two bags) and try to apply Menger’s theorem. In trying to do this, some
issues arise. First, we want the ends of all 2k paths to belong to distinct bags of B. More concerningly,
contracting the bags of the bramble may destroy the connectivity between the bramble and the terminals S
and T . We solve this by throwing away a bounded number of bags from the bramble until we are left with
a sub-bramble that is highly connected to S and from T . In Subsection 5.1, we will show how to find the
first two sets of paths (Lemma 16), modulo finding the sub-bramble (Lemma 17), and the third set of paths
(Lemma 18). Then we show how to put these pieces together to prove Theorem 15. In Subsection 5.2 we
prove Lemma 17.
5.1
Linking into and inside of a depth two bramble
Lemma 16. Let G be a (36k3 + 2k)-strongly connected directed graph and B be a bramble of depth two
and size > 188k3 in G. Let (G, S = (s1 , . . . , sk ), T = (t1 , . . . ,tk )) be a k-linkage problem instance. Then we
can find paths P1s , . . . , Pks , P1t , . . . , Pkt and B 0 ⊆ B satisfying the following:
A1 : For each i, Pis is a directed path from si to some vertex s0i , and Pit is a directed path from some vertex ti0
to ti .
A2 : The vertices s01 , . . . , s0k ,t10 , . . . ,tk0 belong to distinct bags of B 0 , say Bs1 , . . . , Bsk , Bt1 , . . . , Btk , respectively.
A3 : Every vertex belongs to at most two of P1s , . . . , Pks , and if a vertex v does belong to two paths, say Pis
and Pjs (i 6= j), then v = s0i or v = s0j .
A4 : Similarly, every vertex belongs to at most two of P1t , . . . , Pkt , and if a vertex v does belong to two paths,
say Pit and Ptj (i 6= j), then v = ti0 or v = t 0j .
A5 : For each i, the internal vertices of Pis and of Pit belong to at most one bag of B 0 .
10
A6 : For each i, j, ` all distinct , Pis ∩ Ptj ∩ (Bs` ∪ Bt` ) = 0.
/
A7 : Every vertex belongs to at most two of P1s , . . . , Pks , P1t , . . . , Pkt .
Moreover, given the bags of B, we can find the paths P1s , . . . , Pks , P1t , . . . , Pkt in time O(k4 n2 ).
We will prove the following lemma as an intermediate step to Lemma 17 in Section 5.2.
Lemma 17. Let G be a (36k3 + 2k)-strongly connected directed graph and B be a bramble of depth two and
size > 188k3 in G. Let (G, S, T ) be a k-linkage problem instance. Assume B is disjoint from {si ,ti ; 1 ≤ i ≤ k}.
Then there exist brambles BS and BT with BT ⊆ BS ⊆ B such that S is (36k3 + 2k)-connected to KBS in
G(BS ; B) and T is 3k-connected to KBT in G(BT ; BS ). Also |BS | − |BT | < 36k3 . Moreover we can find
BS and BT in time O(k4 n2 ).
But first, let’s see how Lemma 17 implies Lemma 16.
Proof of Lemma 16. Consider the brambles BS and BT given by Lemma 17. Denote by W the vertices in
G(BT ; B) that belong to exactly one bag in BT and to two bags in BS .
Claim. There exist k vertex-disjoint paths P1 , . . . , Pk in G(BT ; B) \ W where Pi links si to vi , for some
vi ∈ KBT .
Suppose not; then by Menger’s theorem there exists a separation (A, B) of order < k in G(BT ; B) \ W
separating S from KBT . But then consider the following separation in G(BS ). Let
A0 = (A ∩V (G(BS ; B))) ∪ {v ∈ KBS : im(v) ∩ A 6= 0}
/ ∪ (KBS \ KBT )
and
B0 = (B ∩V (G(BS ; B))) ∪ {v ∈ KBS : im(v) ∩ B 6= 0}
/ ∪ (KBS \ KBT ).
Intuitively, (A0 , B0 ) is the separation (A, B) viewed in the graph G(BS ; B), plus we add the vertices of
KBS \ KBT to each side. It’s easy to check that (A0 , B0 ) is a separation in G(BS ; B), since every vertex in
V (G(BT ; B)) \V (G(BS ; B)) belongs to im(v) for some v ∈ KBS . Also, we have |A0 ∩ B0 | ≤ 2|A ∩ B| + 36k3
because every vertex belongs to at most two bags of BS and every vertex in W belongs to one bag of BS \BT .
But this contradicts Lemma 17 and proves the claim.
Choose the paths P1 , . . . , Pk so that they are BT -minimal in G(BT ; B). Let us now view these as paths in the
original graph G: Since V (G(BT ; B)) \ KBT ⊆ V (G), each vertex in Pi except vi is a vertex of G, for each
1 ≤ i ≤ k. So choose s0i ∈ im(vi ) such that there exists an edge from the second to last vertex of Pi to s0i . Then
let Pis be the path obtained from Pi by replacing vi with s0i . Notice that Pis is a path in G. The paths P1s , . . . , Pks
are internally disjoint, so they satisfy A3 .
Claim. There exist vertex-disjoint paths Q1 , . . . , Qk in G(BT ; BS ) \ {v1 , . . . , vk , s01 , . . . , s0k } where Qi links wi
to ti for some wi ∈ KBT . Moreover, the vertices v1 , . . . , vk , w1 , . . . , wk are distinct.
Suppose not; then by Menger’s theorem, in the graph G(BT ; BS ) \ {v1 , . . . , vk , s01 , . . . , s0k } there is a separation (A, B) of order < k properly separating KBT from T . But then (A ∪ {v1 , . . . , vk , s01 , . . . , s0k }, B ∪
{v1 , . . . , vk , s01 , . . . , s0k }) has order < 3k and properly separates KBT from T in G(BT ; BS ), contradicting
Lemma 17. This proves the claim.
11
We may also choose the paths Q1 , . . . , Qk to be BT -minimal in G(BT ; BS ). Viewing these paths as paths in
G as above (symmetrically), we obtain paths P1t , . . . , Pkt , with Pit joining ti0 to ti . These paths satisfy A4 .
Let B 0 = {im(v) : v ∈ {v1 , . . . , vk , w1 , . . . , wk }}. For each i, set Bsi = im(vi ) and Bti = im(wi ). We now check
that the paths P1s , . . . , Pks , P1t , . . . , Pkt satisfy the seven assertions in the lemma statement. A1 , A2 , A3 and A4
have already been established.
To see that A5 holds, note that each of P1 , . . . , Pk is internally disjoint from KBT in G(BT ; B). Similarly, the
Q1 , . . . , Qk paths are internally disjoint from KBT in G(BT ; BS ). Moreover, by the definition of G(BT ; B)
and G(BT ; BS ), every vertex not in KBT in either of those graphs belongs to at most one bag of BT and
therefore to at most one bag of B 0 . It follows that for each i, each internal vertex of Pis and Pit belongs to at
most one bag of B 0 , proving A5 .
To see A6 , let 1 ≤ i, j, ` ≤ k be distinct. Suppose for contradiction that some vertex v belongs to Pis ∩ Ptj ∩
(Bs` ∪ Bt` ). If v is an internal vertex of either Pis or Ptj then v belongs to only one bag of B 0 by A5 . Also, if
v = s0i or t 0j then v belongs to two bags of B 0 . We deduce that v is an internal vertex of both Pis and Ptj . Since
we found Pi in the graph G(BT ; B) \W , we know v ∈
/ W so v belongs to one bag in BT and one bag of BS .
But we found Q j in the graph G(BT ; BS ), so v belongs to one bag of BT and two bags of BS . This is a
contradiction, proving A6 .
Finally, let us check A7 . Suppose for contradiction’s sake that some vertex v ∈ V (G) belongs to three paths.
By A3 and A4 , we must have v ∈ Pis ∩ Pjs ∩ P`t or v ∈ Pit ∩ Ptj ∩ P`s for some 1 ≤ i, j, ` ≤ k. If v ∈ Pis ∩ Pjs ∩ P`t ,
then by A3 we may assume without loss of generality that v = s0i . But the path Q` was found in a graph not
containing vi or s0i , so we must have s0i ∈ Bt` ∩ Bsi . Since B 0 is depth two, v ∈
/ Bsj so v is an internal vertex
s
t
t
s
of Pj , contradicting A5 . If v ∈ Pi ∩ Pj ∩ P` , then without loss of generality v = ti0 ∈ Bti = im(wi ). By the
BT -minimality of P1 , . . . , Pk , v cannot be an internal vertex of P` so we have v = s0` ∈ Bs` . Since v belongs to
two bags, A5 implies that v = t 0j , a contradiction.
It remains to check that we can indeed find these paths in time O(k4 n2 ). Indeed finding the brambles BS and
BT takes time O(k4 n2 ) using Lemma 17. Then, the sets of paths P1 , . . . , Pk and Q1 , . . . , Qk can be found in
time O(n2 ) according to Theorem 10, and from these we can easily get P1s , . . . , Pks , P1t , . . . , Pkt in linear time.
The following lemma shows how to solve any linkage problem half-integrally in a depth two bramble, provided the terminals belong to distinct bags.
Lemma 18. For all k ≥ 2, let G be a directed graph and let S0 = (s01 , . . . , s0k ) and T 0 = (t10 , . . . ,tk0 )) be two
ordered k-tuples of vertices in G. Suppose B is a bramble of depth two in G, and s01 , . . . , s0k ,t10 , . . . ,tk0 belong
to distinct bags Bs1 , . . . , Bsk , Bt1 , . . . , Btk , respectively of B. Then there exist paths P1 , . . . , Pk such that Pi links
s0i to ti0 and, additionally, every vertex of G is in at most two distinct paths Pi . Finally, it also holds that
Pi ⊆ Bsi ∪ Bti for each i, and we can find the paths P1 , . . . , Pk in time O(kn2 ).
Proof. For each i, we obtain Pi as follows. By the definition of a bramble, there exist vertices vi ∈ Bsi and
wi ∈ Bti with either vi = wi or (vi , wi ) ∈ E(G). Since Bsi and Bti are both strongly connected, there exist a
directed path from s0i to vi contained in Bsi and a directed path from wi to ti0 contained in Bti . Take Pi to be
the concatenation of these two paths. By construction, each Pi belongs to Bsi ∪ Bti . Further, since the bags
Bs1 , . . . , Bsk , Bt1 , . . . , Btk are distinct, and every vertex in G belongs to at most two distinct bags, it follows that
P1 , . . . , Pk is the desired collection of paths. Each Pi can be found in time O(n2 ), and so the overall running
time of O(kn2 ) follows.
12
We can deduce Theorem 15 from Lemmas 16 and 18 as follows.
Proof of Theorem 15. Let P1s , . . . , Pks , P1t , . . . , Pkt and s01 , . . . , s0k ,t10 , . . . ,tk0 and B 0 = Bs1 , . . . , Bsk , Bt1 , . . . , Btk satisfy
A1 - A7 , as given by Lemma 16.
By A2 , G, S0 = (s01 , . . . , s0k ) and T 0 = (t10 , . . . ,tk0 )) satisfying the hypothesis of Lemma 18. Let P1 , . . . , Pk be
the paths guaranteed by that lemma.
For each 1 ≤ i ≤ k, let Qi = Pis Pi Pit be the concatenation of these three paths. Clearly, each Qi is a directed
walk linking si to ti and therefore contains a directed path from si to ti . We just need to check that the k paths
are half-integral. Suppose for contradiction’s sake that some vertex v ∈ Qi ∩ Q j ∩ Q` for some 1 ≤ i, j, ` ≤ k
all distinct. By symmetry, we can consider four cases.
Case 1: v ∈ Pi ∩ Pj .
Then, by Lemma 18, v ∈ (Bsi ∪ Bti ) ∩ (Bsj ∪ Btj ), so v belongs to two bags of B 0 . Then by A5 v is not an
internal vertex of P`s or P`t , a contradiction.
Case 2: v ∈ Pis ∩ Pjs ∩ P` .
By A3 in Lemma 16, we may assume v = s0i , so v ∈ Pi . Since v ∈ P` , it follows v ∈ Bsi ∩ (Bs` ∪ Bt` ). By A5 , v
is not an internal vertex of Pjs , so v ∈ Bsj as well, a contradiction.
Case 3: v ∈ Pit ∩ Ptj ∩ P` .
By A4 in Lemma 16, we may assume v = ti0 , so v ∈ Pi . Again, since v ∈ P` , it follows v ∈ Bti ∩ (Bs` ∪ Bt` ). By
A5 , v is not an internal vertex of Ptj so v ∈ Bsj as well, a contradiction.
Case 4: v ∈ Pis ∩ Ptj ∩ P` .
By Lemma 18 P` ⊆ (Bs` ∩ Bt` ), but this contradicts A6 . By A6 v ∈
/ Bs` ∩ Bt` so v ∈
/ P` , a contradiction.
The running time bound of O(k4 n2 ) follows from the bounds given by Lemmas 16, 17 and 18.
5.2
Finding a bramble to link to
In this section we prove Lemma 17. We will need the following easy combinatorial lemma.
Lemma 19. For k ≥ 1, let A be a set of k elements and suppose A1 , . . . , Ak+1 are proper subsets of A. Then
there exist i 6= j with Ai ∪ A j 6= A.
Proof. We prove the lemma by induction on k. The base case k = 1 is trivial. Assume that the lemma holds
for k − 1, we will show it must also hold for k. Let A and A1 , . . . , Ak+1 be as in the lemma statement. Choose
an element v ∈ A so that (possibly after relabelling) A1 \ {v}, . . . , Ak \ {v} are proper subsets of A \ v. By
induction, there exist i 6= j with (Ai ∪ A j ) \ {v} 6= A \ {v}. We deduce that Ai ∪ A j 6= A.
We will actually prove Lemma 17 one side at a time, applying the following lemma and then a symmetric
version of it.
Lemma 20. Let k, α, β be integers with β ≥ α ≥ k. Suppose (G, S = (s1 , . . . , sk ), T = (t1 , . . . ,tk )) is a klinkage problem instance where G is a β -strongly connected directed graph and B be a bramble of depth
two and size > 4kα 2 in G. Assume B is disjoint from {si ,ti ; 1 ≤ i ≤ k}. Then there exists a bramble BS ⊆ B
such that S is α-connected to KBS in G(BS ; B). Also |B| − |BS | < 4kα 2 . Moreover we can find BS in time
O(kαn2 ).
13
Proof of Lemma 17. Assume G is (36k3 + 2k)-strongly connected and B is a bramble of depth two and size
> 188k3 in G. By Lemma 20, there is a bramble BS ⊆ B such that S is (36k3 + 2k)-connected to KBS
in G(BS ; B). We have |BS | > 188k4 − (144k3 + 8k2 ) > 36k3 . With a proof symmetric to that of Lemma
20, one can show that there exists a bramble BT ⊆ BS such that T is 3k-connected to KBT in G(BT ; BS ).
Moreover |BS | − |BT | < 36k3 . The running time of O(k4 n2 ) follows from the running time bound given in
Lemma 20.
Proof of Lemma 20. Let G0 = G(B; B) and let K0 = KB be the clique of contracted vertices in G0 . Set
M = 2kα. To find the bramble BS we are looking for, we will generate a sequence of brambles, each
contained in the previous one, until we find one that S is sufficiently highly connected to for our purposes.
Consider Algorithm 1, a procedure to find the sequence of brambles B ⊇ B1 ⊇ · · · ⊇ BM and graphs
G1 , . . . , GM = G(B1 ; B), . . . , G(BM ; B), as well as separations (A0i , B0i ) and (Ai , Bi ) separating S from the
bramble in the graphs Gi−1 and Gi , respectively.
Algorithm 1: Generating the brambles B1 , . . . , BM
1
2
3
4
5
6
7
8
9
10
for i = 1 to M − 1 do
Let (A0i , B0i ) be the separation in Gi−1 that properly separates S from Ki−1 of minimum order.
Ci ←− Ki−1 ∩ (A0i ∩ B0i ).
Ki ←− Ki−1 \Ci .
Ci ←− {im(v) : v ∈ Ci }.
Bi ←− Bi−1 \ Ci .
Gi ←− G(Bi ; B).
Ai ←− (A0i \Ci ) ∪ (∪C∈Ci C).
Bi ←− (B0i \Ci ) ∪ (∪C∈Ci C).
end
We remark that for each i (A0i , B0i ) is a separation in the graph Gi−1 while (Ai , Bi ) is a separation in Gi .
Obesrve that (A0i , B0i ) properly separates S from Ki−1 and (Ai , Bi ) properly separates S from Ki . Observe also
that Ai ∩ Bi is disjoint from Ki .
Our main claim is the following.
Claim 1. For some 1 ≤ m ≤ M, the separation (A0m , B0m ) has order at least α.
Let’s first check that Claim 1 implies the lemma. Using Algorithm 1, we can find Gm−1 and the separation
(A0m , B0m ) given by Claim 1 in time O(Mn2 ) = O(kαn2 ). Take BS = Bm−1 . In each loop of Algorithm 1, we
remove less than 2α bags from the bramble. It follows that |Bm−1 | > |B| − 2Mα ≥ |B| − 4kα 2 .
To prove Claim 1 we need the following intermediate claim. Suppose for the sake of contradiction that
|A0m ∩ B0m | < α for each 1 ≤ m ≤ M.
Claim 2. We can find indices 1 ≤ i < j ≤ M with
1. A0j ∩ B0j ∩C = 0,
/ for each C ∈ Ci and with
2. (Ai ∩ A0j ) \ (Bi ∩ B0j ) nonempty.
14
Proof. We first obtain a set of k + 1 indices J ⊂ [M] such that for each i < j ∈ J we have A0j ∩ B0j ∩C = 0,
/ for
each C ∈ Ci . Algorithm 2 gives a procedure to find J.
Algorithm 2: Generating the set J
1
2
3
4
5
6
7
J1 ←− {M}
M1 ←− [M] \ (J1 ∪ {i : ∃C ∈ Ci with A0M ∩ B0M ∩C 6= 0})
/
for i=2 to k do
Ji ←− Ji−1 ∪ {max ` : ` ∈ Mi−1 }
Mi ←− Mi−1 \ (Ji ∪ {i : ∃C ∈ Ci with A0` ∩ B0` ∩C 6= 0/ for some ` ∈ Ji })
end
J = Jk+1 ←− Jk ∪ {max ` : ` ∈ Mi−1 }
Let us first check that J is well-defined. Observe that |M1 | > M − 2α, since the sets {C` : ` ∈ [M]} correspond
to distinct sets of bags of B and no vertex of A0M ∩ B0M belongs to more than two bags. (Recall that |A0M ∩
B0M | < α.) Similarly |Mi | ≥ |Mi−1 | − 2α for each i ≥ 2. Since M = 2kα, we have |Mk | > M − 2kα ≥ 1. Thus
|J| = k + 1.
Now let us check that any pair of indices i < j ∈ J satisfies the first property in the claim, that is that
A0j ∩ B0j ∩ C = 0/ for each C ∈ Ci . We will show this holds for J1 , . . . , Jk+1 = J inductively. Since |J1 | = 1,
the base case holds trivially. Assuming the property holds for J`−1 , let us show it holds for J` . Observe that
M`−1 ⊂ M`−2 ⊂ · · · ⊂ M. The unique element i ∈ J` \ J`−1 is the smallest index in J` , so we just need to show
that for each j ∈ J`−1 we have A0j ∩ B0j ∩C = 0/ for each C ∈ Ci . But i ∈ M j−1 for each j so this is true by the
choice of i.
Now, we want two indices i < j ∈ J satisfying the second property in the claim. For each ` ∈ J, recall that
S ⊆ A0` and S ⊆ A` and that S ∩ A0` ∩ B0` = S ∩ A` ∩ B` 6= S. It follows from Lemma 19 that there exist indices
i < j with (S ∩ Ai ∩ Bi ) ∪ (S ∩ A0j ∩ B0j ) 6= S. Since S ⊆ Ai ∩ A0j , it follows that (Ai ∩ A0j ) \ (Bi ∩ B0j ) nonempty.
This proves Claim 2.
We can now prove Claim 1.
Proof of Claim 1. Recall our assumption that |A0m ∩ B0m | < α for each 1 ≤ m ≤ M. Fix i and j as in Claim 2.
The separation (A0j , B0j ) is a separation in G j−1 . Let Āi = Ai and let B̄i = Bi \(K j \Ki )∪{C : C ∈ C` : i < ` < j}.
Observe that (Āi , B̄i ) is also a separation in G j−1 and (Āi ∩ B̄i ) does not meet K j−1 . Further, since each C ∈ Ci
is strongly connected and meets B0j and, by the choice of i and j, doesn’t meet A0j ∩ B0j , we have that each
C ∈ Ci belongs to B̄i \ A0j . This implies that
(Āi ∩ B̄i ) ∩ A0j = (Ai ∩ Bi ) ∩ A0j = (A0i ∩ B0i ) ∩ A0j .
Similarly,
(A0j ∩ B0j ) \ B̄i = (A0j ∩ B0j ) \ Bi = (A0j ∩ B0j ) \ B0i .
Now, Āi ∩ A0j ⊆ V (G), that is, it contains no contracted vertices in G j−1 . Thus, while (Āi ∩ A0j , B̄i ∪ B0j ) is
a separation in G j−1 , we may view it as a separation in G, namely as (Ai ∩ A0j ,V (G) \ (Ai ∩ A0j \ Bi ∪ B0j )).
By the second property in Claim 2, this separation is nontrivial. So, by the strong connectivity of G it must
have order at least β ≥ α and we deduce that |(Ai ∩ A0j ) ∩ (Bi ∪ B0j )| ≥ β . Rewriting |(Ai ∩ A0j ) ∩ (Bi ∪ B0j )| as
15
|(Ai ∩ Bi ) ∩ A0j | + |(A0j ∩ B0j ) \ Bi | we conclude that
|(A0i ∩ B0i ) ∩ A0j | + |(A0j ∩ B0j ) \ B0i | ≥ β .
We have |A0i ∩ B0i | < α and |A0j ∩ B0j | < α by assumption, so we can assume that
|(A0i ∩ B0i ) \ A0j | + |(A0j ∩ B0j ) ∩ B0i | < |A0i ∩ B0i |.
(1)
Now, let us turn our attention to the graph Gi−1 and define
A∗j = (A0j ∩V (Gi−1 )) ∪ {u : u ∈ K` and im(u) ∩ A0j 6= 0/ for some j > ` ≥ i}
and
B∗j = (B0j ∩V (Gi−1 )) ∪ {u : u ∈ K` and im(u) ∩ A0j 6= 0/ for some j > ` ≥ i}
In other words we obtain A∗j and B∗j from A0j and B0j by replacing any vertices belonging to bags B that were
‘expanded’ to obtain the intermediate graphs Gi , . . . , G j−1 by the corresponding contracted vertices. Note
that A∗j ∩ A0i = A0j ∩ Āi . We view (A∗j , B∗j ) as a separation in Gi−1 and note that
|(A∗j ∩ B∗j ) ∩ B0i | ≤ |(A0j ∩ B0j ) ∩ B0i |
and
|(A0i ∩ B0i ) \ A∗j | ≤ |(Ai ∩ Bi ) \ A0j |.
Consider the separation (A0i ∪ A∗j , B0i ∩ B∗j ) in Gi−1 . Its order is |(A0i ∩ B0i ) \ A∗j | + |(A∗j ∩ B∗j ) ∩ B0i | ≤ |(A0i ∩ B0i ) \
A0j | + |(A0j ∩ B0j ) ∩ B0i | < |A0i ∩ B0i | by (1). This contradicts the minimality of (A0i , B0i ) in Gi−1 .
This proves Claim 1.
This completes the proof of the theorem.
6
Proofs of Theorems 1 and 3
Given Theorems 14 and 15, it is now easy to complete the proofs of Theorems 1 and 3. We begin with
Theorem 1.
Proof of Theorem 1. Let f1 be the function from Theorem 14. Let t = t(k) be the value necessary for the
size of the bramble in order to apply Theorem 15 and resolve an instance of 12 kDDPP.
Let G be an f (t)-strongly connected graph on n vertices, and let (G, S = (s1 , . . . , sk ), T = (t1 , . . . ,tk )) be an
instance of the 12 kDDPP. We can greedily find a path P with |V (P)| ≥ f (k). Note that any subset of at most
f (k) vertices is well-linked, and thus, V (P) is a well-linked set. By Theorem 14, we can find in time O(nc1 )
a bramble B of size at least t. As f (t) ≥ 36k3 + 2k, by Theorem 15, we can find a solution to (G, S, T ) in
time O(k4 n2 ), completing the proof of the theorem.
For the proof of Theorem 3, we will need two additional results from [5]. Note that in [5], neither statement
is algorithmic, but the existence of the algorithm follows immediately from the constructive proof.
16
Lemma 21 ([5], 4.3). Let G be a directed graph on n vertices and B a bramble in G. Then there is a path
P intersecting every element of B and given G and B in input, we can find the path P in time O(|B|n2 ).
Lemma 22 ([5], 4.4). Let G be a directed graph graph on n vertices, B a bramble of order k(k + 2) and P
a path intersecting every element of B. Then there exists a set X ⊆ V (P) of order 4k which is well-linked.
Given P, B, and G in input, we can algorithmically find X in time |B|nO(k) .
Proof of Theorem 3. Let (G, S = (s1 , . . . , sk ), T = (t1 , . . . ,tk )) be an instance of the 21 kDDPP. Let n = |V (G)|.
Let t be the necessary size of a bramble in order to apply Theorem 15 to resolve an instance of the 12 kDDPP.
Let f be the function in Theorem 14.
By Theorem 7, we can either find a tree decomposition of G of width at most c2 (( f (t) + 2)2 ) or a bramble B of order ( f (t) + 2)2 . Given the tree decomposition, by Corollary 5, we can solve (G, S, T ) in time
2
O(n f1 (c2 ( f (t)+2) ) ) for some function f1 .
If instead with find the bramble B, in order to apply Theorem 15, we will have to convert it to a bramble
2
2
of depth two. By Theorem 7, we may assume that |B| ≤ n2( f (t)+2) . Thus, in time nO( f (t) ) , we can find a
2
path P intersecting every element of B by Lemma 21. By Lemma 22, again in time nO( f (t) ) , we can find
a well-linked subset X ⊆ V (P) with |X| ≥ f (t). Finally, applying Theorem 14, we find can find a bramble
B 0 of size t and depth two. Finally, by Theorem 15, we can resolve (G, S, T ) in time O(k4 n2 ). In total, the
algorithm takes time O(n f2 (k) ) for some function f2 , as desired.
7
Lower Bounds
In this section, we give the proof of Theorem 2.
Proof of Theorem 2. We give a reduction from the satisfiablity problem for Boolean formulas in 3-CNF
(3SAT). The reduction is a variant on the one given in [3] which shows NP-completeness of the integral
2DDPP in directed acyclic graphs.
ε
ε
Fix a formula F with variables x1 , . . . , xn and clauses c1 , . . . , cm . Set k0 = 3+m = n, choose M ∈ {d 1−ε
k0 e, d 1−ε
k0 e+
0
1} so that M is even, and set k = k + M. Noting that M ≥ εk, we will construct a k-linkage problem instance
with strong connectivity M. We construct the graph G(F) = (V, E) as follows (see Figure 1 for an example).
Let
V1 = {vi, j : xi ∈ c j }∪{vi, j : xi ∈ c j }∪{c j , c0j : 1 ≤ j ≤ m}∪{s1 , s2 , s3 ,t1 ,t2 ,t3 }∪{ui , vi , ui , vi }∪{bi , b0i ; 1 ≤ i ≤ n}
and W = {w1 , . . . , wM }. Let V = V1 ∪W .
For each variable i, create a directed path Pi from ui to vi containing internally each vertex of the form vi, j ,
and a directed path Pi from ui to vi containing each vertex of the form vi, j .
17
W
c1
c2
c3
s1
u1
v1,1
v1,2
b3
b2
b1
v1
v2,1
v3,1
v3,2
t1
s2
s3
t2
u1
v1,3
v1
v2,2
v2,3
t3
v3,3
b03
b02
b01
c03
c02
c01
Figure 1: Illustration of the graph G(F) for the formula F = (x1 ∨ x2 ∨ x3 ) ∧ (x1 ∨ x2 ∨ x3 ) ∧ (x1 ∨ x2 ∨ x3 )
Let
E1 = {E(Pi ) ∪ E(Pi ) : 1 ≤ i ≤ n};
E2 = {(vi , ui+1 ), (vi , ui+1 ), vi , ui+1 ), (vi , ui+1 ) : 1 ≤ i < n};
E3 = {(c j , vi, j ), (vi, j , c0j ) : vi, j ∈ V1 } ∪ {(c j , vi, j ), (vi, j , c0j ) : vi, j ∈ V1 };
E4 = {(bi , vi ), (bi , vi ), (vi , bi ), (vi , bi )};
E5 = {(si , u1 ), (si , u1 ) : i ∈ {1, 2, 3}} ∪ {(vn ,ti ), (vn ,ti ) : i ∈ {1, 2, 3}};
E6 = {{(wi , v), (v, wi ) : v ∈ V1 ∪ {w j : j < i}} : 1 ≤ i ≤ M}.
Set E = E1 ∪ E2 ∪ E3 ∪ E4 ∪ E5 ∪ E6 . Observe that G(F) has strong connectivity at least M, since each wi is
adjacent to and from each other vertex.
Consider the k-linkage problem instance (G(F), S, T ) where
S = {s1 , s2 , s3 , c1 , . . . , cm , b1 , . . . , bn , w1 , . . . , w M , w M
2
2 +1
, . . . , wM }
and
T = {t1 ,t2 ,t3 , c01 , . . . , c0m , b01 , . . . , b0n , w M
2 +1
, . . . , wM , w1 , . . . , w M }.
2
This is clearly a polynomial reduction. We need to show that F is satisfiable if and only if (G(F), S, T ) is
half-integrally feasible.
Claim. If F is satisfiable then (G(F), S, T ) is half-integrally feasible.
Fix a satisfying assignment of F and consider the S-T linkage consisting of
• the paths s1 P1 . . . Pnt1 , s2 P1 . . . Pnt2 ;
• the s3 − t3 path whose interior is the concatenation of Pi if xi is set to false or Pi if xi is set to true, for
each 1 ≤ i ≤ n;
18
• for each 1 ≤ j ≤ m, a path c j vi, j c0j if c j contains a variable xi that is true, or the path c j vi, j c0j if c j
contains the negation of a variable xi that is false;
• for each 1 ≤ i ≤ n, the path bi vi b0i if xi is true, or the path bi vi b0i if xi is false;
• the 1-edge paths (wi , w M ) and (w M , wi ) for each 1 ≤ i ≤
2 +i
2 +i
M
2.
It is straightforward to check that each vertex is used at most twice in this linkage, proving that (G(F), S, T )
is half-integrally feasible.
Claim. If (G(F), S, T ) is half-integrally feasible then F is satisfiable.
Suppose (G(F), S, T ) has a half-integral solution. Observe that each vertex in W belongs to both S and T and
therefore cannot belong to any path in the solution for which it is not a terminal. Thus, for j = 1, 2, 3, every
s j − t j path must use either Pi or Pi , for each 1 ≤ i ≤ n. We will show that the following assignment satisfies
F: set variable xi to false if the path Pi is used by two of those paths, and to true if the path Pi is used twice
by those paths.
Every bi − b0i path must use vi or vi . By half-integrality, the bi − b0i path uses vi if Pi is used by two of the
s j − t j paths, and uses v̄i if Pi is used by two of those paths. In particular, each vi and v̄i must be used exactly
twice by the union of the s j − t j and bi − b0i paths. We deduce that each c j − c0j path contains two edges.
Moreover, the middle vertex of that path is either vi, j where the variable xi belongs to ci and is set to true, or
vi, j where the negation of variable xi belongs to ci and xi is set to false. This proves that F is satisfiable.
References
[1] M. Cygan, D. Marx, M. Pilipczuk, and M. Pilipczuk. The planar directed k-vertex-disjoint paths
problem is fixed-parameter tractable. In Foundations of Computer Science (FOCS), 2013 IEEE 54th
Annual Symposium on Foundations of Computer Science, pages 197–206, Oct 2013.
[2] Reinhard Diestel, Tommy R Jensen, Konstantin Yu Gorbunov, and Carsten Thomassen. Highly connected sets and the excluded grid theorem. Journal of Combinatorial Theory, Series B, 75(1):61–73,
1999.
[3] Steven Fortune, John Hopcroft, and James Wyllie. The directed subgraph homeomorphism problem.
Theoretical Computer Science, 10(2):111–121, 1980.
[4] Thor Johnson, Neil Robertson, Paul D Seymour, and Robin Thomas. Directed tree-width. Journal of
Combinatorial Theory, Series B, 82(1):138–154, 2001.
[5] K.-i. Kawarabayashi and S. Kreutzer. The Directed Grid Theorem. ArXiv e-prints, November 2014.
[6] Ken-ichi Kawarabayashi. An improved algorithm for finding cycles through elements. In International
Conference on Integer Programming and Combinatorial Optimization, pages 374–384. Springer, 2008.
[7] Ken-ichi Kawarabayashi, Yusuke Kobayashi, and Stephan Kreutzer. An excluded half-integral grid
theorem for digraphs and the directed disjoint paths problem. In Proceedings of the 46th Annual ACM
Symposium on Theory of Computing, STOC ’14, pages 70–78, New York, NY, USA, 2014. ACM.
19
[8] Ken-Ichi Kawarabayashi and Stephan Kreutzer. An Excluded Grid Theorem for Digraphs with Forbidden Minors, pages 72–81.
[9] Ken-ichi Kawarabayashi and Stephan Kreutzer. The directed grid theorem. In Proceedings of the
Forty-Seventh Annual ACM on Symposium on Theory of Computing, pages 655–664. ACM, 2015.
[10] Karl Menger. Zur allgemeinen kurventheorie. Fundamenta Mathematicae, 1(10):96–115, 1927.
[11] Bruce Reed. Introducing directed tree width. Electronic Notes in Discrete Mathematics, (3):1–8, 2000.
[12] N. Robertson and P.D. Seymour. Graph minors .xiii. the disjoint paths problem. Journal of Combinatorial Theory, Series B, 63(1):65 – 110, 1995.
[13] Alexander Schrijver. Finding k disjoint paths in a directed planar graph. SIAM Journal on Computing,
23(4):780–788, 1994.
[14] Aleksandrs Slivkins. Parameterized tractability of edge-disjoint paths on directed acyclic graphs. SIAM
Journal on Discrete Mathematics, 24(1):146–157, 2010.
[15] Carsten Thomassen. 2-linked graphs. European Journal of Combinatorics, 1(4):371–378, 1980.
[16] Carsten Thomassen. Highly connected non-2-linked digraphs. Combinatorica, 11(4):393–395, 1991.
20
| 8cs.DS
|
arXiv:1708.04770v1 [math.AC] 16 Aug 2017
GENERATORS OF REDUCTIONS OF IDEALS IN A LOCAL
NOETHERIAN RING WITH FINITE RESIDUE FIELD
LOUIZA FOULI AND BRUCE OLBERDING
A BSTRACT. Let (R, m) be a local Noetherian ring with residue field k. While
much is known about the generating sets of reductions of ideals of R if k is
infinite, the case in which k is finite is less well understood. We investigate the
existence (or lack thereof) of proper reductions of an ideal of R and the number
of generators needed for a reduction in the case k is a finite field. When R is
one-dimensional, we give a formula for the smallest integer n for which every
ideal has an n-generated reduction. It follows that in a one-dimensional local
Noetherian ring every ideal has a principal reduction if and only if the number of
maximal ideals in the normalization of the reduced quotient of R is at most |k|.
In higher dimensions, we show that for any positive integer, there exists an ideal
of R that does not have an n-generated reduction and that if n ≥ dim R this ideal
can be chosen to be m-primary.
1. I NTRODUCTION
Let R be a (commutative) Noetherian ring, and let I be an ideal of R. A reduction of I is a subideal J of I such that I n+1 = JI n for some n > 0; equivalently, J is
a subideal of I such that I = J, where denotes the integral closure of the corresponding ideal in R. Northcott and Rees [10] proved that if R is a local Noetherian
ring with infinite residue field and Krull dimension d, then every ideal of R has
a d-generated reduction, that is, a reduction that can be generated by d elements.
This result and its generalizations involving analytic spread underlie many of the
applications of the theory of reductions to local algebra. For example, reductions
and analytic spread have been instrumental in describing the asymptotic properties
of an ideal I of R, the Cohen-Macaulay property of the Rees algebra R (I) of I,
and the blowup Proj R (I) of Spec R along the subscheme defined by I. However,
if the residue field of R is finite, then there may exist ideals of R that do not have
a d-generated reduction, and so the applicability of reductions in the case of local
rings with finite residue field is more limited.
2010 Mathematics Subject Classification. 13A30, 13B22, 13A15.
Key words and phrases. reduction, integral closure, finite field, analytic spread.
The first author was partially supported by a grant from the Simons Foundation, grant #244930.
1
2
L. FOULI AND B. OLBERDING
In this article, we examine the extent to which the result of Northcott and Rees
involving d-generated reductions fails in the case of finite residue field. We prove
two main results, the first of which is devoted to one-dimensional rings and the
second to rings of higher dimension. In the one-dimensional case we find the
optimal choice for replacing d in the result of Northcott and Rees with the smallest
possible positive integer, an integer that depends only on the size of the residue field
and the cardinality | Max(Rred )| of the set of maximal ideals in the normalization
√
Rred of the reduced quotient Rred = R/ 0 of R.
Theorem A. Let R be a one-dimensional local Noetherian ring with finite residue
field k. The smallest positive integer n for which every ideal of R has an n-generated
reduction is
n = ⌈ −1 + log|k| |k| + (|k| − 1) · | Max(Rred )| ⌉.
For a one-dimensional local Noetherian ring R of multiplicity e, every ideal of
R can be generated by e elements [11, Theorem 1.1, p. 49]. Thus, the number n in
the theorem is at most e.
If R is a one-dimensional local Noetherian ring with infinite residue field, then
every ideal of R has a principal reduction. Using Theorem A, we extend this result
to one-dimensional local rings with residue field of any size (see Corollary 3.3).
Corollary. Let R be a one-dimensional local Noetherian ring with residue field k.
Every ideal of R has a principal reduction if and only if | Max(Rred )| ≤ |k|.
In particular, if R is a complete local Noetherian domain, then every ideal of R
has a principal reduction (Corollary 3.7). Thus in dimension one there are interesting local Noetherian rings with finite residue field for which the result of Northcott
and Rees holds, i.e., every ideal of R has a reduction generated by dim(R)-many
elements. Moreover, in dimension one, even if there are ideals without a principal
reduction, we are at least guaranteed the existence of a bound on the number of
elements needed to generate a reduction.
Moving beyond dimension one, we use Theorem A to show (see Theorem 3.8)
that unlike in the case of infinite residue field, no such bound exists for a local
Noetherian ring with finite residue field and dimension at least 2:
Theorem B. Let (R, m) be a local Noetherian ring of dimension d ≥ 2. If the
residue field of R is finite, then for each positive integer n there is an ideal of R that
is minimally generated by n elements and does not have a proper reduction. If also
n ≥ d, this ideal can be chosen to be m-primary.
REDUCTIONS OF IDEALS
3
Notation. Throughout the article, Q(R) denotes the total quotient ring of the ring
R, and R is the integral closure of R in Q(R). We denote by Rred the reduced ring
√
√
R/ 0, where 0 is the nilradical of R. The set of maximal ideals of R is denoted
Max(R).
2. P RELIMINARIES
In this section we develop a criterion for when every ideal in a local Noetherian
ring with finite residue field has an n-generated reduction. This criterion, Proposition 2.4, will be used in the proofs of the main results in the next section. The
first lemma, which is a routine application of well-known properties of reductions,
concerns the transfer of reductions in a ring R to its reduced quotient Rred .
Lemma 2.1. Let R be a ring and let n be a positive integer. A finitely generated
ideal I of R has an n-generated reduction if and only if IRred has an n-generated
reduction.
Proof. It is clear that if every ideal of R has n-generated reduction, then every
ideal of Rred has an n-generated reduction. Conversely, let I be an ideal of R and
suppose that J is an n-generated ideal of R such that JRred is a reduction of IRred .
√
√
Then J ⊆ I + 0. Write J = (x1 , . . . , xn )R. For each i, there is yi ∈ I and zi ∈ 0
√
√
such that xi = yi + zi . Thus J + 0 = (y1 , . . . , yn )R + 0. Let K = (y1 , . . . , yn )R.
Then K ⊆ I and KRred = JRred . Thus KRred is a reduction of IRred . We claim
that K is a reduction of I. Since KRred = IRred we have by [13, Proposition 1.1.5]
√
that KRred = IRred . By [13, Remark 1.1.3(5)], the nilradical 0 of R is contained
in every integrally closed ideal of R, so we conclude that K = I. Thus K is an
n-generated reduction of I.
As discussed in the introduction, if R is a one-dimensional local Noetherian
ring with infinite residue field, then every ideal of R has a principal reduction.
Removing the restriction to infinite residue field, we can assert in general only that
every height 0 ideal has a principal reduction:
Proposition 2.2. Let R be a one-dimensional local Noetherian ring. Then every
height 0 ideal of R has a principal reduction.
Proof. Let I be a height 0 ideal of R. By Lemma 2.1, it suffices to show that IRred
has a principal reduction, so we may assume without loss of generality that R is a
reduced ring and I 6= 0. Let P1 , . . . , Pn denote the minimal prime ideals of R. Since
I has height 0 and R is local and one-dimensional, we may assume P1 , . . . , Pt are
4
L. FOULI AND B. OLBERDING
the minimal prime ideals of I, where 1 ≤ t ≤ n. Since R is reduced and I 6= 0,
we have t < n. By Prime Avoidance there exists x ∈ I \ (Pt+1 ∪ · · · ∪ Pn ). Since
√
I = P1 ∩ · · · ∩ Pt and the only prime ideals of R are P1 , . . . , Pn and the maximal
√
√
ideal of R, it follows that xR = I.
We claim xR = IR. Since I has height 0, every element of I is a zero divisor of R.
Thus the ideal IR of R consists of zerodivisors of R. Since R is a reduced Noetherian
ring, R is a finite product of integral domains R1 , . . . , Rk (see for example [5, p. 64]).
The fact that IR consists of zero divisors implies that for each 1 ≤ i ≤ k, we have
√
√
IRi = 0 or IRi = Ri . Moreover, since xR = I, some power of I is contained
in xR. Thus, in the domain Ri we have IRi = 0 if and only if xRi = 0. Since
R = R1 × · · · × Rk , we conclude that IR = xR. By [13, Proposition 1.6.1], the fact
that R ⊆ R is an integral extension implies I = IR ∩ R and xR = xR ∩ R. Since
IR = xR, we have I = xR. Also, xR ⊆ I, so xR is a reduction of I.
The next lemma and proposition give criteria for when every ideal in a Noetherian ring R has an n-generated reduction. The stronger result, Proposition 2.4, requires that R is also reduced, local and one-dimensional. In light of Theorem B, the
one-dimensional assumption is necessary in the proposition. To state Lemma 2.3,
we recall that the arithmetic rank, ara(I), of a proper ideal I of a Noetherian ring R
p
√
is the least number n such that I = (x1 , . . . , xn )R for some x1 , . . . , xn ∈ I.
Lemma 2.3. Let R be a Noetherian ring, let I be a proper ideal of R and let n be an
√
√
integer such that n ≥ ara(I). If each (n + 1)-generated ideal J ⊆ I with J = I
has an n-generated reduction, then I has an n-generated reduction.
Proof. Let I be a proper ideal of R and let
√
√
F = {J : J ⊆ I, J = I and J is an n-generated ideal of R}.
Since R is a Noetherian ring and n ≥ ara(I), the set F is nonempty and contains
√
√
a maximal element J, where J is an n-generated ideal with J ⊆ I and J = I.
Suppose that J ( I. Then I 6⊆ J and we may choose y ∈ I \ J. Now J + yR is an
(n+ 1)-generated ideal, and hence by assumption there is an n-generated ideal K of
R such that K ⊆ J + yR and K = J + yR. But then J ⊆ K ∈ F , so that the maximality
of J in F forces J = J + yR, a contradiction to the fact that y 6∈ J. Therefore, J = I
and J is an n-generated reduction of I.
Proposition 2.4. Let R be a reduced one-dimensional local Noetherian ring and
let n be a positive integer. Then the following are equivalent:
(1) Every ideal of R has an n-generated reduction.
REDUCTIONS OF IDEALS
5
(2) For all x1 , . . . , xn+1 ∈ R for which (x1 , . . . , xn+1 )R = R, there is an n-generated
R-submodule of (x1 , . . . , xn+1 )R that does not survive in R.
Proof. In the proof we use the fact that every ideal of R is a principal ideal of R.
This can be seen as follows. By [5, Theorem 12.3, p. 66], the fact that R is a reduced
local Noetherian ring of dimension ≤ 2 implies R is a finite product of Noetherian
integrally closed domains. Since also dim R = dim R = 1, we have that R is a finite
product of Dedekind domains. The fact that R is semilocal implies these Dedekind
domains are principal ideal domains. As a finite product of principal ideal domains,
the ring R has the property that every ideal is a principal ideal.
Let m denote the maximal ideal of R. First, suppose every ideal of R has an
n-generated reduction and let x1 , . . . , xn+1 ∈ R such that (x1 , . . . , xn+1 )R = R. Let
A = (x1 , . . . , xn+1 )R. Then there is a nonzerodivisor r ∈ R such that I := rA is
an ideal of R. Since AR = R, we have AQ(R) = Q(R), and hence A contains a
nonzerodivisor of R. Therefore, I = rA also contains a nonzerodivisor of R. By
assumption, there is an n-generated reduction J of I. Let m > 0 be such that I m+1 =
JI m . As we have established, every ideal of R is a principal ideal. Thus IR is a
principal ideal of R that is necessarily generated by a nonzerodivisor of R since
I contains a nonzerodivisor. Since I m+1 R = JI m R and principal ideals generated
by a nonzerodivisor admit no proper reductions, we obtain IR = JR. Therefore,
R = AR = r−1 IR = r−1 JR. Since J ⊆ I, we have r−1 J ⊆ A, and hence r−1 J is an
n-generated R-submodule of A that does not survive in R.
Conversely, suppose that for all x1 , . . . , xn+1 ∈ R with (x1 , . . . , xn+1 )R = R, there
is an n-generated R-submodule of (x1 , . . . , xn+1 )R that does not survive in R. To
prove that every ideal of R has an n-generated reduction, it suffices by Lemma 2.3
to show that every (n + 1)-generated ideal of R has an n-generated reduction, since
ara(I) = 1 for any proper non-trivial ideal I of R. Let a1 , . . . , an+1 ∈ R and let
I = (a1 , . . . , an+1 )R. If I = R, then clearly I has an n-generated reduction. If I has
height 0, then by Proposition 2.2, I has a principal reduction. Thus it remains to
consider the case in which I is m-primary.
Since every ideal of R is a principal ideal, we have IR = tR for some t ∈ R. Since
I is m-primary and R is reduced, I contains a nonzerodivisor, so t is a nonzerodivisor in R. For each i ∈ {1, . . . , n + 1}, let xi = ai t −1 . Then R = t −1 tR = t −1 IR =
(x1 , . . . , xn+1 )R. By assumption there exists an n-generated R-submodule A of
(x1 , . . . , xn+1 )R that does not survive in R. Since we have (x1 , . . . , xn+1 )R = R = AR,
it follows that
IR = t(x1 , . . . , xn+1 )R = tAR.
6
L. FOULI AND B. OLBERDING
Moreover, tA ⊆ t(x1 , . . . , xn+1 )R = I. As in the proof of Lemma 2.2, the fact that
IR = tAR implies I = tA. Since tA ⊆ I, this proves that tA is an n-generated reduction of I.
3. M AIN
RESULTS
In this section we prove the main results of the paper. After proving the first
theorem, which deals with the one-dimensional case, we indicate how Theorem A
of the introduction follows. At the end of the section in Theorem 3.8, we prove
Theorem B of the introduction.
Theorem 3.1. Let R be a one-dimensional local Noetherian ring with finite residue
field k, and let n be a positive integer. Then every ideal of R has an n-generated
reduction if and only if
| Max(Rred )| ≤
|k|n+1 − |k|
.
|k| − 1
Proof. By Lemma 2.1, every ideal of R has an n-generated reduction if and only
if every ideal of Rred has an n-generated reduction. Thus it suffices to prove the
theorem in the case where R is a reduced ring. Throughout the proof, we let U
denote a set of |k|-many elements of R such that R/m = {u + m : u ∈ U }, where m
is the maximal ideal of R. We assume 0 ∈ U . Since R is a local ring, all the nonzero
elements of U are units in R. We denote the elements of the Cartesian product U n
by u = (u1 , . . . , un ). Let
J = {(i, u) ∈ {1, . . . , n + 1} ×U n : u j = 0 for all j ≥ i}.
Then
|J | = 1 + |k| + |k|2 + · · · + |k|n =
Now we prove the theorem. Suppose first that
| Max(R)| >
|k|n+1 − 1
.
|k| − 1
|k|n+1 − |k|
.
|k| − 1
We show there is an ideal of R that does not have an n-generated reduction. By
Proposition 2.4, it suffices to show that there are x1 , . . . , xn+1 ∈ R such that R =
(x1 , . . . , xn+1 )R and every n-generated R-submodule of (x1 , . . . , xn+1 )R survives in
R.
By assumption,
| Max(R)| ≥
|k|n+1 − |k|
|k|n+1 − |k| + |k| − 1
+1 =
= |J |.
|k| − 1
|k| − 1
REDUCTIONS OF IDEALS
7
Therefore, we may index a set {Mi,u : (i, u) ∈ J } of |J |-many maximal ideals of R
by J . Since the ideals Mi,u of R are maximal, the diagonal map
φ : R −→
∏
(i,u)∈J
R/Mi,u : x 7→ (x + Mi,u )
is a surjective ring homomorphism. Thus we may choose x1 , . . . , xn+1 ∈ R such that
for each (i, u) ∈ J and j = 1, . . . , n + 1, we have
if i < j
0 + Mi,u
x j + Mi,u = 1 + Mi,u
if i = j .
u j + Mi,u if i > j
Let K = (x1 , . . . , xn+1 )R. We claim that R = (x1 , . . . , xn+1 )R and every n-generated
R-submodule of K survives in R. To show this we first show that K/mK has dimension n + 1 as a k-vector space. Indeed, suppose i ∈ {1, . . . , n + 1} and
xi + mK = ∑ r j x j + mK
j6=i
for some r j ∈ R. For each j 6= i, the choice of x j implies x j ∈ Mi,0 (with 0 =
(0, . . . , 0)). Since
xi − ∑ r j x j ∈ mK ⊆ Mi,0 ,
j6=i
we conclude that xi ∈ Mi,0 , contrary to the fact that by the choice of xi , we have
xi + Mi,0 = 1 + Mi,0 . This contradiction shows that K/mK has dimension n + 1 as a
k-vector space. In particular, K cannot be generated as an R-module by fewer than
n + 1 elements.
Let I be an n-generated R-submodule of K = (x1 , . . . , xn+1 )R. We claim that I
survives in R. Since I can be generated as an R-module by n elements, the dimension of the k-vector space I/mI is at most n. Since K cannot be generated by fewer
than n + 1 elements, by adding as many of the elements xi to I as needed we can
assume without loss of generality that I is an R-submodule of K such that I can be
generated by n but no fewer elements. In particular, the k-dimension of I/mI is n.
Then Nakayama’s Lemma implies that I is generated by n elements y1 , . . . , yn of
the form
y j = u j,1 x1 + u j,2 x2 + · · · + u j,n+1 xn+1 , where u j,i ∈ U .
Since I/mI is a k-vector space of dimension n, the n × (n + 1) matrix
(u j,i + m : j = 1, . . . , n, i = 1, . . . , n + 1),
whose entries are in the field k, has rank n. Elementary row operations produce a
rank n matrix in reduced row echelon form such that for some i = 1, . . . , n + 1,
8
L. FOULI AND B. OLBERDING
deleting the i-th column yields the n × n identity matrix. It follows from this observation and Nakayama’s Lemma that there are u1 , . . . , ui−1 ∈ U such that
I = ({x j − u j xi : j < i} ∪ {xi+1 , . . . , xn+1 })R.
Let u = (u1 , . . . , ui−1 , 0, . . . , 0) ∈ U n so that (i, u) ∈ J . We claim that I ⊆ Mi,u .
First, observe that xi+1 , . . . , xn+1 ∈ Mi,u by the choice of these elements, so it remains to show that for each j < i, we have x j − u j xi ∈ Mi,u . Let j < i and notice
that by the choice of xi , we have
xi + Mi,u = 1 + Mi,u and x j + Mi,u = u j + Mi,u .
Therefore,
x j − u j xi + Mi,u = u j − u j · 1 + Mi,u = 0 + Mi,u ,
so that x j − u j xi ∈ Mi,u , proving the claim that I ⊆ Mi,u . We conclude that I survives
in R. This shows that every n-generated R-submodule of K = (x1 , . . . , xn+1 )R survives in R. Therefore, R has an ideal that does not have an n-generated reduction.
Conversely, suppose that
| Max(R)| ≤
|k|n+1 − |k|
.
|k| − 1
To prove that every ideal of R has an n-generated reduction, it suffices by Proposition 2.4 to show that for all x1 , . . . , xn+1 ∈ R with (x1 , . . . , xn+1 )R = R, there is an
n-generated R-submodule I of (x1 , . . . , xn+1 )R such that IR = R.
Let x1 , . . . , xn+1 ∈ R such that (x1 , . . . , xn+1 )R = R. For each (i, u) ∈ J with
u = (u1 , . . . , un ), consider the n-generated R-submodule of (x1 , . . . , xn+1 )R defined
by
Ii,u = (x1 − u1 xi , x2 − u2 xi , . . . , xi−1 − ui−1 xi , xi+1 , . . . , xn+1 )R.
We claim first that if (s, u), (t, v) ∈ J and Is,u and It,v are contained in a common
maximal ideal of R, then s = t and u = v. Suppose M is a maximal ideal of R with
Is,u + It,v ⊆ M.
Suppose by way of contradiction that s 6= t. Without loss of generality, we can
assume that s < t. Then xt ∈ Is,u , so that since
x1 − v1 xt , x2 − v2 xt , . . . , xt−1 − vt−1 xt ∈ It,v ⊆ M,
we have x1 , x2 , . . . , xt−1 ∈ M. Also, since s < t, we have xt , . . . , xn+1 ∈ Is,v ⊆ M. But
then R = (x1 , . . . , xn+1 )R ⊆ M, a contradiction that implies s = t.
Next we claim that u = v. Since s = t, we have
x1 − u1 xt , . . . , xt−1 − ut−1 xt ∈ Is,u ⊆ M.
REDUCTIONS OF IDEALS
9
Similarly, we have
x1 − v1 xt , . . . , xt−1 − vt−1 xt ∈ It,v ⊆ M.
Therefore, for each i < t, we have
xi − ui xt − (xi − vi xt ) = (vi − ui )xt ∈ M.
An argument similar to the one in the preceding paragraph shows that xt 6∈ M, since
otherwise every x j ∈ M, for j = 1, 2, . . . , n + 1, a contradiction. Thus, for each i < t
we have vi − ui ∈ M ∩ R = m, where the last equality follows from the fact that as a
maximal ideal of an integral extension of R, M lies over m. Consequently, for each
i < t we have ui + m = vi + m, which, since ui , vi ∈ U , forces ui = vi . Since this
holds for all i < t, we conclude u = v. This proves that if Is,u and It,v are contained
in a common maximal ideal of R, then s = t and u = v.
Next, since no two distinct R-submodules of (x1 , . . . , xn+1 )R from the set {Is,u :
(s, u) ∈ J } are contained in the same maximal ideal of R, it follows that either one
of these R-submodules does not survive in R or there are at least |J |-many maximal
ideals of R. As established at the beginning of the proof,
|J | =
|k|n+1 − 1
,
|k| − 1
so if every Is,u survives in R, we conclude that
|k|n+1 − |k|
|k|n+1 − 1
≤ | Max(R)| ≤
,
|k| − 1
|k| − 1
where the last inequality is given by assumption. This implies
|k|n+1 − 1 ≤ |k|n+1 − |k|,
which is impossible since |k| > 1. This contradiction implies some Is,u does not
survive in R. In particular, some n-generated R-submodule of (x1 , . . . , xn+1 )R does
not survive in R. Therefore, by Proposition 2.4 every ideal of R has an n-generated
reduction.
Theorem A now follows easily from Theorem 3.1: In the setting of the theorem,
we seek the smallest positive integer n such that
|k|n+1 − |k|
;
|k| − 1
that is, we need the smallest positive integer n such that
| Max(Rred )| ≤
|k| + (|k| − 1) · | Max(Rred )| ≤ |k|n+1 .
Equivalently,
log|k| |k| + (|k| − 1) · | Max(Rred )| ≤ n + 1,
10
L. FOULI AND B. OLBERDING
which yields the conclusion of Theorem A.
Remark 3.2. If R is a one-dimensional Noetherian local domain, then the number
of maximal ideals of R is the same as the number of minimal prime ideals in Rb [6,
Corollary 5].
Moreover, if R is a reduced local Noetherian ring with geometrically regular
formal fibers, then the number of maximal ideals of R is the same as the number
of minimal prime ideals of the completion Rb of R [1, Theorem 6.5]. Thus for
excellent local Noetherian rings R, the bound in Theorem 3.1 can be restated using
the completion of Rred rather than the normalization of Rred .
Corollary 3.3. Let R be a one-dimensional local Noetherian ring with residue
field k. Then every ideal of R has a principal reduction if and only if
| Max(Rred )| ≤ |k|.
Proof. Apply Theorem 3.1 in the case n = 1.
Corollary 3.4. Let R be a one-dimensional local Noetherian domain with residue
field k. If the multiplicity of R is at most |k|, then every ideal of R has a principal
reduction.
Proof. The completion of a one-dimensional local Noetherian domain of multiplicity e has at most e minimal prime ideals (e.g., this follows from the multiplicity
formula given in [9, Theorem 14.7]). By Remark 3.2, | Max(R)| ≤ e ≤ |k|. Thus
the corollary is a consequence of Corollary 3.3.
Remark 3.5. It follows from Corollary 3.4 that if a one-dimensional Noetherian
local domain R has multiplicity 2, then every ideal of R has a principal reduction. This is known already for other reasons. In a one-dimensional local CohenMacaulay ring of multiplicity e, every ideal can be generated by e elements [11,
Theorem 1.1, p. 49]. Since R has multiplicity 2, every ideal of R is 2-generated. In
[12, Theorem 3.4], Sally and Vasconcelos prove that a ring in which every ideal is
2-generated has the property that every m-primary ideal has a principal reduction
of reduction number at most 1; see also [8, Lemma 1.11].
Corollary 3.6. Let (R, m) be a one-dimensional Noetherian local domain whose
completion has at most two minimal prime ideals. Then every ideal of R has a
principal reduction.
Proof. Apply Remark 3.2 and Corollary 3.3.
REDUCTIONS OF IDEALS
11
Corollary 3.7. Every ideal of a one-dimensional complete local domain has a
principal reduction.
Proof. Apply Corollary 3.6.
Next we use the one-dimensional case in Theorem 3.1 to show the absence of a
bound on the number of generators of reductions in higher dimensions.
Theorem 3.8. Let (R, m) be a local Noetherian ring of dimension d ≥ 2. If the
residue field of R is finite, then for each positive integer n there is an ideal of R that
is minimally generated by n elements and does not have a proper reduction. If also
n ≥ d, this ideal can be chosen to be m-primary.
Proof. Let n > 0 and let k denote the residue field of R. For the first assertion in
the theorem, it suffices to show that there exists an ideal of R that is generated by
n elements and has no (n − 1)-generated reduction, see for example [13, Proposition 8.3.3]. Since R is Noetherian and R has dimension d > 1, there are infinitely
many prime ideals of R of dimension one. Choose a positive integer t with
|k|n − |k|
.
t >
|k| − 1
Let P1 , . . . , Pt be distinct dimension one prime ideals of R. Let
A = R/(P1 ∩ · · · ∩ Pt ).
Then A is a reduced one-dimensional local Noetherian ring with t minimal prime
ideals. As such, A is a direct product of t integrally closed domains [5, p. 64].
Consequently, A has at least t maximal ideals. By Lemma 2.3 and Theorem 3.1
there is an n-generated ideal I of A that does not have an (n − 1)-generated reduction. Let J be an n-generated ideal of R such that JA = I. If J has an (n − 1)generated reduction K ⊆ J, then there is m > 0 such that J m+1 = J m K. But then
I m+1 = J m+1 A = J m KA = I m KA, so that KA is an (n − 1)-generated reduction of
I, contrary to the choice of I. We conclude that the n-generated ideal J has no
(n − 1)-generated reduction. Hence J is minimally generated by n elements and
has no proper reduction.
It remains to prove the second assertion in the theorem. Suppose n ≥ d. If n = d,
then we may choose any n-generated m-primary ideal of R. By Krull’s height
theorem such an ideal must exist and cannot have an (n − 1)-generated reduction
since such a reduction would be an m-primary ideal generated by d − 1 elements.
Thus, if n = d, the proof is complete.
Assume n > d. With I as above, choose J to be an m-primary ideal of R with
P1 ∩ · · · ∩ Pt ⊆ J and JA = I. Since J is m-primary, Krull’s height theorem implies
12
L. FOULI AND B. OLBERDING
that the arithmetic rank of J is d. Therefore, since n − 1 ≥ d, Lemma 2.3 implies
√
that there is an n-generated ideal J ′ ⊆ J such that J ′ = m and J ′ does not have an
(n − 1)-generated reduction. This shows that J ′ is an n-generated m-primary ideal
with no (n − 1)-generated reduction, so we conclude that J ′ is minimally generated
by n elements and J ′ has no proper reduction.
Remark 3.9. In [3, Example 2.3] an example is given of a two-dimensional CohenMacaulay local ring R such that R has finite residue field and the maximal ideal m
of R fails to have a 2-generated reduction. In [4, Example 5.4], Abhyankar showed
that certain canonically defined m-primary ideals of a two-dimensional regular local ring (R, m) have 2-generated reductions. This motivates the question of Heinzer
and Shannon in [4, Question 5.6] of whether every integrally closed m-primary
ideal of a two-dimensional regular local ring (R, m) with finite residue field has a
2-generated reduction. Theorem 3.8 guarantees the existence of m-primary ideals
without 2-generated reductions, but it is unclear whether these ideals can be chosen
to be integrally closed. In general, a d-dimensional local Noetherian ring can have
the property that every integrally closed ideal has a d-generated reduction yet still
possess non-integrally closed ideals that do not have a d-generated reduction; see
Example 4.1.
4. E XAMPLES
In this section we give several examples to illustrate some of the ideas in Section 3. In the first example, we show that in order for each ideal of a one-dimensional
local Noetherian domain to have a principal reduction, it is not sufficient that every
integrally closed ideal has a principal reduction.
Example 4.1. For each integer n ≥ 1, there exists a one-dimensional local Noetherian domain R such that every integrally closed ideal has a principal reduction,
yet R has an ideal that is minimally generated by n elements and does not have a
proper reduction. Let n ≥ 1. If n = 1, then any one-dimensional local Noetherian domain suffices as an example, since a nonzero principal ideal does not have a
principal reduction. Suppose n > 1. Choose 2n − 1 maximal ideals M1 , . . . , M2n −1
of the ring Z2 [x], and let S = Z2 [x] \ (M1 ∪ · · · ∪ M2n −1 ). Let J denote the Jacobson radical of the localization Z2 [x]S of Z2 [x] at the multiplicatively closed set S,
and let R = Z2 + J. Then R is a one-dimensional local Noetherian domain with
maximal ideal J and normalization R = Z2 [x]S ; see for example [2, Lemma 1.1.4
and Proposition 1.1.7]. Thus R is a PID with | Max(R)| = 2n − 1. Observe that
REDUCTIONS OF IDEALS
13
(J :Q(R) J) = R, and thus since R is a PID, the blow up of the maximal ideal J of
R in the sense of Lipman [8, p. 651] is R. This ring has the property that each localization at a maximal ideal has its embedding dimension (which is 1) equal to its
multiplicity. Therefore, by [8, Theorem 2.2], every integrally closed ideal of R has
a principal reduction of reduction number at most 1. However, since the residue
field k of R has two elements and
|k|n − |k|
| Max(R)| = |k|n − 1 >
,
|k| − 1
Lemma 2.3 and Theorem 3.1 imply that R has an ideal that is minimally generated
by n elements and does not have an (n − 1)-generated reduction. By [13, Proposition 8.3.3], this ideal does not have a proper reduction.
Corollary 3.4 implies that the maximal ideal of a one-dimensional local Noetherian domain of multiplicity 2 has a principal reduction. Corollary 3.3 guarantees that there exist one-dimensional local Cohen-Macaulay rings of multiplicity
3 that have ideals without a principal reduction (and necessarily such a local ring
has residue field the two-element field). The next example, which appears in [13,
Example 8.3.2], shows that there exists a one-dimensional local Cohen-Macaulay
ring of multiplicity 3 whose maximal ideal does not have a principal reduction.
Example 4.2. (See [13, Example 8.3.2]) Let R be the ring given by
Z2 [[x, y]]
.
R=
(xy(x + y))
The multiplicity of R is 3. Let m denote the maximal ideal of R, and let x′ , y′ denote
the images of x and y in R, respectively. Suppose that m has a principal reduction
J = ( f ). Write f = r1 x′ + r2 y′ , for some r1 , r2 ∈ R. Since R = Z2 + m, there are
a1 , a2 ∈ Z2 such that for each i, ri − ai ∈ m. Let f ′ = a1 x′ + a2 y′ , and let J ′ = ( f ′ ).
Then J ⊆ J ′ + mI ⊆ I and thus J ′ = I by [13, Lemma 8.1.8]. In this way, we may
assume that any principal reduction of m is generated by an element of the form
f = a1 x′ + a2 y′ , with a1 , a2 ∈ Z2 . Thus, the only possible generators for principal
reductions of m are x′ , y′ , x′ + y′ , but each of these is in a minimal prime ideal of
R. However, any reduction of m will be an m-primary ideal, and thus m has no
principal reductions.
Although the ring R in Example 4.2 is not a domain, it can be used to produce
similar examples that are domains. We recall a theorem of Lech [7, Theorem 1]:
A complete local Noetherian ring R with maximal ideal m is the completion of a
local Noetherian domain if and only if (i) m = 0 or m 6∈ Ass(R), and (ii) no nonzero
integer of R is a zerodivisor.
14
L. FOULI AND B. OLBERDING
Example 4.3. A one-dimensional local Noetherian domain whose maximal ideal
does not have a principal reduction. Let R be as in Example 4.2. Since R is CohenMacaulay, m 6∈Ass(R), and since the only nonzero integer of R is 1, it follows from
the theorem of Lech that there exists a local Noetherian domain A with completion
R. As in the preceding example, the maximal ideal m of R does not have a principal
reduction. Since the maximal ideal of R is extended from the maximal ideal of A, it
follows that the maximal ideal of A does not have a principal reduction. Moreover,
since R is a complete intersection of multiplicity 3, so is A.
The following example was suggested to us by Bill Heinzer.
Example 4.4. Let R = Z2 [[x, y, z]]/(z2 − xy, xy(x + y)(x + y + z)). One can show
that Min(R) = {(x, z), (y, z), (x + z, y + z), (x + y + z)}. Every linear form belongs to
some minimal prime of R and hence there is no principal reduction of the maximal
ideal m = (x, y, z). On the other hand, the number of minimal primes is 4 and
thus the number of maximal ideals of Rred is 4 by Remark 3.2. Since also |k| < 4,
Corollary 3.3 and Remark 3.2 imply that there exists an ideal of R that does not
have a principal reduction. In this case, m = (x, y, z)R is such an ideal.
The following example was suggested to us by Bernd Ulrich.
Example 4.5. Let R = Z2 [[x, y, z]]/(z2 −xy, xy(x+y)). One can show that Min(R) =
{(x, z), (y, z), (x + z, y + z)}. It is straightforward to show that J1 = (x + y + z) is the
only principal reduction of m = (x, y, z). The number of minimal primes is 3 and
even though |k| < 3 it is the case that m has a principal reduction. By Remark 3.2,
Corollary 3.3 and Remark 3.2 there must exist a different ideal of R that does not
have a principal reduction. One can verify that J2 = (x, y) is such an ideal. Moreover, J2 is another proper reduction of m and there is no principal reduction contained in J2 . Hence there are minimal reductions of m having minimal generating
sets of different sizes.
All the examples in this section have been devoted to the one-dimensional case.
In the next example, we illustrate the failure of 2-generated reductions for the maximal ideal of a two-dimensional ring. As mentioned in Remark 3.9, Heinzer, Ratliff
and Rush [3, Example 2.3] have given for each finite field F an example of a
two-dimensional Cohen-Macaulay local ring R such that R has residue field F, the
associated graded ring of m is Cohen-Macaulay, and the maximal ideal m of R fails
to have a 2-generated reduction. For the case F = Z2 , we give a simple example of
this phenomenon that has the same properties.
REDUCTIONS OF IDEALS
15
Example 4.6. Let R = Z2 [[x, y, z]]/(xy(x+y)(x+y+z)). Then R is a 2-dimensional
Cohen-Macaulay ring and the associated graded ring of m, grm (R) =
∞
L
mi /mi+1 ,
i=0
is Cohen-Macaulay. We claim that m = (x, y, z)R does not have a 2-generated reduction. Indeed, as in Example 4.2, if m has a 2-generated reduction, then one can
assume that the two generators of the reduction are linear forms. Since the residue
field is Z2 there are only 7 linear forms in x, y, and z. It is straightforward to check
that any ideal in R generated by two linear forms has height 1 and therefore cannot
be a reduction of m.
ACKNOWLEDGEMENTS
We thank Bill Heinzer and Bernd Ulrich for suggesting Examples 4.4 and 4.5,
which served as our original motivation for this article.
R EFERENCES
[1] J. Dieudonné, Topics in local algebra, Notre Dame Mathematical Lectures 10, University of
Notre Dame Press, 1967.
[2] M. Fontana, J. Huckaba, and I. Papick, Prüfer Domains, Monographs and Textbooks in Pure
and Applied Mathematics 203, Marcel Dekker, New York, 1997.
[3] W. Heinzer, L. J. Ratliff, Jr., and D. E. Rush, Reductions of ideals in local rings with finite
residue fields, Proc. Amer. Math. Soc., 138 (2010) 1569–1574.
[4] W. Heinzer and D. Shannon, Abhyankar’s work on dicritical divisors, available at
arXiv:1707.06733 [math.AC].
[5] J. Huckaba, Commutative rings with zero divisors, Monographs and Textbooks in Pure and
Applied Mathematics 117, Marcel Dekker, New York, 1988.
[6] D. Katz, On the number of minimal prime ideals in the completion of a local domain, Rocky
Mountain J. Math. 16 (1986), 575–578.
[7] C. Lech, A method for constructing bad noetherian local rings, Lecture Notes in Math. 1183,
Springer, Berlin, 1986, 241–247.
[8] J. Lipman, Stable ideals and Arf rings, Amer. J. Math. 93 (1971), 649–685.
[9] H. Matsumura, Commutative ring theory, Cambridge University Press, 1989.
[10] D.G. Northcott and D. Rees, Reductions of ideals in local rings, Proc. Camb. Phil. Soc. 50
(1954), 145–158.
[11] J. Sally, Numbers of generators of ideals in local rings, Marcel Dekker, Inc., New York-Basel,
1978.
[12] J. Sally, W. Vasconcelos, Stable rings, J. Pure Appl. Algebra 4 (1974), 319–336.
[13] I. Swanson and C. Huneke, Integral Closure of Ideals, Rings, and Modules, London Mathematical Society Lecture Note Series 336, Cambridge University Press, Cambridge, (1994).
D EPARTMENT OF M ATHEMATICAL S CIENCES , N EW M EXICO S TATE U NIVERSITY, L AS C RUCES ,
N EW M EXICO 88003, USA
E-mail address: [email protected]
D EPARTMENT OF M ATHEMATICAL S CIENCES , N EW M EXICO S TATE U NIVERSITY, L AS C RUCES ,
N EW M EXICO 88003, USA
E-mail address: [email protected]
| 0math.AC
|
Zeitschrift für Angewandte Mathematik und Mechanik, 18/4/2016
C 1 -continuous space-time discretization based on
Hamilton’s law of varying action
arXiv:1510.07863v2 [math.NA] 15 Apr 2016
Janine C. Mergel1 , Roger A. Sauer1,∗ , and Sina Ober-Blöbaum2
1
AICES Graduate School, RWTH Aachen University, Templergraben 55, 52056 Aachen, Germany
2
Department of Engineering Science, University of Oxford, Parks Road, Oxford, OX1 3PJ, United Kingdom
Received XXXX, revised XXXX, accepted XXXX
Published online XXXX
Key words Hamilton’s law of varying action, Hermite interpolation, nonlinear elastodynamics, symplectic integration,
variational integrators
MSC (2010) 37M15, 65P10, 70H25, 74B20, 74H15, 74S05
We develop a class of C 1 -continuous time integration methods that are applicable to conservative problems in elastodynamics. These methods are based on Hamilton’s law of varying action. From the action of the continuous system we derive
a spatially and temporally weak form of the governing equilibrium equations. This expression is first discretized in space,
considering standard finite elements. The resulting system is then discretized in time, approximating the displacement by
piecewise cubic Hermite shape functions. Within the time domain we thus achieve C 1 -continuity for the displacement
field and C 0 -continuity for the velocity field. From the discrete virtual action we finally construct a class of one-step
schemes. These methods are examined both analytically and numerically. Here, we study both linear and nonlinear systems as well as inherently continuous and discrete structures. In the numerical examples we focus on one-dimensional
applications. The provided theory, however, is general and valid also for problems in 2D or 3D. We show that the most
favorable candidate — denoted as p2-scheme — converges with order four. Thus, especially if high accuracy of the numerical solution is required, this scheme can be more efficient than methods of lower order. It further exhibits, for linear
simple problems, properties similar to variational integrators, such as symplecticity. While it remains to be investigated
whether symplecticity holds for arbitrary systems, all our numerical results show an excellent long-term energy behavior.
Copyright line will be provided by the publisher
1
Introduction
In this work we derive a class of time integration methods for the computational analysis of deformable solids. We consider
a discrete version of Hamilton’s law of varying action to obtain space-time discretization schemes based on cubic Hermite
functions in time.
∗
Corresponding author
E-mail: [email protected], Phone: +49 241 80 99129, Fax: +49 241 80 628498
Copyright line will be provided by the publisher
J.C. Mergel, R.A. Sauer, and S. Ober-Blöbaum: C 1 -continuous space-time discretization based on Hamilton’s law
2
1.1
Overview of existing methods
One very common approach for the numerical analysis of elastodynamic problems is the application of so-called semidiscrete procedures: Here, the (spatially and temporally) continuous system is discretized in space and time separately.
First, the mechanical equilibrium equations, describing the deformation of the body, are discretized in space by means of
the finite element method (FEM). At this point we refer to any standard literature on nonlinear finite elements for solids; see
e.g. Ref. [8, 78, 79]. The spatially discrete system is then discretized in time, using for instance a finite difference scheme
or collocation based on Taylor series expansion. Discretization schemes of this type include many well-known methods,
such as the Newmark algorithm [61], the HHT-α method [31], the WBZ-α (or Bossak-α) method [77], and the generalizedα method [14]. Besides methods fulfilling the equilibrium equations at single time steps, there exist various approaches
based on weighted residuals; these consider equilibrium in an weighted-average sense and go back to the publication by
Zienkiewicz [80]. A weighted residual approach based on cubic Hermite interpolation in time, for instance, is discussed in
Ref. [21]. See also the generalized method proposed by Modak and Sotelino [60]. In addition, there exists a broad literature
on methods combining finite elements (FE) in both space and time. In general, these solution schemes can be constructed
by forming a (spatially and temporally) weak form of the equations of motion, and discretizing the resulting statement by
means of finite elements. The first approaches accounting for finite elements in time go back to Ref. [1, 20, 65]. Early
publications on space-time FE methods include Ref. [36, 38, 39]. A broader literature review on time integration methods
in structural mechanics can be found e.g. in Ref. [7, 10, 25, 45].
A special class of time integration schemes applied to mechanical systems is formed by geometrical integrators. Geometric integration enables the design of robust methods that provide both quantitatively and qualitatively accurate results.
Since these methods preserve the geometric properties of the flow of a differential equation, they are able to exactly represent the main characteristic properties of the physical process [26, 59, 67]. Geometric integration methods can be mainly
divided into two classes: 1) energy-momentum integrators and 2) symplectic momentum-conserving integrators. The
first class of methods fulfills the conservation laws of energy and momentum automatically; for methods of this type we
refer e.g. to Simo and Tarnow [72], Simo et al. [73], Gonzalez [24], Betsch and Steinmann [10], Groß et al. [25], Leyendecker et al. [54], Hesch and Betsch [29,30], Gautam and Sauer [22] and Krenk [44], Betsch and Janz [9], and the references
therein. See also the generalized energy-momentum method discussed in Ref. [45–47]. The second class preserves both
the symplectic form and — in the presence of symmetries — momentum maps; it additionally shows excellent long-term
energy behavior. Symplectic-momentum integrators can be represented by the class of variational integrators [59, 74]. For
conservative systems, these methods are constructed by forming a discrete version of Hamilton’s principle, choosing both a
finite-dimensional function space and a suitable numerical quadrature. See Ref. [52, 53] for an overview. For dissipative or
controlled mechanical systems, they can be derived from a discrete version of the Lagrange-d’Alembert principle [41, 62].
Within the last years variational integrators have been extended towards constrained [15, 42, 55, 57], non-smooth [19, 40],
stochastic [12], multirate and multiscale [56, 75] systems as well as to electric circuits [64]. For variational integrators in
combination with spatial discretization we refer to Ref. [18, 51, 76] and the references therein. Besides these semi-discrete
approaches, there exists a covariant space-time discretization method by Marsden et al. [58]. This multi-symplectic scheme
allows for symplecticity in both space and time.
Copyright line will be provided by the publisher
ZAMM header will be provided by the publisher
3
In most of the previously mentioned work the solution is approximated by using piecewise Lagrange interpolation.
For a mechanical system, this leads to a smooth approximation of the position, but to discontinuities in the velocity at
the discrete time steps. Besides this approach, Leok and Shingel [50] have developed a variational integrator based on
piecewise Hermite interpolation. In their prolongation-collocation approach, not only the solution of the discrete EulerLagrange equations, but also its time derivatives are approximated with sufficient accuracy. This leads to a globally smooth
approximation of the solution. Note that Ref. [50] does not include the combined discretization in both space and time.
To incorporate any initial conditions of the mechanical system explicitly, early publications on structural dynamics have
considered the so-called Hamilton’s law of varying action; see e.g. Argyris and Scharpf [1], Fried [20], Bailey [2–5],
Simkins [70, 71], and Borri et al. [11]. This law can be regarded as a generalization of Hamilton’s principle: It accounts for
any initial and final velocities by considering non-zero variations in the displacement at the boundaries of the time domain.
Some of the studies mentioned above also include cubic Hermite interpolation in time for the displacement. Based on this
law, a family of methods has been proposed [6, 69] that combines different zero-variations of the displacement or velocity
at initial and final time. In addition, Géradin [23] has constructed from a subsequent application of Hamilton’s law a time
integration method based on Hermite interpolation. While variational integrators are automatically symplectic (due to the
discretized action integral serving as generating function [26]), it is not clear if the same properties hold for integration
schemes constructed from Hamilton’s law.
1.2
Objectives
In this paper we derive a class of space-time integration methods that are based on piecewise cubic Hermite interpolation
in time. To this end, we consider Hamilton’s law of varying action. We thus directly incorporate the additional boundary
terms (arising from the non-zero variations) into our time integration method. Using a semi-discrete approach, we first
discretize our resulting equilibrium equation in space and then in time. Instead of deriving additional conditions on the
time derivatives of the approximated solution (as it is done in Ref. [50]) we consider independent variations of the position
and velocities.
In general, one could then construct a variational integrator by 1) varying the action of the entire temporal domain,
and 2) deriving from the variation a set of discrete Euler-Lagrange equations. Since for a cubic Hermite approximation,
however, this would lead to an (unconditionally) unstable numerical method, we pursue a different approach: We vary the
action for each discrete time interval individually, which leads to an overdetermined system of four equations. By choosing
different combinations of equations, we derive a family of six different one-step methods. One of these schemes coincides
with the method proposed by Géradin [23].
In fact, our time integration methods are not variational in the sense that they are not derived from the virtual action of
the total time domain. We will demonstrate numerically, however, that the most favorable of our schemes — denoted in the
following as p2-scheme — shows similar properties like true variational integrators: an excellent long-term behavior and,
for a simple harmonic oscillator, symplecticity. Interestingly, this is not the case for the variant discussed in Ref. [23].
We emphasize that the aim of this work is to both present the construction of our time integration methods, and to
demonstrate their most important features by means of various numerical examples. A further analytical investigation
Copyright line will be provided by the publisher
J.C. Mergel, R.A. Sauer, and S. Ober-Blöbaum: C 1 -continuous space-time discretization based on Hamilton’s law
4
(including e.g. the proof of the order of convergence) goes beyond the scope of this paper; instead, this should be addressed
in future work.
Note that like other methods based on C 1 -continuous approximations in time, our integration schemes are not favorable
for the simulation of discontinuous changes (such as shock waves) in mechanical systems. Instead, we apply the p2scheme to temporally smooth examples; these include both linear and nonlinear as well as intrinsically discrete and spatially
continuous problems. Here, we focus on one-dimensional applications; the theory, however, is also valid for conservative
systems in 2D and 3D. Compared to both the formulation of Géradin and the method of Leok and Shingel based on cubic
Hermite interpolation, our p2-scheme exhibits a higher order of convergence. In addition, based on the desired accuracy, it
may be more efficient than classical methods like the Newmark algorithm.
1.3
Outline
The remainder of this paper is structured as follows. Section 2 introduces the action integral of a continuous body deforming over time. From Hamilton’s law of varying action a (spatially and temporally) weak form of the mechanical equilibrium
equation is derived. Section 3 briefly outlines its spatial discretization by means of standard finite elements. The temporal
discretization is discussed in Section 4, providing a solution strategy that leads to a class of different integration methods.
These methods are then related to other approaches from the literature. In Section 5, we study the main characteristic properties of our integration schemes, such as symplecticity and the long-term and convergence behavior. The most favorable
scheme is then applied to investigate several numerical examples (Section 6). Section 7 finally concludes this paper.
2
Hamilton’s law of varying action
In this section we summarize the governing equations describing a body undergoing finite motion and deformation. For
the general theory of continuum mechanics, the reader is referred to text books [13, 33]. Consider a body deforming within
the time domain [0, T ]. In the initial configuration, at t = 0, the body is denoted by B0 ; its boundary is denoted by ∂B0 .
The body can be subjected to volumetric loads, B̄ (applied in B0 ), deformations, ϕ̄ (prescribed on ∂ϕ B0 ⊆ ∂B0 ), and
surface loads, T̄ (applied on ∂t B0 ⊆ B0 ).1 At any time t ∈ (0, T ], the deformation of the body is characterized by a unique
mapping of a material point, X ∈ B0 , to its current position, x = ϕ(X, t) ∈ B. The material time derivative v := ∂x/∂t
corresponds to the velocity of a material particle located at x; in short we will also write v = ẋ.
2.1
Action of the continuous system
In the following we assume 1) conservation of mass, 2) hyperelastic material behavior, and 3) that the external forces do
not depend on the deformation. We start with the action integral of the continuous system, defined as
Z T
S=
L(x, v) dt.
(1)
0
The integrand corresponds to the Lagrangian of our system, given by
L(x, v) = K(v) − Π(x).
(2)
1 We assume that ∂ B ∪ ∂ B = ∂B and ∂ B ∩ ∂ B = ∅.
ϕ 0
ϕ 0
t 0
t 0
0
Copyright line will be provided by the publisher
ZAMM header will be provided by the publisher
5
Here, K is the kinetic energy, and Π is the potential energy due to both internal strains and external forces;
Z
1
K(v) =
ρ0 v · v dV,
2 B0
Z
Z
Z
Π(x) =
W (x) dV −
x · ρ0 B̄ dV −
x · T̄ dA.
B0
B0
(3)
(4)
∂t B0
The terms ρ0 and W (x) respectively denote the initial material density and an energy density function characterizing the
material behavior; a detailed description of W (x) for different material models can be found e.g. in Ref. [8, 78, 79].
2.2
Variation of the action
We now consider an admissible variation of the deformation, δx ∈ V,
n
o
V = δx : B0 × [0, T ] → Rd δx(X, t)|∂ϕ B0 = 0 .
(5)
Here, d is the dimension of Euclidean space. Varying the action integral (1) yields
Z T
δS =
δL(x, v) dt,
δL(x, v) = δK(v) − δΠ(x),
(6)
0
where the variations of the energy terms are given by
Z
δK(v) =
δv · ρ0 v dV,
Z
ZB0
Z
δx · ρ0 B̄ dV −
δΠ(x) =
grad δx : σ dv −
B
B0
(7)
δx · T̄ dA.
(8)
∂t B0
The tensor σ denotes the Cauchy stress, which is derived from the strain energy density function, W , appearing in Eq. (4).
If we enforce the deformation at t = 0 and t = T to be fixed (s.t. the variations δx(X, 0) and δx(X, T ) become zero),
we arrive at the classical Hamilton’s principle,
δS = 0
n
o
∀ δx ∈ δx ∈ V δx(X, 0) = 0, δx(X, T ) = 0 ,
(9)
see e.g. the book by Lanczos [48]. Instead, however, we leave the variations δx(X, 0) and δx(X, T ) arbitrary for now. In
this case, δS is equal to the following boundary term evaluated at t = 0 and t = T :
T
∂L(x, v)
∀ δx ∈ V.
δS = δx ,
∂v
0
(10)
The term in angle brackets corresponds to the scalar product of the variation δx and the linear momentum of the body, i.e.
Z
∂L(x, v)
δx ,
=
δx · ρ0 v dV.
(11)
∂v
B0
A more detailed discussion of arbitrary variations at the boundaries can be found in Ref. [48] (see Chapter V.3 there).
Due to the scalar product, Eq. (10) depends on the system’s initial momentum (and thus on its initial velocity) explicitly.
This is the reason why expression (10) has been discussed in several early publications studying initial value problems in
structural dynamics, such as Ref. [1–3, 6, 70, 71]; an overview can also be found in Ref. [66]. Following the terminology
used in several of these papers, we refer to Eq. (10) as “Hamilton’s law of varying action”.2 A very interesting comment
2 Alternatively, this equation has been referred to as “Hamilton’s weak principle” (HWP); see e.g. Ref. [32].
Copyright line will be provided by the publisher
J.C. Mergel, R.A. Sauer, and S. Ober-Blöbaum: C 1 -continuous space-time discretization based on Hamilton’s law
6
on its origin can be found in a paper by Bailey (Ref. [3], p. 434): “When copies of Hamilton’s original papers [27, 28] were
obtained, it was found that Hamilton had furnished what he called the ‘law of varying action’. He did not furnish what is
now known as ‘Hamilton’s principle’. Evidently, in the latter part of the 19th century, application of the concepts of the
variational calculus of Euler and Lagrange reduced Hamilton’s law to Hamilton’s principle.”
We will explain later why the two additional boundary terms (caused by the non-zero variations) are important for the
derivation of our final integration schemes; see Section 4.2. Equation (10) finally represents the (spatially and temporally)
weak form of the governing equilibrium equations. Note that this expression is general and valid for elastodynamic problems in Rd , d ∈ {1, 2, 3}. Nevertheless, since we focus on the development and analysis of a new time integration method,
we numerically investigate only one-dimensional problems with B̄ = T̄ = 0 in this paper. A detailed study of problems
in both 2D or 3D may be the subject of future work.
3
Spatial discretization
We now briefly outline the spatial discretization by means of the finite element method. Regarding nonlinear finite elements
for solids we refer to text books such as Ref. [8, 78, 79]. We spatially discretize Eq. (10) by using nel finite elements; for
each element, Ωe , the initial position, X, the deformation, x(X, t), and the velocity, v(X, t), are approximated by
X h (X) = Ne (X) Xe ,
xh (X, t) = Ne (X) xe (t),
v h (X, t) = Ne (X) ẋe (t),
where the dot indicates the derivative with respect to time. The vectors
ẋ1 (t)
X1
x1 (t)
.
.
.
ẋe (t) = ..
Xe = .. ,
xe (t) = .. ,
ẋnne (t)
xnne (t)
Xnne
(12)
(13)
contain the initial and current positions as well as the velocities of those nne nodes belonging to element Ωe . These
quantities are still continuous with respect to time. The array Ne = N1 I d , . . . , Nnne I d contains the nodal shape
functions N1 – Nnne associated with Ωe . Using an isoparametric concept, we discretize the variations by means of the
same shape functions,
δxh (X, t) = Ne (X) δxe (t),
δv h (X, t) = Ne (X) δ ẋe (t).
(14)
We can further write
grad δxh = Be (X) δxe (t)
(15)
for a suitable definition of the strain operator Be . By inserting these relations into Eq. (10) and (11), we obtain
δS h =
nel h
iT
X
δxT
m
ẋ
,
e
e
e
(16)
0
e=1
where δS h can be computed from
δS h =
Z
0
T
δLh (x, ẋ) dt,
δLh (x, ẋ) =
nel
X
T
e
δ ẋe me ẋe − δxT
;
e f
(17)
e=1
Copyright line will be provided by the publisher
ZAMM header will be provided by the publisher
7
cf. Eq. (6) – (8). Here, the vectors x(t) and ẋ(t) denote the deformation and velocity at all spatial FE nodes. The elemental
e
e
mass matrices, me , and the force vectors, f e := fint
− fext
, are computed through
Z
me :=
ρ0 NT
e Ne dV,
(18)
Ωe0
e
fint
:=
e
fext
:=
Z
Ωe
Z
Ωe0
BT
e σ dv,
(19)
ρ0 NT
e B̄ dV +
Z
Γe0t
NT
e T̄ dA,
(20)
introducing Γe0t = Ωe0 ∩ ∂t B0h . To obtain a shorter notation we will later refer to the global mass matrix, m, and the
global force vectors, f := fint − fext , assembled from the elemental contributions. Equation (16) finally corresponds to the
spatially discrete version of Hamilton’s law of varying action.
4
Temporal discretization
We now discretize the (spatially discrete) virtual action, δS h , in time. To achieve temporal C 1 -continuity, we approximate
the nodal deformation of element Ωe by cubic Hermite shape functions; xe (t) ≈ xte (t) with
e
xte (t) = R1 (t) x̂en + R2 (t) x̂en+1 + H1 (t) v̂ne + H2 (t) v̂n+1
,
t ∈ [tn , tn+1 ],
(21)
where n = 0, . . . , N −1, t0 = 0, and tN = T . The vectors x̂en and x̂en+1 contain the nodal deformations at tn and tn+1 ;
e
are the corresponding nodal velocities. See Appendix A.1 for the definition of the shape functions
the vectors v̂ne and v̂n+1
R1 , R2 , H1 , and H2 . We can further write
e
ẋte (t) = Ṙ1 (t) x̂en + Ṙ2 (t) x̂en+1 + Ḣ1 (t) v̂ne + Ḣ2 (t) v̂n+1
,
(22)
e
δxte (t) = R1 (t) δx̂en + R2 (t) δx̂en+1 + H1 (t) δv̂ne + H2 (t) δv̂n+1
,
(23)
e
δ ẋte (t) = Ṙ1 (t) δx̂en + Ṙ2 (t) δx̂en+1 + Ḣ1 (t) δv̂ne + Ḣ2 (t) δv̂n+1
.
(24)
In order to improve readability, we will also use the assembled counterparts accounting for all spatial FE nodes at once. We
will denote them e.g. by x̂n instead of x̂en ; see the analogy to x(t) / xe (t) introduced in the previous section.
4.1
Virtual action for a single time interval
Approximations (21) – (24) are now inserted into the virtual action for a single time interval, [tn , tn+1 ]; this yields
ht
δSn+1
=
Z
tn+1
δLht xt , ẋt dt,
T
T
δLht xt , ẋt = δ ẋt m ẋt − δxt f xt .
(25)
tn
ht
depends on four variables: x̂n , x̂n+1 , v̂n , and v̂n+1 ; it can thus be reformulated to
The increment δSn+1
ht
δSn+1
= δx̂n ·
ht
ht
ht
ht
∂Sn+1
∂Sn+1
∂Sn+1
∂Sn+1
+ δx̂n+1 ·
+ δv̂n ·
+ δv̂n+1 ·
.
∂ x̂n
∂ x̂n+1
∂ v̂n
∂ v̂n+1
(26)
Following the terminology of Marsden and West [59], we now define the discrete momenta
p̂−
n := −
ht
∂Sn+1
,
∂ x̂n
p̂+
n+1 :=
ht
∂Sn+1
.
∂ x̂n+1
(27)
Copyright line will be provided by the publisher
J.C. Mergel, R.A. Sauer, and S. Ober-Blöbaum: C 1 -continuous space-time discretization based on Hamilton’s law
8
We further introduce two analogous variables that appear due to the Hermite discretization (21),
q̂−
n := −
ht
∂Sn+1
,
∂ v̂n
q̂+
n+1 :=
ht
∂Sn+1
.
∂ v̂n+1
(28)
+
Since q̂−
n and q̂n+1 have the unit “momentum × time”, we refer to them as discrete “pseudo-momenta”. The four terms
ht
can be computed from the discrete action, Sn+1
, given in Appendix A.2; this results in
p̂−
n =−
q̂−
n =−
Z
tn+1 h
tn
Z tn+1 h
Ṙ1 (t) m ẋt − R1 (t) f xt
i
Ḣ1 (t) m ẋt − H1 (t) f xt
i
dt,
dt,
p̂+
n+1 =
q̂+
n+1 =
Z
tn+1 h
tn
Z tn+1 h
Ṙ2 (t) m ẋt − R2 (t) f xt
i
dt,
(29)
i
dt.
(30)
Ḣ2 (t) m ẋt − H2 (t) f xt
tn
tn
By inserting these (pseudo-)momenta, we can simplify Eq. (26) to
+
+
ht
−
δSn+1
= −δx̂n · p̂−
n + δx̂n+1 · p̂n+1 − δv̂n · q̂n + δv̂n+1 · q̂n+1 .
(31)
ht
In addition, since the variations δx̂n and δx̂n+1 remain arbitrary, δSn+1
must fulfill Hamilton’s law of varying action
applied to the interval [tn , tn+1 ],
ht
δSn+1
= δx̂n+1 · m v̂n+1 − δx̂n · m v̂n .
(32)
After inserting Eq. (31), we finally obtain the spatially and temporally discrete version of Eq. (10) for a single time interval:
h
i
h
i
h
i
h
i
+
+
−
δx̂n · m v̂n − p̂−
n + δx̂n+1 · p̂n+1 − m v̂n+1 + δv̂n · −q̂n + δv̂n+1 · q̂n+1 = 0
(33)
for arbitrary variations δx̂n , δx̂n+1 , δv̂n , and δv̂n+1 . Note that we derive Eq. (33) by first varying the continuous action
integral, and then discretizing its variation in space and time. We would, however, obtain the same expression if we first
discretized the action itself, and then varied the discrete action for a single time interval.
4.2
Solution strategy
In general, variational integrators are constructed by 1) summing up the discrete action for all time intervals, 2) taking its
variation, and 3) re-arranging the summands. Doing so for Eq. (33), we would arrive at the following statement:
−1
h
i NX
h
i
h
i
+
−
δx̂0 · m v̂0 − p̂−
δx̂n · p̂+
n − p̂n + δx̂N · m v̂N − p̂N
0 +
n=1
−1
h
i NX
h
i
h i
+
−
+ δv̂0 · − q̂−
+
δv̂n · q̂+
n − q̂n + δv̂N · q̂N = 0.
0
(34)
n=1
This expression is equivalent to Eq. (33) summed up over the entire time domain. Solving Eq. (34) subsequently, however,
results in a two-step method that is unconditionally unstable (i.e. for which the spectral radius is larger than one). This
observation is also discussed in Ref. [68]. Instead, we develop a class of one-step methods arising from the virtual action
for the individual time interval [tn , tn+1 ]. If the virtual displacements and velocities are presumed to be arbitrary, Eq. (33)
Copyright line will be provided by the publisher
ZAMM header will be provided by the publisher
9
provides us with (4 · d nno ) equations,
p̂−
n = m v̂n ,
(35)
p̂+
n+1 = m v̂n+1 ,
(36)
q̂−
n = 0,
(37)
q̂+
n+1 = 0,
(38)
where nno is the number of finite element nodes. Physically, the first two equations relate the discrete momenta, p̂−
n
and p̂+
n+1 , to the linear momenta at tn and tn+1 ; see Fig. 1. The second two equations arise from the chosen Hermite
approach. Assuming that the displacement and velocity of the previous time step, x̂n and v̂n , are given, we need only
given:
find:
interval
Fig. 1 Equilibrium equations for time interval [tn , tn+1 ].
(2 · d nno ) equations to determine the new state, x̂n+1 and v̂n+1 . The system (35) – (38) is thus over-determined. For
this reason, we set two of the (so far arbitrary) variations, δx̂n , δx̂n+1 , δv̂n , and δv̂n+1 , to zero; this approach is further
motivated in the following. The new deformation and velocity are then computed from the remaining two equations. We
finally obtain six methods, which are illustrated in Fig. 2.
p2-scheme:
δv̂n = δv̂n+1 = 0
p̂−
n = m v̂n ,
p̂+
n+1 = m v̂n+1 .
(39)
This seems to be the most promising approach: From its definition follows that it enforces a matching of the momenta
+
at the discrete time steps, i.e. that p̂−
n = p̂n = m v̂n .
q2-scheme:
δx̂n = δx̂n+1 = 0
q̂−
n = 0,
q̂+
n+1 = 0.
(40)
This method can be seen as the counterpart of the p2-scheme.
p+ q− -scheme:
δx̂n = δv̂n+1 = 0
p̂+
n+1 = m v̂n+1 ,
q̂−
n = 0.
(41)
This is one of four mixed methods, varying once the displacement and once the velocity.
Copyright line will be provided by the publisher
J.C. Mergel, R.A. Sauer, and S. Ober-Blöbaum: C 1 -continuous space-time discretization based on Hamilton’s law
10
p+ q+ -scheme:
δx̂n = δv̂n = 0
p̂+
n+1 = m v̂n+1 ,
q̂+
n+1 = 0.
(42)
This scheme corresponds to the formulation proposed by Géradin [23].
p− q− -scheme:
δx̂n+1 = δv̂n+1 = 0
p̂−
n = m v̂n ,
p− q+ -scheme:
q̂−
n = 0.
(43)
δx̂n+1 = δv̂n = 0
p̂−
n = m v̂n ,
q̂+
n+1 = 0.
(44)
(a) p2-scheme
(b) q2-scheme
(c) p+ q− -scheme
(d) p+ q+ -scheme
(e) p− q− -scheme
(f) p− q+ -scheme
Fig. 2 Illustration of the six integration schemes.
Interestingly, the resulting six methods have completely different characteristics; it is not surprising that some of them are
more favorable than others. In Section 5 we numerically investigate the properties of each scheme in terms of preservation
of energy and convergence behavior. Furthermore, for linear systems we analyze the stability and symplecticity of the
schemes. For several reasons we especially focus on the first, i.e. on the p2-scheme. Here, the discrete velocities, v̂n and
v̂n+1 — appearing in the Hermite ansatz (21) — are connected to the displacements by setting the linear momenta, m v̂n
+
and m v̂n+1 , equal to the discrete momenta, p̂−
n and p̂n+1 . Note that this approach is only possible because we explicitly
Copyright line will be provided by the publisher
ZAMM header will be provided by the publisher
11
account for the boundary terms appearing in Hamilton’s law of varying action. Second, since from definition (39) directly
−
+
−
follows that p̂+
n = p̂n , the discrete Euler-Lagrange equations, p̂n − p̂n = 0, are fulfilled automatically (cf. Chapter
VI, Eq. (6.7) of Hairer et al. [26]). Interestingly, the p2-scheme satisfies balance of linear momentum (a generalization of
conservation of momentum) averaged over the time step (Appendix A.4):
Z
tn+1
m ẍt + f xt dt = 0;
(45)
tn
Nevertheless, we must point out that our six integrators are not variational; this becomes apparent especially for the four
mixed methods, where we use one zero-variation for the displacement, and one for the velocity. We will show, however,
that — at least for simple linear problems — the first two methods (39) and (40) have similar properties like variational
integrators, such as symplecticity. In addition, we numerically demonstrate that these methods show a very good energypreserving behavior even for nonlinear problems with multiple degrees of freedom. In future work it would be interesting
to compare these schemes with variational integrators in more detail, and to investigate symplecticity for arbitrary systems.
It would be further interesting to examine whether the p2-schemes conserves momentum maps associated with symmetries
of the Lagrangian.
4.3
Implementation
In general, equations (35) – (38) are nonlinear. They thus must be linearized by using e.g. Newton’s method; this provides
a system of linear equations that is iteratively solved for the new positions and velocities of the finite element nodes.
For linearization, the derivatives of the discrete (pseudo-)momenta are required; see Appendix A.3. In analogy to the
+
+
−
force vectors and the mass matrix, the terms p̂−
n , p̂n+1 , q̂n , and q̂n+1 can be computed by assembling the contributions
+
+
−
of each spatial element, denoted by p̂−
e,n , p̂e,n+1 , q̂e,n , and q̂e,n+1 . Where possible, the integrals should be computed
analytically; this can be done for the contributions due to 1) the kinetic energy and due to 2) any linear elastic internal
energy (Appendix A.3). The remaining integrals are evaluated by Gaussian quadrature, choosing a sufficient number of
quadrature points. For our time integration schemes, neither the kinetic/potential energy nor the total energy of the system
must be evaluated explicitly. Since we want to investigate them in our numerical examples, however, we discuss these
quantities in Appendix A.2.
4.4
Relation to other methods
The idea of applying Hamilton’s law of varying action to initial value problems in structural dynamics goes back to the
first approaches using finite elements in both space and time; see e.g. Ref. [1, 2, 20]. Instead of zero-variations of the
displacement at initial and final time (as required for Hamilton’s principle), these publications account for zero-variations
of both the initial displacement and velocity: δx(0) = δ ẋ(0) = 0.
This idea has motivated Baruch and Riff [6] to combine different zero-variations of either the displacement or velocity
at both t = 0 and t = T . Their approach results in six different methods that can be related — with several important
differences — to our schemes. Since the same authors have discovered in a previous work [68] the instability of the solution
scheme given by Eq. (34), they propose a modified discretization of the virtual displacements in Ref. [69]. In their approach,
δx(t) is discretized by considering the second derivatives of the shape functions, R̈• (t) and Ḧ• (t); the variation of the
Copyright line will be provided by the publisher
J.C. Mergel, R.A. Sauer, and S. Ober-Blöbaum: C 1 -continuous space-time discretization based on Hamilton’s law
12
displacement is thus approximated by a linear (instead of a cubic) function in time. This modification leads to different
partial derivatives of the action, and therefore to a different integration method. An even more important difference to our
schemes concerns the zero-variations at the boundaries: While Riff and Baruch [69] consider the boundaries of the entire
time domain (t = 0 and t = T ), we derive our schemes from zero-variations within each time interval, [tn , tn+1 ]. Our
approach results in six different one-step methods solving the equations subsequently. In contrast, this is possible only for
the so-called F4-method of Ref. [6, 69], where δx(T ) = δ ẋ(T ) = 0. For the remaining formulations in Ref. [6, 69], all
equations would have to be solved simultaneously. In summary, one could loosely relate our six schemes to a subsequent
application of the methods by Riff and Baruch [69] for each time interval, [tn , tn+1 ]. The underlying equations, however,
are approximated differently. Besides that, the references mentioned above discuss only linear dynamic systems (where the
forces depend on the displacement linearly).
Recently, Leok and Shingel [50] have proposed a variational integrator based on Hermite finite elements in time. Their
formulation is derived from a prolongation-collocation approach: In addition to the discrete Euler-Lagrange equations this
method accounts for the system’s equation of motion in strong form,
m ẍt (t• ) + f xt (t• ) = 0,
• ∈ {n, n + 1}.
(46)
For cubic Hermite shape functions — as they are used in our schemes — the velocities, v̂n and v̂n+1 , are computed from
Eq. (46), using
ẍt (t) = R̈1 (t) x̂n + R̈2 (t) x̂n+1 + Ḧ1 (t) v̂n + Ḧ2 (t) v̂n+1 ;
(47)
t
cf. Eq. (22). These expressions are inserted into the temporally discrete action for one time interval, Sn+1
, which then
depends only on the displacements, x̂n and x̂n+1 . The final time integration method of Leok and Shingel is obtained by
t
1) varying the incremental action, Sn+1
, with respect to the displacements, and 2) setting the total virtual action to zero.
Compared to our six Hermite formulations, the resulting method requires only half the number of unknowns to be solved
within each time step. Its rate of convergence, however, is lower than the best of our schemes; see Section 5.4. Note that
the combination of both spatial and temporal discretizations is not discussed in Ref. [50].
5
Properties of the six schemes
We now investigate the different properties of the six formulations, first focusing on a linear problem with a single degree
of freedom.
5.1
Long-term behavior
Consider a simple harmonic oscillator (i.e. a spring pendulum) with mass m and stiffness k. For an initial elongation
of u0 = u(0), the displacement and velocity of the oscillator can be computed analytically: uan (t) = u0 cos(ω t) and
p
van (t) = −ω u0 sin(ω t). The frequency of oscillation is given by ω = k/m; the period length (i.e. the duration of one
oscillation) is determined through T0 = 2π/ω. In the following, the numerical results are normalized by u0 , ω, T0 , and by
the initial energy of the system, E0 =
1
2
k u20 .
Copyright line will be provided by the publisher
ZAMM header will be provided by the publisher
13
We now compare our Hermite schemes with the implicit Newmark algorithm [61], choosing the Newmark parameters
as β = 1/4 and γ = 1/2. Regarding linear systems, this method is then not only unconditionally stable; it can further
conserve the energy (see e.g. Ref. [35] or [43]). Besides, it is discussed in Ref. [41] that for γ = 1/2, the Newmark is
variational. In addition to the Newmark algorithm, we consider a variational integrator based on linear finite elements in
time (see Appendix B). We will refer to this method as L1-integrator.
Fig. 3 shows the displacement and the velocity of the oscillator for three periods and a very coarse time discretization.
As expected, for our six schemes the displacement (velocity) is C 1 -continuous (C 0 -continuous) at the discrete time steps.
In contrast, the L1-integrator approximates the velocity as a constant along each time interval; this leads to discontinuities
at the interval boundaries. For the Newmark algorithm, the displacement and velocity are evaluated only at the discrete
1.5
1.5
1
1
0.5
0.5
v / (ω u0)
u / u0
time steps, tn . We observe that the oscillation period increases for both Newmark’s method and the L1-integrator.
0
−0.5
0
−0.5
−1
−1.5
−2
0
−1
NM, β = 1/4, γ = 1/2
L1 integrator
p+q+ / Geradin (’74)
p+q− scheme
0.5
1
−1.5
1.5
t / T0
2
2.5
3
−2
0
p2 scheme
q2 scheme
p−q+ scheme
p−q− scheme
0.5
1
(a) Displacement
1.5
t / T0
2
2.5
3
(b) Velocity
Fig. 3 Harmonic oscillator: Displacement and velocity for three periods of oscillation; the six Hermite schemes are compared with the
Newmark algorithm (NM) and with the L1-integrator; ∆t = T0 /6.
Regarding the maximum displacement, for two of our mixed methods (p− q+ and p− q− ) the amplitude of oscillation
seems to increase remarkably (Fig. 3(a)). This indicates that these methods may be unstable. For the remaining mixed
schemes (p+ q+ /Géradin and p+ q− ) the amplitudes in both the displacement and velocity decrease. In contrast, both the
amplitude and the period of oscillation are well-preserved for our p2- and q2-schemes.
Fig. 4 shows the total energy of the system over 100 oscillation periods. We observe that the schemes p− q+ and p− q−
are unstable, while the schemes p+ q+ and p+ q− are strongly dissipative. This agrees with the results shown in Fig. 3.
As expected, for both the remaining two schemes, p2 and q2, and for the L1-integrator, the total energy is qualitatively
preserved. Interestingly, the p2- and the q2-scheme are more accurate; this is indicated by smaller amplitudes of oscillation
in Fig. 4(b). Compared to the linear integrator, the maximum relative errors are smaller by one order of magnitude for the
q2-scheme, and even by two orders for the p2-scheme; see Table 1.
Copyright line will be provided by the publisher
J.C. Mergel, R.A. Sauer, and S. Ober-Blöbaum: C 1 -continuous space-time discretization based on Hamilton’s law
14
1.6
1.01
1.4
1
1.2
E / E0
E / E0
1
0.8
0.6
0.4
0.2
0.98
L1 integrator
p+q+ / Geradin (’74)
p+q− scheme
p2 scheme
q2 scheme
p−q+ scheme
p−q− scheme
0
0
20
40
0.99
0.97
t / T0
60
80
(a) Total energy, E = K + Π
100
0.96
80
L1 integrator
p2 scheme
q2 scheme
85
90
t / T0
95
100
(b) Zoom of (a)
Fig. 4 Harmonic oscillator: Long-term energy behavior for 100 periods of oscillation; the six Hermite schemes are compared with the
L1-integrator; ∆t = T0 /10.
Table 1 Harmonic oscillator: Maximum error in the total energy for the methods shown in Fig. 4(b).
emax
E
5.2
p2
q2
L1
0.023 %
0.653 %
3.29 %
Stability
We now investigate the stability of the six schemes by means of the harmonic oscillator. For this purpose, we introduce the
normalized time step γ := ω ∆t and insert it into Eq. (39) – (44). Following Ref. [49,63], the six schemes can be expressed
in the form
vn+1
vn
= A
,
ω un+1
ω un
(48)
where A is the amplification matrix given in Appendix C. The terms un and vn denote the displacement and the velocity
at time step tn . Fig. 5 shows the spectral radius, ρ(A), for each of the schemes. Table 2 shows the maximum permitted
time step, ∆tstab , for which the schemes are stable, i.e. for which ρ(A) ≤ 1. Both the table and Fig. 5(b) show that the
last two schemes, p− q− and p− q+ , are unstable even for very small time steps. In contrast, the schemes p+ q− and p+ q+
seem to be stable for large time steps; nevertheless, these methods are numerically dissipative because of ρ(A) < 1 for
γ < γstab (see Fig. 5(b)). The most promising methods seem to be the p2-scheme and the q2-scheme; they show both
excellent stability and energy preservation.
5.3
Symplecticity
As discussed in Section 5.1, for the linear oscillator both the p2-scheme and the q2-scheme preserve the energy of the system well. This motivates us to investigate whether these methods are generally symplectic. One way to prove symplecticity
is to investigate the derivatives of the phase state, (p̂n+1 , x̂n+1 ), w.r.t. the previous state, (p̂n , x̂n ). Here, p̂• is the linear
Copyright line will be provided by the publisher
ZAMM header will be provided by the publisher
15
1+1e−9
p+q+ / Geradin (’74)
p+q− scheme
p2 scheme
q2 scheme
p−q+ scheme
p−q− scheme
1
p+q+ / Geradin (’74)
p+q− scheme
p2 scheme
q2 scheme
p−q+ scheme
p−q− scheme
ρ(A)
ρ(A)
10
1
0
10
−2
−1
10
0
10
1
10
10
1−1e−9 −2
10
−1
10
γ
γ
(a) Spectral radius of A
(b) Zoom of (a); the purple line is hidden behind the blue one
Fig. 5 Harmonic oscillator: Spectral radius for the six Hermite schemes as a function of the normalized time step, γ.
Table 2 Harmonic oscillator: Maximum time steps, γstab and ∆tstab , for which the schemes are stable.
p2
q2
p+ q−
p+ q+
p− q−
p− q+
γstab [−]
3.144
3.055
3.083
9.165
–
–
∆tstab [T0 ]
0.500
0.486
0.491
1.459
–
–
momentum, given by p̂• = m v̂• . This results in the Jacobian
∂ p̂n+1
∂ p̂n
:=
∂ x̂n+1
∂ p̂n
Bn+1
∂ p̂n+1
∂ v̂n+1 −1
m
m
∂ x̂n
∂ v̂n
=
∂ x̂n+1 ∂ x̂n+1 −1
m
∂ x̂n
∂ v̂n
∂ v̂n+1
∂ x̂n
.
∂ x̂n+1
∂ x̂n
m
(49)
According to Ref. [26], the mapping (p̂n , x̂n ) 7→ (p̂n+1 , x̂n+1 ) is symplectic if Bn+1 is symplectic, i.e. if
0
I
d·n
no
,
BT
J=
n = 0, . . . , N −1,
n+1 J Bn+1 = J,
−I d·nno
0
where I d·nno is the identity matrix of dimension d · nno . For the harmonic oscillator, the Jacobian reduces to
∂vn+1
∂vn+1
m
∂vn
∂un
.
Bn+1 =
1 ∂u
∂un+1
n+1
m ∂vn
∂un
(50)
(51)
In this case one can show that the determinant of Bn+1 is equal to the determinant of the amplification matrix introduced
in Eq. (48): det (Bn+1 ) = det (A). With this relation the condition for symplecticity (50) is fulfilled if the determinant
of A is equal to one,
0
− det (A)
det (A)
0
=
0
−1
1
0
⇒
det (A) = 1.
(52)
Copyright line will be provided by the publisher
J.C. Mergel, R.A. Sauer, and S. Ober-Blöbaum: C 1 -continuous space-time discretization based on Hamilton’s law
16
By means of Appendix C one can show that both our p2-scheme and q2-scheme fulfill det (A) = 1 ∀ γ. This implies
that — at least for the harmonic oscillator — these schemes are symplectic. It remains to be subject of further investigation
whether this is also true for arbitrary systems with multiple degrees of freedom.
5.4
Convergence behavior
We now focus on the four stable schemes: p2, q2, p+ q− , and p+ q+ (which is equivalent to the method of Géradin [23]).
In order to study convergence for the harmonic oscillator we consider the maximum errors of the displacement, velocity,
and total energy at the discrete time steps; these are given by
emax
=
•
max |e• (tn )|,
• ∈ {u, v, E},
n=0,...,N
(53)
where
eu (t) = |u(t) − uan (t)| / |u0 |,
ev (t) = |v(t) − van (t)| / |ω u0 |,
eE (t) = |E(t) − E0 | / E0 .
(54)
Fig. 6 shows the convergence behavior of these errors for our stable schemes, the Newmark algorithm, and the L1-integrator.
In addition, we account for the results discussed in Ref. [50] for cubic Hermite interpolation. Note that compared to the
last three methods, our time integration schemes must account for twice the number of unknowns in each step: the nodal
displacements and the nodal velocities. This has been considered in the scaling of the abscissae by introducing the factor
cDOF (cDOF = 1/2 for our schemes, otherwise cDOF = 1).
As already shown in Fig. 3(b), for the L1-integrator the piecewise approximation of the velocity is discontinuous at
the discrete time steps. Nevertheless, we can determine the discrete velocity at tn by computing the momentum from a
discrete Legendre transformation [26]. This approach is also discussed by Ober-Blöbaum et al. [62] considering boundary
conditions for the velocity. For the L1-integrator the maximum errors in both the velocity and in the energy (Fig. 6(b) and
6(c)) are finally obtained by
v̂n = m−1 p̂n ,
p̂n := −
ht
∂Sn+1
(x̂n , x̂n+1 )
∂Snht (x̂n−1 , x̂n )
=
.
∂ x̂n
∂ x̂n
(55)
As expected, Newmark’s method can conserve the energy of the (linear) system; the corresponding error thus lies in the
range of machine precision. The errors in the displacement and in the velocity, however, are of order p = 2. The same
rate of convergence can be observed for the L1-integrator. This agrees with the discussion in Ref. [63] that integrators
interpolating the displacement linearly in time can be at most of second order.
Our schemes p+ q+ [23], p+ q− , and q2 also show an order of p = 2. Since their computational effort is higher than
for both Newmark’s method and the L1-integrator, these methods are not favorable. In comparison, Leok and Shingel
prove convergence with order three for their integrator based on cubic Hermite interpolation (i.e. n = 2 in Theorem 2 of
Ref. [50]). An even better rate of convergence is achieved with our favorite candidate, the p2-scheme: All of the errors in
the displacement, the velocity, and the energy converge with order p = 4.
Copyright line will be provided by the publisher
ZAMM header will be provided by the publisher
17
0
0
10
10
p=2
p=2
−5
−5
emax
v
10
max
10
eu
p=4
−10
−15
10
−10
NM, β = 1/4, γ = 1/2
L1 integrator
p+q+ / Geradin (’74)
p+q− scheme
p2 scheme
q2 scheme
10
0
−1
10
10
p=4
10
εN
εN
steps
steps
−15
−2
10
(∆t / ∆t0) ⋅ cDOF
−3
−4
10
10
10
(a) Maximum error in the displacement
0
−1
10
10
−2
10
(∆t / ∆t0) ⋅ cDOF
−3
10
−4
10
(b) Maximum error in the velocity
0
10
p=2
−5
eE
max
10
p=4
−10
10
ε Nsteps
−15
10
0
−1
10
10
−2
10
(∆t / ∆t0) ⋅ cDOF
−3
10
−4
10
(c) Maximum error in the energy
Fig. 6 Harmonic oscillator: Convergence behavior for T = 2 T0 , ∆t0 = T0 /8; cDOF = 1/2 for our Hermite schemes, otherwise
cDOF = 1; the dashed line labeled with ε Nsteps denotes the estimated machine precision multiplied with the number of computed steps.
6
Numerical results for 1D elastodynamics
As shown in the previous sections, our p2-scheme is symplectic for the harmonic oscillator; it further possesses the highest
rate of convergence. So far we have studied only linear problems with a single degree of freedom. In the following section
we apply the p2-scheme to spatially continuous problems.
6.1
Axial vibration of a linear elastic bar
We now discuss the free axial vibration of a linear elastic bar (Appendix D). For the spatial discretization we either use
linear Lagrange or cubic Hermite finite elements; see Appendix A.1 and text books such as Ref. [78]. The second type of
element yields a C 1 -continuous approximation of the displacement also in space.
Copyright line will be provided by the publisher
J.C. Mergel, R.A. Sauer, and S. Ober-Blöbaum: C 1 -continuous space-time discretization based on Hamilton’s law
18
Consider the bar vibrating in the first (i.e. the lowest) characteristic eigenmode. For this test case, the corresponding
displacement, velocity, and energy can be analytically computed from the one-dimensional wave equation; this yields
uan (X, t) = u0 · cos
πX
L
· cos (ωan t),
van (X, t) = −u0 · cos
πX
L
· ωan sin (ωan t) ,
and E0 = EA/(4 L) · (π u0 )2 . Here, X ∈ [0, L], u0 is the amplitude of oscillation, and ωan = π
p
(56)
E/(ρ0 L2 ) is the first
natural frequency. The deformation of the bar vibrating in the first mode is shown in Fig. 7. Here, six linear elements are
used for the spatial discretization. As expected, the bar performs sinusoidal oscillations. Due to the coarse finite element
mesh, however, the structure oscillates with a frequency slightly higher than the analytical solution: |ω − ωan | / ωan ≈
1.15 %. The oscillation period of the discrete system, T0 , is thus smaller than the analytical solution, T0an = 2π/ωan .
2
t / T0
1.5
1
0.5
0
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
x/L
Fig. 7 Linear elastic bar: Two oscillations in the first natural frequency using the p2-scheme and linear FE.
Like variational integrators, our p2-scheme does not incorporate any numerical dissipation to damp spurious oscillations;
we thus must carefully adjust the size of the time step, ∆t, to the spatial FE element mesh. Since a stability analysis
including spatial discretization can be quite tedious, we roughly estimate the maximum permitted time step for either linear
or Hermite elements. We therefore consider the Courant-Friedrichs-Lewy (CFL) condition for one-dimensional problems,
CCFL ≤ C max ,
where cW =
CCFL :=
cW · ∆t
,
∆L
(57)
p
E/ρ0 is the velocity of wave propagation, and ∆L is the characteristic discretization length. For a linearly
interpolated element applies ∆L := Le ; for a Hermite finite element, we choose ∆L := Le /2 to take into account that it
has twice the number of unknowns (and thus higher accuracy).
We now vary the CFL number for a bar vibrating for at least 1000 oscillations. The estimated maximum values are useful
to choose appropriate parameters for the following numerical examples. Of course we cannot ensure, however, stability
for arbitrary CFL numbers smaller than these estimates. As shown in the stability analysis for a single degree of freedom
(Section 5.2), the methods may also become unstable for small ranges of parameters. This becomes apparent in Fig. 5(a),
where the spectral radius of the p2-scheme exceeds the limit (one) for a small range of time steps, while being stable again
for larger steps. Apart from that, the CFL condition does not serve as sufficient condition for stability. For linear finite
max
max
elements we estimate CL1
≈ 1.00 and Cp2
≈ 0.90; this implies that the time step should fulfill ∆t ≤ 1.00 ∆L/cW
Copyright line will be provided by the publisher
ZAMM header will be provided by the publisher
19
max
for the L1-integrator, and ∆t ≤ 0.90 ∆L/cW for the p2-scheme. For Hermite finite elements we obtain CL1
≈ 0.72 and
max
Cp2
≈ 0.96. In the case of linear problems Newmark’s method is unconditionally stable if β = 1/4 and γ = 1/2.
6.2
Convergence for the linear bar
Reconsider the axial vibration discussed in the previous section. Following Ref. [18], we introduce for the displacement
and velocity discrete L2 -norms including the relative errors at all time steps and finite element nodes:
v
u N nno
u X X |e• (XI , tn )|2
,
• ∈ {u, v}.
||e• ||Σ := t
(N +1) · nno
n=0
(58)
I=1
This corresponds to the Frobenius norm of the arrays eu (XI , tn ) and ev (XI , tn ) normalized by (the square roots of) the
numbers of nodes and time steps. The relative errors are defined as
eu (X, t) = |u(X, t) − uan (X, t)| / |u0 |,
ev (X, t) = |v(X, t) − van (X, t)| / |ωan u0 |.
In analogy to Eq. (58), we define a discrete L2 -norm for the error in the energy,
v
uN
u X |eE (tn )|2
,
||eE ||Σ := t
N +1
n=0
(59)
(60)
using eE (t) from Eq. (54). Fig. 8 shows the convergence behavior of the displacement, the velocity, and the energy. For
the spatial discretization, we either use linear and Hermite elements. We further consider two fixed CFL numbers, while
refining both the mesh and the time step simultaneously.
For a linear finite element mesh (left column of Fig. 8), the three time discretization methods converge with the same
order. This indicates that for this specific problem, the error caused by the spatial discretization predominates. In contrast,
the error due to the spatial Hermite discretization (right column of Fig. 8) carries considerably less weight. The p2-scheme
yields a significantly higher convergence than for both the Newmark algorithm and the L1-integrator.
Note that if the bar is discretized with linear elements, the resulting system can be treated as a naturally discrete springmass system consisting of linear springs. For such a spring-mass system, the (temporally) analytical solution is
π XI
π XI
h
h
· cos (ω t),
van (XI , t = −u0 · cos
· ω sin (ω t),
uan (XI , t = u0 · cos
L
L
(61)
I = 1, . . . , nno ; cf. Eq. (56). The natural frequency, ω, can be determined by analyzing the eigenmodes of the discrete system. Fig. 9(a) and 9(b) show the maximum errors in the displacement and velocity arising from the temporal discretization.
Here, the errors ehu and ehv are determined from Eq. (59), inserting the analytical solutions given by Eq. (61). As expected,
the orders of convergence agree with those studied for a single degree of freedom (Section 5.4).
Table 3 compares the computational cost of the three methods for those results shown in the right column of Fig. 8,
CCFL = 0.5. Here, we measure the computation time that is required to obtain an error in the displacement smaller either
than 1 %, 0.1 %, or 0.01 %; see Fig. 8(b). Although the p2-scheme must account for twice the number of unknowns within
each time step, it takes — due to its higher order of convergence — less computation time than both the L1-integrator and
Newmark’s method. Note that for this linear example, the integrals in the discrete momenta can be evaluated analytically,
i.e. without numerical quadrature. For an example requiring quadrature also in time, we refer to the following section.
Copyright line will be provided by the publisher
J.C. Mergel, R.A. Sauer, and S. Ober-Blöbaum: C 1 -continuous space-time discretization based on Hamilton’s law
20
NM, β = 1/4, γ = 1/2
L1 integrator
p2 scheme
−1
10
NM, β = 1/4, γ = 1/2
L1 integrator
p2 scheme
−2
10
1
−2
10
|| eu ||Σ
|| eu ||Σ
−4
10
(linear FE)
(Hermite FE)
−6
10
−3
10
2
2
4.5
−8
10
1
−4
10
−1
1
−2
10
−3
10
∆s
−1
10
−2
10
−3
10
10
∆s
(a) Error in the displacement for spatially linear FE
(b) Error in the displacement for spatial Hermite FE
−2
10
−1
10
1
10
|| ev ||Σ
|| ev ||Σ
−4
10
(linear FE)
−2
(Hermite FE)
−6
10
−3
10
2
3.5
2
1
−8
10
1
−4
10
−1
−2
10
−3
10
∆s
−1
10
−2
10
−3
10
10
∆s
(c) Error in the velocity for spatially linear FE
(d) Error in the velocity for spatial Hermite FE
−2
10
−1
10
1
10
|| eE ||Σ
|| eE ||Σ
−4
10
(linear FE)
−2
−3
10
2
(Hermite FE)
−6
10
2
4
−8
10
1
−4
10
−1
10
−2
10
∆s
(e) Error in the energy for spatially linear FE
−3
10
1
−1
10
−2
−3
10
10
∆s
(f) Error in the energy for spatial Hermite FE
Fig. 8 Linear elastic bar: Convergence behavior of the displacement, the velocity, and the energy refining both the mesh and time step;
the parameter ∆s is given by ∆s = ∆t/T0an = CCFL · ∆L/(2 L); dashed line: CCFL = 0.5, solid line: CCFL = 0.125; T = 1 T0an .
Copyright line will be provided by the publisher
ZAMM header will be provided by the publisher
21
0
0
10
10
p=2
p=2
−5
−5
|| ehv ||Σ
10
|| ehu ||Σ
10
p=4
−10
10
10
(linear FE)
NM, β = 1/4, γ = 1/2
L1 integrator
p2 scheme
−15
10
p=4
−10
0
−1
10
10
(linear FE)
−15
−2
−3
10
∆t / T0an
10
−4
10
10
0
10
(a) Temporal error in the displacement
−1
−2
10
10
∆t / T0an
−3
10
−4
10
(b) Temporal error in the velocity
Fig. 9 Linear elastic bar: Convergence of the temporal errors in the displacement and the velocity for linear Lagrange FE; T = 1 T0an ,
∆t0 = T0an /4.
Table 3 Linear elastic bar: Step size, ∆s, and computation time, Tct , of the test cases from Fig. 8(b) (CCFL = 0.5), for
which the error in the displacement is smaller than 1 %, 0.1 %, and 0.01 %; Tct denotes the time for one oscillation.
Newmark
∆s
Tct [ms]
L1-integrator
∆s
Tct [ms]
p2-scheme
∆s
Tct [ms]
||eu ||Σ < 1 %
1/32
79.6
1/32
105.0
1/16
44.9
||eu ||Σ < 0.1 %
1/128
815.8
1/64
416.3
1/32
135.7
||eu ||Σ < 0.01 %
1/256
2,919.4
5,381.3
1/64
574,3
6.3
1/256
Vibration of a nonlinear bar
The numerical examples discussed in the previous sections cover both naturally discrete and (spatially discretized) continuum systems. So far only linear problems (for which the internal forces depend on the displacement linearly) have been
investigated. We therefore consider a nonlinear Neo-Hooke material behavior, which is described in Appendix D. The bar
is initially deformed by prescribing the same displacement and velocity as in Eq. (56). Fig. 10 shows the deformation of
the bar (a) at the very beginning and (b) after a long period of oscillations. Here, a Hermite finite element mesh with eight
elements is chosen. For a better comparison with the results from the previous section, the time step is normalized by
the period length of the first eigenmode, T0an . Since the mechanical response of the system differs from the linear case,
however, the initially sinusoidal oscillations turn into a set of different interfering oscillations (Fig. 10(b)).
Fig. 11 shows the long-term behavior for a nonlinear bar, comparing the system’s total energy for the p2-scheme, the
Newmark algorithm, and the L1-integrator. Like for the harmonic oscillator from Section 5.1, the energy oscillates while
being qualitatively preserved. In comparison with the other methods the relative error of the p2-scheme is smaller by five
orders of magnitude.
Copyright line will be provided by the publisher
J.C. Mergel, R.A. Sauer, and S. Ober-Blöbaum: C 1 -continuous space-time discretization based on Hamilton’s law
3
200
2.5
199.5
2
199
0
t / T an
t / T0an
22
1.5
198.5
1
198
0.5
197.5
0
197
0
0.25
0.5
x/L
0.75
1
0
(a) Deformation for t / T0an ∈ [0, 3]
0.25
0.5
x/L
0.75
1
(b) Deformation for t / T0an ∈ [197, 200]
Fig. 10 Nonlinear bar: Deformation of the bar (a) at the beginning of the oscillation, and (b) after a long time period, using the
p2-scheme and Hermite FE; T = 200 T0an , ∆t = T0an /512, Le = L/8.
−5
x 10
(E − E0) / E0
1
0.5
0
−0.5
−1
0
NM, β = 1/4, γ = 1/2
L1 integrator
p2 scheme
20
40
60
80
100
t / T0an
120
140
160
180
200
Fig. 11 Nonlinear bar: Long-term energy behavior for approximately 200 periods of oscillation using Hermite FE; the maximum relative
error of the p2-scheme is 3.7 · 10−10 ; ∆t = T0an /256, Le = L/4.
In addition, we investigate the convergence behavior for a nonlinear bar, considering Hermite finite elements in space.
Since for this case the deformation cannot be computed analytically, we compare our results with a fine reference solution,
using both a very fine FE mesh and a small time step. Like for the examples shown in the previous section, we refine
both discretizations simultaneously. The results are shown in Fig. 12(a) – 12(c). The accuracy of our scheme becomes
most apparent for the displacement and the total energy; for these quantities we observe a significantly higher convergence.
Note that caused by the nonlinear material law, both the p2-scheme and the L1-integrator require numerical quadrature to
evaluate the time integral over the internal forces; see also Appendix A.3. If we demand, however, a sufficient high accuracy
w.r.t. the fine solution, the computational cost may still be lower for the p2-scheme than for the Newmark algorithm. An
error in the displacement smaller than 10−5 (Fig. 12(a)), for instance, requires a step size of ∆s = 1/2048 for Newmark’s
Copyright line will be provided by the publisher
ZAMM header will be provided by the publisher
23
method. In contrast, the p2-scheme achieves this accuracy already for ∆s = 1/512. Therefore, the measured computation
time is significantly lower: 48.7% of the time required for the Newmark algorithm.
We finally investigate how well the (spatially discrete) initial energy is preserved over time; see Fig. 12(d). The error
plotted here thus arises only from the temporal discretization. A comparison with Fig. 6(c) shows that — for both the linear
oscillator and the nonlinear bar — we achieve with the p2-scheme the same high order of convergence: p = 4.
−2
−2
10
10
NM, β = 1/4, γ = 1/2
L1 integrator
p2 scheme
−4
−4
|| eref
||Σ
v
10
|| eref
||Σ
u
10
−6
−6
10
10
(Hermite FE)
(Hermite FE)
−8
10
−8
−2
−3
10
10
∆s
10
−4
10
−2
−3
10
(a) Error in the displacement (w.r.t. reference solution)
10
∆s
−4
10
(b) Error in the velocity (w.r.t. reference solution)
−4
−5
10
10
−6
10
|| ehE ||Σ
|| eref
||Σ
E
p=2
−8
10
−10
10
p=4
−10
10
(Hermite FE)
(Hermite FE)
−12
10
−15
−2
−3
10
10
∆s
(c) Error in the total energy (w.r.t. reference solution)
−4
10
10
−2
10
−3
10
∆s
−4
10
(d) Error in the total energy (w.r.t. E0 )
Fig. 12 Nonlinear bar: Relative errors with respect to (a) – (c) a very fine reference solution and (d) to the initial energy of the spatially
discrete system, refining both the mesh and time step; ∆s = ∆t/T0an = ∆L/(16 L); shown is CCFL = 0.125 for Hermite FE and
ref
T = 2 T0an ; the reference solution is computed with the p2-scheme using Lref
= T0an /8192.
e = L/256 and ∆t
7
Conclusion
In this work we derive a class of C 1 -continuous time integration methods for conservative elastodynamic problems. Using
piecewise Hermite interpolation, we approximate the displacement of a deformable solid by C 1 -continuous functions in
time. The velocity of the body is thus C 0 -continuous in the entire time domain. To explicitly account for any initial velocities in the system, we consider a generalization of Hamilton’s principle, referred to as Hamilton’s law of varying action.
Copyright line will be provided by the publisher
J.C. Mergel, R.A. Sauer, and S. Ober-Blöbaum: C 1 -continuous space-time discretization based on Hamilton’s law
24
From the action integral of the continuous system we derive the spatially and temporally weak form of the equilibrium
equation in elastodynamics. This expression is first discretized in space, using a standard (Galerkin) finite element method.
Afterwards, the spatially discrete system is discretized in time by 1) subdividing the temporal domain into a set of smaller
time intervals, and 2) approximating the displacement by cubic Hermite shape functions.
Generally, methods belonging to the class of variational integrators can be constructed by varying the action integral
over the entire time domain. From this equation one can derive the discrete Euler-Lagrange equations in order to develop
a subsequent time integration method. For cubic Hermite interpolation, however, the resulting variational integrator is
unconditionally unstable; this issue has been discussed also by Riff and Baruch [68]. Therefore, we first vary the action of
each time interval individually, and then derive different one-step methods to solve for the new (unknown) displacement and
velocity. This yields six time integration schemes with different properties. Technically, these methods are not variational
in the sense that they are not derived from the virtual action of the total domain. Interestingly, the most favorable of these
methods — denoted as p2-scheme — offers similar advantageous properties like variational integrators: 1) a qualitatively
accurate long-term behavior, and 2) symplecticity for simple linear systems.
We first investigate the properties of our schemes by considering a harmonic oscillator. We then demonstrate numerically
that the p2-scheme shows both conditional stability and convergence of order four. Afterwards, we examine both linear
and nonlinear problems accounting either for inherently continuous or spatially discrete systems. Our results show that for
an appropriate spatial discretization, the p2-scheme provides both reasonable computational effort and remarkably higher
accuracy than variational integrators based on linear interpolation in time. This scheme also well-preserves the energy of the
system for long-time integration. It remains to be seen whether our method is symplectic for arbitrary nonlinear systems.
This and other properties (such as the convergence behavior and the preservation of momentum maps in the presence of
symmetries) should be addressed in an analytical study in the future.
Since the scope of this work is the construction of time integration methods, we here focus on one-dimensional elastodynamic problems. An elaborate study of elastodynamic problems in either two or three dimensions may be the scope
of future work. In addition, since the numerical results obtained within this study for spatial Hermite finite elements look
very promising, it would be interesting to apply our time integrator in combination with other spatially C 1 -continuous FE
discretizations and — in particular — to isogeometric approaches [16, 34]. In contrast to higher-order Lagrangian finite
elements, isogeometric elements can prevent high-frequency errors [17, 37] that require unnecessarily small time steps.
Acknowledgements We are grateful to the German Research Foundation (DFG) supporting J. C. Mergel and R. A. Sauer through grants
SA1822/5-1 and GSC111. We further thank Prof. Yuri B. Suris (Technical University of Berlin), Prof. Sachin S. Gautam (IIT Guwahati),
and Marcus G. Schmidt (formerly RWTH Aachen University) for their helpful comments.
A
Implementation of our schemes
This section contains helpful details for the implementation of our time integration schemes.
Copyright line will be provided by the publisher
ZAMM header will be provided by the publisher
A.1
25
Shape functions
In both the spatial and temporal discretizations, we consider either linear Lagrange or cubic Hermite shape functions.
Usually, these functions are defined on master domains, denoted e.g. by τ ∈ [−1, 1] for the discretization in time. The
mapping from the master to the temporal domain, τ 7→ t (t ∈ [tn , tn+1 ]), is characterized by the Jacobian determinant
Jτ := ∂t/∂τ ; this determinant is given by Jτ = (tn+1 − tn )/2 for both the L1-integrator and our six schemes.
• Linear Lagrange shape functions (for the L1-integrator):
R1 (τ ) =
1
(1 − τ ),
2
R2 (τ ) =
1
(1 + τ ).
2
(62)
• Cubic Hermite shape functions (for our schemes):
R1 (τ )
H1 (τ )
A.2
=
=
1
(2 + τ )(1 − τ )2 ,
4
Jτ
(τ + 1)(1 − τ )2 ,
4
R2 (τ )
H2 (τ )
=
=
1
(2 − τ )(1 + τ )2 = 1 − R1 (τ ),
4
Jτ
(τ − 1)(1 + τ )2 .
4
(63)
Discrete Lagrangian and action integral
The discrete (pseudo-)momenta (27) and (28) are obtained by taking the derivatives of the spatially and temporally discrete
action
S ht =
N
−1
X
ht
Sn+1
,
ht
Sn+1
=
n=0
Z
tn+1
Lht xt , ẋt dt.
(64)
tn
Here, the discrete Lagrangian, Lht xt , ẋt = K ht ẋt − Πht xt , can be determined from
K ht ẋ
t
nel
X
1 t T
t
=
ẋ
me ẋe ,
2 e
e=1
Πht x
t
=
"Z
nel
X
e=1
Ωe0
#
t
t T
W xe dV − xe
t
e
fext
xe
.
(65)
In Eq. (29) and (30) we additionally use the relation f xt = ∂Πht xt /∂xt .
If required, the total energy of the system can be computed from E ht xt , ẋt = K ht ẋt + Πht xt .
A.3
Implementation of the integrals
As pointed out in Section 4.3, some of the integrals in the discrete (pseudo-)momenta can be computed analytically. For
this purpose, we split the internal force into a linear and into a nonlinear part,
fint xt = k ut + fnlin xt ,
(66)
where ut = xt − X, and where k is a linear stiffness matrix. The case fnlin = 0 corresponds to linear elasticity, while
k = 0 represents the fully nonlinear case. Further let
e
f xt := fnlin xt − fext .
(67)
Copyright line will be provided by the publisher
J.C. Mergel, R.A. Sauer, and S. Ober-Blöbaum: C 1 -continuous space-time discretization based on Hamilton’s law
26
Inserting Eq. (21) and (22) into Eq. (29) and (30) yields with ∆t := tn+1 − tn
h
i Z tn+1
1
m 12 ûn+1 − 12 ûn − ∆t v̂n+1 − ∆t v̂n +
R1 (t) e
f xt dt
10 ∆t
tn
i
∆t h
k 54 ûn+1 + 156 ûn − 13 ∆t v̂n+1 + 22 ∆t v̂n ,
+
420
h
i Z tn+1
1
m 12 ûn+1 − 12 ûn − ∆t v̂n+1 − ∆t v̂n −
R2 (t) e
f xt dt
=
10 ∆t
tn
i
∆t h
+
k − 156 ûn+1 − 54 ûn + 22 ∆t v̂n+1 − 13 ∆t v̂n
420
p̂−
n =
p̂+
n+1
(68)
(69)
and
h
i Z tn+1
1
H1 (t) e
f xt dt
m 3 ûn+1 − 3 ûn + ∆t v̂n+1 − 4 ∆t v̂n +
30
tn
h
i
2
∆t
k 13 ûn+1 + 22 ûn − 3 ∆t v̂n+1 + 4 ∆t v̂n ,
+
420
h
i Z tn+1
1
m − 3 ûn+1 + 3 ûn + 4 ∆t v̂n+1 − ∆t v̂n −
H2 (t) e
f xt dt
=
30
tn
i
∆t2 h
+
k 22 ûn+1 + 13 ûn − 4 ∆t v̂n+1 + 3 ∆t v̂n .
420
q̂−
n =
q̂+
n+1
(70)
(71)
The derivatives with respect to x̂n+1 and v̂n+1 (required for linearization, see Section 4.3) are given by
∂ p̂+
n+1
∂ x̂n+1
∂ q̂−
n
∂ x̂n+1
∂ q̂+
n+1
∂ x̂n+1
tn+1
∂e
f xt
R1 (t) R2 (t)
dt,
∂xt
tn
Z tn+1
∂e
f xt
6
13 ∆t
2
R2 (t)
=
m−
k−
dt,
5 ∆t
35
∂xt
tn
Z tn+1
∂e
f xt
13 ∆t2
1
m+
k+
H1 (t) R2 (t)
=
dt,
10
420
∂xt
tn
Z tn+1
∂e
f xt
1
11 ∆t2
=− m+
k−
H2 (t) R2 (t)
dt
10
210
∂xt
tn
∂ p̂−
6
9 ∆t
n
=
m+
k+
∂ x̂n+1
5 ∆t
70
Z
(72)
(73)
(74)
(75)
and
∂ p̂+
n+1
∂ v̂n+1
∂ q̂−
n
∂ v̂n+1
∂ q̂+
n+1
∂ v̂n+1
A.4
tn+1
∂e
f xt
dt,
∂xt
tn
Z tn+1
∂e
f xt
1
11 ∆t2
=− m+
k−
R2 (t) H2 (t)
dt,
10
210
∂xt
tn
Z tn+1
∂e
f xt
∆t
∆t3
=
m−
k+
H1 (t) H2 (t)
dt,
30
140
∂xt
tn
Z tn+1
∂e
f xt
2 ∆t
∆t3
2
=
m−
k−
H2 (t)
dt.
15
105
∂xt
tn
1
13 ∆t2
∂ p̂−
n
=− m−
k+
∂ v̂n+1
10
420
Z
R1 (t) H2 (t)
(76)
(77)
(78)
(79)
Alternative representation of the p2-scheme
+
By integrating p̂−
n and p̂n+1 from Eq. (29) by parts, we obtain
p̂−
n
Z
tn+1
h
t
t
R1 (t) ẍ + f x
= m v̂n +
tn
i
dt,
p̂+
n+1
Z
tn+1
= m v̂n+1 −
h
i
R2 (t) ẍt + f xt dt;
(80)
tn
Copyright line will be provided by the publisher
ZAMM header will be provided by the publisher
27
the p2-scheme (39) thus fulfills
Z
tn+1
h
i
R1 (t) ẍt + f xt dt = 0,
tn
Z
tn+1
h
i
R2 (t) ẍt + f xt dt = 0.
(81)
tn
Since further R1 (t), R2 (t) ≥ 0 and R1 (t) + R2 (t) = 1 (see Eq. (63)), we arrive at Eq. (45).
B
Linear variational integrator (L1)
This section outlines the linear variational integrator (L1) that we use for comparison in Section 5 and 6. Here, the deformation and velocity are approximated by x(t) ≈ xt (t) and ẋ(t) ≈ ẋt (t),
xt (t) = R1 (t) x̂n + R2 (t) x̂n+1 ,
ẋt (t) = Ṙ1 (t) x̂n + Ṙ2 (t) x̂n+1 ,
t ∈ [tn , tn+1 ],
(82)
where R1 and R2 are defined by Eq. (62). In analogy, we introduce
δxt (t) = R1 (t) δx̂n + R2 (t) δx̂n+1 ,
δ ẋt (t) = Ṙ1 (t) δx̂n + Ṙ2 (t) δx̂n+1 .
(83)
With this interpolation the discretized action becomes S h ≈ S ht ,
S ht =
N
−1
X
ht
Sn+1
(x̂n , x̂n+1 ) .
(84)
n=0
From the discretization of the virtual action then follows that
N
−1
X
n=0
ht
ht
∂Sn+1
∂Sn+1
δx̂n +
δx̂n+1 = δxN · m ẋt (T ) − δx0 · m ẋt (0).
∂ x̂n
∂ x̂n+1
(85)
Since the nth time step belongs to two time intervals, [tn−1 , tn ] and [tn , tn+1 ], we obtain
∂S1ht (x̂0 , x̂1 )
= 0,
∂ x̂0
ht
(x̂n , x̂n+1 )
∂Snht (x̂n−1 , x̂n ) ∂Sn+1
+
= 0,
∂ x̂n
∂ x̂n
m ẋt (0) +
(86)
n = 1, . . . , N −1.
(87)
ht
The integrals within the partial derivatives of Sn+1
are also evaluated by Gaussian quadrature.
Copyright line will be provided by the publisher
J.C. Mergel, R.A. Sauer, and S. Ober-Blöbaum: C 1 -continuous space-time discretization based on Hamilton’s law
28
C
Amplification matrices for the simple harmonic oscillator
The amplification matrices of our six schemes (39) – (44) (required for Eq. (48)) are given by
26γ 4 −876γ 2 +2016
204γ 3 −2016γ
1
,
Ap2 = 4
8γ +132γ 2 +2016 3γ 5 −204γ 3 +2016γ 26γ 4 −876γ 2 +2016
Aq2
Ap+ q−
4
2
3
7γ
−192γ
+420
45γ
−420γ
1
= 4
,
2γ +18γ 2 +420
γ 5 −52γ 3 +420γ
7γ 4 −192γ 2 +420
4
2
3
65γ
−1692γ
+3780
390γ
−3780γ
1
=
,
26γ 4 +198γ 2 +3780
7γ 5 −432γ 3 +3780γ
46γ 4 −1692γ 2 +3780
Ap+ q+
Ap− q−
Ap− q+
D
13γ 4 −291γ 2 +630
1
=
10γ 4 +24γ 2 +630
γ 5 −81γ 3 +630γ
10γ 4 −582γ 2 +1260
1
= 4
γ +48γ 2 +1260 2γ 5 −162γ 3 +1260γ
60γ 3 −630γ
,
5γ 4 −291γ 2 +630
3
120γ −1260γ
,
26γ 4 −582γ 2 +1260
46γ 4 −1692γ 2 +3780
1
=
10γ 4 +198γ 2 +3780
7γ 5 −432γ 3 +3780γ
390γ 3 −3780γ
.
65γ 4 −1692γ 2 +3780
Equations for the 1D bar
In Section 6 we investigate the axial deformation of a thin bar that is characterized by its length, L, density, ρ0 , cross section
area, A, and Young’s modulus, E. We study both linear and nonlinear problems by considering different material behavior
for the bar. Let X, x, and v denote the 1D equivalents of the terms written in bold font. We further assume that B̄ = T̄ = 0.
For the 1D case the virtual kinetic energy, given by Eq. (7), reduces to
Z L
δK(v) = ρ0 A
δv · v dX.
(88)
0
For a linear elastic bar (discussed in Section 6.1 and 6.2), the virtual work (8) is given by
Z L
δΠ(x) = A
δε · E ε dX,
ε := λ − 1,
(89)
0
where λ := ∂x/∂X. In Section 6.3 we consider a Neo-Hooke material with Poisson’s ratio ν = 0; see e.g. Ref. [79]:
Z L
E
δΠ(x) = A
δλ · P dX,
P :=
λ − λ−1 .
(90)
2
0
For the spatial discretization of Eq. (88) – (90) by means of 1D finite elements, we refer e.g. to Ref. [78].
References
[1] J. H. Argyris and D. W. Scharpf, Finite elements in time and space, Nucl. Eng. Des. 10(4), 456–464 (1969).
[2] C. D. Bailey, Application of Hamilton’s law of varying action, AIAA J. 13(9), 1154–1157 (1975).
[3] C. D. Bailey, A new look at Hamilton’s principle, Found. Phys. 5(3), 433–451 (1975).
Copyright line will be provided by the publisher
ZAMM header will be provided by the publisher
29
[4] C. D. Bailey, Hamilton, Ritz, and elastodynamics, J. Appl. Mech. 43(4), 684–688 (1976).
[5] C. D. Bailey, The method of Ritz applied to the equation of Hamilton, Comput. Methods Appl. Mech. Eng. 7, 235–247 (1976).
[6] M. Baruch and R. Riff, Hamilton’s principle, Hamilton’s law – 6n correct formulations, AIAA J. 20(5), 687–692 (1982).
[7] O. A. Bauchau and T. Joo, Computational schemes for non-linear elasto-dynamics, Int. J. Numer. Methods Eng. 45, 693–719
(1999).
[8] T. Belytschko, W. K. Liu, and B. Moran, Nonlinear Finite Elements for Continua and Structures (Wiley, Chichester, 2000).
[9] P. Betsch and A. Janz, An energy-momentum consistent method for transient simulations with mixed finite elements developed
in the framework of geometrically exact shells, Int. J. Numer. Methods Eng. (2016), DOI: 10.1002/nme.5217.
[10] P. Betsch and P. Steinmann, Conservation properties of a time FE method – part II: Time-stepping schemes for non-linear elastodynamics, Int. J. Numer. Methods Eng. 50, 1931–1955 (2001).
[11] M. Borri, G. L. Ghiringhelli, M. Lanz, P. Mantegazza, and T. Merlini, Dynamic response of mechanical systems by a weak
Hamiltonian formulation, Comput. Struct. 20(1–3), 495–508 (1985).
[12] N. Bou-Rabee and H. Owhadi, Stochastic variational integrators, IMA J. Numer. Anal. 29, 421–443 (2008).
[13] P. Chadwick, Continuum Mechanics: Concise Theory and Problems (Dover Publications, 1999).
[14] J. Chung and G. M. Hulbert, A time integration algorithm for structural dynamics with improved numerical dissipation: The
generalized-α method, J. Appl. Mech. 60, 371–375 (1993).
[15] J. Cortés and S. Martı́nez, Nonholonomic integrators, Nonlinearity 14(5), 1365–1392 (2001).
[16] J. A. Cottrell, T. J. Hughes, and Y. Bazilevs, Isogeometric Analysis: Toward Integration of CAD and FEA (Wiley, 2009).
[17] J. A. Cottrell, A. Reali, Y. Bazilevs, and T. J. R. Hughes, Isogeometric analysis of structural vibrations, Comput. Methods Appl.
Mech. Eng. 195, 5257–5296 (2006).
[18] F. Demoures, F. Gay-Balmaz, S. Leyendecker, S. Ober-Blöbaum, T. S. Ratiu, and Y. Weinand, Discrete variational Lie group
formulation of geometrically exact beam dynamics, Numer. Math. 130(1), 73–123 (2015).
[19] R. C. Fetecau, J. E. Marsden, M. Ortiz, and M. West, Nonsmooth Lagrangian mechanics and variational collision integrators,
SIAM J. Appl. Dyn. Syst. 2(3), 381–416 (2003).
[20] I. Fried, Finite-element analysis of time-dependent phenomena, AIAA J. 7(6), 1170–1173 (1969).
[21] T. C. Fung, Unconditionally stable higher-order accurate Hermitian time finite elements, Int. J. Numer. Methods Eng. 39, 3475–
3495 (1996).
[22] S. S. Gautam and R. A. Sauer, An energy-momentum-conserving temporal discretization scheme for adhesive contact problems,
Int. J. Numer. Methods Eng. 93(10), 1057–1081 (2013).
[23] M. Géradin, On the variational method in the direct integration of the transient structural response, J. Sound Vib. 34(4), 479–487
(1974).
[24] O. Gonzalez, Exact energy and momentum conserving algorithms for general models in nonlinear elasticity, Comput. Methods
Appl. Mech. Eng. 190, 1763–1783 (2000).
[25] M. Groß, P. Betsch, and P. Steinmann, Conservation properties of a time FE method. Part IV: Higher order energy and momentum
conserving schemes, Int. J. Numer. Methods Eng. 63, 1849–1897 (2005).
[26] E. Hairer, C. Lubich, and G. Wanner, Geometric Numerical Integration, 2nd edition (Springer, Berlin Heidelberg, 2006).
[27] W. R. Hamilton, On a general method in dynamics; by which the study of the motions of all free systems of attracting or repelling
points is reduced to the search and differentiation of one central relation, or characteristic function, Phil. Trans. R. Soc. London
124, 247–308 (1834).
[28] W. R. Hamilton, Second essay on a general method in dynamics, Phil. Trans. R. Soc. London 125, 95–144 (1835).
[29] C. Hesch and P. Betsch, A mortar method for energy-momentum conserving schemes in frictionless dynamic contact problems,
Int. J. Numer. Methods Eng. 77, 1468–1500 (2009).
Copyright line will be provided by the publisher
J.C. Mergel, R.A. Sauer, and S. Ober-Blöbaum: C 1 -continuous space-time discretization based on Hamilton’s law
30
[30] C. Hesch and P. Betsch, Transient three-dimensional domain decomposition problems: Frame-indifferent mortar constraints and
conserving integration, Int. J. Numer. Methods Eng. 82, 329–358 (2010).
[31] H. M. Hilber, T. J. R. Hughes, and R. L. Taylor, Improved numerical dissipation for time integration algorithms in structural
dynamics, Earthq. Eng. Struct. Dyn. 5, 283–292 (1977).
[32] D. H. Hodges and R. R. Bless, Weak Hamiltonian finite element method for optimal control problems, J. Guidance Control Dyn.
14(1), 148–156 (1991).
[33] G. A. Holzapfel, Nonlinear Solid Mechanics: A Continuum Approach for Engineering (Wiley, Chichester, 2000).
[34] T. J. R. Hughes, , J. A. Cottrell, and Y. Bazilevs, Isogeometric analysis: CAD, finite elements, NURBS, exact geometry and mesh
refinement, Comput. Methods Appl. Mech. Eng. 194, 4135–4195 (2005).
[35] T. J. R. Hughes, Stability, convergence and growth and decay of energy of the average acceleration method in nonlinear structural
dynamics, Comput. Struct. 6(4–5), 313–324 (1976).
[36] T. J. R. Hughes and G. M. Hulbert, Space-time finite element methods for elastodynamics: Formulations and error estimates,
Comput. Methods Appl. Mech. Eng. 66, 339–363 (1988).
[37] T. J. R. Hughes, A. Reali, and G. Sangalli, Duality and unified analysis of discrete approximations in structural dynamics and
wave propagation: Comparison of p-method finite elements with k-method NURBS, Comput. Methods Appl. Mech. Eng. 197,
4104–4124 (2008).
[38] G. M. Hulbert, Time finite element methods for structural dynamics, Int. J. Numer. Methods Eng. 33, 307–331 (1992).
[39] G. M. Hulbert and T. J. R. Hughes, Space-time finite element methods for second-order hyperbolic equations, Comput. Methods
Appl. Mech. Eng. 84, 327–348 (1990).
[40] G. Johnson, S. Leyendecker, and M. Ortiz, Discontinuous variational time integrators for complex multibody collisions, Int. J.
Numer. Methods Eng. 100, 871–913 (2014).
[41] C. Kane, J. E. Marsden, M. Ortiz, and M. West, Variational integrators and the Newmark algorithm for conservative and dissipative
mechanical systems, Int. J. Numer. Methods Eng. 49, 1295–1325 (2000).
[42] M. Kobilarov, J. E. Marsden, and G. S. Sukhatme, Geometric discretization of nonholonomic systems with symmetries, Discrete
Contin. Dyn. Syst. Ser. S 3(1), 61–84 (2010).
[43] S. Krenk, Energy conservation in Newmark based time integration algorithms, Comput. Methods Appl. Mech. Eng. 195, 6110–
6124 (2006).
[44] S. Krenk, Global format for energy-momentum based time integration in nonlinear dynamics, Int. J. Numer. Methods Eng. 100,
458–476 (2014).
[45] D. Kuhl and M. A. Crisfield, Energy-conserving and decaying algorithms in non-linear structural dynamics, Int. J. Numer. Methods Eng. 45, 569–599 (1999).
[46] D. Kuhl and E. Ramm, Generalized energy-momentum method for non-linear adaptive shell dynamics, Comput. Methods Appl.
Mech. Eng. 178, 343–366 (1999).
[47] D. Kuhl and E. Ramm, Time integration in the context of energy control and locking free finite elements, Arch. Comput. Methods
Eng. 7(3), 299–332 (2000).
[48] C. Lanczos, The Variational Principles of Mechanics, 4th edition (Dover Publications, New York, 1970).
[49] B. Leimkuhler and S. Reich, Simulating Hamiltonian Dynamics (Cambridge University Press, Cambridge, 2005).
[50] M. Leok and T. Shingel, Prolongation-collocation variational integrators, IMA J. Numer. Anal. 32(3), 1194–1216 (2012).
[51] A. Lew, J. E. Marsden, M. Ortiz, and M. West, Asynchronous variational integrators, Arch. Rational Mech. Anal. 167(2), 85–146
(2003).
[52] A. Lew, J. E. Marsden, M. Ortiz, and M. West, An overview of variational integrators, in: Finite Element Methods: 1970’s and
Beyond, edited by L. P. Franca, T. E. Tezduyar, and A. Masud (CIMNE, Barcelona, 2004).
Copyright line will be provided by the publisher
ZAMM header will be provided by the publisher
31
[53] A. Lew, J. E. Marsden, M. Ortiz, and M. West, Variational time integrators, Int. J. Numer. Methods Eng. 60, 153–212 (2004).
[54] S. Leyendecker, P. Betsch, and P. Steinmann, Objective energy-momentum conserving integration for the constrained dynamics
of geometrically exact beams, Comput. Methods Appl. Mech. Eng. 195, 2313–2333 (2006).
[55] S. Leyendecker, J. E. Marsden, and M. Ortiz, Variational integrators for constrained dynamical systems, Z. angew. Math. Mech.
(ZAMM) 88(9), 677–708 (2008).
[56] S. Leyendecker and S. Ober-Blöbaum, A variational approach to multirate integration for constrained systems, in: Multibody
Dynamics, edited by J. C. Samin and P. Fisette, Computational Methods in Applied Sciences Vol. 28 (Springer Netherlands,
2013), pp. 97–121.
[57] S. Leyendecker, S. Ober-Blöbaum, J. E. Marsden, and M. Ortiz, Discrete mechanics and optimal control for constrained systems,
Optim. Control Appl. Methods 31(6), 505–528 (2010).
[58] J. E. Marsden, G. W. Patrick, and S. Shkoller, Multisymplectic geometry, variational integrators, and nonlinear PDEs, Commun.
Math. Phys. 199(2), 351–395 (1998).
[59] J. E. Marsden and M. West, Discrete mechanics and variational integrators, Acta Numer. 10, 357–514 (2001).
[60] S. Modak and E. D. Sotelino, The generalized method for structural dynamics applications, Adv. Eng. Softw. 33, 565–575 (2002).
[61] N. M. Newmark, A method of computation for structural dynamics, ASCE J. Eng. Mech. Div. 85(EM3), 67–94 (1959).
[62] S. Ober-Blöbaum, O. Junge, and J. E. Marsden, Discrete mechanics and optimal control: An analysis, ESAIM Control Optim.
Calc. Var. 17(2), 322–352 (2011).
[63] S. Ober-Blöbaum and N. Saake, Construction and analysis of higher order Galerkin variational integrators, Adv. Comput. Math.
41, 955–986 (2015).
[64] S. Ober-Blöbaum, M. Tao, M. Cheng, H. Owhadi, and J. E. Marsden, Variational integrators for electric circuits, J. Comput. Phys.
242, 498–530 (2013).
[65] J. T. Oden, A general theory of finite elements II. Applications, Int. J. Numer. Methods Eng. 1, 247–259 (1969).
[66] D. A. Peters and A. P. Izadpanah, hp-version finite elements for the space-time domain, Comput. Mech. 3, 73–88 (1988).
[67] S. Reich, Momentum conserving symplectic integrators, Physica D 76, 375–383 (1994).
[68] R. Riff and M. Baruch, Stability of time finite elements, AIAA J. 22(8), 1171–1173 (1984).
[69] R. Riff and M. Baruch, Time finite element discretization of Hamilton’s law of varying action, AIAA J. 22(9), 1310–1318 (1984).
[70] T. E. Simkins, Unconstrained variational statements for initial and boundary-value problems, AIAA J. 16(6), 559–563 (1978).
[71] T. E. Simkins, Finite elements for initial value problems in dynamcs, AIAA J. 19(10), 1357–1362 (1981).
[72] J. C. Simo and N. Tarnow, The discrete energy-momentum method. Conserving algorithms for nonlinear elastodynamics, Z.
angew. Math. Phys. (ZAMP) 43, 757–792 (1992).
[73] J. C. Simo, N. Tarnow, and K. K. Wong, Exact energy-momentum conserving algorithms and symplectic schemes for nonlinear
dynamics, Comput. Methods Appl. Mech. Eng. 100, 63–116 (1992).
[74] Y. B. Suris, Hamiltonian methods of Runge-Kutta type and their variational interpretation, Math. Model. 2(4), 78–87 (1990).
[75] M. Tao, H. Owhadi, and J. E. Marsden, Non-intrusive and structure preserving multiscale integration of stiff ODEs, SDEs and
Hamiltonian systems with hidden slow dynamics via flow averaging, Multiscale Model. Simul. 8(4), 1269–1324 (2010).
[76] S. Wolff and C. Bucher, Asynchronous variational integration using continuous assumed gradient elements, Comput. Methods
Appl. Mech. Eng. 255, 158–166 (2013).
[77] W. L. Wood, M. Bossak, and O. C. Zienkiewicz, An alpha modification of Newmark’s method, Int. J. Numer. Methods Eng. 15,
1562–1566 (1980).
[78] P. Wriggers, Nonlinear Finite Element Methods (Springer, Berlin Heidelberg, 2008).
[79] O. C. Zienkiewicz and R. L. Taylor, The Finite Element Method for Solid and Structural Mechanics, 6th edition (Elsevier Ltd.,
Butterworth-Heinemann, 2005).
[80] O. C. Zienkiewicz, A new look at the Newmark, Houbolt and other time stepping formulas. A weighted residual approach, Earthq.
Eng. Struct. Dyn. 5, 413–418 (1977).
Copyright line will be provided by the publisher
| 5cs.CE
|
Towards a Constructive Version of Banaszczyk’s Vector Balancing
Theorem
Daniel Dadush∗1 , Shashwat Garg†2 , Shachar Lovett‡3 , and Aleksandar Nikolov§4
arXiv:1612.04304v1 [cs.DS] 13 Dec 2016
1
Centrum Wiskunde & Informatica, Amsterdam, [email protected]
2
3
Department of Mathematics and Computer Science, Eindhoven University of Technology, [email protected]
Department of Computer Science and Engineering, University of California, San Diego, [email protected]
4
Department of Computer Science, University of Toronto, [email protected]
Abstract
An important theorem of Banaszczyk (Random Structures & Algorithms ‘98) states that for any
sequence of vectors of `2 norm at most 1/5 and any convex body K of Gaussian measure 1/2 in Rn , there
exists a signed combination of these vectors which lands inside K. A major open problem is to devise a
constructive version of Banaszczyk’s vector balancing theorem, i.e. to find an efficient algorithm which
constructs the signed combination.
We make progress towards this goal along several fronts. As our first contribution, we show an
equivalence between Banaszczyk’s theorem and the existence of O(1)-subgaussian distributions over
signed combinations. For the case of symmetric convex bodies, our equivalence implies the existence of a
universal signing algorithm (i.e. independent of the body), which simply samples from the subgaussian
sign distribution and checks to see if the associated combination lands inside the body. For asymmetric
convex bodies, we provide a novel recentering procedure, which allows us to reduce to the case where the
body is symmetric.
As our second main contribution,
√ we show that the above framework can be efficiently implemented
when the vectors have length O(1/ log n), recovering Banaszczyk’s results under this stronger assumption. More precisely, we use random walk techniques
to produce the required O(1)-subgaussian signing
√
distributions when the vectors have length O(1/ log n), and use a stochastic gradient ascent method to
implement the recentering procedure for asymmetric bodies.
1
Introduction
Given a family of sets S1 , . . . , Sm over a universe U = [n], the goal of combinatorial discrepancy minimization isPto find a bi-coloring χ : U → {−1, 1} such that the discrepancy, i.e. the maximum imbalance
maxj∈[m] | i∈Sj χ(i)|, is made as small as possible. Discrepancy theory, where discrepancy minimization
plays a major role, has a rich history of applications in computer science as well as mathematics, and we
refer the reader to [22, 11, 12] for a general exposition.
A beautiful question regards the discrepancy of sparse set systems, i.e. set systems in which each element
appears in at most t sets. A classical
√ theorem of Beck and Fiala [8] gives an upper bound of 2t − 1 in this
setting. They also conjectured an O( t) bound, which if true would be tight. An improved Beck-Fiala bound
of 2t−log∗ t was given by Bukh [10], where log∗ t is√the iterated logarithm function in base 2. Recently, it was
shown by Ezra and Lovett [15] that a bound of O( t log t) holds with high probability when m ≥ n and each
element is assigned to t sets uniformly at random. The best general bounds having sublinear dependence
in t currently depend on n or m. Srinivasan [30] used Beck’s partial coloring method [7] to give a bound
∗ Supported
by
by
‡ Supported by
§ Supported by
† Supported
the NWO Veni grant 639.071.510.
the Netherlands Organisation for Scientific Research (NWO) under project no. 022.005.025.
an NSF CAREER award 1350481 and a Sloan fellowship
an NSERC Discovery Grant
1
√
of O( t log min {n, m}). Using techniques from convex
p geometry, Banaszczyk [2] proved a general result on
vector balancing (stated below) which implies an O( t log min {n, m}) bound.
The proofs of both Srinivasan’s and Banaszczyk’s bounds were non-constructive, that is, they provided no
efficient algorithm to construct the guaranteed colorings, short of exhaustive enumeration. In the last 6 years,
tremendous progress has been made on the question of matching classical discrepancy bounds algorithmically.
Currently, essentially all discrepancy bounds proved using the partial coloring method, including Srinivasan’s,
have been made constructive [4, 21, 17, 27, 14]. Constructive versions of Banaszczyk’s result have, however,
proven elusive until very recently. In recent work [5], the first and second named authors jointly with Bansal
gave a constructive algorithm for recovering Banaszczyk’s bound in the Beck-Fiala setting as well as the more
general Komlós setting. An alternate algorithm via multiplicative weight updates was also given recently in
[19]. However, finding a constructive version of Banaszczyk’s more general vector balancing theorem, which
has further applications in approximating hereditary discrepancy, remains an open problem. This theorem
is stated as follows:
m
Theorem 1 (Banaszczyk [2]). Let v1 , . . . , vn ∈ Rm satisfy kvi k2 ≤ 1/5.
PnThen for any convex body K ⊆ R
n
of Gaussian measure at least 1/2, there exists χ ∈ {−1, 1} such that i=1 χi vi ∈ K.
The lower bound 1/2 on the Gaussian measure of K is easily seen to be tight. In particular, if all
the vectors are equal to 0, we must have that 0 ∈ K. If we allow Gaussian measure < 1/2, then K =
{x ∈ Rn : x1 ≥ ε}, for ε > 0 small enough, is a clear counterexample. On the other hand, it is not hard to
see that if K has Gaussian measure 1/2 then 0 ∈ K. Otherwise, there exists a halfspace H containing K
but not 0, where H clearly has Gaussian measure less than 1/2.
Banaszczyk’s theorem gives the best known bound for the notorious Komlós conjecture [29], a generalization of the Beck-Fiala conjecture, which states thatPfor any sequence of vectors v1 , . . . , vn ∈ Rm of `2
n
n
norm at most 1, there exists χ ∈ {−1, 1} such that k i=1√
χi vi k∞ is a constant independent
of m and n.
√
In this context, Banaszczyk’s theorem gives a bound of O( log m), because an O( log m) scaling of the
unit ball of `m
∞ has Gaussian measure 1/2. Banaszczyk’s theorem together with estimates on the Gaussian
√
measure of slices of the `m
∞ ball due to Barthe, Guedon, Mendelson, and Naor [6] give a bound of O( log d),
where d ≤ min{m, n} is the dimension of the span of v1 , . . . , vn . A well-known
reduction (see e.g. Lecture
p
9 in [29]), shows that this bound for the Komlós problem implies an O( t log min{m, n}) bound in the
Beck-Fiala setting.
While the above results only deal with the case of K being a cube, Banaszczyk’s theorem has also been
applied to other cases. It was used in [3] to give the best known bound on the Steinitz conjecture. In this
problem, the input is a set of vectors v1 , . . . , vn in Rm of norm at most one and summing to 0. The aim is to
find a permutation π : [n] → [n] to minimise the maximum sum prefix of the vectors rearranged according
Pk
to √
π i.e. to minimize maxk∈[n] k i=1 vπ(i) k. The Steinitz conjecture is that this bound should always be
O( m), irrespective
√of the number of vectors, and using the vector balancing theorem Banaszczyk proved a
√
bound of O( m + log n) for the `2 norm.
More recently, Banaszczyk’s theorem was applied to more general symmetric polytopes in Nikolov and
Talwar’s approximation algorithm [25] for a hereditary notion of discrepancy. Hereditary discrepancy is
defined as the maximum discrepancy of any restriction of the set system to a subset of the universe. In [25]
it was shown that an effan efficiently computable quantity, denoted γ2 , bounds hereditary discrepancy from
above and from below for any given set system, up to polylogarithmic factors. For the upper bound they
used Banaszczyk’s theorem for a natural polytope associated with the set system. However, since there is
no known algorithmic version of Banaszczyk’s theorem for a general body, it is not known how to efficiently
compute colorings that achieve the discrepancy upper bounds in terms of γ2 . The recent work on algorithmic
bounds in the Komlós setting does not address this more general problem.
Banaszczyk’s proof of Theorem 1 follows an ingenious induction argument, which folds
effect of
Pthe
n
choosing the sign of vn into the body K. The first observation is that finding a point of the set i=1 {−vi , vi }
Pn−1
inside K is equivalent to finding a point of i=1 {−vi , vi } in K − vn ∪ K + vn . Inducting on this set is
not immediately possible because it may no longer be convex. Instead, Banaszczyk shows that a convex
subset K ∗ vn of (K − vn ) ∪ (K + vn ) has Gaussian measure at least that of K, as long as K has measure
at least 1/2, which allows him to induct on K ∗ vn . In the base case, he needs to show that a convex body
of Gaussian measure at least 1/2 must contain the origin, but this fact follows easily from the hyperplane
2
separation theorem, as indicated above. While extremely elegant, Banaszczyk’s proof can be seen as relatively
mysterious, as it does not seem to provide any tangible insights as to what the colorings look like.
1.1
Our Results
As our main contribution, we help demystify Banaszczyk’s theorem, by showing that it is equivalent, up to
a constant factor in the length of the vectors, to the existence of certain subgaussian coloring distributions.
Using this equivalence, as our p
second main contribution, we give an efficient algorithm that recovers Banaszczyk’s theorem up to a O( log min {m, n}) factor for all convex bodies. This improves upon the best
previous algorithms of Rothvoss [27], Eldan and Singh [14], which only recover the theorem for symmetric
convex bodies up to a O(log min {m, n}) factor.
As a major consequence of our equivalence, we show that for any sequence v1 , . . . , vn ∈ Rm of short enough
n
vectors there exists a probability distribution χ ∈ {−1, 1} over colorings P
such that, for any symmetric
n
m
convex body K ⊆ R of Gaussian measure at least 1/2, the random variable i=1 χi vi lands inside K with
probability at least 1/2. Importantly, if such a distribution can be efficiently sampled, we immediately get a
universal sampler for constructing Banaszczyk colorings for all symmetric convex bodies (we remark that the
recent work of [5] constructs a more restricted form of such distributions). Using random walk techniques,
we show how to implement an approximate version
of this sampler efficiently, which guarantees the same
p
conclusion when the vectors are of length O(1/ log min {m, n}). We provide more details on these results
in Sections 1.1.1 and 1.1.2.
To extend our results to asymmetric convex bodies, we develop a novel recentering procedure and a
corresponding efficient implementation which allows us to reduce the asymmetric setting to the symmetric
one. After this reduction, a slight extension of the aforementioned sampler again yields the desired colorings.
We note that our recentering procedure in fact depends on the target convex body, and hence our algorithms
are no longer universal in this setting. We provide more details on these results in Sections 1.1.3 and 1.1.4.
Interestingly, we additionally show that this procedure can be extended
to yield a completely different
p
coloring algorithm, i.e. not using the sampler, achieving the same O( log min {m, n}) approximation factor.
Surprisingly, the coloring outputted by this procedure is essentially deterministic and has a natural analytic
description, which may be of independent interest.
Before we continue with a more detailed description on our results, we begin with some terminology and a
well-known reduction. Given a set of vectors v1 , . . . , vn ∈ Rm , we shall call a property hereditary if it holds for
all subsets of the vectors. We note that Banaszczyk’s vector balancing bounds restricted to a set of vectors are
hereditary, since a bound on the maximum `2 norm of the vectors is hereditary. We shall say that a property of
Pn
def Pn
n
colorings holds in the linear setting, if when given any shift t ∈ i=1 [−vi , vP
i] = {
i=1 λi vi : λ ∈ [−1, 1] }, one
n
n
can find a coloring (or distribution on colorings) χ ∈ {−1, 1} such that i=1 χi vi − t satisfies the property.
It is well-known that Banaszczyk’s theorem also extends by standard arguments to the linear setting after
reducing the `2 norm bound from 1/5 to 1/10 (a factor 2 drop). This follows, for example, from the general
inequality between hereditary and linear discrepancy proved by Lovasz, Spencer, and Vesztergombi [20].
All the results in this work will in fact hold in the linear setting. When treating the linear setting, it is well
known that one can always reduce to the case where the vectors v1 , . . . , vn P
are linearly independent, and in
n
our setting, when m = n. In particular, assume we are given some shift t ∈ i=1 [−vi , vi ] and that v1 , . . . , vn
are not linearly independent. Then,
Pn using a standard linear algebraic technique, we can find a “fractional
coloring” x ∈ [−1, 1]n such that i=1 xi vi = t, and the vectors (vi : i ∈ Ax ) are linearly independent, where
def
Ax = {i : xi ∈ (−1, 1)} is the set of fractional coordinates (see Lecture 5 in [29], or Chapter 4 in [22]). We
can think of this as a reduction to coloring the linearly independent vectors indexed by Ax . Specifically,
given x as above, define the lifting function Lx : [−1, 1]Ax → [−1, 1]n by
(
zi : i ∈ Ax
Lx (z)i =
, ∀i ∈ [n] .
(1)
xi : i ∈ [n] \ Ax
A
This map takes any coloring χP
∈ {−1, 1} x P
and “lifts” it to a full coloring Lx (χ) ∈ {−1, 1}n . It also satisfies
Ax
the
P property that
P Lx (χ)−t = i∈Ax χi vi − i∈Ax xi vi . So, if we can find a coloring χ ∈ {−1, 1} such that
K, then we would
if we define W as the
i∈Ax χi vi −
i∈Ax xi vi ∈P
P have Lx (χ) − t ∈ K as well.
P Moreover, P
span of (vi : i ∈ Ax ), then i∈Ax χi vi − i∈Ax xi vi ∈ K if and only if i∈Ax χi vi − i∈Ax xi vi ∈ K ∩ W ,
3
so we can replace K with K ∩ W , and work entirely inside W . For convex bodies K with Gaussian measure
at least 1/2, the central section K ∩ W has Gaussian measure that is at least as large, so we have reduced
the problem to the case of |Ax | linearly independent vectors in an |Ax |-dimensional space. (See Section 2 for
the full details.) We shall thus, for simplicity, state all our results in the setting where the vectors v1 , . . . , vn
are in Rn and are linearly independent.
1.1.1
Symmetric Convex Bodies and Subgaussian Distributions
In this section, we detail the equivalence of Banaszczyk’s theorem restricted to symmetric convex bodies
with the existence of certain subgaussian distributions. We begin with the main theorem of this section,
which we note holds in a more general setting than Banaszczyk’s result.
Theorem 2 (Main Equivalence). Let T ⊆ Rn be a finite set. Then, the following parameters are equivalent
up to a universal constant factor independent of T and n:
1. The minimum sb > 0 such that for any symmetric convex body K ⊆ Rn of Gaussian measure at least
1/2, we have that T ∩ sb K 6= ∅.
2. The minimum sg > 0 such that there exists an sg -subgaussian random variable Y supported on T .
We recall that a random vector Y ∈ Rn is s-subgaussian, or subgaussian with parameter s, if for any unit
2
vector θ ∈ S n−1 and t ≥ 0, Pr[|hY, θi| ≥ t] ≤ 2e−(t/s) /2 . In words, Y is subgaussian if all its 1-dimensional
marginals satisfy the same tail bound as the 1-dimensional
Gaussian of mean 0 and standard deviation s.
Pn
To apply the above to discrepancy, we set T = i=1 {−vi , vi }, i.e. all signed combinations of the vectors
v1 , . . . , vn ∈ Rn . In this context, Banaszczyk’s theorem directly implies that sb ≤ 5 maxi∈[n] kvi k2 , and hence
by our
that sg = O(1)
i∈[n] kvi k2 . Furthermore, the above extends to the linear setting letting
Pequivalence
Pmax
n
n
T = i=1 {−vi , vi } − t, for t ∈ i=1 [−vi , vi ], because, as mentioned above, Banaszczyk’s theorem extends
to this setting as well.
The existence of the universal sampler claimed in the previous section is in fact the proof that sb = O(sg )
in the above Theorem. In particular, it follows directly from the following lemma.
Lemma 3. Let Y ∈ Rn be an s-subgaussian random variable. There exists an absolute constant c > 0, such
for any symmetric convex body K ⊆ Rn of Gaussian measure at least 1/2, Pr[Y ∈ s · cK] ≥ 1/2.
Pn
Here, if Y is the sg -subgaussian distribution
Pn supported on i=1 {−vi , vi } − t as above, we simply let χ
denote the random variable such that Y = i=1 χi vi −t. That χ now yields the desired universal distribution
on colorings is exactly the statement of the lemma.
As a consequence of the above, we see that to recover Banaszczyk’s theorem for symmetric convex
bodies,
it suffices to be ableP
to efficiently sample from an O(1)-subgaussian distribution over sets of the type
Pn
n
n
i=1 {−vi , vi } − t, for t ∈
i=1 [−vi , vi ], when v1 , . . . , vn ∈ R are linearly independent and have `2 norm
at
Pnmost 1. Here we rely on homogeneity, that is,Pifn Y is an s-subgaussian random variable supported on
i=1 {−vi , vi } − t then αY is αs-subgaussian on
i=1 {−αvi , αvi } − αt, for α > 0.
The proof of Lemma 3 (see section 3 for more details) follows relatively directly from well-known convex geometric estimates combined with Talagrand’s majorizing measures theorem, which gives a powerful
characterization of the supremum of any Gaussian process.
Unfortunately, Lemma 3 does not hold for asymmetric convex bodies. In particular, if Y = −e1 , the
negated first standard basis vector, and K = {x ∈ Rn : x1 ≥ 0}, the conclusion is clearly false no matter how
much we scale K, even though Y is O(1)-subgaussian and K has Gaussian measure 1/2. One may perhaps
hope that the conclusion still holds if we ask for either Y or −Y to be in s · cK in the asymmetric setting,
though we do not know how to prove this. We note however that this only makes sense when the support of
Y is symmetric, which does not necessarily hold in the linear discrepancy setting.
We now describe the high level idea of the proof for the reverse direction, namely, that sg = O(sb ). For
this purpose, we show that the existence of a O(sb )-subgaussian distribution on T can be expressed as a two
player zero-sum game, i.e. the first player chooses a distribution on T and the second player tries to find
a non-subgaussian direction. Here the value of the game will be small if and only if the O(sb )-subgaussian
distribution exists. To bound the value of the game, we show that an appropriate “convexification” of
the space of subgaussianity tests for the second player can be associated with symmetric convex bodies of
4
Gaussian measure at least 1/2. From here, we use von Neumann’s minimax principle to switch the first and
second player, and deduce that the value of the game is bounded using the definition of sb .
1.1.2
The Random Walk Sampler
From the algorithmic perspective, it turns out that subgaussianity is a very natural property in the context
of random walk approaches to discrepancy minimization. Our results can thus be seen as a good justification
for the random walk approaches to making Banaszczyk’s theorem constructive.
At a high level, in such approaches one runs a random walk over the coordinates of a “fractional coloring”
χ ∈ [−1, 1]n until all the coordinates hit either 1 or −1. The steps of such a walk usually come from
Gaussian increments (though not necessarily spherical), which try to balance the competing goals of keeping
n
discrepancy low and moving the fractional coloring χ closer to {−1, 1} . Since a sum of small centered
Gaussian increments is subgaussian with the appropriate parameter, it is natural to hope that the output of
a correctly implemented random walk is subgaussian. Our main result in this setting is that this is indeed
possible to a limited extent, with the main caveat being that the walk’s output will not be “subgaussian
enough” to fully recover Banaszczyk’s theorem.
Pn
Theorem 4. Let v1 , . . . , vn ∈ Rn be vectors of `2 norm at most 1 and let t ∈ i=1 [−vi , vi ]. Then, there is
n
an expected
algorithm which outputs a random coloring χ ∈ {−1, 1} such that the random
Pnpolynomial time √
variable i=1 χi vi − t is O( log n)-subgaussian.
To achieve the above sampler, we guide our random walk using solutions to the so-called vector Kómlos
program, whose feasibility was first given by Nikolov [24], and show subgaussianity using well-known martingale concentration bounds. Interestingly, the random walk’s analysis does not rely on phases, and is instead
based on a simple relation between the walk’s convergence time and the subgaussian parameter. As an added
bonus, we also give a new and simple constructive proof of the feasibility of the vector Kómlos program (see
section 10 for details) which avoids the use of an SDP solver.
Given the results of the previous section, the above random walk is a universal sampler for constructing
the following colorings.
Pn
Corollary 5. Let v1 , . . . , vn ∈ Rn be vectors of `2 norm at most 1, let t ∈ i=1 [−vi , vi ], and let K ⊆ Rn be a
symmetric convex body of Gaussian measure 1/2 (given by a membership oracle).
Pn Then, there is√an expected
n
polynomial time algorithm which outputs a coloring χ ∈ {−1, 1} such that i=1 χi vi − t ∈ O( log n)K.
As mentioned previously, the best previous algorithms in this setting are due to Rothvoss [27], Eldan
and Singh [14], which find a signed combination inside O(log n)K. Furthermore, these algorithms are not
universal, i.e. they heavily depend on the body K. We note that these algorithms are in fact tailored to
find partial colorings inside a symmetric convex body K of Gaussian measure at least 2−cn , for c > 0 small
enough, a setting in which our sampler does not provide any guarantees.
We now recall prior work on random walk based discrepancy minimization. The random walk approach
was pioneered by Bansal [4], who √
used a semidefinite program to guide the walk and gave the first efficient
algorithm matching the classic O( n) bound of Spencer [28] for the combinatorial discrepancy of set systems
satisfying m = O(n). Later, Lovett and Meka [21] provided a greatly simplified walk, removing the need for
the semidefinite program, which recovered the full power of Beck’s entropy method for constructing partial
colorings. Harvey, Schwartz, and Singh [17] defined another random walk based algorithm, which, unlike
previous work and similarly to our algorithm, doesn’t explicitly use phases or produce partial colorings. The
random walks of [21] and [17] both depend on the convex body K; the walk in [21] is only well-defined
in a polytope, while the one in [17] remains well-defined in any convex body, although the analysis still
applies only to the polyhedral setting. Most directly related to this paper is the recent work [5], which
gives a walk that can be viewed as a randomized variant of the original 2t − 1 Beck-Fiala P
proof. This
n
n
walk induces a distribution χ ∈ {−1, 1} on colorings for which each coordinate of the output i=1 χi vi is
O(1)-subgaussian. From the discrepancy perspective, this gives a sampler which finds colorings inside any
axis parallel box of Gaussian measure at least 1/2 (and their rotations, though not in a universal manner),
matching Banaszczyk’s result for this class of convex bodies.
5
1.1.3
Asymmetric Convex Bodies
In this section, we explain how our techniques extend to the asymmetric setting. The main difficulty in the
asymmetric setting is that one cannot hope to increase the Gaussian mass of an asymmetric convex body by
simply scaling it. In particular, if we take K ⊆ Rn to be a halfspace through the origin, e.g. {x ∈ Rn : x1 ≥ 0},
then K has Gaussian measure exactly 1/2 but sK = K for all s > 0. At a technical level, the lack of any
measure increase under scaling breaks the proof of Lemma 3, which is crucial for showing that subgaussian
coloring distributions produce combinations that land inside K.
The main idea to circumvent this problem will be to reduce to a setting where the mass of K is “symmetrically distributed” about the origin, in particular, when the barycenter of K under the induced Gaussian
measure is at the origin. For such a body K, we show that a constant factor scaling of K ∩ −K also has
Gaussian mass at least 1/2, yielding a direct reduction to the symmetric setting.
To achieve this reduction, we will use a novel recentering procedure, which will both carefully fix certain
coordinates of the coloring as well as shift the body K to make its mass more “symmetrically distributed”.
The guarantees of this procedure are stated below:
Pn
Theorem 6 (Recentering Procedure). Let v1 , . . . , vn ∈ Rn be linearly independent, t ∈
i=1 [−vi , vi ],
and K ⊆ Rn be a convex body P
of Gaussian measure at least 1/2. Then, there exists a fractional coloring
n
x ∈ [−1, 1]n , such that for p = i=1 xi vi − t, Ax = {i ∈ [n] : xi ∈ (−1, 1)} and W = span(vi : i ∈ Ax ), the
following holds:
1. p ∈ K.
2. The Gaussian measure of (K − p) ∩ W on W is at least the Gaussian measure of K.
R
2
3. The barycenter of (K − p) ∩ W is at the origin, i.e. (K−p)∩W ye−kyk /2 dy = 0.
n
By convention, if the procedure returns a full coloring x ∈ {−1, 1} (in which case, since p ∈ K, we are
done), we shall treat conditions 2 and 3 as satisfied, even though W = {0}. At a high level, the recentering
procedure allows us to reduce the initial vector balancing problem to one in a possibly lower dimension
with respect to “well-centered” convex body of no smaller Gaussian measure, and in particular, of Gaussian
measure at least 1/2. Interestingly, as mentioned earlier in the introduction, the recentering procedure can
also be extended to yield a full coloring algorithm. We explain the high level details of its implementation
together with this extension in the next subsection.
To explain how to use the fractional coloring x from Theorem 6 to get a useful reduction, recall the
lifting function Lx : [−1, 1]Ax → [−1, 1]n defined in (1). WePreduce the initial
P vector balancing problem to
Ax
such
that
χ
v
−
the problem
of
finding
a
coloring
χ
∈
{−1,
1}
i∈Ax i i
i∈Ax xi vi ∈ (K − p) ∩ W (note
P
P
that i∈Ax χi vi − i∈Ax xi vi ∈ W by construction). Then we can lift this coloring to Lx (χ), which satisfies
X
i∈Ax
χ i vi −
X
xi vi ∈ (K − p) ∩ W ⇔
n
X
Lx (χ)i vi − t ∈ K.
i=1
i∈Ax
def
From here, the guarantee that K 0 = (K − p) ∩ W has Gaussian measure at least 1/2 and barycenter at
the origin allows a direct reduction to the symmetric setting. Namely, we can replace K 0 by the symmetric
convex body K 0 ∩ −K 0 without losing “too much” of the Gaussian measure of K 0 . This is formalized
by the following extension of Lemma 3, which directly implies a reduction to subgaussian sampling as in
section 1.1.1.
Lemma 7. Let Y ∈ Rn be an s-subgaussian random variable. There exists an absolute constant c > 0,
such for any convex body K ⊆ Rn of Gaussian measure at least 1/2 and barycenter at the origin, Pr[Y ∈
s · c(K ∩ −K)] ≥ 1/2.
P
P
A
In particular, if there exists a distribution over colorings χ ∈ {−1, 1} x such that i∈Ax χi vi − i∈Ax xi vi
as above is 1/c-subgaussian, Lemma 7 implies that the random signed combination lands inside K 0 with
probability at least 1/2. Thus, the asymmetric setting can be effectively reduced to the symmetric one, as
claimed.
6
Crucially, the recentering procedure in Theorem 6 can be implemented in probabilistic polynomial time if
one relaxes the barycenter condition from being exactly 0 to having “small” norm (see section 6 for details).
Furthermore, the estimate in Lemma 7 will be robust to such perturbations. Thus, to constructively recover
the colorings in the asymmetric setting, it will still suffice to be able to generate good subgaussian coloring
distributions.
Combining the sampler from Theorem 4 together with the recentering
procedure, we constructively
√
recover Banaszczyk’s theorem for general convex bodies up to a O( log n) factor.
Theorem 8 (Weak Constructive Banaszczyk). There exists a probabilistic polynomial time
√ algorithm which,
n
on input a linearly
independent
set
of
vectors
v
,
.
.
.
,
v
∈
R
of
`
norm
at
most
c/
log n, c > 0 small
1
n
2
Pn
enough, t ∈ i=1 [−vi , vi ], and a (not necessarily symmetric) convex body K ⊆ Rn of Gaussian measure at
n
least
Pn 1/2 (given by a membership oracle), computes a coloring χ ∈ {−1, 1} such that with high probability
i=1 χi vi − t ∈ K.
As far as we are aware, the above theorem gives the first algorithm to recover Banaszczyk’s result for
asymmetric convex bodies under any non-trivial restriction. In this context, we note that the algorithm of
Eldan and Singh [14] finds “relaxed” partial colorings, i.e. where the fractional coordinates of the coloring
are allowed to fall outside [−1, 1], lying inside an n-dimensional convex body of Gaussian measure at least
2−cn . However, it is unclear how one could use such partial colorings to recover the above result, even with
a larger approximation factor.
1.1.4
The Recentering Procedure
In this section, we describe the details of the recentering procedure. We leave a thorough description of its
algorithmic implementation however to section 6, and only provide its abstract instantiation here.
Before we begin, we give a more geometric view of the vector balancing problem and the recentering
procedure,
which help clarify the exposition. Let v1 , . . . , vn ∈ Rn be linearly independent vectors and
Pn
t ∈ i=1 [−vi , vi ]. Given the target body K ⊆ Rn of Gaussian measure at least 1/2, weP
can restate the
n
vector balancing problem geometrically as that of finding a vertex of the parallelepiped P = i=1 [−vi , vi ]−t
lying inside K. Here, the choice of t ensures that 0 ∈ P . Note that this condition is necessary, since otherwise
there exists a halfspace separating P from 0 having Gaussian measure at least 1/2.
Recall now that in the linear setting, and using this geometric language, Banaszczyk’s theorem implies
that if P contains the origin, and maxi∈[n] kvi k2 ≤ 1/10 (which we do not need to assume here), then any
convex body of Gaussian measure at least 1/2 contains a vertex of P . Thus, for our given target body K,
we should make our situation better replacing P and K by P − q and K − q, if q ∈ P is a shift such that
K − q has higher Gaussian measure than K. In particular, given the symmetry of Gaussian measure, one
would intuitively expect that if the Gaussian mass of K is not symmetrically distributed around 0, there
should be a shift of K which increases its Gaussian measure.
In the current language, fixing a color
Pχi ∈ {−1, 1} for vector vi , corresponds to restricting ourselves to
finding a vertex in the facet F = χi vi + j6=i [−vj , vj ] − t of P lying inside K. Again intuitively, restricting
to a facet of P should improve our situation if the Gaussian measure of the corresponding slice of K in the
lower dimension is larger than that of K. To make this formal, note that when inducting on a facet F of P
(which is an n − 1 dimensional parallelepiped), we must choose a center q ∈ F to serve as the new origin in
the lower dimensional space. Precisely, this can be expressed as inducting on the parallelepiped F − q and
shifted slice (K − q) ∩ span(F − q) of K, using the n − 1 dimensional Gaussian measure on span(F − q).
With the above viewpoint, one can restate the goal of the recentering procedure as that of finding a point
q ∈ P ∩K, such that smallest facet F of P containing q, satisfies that (K −q)∩span(F −q) has its barycenter
at the origin and Gaussian measure no smaller than that of K. Recall that as long as (K − q) ∩ span(F − q)
has Gaussian measure at least 1/2, we are guaranteed that 0 ∈ K − q ⇒ q ∈ K. With this geometry in
mind, we implement the recentering procedure as follows:
Compute q ∈ P so that the Gaussian mass of K − q is maximized. If q is on the boundary of P , letting
F denote a facet of P containing q, induct on F − q and the slice (K − q) ∩ span(F − q) as above. If q is in
the interior of P , replace P and K by P − q and K − q, and terminate.
We now explain why the above achieves the desired result. Firstly, if the maximizer q is in a facet F of P ,
then a standard convex geometric argument reveals that the Gaussian measure of (K − q) ∩ span(F − q) is no
7
smaller than that of K − q, and in particular, no smaller than that of K. Thus, in this case, the recentering
procedure fixes a color for “free”. In the second case, if q is in the interior of P , then a variational argument
gives
that the barycenter of K − q under the induced Gaussian measure must be at the origin, namely,
R
−x2 /2
xe
dx = 0.
K−q
To conclude this section, we explain how to extend the recentering procedure to directly produce a
deterministic
√ coloring satisfying Theorem 8. For this purpose, we shall assume that v1 , . . . , vn have length
at most c/ log n, for a small enough constant c > 0. To begin, we run the recentering procedure as above,
which returns P and K, with K having its barycenter at the origin. We now replace P, K by a joint scaling
αP, αK, for α > 0 a large enough constant, so that αK has Gaussian mass at least 3/4. At this point,
we run the original recentering procedure again with the following modification: every time we get to the
situation where K has its barycenter at the origin, induct on the closest facet of P closest to the origin.
More precisely, in this situation, compute a point p on the boundary of P closest to the origin, and, letting
F denote the facet containing p, induct on F − p and (K − p) ∩ span(F − p). At the end, return the final
found vertex.
Notice that, as claimed, the coloring (i.e. vertex) returned by the algorithm is indeed deterministic. The
reason the above algorithm works is the following. While we cannot guarantee, as in the original recentering
procedure, that the Gaussian mass of (K − p) ∩ span(F − p) does not√decrease, we can instead show that
it decreases only very slowly. In particular, we use the bound of O(1/ log n) on the length of the vectors
v1 , . . . , vn to show that every time we induct, the Gaussian mass drops by at most a 1 − c/n factor. More
generally, if the vectors had length at most d > 0, for d small enough, the drop would be of the order
2
1 − ce−1/(cd) , for some constant c > 0. Since we “massage” K to have Gaussian mass at least 3/4 before
applying the modified recentering algorithm, this indeed allows to induct n times while keeping the Gaussian
mass above 1/2, which guarantees that the final vertex is in K. To derive the bound on the rate of decrease
of Gaussian mass, we prove a new inequality on the Gaussian mass of sections of a convex body near the
barycenter (see Theorem 41), which may be of independent interest.
As a final remark, we note that unlike the subgaussian sampler, the recentering procedure is not scale
invariant. Namely, if we jointly scale P and K by some factor α, the output of the recentering procedure will
not be an α-scaling of the output on the original K and P , as Gaussian measure is not homogeneous under
scalings. Thus, one must take care to appropriately normalize P and K before applying the recentering
procedure to achieve the desired results.
We now give the high level overview of our recentering step implementation. The first crucial observation
in this context, is that the task of finding t ∈ P maximizing the Gaussian measure of K − t is in fact a convex
program. More precisely, the objective function (Gaussian measure of K − t) is a logconcave function of t
and the feasible region P is convex. Hence, one can hope to apply standard convex optimization techniques
to find the desired maximizer.
It turns out however, that one can significantly simplify the required task by noting that the recentering
strategy does not in fact necessarily need an exact maximizer, or even a maximizer in P . To see this, note
that if p is a shift such that K − p has larger Gaussian measure than K, then by logconcavity the shifts
K − αp, 0 < α ≤ 1, also have larger Gaussian measure. Thus, if a we find a shift p ∈
/ P with larger Gaussian
measure, letting αp be the intersection point with the boundary ∂P , we can induct on the facet of P − αp
containing 0 and the corresponding slice of K − αp just as before. Given this, we can essentially “ignore”
the constraint p ∈ P and we treat the optimization problem as unconstrained.
This last observation will allow us to use the following simple gradient ascent strategy. Precisely, we
simply take steps in the direction of the gradient until either we pass through a facet of P or the gradient
becomes “too small”. As alluded to previously, the gradient will exactly equal a fixed scaling of the barycenter
of K − p, p the current shift, under the induced Gaussian measure. Thus, once the gradient is small, the
barycenter will be very close to the origin, which will be good enough for our purposes. The last nontrivial
technical detail is how to efficiently estimate the barycenter, where we note that the barycenter is the
expectation of a random point inside K − p. For this purpose, we simply take an average of random samples
from K − p, where we generate the samples using rejection sampling, using the fact that the Gaussian
measure of K is large.
Conclusion and Open Problems In conclusion, we have shown a tight connection between the existence
of subgaussian coloring distributions and Banaszczyk’s vector balancing theorem. Furthermore, we make use
8
of this connection to constructively recover a weaker version of this theorem. The main open problem we
leave is thus to fully recover Banaszczyk’s result. As explained above, this reduces to finding a distribution
on colorings such that the output random signed combination is O(1)-subgaussian, when the input vectors
have `2 norm at most 1. We believe this approach is both attractive and feasible, especially given the recent
work [5], which builds a distribution on colorings for which each coordinate of the output random signed
combination is O(1)-subgaussian.
Organization In section 2, we provide necessary preliminary background material. In section 3, we give
the proof of the equivalence between Banaszczyk’s vector balancing theorem and the existence of subgaussian
coloring distributions. In section 5, we give our random walk based coloring algorithm. In section 6, we
describe the implementation of the recentering procedure. In section 7, we give the algorithmic reduction
from asymmetric bodies to symmetric bodies, giving the proof of Theorem 8. In section 8, we show how
extend the recentering procedure to a full coloring algorithm. In section 9, we prove the main technical
estimate on the Gaussian measure of slices of a convex body near the barycenter, which is needed for the
algorithm in 8. Lastly, in section 10, we give our constructive proof of the feasibility of the vector Kómlos
program.
Acknowledgments We would like to thank the American Institute for Mathematics for hosting a recent
workshop on discrepancy theory, where some of this work was done.
2
Preliminaries
Basic Concepts We write log x and logp
for the logarithm base e and base 2 respectively.
2 x,
Pnx > 0,
2 to be its Eucliean norm. Let B n = {x ∈ Rn : kxk ≤ 1}
x
For a vector x ∈ Rn , we define kxk2 =
2
2
i=1 i
denote the unit Euclidean ball and S n−1 = {x ∈ P
Rn : kxk2 = 1} = ∂Bn2 denote the unit sphere in Rn . For
n
x, y ∈ Rn , we denote their inner product hx, yi = i=1 xi yi .
For subsets A, B ⊆ Rn , we denote their Minkowski sum A + B = {a + b : a ∈ A, b ∈ B}. Define span(A)
to be the smallest linear subspace containing A. We denote the boundary of A by ∂A. We use the phrase ∂A
relative to span(A) to specify that we are computing the boundary with respect to the subspace topology
on span(A).
A set K ⊆ Rn is convex if for all x, y ∈ K,λ ∈ [0, 1], λx + (1 − λ)y ∈ K. K is symmetric if K = −K. We
shall say that K is a convex body if additionally it is closed and has non-empty interior. We note that the
usual terminology, a convex body is also compact (i.e. bounded), but we will state this explicitly when it is
necessary. If convex body contains the origin in its interior, we say that K is 0-centered.
We will need the concept of a gauge function for 0-centered convex bodies. For bounded symmetric
convex bodies, this functional will define a standard norm.
Proposition 9. Let K ⊆ Rn be a 0-centered convex body. Defining the gauge function of the body K by
kxkK = inf {s ≥ 0 : x ∈ sK}, the following holds:
1. Finiteness: kxkK < ∞, for x ∈ Rn .
2. Positive homogeneity: kλxkK = λkxkK , for x ∈ Rn , λ ≥ 0.
3. Triangle inequality: kx + ykK ≤ kxkK + kykK , for x, y ∈ Rn .
Furthermore, if K is additionally bounded and symmetric, then k·kK is a norm which we call the norm induced
by K. In particular, k · kK additionally satisfies that kxkK = 0 iff x = 0 and kxkK = k − xkK ∀x ∈ Rn .
Gaussian and subgaussian random variables We define n-dimensional standard Gaussian X ∈ Rn to
1
−kxk22 /2
be the random variable with density √2π
for x ∈ Rn .
ne
Definition 10 (Subgaussian Random Variable). A random variable X ∈ R is σ-subgaussian, for σ > 0, if
∀t ≥ 0,
2
1
Pr[|X| ≥ t] ≤ 2e− 2 (t/σ) .
9
We note that the canonical example of a 1-subgaussian distribution is the 1-dimensional standard Gaussian
itself.
For a vector valued random variable X ∈ Rn , we say that X is σ-subgaussian if all its one dimensional
marginals are. Precisely, X is σ-subgaussian if ∀θ ∈ S n−1 , the random variable hX, θi is σ-subgaussian.
We remark that from definition 10, it follows directly that if X is σ-subgaussian then αX is |α|σsubgaussian for any α ∈ R.
The following standard lemma allows us to deduce subgaussianity from upper bounds on the Laplace
transform of a random variable. We include a proof in the appendix for completeness.
Lemma 11. Let cosh(x) = 21 (ex + e−x ) for x ∈ Rn . Let X ∈ Rn be a random vector. Assume that
2
E[cosh(hw, Xi)] ≤ βekσwk2 /2 , ∀w ∈ Rn ,
p
for some σ > 0 and β ≥ 1. Then X is σ log2 β + 1-subgaussian. Furthermore, for X ∈ Rn standard
2
Gaussian, E[cosh(hw, Xi)] = ekwk2 /2 for w ∈ Rn .
Gaussian measure We define γn to be the n-dimensional Gaussian measure on Rn . Precisely, for any
measurable set A ⊆ Rn ,
Z
2
1
γn (A) = √ n
e−kxk2 /2 dx,
(2)
2π A
noting that γn (Rn ) = 1. We will also need lower dimensional Gaussian measures restricted to linear subspaces
of Rn . Thus, if A ⊆ W , W ⊆ Rn a linear subspace of dimension k, then γk (A) should be understood as
the Gaussian measure of A within W , where W is treated as the whole space. When convenient, we will
also use the notation γW (A) to denote γdim(W ) (A ∩ W ). When treating one dimensional Gaussian measure,
we will often denote γ1 ((a, b)), where (a, b) is an interval, simply by γ1 (a, b) for notational convenience. By
convention, we define γ0 (A) = 1 if 0 ∈ A and 0 otherwise.
An important concept used throughout the paper is that of the barycenter under the induced Gaussian
measure.
Definition 12 (Barycenter). For a convex body K ⊆ Rn , we define its barycenter under the induced Gaussian
measure, by
Z
2
1
dx
b(K) = √ n
xe−kxk2 /2
.
γn (K)
2π K
2
1
−kxk2 /2
Note that b(K) = E[X], if X is the random variable supported on K with probability density √2π
/γn (K).
ne
n
Extending the definition to slices of K, for any linear subspace W ⊆ R , we refer to the barycenter of K ∩ W
to denote the one relative to the dim(W )-dimensional Gaussian measure on W (i.e. treating W as the whole
space).
Throughout the paper, we will need many inequalities regarding the Gaussian measure. The first important inequality is the Prékopa-Leindler inequality, which states that for λ ∈ [0, 1] and A, B, λA+(1−λ)B ⊆ Rn
measurable subsets, that
γn (λA + (1 − λ)B) ≥ γn (A)λ γn (B)1−λ .
(3)
We note that the Prékopa-Leindler inequality applies more generally to any logconcave measure on Rn , i.e. a
measure defined by a density whose logarithm is concave. Importantly, this inequality directly implies that
if A ⊆ Rn is convex, then log γn (A + t), for t ∈ Rn , is a concave function of t.
We will need the following powerful inequality of Ehrhard, which provides a crucial strengthening of
Prékopa-Leindler for Gaussian measure.
Theorem 13 (Ehrhard’s inequality [13, 9]). For Borel sets A, B ⊆ Rn and 0 ≤ λ ≤ 1,
Φ−1 (γn (λA + (1 − λ)B)) ≥ λΦ−1 (γn (A)) + (1 − λ)Φ−1 (γn (B))
where Φ(a) = γ1 ((−∞, a]) for all a ∈ R.
10
The power of the Ehrhard inequality is that it allows us to reduce many non-trivial inequalities about
Gaussian measure to two dimensional ones.
One can use it to show the following standard inequality on the Gaussian measures of slices of a convex
body. We include a proof for completeness.
Lemma 14. Given a convex body K ⊆ Rn with γn (K) ≥ 1/2, and a linear subspace H ⊆ Rn of dimension
k. Then, γk (K ∩ H) ≥ γn (K).
Proof. Clearly it suffices to prove the lemma for k = n − 1. Since Gaussian distribution is rotation invariant,
without loss of generality, H = {x ∈ Rn : x1 = 0}. Let Kt = {x ∈ K : x1 = t} denote a slice of K at x1 = t.
Then,
Z ∞ −t2 /2
e
√
γn−1 (Kt − te1 )dt
γn (K) =
2π
−∞
where γn−1 (Kt − te1 ) = 0 outside support of K.
Define W ⊆ R2 as W = {(x, y) : y ≤ f (x)} where f : R → R is defined as
f (t) = Φ−1 (γn−1 (Kt − te1 ))
and f (t) = −∞ outside the support of K. It follows that γ2 (W ) = γn (K) ≥ 1/2. By Ehrhard’s inequality,
f is concave on its support. Hence, W is a closed convex body.
Let g = Φ−1 (γn (K)) ≥ 0. γn−1 (K ∩ H) ≥ γn (K) is then equivalent to showing (0, g) ∈ W . If (0, g) 6∈ W ,
then there exists a halfspace L such that W ⊆ L and (0, g) 6∈ L. Let d be the distance of origin (0, 0) from
∂L, the boundary of L. Since (0, g) 6∈ L and γ2 (L) ≥ 1/2, d < g. But this implies
γ2 (L) = γ1 (−∞, d) < γ1 (−∞, g) = γn (K) = γ2 (W )
contradicting W ⊆ L.
Vector Balancing: Reduction to the Linearly Independent Case In this section, we detail the
standard vector balancing reduction to the case where the vectors are linearly independent. We will also
cover some useful related concepts and definitions, which will be used throughout the paper.
Definition 15 (Lifting Function). For a fractional coloring x ∈ [−1, 1]n , denote the set of fractional coordinates by Ax = {i ∈ [n] : xi ∈ (−1, 1)}. From here, for z ∈ [−1, 1]Ax , we define the lifting function
Lx : [−1, 1]Ax → [−1, 1]n by
(
zi : i ∈ Ax
Lx (z)i =
, ∀i ∈ [n] .
xi : i ∈ [n] \ Ax
Importantly, for χ ∈ {−1, 1}
n
to full colorings in {−1, 1} .
Ax
n
we have that Lx (χ) ∈ {−1, 1} . Thus, Lx sends full colorings in {−1, 1}
Ax
The lifting function above is useful in that it allows us, given a fractional coloring x ∈ [−1, 1] with some
of its coordinates set to {−1, 1}, to reduce any linear vector balancing problem to one on a smaller number
of coordinates. We detail this in the following lemma.
Pn
n
Lemma 16. Let v1 ,P
. . . , v m ∈ Rm , t ∈
i=1 [−vi , vi ], and K ⊆ R . Then given a fractional coloring
n
n
x ∈ [−1, 1] and p = i=1 xi vi − t, the following holds:
1. For z ∈ [−1, 1]Ax , we have that
X
zi vi −
i∈Ax
X
x i vi + p =
n
X
Lx (z)i vi − t .
i=1
i∈Ax
2. For z ∈ [−1, 1]Ax , we have that
X
i∈Ax
zi vi −
X
x i vi ∈ K − p ⇔
n
X
i=1
i∈Ax
11
Lx (z)i vi − t ∈ K .
Proof. The first part follows from the computation
X
X
Lx (z)i vi − t =
zi vi +
i∈Ax
i∈[n]
=
X
X
xi v i − t
i∈[n]\Ax
zi vi −
i∈Ax
=
X
X
x i vi + (
i∈Ax
zi vi −
i∈Ax
X
X
xi vi − t)
i∈[n]
x i vi + p .
i∈Ax
The second part follows since
X
zi vi −
X
x i vi ∈ K − p ⇔
i∈Ax
i∈Ax
i∈Ax
X
zi vi −
X
x i vi + p ∈ K ⇔
n
X
Lx (z)i vi − t ∈ K ,
i=1
i∈Ax
where the last equivalence is by part (1).
In terms of a reduction, the above lemma says in words that the linear vector balancing problem with
respect to the vectors
P (vi : i ∈ [n]), shift t and set K, reduces to the linear discrepancy problem on
(vi : i ∈ Ax ), shift i∈Ax xi vi and set K − p.
We now give the reduction to the linearly independent setting.
Pn
Lemma 17. Let v1 , . . . , vn ∈ Rm , t ∈ i=1 [−vi , vi ]. Then, there is a polynomial time algorithm computing
a fractional coloring x ∈ [−1, 1]n such that:
Pn
1.
i=1 xi vi = t.
2. The vectors (vi : i ∈ Ax ) are linearly independent.
P
P
Pn
3. For z ∈ [−1, 1]Ax , i∈Ax zi vi − i∈Ax xi vi = i=1 Lx (z)i vi − t.
Proof. Let x denote a basic feasible solution to the linear system
(
)
n
X
n
y∈R :
yi vi = t, yi ∈ [−1, 1] ∀i ∈ [n] ,
i=1
which clearly can be computed in polynomial time. Note the system is feasible by construction of t. We now
show that x satisfies the required conditions.
Let r ≤ n denote the rank of the matrix (v1 , . . . , vn ). Since x is basic, it must satisfy at least n least of
the constraints at it equality. In particular, at least n − r of the bound constraints must be tight. Thus,
since Ax is the set of fractional coordinates, we must have |Ax | ≤ r. Furthermore, the vectors (vi : i ∈ Ax )
must be linearly independent, since otherwise x is not basic. Finally, for z ∈ [−1, 1]Ax , we have that
X
X
X
X
X
X
zi vi −
x i vi =
zi vi +
x i vi −
x i vi =
Lx (z)i vi − t ,
i∈Ax
i∈Ax
i∈Ax
i∈[n]\Ax
i∈[n]
i∈[n]
as needed.
Let us now apply the above lemma to both the vector balancing problem and the subgaussian sampling
problem.
we have a vector balancing problem with respect to v1 , . . . , vn ∈ Rm , shift
Pn First assume that m
t ∈ i=1 [−vi , vi ], and K ⊆ R a convex body of Gaussian measure at least 1/2. Then applying the above
n
lemma,
P we get x ∈ [−1,
P 1] , such that our vector balancing reduces to the one with respect to (vi : i ∈ Ax ),
shift i∈Ax xi vi ∈ i∈Ax [−vi , vi ], and K. This follows directly from Lemma 17 part 3 using the lifting
function Lx . Now let W = span(vi : i ∈ Ax ), where dim(W ) = |Ax | by linear independence. Clearly, the
reduced vector balancing problem looks for signed combinations in W , and hence we may replace K by
K ∩ W . Here, note that by Lemma 14, γ|Ax | (K ∩ W ) ≥ γm (K) ≥ 1/2. Hence, this reduction reduces to
a problem of the same type, where in addition, the vectors form a basis of the ambient space W . For the
12
n
subgaussian sampling problem, by the identity 3 in Lemma 17, sampling a random coloring χ ∈ {−1, 1}
Pn
Ax
such P
that i=1 χi vP
i − t is subgaussian clearly reduces to sampling
Pn a random coloring χ ∈ {−1, 1} such
that i∈Ax χi vi − i∈Ax xi vi is subgaussian since this equals i=1 Lx (χ)i vi − t. Furthermore, since the
support
support distribution lives in W , to test subgaussianity we need only check the marginals
P of such aP
hθ, i∈Ax χi vi − i∈Ax xi vi i for θ ∈ W ∩ S m−1 . Thus, we may assume that W is the full space. This
completes the needed reductions.
Computational Model To formalize how our algorithms interact with convex bodies, we will use the
following computational model.
To interact algorithmically with a convex body K ⊆ Rn , we will assume that K is presented by a
membership oracle. Here a membership oracle OK on input x ∈ Rn , outputs 1 if x ∈ K and 0 otherwise.
Interestingly, since we will always assume that our convex bodies have Gaussian measure at least 1/2, we
will not need any additional centering (known point inside K) or well-boundedness (inner contained and
outer containing ball) guarantees.
The runtimes of our algorithms will be measured by the number of oracle calls and arithmetic operations
they perform. We note that we use a simple model of real computation here, where we assume that our
algorithms can perform standard operations on real numbers (multiplication, division, addition, etc.) in
constant time.
3
Banaszczyk’s Theorem and Subgaussian Distributions
In this section, we give the main equivalences between Banaszczyk’s vector balancing theorem and the
existence of subgaussian coloring distributions.
The fundamental theorem which underlies these equivalences is known as Talagrand’s majorizing measure
theorem, which provides a nearly tight characterization of the supremum of any Gaussian process using
chaining techniques. We now state an essential consequence of this theorem, which will be sufficient for our
purposes. For a reference, see [31].
Theorem 18 (Talagrand). Let K ⊆ Rn be a 0-centered convex body and Y ∈ Rn be an s-subgaussian random
vector. Then for X ∈ Rn the n-dimensional standard Gaussian, we have that
E[kY k] ≤ s · CT · E[kXk] ,
where CT > 0 is an absolute constant.
As a consequence of the above theorem together with geometric estimates proved in subsection 4.2, we
derive the following lemma, which will be crucial to our equivalences and reductions.
Lemma 19 (Reduction to Subgaussianity). Let Y ∈ Rn be s-subgaussian. Then,
1. If K ⊆ Rn is a symmetric convex body with γn (K) ≥ 1/2, then
E[kY kK ] ≤ 1.5 · CT · s.
In particular, Pr[Y ∈ 3 · CT · sK] ≥ 1/2.
2. If K ⊆ Rn is a convex body with γn (K) ≥ 1/2 and kb(K)k2 ≤
1
√
,
32 2π
then
√
E[kY kK∩−K ] ≤ 2(1 + π 8 ln 2) · CT · s.
√
In particular, Pr[Y ∈ 4(1 + π 8 ln 2) · CT · s(K ∩ −K)] ≥ 1/2.
Proof of Lemma 19. The proof follows immediately by combining Lemmas 26, 30 and Theorem 18. We note
that the lower bounds on the probabilities follow directly by Markov’s inequality.
To state our equivalence, we will need the definitions of the following geometric parameters.
13
Definition 20 (Geometric Parameters). Let T ⊆ Rn be a finite set.
• Define sg (T ) > 0 to be least number s > 0 such that there exists an s-subgaussian random vector Y
supported on T .
• Define sb (T ) > 0 to be the least number s > 0 such that for any symmetric convex body K ⊆ Rn ,
γn (K) ≥ 1/2, T ∩ sK 6= ∅.
We now state our main equivalence, which gives a quantitative version of Theorem 2 in the introduction.
Theorem 21. For T ⊆ Rn be a finite set, the following holds:
1. sb (T ) ≤ 1.5CT · sg (T ).
√
2. sg (T ) ≤ 2 · sb (T ).
Using the above language, we can restate Banaszczyk’s vector balancing theorem restricted to symmetric
convex bodies as follows:
Pn
Theorem 22 ([2]). Let v1 , . . . , vm ∈ Rn . Then sb ( i=1 {−vi , vi }) ≤ 5 maxi∈[m] kvi k2 .
As an immediate corollary of Theorems 21 and 22 (extended to the linear setting) we deduce:
√
Pn
Corollary 23. Let v1 , . . . , vm ∈ Rn . Then sg ( i=1 {−vi , vi }) ≤ 2 · 5 maxi∈[m] kvi k2 . Furthermore, for
√
Pm
Pn
t ∈ i=1 [−vi , vi ], sg ( i=1 {−vi , vi } − t) ≤ 2 · 10 maxi∈[m] kvi k2 .
As explained in the introduction, the above equivalence shows the existence of a universal sampler for
recovering Banaszczyk’s vector balancing theorem for symmetric convex bodies up to a constant factor in
the length of the vectors. Precisely, this follows directly from Lemma 19 part 1 and Corollary 23 (for more
details see the proof of Theorem 21 below).
The following theorem, which we will need, is the classical minimax principle of Von-Neumann.
Theorem 24 (Minimax Theorem [23]). Let X ⊆ Rn , Y ⊆ Rm be compact convex sets. Let f : X × Y → R
be a continuous function such that
1. f (·, y) : X → R is convex for fixed y ∈ Y .
2. f (x, ·) : Y → R is concave for fixed x ∈ X.
Then,
min max f (x, y) = max min f (x, y) .
x∈X y∈Y
y∈Y x∈X
We now proceed to the proof of Theorem 21.
Proof of Theorem 21.
Proof of 1: Let Y ∈ T be the sg (T )-subgaussian random variable. Let K ⊆ Rn be a symmetric convex
body such that γn (K) ≥ 1/2. By Lemma 19 part 1, we have that
E[kY kK ] ≤ 1.5CT · sg (T ).
Thus, there exists x ∈ T such that x ∈ 1.5CT · sg (T )K. Since this holds for all such K, we have that
sb (T ) ≤ 1.5CT · sg (T ) as needed.
14
Proof of 2: Recall the definition of cosh(x) = 12 (ex + e−x ) for x ∈ R. Note that cosh is convex,
symmetric (cosh(x) = cosh(−x)), and non-negative. For w ∈ Rn , define gw : Rn → R≥0 by gw (x) =
2
cosh(hx, wi)/ekwk2 /2 . By Lemma 11, note that E[gw (X)] = 1 for X an n-dimensional standard Gaussian.
Let D denote√the set of probability distributions on T . Our goal is to show that there exists D ∈ D such
that Y ∼ D is 2 · sb (T )-subgaussian. By homogeneity, we may replace T by T /sb (T ), and thus assume
that sb (T ) = 1. To show the existence of the subgaussian distribution, we will show that
inf sup
E [gw (Y )] ≤ 2 .
(4)
D∈D w∈Rn Y ∼D
√
Before proving the bound (4), we show that this suffices to show the existence of the desired 2subgaussian distribution. Let D∗ ∈ D denote the minimizing distribution for (4). Then by definition of
gw , we have that
2
E ∗ [cosh(hw, Y i)] ≤ 2ekwk2 /2 ∀ w ∈ Rn .
(5)
Y ∼D
With
the bounds
p
√ on the Laplace transform in (5), by Lemma 11 with β = 2 and σ = 1, we have that Y is
log2 2 + 1 = 2-subgaussian as needed.
We now prove the estimate in (4). Let C denote the closed convex hull of the functions gw . More
precisely, C is the closure of the set of functions
(
)
k
k
X
X
n
f : T → R≥0 |f =
λi gwi , k ∈ N,
λi = 1, λi ≥ 0 ∀ i ∈ [k], wi ∈ R ∀ i ∈ [k] .
i=1
i=1
By continuity, we clearly have that
inf sup
E [gw (Y )] = inf sup E [f (Y )] .
D∈D w∈Rn Y ∼D
D∈D f ∈C Y ∼D
(6)
The strategy will now be to apply the minimax theorem 24 to (6). For this to hold, we first need that both
D and C are both convex and compact. This is clear for D, since D can be associated with the standard
simplex in R|T | . By construction C is also convex, hence we need only prove compactness. Since T is finite
and C is a closed subset of non-negative functions on T , C can be associated in the natural way with a
|T |
closed subset of R≥0 . To show compactness, it suffices to show that this set is bounded. In particular, it
suffices to show that for f ∈ C, maxx∈T f (x) ≤ M for some universal constant M < ∞. Since every f ∈ C
is a limit of convex combinations of the functions gw , w ∈ Rn , it suffices to show that
sup max gw (x) ≤ ∞ .
w∈Rn x∈T
We prove this with the following computation:
cosh(hw, xi)
2
ekwk2 /2
cosh(maxx∈T kxkkwk2 )
≤ sup
2
n
ekwk2 /2
w∈R
sup max gw (x) = sup max
w∈Rn
x∈T
w∈Rn
x∈T
2
≤ sup e(maxx∈T kxk2 )kwk2 −kwk2 /2
w∈Rn
= e(maxx∈T kxk2 )
2
/2
< ∞.
Thus C is compact as needed. Lastly, note that the function EY ∼D [f (Y )] from D × C is bilinear, and hence
is both continuous and satisfies (trivially) the convexity-concavity conditions in Theorem 24.
By compactness of D and C and continuity, we have that
inf sup E [f (Y )] = min max E [f (Y )] .
D∈D f ∈C Y ∼D
D∈D f ∈C Y ∼D
15
Next, by the minimax theorem 24, we have that
min max E [f (Y )] = max min E [f (Y )] = max min f (x)
D∈D f ∈C Y ∼D
f ∈C D∈D Y ∼D
=
f ∈C x∈T
sup
P
min f (x) .
f= k
i=1 λi gwi
i=1 λi =1, λi ≥0 ∀i∈[k]
wi ∈Rn ∀ i∈[k], k∈N
x∈T
Pk
Pk
Take f = i=1 λi gwi as above (now as a function on Rn ) and let Kf = {x ∈ Rn : f (x) ≤ 2}. Our task now
reduces to showing that ∃x ∈ T such that x ∈ Kf . Since sb (T ) ≤ 1, it suffices to show that γn (Kf ) = Pr[X ∈
Kf ] ≥ 1/2, for X the n-dimensional standard Gaussian, and that Kf is symmetric and convex. Since f is a
convex combination of symmetric and convex functions, it follows that Kf is symmetric and convex. Since
f is non-negative, by Markov’s inequality
k
Pr[X ∈
/ Kf ] = Pr[f (X) ≥ 2] ≤
k
1X
1X
1
E[f (X)]
=
λi E[gwi (X)] =
λi = .
2
2 i=1
2 i=1
2
Hence Pr[X ∈ Kf ] = 1 − Pr[X ∈
/ Kf ] ≥ 1/2, as needed.
4
Analysis of the Recentering Procedure
We now give the crucial tool to reduce the asymmetric setting to the symmetric setting, namely, the recentering procedure corresponding to Theorem 6 in the introduction. In the next subsection (subsection 4.1),
we detail how to use this procedure to yield the desired reduction.
Proof of Theorem 6 (Recentering Procedure).
We first recall the desired guarantees. For linearly independent
Pn
vectors v1 , . . . , vn ∈ Rn , a shift t ∈ i=1 [−vi , vi ], and a convex body K ⊆ RnPof Gaussian measure at least
n
1/2, we would like to find a fractional coloring x ∈ [−1, 1]n , such that for p = i=1 xi vi − t and the subspace
W = span(vi : i ∈ Ax ), the following holds:
1. p ∈ K.
2. γ|Ax | ((K − p) ∩ W ) ≥ γn (K).
3. b((K − p) ∩ W ) = 0.
We shall prove this by induction on n. Note that the base case n = 0, reduces to the statement that
0 ∈ K, which is trivial.
For a fractional coloring x ∈ [−1, 1]n , we remember first that Ax denotes the set of fractional coordinates
and that Lx P
: [−1, 1]Ax → [−1, 1]n is the lifting function (see Definition 15 for details).
n
Let P = i=1 [−vi , vi ] − t. Define the function
Pnf (p) = log γn (K − p) for p ∈ P . Compute the maximizer
p of f over P . Let x ∈ [−1, 1]n satisfy that p = i=1 xi vi − t and let W = span(vi : i ∈ Ax ). Note first that
since γn (K − p) ≥ γn (K) ≥ 1/2, by Lemma 27 part 1 we have that 0 ∈ K − p ⇒ p ∈ K.
Assume first that p is in the interior of P . Then, since p is a maximizer and does not touch the boundary
of P , by the KKT conditions we must have that ∇f (p) = 0. From here, direct computation reveals that
∇f (p) = b(K − p). Again, since y does not touch again constraints of P , we see that Ax = [n] and hence
W = Rn . Thus, as claimed, x satisfies the conditions of the theorem.
Assume now that p ∈ ∂P . From here, we must have that |Ax | < n and hence dim(W ) = |Ax | < n. Next,
by Lemma 14, we see that
γ|Ax | ((K − p) ∩ W ) ≥ γn (K − p) ≥ γn (K) ≥ 1/2 .
By Lemma 16 part 2, for z ∈ [−1, 1]Ax , we have that
X
i∈Ax
zi vi −
X
xi vi ∈ K − p ⇔
n
X
i=1
i∈Ax
16
Lx (z)i vi − t ∈ K .
P
Thus, we may apply induction on the vectors (vi : i ∈ Ax ), the shift i∈Ax xi vi and convex body (K −p)∩W ,
and recover z ∈ [−1, 1]Ax , such that for Wz = span(vi : i ∈ Az ), we get
X
zi vi −
X
def
xi vi = pz ∈ K − p ,
(7)
i∈Ax
i∈Ax
and
γ|Az | ((K − p − pz ) ∩ Wz ) ≥ γ|Ax | (K − p) ,
(8)
b((K − p − pz ) ∩ Wz ) = 0 .
(9)
and
We now claim that w = Lx (z) satisfies the conditions of the theorem. To see this, note that by Lemma 16
part 1,
n
n
X
X
X
X
x i vi + p = p z + p .
wi vi − t =
zi vi −
Lx (z)i vi − t =
i=1
i=1
i∈Ax
i∈Ax
Furthermore, since pz ∈ K − p we have that pz + p ∈ K. Next, clearly Aw = Az and hence span(vi : i ∈
Aw ) = Wz . The claim thus follows by combining (7), (8), (9).
4.1
Reduction from Asymmetric to Symmetric Convex Bodies
As explained in the introduction, the recentering procedure allows us to reduce Banaszczyk’s vector balancing
theorem for all convex bodies to the symmetric case, and in particular, to the task of subgaussian sampling.
We give this reduction in detail below.
n
Let v1 , . . . , vn , t, and K be as in Theorem 6, and let
Pnx ∈ [−1, 1] the fractional coloring guaranteed by
the recentering procedure. As in Theorem 6, let p = i=1 xi vi − t and W = span(vi : i ∈ Ax ). We shall
now assume that maxi∈[n] kvi k2 ≤ c, a constant to be chosen later. From here, by Lemma 16 part 2, for
A
χ ∈ {−1, 1} x ,
X
X
X
χi vi −
xi vi ∈ (K − p) ∩ W ⇔
Lx (χ)i vi − t ∈ K .
i∈Ax
i∈Ax
i∈[n]
Let C = (K − p) ∩ W and d = |Ax |. By the guarantees on the recentering procedure, we know that
γd (C) ≥ 1/2 and b(C) = 0. Then by Lemma 30 in section 4.2, for X ∈ W the d-dimensional standard
Gaussian on W , we have that
√
E[kXkC∩−C ] ≤ 2 E[kXkC ] ≤ 2(1 + π 8 ln 2) .
√
Hence by Markov’s inequality, Pr[X ∈ 4(1 + π 8 ln 2)(C ∩ −C))] ≥ 1/2. At this point, using Banaszczyk’s
theorem in the linear setting for symmetric bodies (which loses a factor of 2), if the `2 norm bound c satisfies
√
1/c ≥ 10 · 4(1 + π 8 ln 2) ,
Ax
then by homogeneity there exists χ ∈ {−1, 1}
X
i∈Ax
χi vi −
X
such that
xi vi ∈ C ∩ −C ⇒
n
X
Lx (χ)i vi − t ∈ K .
i=1
i∈Ax
Hence, the reduction to the symmetric case follows.
We can also achieve the same with a subgaussian sampler, though the vectors should be shorter. In
particular, applying corollary 23, if
√
√
1/c ≥ 2 · 10 · CT · 4(1 + π 8 ln 2),
17
P
P
A
then there exists a distribution on colorings χ ∈ {−1, 1} x such that i∈Ax χi vi − i∈Ax xi vi is (CT · 4(1 +
√
π 8 ln 2))−1 -subgaussian. From here, by Lemma 19 part 2 applied to C,
"
#
X
X
Pr (
χ i vi −
xi vi ) ∈ C ∩ −C ≥ 1/2 ,
i∈Ax
i∈Ax
as needed.
4.2
Geometric Estimates
In this section, we present the required estimates for the proof of Lemma 19.
The following theorem of Latala and Oleszkiewicz will allow us to translate bounds on Gaussian measure
to bounds on Gaussian norm expectations.
Theorem 25 ([18]). Let X ∈ Rn be a standard n-dimensional Gaussian. Let K ⊆ Rn be a symmetric
convex body, and let α ≥ 0 be chosen such that Pr[X ∈ K] = Pr[|X1 | ≤ α]. Then the following holds:
1. For t ∈ [0, 1], Pr[X ∈ tK] ≤ Pr[|X1 | ≤ tα].
2. For t ≥ 1, Pr[X ∈ tK] ≥ Pr[[X1 | ≤ tα].
Using the above theorem we derive can derive bound goods bounds on Gaussian norm expectations. We
note that much weaker and more elementary estimates than those given in 25 would suffice (e.g. Borell’s
inequality), however we use the stronger theorem to achieve a better constant.
Lemma 26. Let X ∈ Rn be a standard n-dimensional Gaussian. Let K ⊆ Rn be a symmetric convex body
such that Pr[X ∈ K] ≥ 1/2. Then E[kXkK ] ≤ 1.5.
Proof. Let m > 0 satisfy Pr[X ∈ mK] = Pr[kXkK ≤ m] = 21 . Note that m ≤ 1 by our assumption on
K. Let α > 0 denote the number such that Pr[|X1 | ≤ α] = γ1 ([−α, α]) = 21 . Here a numerical calculation
reveals α ≥ .67.
By Theorem 25, we have that Pr[kXkK ≥ tm] ≤ Pr[|X1 | ≥ tα] for t ≥ 1. Thus,
Z ∞
Z ∞
Z ∞
αt
E[kXkK ] =
]dt
Pr[kXkK ≥ t]dt ≤ m +
Pr[kXkK ≥ t]dt ≤ m +
Pr[|X1 | ≥
m
0
m
m
Z
Z
1 ∞
1 ∞
= (1 +
Pr[|X1 | ≥ t]dt)m ≤ 1 +
Pr[|X1 | ≥ t]dt
α α
α α
r
r
Z
2
1 ∞ 2
1
2 −α2 /2
2 e−α /2
−t2 /2
√ (t − α)e
=1+
dt = 1 + (
e
− α/2) = 1/2 +
α α
α
π
π α
2π
r
2
2 e−(.67) /2
≤ 1/2 +
≤ 1.5
π
.67
The following lemma shows that we can find a large ball in K centered around the origin, if either its
Gaussian mass is large or its barycenter is close to the origin.
Lemma 27. Let K ⊆ Rn be a convex body. Then the following holds:
1. If for r ≥ 0, γ1 ((−∞, r]) ≤ γn (K), then rB2n ⊆ K.
√
In particular, if γn (K) = 1/2 + ε, for ε ≥ 0, this holds for r = 2πε.
2. If for 0 ≤ r ≤ 1/2, γ1 ([−2r, 2r]) ≤ γn (K) and kb(K)k2 γn (K) ≤
this holds for r = 1/4 if γn (K) ≥ 1/2 and kb(K)k2 ≤ 32√1 2π .
18
√1
2π
2
· r2 , then rB2n ⊆ K. In particular,
Proof. We begin with part (1). Assume for the sake of contradiction that there exist x ∈ Rn , kxk2 ≤ r, such
that x ∈
/ K. Then, by the separator theorem there exists a unit vector θ ∈ S n−1 , such that maxz∈K hθ, zi <
hθ, xi. In particular, K is strictly contained in the halfspace H = {z ∈ Rn : hθ, zi ≤ g} where g = hθ, xi. Thus
γn (K) < γn (H) = γ1 ((−∞, g]). But note that by Cauchy-Schwarz g ≤ kθk2 kxk2 = r, a clear contradiction
to the assumption on r.
For the furthermore, we first see that
√
1
√
2π
Z
2πε
e
−x2 /2
0
√
( 2πε)
dx ≤ √
=ε.
2π
√
Hence, γ1 ((−∞, 2πε]) ≤ 1/2 + ε = γn (K), as needed.
We now prove part (2). Similarly to the above, if K does not contain a ball of radius r, then there
exists a halfspace H = {z ∈ Rn : hθ, zi ≤ r − ε}, for some 0 < ε ≤ r, such that K ⊆ H. By rotational
invariance of the Gaussian measure, we may assume that θ = e1 . Now let Kt = {x ∈ K : x1 = t} and let
f (t) = γn−1 (Kt − te1 ), where clearly f (t) ∈ [0, 1]. From here, we see that
Z
r−ε
−∞
2
f (t)
e−t /2
dt =
t √
R
2 /2
2π r−ε e−t
√
f (t)dt
−∞
2π
Z
r−ε
−∞
2
e−t /2 γn−1 (Kt − te1 )
t √
dt
γn (K)
2π
(10)
= b(K)1 ≥ −kb(K)k2 .
Thus to get a contradiction, it suffices to show that
Z
r−ε
−∞
2
e−t /2
f (t)dt < −γn (K)kb(K)k2 ,
t √
2π
(11)
for any function f : (−∞, r − ε] → [0, 1] satisfying
Z
r−ε
−∞
2
e−t /2
√
f (t)dt = γn (K).
2π
(12)
From here, it is easy to see that the function f maximizing the left hand side of (11) satisfying (12) must be
the indicator function of an interval with right end point r − ε, i.e. the function f which pushes mass “as far
to the right” as possible. Now let l ≤ r − ε denote the unique number such that γ1 ([l, r − ε]) = γn (K), noting
that the optimizing f is now the indicator function of [l, r − ε]. From here, a direct computation reveals that
Z
l
r−ε
2
2
2
2
2
e−t /2
1
1
t √
= √ (e−l /2 − e−(r−ε) /2 ) < √ (e−l /2 − e−r /2 ) .
2π
2π
2π
Now since γ1 ([l, r]) > γn (K) ≥ γ1 ([−2r, 2r]), we must have that l ≤ −2r. Using the inequalities 1 + x ≤
ex ≤ 1 + x + x2 , for |x| ≤ 1/2, we have that
2
2
2
2
1
1
√ (e−l /2 − e−r /2 ) ≤ √ (e−(2r) /2 − e−r /2 )
2π
2π
1
≤ √ (1 − (2r)2 /2 + (2r)4 /4 − (1 − r2 /2))
2π
1
1
= √ ((4r2 )r2 − 3r2 /2) ≤ − √ · r2 /2 . ( since r ≤ 1/2 )
2π
2π
But by assumption − √12π · r2 /2 ≤ −γn (K)kb(K)k2 , yielding the desired contradiction.
For the furthermore, it follows by a direct numerical computation.
We will now extend the bound to asymmetric convex bodies having their barycenter near the origin. To
do this, we will need the standard fact that the gauge function of a body is Lipschitz when it contains a
large ball. We recall that a function f : Rn → R is L-Lipschitz if for x, y ∈ Rn , |f (x) − f (y)| ≤ Lkx − yk2 .
19
Lemma 28. Let K ⊆ Rn be a convex body satisfying rB2n ⊆ K for some r > 0. Then, the gauge function
k · kK : Rn → R+ of K is (1/r)-Lipschitz.
Proof. We need to show
|kxkK − kykK | ≤
1
kx − yk2
r
for all x, y ∈ Rn .
To see this, note that
kxkK
= k(x − y) + ykK
≤
kx − ykK + kykK
≤
kx − ykrB2n + kykK
1
=
kx − yk2 + kykK ,
r
(by triangle inequality)
(since rB2n ⊆ K )
yielding kxkK − kykK ≤ 1r kx − yk2 . The other inequality follows by switching x and y.
We will also need the following concentration inequality of Maurey and Pisier.
Theorem 29 (Maurey-Pisier). Let f : Rn → R be an L-Lipschitz function. Then for X ∈ Rn standard
Gaussian and t ≥ 0, we have the inequalities
2t2
Pr[f (X) − E[f (X)] ≥ tL] ≤ e− π2
and
2t2
Pr[f (X) − E[f (X)] ≤ −tL] ≤ e− π2 .
We now prove the main estimate for asymmetric convex bodies.
Lemma 30. Let K ⊆ Rn be a 0-centered convex body and X ∈ Rn be the standard n-dimensional Gaussian.
Then the following holds:
1. E[kXkK∩−K ] ≤ 2 E[kXkK ].
2. If γn (K) ≥ 1/2 and kb(K)k2 ≤
1
√
,
32 2π
√
then E[kXkK ] ≤ (1 + π 8 ln 2).
Proof. We prove part (1). By symmetry of the Gaussian measure
E[kXkK∩−K ] = E[max {kXkK , k − XkK }] ≤ E[kXkK + k − XkK ] = 2 E[kXkK ] ,
as needed.
√
We prove part (2). Let c = π 8 ln 2. First, by Lemma 27 part 2 and our assumptions on K, we have
that (1/4)B2n ⊆ K. Thus, by Lemma 28 k · kK is 4-Lipschitz. Assume for the sake of contradiction that
E[kXkK ] > 1 + c. Then, since
1/2 ≤ γn (K) = Pr[X ∈ K] = Pr[kXkK ≤ 1] ,
we must have that Pr[kXkK − E[kXkK ] ≤ −c] > 1/2. But by Theorem 29 and the Lipschitz proporties of
k · kK ,
Pr[kXkK − E[kXkK ] ≤ −c] ≤ e−
2(c/4)2
π2
= 1/2 ,
a clear contradiction.
5
√
An O( log n)-subgaussian Random Walk
√
The O( log n)-subgaussian random walk algorithm is given as Algorithm 1. Step 8 can be executed in
polynomial time by either calling an SDP solver, or executing the algorithm from Section 10. The feasibility
of the program is guaranteed by Theorem 49, and also by the results of [24]. The matrix U (t) in step 10 can
be computed by Cholesky decomposition.
Let us first make some observations that will be useful throughout the analysis. Notice first that the
random process χ(0), . . . , χ(T ) is Markovian. Let ui (t) be the i-th row of U (t). By the definition of Σ(t)
20
√
Algorithm 1 O( log n)-subgaussian Random Walk Algorithm
Pn
Input: v1 , . . . , vn ∈ Rm such that maxni=1 kvi k2 ≤ 1; a vector
, vi ].
Pn y ∈ i=1 [−vi√
Output: random signs χ1 , . . . , χn ∈ {−1, 1} such that i=1 χi vi − y is O( log n)-subgaussian.
Let V = (vi )ni=1 .
√
2
2 ln 2 log 2n
2
2
2 2n
Let γ = 2 log
;δ=
n
n5/2
P; T = d2/γ e · dlog2 2ne;
n
Let χ(0) ∈ [−1, 1] be such that i χ(0)i vi = y.
Let A(1) = {1, . . . , n}
for t = 1, . . . , T do
Compute Σ(t) ∈ Rn×n , Σ(t) 0, such that
Σ(t)ii = 1 ∀i ∈ A(t)
Σ(t)ii = 0 ∀i 6∈ A(t)
V Σ(t)V T Im ,
Pick r(t) ∈ {−1, +1}n uniformly at random.
χ(t) = χ(t − 1) + γU (t)r(t), where U (t)U (t)| = Σ(t).
A(t + 1) = {i : |χ(t)i | < 1 − δ}.
end for
Set χi = sign(χ(T )i ) for all i ∈ {1, . . . , n}.
if A(T ) = ∅ then
Return χ
else
Restart algorithm from line 5.
end if
and U (t), kui (t)k2 = Σ(t)ii equals 1 if i ∈ A(t), and 0 otherwise. We have χ(t)i − χ(t − 1)i = γhui (t), r(t)i,
and, because r(t) ∈ {−1, 1}n , by Cauchy-Schwarz we get
( √
γ n if i ∈ A(t)
|χ(t)i − χ(t − 1)i | ≤
.
(13)
0
otherwise
We first analyze the convergence of the algorithm: we show that, with constant probability, the random
walk fixes all coordinates to have absolute value between 1 − δ and 1. First we prepare a lemma.
Lemma 31. Let X0 , X1 , X2 , . . . form a martingale sequence adapted to the filtration {Ft } such that X0 ∈
[−1, 1], and for every t ≥ 1 we have E[(Xt − Xt−1 )2 | Ft−1 ] = σ 2 , and |Xt − Xt−1 | ≤ δ. Denote τ = inf{t :
1−X 2
|Xt | ≥ 1 − δ}. Then E[τ ] < σ2 0 .
Proof. Define Y0 , Y1 , Y2 , . . . to be a martingale with respect to X0 , X1 , X2 , . . . defined by Yt = Xmin{t,τ } .
Because |Xt − Xt−1 | < δ, we easily see by induction that |Yt | < 1 for all t ≥ 0. Therefore, for any t ≥ 1, we
compute
1>
E Yt2
=
=
t
X
s=1
t
X
E E[(Yt − Yt−1 )2 | Ft−1 ] + X02
σ 2 Pr[τ ≥ s] + X02
s=1
= σ 2 E[min{t, τ }] + X02 .
By the monotone convergence theorem, we have that σ 2 E[τ ] < 1 − X02 , which was to be proved.
The next lemma gives our convergence analysis of the random walk.
21
Lemma 32. With probability 1, |χ(t)i | ≤ 1 for all 1 ≤ t ≤ T and all 1 ≤ i ≤ n. With probability at least
1/2, |χ(T )i | ≥ 1 − δ for all 1 ≤ i ≤ n.
Proof. We prove the first claim by induction on t. It is clearly satisfied for t = 0; assume then that the claim
holds up to t − 1, and we will prove it for t. If i 6∈ A(t), then χ(t)i = χ(t − 1)i by (13), and the claim follows
by the inductive hypothesis.
If i ∈ A(t), then |χi (t)| < 1 − δ, and by (13) √
and the triangle inequality, we
√
have |χi (t)| < 1 − δ + γ n < 1, where the final inequality follows because γ n < δ.
To prove the second claim, we will show that Pr[|χ(T )i | < 1 − δ] ≤ 1/2n holds for every i. The claim
will then follow by a union bound. Let us fix an arbitrary i ∈ {1, . . . , n}. Define ∆ = d2/γ 2 e, and let
Ej , 0 ≤ j ≤ (T − ∆)/∆ be the event that |χ(t)i | < 1 − δ for all t ∈ [j∆ + 1, (j + 1)∆]. Observe that if
|χ(t)i | ≥ 1 − δ, then |χ(s)i | = |χ(t)i | ≥ 1 − δ for all t ≤ s ≤ T , and, therefore χ(T )i < 1 − δ if and only if all
the events E0 , . . . , E(T −∆)/∆ hold simultaneously. By this observation, and the Markov property, we have
T /∆−1
Pr[|χ(T )i | < 1 − δ] =
Y
Pr[Ej | χ(j∆)].
(14)
j=0
Let τj = (j + 1)∆ if Ej holds and τj = min{t ≥ j∆ : |χ(t)i | ≥ 1 − δ}, otherwise. The sequence
χ(j∆)i , . . . , χ(τj )i , conditioned on χ(j∆), is a martingale. Moreover, since for any t ∈ [j∆, τj ] we have
i ∈ A(t), for all such t we get
E[(χ(t)i − χ(t − 1)i )2 | χ(t − 1)] = γ 2 E |hui (t), r(t)i|2 = γ 2 kui k22 = γ 2 .
(15)
By (13) and (15), the sequence χ(j∆)i , . . . , χ(τj )i , conditioned on χ(j∆), satisfies the assumptions of
Lemma 31. By the lemma, we have
E[τj − j∆ | χ(j∆)] ≤
1
1 − χ(j∆)2i
≤ 2.
γ2
γ
Since the event Ej holds only if τj ≥ (j + 1)∆, by Markov’s inequality we have
Pr[Ej | χ(j∆)] ≤
1
γ2∆
≤
1
.
2
This bound and (14) imply that Pr[|χ(T )i | < 1 − δ] ≤ 2−T /∆ ≤ 1/2n, which was to be proved.
To prove that the walk is subgaussian, we will need the following martingale concentration inequality
due to Freedman.
Theorem 33 ([16]).
Pt Let Z1 , . . . , ZT be a martingale
Ptadapted to the filtration {Ft } |Zt − Zt−1 | ≤ M for all
t, and let Wt = j=1 Ej−1 [(Zj − Zj−1 )2 | Fj−1 ] = j=1 Var[Zj | Fj−1 ]. Then for all λ ≥ 0 and σ 2 ≥ 0, we
have
λ2
2
.
Pr[∃t s.t. |Zt − Z0 | ≥ λ and Wt ≤ σ ] ≤ 2 exp −
2(σ 2 + M λ)
Next we state the main lemma, which, together with an estimate on the error due to rounding, implies
subgaussianity.
√
Pn
Lemma 34. The random variable i=1 χ(T )i vi − y is (γ 2T )-subgaussian.
Pn
Proof. Define Yt = i=1 χ(t)i vi for all t = 1, . . . , T . Notice that Y0 = y. Let us fix a θ ∈ S n−1 once and for
2
2
all, and let Zt = hθ, Yt i for t = 0, . . . , T . We need to show that for every λ > 0, Pr[|ZT | ≥ λ] ≤ 2e−λ /2σ .
We first
Pnobserve that Zt is bounded, so we only need to consider λ in a finite range.
Pn Indeed, by Lemma 32,
Yt ∈ i=1 [−vi , vi ] with probability 1, so by the triangle inequality, kYt k2 ≤ i=1 kvi k2 ≤ n. Then, by
Cauchy-Schwarz, |Zt | ≤ n as well, and, therefore, Pr[|ZT | > n] = 0. For the rest of the proof we will assume
that 0 < λ ≤ n.
22
Observe that Z0 , . . . , ZT is a martingale. First we prove that the increments are bounded: this follows
from the boundedness of the increments of the coordinates of χ(t). Indeed, by the triangle inequality and
(13),
n
n
X
X
kYt − Yt−1 k2 =
(χ(t)i − χ(t − 1)i )vi ≤
|χ(t)i − χ(t − 1)i |kvi k2 ≤ γn3/2 .
i=1
2
i=1
Then, it follows from Cauchy-Schwarz that |Zt − Zt−1 | ≤ γn3/2 .
Next we bound the variance of the increments. By the Markov property of the random walk, Zt − Zt−1
is entirely determined by χt−1 . Denoting V = (vi )ni=1 as in the description of Algorithm 1, we have
E[(Zt − Zt−1 )2 | χt−1 ] = θ| E[(Yt − Yt−1 )(Yt − Yt−1 )| ]θ
= θ| V E[(χ(t) − χ(t − 1))(χ(t) − χ(t − 1))| ]V | θ
= γ 2 θ| V U (t) E[r(t)r(t)| ]U (t)| V | θ
= γ 2 θ| V Σ(t)V | θ
≤ γ2.
The penultimate equality follows because E[r(t)r(t)| ] = In and U (t)U (t)| = Σ(t), and the final inequality
follows because Σ(t) was chosen so that V Σ(t)V | Im .
We are now ready to apply Theorem 33. Using the notation from the theorem, we have shown that
M ≤ γn3/2 , and that Wt ≤ γ 2 t for all t, and both bounds hold with probability 1. Let σ 2 = γ 2 T . First we
claim that for any λ ≤ n, M λ ≤ σ 2 . Indeed,
M λ ≤ γn5/2 ≤ 2 log2 (2n) ≤ γ 2 T = σ 2 .
2
Now, Theorem 33 and the above calculation imply that Pr[|ZT − Z0 | ≥ λ] ≤ 2e−λ
This proves the lemma.
/4σ 2
for all 0 < λ ≤ n.
Finally we state our main theorem.
Theorem 35 (Restatement of Theorem 4). Algorithm
1 runs in expected
polynomial time, and outputs a
√
Pn
random vector χ such that the random variable i=1 χi vi − y is O( log n)-subgaussian.
Proof. Let E be the event that for all i, |χ(T )i | ≥ 1 − δ (equivalently, that A(T ) = ∅). The algorithm takes
returns if E holds, and otherwise it restarts. By Lemma 32 this event occurs with probability at least 1/2,
so there will be a constant number of restarts in expectation. Since the random walk talks T steps, where
T is polynomial in the input size, and each step can also be executed in polynomial time, it follows that the
expected running time of the algorithm is polynomial.
Because the algorithm returns an output exactly when E holds, the output is distributed as the random
vector
χ conditioned on E. Let us fix a vector θ ∈ S n−1 once and for all. Let Y be the random variable
√
Pn
i=1 χi vi − y and let Z = hθ, Y i. Let Yt and Zt be defined as in the proof of Lemma 34. Let s = γ 2T
be the parameter with which we proved ZT − Z0 is subgaussian in Lemma 34. We will show that Z − Z0 ,
2
2
conditioned on E, is 2s-subgaussian, i.e. we will √
prove that Pr[|Z − Z0 | ≥ λ | E] ≤ 2e−λ /8s . Observe that
this inequality is trivially satisfied for λ ≤ λ0 = 2 2 ln 2s, since the right hand side is at least 1 in this range.
For the rest of the proof we will assume that λ > λ0 .
Conditional on E, and using the triangle inequality, we can bound the distance between Y and Y (T ) by
kY − Y (T )k2 =
≤
n
X
(sign(χ(T )i ) − χ(T )i )vi
i=1
n
X
2
(1 − |χ(T )i |)kvi k2 ≤ δn.
i=1
By Cauchy-Schwarz, we get that, conditional on E, |Z−ZT | ≤ δn, which implies that |Z−Z0 | ≤ |ZT −Z0 |+δn,
by the triangle inequality. Then, we have that, conditional on E, |Z − Z0 | ≥ λ ⇒ |ZT − Z0 | ≥ λ − δn. For
23
p
every λ > λ0 , δn ≤ 2 2 ln 2 log2 2n ≤ λ0 /2 < λ/2. Therefore, conditional on E and for every λ > λ0 ,
|Z − Z0 | ≥ λ ⇒ |ZT − Z0 | ≥ λ/2. By Lemma 34 we have
2
Pr[|Z − Z0 | ≥ λ | E] ≤ Pr[|ZT − Z0 | ≥ λ/2 | E] ≤ 2 Pr[|ZT − Z0 | ≥ λ/2] ≤ 4e−λ
2
2
2
/4s2
.
2
Recall that for every λ > λ0 , e−λ /8s < 1/2, so the right hand side above √
is at most 2e−λ /8s , as claimed.
Therefore, Z − Z0 , conditioned on E, is 2s-subgaussian, and, since s = O( log n), this suffices to prove the
theorem.
6
Recentering procedure
In this section we will give an algorithmic variant of the recentering procedure in Theorem 6.
Given a convex body K ⊆ Rn , let b be its barycenter under the Gaussian distribution. The following
lemma shows that if we have an estimate b0 of the barycenter, which is close to b but farther from the origin,
then shifting K to K − b0 , increases the Gaussian volume of K.
Lemma 36. Let b be the barycenter of K ⊆ Rn and b0 a point in Rn satisfying kb −b0 k2 ≤ δ/3 and kb0 k2 ≥ δ.
2
Then, γn (K − b0 ) ≥ eδ /6 γn (K)
Proof. Let η = b − b0 .
γn (K − b0 )
γn (K)
=
=
1 n/2
( 2π
)
R
2
x∈K−b0
e−kxk2 /2 dx
2
1 n/2
( 2π
)
e−kxk2 /2 dx
x∈K
R
2
0 2
0
1 n/2
( 2π
)
e−kyk2 /2−kb k2 /2+hy,b i dy
y∈K
R
2
1 n/2
( 2π
)
e−kxk2 /2 dx
x∈K
R
(change of variables y = x + b0 )
Let Y be a random variable drawn from the n-dimensional Gaussian distribution restricted to the body K.
Then the right hand side above is equal to
0 2
0
e−kb k2 /2 E[ehY,b i ] ≥
0 2
= e
6.1
0
e−kb k2 /2 ehE[Y ],b i
−kb0 k22 /2+hb,b0 i
2
(by Jensen’s inequality)
−kb0 k22 /2+hb0 ,b0 i+hη,b0 i
=e
0 2
0
= ekb k2 /2+hη,b i
2
=
ekbk2 /2−kηk2 /2
≥
e(2δ/3)
2
/2−(δ/3)2 /2
= eδ
2
/6
(kbk2 ≥ kb0 k2 − kηk2 ≥ 2δ/3)
Algorithm
In our recentering P
algorithm we use the geometric language of section 1.1.4. Instead
Pnof the vectors v1 , . . . , vn
n
and the shift t ∈ i=1 [−vi , vi ], we work directly with the parallelepiped P = i=1 [−vi , vi ] − t. Notice
that a facet of P corresponds to a fractional coloring with some coordinates fixed.
P Indeed, a facet
P F of P is
determined by a subset S ⊆ [n], and a coloring χ ∈ {−1, 1}S , and equals F = i6∈S [−vi , vi ] + i∈S χi vi − t.
The size of the set S is equal to the co-dimension of F , so a vertex (face of dimension 0) is equivalent to
full coloring χ ∈ {−1, 1}n . The edges (faces of dimension 1) are linear segments that have length exactly
twice the length of the corresponding vectors. We say that P has side lengths at most ` if each edge of P
has length at most `: this corresponds to requiring that maxi kvi k2 ≤ `/2. Given a point p ∈ P , we denote
by FP (p) the face of P that contains p and has minimal dimension. We denote by WP (p) the subspace
span(FP (p) − p)
In this language, the (linear) discrepancy problem is translated to the problem of finding a vertex of P
inside K. The recentering problem can also be expressed in this way: we are looking for a point p ∈ P ∩K such
that the Gaussian measure of (K−p)∩WP (p), restricted to WP (p), is at least that of K, and b((K−p)∩WP (p))
is close to 0. To do this, we start out by approximating b = b(K), the barycenter of K. If b is close to the
origin, then we are already done and can return. If b is far from origin, then moving the origin to b (i.e.
24
shifting K and P to K − b, P − b respectively), should only help us by increasing the Gaussian volume of K.
But we cannot make this move if b lies outside P . In this case, we start moving towards b; when we hit ∂P ,
the boundary of P , we stop and induct on the facet we land on, choosing the point on boundary of P we
stopped on as our new origin. We show that even this partial move towards b does not decrease the volume
of K. Moreover, it ensures that the origin always stays inside P .
One difficulty is that we cannot efficiently compute the barycenter of K exactly. To get around this,
we use random sampling from Gaussian distribution restricted to K to estimate the barycenter with high
accuracy. We will then return a shift of the body K such that its barycenter is δ-close to the origin, where
the running time is polynomial in n and (1/δ) and it suffices to choose δ as inversely polynomial in n. We
assume that we have access to a membership oracle for the convex body K.
Algorithm 2 Recentering procedure
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
Input: Convex body K ⊆ Rn with γn (K) ≥ 1/2, an n-dimensional parallelepiped P 3 0, δ ≥ 0 and
error probability ε ∈ (0, 1).
Output: See statement of Theorem 37.
If 0 6∈ P ∩ K, return FAIL.
Set N = d24/δ 2 e + n.
Set q = 0, W = WP (0), K̄ = K ∩ W , P̄ = FP (0).
for i = 1, . . . , N do
Compute an estimate b0 of the barycenter b of K̄ restricted to the subspace W ,
satisfying kb − b0 k2 ≤ δ/6 with probability at least 1 − ε/N .
If b0 6∈ K̄, return FAIL, otherwise continue.
if kb0 k2 ≤ δ/2 then Return q.
else if kb0 k2 > δ/2 and b0 6∈ P̄ then
Compute λ ∈ (0, 1) be such that λb0 ∈ ∂ P̄ relative to W .
Set s = λb0 .
else
Set s = b0 .
end if
Set q = q + s.
Set W = WP̄ (s), P̄ = FP̄ (s) − s, K̄ = (K̄ − s) ∩ W .
end for
Return FAIL.
The following theorem is an algorithmic version of Theorem 6. We note that the guarantees of the
algorithm are relatively robust. This is to make it simpler to use within other algorithms, since it may be
called on invalid inputs as well as output incorrectly with small probability.
Theorem 37. Let P be a parallelepiped in Rn containing the origin and K ⊆ Rn be a convex body of Gaussian
measure at least 1/2, given by a membership oracle, and let δ ≥ 0 and ε ∈ (0, 1). Then, Algorithm 2 on these
inputs either returns FAIL or a point p ∈ P ∩ K. Furthermore, if the input is correct, then with probability
at least 1 − ε, it returns p satisfying
1. The Gaussian measure of (K − p) ∩ WP (p) on WP (p), is at least that of K;
2. kb((K − p) ∩ WP (p))k2 ≤ δ.
Moreover, Algorithm 2 runs in time polynomial in n, 1/δ and ln(1/ε).
Proof. Firstly, it easy to check by induction, that at the beginning of each iteration of the for loop that
q ∈ P ∩ K,
W = WP (q),
K̄ = (K − q) ∩ WP (q),
P̄ = FP (q) − q .
(16)
To prove correctness of the algorithm, we must show that the algorithm returns a point q satisfying the
conditions of Theorem 37 with probability at least 1 − ε.
25
For this purpose, we shall condition on the event that all the barycenter estimates computed on line 7 are
within distance δ/6 of the true barycenters, which we denote by E. Since we run the barycenter estimator
at most N times, by the union bound, E occurs with probability at least 1 − ε. We defer the discussion of
how to implement the barycenter estimator till the end of the analysis.
With this conditioning, we prove a lower bound on the Gaussian mass as a function of the number of
iterations, which will be crucial for establishing the correctness of the algorithm.
Claim 38. Let W, K̄, P̄ denote the state after t ≥ 0 non-terminating iterations. Let kt ≥ 0 denote number
of iterations before time t, where the dimension of W decreases. Then, conditioned on E, we have that
γW (K̄) ≥ e(t−kt )δ
2
/24
γn (K) .
Proof. We prove the claim by induction on t. At the base case t = 0 (i.e. at the beginning of the first
iteration), note that kt = 0 by definition. If W = Rn , the inequality clearly holds since K̄ = K. If W ⊂ Rn ,
then since γn (K) ≥ 1/2 by Lemma 14, we have γW (K̄) ≥ γn (K). The base case holds thus holds.
We now assume that the bound holds at time t and prove it for t + 1, assuming that iteration t + 1 is
non-terminating. Let b,b0 ,s denote the corresponding loop variables, and W 0 , K̄ 0 , P̄ 0 denote the new values
of W, K̄, P̄ after line 16.
Since the iteration is non-terminating, we have that kb0 k2 > δ/2. Since by our conditioning kb0 −bk2 ≤ δ/6,
by Lemma 36 and the induction hypothesis, we have that
γW (K̄ − b0 ) ≥ eδ
2
/24
γW (K̄) ≥ e(t+1−kt )δ
2
/24
γn (K) .
(17)
Note that we drop in dimension going from W to W 0 if and only if s lies on the boundary of P̄ relative to
W (since then the minimal face of P̄ containing s is lower dimensional).
We now examine two cases. In the first case, we assume b0 is in the relative interior of P̄ . In this case,
we have s = b0 , and hence W = W 0 and K̄ 0 = K̄ − b0 . Given this, kt+1 = kt (no drop in dimension) and the
desired bound is derived directly from Equation (17).
In the second case, we assume that b0 is not in the interior of P̄ relative to W . In this case, s = λb0 ∈ ∂ P̄
relative to W , for some λ ∈ [0, 1]. Furthermore, W 0 ⊂ W and kt+1 = kt + 1. From here, by Ehrhard’s
inequality and Equation 17, we get that
γW (K̄ − s) ≥ Φ((1 − λ)Φ−1 (γW (K̄)) + λΦ−1 (γW (K̄ − b0 ))) ≥ γW (K̄)
≥ e(t−kt )δ
2
/24
γn (K) = e(t+1−kt+1 )δ
2
/24
γn (K) ≥ 1/2 .
(18)
Lastly, by Lemma 14 since γW (K̄ − s) ≥ 1/2, we have that
γW 0 (K̄ 0 ) = γW 0 (K̄ − s) ≥ γW (K̄ − s) .
(19)
The desired bound now follows combining Equations (18),(19).
We now prove correctness of the algorithm conditioned on E. We first show that conditioned on E, the
algorithm returns q from line 8 during some iteration of the for loop. For the sake of contradiction, assume
instead that the algorithm returns FAIL. Let W, K̄, P̄ denote the state after the end of the loop. Then, by
Claim 38, we have that
γW (K̄) ≥ e(N −kN )δ
2
/24
γn (K) ≥ e(N −n)δ
2
/24
γn (K) ≥ eγn (K) > 1,
where we used that fact kN ≤ n, since dimension cannot drop more than n times. This is clear contradiction
however, since Gaussian measure is always at most 1.
Given the above, we can assume that the algorithm returns q during some iteration of the for loop. Let
W, K̄, P̄ , b0 denote the state at this iteration. Since we return at this iteration, we must have that kb0 k2 ≤ δ/2.
Given E, we have that the barycenter b of K̄ satisfies
kbk2 ≤ kb0 k2 + kb − b0 k2 ≤ δ/2 + δ/6 < δ .
By Claim 38, we also know that γW (K̄) ≥ γn (K). Since by Equation 16, q ∈ P and K̄ = (K − q) ∩ WP (q),
the correctness of the algorithm follows.
26
For the runtime, we note that it is dominated by the N = O(1/δ 2 + n) calls to the barycenter estimator.
Thus, as long as the estimator runs in poly(n, ln(1/(εδ))) time, the desired runtime bound holds.
It remains to show that we can estimate the barycenter efficiently. We show how to do this in appendix
in Theorem 54 with failure probability at most ε/N in time poly(n, 1/δ, ln(N/ε)) = poly(n, 1/δ, ln(1/ε)), as
needed.
7
Algorithmic Reduction from Asymmetric to Symmetric Banaszczyk
In this section, we make algorithmic the reduction in section 4.1 from the asymmetric to the symmetric
case. This will directly imply that given an algorithm to return a vertex of P contained in a symmetric
convex body K of Gaussian volume at least a half, we can also efficiently find a vertex of P contained in an
asymmetric convex body of Gaussian measure at least a half.
Definition 39 (Symmetric Body Coloring Algorithm). We shall say that A is a symmetric body coloring
algorithm, if given as input an n-dimensional parallelepiped P 3 0 of side lengths at most lA (n), lA a nonnegative non-increasing function of n, and a symmetric convex body K ⊆ Rn satisfying γn (K) ≥ 1/2, given
by a membership oracle, it returns a vertex of P contained in K with probability at least 1/2.
√
Let α = 4(1 + π 8 ln 2). We now present an algorithm, which uses A as a black box and achieves the
same guarantee for asymmetric convex bodies, with only a constant factor loss in the length of the vectors.
Algorithm 3 Reducing asymmetric convex bodies to symmetric convex bodies
1:
2:
3:
4:
5:
Input: Algorithm A as in (39), K ⊆ Rn convex body, given by membership oracle, with γn (K) ≥ 1/2,
P 3 0 an n-dimensional parallelepiped of side lengths at most lA (n)/α.
Output: A vertex v of P contained in K.
Call Recentering Procedure on K and P and δ = 32√1 2π and ε = 1/4.
Restart from line 3 if the call outputs FAIL, and otherwise let q be the output.
Call A on α(FP (q) − q) and α(K − q) ∩ (q − K) ∩ WP (q) inside WP (q).
Let v be the output.
If v/α + q ∈ K and is a vertex of P , return v/α + q. Else, restart from line 3.
Theorem 40. Algorithm 3 is correct and runs in expected polynomial time.
Proof. Clearly, by line 5 correctness is trivial, so we need only argue that it runs in expected polynomial time.
Since the runtime of the recentering procedure (Algorithm 2) is polynomial, and the runs are independent,
we need only argue that line 5 accepts with constant probability. Since the recentering procedure outputs
correctly with probability at least 1 − ε = 3/4, we may condition on the correctness of the output q in line 3.
Under this conditioning, by the guarantees of the recentering algorithm, letting d = dim(FP (q)), W =
WP (q) and C = (K − q) ∩ W , we have that
γW (C) ≥ 1/2
kb(C)k2 ≤
and
1
√
.
32 2π
Thus by Lemma 30, for X ∈ W the d-dimensional standard Gaussian on W , we have that
√
E[kXkC∩−C ] ≤ 2 E[kXkC ] ≤ 2(1 + π 8 ln 2) .
√
Hence by Markov’s inequality, Pr[X ∈ α(C ∩ −C)] = Pr[X ∈ 4(1 + π 8 ln 2)(C ∩ −C))] ≥ 1/2.
Now by construction αP has side lengths at most lA (n), and hence α(FP (q) − q) also has side lengths
at most lA (n) ≤ lA (d). Thus, A on input α(FP (q) − q) and α(C ∩ −C), outputs a vertex v of α(FP (q) − q)
contained in α(C ∩ −C) ⊆ α(K − p) ∩ W with probability at least 1/2. Hence, the check in line 5 succeeds
with constant probability, as needed.
The above directly implies Theorem 8, as shown below.
27
Proof of√Theorem 8. Let P ⊆ Rn be an n-dimensional parallelepiped P
containing the origin of side lengths at
n
c
,
and
t
∈
most c/ log n. Let v1 , . . . , vn ∈ Rn , maxi∈[n] kvi k ≤ √log
i=1 [−vi , vi ], denote the vectors such
n
Pn
that P = i=1 [−vi , vi ] − t.
On input v1 , . . . , vn and t, theP
random walk sampler (Algorithm 1) outputs in expected polynomial time
n
n
a random χ ∈ {−1, 1} such that i=1 χi vi −t is O(c)-subgaussian and supported on the vertices of P . Thus
n
by Lemma 19 part 1, we can pick
Pn c > 0 small enough such that for any symmetric convex body K ⊆ R
with γn (K) ≥ 1/2, we have Pr[ i=1 χi vi − t ∈ K] ≥ 1/2.
Thus, letting A denote the above sampler, we see that A satisfies the conditions 39 with lA (n) =
√
2c/ log n. The theorem now follows by combining Algorithm 3 with A.
8
Body Centric Algorithm for Asymmetric Convex Bodies
In this section, we give the algorithmic implementation of the extended recentering procedure, which returns
full colorings matching the guarantees of Theorem 8. Interestingly, the coloring output by the procedure
will be essentially deterministic. The only randomness will be in effect due to the random errors incurred in
estimating barycenters.
For a convex body K ⊆ Rn , unit vector θ ∈ Rn ,kθk2 = 1, and v ∈ R, we define the shifted slice
θ
Kv = (K − vθ) ∩ {x ∈ Rn : hθ, xi = 0}. The main technical estimate we will require in this section, is the
following lower bound on the gaussian measure of shifted slices. We defer the proof of this estimate to
section 9.
Theorem 41. There exists universal constants v0 , η0 , c0 > 0, such that for any n ≥ 1, convex body K ⊆ Rn
satisfying kb(K)k2 = η ≤ η0 and γn (K) = α ≥ 3/5, v ∈ [−v0 , v0 ] and θ ∈ Rn , kθk2 = 1, we have that
1
γn−1 (Kvθ ) ≥ (α − c0 η)(1 −
e− 100v2
√
).
4 2π
The above inequality says that if barycenter of K is close to the origin, then the Gaussian measure of
parallel slices of K does not fall off too quickly as we move away from the origin.
Recall that the problem can be recast as P
finding a vertex of a parallelepiped
P contained inside the
Pn
n
convex body K, where the parallelepiped P = i=1 [−vi , vi ] − t and t ∈ i=1 [−vi , vi ]. Thus, 0 ∈ P .
We start out by calling the recentering procedure to get the barycenter, b, close to the origin. This
recentering allows us to rescale K by a constant factor
√ such that√the Gaussian volume of K increases i.e. we
replace P by βP and K by βK where β = 1 + π 8 log 2 + 4π log 2 is chosen such that the volume of K
after rescaling is at least 3/4. Then we find a point q ∗ on ∂P , the boundary of P which is closest to the
origin. We recurse by taking a (n − 1)-dimensional slice of K(here we abuse notation by calling the convex
body after rescaling as also K) with the facet containing q ∗ . A crucial point here is that we choose q ∗ as the
origin of the (n − 1)-dimensional space we use in the induction step. This is done to maintain the induction
hypothesis that the parallelepiped contains the origin. Theorem 41 guarantees that in doing so, we do not
lose too much Gaussian volume.
Lemma 42. Given a convex body K in Rn such that γn (K) ≥ 1/2 and kb(K)k2 ≤
√
√
3/4, where β = 1 + π 8 log 2 + 4π log 2.
1
√
,
32 2π
then γn (βK) ≥
√
Proof. Let X be the standard n-dimensional Gaussian. From Lemma 30, E[kXkK ] ≤ 1 + π 8 log 2. This
gives
Pr[kXkK > β]
Pr[kXkK − E[kXkK ] > β − E[kXkK ]]
p
≤ Pr[kXkK − E[kXkK ] > β − 1 − π 8 log 2]
=
By Lemma 27 and Lemma 28, the function k.kK is 4−Lipschitz. Then, by Theorem 29
2
Pr[kXkK > β] ≤ e− π2 (
√
β−1−π 8 log 2 2
)
4
= 1/4 .
Thus, γn (βK) = Pr[X ∈ βK] = 1 − Pr[kXkK > β] ≥ 1 − 1/4 = 3/4, as needed.
28
Algorithm 4 Body centric algorithm for general convex bodies
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
Input: Convex body K ⊆ Rn , given by a membership oracle, with γn (K) ≥ 1/2, an n-dimensional
parallelepiped P 3 0 of side lengths at most 2αn .
Output: A vertex of P contained in K.
1
.
Call Recentering Procedure on K and P with parameters δ = ηn and ε = 2(n+1)
Restart from line 3 if the call outputs FAIL, and otherwise let q denote the output.
Set q̄ = 0, W̄ = WP (q), K̄ = β((K − q) ∩ W̄ ), P̄ = β(FP (q) − q).
repeat
1
.
Call Recentering Procedure on P̄ and K̄ with δ = ηn and ε = 2(n+1)
Restart from line 3 if the call outputs FAIL, and otherwise let s denote the output.
Set q̄ = q̄ + s, W̄ = WP̄ (s), K̄ = (K̄ − s) ∩ W̄ , P̄ = (FP̄ (s) − s).
if dim(W̄ ) 6= 0 then
Compute s ∈ argmin kpk2 : p ∈ ∂ P̄ relative to W̄ .
Set q̄ = q̄ + s, W̄ = WP̄ (s), K̄ = (K̄ − s) ∩ W̄ , P̄ = (FP̄ (s) − s).
end if
until dim(W̄ ) = 0
If q + q̄/β ∈ K and is a vertex of P , return q + q̄/β. Else, restart from line 3.
For Algorithm 4, we use the parameters
(
)
1
αn = min v0 , p
,
10 log(2n)
ηn = min η0 ,
1
32 2π 14c0 n
1
√
,
,
where v0 , η0 , c0 are as in Theorem 41. We now give the formal analysis of the algorithm.
We begin by explaining how to compute the minimum norm point on the boundary of a parallelepiped.
Pk
n
Lemma 43. Let P =
i=1 [−vi , vi ] − t ⊂ R be a parallelepiped of side lengths at most 2α containing
the origin, with v1 , . . . , vk linearly independent. Let v1∗ , . . . , vk∗ denote the dual basis, i.e. the unique set of
vectors lying inside W := span(v1 , . . . , vk ) satisfying
hvi∗ , vj i = 1 if i o
= j and 0 otherwise, and let s denote
n
an element of minimum `2 norm in the set
±1+hvi∗ ,ti
kvi∗ k2
·
vi∗
kvi∗ k2
: i ∈ [k] . Then, the following holds:
1. s ∈ argmin {kpk : p ∈ ∂P relative to W }.
2. WP (s) ⊆ {x ∈ W : hs, xi = 0}.
3. ksk2 ≤ α.
Furthermore, s can be computed in polynomial time.
Pk
Pk
Proof. Note that for any x ∈ W , we have that x = i=1 hvi∗ , xivi . From here, given that P = i=1 [−vi , vi ]−
t, it is easy to check that
P = {x ∈ W : −1 + hvi∗ , ti ≤ hvi∗ , xi ≤ 1 + hvi∗ , ti, ∀i ∈ [k]} .
(20)
We now show that s ∈ argmin {kpk2 : p ∈ ∂P relative to W }. Since 0 ∈ P , we must show that x ∈ P
if kxk2 ≤ ksk2 and x ∈ W , and that s ∈ ∂P relative to W . Given that the vectors vi∗ /kvi∗ k2 have unit `2
norm, the norm of s is equal to
| ± 1 + hvi∗ , ti|
ω := min
:
i
∈
[k]
.
kvi∗ k2
Now assume x ∈ W and kxk2 ≤ ω. Since by assumption 0 ∈ P , we must have −1 + hvi∗ , ti ≤ 0 ≤ 1 + hvi∗ , ti,
∀i ∈ [k]. Therefore, for i ∈ [k], by Cauchy-Schwarz
hvi∗ , xi ≤ kvi∗ k2 · ω ≤ 1 + hvi∗ , ti,
hvi∗ , xi ≥ −kvi∗ k2 · ω ≥ −1 + hvi∗ , ti .
29
Hence x ∈ P , as needed. Next, we must show that s ∈ ∂P relative to W . Firstly, clearly s ∈ W since
each v1∗ , . . . , vk∗ ∈ W , and thus by the above argument s ∈ P . Now choose i ∈ [k], r ∈ {−1, 1} such that
v∗
r+hv ∗ ,ti
s = kv∗ ik2 · kv∗ik2 . Then, by a direct calculation hvi∗ , si = r + hvi∗ , ti, and hence s satisfies one of the
i
i
inequalities of P (see Equation 20) at equality. Thus, s ∈ ∂P relative to W (note that P is full-dimensional
in W ), as needed.
We now show that WP (s) ⊆ {x ∈ W : hs, xi = 0}. By the above paragraph, every element x of the
minimal face FP (s) of P containing s satisfies hvi∗ , xi = r + hvi∗ , ti = hvi∗ , si. In particular hvi∗ , x − si = 0.
Since s is collinear with vi∗ (s may be 0), we have hvi∗ , x − si = 0 ⇒ hs, x − si = 0. The claim now follows
since WP (s) is the span of FP (s) − s and FP (s) − s ⊆ {x ∈ W : hs, xi = 0} by the previous statement.
We now show that ksk2 ≤ α. Firstly, by minimality of s, note that |r + hvi∗ , ti| ≤ 1, for r and i as above.
Thus, by Cauchy-Schwarz,
ksk =
1
hvi , vi∗ i
kvi k2 kvi∗ k2
r + hvi∗ , ti
≤
=
≤
= kvi k .
kvi∗ k2
kvi∗ k
kvi∗ k2
kvi∗ k2
Since P has side lengths at most 2α, we have kvi k ≤ α. Thus, ksk2 ≤ α, as claimed.
We now prove the furthermore. Let V denote the matrix whose columns are v1 , . . . , vk . By linear
independence of v1 , . . . , vk , the matrix V T V is invertible. Since then (V (V T V )−1 )T V = Ik , we see that
v1∗ , . . . , vk∗ are the columns of V (V T V )−1 (note that these lie in W by construction), and hence can be
constructed in polynomial time. Since s can clearly be constructed in polynomial time from the dual basis
and t, the claim is proven.
Theorem 44. Algorithm 4 is correct and runs in expected polynomial time.
Proof. Clearly, by the check on line 13, correctness is trivial. So we need only show that the algorithm
terminates in expected polynomial time. In particular, it suffices to show the probability that a run of the
algorithm terminates without a restart is at least 1/2.
For this purpose, we will show that the algorithm terminates correctly conditioned on the event that
each call to the recentering procedure terminates correctly, which we denote E. Later, we will show that this
event occurs with probability at least 1/2, which will finish the proof.
Let W1 , K1 , P1 denote the values W̄ , K̄, P̄ after line 4.
During the iterations of the repeat loop, it is easy to check by induction that after the execution of either
line 7 or 10, the variables q̄, W̄ , K̄, P̄ satisfy:
q̄ ∈ P1 ,
W̄ = WP1 (q̄),
K̄ = (K1 − q̄) ∩ WP1 (q̄),
P̄ = FP1 (q̄) − q .
(21)
We now establish the main invariant of the loop, which will be crucial in establishing correctness conditioned on E:
Claim 45. Let W̄ , K̄, P̄ denote the state after k ≥ 0 successful iterations of the repeat loop. Then, the
following holds:
1. dim W ≤ n − k.
2. Conditioned on E, γW̄ (K̄) ≥ 3/4 −
k
7n
> 3/5.
Proof. We prove the claim by induction on k.
For k = 0, the state corresponds to W1 , K1 and P1 . Trivially, dim(W1 ) ≤ n = n − k, so the first
condition holds. Conditioned on E, we have that K1 /β has Gaussian mass at least 1/2 restricted to W1 and
its barycenter has `2 norm at most ηn . Since ηn ≤ 32√1 2π by Lemma 42, we have that γW1 (K1 ) ≥ 3/4. Thus,
the second condition holds as well.
We now assume the statement holds after k iterations, and show it holds after iteration k + 1, assuming
that we don’t terminate after iteration k and that we successfully complete iteration k + 1. Here, we denote
the state at the beginning of iteration k + 1 by W̄ , K̄, P̄ , after line 7 by W̄1 , K̄1 , P̄1 and at the end the
iteration by W̄2 , K̄2 , P̄2 .
30
We first verify that W̄2 ≤ n − (k + 1). By the induction hypothesis n − k ≥ dim(W̄ ) and by construction
W̄2 ⊆ W̄1 ⊆ W̄ . Thus, we need only show that dim(W̄2 ) < dim(W̄ ). Given that we successfully complete
the iteration, namely the call to the recentering algorithm on line 6 doesn’t return FAIL, we may distinguish
two cases. Firstly, if dim(W̄2 ) = 0, then we must have dim(W̄2 ) < dim(W̄ ), since otherwise dim(W̄ ) = 0
and the loop would have exited after the previous iteration. Second if dim(W̄2 ) > 0, we must have entered
the if statement on line 8 since dim(W̄2 ) ≤ dim(W̄1 ). From here, we see that dim(W̄2 ) corresponds to the
dimension of the minimal face of P̄1 containing s. Since s is on the boundary of P̄1 relative to W̄1 , we get
that dim(W̄2 ) < dim(W̄1 ) ≤ dim(W̄ ), as needed. Thus, condition 1 holds at the end of the iteration as
claimed.
We now show that conditioned E, γW̄2 (K̄2 ) ≥ 3/4 − (k + 1)/(7n). By the induction hypothesis, recall that
γW̄ (K̄) ≥ 3/4−k/(7n), thus it suffices to prove that γW̄2 (K̄2 ) ≥ γW̄ (K̄)−1/(7n). Note that since we decrease
dimension at every iteration (as argued in the previous paragraph), the number of iterations of the loop can
never exceed n. Thus, after any valid number of iterations l, we always have 3/4 − l/(7n) ≥ 3/4 − 1/7 > 3/5.
In particular, we have γW̄ (K̄) ≥ 3/5.
We now track the change in Gaussian mass going from K̄ to K̄2 . Since the recentering procedure on line
6 terminates correctly by our conditioning on E, we get that
γW̄1 (K̄1 ) ≥ γW̄ (K̄)
and
kb(K̄1 )k2 ≤ ηn .
If dim(W̄1 ) = 0, then clearly W̄2 = W̄1 and K̄2 = K̄1 , and hence γW̄2 (K̄2 ) ≥ γW̄ (K̄) as needed. If
dim(W̄1 ) > 0, we enter the if statement at line 8. Since P̄1 is a parallelepiped containing 0 of side length
2αn , by Lemma 43 we have that ksk ≤ αn and W̄2 = WP̄1 (s) ⊆ Hs where Hs := x ∈ W̄1 : hs, xi = 0 . Now
if s = 0, then K̄2 = K̄1 ∩ W̄2 , and thus by Lemma 14, we have γW̄2 (K̄2 ) ≥ γW̄1 (K̄1 ) ≥ γW̄ (K̄), as needed. If
s 6= 0, given that ksk ≤ αn ≤ v0 , kb(K̄1 )k2 ≤ ηn ≤ η0 and γW̄1 (K̄1 ) ≥ 3/5, by applying Theorem 41 on K̄1
with v = ksk2 and θ = s/v, we get that
−
1
e 100α2n
)
γHs ((K̄1 − s) ∩ Hs ) ≥ (γW̄1 (K̄1 ) − c0 ηn )(1 − √
4 2π
−
1
(22)
1
e 100α2n
≥ γW̄ (K̄) −
≥ γW̄ (K̄) − c0 ηn − √
.
7n
4 2π
Since K̄2 = (K̄1 − s) ∩ W̄2 , W̄2 ⊆ Hs and γHs ((K̄1 − s) ∩ Hs ) ≥ 3/4 − (k + 1)/(7n) > 1/2, by Lemma 14
γW̄2 (K̄2 ) ≥ γHs ((K̄1 − s) ∩ Hs ).
(23)
The desired estimate follows combining (22) and (23).
By Claim 45, we see that the number of iterations of the repeat loop is always bounded by n. Furthermore,
conditioned on E, the loop successfully terminates with W̄ ,K̄,P̄ satisfying γW̄ (K̄) > 0 and dim(W̄ ) = 0.
Since dim(W̄ ) = 0, this implies that W̄ = K̄ = {0}. Furthermore, by Equation 21, this implies that
q̄ ∈ K1 ∩ P1 and dim(WP1 (q̄)) = 0, and hence q̄ is a vertex of P1 . Since K1 = β(K − q) and P1 = β(P − q),
we get that q + q̄/β is a vertex of P contained in K, as needed. Thus, conditioned on E, the algorithm
returns correctly.
To lower bound E, by the above analysis, note that we never call the recentering procedure more than
n + 1 times, i.e. once on line 3 and at most n times on line 6. By the union bound, the probability that one
of these calls fails is at most (n + 1) · 1/(2(n + 1)) = 1/2. Thus, E occurs with probability at least 1/2, as
needed.
9
An estimate on the Gaussian measure of slices
In this section, we prove Theorem 41. We will need the following estimate on Gaussian tails [1, Formula 7.1.13].
31
Lemma 46 (Gaussian Tailbounds). Let X ∼ N (0, 1). Then for any t ≥ 0,
r
r
2
2
2 e−t /2
2
e−t /2
√
p
≤ Pr[X ≥ t] ≤
.
π t + t2 + 4
π t + t2 + 8/π
Before proving Theorem 41, we first prove a similar result for a special class of convex bodies in R2 .
We define a convex body K in R2 to be downwards closed if (x, y) ∈ K implies (x, y 0 ) ∈ K for all y 0 ≤ y.
For notational convenience, we shall denote the first and second coordinate of a vector in R2 respectively as
the x and y coordinates. We shall say the slice of K at x = t or y = t to denote either the vertical slice of
K having x-coordinate t or horizontal slice having y-coordinate t. We define the height of K at x = t to be
maximum y-coordinate of any point (t, y) ∈ K. By convention, we let the height of K at x = t be −∞ if K
does not a contain a point with x-coordinate t.
Lemma 47. Let K ⊆ R2 be a downwards closed convex body with γ2 (K) = α ≥ 1/2 and barycenter b = b(K)
satisfying b1 ≥ 0, and let g = Φ−1 (α) ≥ 0. Then, there exists auniversal constant v0> 0 such that for all
0 ≤ v ≤ v0 , the height of K at x = v is least f (v, g) := g − min e
g2
2
1
− 8ev
2
2
, (4e + 2) vg
.
Proof.
Step 1: Reduction to a wedge We first show that the worst-case bodies for the lemma are “wedgeshaped” (see the illustration in Figure 1). Namely, the worst case down closed convex bodies are of form
(x, y) ∈ R2 : x ≥ −c, sx + ty ≤ d
where d, s, t ≥ 0, s2 + t2 = 1, c ∈ R≥0 ∪ {∞} .
(24)
More precisely, we will show that given any body K satisfying the conditions of the theorem, there exists a
wedge W satisfying the conditions of the theorem whose height at x = v is at most that of K.
Let K ⊆ R2 satisfy the conditions of the theorem. We first show that K contains a point on the line at
x = v. If not, we claim that K has Gaussian mass at most γ1 (−v, v) ≤ γ1 (−v0 , v0 ) < 1/2 by choosing v0
small enough, a clear contradiction. To see this, note that by pushing the mass of K to the right as much
as possible towards the line x = v, we can replace K by a band [a, v] × R with the same Gaussian mass, and
barycenter to the right of b(K). Clearly, such a band has barycenter to the right of the y-axis iff a ≥ −v,
and hence K has Gaussian mass at most γ2 ([−v, v] × R) = γ1 (−v, v), as needed.
Now assume that K has height at least g at x = v, where we recall that g = Φ−1 (α). Note now that the
band W = R × (−∞, g] (corresponding to s = 0, t = 1, d = g, c = ∞) has height at x = v at most that of K
and satisfies the conditions of the theorem. Thus, we may assume that the height of K at x = v is f , where
−∞ < f < g. Note that (v, f ) is now a point on the boundary of K.
Let g 0 denote the height of K at x = 0. Since γ2 (K) ≥ 1/2, by Lemma 14 we have that γ1 (∞, g 0 ) ≥ γ2 (K),
and hence g 0 ≥ g ≥ 0. Thus g 0 ≥ g > f , and hence v > 0 (since otherwise we would have f = g 0 ).
By convexity of K, we may choose a line ` tangent
to K passing through (v, f ). We may now choose
t ≥ 0, s, d ∈ R, such that s2 + t2 = 1 and ` = (x, y) ∈ R2 : sx + ty = d . Since K is downwards-closed,
t ≥ 0 and ` is tangent to K, we must have that K ⊆ H` := {(x, y) : sx + ty ≤ d}. Since 0 is below (0, g 0 ) ∈ K,
we have that 0 ∈ H` , and hence d ≥ 0. Given the (0, g 0 ) ⊆ H` , we have that tg 0 ≤ d, and, because ` is
tangent at (v, f ), also sv + tf = d; using that v > 0and g 0 > f , we conclude that s > 0.
We will now show that the wedge W = H` ∩ (x, y) ∈ R2 : x ≥ −c satisfies our requirements for an
appropriate
choice of c (note the conditions
for s, t, d are already satisfied by the above paragraph). Let
Bv− = (x, y) ∈ R2 : x ≤ v and Bv+ = (x, y) ∈ R2 : x ≥ v . Choose c ≥ −v such that γ2 (W ∩ Bv− ) =
γ2 (K ∩ Bv− ). Note that such a c must exist since K ⊆ H` . Now by construction, note that W has the
same height as K at x = v, so it remains to check that c ≥ 0, γ2 (W ) ≥ 1/2 and b(W )1 ≥ 0. To bound the
Gaussian mass, again by construction, we have that
γ2 (W ) = γ2 (W ∩ Bv− ) + γ2 (W ∩ Bv+ ) = γ2 (K ∩ Bv− ) + γ2 (W ∩ Bv+ )
≥ γ2 (K ∩ Bv− ) + γ2 (K ∩ Bv+ ) = γ2 (K) ≥ 1/2 .
32
Figure 1: B is the triangular region beneath the red line
Given that γ2 (W ) ≥ 1/2, we must have that 0 ∈ W and hence c ≥ 0, as needed. It remains to check that
b(W )1 ≥ 0. For this purpose, note first that we can transform K ∩ Bv− into W ∩ Bv− by only pushing mass
to the right, and hence
Z
Z
2
2
xe−(x+y) /2 dy dx .
(25)
xe−(x+y) /2 dy dx ≤
W ∩Bv−
K∩Bv−
Since v ≥ 0 and K ∩ Bv+ ⊆ W ∩ Bv+ , we also immediately get that
Z
Z
2
−(x+y)2 /2
xe−(x+y) /2 dy dx .
xe
dy dx ≤
(26)
W ∩Bv+
K∩Bv+
We derive b(W )1 ≥ 0 by combining (25), (26) and our assumption that b(K)1 ≥ 0.
above reduction, we may now assume that K is a wedge of the form
Given the
(x, y) ∈ R2 : x ≥ −c, sx + ty ≤ d , d, s, t ≥ 0, s2 + t2 = 1, as in equation (24). We first take care of some
trivial cases. Firstly, if s = 0, t = 1, we have K = [−c, ∞) × (−∞, d]. Then the height at x = v is clearly d,
and since γ1 (−∞, d) ≥ γ2 (K), we get d ≥ g as is needed. Now assume that s = 1, t = 0, then K = [−c, d]×R,
and hence the height at x = v is infinite (note that K always intersects the line at x = v by the first part),
and thus the desired bound trivially
holds. We may thus assume that both s, t > 0.
In this setting, the line ` = (x, y) ∈ R2 : sx + ty = d intersects the x-axis at a = d/s and forms an
angle θ ∈ (0, π/2) with the x-axis as in Figure 1. Given the normalization s2 + t2 = 1, note that d is the
perpendicular distance from 0 to the edge K ∩ ` of K. In what follows, we maintain the parametrization of
the wedge K in terms of a,θ,c, using the relations d = a sin θ, s = sin θ and t = cos θ to recover d, s, t when
needed.
Recall that γ2 (K) ≥ α = γ1 (−∞, g) and b(K)1 ≥ 0. Let f ∗ = f (v, g) and let f be the height of K at x = v.
We want to prove that f ≥ f ∗ . If f ≥ g, we are already done since f ∗ ≤ g. Note that
by Lemma 14, f ≥ gif
v = 0 and hence we may assume v > 0. Our goal is now to show that g − f ≤ min e
g2
2
1
− 8ev
2
2
, (4e + 2) vg
.
Step 2: Using the barycenter condition We now derive a bound on how large d must be, given c and
θ, such that the x-coordinate ofbarycenter is non-negative.
Let H denote the halfspace (x, y) ∈ R2 : sx + ty ≤ d and let E = H \ K. A simple computation shows
R
−d2 /2
2
2
x 1 e−1/2(x +y ) dy dx = − e √2π sin θ. Now from the fact that b(K)1 ≥ 0, and E = H \ K, we get
H 2π
−d2 /2
− e √2π sin θ ≥
1
R
E
2
It follows that sin θ ≤ e− 2 (c
1 −1/2(x
x 2π
e
−d2 )
2
+y 2 )
dy dx ≥
R
x≤−c
2
√1 xe−1/2x dx
2π
−c2 /2
= − e √2π .
. When c2 ≥ −2 log sin θ, we get the following bound on d:
p
d ≥ c2 + 2 log sin θ .
33
(27)
When c2 < −2 log sin θ, this gives no useful bound on d since in that case the barycenter is non-negative
even for d = 0. But d ≥ 0 always as the Gaussian measure of K is at least half. Thus,
p
(28)
d ≥ max{0, c2 + 2 log sin θ} .
Step 3: Getting a bound on f By construction of K, the point (v, f ) lies on the boundary of H, and
hence
vs + f t = v sin θ + f cos θ = a sin θ = d .
(29)
Now,
1 2
e− 2 c
1
1
√
√
≤ γ1 (c, ∞) ≤ γ2 (E)
2
2π c + c2 + 4
(using lemma 46) .
Also,
1 2
1
γ2 (E) = γ2 (H) − γ2 (K) = γ1 (−∞, d) − γ1 (−∞, g) = γ1 (g, d) ≤ √ (d − g)e− 2 g
2π
(since d ≥ g ≥ 0) .
Combining the above two, we get
1
2
2
e 2 (g −c )
√
≤d−g .
c + c2 + 4
(30)
From (28) and (30),
2
2
1
p
e 2 (g −c )
√
}.
d ≥ max{ max{0, c2 + 2 log sin θ}, g +
c + c2 + 4
Putting the above in (29),
p
max{0, c2 + 2 log sin θ} − v sin θ g +
,
f ≥ max
cos θ
1
2
2
−c )
e 2 (g
√
c+ c2 +4
cos θ
− v sin θ
.
Observe that
γ1 (−∞, g) = γ2 (K) ≤ γ1 (−c, ∞) = γ1 (−∞, c),
giving c ≥ g. Also θ ∈ [0, π/2]. Thus, the above lower bound on f holds if we minimize over all c ≥ g and
θ ∈ [0, π/2].
1 (g 2 −c2 )
pmax{0, c2 + 2 log sin θ} − v sin θ g + e 2 √ 2 − v sin θ
c+ c +4
f≥
min
max
,
.
cos θ
cos θ
c≥g,θ∈[0,π/2]
We will first minimize with respect to c. For this, we make the following observations:
• for a fixed θ, the first term inside the maximum is a non-decreasing function of c while the second is a
decreasing function of c.
p
• for c = g 2 − 2 log sin θ ≥ g, the first term is smaller than the second term
p
• for c = g 2 − 2 log sin θ + 1 ≥ g, the first term is greater than the second term.
Thus, the two terms must become equal somewhere in the range
p
p
c ∈ [ g 2 − 2 log sin θ, g 2 − 2 log sin θ + 1] .
34
In particular, substituting c =
f
≥
p
g 2 − 2 log sin θ + 1 in the second term provides a lower bound for f :
√
sin θ √
g 2 −2 log sin θ+1+ g 2 −2 log sin θ+5)
g+
√
g+
sin θ
√ √
2 e g 2 −2 log sin θ+5
min
e(
cos θ
θ∈[0,π/2]
≥
− v sin θ
min
− v sin θ
.
cos θ
θ∈[0,π/2]
This expression goes to g as θ → 0 and to ∞ as θ → π/2. If it is increasing in this whole interval, we are
already done. Else, it achieves its minimum somewhere in (0, π/2). Let this be at θ∗ . Setting the derivative
to zero, we get
f
sin 2θ∗
≥ g cos θ∗ − √ p
3
4 e g 2 − 2 log sin θ∗ + 5
sin 2θ∗
= g − 2g sin2 (θ∗ /2) − √ p
3 ,
4 e g 2 − 2 log sin θ∗ + 5
where θ∗ satisfies
(31)
1
v = g sin θ∗ + √ p
.
2
2 e g − 2 log sin θ∗ + 5
From the two terms above, we can get two upper bounds on sin θ∗ :
sin θ∗
≤
sin θ∗
≤
v/g ,
eg
2
/2+5/2
.
1
e 8ev2
Using these, we can simplify (31) as
f
≥
g − 2g sin2 (θ∗ /2) − 2e sin 2θ∗ (v − g sin θ∗ )3
≥
g − 2g sin2 (θ∗ /2) − 2ev sin 2θ∗
≥
g − 2g sin2 (θ∗ ) − 4ev sin θ∗ .
We derive two bounds on the above expression, one which will be useful when g is small and other when
g is large. For the small g bound, using that v < v0 for v0 small enough,
2
2g sin2 (θ∗ ) + 4ev sin θ∗ ≤ (2g
2
eg /2+5/2
eg /2
v
+ 4ev)
≤
.
1
1
g
e 8ev2
e 8ev2
For the large g bound,
2g sin2 (θ∗ ) + 4ev sin θ∗ ≤ 2g
v2
v2
4ev 2
= (4e + 2)
.
+
2
g
g
g
Thus,
f ≥ g − min{e
g2
2
1
− 8ev
2
, (4e + 2)
v2
} = f ∗ , as needed.
g
We now prove Theorem 41 in the special case where the barycenter lies to the right of the hyperplane
θ⊥ . We show later how to reduce Theorem 41 to this case.
Lemma 48. There exists universal constants v0 , c0 > 0, such that for any n ≥ 1, v ∈ [0, v0 ] and θ ∈ Rn ,
kθk2 = 1, convex body K ⊆ Rn satisfying γn (K) = α ≥ 1/2 and hb(K), θi ≥ 0, we have that
1
γn−1 (Kvθ ) ≥ α(1 −
35
e− 100v2
√
).
4 2π
Proof. We split the proof into two steps. In step one, we reduce to a 2-dimensional problem and show that it
suffices to prove our theorem for a downwards closed convex body K 0 ⊆ R2 . This reduction will guaranteee
that K 0 has barycenter on the y-axis and that the Gaussian measure of slices of K 0 parallel to the y-axis will
correspond in the natural way to that of slices of K parallel to the hyperplane θ⊥ = {x ∈ Rn : hx, θi = 0}.
We then invoke Lemma 47 to get a lower bound on the height of K 0 at x = v. Lastly, in step 2, we show
that implies the required lower bound on the slice measure.
Let g be s.t. γ1 (−∞, g) = α, i.e. g = Φ−1 (α). Note that g ≥ 0 since α ≥ 1/2.
Step 1: reduction to a 2-dimensional case We will reduce our problem to one for a 2-dimensional
downwards closed convex body K 0 . To specify K 0 , we need only specify the height of the boundary at each
x-coordinate. At x-coordinate t, we define the height of K 0 to be the yt satisfying γ1 (−∞, yt ) = γn−1 (Ktθ ).
From Ehrhard’s inequality, we see that K 0 is in fact convex. Furthermore, it is easy to check that γ2 (K 0 ) =
γn (K) and b(K 0 )1 = hb(K), θi ≥ 0.
Thus, K 0 is a downwards closed convex body in R2 with γ2 (K 0 ) = α, b(K 0 )1 ≥ 0. From here, we may
invoke Lemma 47 to conclude that the height of K 0 at x = v is at least f ∗ := f (v, g). We now have that
γn−1 (Ktθ ) = γ1 (−∞, yt ) ≥ γ1 (−∞, f ∗ ) .
From the above, it suffices to give a lower bound on γ1 (−∞, f ∗ ) in order to derive the theorem.
−
1
2
100v
to show γ1 (f ∗ , g) ≤ α e 4√
2π
1
−
Step 2: Bounding γ1 (−∞, f ∗ )
2
100v
). Clearly, it suffices
Our goal is to show that γ1 (−∞, f ∗ ) ≥ α(1 − e 4√
2π
2
g
1
2
. Let εg = g − f ∗ = min e 2 − 8ev2 , (4e + 2) vg . We split the analysis in two
cases depending on whether g is small or big.
Step 2a: g ≤
1
5v
g2
1
1
1
1
1
e 2 − 8ev2
e 50v2 − 8ev2
e− 100v2
e− 100v2
εg
√
≤ √
≤
≤ √
≤α √
.
γ1 (f ∗ , g) ≤ √
2π
2π
2π
8 2π
4 2π
The penultimate inequality holds for an appropriate choice of v0 , and the last inequality uses α ≥ 1/2.
Step 2b: g >
1
5v
Here we will use the other bound for εg .
γ1 (f ∗ , g) ≤
≤
∗ 2
2
2
ε
ε
√ g e−(f ) /2 = √ g e−g /2 egεg −εg /2
2π
2π
2
ε
(4e + 2)v 2 −g2 /2+(4e+2)v2
√
√ g e−g /2+gεg ≤
e
2π
g 2π
1
≤
1
e− 100v2
e− 100v2
5(4e + 2)v 3 − 1 2 +(4e+2)v2
√
e 50v
≤ √
≤α √
.
2π
8 2π
4 2π
The penultimate inequality holds for an appropriate choice of v0 , and the last inequality uses α ≥ 1/2.
We now come to the proof of Theorem 41.
Theorem 41 (restated): There exist universal constants v0 , η0 , c0 > 0, such that for any n ≥ 1, convex
body K ⊆ Rn satisfying kb(K)k2 = η ≤ η0 and γn (K) = α ≥ 3/5, v ∈ [−v0 , v0 ] and θ ∈ Rn , kθk2 = 1, we
have that
1
e− 100v2
θ
√
γn−1 (Kv ) ≥ (α − c0 η)(1 −
).
4 2π
36
Proof. By rotational invariance, we may assume that θ = e1 , the first standard basis vector. By possibly
replacing K by −K, we may also assume that v ≥ 0.
If b(K)1 ≥ 0, the desired lower bound follows directly from Lemma 48. Given this, we may assume that
−η ≤ b(K)1 < 0. To deal with this second case, the main idea is to remove some portion of K lying to the
n
⊥
left of the hyperplane e⊥
1 = {x ∈ R : x1 = 0} so that the barycenter of the remaining body lies on e1 . After
this, we apply Lemma 48 again on the truncated body.
Define
Z
2
1
e− 2 kxk2
b := b(K)1 γn (K) =
x1 √ n dx < 0 .
2π
K
−
+
n
n
Let Ht = {x ∈ R : x1 ≤ t} and Ht = {x ∈ R : x1 ≥ t} for t ∈ R. Let z < 0 be defined as the smallest
negative number satisfying
Z
2
1
e− 2 kxk2
(32)
x1 √ n dx = 0 .
2π
K∩Hz+
By continuity, such a z must exists, since as z → −∞ the left hand side tends to b < 0 and at z = 0 it is
positive. Given the above, note that b(K ∩ Hz+ )1 = 0. We will now show that γn (K ∩ Hz+ ) ≥ α − c0 η if
η ≤ η0 , for c0 , η0 appropriately chosen constants.
By our choice of z, we have the equality
Z
2
1
e− 2 kxk2
x1 √ n dx = b .
2π
K∩Hz−
From here, we see that
γn (K ∩
Hz− )
2
1
Z
=
K∩Hz−
e− 2 kxk2
√ n dx ≤
2π
2
1
Z
K∩Hz−
x1 e− 2 kxk2
b
b
.
√ n dx = =
z
z
z
2π
Given this, we get that
η
b
≥α−
.
z
z
γn (K ∩ Hz+ ) = γn (K) − γn (K ∩ Hz− ) = α − γn (K ∩ Hz− ) ≥ α −
We now show that there exists a constant c0 s.t. 1/|z| ≤ c0 . Let β = γn (K ∩ H0+ ), and note that
β = γn (K) − γn (K ∩ H0− ) ≥ α − 1/2 ≥ 3/5 − 1/2 = 1/10 .
Let τ > 0 be positive number satisfying γ1 (0, τ ) = β, i.e. τ = Φ−1 (1/2 + β). By pushing the mass of K ∩ H0+
to the left towards e⊥
1 as much as possible, we see that
Z
Z
2
2
1
1
2
e− 2 kxk2
e− 2 kxk2
1
x1 √ n dx ≥
x1 √ n dx = √ (1 − e−τ /2 ) .
(33)
+
n
2π
2π
2π
K∩H0
{x∈R :0≤x1 ≤τ }
Next, by inclusion
Z
2
1
e− 2 kxk2
x1 √ n dx ≥
2π
K∩{x∈Rn :z≤x≤0}
1
2
2
e− 2 kxk2
1
x1 √ n dx = √ (e−z /2 − 1) .
n
2π
2π
{x∈R :z≤x≤0}
Z
Given that z satisfies (32), combining equations (33), (34), we must have that
0 ≥ e−z
2
/2
− e−τ
2
Thus, we may set c0 = 1/Φ−1 (6/10). Set η0 =
/2
⇒ |z| ≥ τ ≥ Φ−1 (6/10) > 0 .
1
10c0 .
Since η ≤ η0 , we have that
γn (K ∩ Hz+ ) ≥ α − c0 η ≥ 3/5 − 1/10 = 1/2 .
Lastly, using Lemma 48 on K ∩ Hz+ , we now get that
1
γn−1 (Kve1 )
= γn−1 ((K ∩
Hz+ )ev1 )
as needed.
37
e− 100v2
≥ (α − c0 η)(1 − √
),
4 2π
(34)
10
Constructive Vector Komlós
In this section we give a new proof of the main result of [24] that the natural SDP for the Komlós problem
has value at most 1. While the proof in [24] used duality, our proof is direct and immediately yields an
algorithm to compute an SDP solution which only uses basic linear algebraic operations, and does not need
a general SDP solver. We state the main theorem next.
Theorem 49. Let v1 , . . . , vn ∈ Rm be vectors of Euclidean length at most 1, and let α1 , . . . , αn ∈ [0, 1].
There exists an n × n PSD matrix X such that
Xii = αi
V XV
T
∀1 ≤ i ≤ n
Im ,
where V = (v1 , . . . , vm ) is the n × m matrix whose columns are the vectors vi .
To prove Theorem 49 we make use of a basic identity about inverses of block matrices. This is a standard
use of the Schur complement and we will not prove it here.
Lemma 50. Let
A=
A11
A21
A12
A22
be a (k + `) × (k + `) block matrix, where A11 is a k × k matrix, A12 is a k × ` matrix, A21 is a ` × k matrix,
and A22 is a ` × ` matrix. Assume A and A22 are invertible, and write B = A−1 in block form as
B11 B12
B=
,
B21 B22
−1
where Bij has the same dimensions as Aij . Then B11 = (A11 − A12 A−1
(i.e. the inverse of the Schur
22 A21 )
complement of A11 in A).
From Lemma 50 we derive the main technical claim used in the proof of Theorem 49.
Lemma 51. Let A = V T V be an n × n positive definite matrix, and let v1 , . . . , vn be the columns of V . Let
B = A−1 . Then, for each 1 ≤ i ≤ n
Bii =
1
1
≥
,
2
k(I − Π−i )vi k2
kvi k22
where Π−i is the orthogonal projection matrix onto span{vj : j 6= i}.
Proof. It is sufficient to prove the lemma for i = 1. Let U be the matrix with columns v2 , . . . , vn . Since
A is positive definite, the principal minor U T U is positive definite as well, and, therefore, invertible. By
Lemma 50,
1
B11 =
.
kv1 k22 − v1T U (U T U )−1 U T v1
Let Π = U (U T U )−1 U T . Since Π is symmetric and idempotent (i.e. Π2 = Π), it is an orthogonal projection
matrix. Moreover ΠU = U and Π has the same rank as U , so Π is the orthogonal projection matrix onto
the column span of U , i.e. U (U T U )−1 U T = Π−1 and the lemma follows.
Proof of Theorem 49. We prove the theorem by induction on n.
In the base case m = 1, we have a single vector v ∈ Rm , kvk2 ≤ 1, and an α ∈ [0, 1]. We set x = α, and
we clearly have vxv T αI I.
We now proceed with the inductive step. Consider first the case that V T V is singular. Then there exists
a vector x 6= 0 such that V x = 0. Scale x so that x2i ≤ αi for all i, and there exists k such that x2k = αk2 .
Apply the inductive hypothesis to the vectors (vi : i 6= k) and the reals (αi0 = αi − x2i : i 6= k) to get a matrix
Y ∈ R([m]\{k})×([n]\{k}) . Extend Y to a matrix Ỹ ∈ Rn×n by padding with 0’s, i.e. Ỹij = Yij if i, j 6= k
and Ỹij = 0, otherwise. Define X = xxT + Ỹ : it is easy to verify that both conditions of the theorem are
satisfied.
38
Finally, assume that V T V is invertible, and let B = (V T V )−1 . Define
β = min αi /Bii ,
i
k = arg min αi /Bii ,
i
γ = max αi − βBii .
i
Apply the inductive hypothesis to the vectors (vi : i 6= k) and the reals (αi0 = (αi − βbii )/γ : i 6= k) to get a
matrix Y ∈ R([n]\{k})×([n]\{k}) , which we then pad with 0’s to an n × n matrix Ỹ , as we did in the first case
above. Define X as X = βB + γ Ỹ . It is easy to verify that Xii = αi for all i. We have
V XV T = βV BV T + γV T Ỹ V = βV (V T V )−1 V T + γU T Y U,
where U is the submatrix of V consisting of all columns of V except vk . U T Y U I by the induction
hypothesis. Since V (V T V )−1 V T is symmetric and idempotent, it is an orthogonal projection matrix, and
therefore V (V T V )−1 V T I. Because Bii ≥ kvi k−2
2 ≥ 1 by Lemma 51, we have γ ≤ maxi αi − β. Therefore,
V XV T = βV (V T V )−1 V T + γU T Y U (β + γ)In (max αi )In In .
i
This completes the proof.
Observe that the proof of Theorem 49 can be easily turned into an efficient recursive algorithm.
References
[1] Milton Abramowitz and Irene A. Stegun, editors. Handbook of mathematical functions with formulas,
graphs, and mathematical tables. Dover Publications, Inc., New York, 1992. Reprint of the 1972 edition.
[2] Wojciech Banaszczyk. Balancing vectors and Gaussian measures of n-dimensional convex bodies. Random Structures Algorithms, 12(4):351–360, 1998.
[3] Wojciech Banaszczyk. On series of signed vectors and their rearrangements. Random Struct. Algorithms,
40(3):301–316, 2012. URL: http://dx.doi.org/10.1002/rsa.20373, doi:10.1002/rsa.20373.
[4] Nikhil Bansal. Constructive algorithms for discrepancy minimization. In 2010 IEEE 51st Annual
Symposium on Foundations of Computer Science FOCS 2010, pages 3–10. IEEE Computer Soc., Los
Alamitos, CA, 2010.
[5] Nikhil Bansal, Daniel Dadush, and Shashwat Garg. An algorithm for Komlós conjecture matching
Banaszczyk’s bound. In 57th Annual IEEE Symposium on Foundations of Computer Science, FOCS
2016, New Brunswick, NJ, USA, October 9-11, 2016, pages 788–799, 2016.
[6] Franck Barthe, Olivier Guédon, Shahar Mendelson, and Assaf Naor. A probabilistic approach to the
geometry of the lpn -ball. Ann. Probab., 33(2):480–513, 2005.
[7] József Beck. Roth’s estimate of the discrepancy of integer sequences is nearly sharp. Combinatorica,
1(4):319–325, 1981.
[8] József Beck and Tibor Fiala. “Integer-making” theorems. Discrete Appl. Math., 3(1):1–8, 1981.
[9] Christer Borell. The Ehrhard inequality. C. R. Math. Acad. Sci. Paris, 337(10):663–666, 2003.
[10] Boris Bukh. An improvement of the Beck-Fiala theorem. CoRR, abs/1306.6081, 2013.
[11] Bernard Chazelle. The discrepancy method. Cambridge University Press, Cambridge, 2000. Randomness
and complexity.
[12] William Chen, Anand Srivastav, Giancarlo Travaglini, et al. A Panorama of Discrepancy Theory,
volume 2107. Springer, 2014.
39
[13] Antoine Ehrhard. Symétrisation dans l’espace de Gauss. Math. Scand., 53(2):281–301, 1983.
[14] Ronen Eldan and Mohit Singh. Efficient algorithms for discrepancy minimization in convex sets. CoRR,
abs/1409.2913, 2014.
[15] Esther Ezra and Shachar Lovett. On the Beck-Fiala conjecture for random set systems. Electronic
Colloquium on Computational Complexity (ECCC), 22:190, 2015.
[16] David A. Freedman. On tail probabilities for martingales. Ann. Probability, 3:100–118, 1975.
[17] Nicholas J. A. Harvey, Roy Schwartz, and Mohit Singh. Discrepancy without partial colorings. In
Approximation, randomization, and combinatorial optimization, volume 28 of LIPIcs. Leibniz Int. Proc.
Inform., pages 258–273. Schloss Dagstuhl. Leibniz-Zent. Inform., Wadern, 2014.
[18] Rafal Latala and Krzysztof Oleszkiewicz. Gaussian measures of dilatations of convex symmetric sets.
Ann. Probab., 27(4):1922–1938, 1999.
[19] Avi Levy, Harishchandra Ramadas, and Thomas Rothvoss. Deterministic discrepancy minimization via
the multiplicative weight update method. arXiv preprint arXiv:1611.08752, 2016.
[20] L. Lovász, J. Spencer, and K. Vesztergombi. Discrepancy of set-systems and matrices. European
J. Combin., 7(2):151–160, 1986. URL: http://dx.doi.org/10.1016/S0195-6698(86)80041-5, doi:
10.1016/S0195-6698(86)80041-5.
[21] Shachar Lovett and Raghu Meka. Constructive discrepancy minimization by walking on the edges.
SIAM J. Comput., 44(5):1573–1582, 2015. Preliminary version in FOCS 2012.
[22] Jiřı́ Matoušek. Geometric discrepancy, volume 18 of Algorithms and Combinatorics. Springer-Verlag,
Berlin, 1999. An illustrated guide.
[23] John Von Neumann. Zur theorie der gesellschaftsspiele. Mathematische Annalen, 100(1):295–320, 1928.
[24] Aleksandar Nikolov. The Komlós conjecture holds for vector colorings. arXiv preprint arXiv:1301.4039,
2013.
[25] Aleksandar Nikolov and Kunal Talwar. Approximating hereditary discrepancy via small width ellipsoids.
In Symposium on Discrete Algorithms, SODA, pages 324–336, 2015.
[26] G. Paouris. Concentration of mass on convex bodies. Geom. Funct. Anal., 16(5):1021–1049, 2006.
[27] Thomas Rothvoss. Constructive discrepancy minimization for convex sets. In 55th Annual IEEE
Symposium on Foundations of Computer Science—FOCS 2014, pages 140–145. IEEE Computer Soc.,
Los Alamitos, CA, 2014.
[28] Joel Spencer. Six standard deviations suffice. Trans. Amer. Math. Soc., 289(2):679–706, 1985.
[29] Joel Spencer. Ten lectures on the probabilistic method, volume 52. SIAM, 1987.
[30] Aravind Srinivasan. Improving the discrepancy bound for sparse matrices: better approximations for
sparse lattice approximation problems. In Proceedings of the Eighth Annual ACM-SIAM Symposium
on Discrete Algorithms (New Orleans, LA, 1997), pages 692–701. ACM, New York, 1997.
[31] Michel Talagrand. The generic chaining. Springer Monographs in Mathematics. Springer-Verlag, Berlin,
2005. Upper and lower bounds of stochastic processes.
40
11
11.1
Appendix
Estimating the Barycenter
In this section we show how to efficiently estimate the barycenter of K up to a small accuracy in `2 -norm.
For a convex body K ⊆ Rn , we let γK denote the Gaussian measure restricted to K. For a random variable
X in Rn , we denote the covariance of X by cov[X] = E[(X − E[X])(X − E[X])T ].
The following lemma shows that the covariance of a Gaussian random vector shrinks when restricted to
a convex body. We include a short proof for completeness.
Lemma 52. Given a convex body K in Rn , let γK be the Gaussian distribution restricted to K, and let X
be a random variable distributed according to γK . Then, cov[X] In .
Proof. Consider f (t) = ln γn (K + t). f is concave in t. This follows from log-concavity of γn , an easy
consequence of the Prekopa-Leindler inequality. Hence, the Hessian of f , H(f ), is negative semi-definite. It
can be calculated that
H(f ) = H(ln γn (K + t)) = cov[X + t] − In ,
where X ∼ γK . Setting t = 0 completes the proof.
We will also need to use Paouris’ inequality [26], which we restate slightly:
Theorem 53. If X ⊆ Rn is a log-concave random vector with mean 0 and positive-definite covariance matrix
C, then for every t ≥ 1,
√
√
√
Pr[ X T C −1 X ≥ βt n] ≤ e−t n
where β > 0 is an absolute constant.
Theorem 54. Let K be a convex body in Rn , given by a membership oracle, with γn (K) ≥ 1/2. For any
δ > 0 and ε ∈ (0, 1), there is an algorithm which computes the barycenter of K within accuracy δ in `2 -norm
with probability at least 1 − ε in time polynomial in n, 1/δ and log(1/ε).
Proof. Let b be the barycenter of K and Xi for 1 ≤ i ≤ N be i.i.d generated from γK , where N =
d(β/δ)2 log2 (e/ε)ne. Here β is the constant from Theorem 53. Defining the following quantities
N
1 X
Xi
N i=1
b0
=
Yi
= Xi − b
N
1 X
=
Yi
N i=1
Y
C
=
E [(X − b)(X − b)T ]
X∼γK
we can see that b0 is an estimate of the barycenter, generated by averaging random samples from the
distribution γK and Y is the difference vector between the true barycenter and b0 . Thus it suffices to
bound the probability that Y is large and then show how to efficiently generate random samples from the
distribution γK . It holds that
E[Yi ] = E[Xi − b] = b − b = 0
Also, using Lemma 52,
E[Yi YiT ] = cov[Xi ] = C In
Thus,
E[Y ] = 0 and cov[Y ] = C/N In /N
41
Since γK is a log-concave distribution, Xi and hence Yi are log-concave random vectors. It is easily checked
(using the Prekopa-Leindler inequality) that the average of log-concave random variables is also log-concave
and hence Y is a log-concave random vector. Now,
s
−1
√
I
n
Y ≥ δ N
Pr[kY k2 ≥ δ] = Pr Y T
N
s
−1
√
C
≤ Pr Y T
Y ≥ δ N
(using C/N In /N )
N
Putting N = d(β/δ)2 log2 (e/ε)ne and using Theorem 53 with t = log(e/ε), we get
Pr[kY k2 ≥ δ] ≤ e−
√
n log(e/ε)
≤ ε/e ≤ ε/2.
(35)
We can generate the random points Xi using rejection sampling. For each i, we generate a sequence of
(1)
(k)
i.i.d. standard Gaussian random variables Xi , . . . , Xi ∈ Rn , k = dlog2 (2N/ε)e. We set Xi to the first
(j)
(j)
Xi in the sequence that belongs to K; if no such Xi exists, we set Xi arbitrarily. Clearly, conditional on
(j)
the existence of a j ≤ k such that Xi ∈ K, Xi ∼ γK . Furthermore, because K has Gaussian measure at
(j)
least 1/2, for every j we have Pr[Xi 6∈ K] ≤ 1/2, so
(j)
Pr[∀j : Xi
6∈ K] ≤ 2−k ≤ ε/2N.
By a union bound, with probability at least 1 − ε/2, all Xi are distributed according to γK ; let us call this
event E. Conditional on E, inequality (35) holds, and
Pr[kb − b0 k2 ≥ δ] = Pr[kY k2 ≥ δ]
= Pr[kY k2 ≥ δ | E] · Pr[E] + Pr[kY k2 ≥ δ | E c ](1 − Pr[E])
≤ ε/2 + ε/2 = ε.
The algorithm needs to generate O(N log(N/ε)) d-dimensional Gaussian random variables, check membership
in K for each of them, and compute the average of N points. Since each of these operations takes polynomial
time, and N is polynomial in n, δ, and log(1/ε), the running time of the algorithm is polynomial.
11.2
Proof of Lemma 11
Proof. We first prove subgaussianity. Let θ ∈ S n−1 , t ≥ 0. By assumption on X,
2
Pr[|hX, θi| ≥ t] = min Pr[cosh(λhX, θi) ≥ cosh(λt)] ≤ min β ·
λ>0
λ>0
≤ min 2β · e
σ 2 λ2 /2−λt
λ>0
≤ 2β · e
− 21 (t/σ)2
2
eσ λ /2
cosh(λt)
,
2
where the last
p inequality follows by setting λ = t/σ .
Let α = log2 β + 1. To prove that X is ασ-subgaussian, since probabilities are always at most one, it
suffices to prove that
n
o
2
2
1
1
min 1, 2β · e− 2 (t/σ) ≤ 2e− 2 (t/(ασ)) , ∀t ≥ 0.
√
Replacing t ← 2σt, the above simplifies to showing
n
o
2
2
min 1, 2β · e−t ≤ 2e−(t/α) , ∀t ≥ 0.
(36)
From here, we see that
β·e
−t2
≤e
−(t/α)2
⇔β≤e
(1−1/α2 )t2
42
r
⇔t≥
ln β ·
p
α2
= ln(2β).
2
α −1
(37)
p
2
2
Let r = ln(2β), noting that 1 = 2β · e−r = 2e−(r/α) , we have that for t ≤ r, the LHS of 36 is 1 and
2
the RHS is at least 1, for t > r, the LHS is equal to 2β · e−t and the RHS is larger by 37. Thus, X is
ασ-subgaussian as needed.
We now prove the furthermore. For X an n-dimensional standard Gaussian, note that hX, wi is distributed like σY , where Y ∼ N (0, 1) and σ = kwk2 . Hence,
Z ∞
2
1
eσx e−x /2 dx
E[ehX,wi ] = E[eσY ] = √
2π −∞
Z ∞
2
2
2
1
= eσ /2 √
e−(x−σ) /2 dx = eσ /2 ,
2π −∞
as needed.
43
| 8cs.DS
|
Modeling the Formation of Social Conventions in
Multi-Agent Populations
Ismael T. Freire¹ ³
, Clément Moulin-Frier², Marti Sanchez-Fibla²,
Xerxes D. Arsiwalla¹ ² ³, Paul Verschure¹ ² ³ *
¹ IBEC, Instituto de Bioingeniería de Cataluña, Barcelona, Spain
² UPF, Universitat Pompeu Fabra, Barcelona, Spain
³ BIST, Barcelona Institue of Science and Technology, Barcelona, Spain
* ICREA Institució Catalana de Recerca i Estudis Avançats, Barcelona, Spain
Abstract
In order to understand the formation of social conventions we need to know the specific
role of control and learning in multi-agent systems. To advance in this direction, we
propose, within the framework of the Distributed Adaptive Control (DAC) theory, a
novel Control-based Reinforcement Learning architecture (CRL) that can account for
the acquisition of social conventions in multi-agent populations that are solving a
benchmark social decision-making problem. Our new CRL architecture, as a concrete
realization of DAC multi-agent theory, implements a low-level sensorimotor control
loop handling the agent’s reactive behaviors (pre-wired reflexes), along with a layer
based on model-free reinforcement learning that maximizes long-term reward. We
apply CRL in a multi-agent game-theoretic task in which coordination must be achieved
in order to find an optimal solution. We show that our CRL architecture is able to both
find optimal solutions in discrete and continuous time and reproduce human
experimental data on standard game-theoretic metrics such as efficiency in acquiring
rewards, fairness in reward distribution and stability of convention formation.
[email protected]
I. Introduction
The formation of social conventions
In his seminal work “Convention” 1, David Lewis defines social conventions as
regularities in action that emerge to solve coordination problems and possess two main
characteristics: conventions are self-sustaining and arbitrary. Self-sustaining, in the
sense that a group of agents in a given population will continue to conform to it as long
as they expect the others to do so; and arbitrary, because there are other equally
acceptable possibilities to solve the same coordination problem.
Understanding under which conditions these conventions can be formed and how, is
still an open question, traditionally studied in cooperation games under the umbrella of
Game Theory studies 2, and later within Evolutionary Game Theory 3–5.
The most iconic example of Game Theory is the Prisoner’s Dilemma. The game
proposes a situation in which two arrested criminals have to decide, without
communication, whether to testify against the other (defect) or to remain silent
(cooperate). If both cooperate, they will spend just 1 year in jail. If both defect, they will
spend 6 years. If one cooperates and the other defects, the former will spend 10 years in
jail while the latter will be freed. In game theory, the combination of possible actions
and its subsequent rewards it’s represented by a payoff matrix like the one shown in
Figure 1.
One key concept on game theory is the so-called Nash Equilibrium 6, that is a set of
strategies in which no player has the interest to change unilaterally her or his own strategy.
In the case of the classical Prisoner’s Dilemma, and if the game is only played once, the
only Nash Equilibrium is mutual defection, because ‘defect’ will always render a higher
payoff, regardless of what the other player chooses.
Figure 1: Payoff matrix of the Prisoner’s Dilemma. Each cell of the matrix shows
the reward that each player gets for any combination of their respective choices. For
example, if player 1 (red) chooses to cooperate and player 2 (blue) chooses to defect
(as in the upper right cell), player 1 receives a reward of -10 and player 2 receives
a reward of 0.
There is a particularly interesting version of this game called Iterated Prisoner's Dilemma
(IPD), thoroughly studied by Robert Axelrod, that consists in playing several consecutive
rounds of the game. If one player chooses to ‘defect’ in a given round, the other player
has the chance to retaliate in the next round by ‘defecting’ back. In this case, if the number
of rounds is random or unknown to the players, ‘mutual defection’ is no longer the only
Nash Equilibrium, because the incentive to defect is countered by the fear of retaliation,
thus, leading to the emergence of cooperation as a Nash Equilibrium. As proven by
Axelrod’s seminal work 7, the IPD shows that cooperation can emerge out of repeated
interactions by pure natural selection. Following Axelrod’s findings, Russell Hardin
postulated that social conventions could be originated in large populations through the
same type of repeated dyadic interactions due to the capacity of each individual to
generalize to similar cases and the overlapping nature of group activities 8.
However, although the classical game theoretic approach remained hegemonic for more
than three decades in studying how humans and other animals cooperate 9,10, flaws in this
approach led to alternate models
11,12
. One of the major concerns was related to the
ecological validity of the experiments based on the IPD, arguing that the conditions under
which these experiments are conducted are hardly (if ever) found in nature
9,11,13
.
Particularly, many studies pointed to the fact that cooperation between animals (humans
and non-humans) usually requires a continuous exchange of information in order to
emerge, a feature that the IPD and other similar cooperation games lack, precisely because
they are based on discrete-time turns that impose a significant delay between actions 13–
15
.
In order to tackle this issue, several investigations have devised ways to modify standard
game theory discrete-time tasks into dynamic versions where individuals can respond to
each other’s actions in continuous-time 16–21. Their results point out that cooperation can
be more easily achieved in the dynamic version of the task due to the rapid flow of
information between individuals and their capacity to react in continuous-time 20,22.
One example of this can be found in
23
, where Hawkins and Goldstone show that
continuous-time interaction helps to converge to more stable strategies in a game
theoretical task compared to the same task modeled in discrete-time. They also show that
the involved payoffs affect the formation of social conventions. According to these
results, they suggest that real-life coordination problems can be solved either by
convention or through spontaneous coordination and that these solutions depend on what
is at stake if the coordination fails. To illustrate this point, they make the comparison
between two real-life examples of a coordination problem: On one hand, when we drive
a car, the stakes are high because if we fail to coordinate the outcome could be fatal, so
we resort to a convention – e.g. to drive on the right side of the road. On the other hand,
when we try to avoid people on a crowded street, we do it “on the fly” because the stakes
are low, so it’s not risky to rely on purely reactive behaviors (e.g. avoidance behavior) to
solve it.
In this paper, we propose a computational model of cognitive agents involved in a social
decision-making task called The Battle of the Exes and we confront our results to human
data obtained in 23. For this purpose, we develop a Control-based Reinforcement Learning
(CRL) cognitive architecture based on the principles of the Distributed Adaptive Control
multi-agent theory (DACma). Our architecture integrates both a low-level reactive
control loop to manage within-round conflicts, and a policy learning algorithm to acquire
across-round strategies. We run simulations showing that the modeled cognitive agents
rely more on across-round policy learning when the stakes of the game are higher and
that reactive control helps to acquire better outcomes in terms of efficiency and fairness.
This provides a computational hypothesis explaining key aspects of the emergence of
social conventions (e.g. turn-taking or pure dominance) in game-theoretic setups and
provides new experimental predictions to be tested in human coordination tasks.
Multi-Agent Reinforcement Learning
Regarding the computational modeling of game-theoretical tasks, there is already
extensive literature where the study of the emergence of conflict and cooperation in agent
populations has been tackled, especially through the use of multi-agent reinforcement
learning (for extensive reviews, check 24–26).
In this direction, a lot of focus has been put recently on developing enhanced versions of
the Deep Q-Learning Network architecture proposed in 27, particularly on their extensions
to the social domain 27–31. This architecture uses a reinforcement learning algorithm that
extracts abstract features from raw pixels through a deep convolutional network. Along
those lines, some researchers
28–30
are modeling the type of conflicts represented in the
classic game-theoretic tasks (e.g. the IPD) into more ecologically valid environments
28
where agent learning is based on deep Q-networks 29,30. For instance, agents based on this
cognitive model are already capable or learning how to play a two-player video game
such as Pong from raw sensory data and achieve human-level performance
cooperative and competitive modes
31
27
, both in
. Other approaches are paying more attention to
how to develop agents that achieve good outcomes in general-sum games and complex
social dilemmas, by focusing on maintaining cooperation 32, by making an agent prosocial
(taking into account other’s rewards)33 or by conditioning its behavior solely on its
outcomes 34.
While the computational approaches described above are able to reproduce some of the
key aspects of the formation of social behaviors in agent populations, they are not suited
for disentangling the role of reactive control and policy learning in this process. Even
when the coordination task is being modeled in real-time and the agents are biologically
and environmentally constrained, these approaches are not considering lower-level
sensorimotor control loops bootstrapping learning in higher levels of the cognitive
architecture.
In contrast, the Control-based Reinforcement Learning (CRL) architecture we propose is
based on the Distributed Adaptive Control multi-agent (DACma)35,36 theory, where
learning processes are bootstrapped from sensorimotor control loops, as we will see in
the next section. Moreover, we systematically confront our results to the experimental
human data collected in 23 for studying the conditions under which such agents are able
to converge towards social conventions. For this purpose, we use an already designed and
tested game-theoretical task called The Battle of the Exes 23, which we explain in the end
of the following section. In Section II, we describe the CRL architecture and its two
layers: one dealing with the low-level intrinsic behaviors of the agent and another based
on model-free reinforcement learning, allowing the agents to acquire rules for
maximizing long-term reward
37
. In Section III we compare the results of our model
against existing human data. Finally, we conclude this study by analyzing in Section IV
its main results and implications, by discussing the limitations and possible extensions of
the current model and by posing several experimental predictions.
Foundations of our approach
Our approach relies on two main contributions from the literature that we describe in
detail in this section. Firstly, the CRL architecture of our agents follows the principles
of the DACma theory, where learning processes are bootstrapped from pre-existing
reactive control loops. Secondly, we use an existing game-theoretical task called The
Battle of the Exes 23, for which human data is available in various experimental
conditions.
A. The Distributed Adaptive Control theory
DAC is a theory of brain and mind that proposes that cognition is based on four control
layers that operate at different levels of abstraction 35,36,38.
The first level, the Soma layer, contains the whole body of the agent with all the sensors
and actuators and represents the interface between the agent and its environment. This
layer also contains the physiological needs of the agent, which are the driving force of the
whole system. In the Reactive layer, those physiological needs are satisfied through the
self-regulation of internal drives, implemented as reactive sensorimotor loops for
maintaining them stable (in homeostasis). These reactive interactions bootstrap the
learning of instantaneous policies implemented in the Adaptive layer for acquiring a statespace of the agent-environment interaction. Outside the scope of this paper, the
Contextual layer acquires temporally extended policies that contribute to the acquisition
of more abstract cognitive abilities such as goal selection, memory and planning 36. These
higher-level representations, in turn, affect the behavior of the lower layers in a top-down
fashion. Control in this architecture is therefore distributed between all layers thanks to
the interactions in both directions, top-down and bottom-up, as well as laterally within
each layer.
DAC makes explicit the distinction between real-time control on one hand (Reactive
layer) and perceptual and behavioral learning on the other hand. It is, therefore, an
adequate theoretical framework for understanding the specific roles of reactive control
and policy learning in the formation of social conventions, which is the aim of the present
paper. This allows the identification of the functions the agents will need in both the
ballistic and the dynamic conditions of the Battle of the Exes. On one hand, in the ballistic
condition, where the players can only make a decision at the beginning of each round, the
cognitive agents will only need to use the adaptive layer for solving the task. On the other
hand, in the dynamic condition, the agents will need both the reactive and the adaptive
layer, since they will be moving through the environment, sensing and acting in
continuous-time, and not only making abstract discrete decisions.
In section II, we describe in detail the novel Control-based Reinforcement Learning
(CRL) architecture, composed of both a reactive and an adaptive control layer. This
results in a hybrid control-learning model where reinforcement learning interacts with a
feedback controller by inhibiting specific reactive behaviors.
B. Benchmark
In 23, Hawkins and Goldstone conducted a study to investigate how two factors – the
continuity of the interaction (ballistic vs. dynamic) and the stakes of the interaction
(high vs. low payoff) – affected the formation of social conventions between humans in
a social decision-making task called The Battle of the Exes.
The Battle of the Exes is a coordination game similar to the classic Battle of the Sexes 39,
that proposes the following situation: A couple just broke up and they don’t want to see
each other. Both have their coffee break at the same time, but there are only two coffee
shops in the neighborhood: one with great coffee (high reward) and one with okay coffee
(low reward). If both go to the great coffee shop they will come across into each other
and will not enjoy the break at all. So basically, if they want to enjoy their coffee break,
they will have to coordinate in a way that they avoid each other every day. This situation
can be modeled within the framework of game theory with a payoff relation such as a >
b > 0; where a is the payoff for getting the great coffee, b the payoff of the okay coffee
and 0 the payoff for both players if they go to the same location.
Regarding the stakes of the interaction, the payoff matrix is manipulated to create two
different conditions: high, where the difference between payoffs is higher; and low,
where the difference is lower. The payoff matrices in Figure 2 illustrate these two
conditions.
Figure 2: Payoff matrices of the original “Battle of the Exes” game. The numbers
indicate the reward received by each player (red and blue). Reproduced from 23.
As for the continuity of the interaction, the experiment has a ballistic and a dynamic
condition. In the ballistic condition, as in classical game theory, the players can only
choose an action at the beginning of every round of the game, without any further
control on the outcome. However, in the dynamic condition, the players can freely
change the course of their avatars until one of them reaches a reward (for a visual
example of the difference between conditions, check the original videos here). In both
conditions, the round ends when one of the players reaches one of the reward spots that
represent the coffee shops. Altogether, this results in four conditions in total, two for the
stakes of the interaction (high vs. low) combined with two for the continuity of the
interaction (ballistic vs. dynamic).
For the experiment, they pair the human players in dyads that depending on the payoff
condition, play 50 (high) or 60 (low) consecutive rounds together. In order to analyze the
coordination between the players of each dyad, they use three measures -efficiency,
fairness, and stability- based on Binmore’s three levels of priority 40:
•
Efficiency –It measures the cumulative sum of rewards that players were able to
earn collectively in each round, divided by the total amount of possible rewards.
If the efficiency value is 1, it means that the players got the maximum amount of
reward.
•
Fairness – It quantifies the balance between the earnings of the two players. If the
fairness value is 1, it means that both players earned the higher payoff the same
amount of times.
•
Stability – It measures how well the strategy is maintained over time. In other
words, it quantifies how predictable are the outcomes of the following rounds
based on previous results by “using the information-theoretic measure of
surprisal, which Shannon defined as the negative logarithm of the probability of
an event” 23.
The results show that players in the dynamic condition achieve greater efficiency and
fairness than their counterparts in the ballistic condition, both in the high payoff and low
payoff setups (see Figure 3). However, their key finding is that in the dynamic
condition, the players coordinate more “on the fly” (i.e. without the need of a long-term
strategy) when the payoff is low, but when the payoff is high, the participants
coordinate into more stable strategies. In other words, they identified the stakes of the
interaction as a crucial factor in the formation of social conventions when the
interaction happens in real-time.
The present paper attempts at replicating the results we have just described with a
computational model based on an integrated cognitive architecture, in order to identify
what are the specific roles of reactive control and policy learning in the emergence of
social conventions.
Figure 3: Human decision making in the Battle of the Exes task
23
, analyzed by
Efficiency (left), Fairness (center) and Stability (right). Within every plot, the
experimental conditions are divided into two columns: ballistic (left) and dynamic
(right); and by color: high (blue) and low (orange). Note that stability is measured
by the level of ‘surprisal’ (i.e. the negative logarithm of the probability of an event),
which means that lower surprisal values imply higher stability. Adapted from 23.
II. Methods
Model
The CRL architecture we propose in this study is a model-free approach to
reinforcement learning, but with the addition of a reactive controller (for a model-based
approach, see 41). The CRL is composed of two layers, a Reactive and an Adaptive
layer. The former governs the sensorimotor contingencies of the agent within the rounds
of the game, whereas the latter is in charge of learning across rounds.
A. Reactive Layer
The Reactive Layer (RL) represents the agent sensorimotor control system and is
supposed to be prewired (typically from evolutionary processes in a biological
perspective). In the Battle of the Exes game that we are considering here, we equip the
agents with two predefined reactive behaviors ‘orienting towards rewards’ and ‘escape
from agents’. This means that, even in the absence of any learning process, the agents are
intrinsically attracted to the reward spots and repulsed from each other. This intrinsic
dynamic will bootstrap learning in the Adaptive Layer, as we will see.
To model this layer, we follow an approach inspired by Valentino Braitenberg’s Vehicles
42
. These simple vehicles consist of just a set of sensors and actuators (e.g. motors) that,
depending on the type of connections created between them, can perform complex
behaviors. For a visual depiction of the two behaviors (‘orienting towards rewards’ and
‘escape from agents’), see Figure 4 (or this video).
Figure 4. A: Top view of an agent’s body, as represented by the dark-blue large
circle. Here the agent is facing toward the top of the page. The two thin black
rectangles at the sides represent the two wheels, controlled by their speed. On
its front, the agent is equipped with three types of sensors: A: agent sensors
(sensing the proximity of the other agent), L: low reward sensors, H: high
reward sensors. For each type, the agent is able to sense the proximity of the
corresponding entity both on its left and right side (hence six sensors in total).
B, C and D: Representation of the specific set of excitatory and inhibitory
connections between sensors and actuators that provides the agents with the
fundamental reactive behaviors: “escape from agents” (B), represented by red
lines; and “orienting towards rewards” for the low (C) and high rewards (D),
represented by green lines.
•
The ‘orienting towards rewards’ behavior is made by a combination of a crossed
excitatory connection and a direct inhibitory connection between the reward spot
sensors (𝑠 𝑋 ) and the motors (m), plus a forward speed constant value f set to 0.3:
𝑋
𝑋
𝑚𝑙𝑒𝑓𝑡 = 𝑓 + 𝑠𝑟𝑖𝑔ℎ𝑡
− 𝑠𝑙𝑒𝑓𝑡
𝑋
𝑋
𝑚𝑟𝑖𝑔ℎ𝑡 = 𝑓 + 𝑠𝑙𝑒𝑓𝑡
− 𝑠𝑟𝑖𝑔ℎ𝑡
𝑋
where 𝑠𝑙𝑒𝑓𝑡
is the sensor positioned on the left side of the robot indicating the
proximity of a reward spot, and 𝑋 is either the high (H) or the low reward (L)
sensor (see Figure 4, C and D). The sensors perceive the proximity of the spot.
The closer the reward spots, the higher the sensors will be activated.
𝑋
𝑋
Therefore, if no reward spot is detected (𝑠𝑙𝑒𝑓𝑡
= 𝑠𝑟𝑖𝑔ℎ𝑡
= 0), the robot will go forward at
speed f. Otherwise, the most activated sensor (left or right) will make the robot turn in
the direction of the corresponding reward spot.
•
The ‘escape from agents’ behavior is made by the opposite combination: a direct
excitatory connection and a crossed inhibitory connection, but in this case
between the agent sensors (𝑠 𝐴 ) and the motors (m):
𝐴
𝐴
𝑚𝑙𝑒𝑓𝑡 = 𝑓 + 𝑠𝑙𝑒𝑓𝑡
− 𝑠𝑟𝑖𝑔ℎ𝑡
𝐴
𝐴
𝑚𝑟𝑖𝑔ℎ𝑡 = 𝑓 + 𝑠𝑟𝑖𝑔ℎ𝑡
− 𝑠𝑙𝑒𝑓𝑡
𝐴
where 𝑠𝑙𝑒𝑓𝑡
is the sensor positioned on the left side of the robot indicating the
proximity of the other agent (see Figure 4B). The closer the reward spots, the
higher the sensors will be activated.
𝐴
𝐴
In this case as well, if no agent is detected (𝑠𝑙𝑒𝑓𝑡
= 𝑠𝑟𝑖𝑔ℎ𝑡
= 0), the robot will go forward
at the speed f. Otherwise, the most activated sensor will make the robot turn in the
opposite direction of the other agent, thus avoiding it.
B. Adaptive Layer
The agent’s Adaptive layer (AL) is based on a model-free reinforcement learning
algorithm that endows the agent with learning capacities for maximizing long-term
reward. Functionally, it decides the agent’s action at the beginning of the round, based on
the state of the previous round and its policy. The possible states 𝑆 are three: ‘high’, ‘low’
and ‘tie’; and indicate the outcome of the previous round for each agent. That is, if an
agent got the high reward on the previous round, the state is ‘high’; if it got the low
reward, the state is ‘low’; and if both agents went to the same reward, the state is ‘tie’.
The actions 𝐴 are three as well: ‘go to the high’, ‘go to the low’ and ‘none’.
In the ballistic condition, in which the agents only use this layer to operate, the two first
actions will take the agent directly to the respective reward spots (high and low), while
the ‘none’ action will choose randomly between them. In each round, the action chosen
by the AL is sampled according to 𝑃(𝐴 = 𝑎 | 𝑆 = 𝑠), where s is the actual state observed
by the agent.
Figure 5: Representation of the complete cognitive architecture of the agent. On top
the Adaptive layer (reinforcement learning control loop) composed of a Critic or
value function (V), an Actor or action policy (P) and an inhibitor function (i). In the
bottom, the Reactive layer (sensorimotor control loop), composed of three sets of
sensors 𝑠 𝐻 , 𝑠 𝐿 , 𝑠 𝐴 (corresponding to High/Low reward and the other Agent,
respectively), three functions 𝑓 𝐻 , 𝑓 𝐿 , 𝑓 𝐴 (corresponding to ‘orienting towards
High /Low reward’ and ‘escape from agents’ behaviors, respectively) and two
motors m𝑙 , m𝑟 (corresponding to the left and right motors). The action selected by
the AL is passed through the inhibitor function that will turn off one of the attraction
behaviors of the RL depending on the action selected. If the action is ‘go to the high’,
the ‘orienting towards low reward’ reactive behavior will be inhibited. If the AL
selects ‘go to the low’, the RL will inhibit its ‘orienting towards high reward’
behavior. If the AL selects ‘none’, the RL will act normally without any inhibition.
In the dynamic condition, the agent uses the whole architecture, with the Adaptive and
the Reactive layer working together (see Figure 5). As in the previous condition, the
agent’s AL chooses an action at the beginning of the round, based on the state of the
previous round and its policy. This action is then signaled to the RL, that will inhibit the
opposite reward-attraction reactive behavior according to the action selected by the AL.
In the case that the AL chooses the action ‘go to the high’, the RL will inhibit the
‘orienting towards low reward’ behavior, allowing the agent to focus only on the high
reward. Conversely, if the AL chooses the action ‘go to the low’, the reactive attraction
to the high reward will be inhibited. In both cases, the ‘agent avoidance’ reactive behavior
still operates. Finally, if the action ‘none’ is selected, instead of choosing randomly
between the other two actions as in the ballistic condition, the AL will not act at all, thus
relying completely on the behaviors of the RL to play that round of the game.
The Adaptive Layer is implemented as a model-free reinforcement learning algorithm for
maximizing accumulated reward over rounds through action, similar to the one
implemented in
43
(see Figure 6) and adapted to operate on discrete state and action
spaces.
Figure 6: Representation of the TD-learning algorithm implementing the Adaptive
layer. At each time step, the Critic (or value function, V) receives the state of the
environment and the current reward as an input and outputs a TD error (e) signal
informing the agent whether its last action has performed better or worse than
expected (see details below). This signal is sent to the Actor (or action policy, P) and
back to the Critic. The Actor uses the TD error to update its policy in order to
converge to the optimal action for each state, while the Critic uses it to update the
value function.
More specifically, we use an Actor-Critic Temporal Difference Learning algorithm (TDlearning), which is based on the interaction between two main components:
•
an Actor, or action policy π, which learns the mapping from states (s ϵ S) to actions
(a ϵ A) and define what is the action (a) -based on a probability (P)- to be
performed in each state (s);
π ∶ 𝑆 x 𝐴 → [0,1]
π(𝑎|𝑠) = 𝑃(𝑎𝑡 = 𝑎|s𝑡 = 𝑠)
•
and a Critic, or value function 𝑉π (𝑠), that estimates the expected accumulated
reward (𝐸[𝑅]) of a state (s) following a policy π;
𝑡
𝑉π (𝑠) = 𝐸[𝑅] = 𝐸[∑∞
𝑡=0 γ 𝑟𝑡 |𝑠0 = 𝑠],
where γ ϵ [0,1] is the discount factor, and 𝑟𝑡 is the reward at step t.
The Critic also estimates if the Actor performed better or worse than expected, by
comparing the observed reward with the prediction of 𝑉π (𝑠). This provides a learning
signal to the actor for optimizing it, where actions performing better (resp. worse)
than expected are reinforced (resp. diminished). This learning signal is called the
temporal-difference error (TD error). The TD error 𝑒(s𝑡 ) is computed as a function of the
prediction from value function 𝑉π (𝑠) and the currently observed reward of a given state
𝑟(s𝑡 ).
𝑒(s𝑡 ) = 𝑟(s𝑡 ) + γ 𝑉π (s𝑡 ) − 𝑉π (s𝑡−1 )
where γ is a discount factor that is empirically set to 0.40.
When e(s) > 0 (resp. e(s) < 0), this means that the action performed better (resp. worse)
than expected. The TD error signal is then sent both to the Actor and back to the Critic
for updating their current values.
The update of the Actor is done in two steps. First, a matrix 𝐶𝑜𝑢𝑛𝑡(a𝑡 , s𝑡−1 ), with rows
indexed by discrete actions and columns by discrete states, is updated according to the
TD error,
𝐶𝑜𝑢𝑛𝑡(a𝑡 , s𝑡−1 ) = 𝐶𝑜𝑢𝑛𝑡(a𝑡 , s𝑡−1 ) + 𝛿 𝑒(s𝑡 )
where 𝛿 is a learning rate, that is set to 0.45, a𝑡 is the current action and s𝑡−1 the previous
state.
𝐶𝑜𝑢𝑛𝑡(a𝑡 , s𝑡 ) integrates the observed TD errors when executing the action a𝑡 in the state
s𝑡 . It is initialized to 0 for all a𝑡 , s𝑡 and kept to a lower bound of 0.
𝐶𝑜𝑢𝑛𝑡(a𝑡 , s𝑡 ) is then used for updating the probabilities by applying Laplace’s Law of
Succession 44:
𝑃(𝐴 = 𝑎 | 𝑆 = 𝑠) =
𝐶𝑜𝑢𝑛𝑡(a𝑡 , s𝑡 ) + 1
(∑𝑎∈𝐴 𝐶𝑜𝑢𝑛𝑡(a𝑡 , s𝑡 )) + 𝑘
where k is the number of possible actions.
Laplace’s Law of Succession is a generalized histogram (frequency count) where it is
assumed that each value has already been observed once prior to any actual observation.
By doing so it prevents null probabilities (when no data has been observed, it returns a
uniform probability distribution). Therefore, the higher 𝐶𝑜𝑢𝑛𝑡(a𝑡 , s𝑡 ), the more probable
a𝑡 will be executed in s𝑡 .
Using these equations, actions performing better than expected (e(s) > 0) will increase
their probability to be chosen the next time the agent will be in state s𝑡−1 . When
e(s) < 0, the probability will decrease.
Finally, the Critic (value function) is updated following,
𝑉π (s𝑡−1 ) = 𝑉π (s𝑡−1 ) + 𝜂 𝑒(s𝑡 )
where 𝜂 is a learning rate that is set to 0.15.
Experimental Setup
We follow, as in the Battle of the Exes benchmark 23, a 2x2 between-subjects experimental
design. One dimension represents the ballistic and dynamic versions of the game, whereas
the other dimension is composed of the high and low difference between payoffs.
Each condition is played by 50 agents who are paired in dyads and that play together 50
rounds of the game if they are in one of the high payoff conditions (ballistic or dynamic),
or 60 rounds if they are in one of the low payoff conditions.
For the task, we have developed the two versions (ballistic and dynamic) of the Battle of
the Exes in a 2D simulated robotic environment (the source code is available online at
https://github.com/IsmaelTito/PythonRobot2DSim). A representation of the navigational
interface of the experiment can be seen below, in Figure 7.
Figure 7: Screenshot of the Experimental Setup (top view). In blue, the two cognitive
agents in their initial position at the start of a round and equipped with different
sensors, described below. In green, the two reward spots; the bigger one
representing the high reward and the smaller, the low reward (i.e. lower payoff). In
white, the circles that delimit the tie area.
The rules of the game are as follows: A round of the game finishes when one of the agents
reaches a reward spot. If both agents are within the white circle area when this happens,
it’s considered a tie, and both get 0 points. The small spot always gives a reward of 1,
whereas the big spot gives 2 or 4 depending on the payoff condition (low or high
respectively, see Figure 2). The reward spots are allocated randomly between the two
positions at the beginning of each round.
III. Results
We report the main results of our experiment in relation to human performance in the
Battle of the Exes task22, which are analyzed using: efficiency, fairness, and stability 40.
For each of these measures, we report the results of the model and plot them in contrast
with the human data from 22. Then, we interpret those results and after that, we analyze
the role of each layer of the CRL architecture in relation to the data obtained in each
condition.
Efficiency
The efficiency scores of the model followed a normal distribution in the four conditions,
so first, a one-way ANOVA was performed, showing a statistically significant difference
between groups (F (3, 46) = 755, p < .001). Post-hoc independent samples t-tests showed
that dynamic (high and low) conditions, (M = 0.88, M = 0.86), are as well significantly
more efficient than their ballistic counterparts (M = 0.46, M = 0.45); between both high
(t (98) = -33, p < .001) and low conditions (t (98) = -33, p < .001). We can observe the
same statistical tendencies than in the human benchmark data (see Figure 8).
Figure 8: Efficiency results comparison between human data (left) and model (right) in the
four experimental conditions: ballistic high, ballistic low, dynamic high and dynamic low. In
both cases, the dynamic conditions achieved higher efficiency scores than the ballistic ones.
Fairness
As for the fairness scores of the model, a one-way ANOVA showed a statistically
significant difference between groups (F (3, 46) = 6.88, p < .001). Post-hoc independent
samples t-tests showed that, as in the benchmark, the high ballistic condition (M = 0.50)
is significantly less fair than the two dynamic conditions, high (t (98) = -3.73, p < .001)
and low (t (98) = -3.99, p < .001), both with a mean fairness score of 0.68. On the other
hand, there is no statistically significant difference between the low ballistic condition (M
= 0.61) and the low dynamic (p = .086), but, as opposed to the human results, the low and
high ballistic conditions of the model are significantly different (t (98) = -2.17, p < .05).
Again, the same statistical tendencies are observed in humans and model (see Figure 9).
Figure 9: Fairness results comparison between human data (left) and model (right) in
the four experimental conditions: ballistic high, ballistic low, dynamic high and
dynamic low. In both cases, the dynamic conditions achieved higher fairness scores
than their ballistic counterparts. The difference between both ballistic conditions is
similar as well in both human and model data.
Stability
Regarding stability, since the results of the four conditions showed a non-Gaussian
distribution, a non-parametric Kruskal-Wallis H-test was performed, showing a
statistically significant difference between groups (H (3) = 42.3, p < .001). Post-hoc
Mann-Whitney U-tests showed that the difference between ballistic low (M = 1.18) and
dynamic low (M = 1.17) conditions is not statistically significant (p = .10), but that the
dynamic high condition (M = 1.09) was significantly more stable than the ballistic high
(M = 1.16; p < .001). There was as well a statistically significant difference between the
two dynamic conditions (p < .001), but not between the two ballistic (p = .26). Similar
statistical tendencies are observed in humans and model (see Figure 10).
Figure 10: Comparison between stability results of human (left) and model (right) in
the four experimental conditions: ballistic high, ballistic low, dynamic high and
dynamic low. Note that stability is measured by the level of ‘surprisal’, which means
that lower surprisal values imply higher stability. In both cases, there is a difference
between dynamic high and low conditions, being the high condition more stable than
the low.
Analysis
Overall, the model achieved a good fitting with the benchmark. Like in the human
experiment, we can observe that the dynamic (continuous-time) version of the model
achieves better results in efficiency and fairness and that this improvement is consistent
regardless the manipulation of the payoff difference.
The remarkable increase in efficiency in the dynamic condition is due to the Reactive
layer key role in avoiding within-round conflict when both agents have chosen to go to
the same reward, a feature that a ballistic model of the game cannot achieve.
As for the results in stability, the model was overall less stable than the benchmark,
although it reflected a similar relation between conditions: no difference between high
and low payoff conditions in the ballistic setup, and an increase in stability in the high
dynamic condition compared to the low dynamic (see Figure 10). Nonetheless, our
results show the formation of social conventions such as turn-taking and dominance, as
shown in Figure 11. The examples shown in the figure illustrate how two conventions
were formed in the dynamic high condition, where this type of equilibria ocurred more
often and during more rounds than in the other three conditions, thus explaining the
higher stability in this condition.
To conclude, besides the mentioned differences between the human benchmark and the
model, our results are consistent with human data in that dynamic, continuous-time
interaction helps to converge to more efficient, fair and stable strategies when the stakes
are high.
Figure 11. Top panel: Outcomes of two dyads in the high dynamic condition
showing the formation of turn-taking (left) and pure dominance (right)
equilibria. Each bar represents the outcome of a round of the game. A red
bar means that player1 got the high reward, and a blue bar means that player
2 got the high reward. Black bars represent ties. Middle panel: Surprisal
measure over rounds of play. When the convention is formed, the surprisal
drops down because the outcomes start to be predictable. Bottom panel:
Frequency count of suprisal values in a histogram. The y axis represents
increasingly higher values of surprisal (in bits). The longer the convention is
maintained, the larger the initial bars of the frequency count (that represents
no surprisal).
Reliance on the Adaptive layer
In order to identify the specific role of each layer in the formation of the social
conventions, we run an analysis that estimates their participation (through the
introduction of the “none” action) as a function of the payoff difference. Based on the
results of benchmark and model in the dynamic condition, where higher payoff
differences help to achieve higher stability, we expect that the more we increase this
difference between payoffs, the more the agents will rely on the Adaptive layer.
For testing this prediction, we performed a simulation with six different conditions with
varying levels of payoff difference (high vs. low reward value), from 1-1 to 32-1. To
measure the level of reliance on each layer, we logged the number of times each agent
outputted a ‘none’ action, that is the action in which the agent relies completely on the
Reactive layer to solve the round.
Figure 12: Mean of the percentage of ‘not-none’ actions (ie. ‘go to the high’ and
‘go to the low’ actions) selected by the agents plotted against 6 conditions with an
increasing difference between high and low payoffs.
Considering that there are only 3 possible actions, if the Adaptive layer was randomly
choosing the actions, we should observe that the agent selects each action the same
amount of times (i.e. prior to any learning, when the action policy returns uniform
distributions over A). That means that, at the beginning of each dyad, the reliance on the
Reactive layer would be 33% and the reliance on the Adaptive layer 66%. Starting from
this point, if our hypothesis is correct, we will expect to observe an increase in the
reliance of the Adaptive layer as the payoff difference increases. Indeed, the results
show, as seen in Figure 12, that there is a steady increase in the percentage of selection
of the adaptive layer as the payoff difference augments.
IV. Discussion
In this paper, we have investigated the specific role that real-time control and learning
have on the formation of social conventions in multi-agent setups. We have proposed,
based on the principles of DAC, a novel Control-based Reinforcement Learning (CRL)
cognitive architecture. The CRL is a model-free approach to reinforcement learning, but
with the addition of a reactive controller. Our CRL architecture is composed of a
module based on an actor-critic TD learning algorithm that endows the agent with
learning capacities for maximizing long-term reward, and a low-level sensorimotor
control loop handling the agent’s reactive behaviors. This integrated cognitive
architecture is applied to a multi-agent game-theoretic task, the Battle of the Exes, in
which coordination between two agents can be achieved. We have shown that our
model is able to converge to both turn-taking and pure dominance equilibria. Moreover,
we have demonstrated that real-time agent interaction does affect the formation of more
stable, fair and effective social coventions when compared to the same task modeled in
discrete-time. These results are consistent with the ones Hawkins and Goldstone
obtained with human subjects in 23.
In addition, we can interpret our results in the context of a functional cognitive model,
the CRL, in which we can determine the specific role that reactive and adaptive control
loops play in the formation of social conventions and of spontaneous coordination. In
that sense, we can assess that the Reactive layer has a significant role in avoiding
within-round conflict (spontaneous coordination), while the Adaptive layer is required
to achieve across-round coordination (social conventions). In addition, the model
supports our hypothesis that higher payoff differences will increase the reliance on the
Adaptive layer.
To our knowledge, this is the first model that is able to reproduce human behavioral
data in a social decision-making game that was presented both in discrete-time and
continuous-time setups. Furthermore, similar experiments don’t take into account the
role that sensorimotor control loops have in solving the modeled problems in real life
scenarios, an issue that we do consider in our study and constitutes a fundamental
requirement for the development of a fully embodied AI.
Moreover, as a result of this work, we can now further study how specific
manipulations of the experimental conditions (like increasing the differences between
payoffs, as presented in Figure 12) can affect both the outcome of the task and the
implication of both control loops in solving it. These results allow us to make
predictions that can later be tested in human experiments.
For future work, there are several directions in which we can continue to develop the
work presented in this paper. One possibility would be to focus on a more biologically
detailed approach. We could implement a Reactive layer that would be also more
consistent with the DAC theory by taking inspiration in the work done in
45
, where a
reactive controller model is presented that also regulates the agent’s internal needs. In this
direction, we could also implement a Contextual layer like the one presented in 46,47 that
allows the agent the learning of rules to maximize delayed reward. Such models allow to
build causal models of the world and to take into account context in the learning of
optimal action policies. Extending our agent with such functionalities could allow solving
much more complicated social coordination problems with delayed reward.
Another interesting avenue concerns the emergence of communication. We could extend
the model by adding signaling behaviors to the agents and test them in experimental
setups similar to the seminal sender-receiver games proposed by Lewis 1, or following a
more robot centered approach such as Luc Steels 48,49. With the addition of a Contextual
layer as proposed above, we could study the emergence of complex communicative
systems embedding a proto-syntax, as outlined in 43,50.
Finally, this work aims to help in the advance towards the development of a functional
embodied-AI that is able to survive in a complex multi-agent world, as outlined in 51. For
this purpose, we could extend this model to study not only other aspects of cooperation,
like the wolfpack hunting behavior of 52,53 but also of competing agent populations (e.g..
predator-prey). In ongoing work, we are developing a setup in which the simulated
cognitive agents presented in this work will have to compete for limited resources in a
massively multi-agent environment. This setup will also allow us to test the hypothesis
proposed in 54–56 about the role of consciousness as a game-theoretic evolutionary strategy
that could emerge out of a coevolutionary process triggered by a cognitive arms-race
between goal-oriented agents.
Acknowledgments
The research leading to these results has received funding from the European
Commission’s Horizon 2020 socSMC project (socSMC-641321H2020-FETPROACT2014) and by the European Research Council’s CDAC project (ERC-2013-ADG
341196).
V. References
1.
Lewis, D. Convention: A philosophical study. (John Wiley & Sons, 2008).
2.
Von Neumann, J. & Morgenstern, O. Game theory and economic behavior. Joh
Wiley Sons, New York (1944).
3.
Smith, J. M. Game theory and the evolution of fighting. Evol. 8–28 (1972).
4.
Smith, J. M. & Price, G. R. The Logic of Animal Conflict. Nature 246, 15–18
(1973).
5.
Smith, J. M. Evolution and the Theory of Games. (Cambridge University Press,
1982).
6.
Nash, J. F. Equilibrium Points in n-Person Games. Source Proc. Natl. Acad. Sci.
United States Am. 36, 48–49 (1950).
7.
Axelrod, R. & Hamilton, W. D. The evolution of cooperation. Science (80-. ).
211, 1390–1396 (1981).
8.
Hardin, R. Collective action. (Published for Resources for the Future by the
Johns Hopkins University Press, 1982).
9.
Clements, K. C. & Stephens, D. W. Testing models of non-kin cooperation:
mutualism and the Prisoner’s Dilemma. Anim. Behav. 50, 527–535 (1995).
10.
Riehl, C. & Frederickson, M. E. Cheating and punishment in cooperative animal
societies. Philos. Trans. R. Soc. London B Biol. Sci. 371, (2016).
11.
Connor, R. C. Altruism among non-relatives: alternatives to the ‘Prisoner’s
Dilemma’. Trends Ecol. Evol. 10, 84–86 (1995).
12.
Gintis, H., Bowles, S., Boyd, R. & Fehr, E. Explaining altruistic behavior in
humans. Evol. Hum. Behav. 24, 153–172 (2003).
13.
Noë, R. Cooperation experiments: coordination through communication versus
acting apart together. Anim. Behav. 71, 1–18 (2006).
14.
Miller, N., Garnier, S., Hartnett, A. T. & Couzin, I. D. Both information and
social cohesion determine collective decisions in animal groups. Proc. Natl.
Acad. Sci. U. S. A. 110, 5263–8 (2013).
15.
Taborsky, M., Frommen, J. G. & Riehl, C. Correlated pay-offs are key to
cooperation. Philos. Trans. R. Soc. London B Biol. Sci. 371, (2016).
16.
Oprea, R., Henwood, K. & Friedman, D. Separating the Hawks from the Doves:
Evidence from continuous time laboratory games. J. Econ. Theory 146, 2206–
2225 (2011).
17.
Friedman, D. & Oprea, R. A Continuous Dilemma. Am. Econ. Rev. 102, 337–363
(2012).
18.
Oprea, R., Charness, G. & Friedman, D. Continuous time and communication in
a public-goods experiment. J. Econ. Behav. Organ. 108, 212–223 (2014).
19.
Bigoni, M., Casari, M., Skrzypacz, A. & Spagnolo, G. Time Horizon and
Cooperation in Continuous Time. Econometrica 83, 587–616 (2015).
20.
van Doorn, G. S., Riebli, T. & Taborsky, M. Coaction versus reciprocity in
continuous-time models of cooperation. J. Theor. Biol. 356, 1–10 (2014).
21.
Kephart, C. & Friedman, D. Hotelling revisits the lab: equilibration in continuous
and discrete time. J. Econ. Sci. Assoc. 1, 132–145 (2015).
22.
Hawkins, R. X. D. Conducting real-time multiplayer experiments on the web.
Behav. Res. Methods 47, 966–76 (2015).
23.
Hawkins, R. X. D. & Goldstone, R. L. The Formation of Social Conventions in
Real-Time Environments. PLoS One 11, e0151670 (2016).
24.
Busoniu, L., Babuska, R. & De Schutter, B. A Comprehensive Survey of
Multiagent Reinforcement Learning. IEEE Trans. Syst. Man, Cybern. C Appl.
Rev. 38 (2), 2008 (2008).
25.
Claus, C. & Boutilier, C. The Dynamics of Reinforcement Learning in
Cooperative Multiagent Systems. AAAI/IAAI 752 (1998).
26.
Tan, M. Multi-agent reinforcement learning: Independent vs. cooperative agents.
in Proceedings of the tenth international conference on machine learning 330–
337 (1993).
27.
Mnih, V. et al. Human-level control through deep reinforcement learning. Nature
518, 529–533 (2015).
28.
Kleiman-Weiner, M., Ho, M. K., Austerweil, J. L., Littman, M. L. & Tenenbaum,
J. B. Coordinate to cooperate or compete: Abstract goals and joint intentions in
social interaction. COGSCI (2016).
29.
Leibo, J. Z., Zambaldi, V., Lanctot, M., Marecki, J. & Graepel, T. Multi-agent
Reinforcement Learning in Sequential Social Dilemmas. in Proceedings of the
16th Conference on Autonomous Agents and MultiAgent Systems 464–473
(International Foundation for Autonomous Agents and Multiagent Systems,
2017).
30.
Pérolat, J. et al. A multi-agent reinforcement learning model of common-pool
resource appropriation. in Advances in Neural Information Processing Systems
30 (eds. Guyon, I. et al.) 3646–3655 (Curran Associates, Inc., 2017).
31.
Tampuu, A. et al. Multiagent Cooperation and Competition with Deep
Reinforcement Learning. PLoS One 12, e0172395 (2015).
32.
Lerer, A. & Peysakhovich, A. Maintaining cooperation in complex social
dilemmas using deep reinforcement learning. Arxiv abs/1707.0, (2017).
33.
Peysakhovich, A. & Lerer, A. Prosocial learning agents solve generalized Stag
Hunts better than selfish ones. Arxiv abs/1709.0, (2017).
34.
Peysakhovich, A. & Lerer, A. Consequentialist conditional cooperation in social
dilemmas with imperfect information. Arxiv abs/1710.0, (2017).
35.
Verschure, P. F. M. J., Voegtlin, T. & Douglas, R. J. Environmentally mediated
synergy between perception and behaviour in mobile robots. Nature 425, 620–
624 (2003).
36.
Verschure, P. F. M. J., Pennartz, C. M. A. & Pezzulo, G. The why, what, where,
when and how of goal-directed choice: neuronal and computational principles.
Philos. Trans. R. Soc. London B Biol. Sci. 369, (2014).
37.
Moulin-Frier, C. et al. Top-Down and Bottom-Up Interactions between LowLevel Reactive Control and Symbolic Rule Learning in Embodied Agents. in
(2016).
38.
Verschure, P. F. M. J. Synthetic consciousness: the distributed adaptive control
perspective. Philos. Trans. R. Soc. London B Biol. Sci. 371, (2016).
39.
Fudenberg, D., Tirole, J., Fudenberg, D. & Tirole, J. Game theory. 1, (MIT Press,
1991).
40.
Binmore, K. Natural justice. (Oxford University Press, 2005).
41.
Herreros, I., Arsiwalla, X. & Verschure, P. A forward model at Purkinje cell
synapses facilitates cerebellar anticipatory control. In Advances in Neural
Information Processing Systems (pp. 3828-3836). (2016).
42.
Braitenberg, V. Vehicles: Experiments in Synthetic Psychology. (MIT Press,
1986).
43.
Moulin-Frier, C., Sanchez-Fibla, M. & Verschure, P. F. M. J. Autonomous
development of turn-taking behaviors in agent populations: A computational
study. in 2015 Joint IEEE International Conference on Development and
Learning and Epigenetic Robotics (ICDL-EpiRob) 188–195 (IEEE, 2015).
44.
Jaynes, E. T. & Bretthorst, G. L. Probability theory : the logic of science.
(Cambridge University Press, 2003).
45.
Sanchez-Fibla, M. et al. Allostatic control for robot behavior regulation: a
comparative rodent-robot study. Adv. Complex Syst. 13, 377–403 (2010).
46.
Duff, A., Sanchez Fibla, M. & Verschure, P. F. M. J. A biologically based model
for the integration of sensory–motor contingencies in rules and plans: A
prefrontal cortex based extension of the Distributed Adaptive Control
architecture. Brain Res. Bull. 85, 289–304 (2011).
47.
Marcos, E., Ringwald, M., Duff, A., Sánchez-Fibla, M. & Verschure, P. F. M. J.
The Hierarchical Accumulation of Knowledge in the Distributed Adaptive
Control Architecture. in Computational and Robotic Models of the Hierarchical
Organization of Behavior 213–234 (Springer Berlin Heidelberg, 2013).
doi:10.1007/978-3-642-39875-9_10
48.
Steels, L. Language games for autonomous robots. IEEE Intell. Syst. 16, 16–22
(2001).
49.
Steels, L. Evolving grounded communication for robots. Trends Cogn. Sci. 7,
308–312 (2003).
50.
Moulin-Frier, C. & Verschure, P. F. M. J. Two possible driving forces supporting
the evolution of animal communication. Phys. Life Rev. 16, 88–90 (2016).
51.
Moulin-Frier, C., Puigbò, J.-Y., Arsiwalla, X. D., Sanchez-Fibla, M. &
Verschure, P. F. M. J. Embodied Artificial Intelligence through Distributed
Adaptive Control: An Integrated Framework. (2017).
52.
Weitzenfeld, A., Vallesa, A. & Flores, H. A Biologically-Inspired Wolf Pack
Multiple Robot Hunting Model. in 2006 IEEE 3rd Latin American Robotics
Symposium 120–127 (IEEE, 2006). doi:10.1109/LARS.2006.334327
53.
Muro, C., Escobedo, R., Spector, L. & Coppinger, R. P. Wolf-pack (Canis lupus)
hunting strategies emerge from simple rules in computational simulations. Behav.
Processes 88, 192–197 (2011).
54.
Arsiwalla, X. D., Herreros, I., Moulin-Frier, C. & Verschure, P. Consciousness as
an Evolutionary Game-Theoretic Strategy. in 509–514 (Springer, Cham, 2017).
doi:10.1007/978-3-319-63537-8_43
55.
Arsiwalla, X. D., Herreros, I., Moulin-Frier, C., Sanchez, M. & Verschure, P. F.
Is consciousness a control process. Artif. Intell. Res. Dev. 233–238 (2016).
56.
Arsiwalla, X. D., Moulin-Frier, C., Herreros, I., Sanchez-Fibla, M. & Verschure,
P. The Morphospace of Consciousness. Arxiv abs/1705.1, (2017).
| 2cs.AI
|
1
ChineseFoodNet: A Large-scale Image Dataset for
Chinese Food Recognition
arXiv:1705.02743v3 [cs.CV] 15 Oct 2017
Xin Chen† , Yu Zhu† , Hua Zhou, Liang Diao, and Dongyan Wang*
Abstract—In this paper, we introduce a new and challenging
large-scale food image dataset called “ChineseFoodNet”, which
aims to automatically recognizing pictured Chinese dishes. Most
of the existing food image datasets collected food images either
from recipe pictures or selfie. In our dataset, images of each
food category of our dataset consists of not only web recipe
and menu pictures but photos taken from real dishes, recipe and
menu as well. ChineseFoodNet contains over 180,000 food photos
of 208 categories, with each category covering a large variations
in presentations of same Chinese food. We present our efforts
to build this large-scale image dataset, including food category
selection, data collection, and data clean and label, in particular
how to use machine learning methods to reduce manual labeling
work that is an expensive process. We share a detailed benchmark
of several state-of-the-art deep convolutional neural networks
(CNNs) on ChineseFoodNet. We further propose a novel two-step
data fusion approach referred as “TastyNet”, which combines
prediction results from different CNNs with voting method. Our
proposed approach achieves top-1 accuracies of 81.43% on the
validation set and 81.55% on the test set, respectively. The latest
dataset is public available for research and can be achieved at
https://sites.google.com/view/chinesefoodnet/.
Index Terms—dish recognition, deep learning, ChineseFoodNet, TastyNet
Fig. 1. Example images from our dataset. Each row shows five images from
one category of Chinese food. From top to bottom, the food names are Sichuan
noodles with peppery sauce, Mapo tofu, potato silk, and scrambled egg with
tomato, respectively. Variations in visual appearance of images of Chinese
food caused by complex background, various illumination, different angle of
view, different ingredients of the same category, etc. show challenges of visual
food recognition. All of these image keep their original size.
I. I NTRODUCTION
F
OOD plays an essential role in everyone’s lives, and
the behaviour of diet and eating impacts everyone’s
health [1]. Underestimating food intake directly relates to
diverse psychological implications [2]. In recent years, photographing foods and sharing them on social networks have
become a part of daily life. Consequently, several applications
have been developed to record daily meal activities in personal
food log system [3] [4] [5], which are employed to computeraided dietary assessment [6], further usage preference experiments [7] [8], calorie measurement [9] and nutrition balance
estimation [10] [11]. As one of user-friendly ways to input of
the food log, automatic recognition of dish pictures gives rise
of a research field of interest.
Deep convolutional neural networks (CNNs) have achieved
state-of-the-art in a variety of computer vision tasks [12] [13].
The visual dish recognition task is the same situation [14]. The
quality of training datasets always plays an important role for
†
These authors contributed equally to this work.
* means corresponding author.
Xin Chen, Hua Zhou, Yu Zhu and Dongyan Wang are with Midea
Emerging Technology Center, San Jose, 95134, USA. Xin’s email:
[email protected], Yu’s email: [email protected], Hua’s email:
[email protected], and Dongyan’s email: [email protected].
Liang Diao is with Midea Artificial Intelligence Research Institute, Shenzhen, Guangdong, 528311, P. R. China. email: [email protected].
training a deep neural network, where the high performance
of the deep model is still data-driven to some extent [15] [16].
However, to the best of our knowledge, there still exist no
effective Chinese food recognition system matured enough to
be used in real-world. The major reason is absence of largescale and high quality image datasets. In [17], the Chinese
food dataset includes 50 categories, each of which has only
100 images. Obviously, the size of this dataset is not sufficient
to satisfy deep learning training requirements.
The visual dish recognition problem has widely been considered as one of challenging computer vision and pattern
recognition tasks [14] [18]. Compared to other types of food
such as Italian food and Japanese food, it is more difficult to
recognize the images of Chinese dish as the following reasons:
1) The images of same category appear differently. Since
most of the same Chinese dish have different ingredients
and different cooking methods, the images are greatly
visual different, even for human vision;
2) The noise of images of Chinese dishes is hard to
model because of complex noise and a variety of
backgrounds.The images of Chinese food are taken in
various environment and complex background, for example dim light, vapour environment, strong reflection,
various utensils of Chinese dishes such as color, shape,
2
ornament, etc.
In order to give impetus to the progress of visual food
classification and related computer vision tasks, we build a
large-scale image dataset of Chinese dish, named by ChineseFoodNet. This dataset contains 185,628 images of 208 food
categories covering most of popular Chinese food, and these
images include web images and photos taken in real world
under unconstrained conditions. To the best of our knowledge,
ChineseFoodNet is the largest and most comprehensive dataset
for visual Chinese food recognition. Some of images of
ChineseFoodNet are shown in Figure. 2.
We benchmark nine CNNs models of four state-of-the-art
deep CNNs, SqueezeNet [19], VGG [20], ResNet [21], and
DenseNet [22], on our dataset. Experimental results reveal that
ChineseFoodNet is capable of learning complex models.
In this paper, we also propose a novel two-step data fusion
approach with voting. Although simple, voting is an effective
way to fuse results [23] [24]. Guided by our benchmarks, we
try some combination of different CNNs models Based on
results on ChineseFoodNet, we take ResNet152, DenseNet121,
DeneseNet169, DenseNet201 and VGG19-batch normalization (BN) [25] as our predictive models. 1 Then we fusing
these results with voting as a final result. This method is
designated as“ TastyNet”. Our proposed method has achieved
top-1 accuracy 81.43% in validation set and 81.55% in test
set, respectively. Compared to best results of the approaches
with a single network structure, the improvements of 2.38%
in validation set and 2.33% in these sets have been achieved,
respectively.
This paper takes three major contributions as following:
1) We present a large-scale image dataset, ChineseFoodNet,
for Chinese food recognition tasks. ChineseFoodNet is
made up with 185,628 images of 208 categories, and
most of the food image are from users’ daily life. It is
public available for research in related topics. 2
2) We provide a benchmark on our dataset. Totally nine
different models of four state-of-the-art CNNs architectures are evaluated. We presents the details of the
methodology used in the evaluation and the pre-trained
models will be public available for further research.
3) We propose a novel two-step data fusion approach for
visual food recognition, which combines predictive results of different CNNs with voting. Experimental results
on ChineseFoodNet have shown that approach improves
the performance compared to one deep CNNs model. It
has shown that data fusion should be an alternative way
to improve accuracy instead of only increasing numbers
of layers in CNNs.
The paper is organized as follows. Section II briefly reviews
some public food datasets and the state-of-the-art visual food
recognition methods. Section III describes the procedure of
building and tagging the ChineseFoodNet dataset. In section IV, several state-of-the-art CNNs methods are benchmarked on ChineseFoodNet. Section V details our proposed
1 The name of CNNs networks consists of letters+numbers. Letters are type
of CNNs, and following numbers are the number of layers.
2 Our dataset can be accessed from https://sites.google.com/view/
chinesefoodnet/.
data fusion approach and present our results on ChineseFoofNet. This paper closes with a conclusion of our work
and some future directions in section VI.
II. R ELATED W ORK
A. Food Dataset
The scholars have developed some public food datasets3 for
food-related applications such as dietary assessment, computational cooking, food recipe retrieval and so on. Pittsburgh Food
Image Dataset (PFID) collects 4,556 fast food images [26].
The UNICT-FD889 dataset of 3,583 images related to 889
distinct dishes are used for Near Duplicate Image retrieval
(NDIR) [18]. UEC-Food100 [27] and UEC-Food256 [28] are
both Japanese food datasets and contain 100 and 256 categories, respectively. The UPMC-FOOD-101 [29] and ETHZFOOD-101 [30] datasets are twin datasets and have same 101
food categories but different images. The images of UPMCFOOD-101 are recipe images, in which each has the additional
textual information, and the images of ETHZ-FOOD-101 are
selfies. VIREO-172 [31] is a Chinese Food dataset containing
a total of 353 ingredient labels and 110,241 images. However,
it aims at cooking recipe retrieval with ingredient recognition.
B. Visual Dish Recognition
Before introducing deep learning techniques to classification, traditional approaches with hand-crafted features have
been applied to visual food recognition, including the pairwise
feature distribution (PED) [32], Gabor filters [33], SIHT-based
Bag of Visual Words (BoW) [34] [4], optimized bag-offeatures model [35], co-occurrence [36], textons [37], Random
Forests (RF) [30], and Fisher Vector [38]. Like deep learning
applied to other computer vision tasks, CNNs models have
outperformed all of traditional methods and achieve higher
and higher accuracy with deeper and deeper CNNs [4] [6]
[14] [39] [40] [41].
However, all of these approaches of both traditional methods
and deep learning haven’t been tested on a large-scale image
dataset of Chinese food.
III. C HINESE F OOD N ET: A L ARGE - SCALE C HINESE F OOD
I MAGE DATASET
To the best of our knowledge, there is no such largescale image datasets for Chinese dish recognition which is
mature enough to provided necessary resources for the datadriven techniques, e.g. deep learning, to train complex food
recognition models. In this section, we present our procedures
to build ChineseFoodNet. Labelling image is an expensive
step in building large-scale dataset. In this paper, we design
and develop a semi-supervised method to accelerate the whole
process.
3 In order to review fairly, we only discuss the data that are available for
download in this paper. The last access date is June 1, 2017
3
Fig. 2. Fifty sample images of ChineseFoodNet dataset. The dataset contains 185,628 Chinese food images organized into 208 categories. All images in the
dataset are color. Images are resized for better presentation
A. Category Selection
C. Data Clean and Label
Various cooking styles exist in Chinese food culture, such as
Sichuan cuisine, Canton cuisine, etc. Our Chinese food dataset
must cover the most popular of Chinese cuisines from different
styles of cooking. In this subsection, we present our efforts to
meet this goal.
First, 250 food categories are gathered from the internet.4
However, some dishes are missed in search engine yet because they are too popular to be searched such as Tomato
omelette. In order to cover them, we conduct a survey of
favorite Chinese dishes within our group. Combining with
results of the survey, we select about 300 categories. Since
Chinese cruise categories is complex and some dishes are
very similar visually, such as Braised Chicken Wings and Cola
Chicken Wings, we manually merge related categories. After
this process, 208 categories of Chinese dish are taken.5
After collecting large number of food images, the next
step is to clean these data and generate proper labels for
each image. In this step, we first remove the images with
irregular height or width (too large or too small) which usually
are irrelevant images. Then we use entropy to clean the
images without content. Entropy is a quantitative metric of
image content [42]. We calculate the value of entropy of each
channel. If the value of any channel is small, we remove it
because the image doesn’t have enough useful information.
The following step is to remove duplicate and/or very similar
images with two steps. First, we calculate 1,024 deep features
with the last full connection layer of AlexNet [43]. Second,
we calculate the Euclidean distance to measure the similarity.
If the distance is below a threshold, we consider the images
are very similar and remove one.
Some of these images are clearly categorized with specific
Chinese food name, such as most of recipe and menu images.
The ground truth of this type of image can be directly
extracted. However, the number of such images is very limited
and the quality of those images are usually very high, e.g.,
the images are shot with sufficient light condition, good
presentation of the food, and good angels, etc. Thus this type
of images shows very different distributions comparing to the
images captured in daily life, and brings a potential impact for
the food recognition tasks in real life.
The other images are usually not well-labeled, and the food
photos are taken in real world conditions. Those images are
mainly from the users’ daily uploads which show very preferred data distributions in food images in the wild. Besides,
this type of images is usually associated with metadata. The
metadata can be viewed as an description of each image in
text format, which often describes the name, cooking recipe
B. Data Collection
There are two resources of our images, web images and
taken photos. The web images in our dataset are coming from
social network of the Chinese food and drink/cooking,6 where
users uploaded their Chinese food pictures and also provided
the tags (labels) of the image. Also some partial of the images
in this dataset are collected by our group in daily life.
After these steps, the number of images we brought together
achieves more than 500,000. However, those images may
contain missing labels, incorrect labels or unclear labels.
4 www.top.baidu.com
5 The names of Chinese dish in ChineseFoodNet are also listed at https:
//sites.google.com/view/chinesefoodnet/
6 www.douguo.com
4
Fig. 3. We show basic architectures of four well-known CNNs in our evaluation. From left to right, the architectures are VGG, Resnet, Densenet, and
Squeezenet, respectively.
and other information about the food in that image. In our
procedure, this metadata is utilized to filter the useful images
with correct labels. Particularly, we manually generate a set
of keywords for each food class in our database, and use each
set of the keyword to match the image metadata. Images with
metadata which contains the keywords of certain class are
selected and labeled with that class.
It should be noted that, after the aforementioned step, there
are still a number of incorrect labels, which are either caused
by unclear descriptions in metadata or irrelevant images. Label
validation by human labor on this large number of images
is an expensive task in terms of both time and costs. Here
we accelerate to label these image by some already labeled
samples in advance. We first collect a small database of
food images using the crowd-sourcing platform (no overlap
between our current dataset) with same class labels. Then a
shallow CNN model is trained for the food recognition task on
this small database. Given this CNNs model, we classify our
collected images into different classes representing candidate
labels. Specifically, top n (e.g., 5) predictions from the shallow
network are selected as the candidate labels for one image.
Finally, we perform manually label validation to finalize the
dataset by eliminating the wrong labelled images.
D. Dataset Description
After work of category selection, data collection, the data
collection and cleaning mentioned in previous subsections,
finally the ChineseFoodNet dataset contains 185,628 images,
with total size of 19.4 Gigabyte (GB). Images in the dataset
are kept their original size without any processing and color.
The total number of categories is 208 for the current version
of dataset, and each image is labelled with only one label from
0 to 207.
We split the whole dataset into training, testing and validation sets, approximately in the ratio 80%, 10% and 10%,
respectively. Specifically, there are 145,066, 20,254 and 20,310
images for training, validation and testing set, respectively.
Figure 1 and Figure 2 show some example images in our
dataset.
IV. B ENCHMARK ON C HINESE F OOD N ET DATASET
In this section, we conducted benchmark experiments for
the ChineseFoodNet dataset. First the experimental settings
are described, then we introduce the experimental protocol
and finally we provide the experimental results and analysis.
A. Experimental settings
Our experiments were all conducted using PyTorch [44]
deep learning framework. In the training phase, the initial
learning rate is set to 0.01, momentum is set 0.9, and
weight decay is set to 1e-4. We set the learning rate to the
initial learning rate decayed by 10 every 30 epoch. The number
of epoch for the training is set to 90. Training optimization
5
TABLE I
R ECOGNITION RATES OF DIFFERENT DEEP NETWORKS ON OUR FOOD DATASET. B OTH TOP -1 AND TOP -5 ACCURACY ARE SHOWN ON VALIDATION SET
AND TEST SET.
Method
Squeezenet1 1
VGG19-BN
ResNet18
ResNet34
ResNet50
ResNet152
DenseNet121
DenseNet169
DenseNet201
Validation
Top-1 Accuracy
Top-5 Accuracy
58.42%
85.02%
78.96%
95.73%
73.64%
93.53%
75.51%
94.29%
77.31%
95.20%
78.34%
95.51%
78.07%
95.42%
78.87%
95.80%
79.05%
95.79%
method is selected to stochastic gradient descent (SGD) with
momentum. No augmentation process is applied except the
resizing and mirror. Training images are firstly resized to
256x256, then a random crop of size 224x224 with hoizontal
flip (probability 0.5) is applied. We have used the pretrained
models from imagenet dataset [16] and fine-tuned the network
with our food data. During the testing, images are resized
to 256x256 and then we use center crop of size 224x224 to
feed into the network. All the experiments were conducted on
CentOS 7 operation system, with Intel Xeon E5 CPU (2.2G),
128GB RAM and Nvidia P100 Tesla GPUs hardware with
16G memory.
B. Experimental Protocol
The dataset is split into training, validation, and test sets by
random selection. There are 145,065 images in the training set.
There are 20254 images in the validation set and the rest 20310
images are used for testing. Comprehensive experiments have
been conducted using various popular deep learning network
architectures with different structures and different number of
layers. Specifically, we have benchmarked the performance
of: Squeezenet (version 1.1) [19], VGG19 (with BN layer)
[20], Resnet (18, 34, and 50) [21], DenseNet (121, 169,
and 201) [22]. In order to have a fair comparison, all the
experiments are using same input image size and same preprocessing/postprocessing procedures. Some implementation
details of ResNet and Squeezenet are illuminated in Figure. 4.
Fig. 4. Illustration of residual block in ResNet and fire block in Squeezenet.
C. Experimental Results
The recognition performance of different deep networks are
shown in Table I, both top 1 accuracy and top 5 accuracy
Test
Top-1 Accuracy
Top-5 Accuracy
58.24%
85.43%
79.22%
95.99%
73.67%
93.62%
75.82%
94.56%
77.84%
95.44%
79.00%
95.79%
78.25%
95.53%
78.72%
95.83%
78.78%
95.72%
are presented. Table I has shown that the best top-1 performance on validation set is 79.05%, which is achieved by
DenseNet201. The accuracies of VGG19 and DenseNet169
are also very close to the best results. On the test set, the best
recognition rate is 79.22%, obtained by VGG19, the second
best results is obtained by Resnet152, which is 0.22% lower
than then VGG19.
Deeper CNNs models generally achieve better performance [45] [46]. From the results, we can see that, CNN
models obtains significant improvements in performance when
number of layers in same network architecture are increased.
E.g., ResNet with 18 layers has recognition rate 73.64%,
while the deeper mode ResNet with 152 layers achieves about
5% improvement in both validation and test sets. Similar
results can be observed in DenseNet architecture. On the other
hand, deep models with wider structure also shows promising
performance, e.g., VGG19-BN obtains the best results in test
set, and the worst result (58.42 % and 58.24% on validation
and test sets, respectively) is achieved by Squeezenet v1.1,
which is a shallow and narrow network structure designed for
fast and efficient inference.
V. TASTY N ET: A TWO - STEP DATA FUSION APPROACH
A. Methodology
As shown in Table I, the accuracy has higher and higher
with deeper and deeper model. If we would improve furthermore, a possible way to use much deeper CNNs models. However, it needs much computation and memory resources. What
is more, deeper models easily lead to overfitting problem. The
alternative way is the data confusion approach. Its idea is to
fuse the inference results of different models. As shown in
Figure. 5, predictions from different networks are gathered
and a voting approach is utilized to obtain the final fused
prediction.
Based on some results of different combinations, as shown
in Table. II, we select the combination of models that achieves
the best top 1 result, ResNet152, DenseNet121, DenseNet169,
DenseNet201 and VGG19-BN. The voting method is to average the results of all models. The algorithm is details in
Algorithm.!1.
B. Results and Analysis
Different combinations of network architectures are applied
and the experimental results are shown in Table II. From
6
Fig. 5. Basic scheme of the two-step data fusion approach. The first one is to obtain some predictive results from different models. In TastyNet, we use
Resnet152, DenseNet121, DenseNet169, DenseNet201 and VGG19-BN. The second one is to combine these result to one final result with voting policy. It
his paper, we use weighted coefficient of the results of the first step.
TABLE II
E XPERIMENTAL RESULTS ( RECOGNITION ACCURACIES ) OF DIFFERENT FUSION SCHEMES
Fusion Method
ResNet (18 + 34 + 50 + 152)
DenseNet (121 + 169 + 201)
ResNet (18 + 34 + 50 + 152) + Densenet (121 + 169 + 201)
ResNet (18 + 34 + 50 + 152) + Densenet (121 + 169 + 201) + VGG19-BN
ResNet152 + DenseNet (121 + 169 + 201) + VGG19-BN
Top 1 Accuracy
Validation
Test
79.19%
79.46%
80.47%
80.17%
80.89%
81.08%
81.23%
81.12%
81.43%
81.55%
Top 5 Accuracy
Validation
Test
96.03%
96.16%
96.26%
96.30%
96.60%
96.67%
96.79%
96.76%
96.73%
96.76%
Algorithm 1 Algorithm of TastyNet.
1: Input:
2: Image
3: Output:
4: Number
. Range from 0-207
5: Predictive result from Resnet152, p(i), i from 0 to 207
6: Predictive result from DenseNet121, p(i), i from 0 to 207
7: Predictive result from DenseNet169, p(i), i from 0 to 207
8: Predictive result from DenseNet201, p(i), i from 0 to 207
9: Predictive result from Resnet152, p(i), i from 0 to 207
10: Get average result p(i) of all p(i), i from 0 to 207
11: Find maximum p(i) and get i
12: The output is number i
network on validation and test set, respectively. Based on the
experimental results, we select five CNNs models ,Resnet152,
DenseNet121, DenseNet169, DenseNet201 and VGG19-BN,
as components of TastyNet.
From our proposed approach, we get two conclusions as
followings:
1) By applying data fusing approach on different deep
networks, the overall performance can be further boosted
than using the single deep network;
2) Combination of different network architectures show
more benefits in improving the performance than the
combinations with same network architectures, and combination of deeper and wider networks obtains the best
results in our evaluation;
this table, we can conclude that the overall performance is
generally increasing for different combinations with ensemble
more deep networks. The fusion results of ResNet with different number of layers, obtained higher performance (79.46%
top 1 accuracy on test set) than single ResNet (ResNet 152,
77.84% top 1 accuracy on test set). Also the fusion results on
DenseNet achieved a 1.12% improvement on test set than the
best results achieved for single DenseNet architecture. Furthermore, combination of different types of CNNs networks (e.g.,
ResNets, DenseNets and VGG shown in Row 3 and 4 in Table
II) further improves the overall recognition performance. The
best result is obtained by fusing Resnet152 and Densenet 121,
Densenet 169, Densenet 201, and VGG19-BN, the recognition
accuracy is 81.43% on the validation set and 81.55% for the
test set. This results is 2.38% and 2.33% higher than the single
VI. C ONCLUSION AND F UTURE W ORK
In this paper, we have successfully created a very large-scale
image dataset for Chinese dish recognition, ChineseFoodNet.
It contains 185,628 images of 208 food categories, in which
the images are from not only web images but also real
world. As a consequence, the models trained on our dataset
should have covered most of food recognition applications.
Also, we present the benchmarks of nine state-of-the-art
CNNs models of four well-known CNNs architectures on
ChineseFoodNet. Finally, we propose a novel two-step data
fusion approach, ”TastyNet”. Based on experimental results,
we select Resnet 152, Densenent 121, Densenet 169, Densenet
201 and VGG19+BN models. After voting the results of
these model, we obtain final inference result. It has shown the
state-of-the-art results on ChineseFoodNet. What is more, our
7
proposed approach has shown that data fusion is an effective
way to obtain a better result instead of only working on one
type CNNs model.
For our future work, we are extending the number of
food category to over 500 that should be applied in much
applications. Also, we will investigate new fusion methods to
fuse the different results with different models to obtain the
better performance.
ACKNOWLEDGMENT
The authors would like to thank Zhi Zhang, Xukai Zhang,
Zigang Wang, Xiangping Zeng, Xiaofei Xu, Dangdang Mi and
Qingqing Chang for their discussions and efforts to collect
data. All of them are with Midea Health and Nutrition Institute.
R EFERENCES
[1] A. Mesas, M. Muñoz-Pareja, E. López-Garcı́a, and F. Rodrı́guezArtalejo, “Selected eating behaviours and excess body weight: a systematic review,” Obesity Reviews, vol. 13, no. 2, pp. 106–135, 2012.
[2] M. B. E. Livingstone and A. E. Black, “Markers of the validity of
reported energy intake,” The Journal of nutrition, vol. 133, no. 3, pp.
895S–920S, 2003.
[3] K. Aizawa, “Multimedia foodlog: Diverse applications from selfmonitoring to social contributions,” ITE Transactions on Media Technology and Applications, vol. 1, no. 3, pp. 214–219, 2013.
[4] Y. Kawano and K. Yanai, “Real-time mobile food recognition system,”
in Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition Workshops, 2013, pp. 1–7.
[5] O. Beijbom, N. Joshi, D. Morris, S. Saponas, and S. Khullar, “Menumatch: restaurant-specific food logging from images,” in Applications
of Computer Vision (WACV), 2015 IEEE Winter Conference on. IEEE,
2015, pp. 844–851.
[6] C. Liu, Y. Cao, Y. Luo, G. Chen, V. Vokkarane, and Y. Ma, “Deepfood:
Deep learning-based food image recognition for computer-aided dietary
assessment,” in International Conference on Smart Homes and Health
Telematics. Springer, 2016, pp. 37–48.
[7] K. Aizawa, K. Maeda, M. Ogawa, Y. Sato, M. Kasamatsu, K. Waki,
and H. Takimoto, “Comparative study of the routine daily usability
of foodlog a smartphone-based food recording tool assisted by image
retrieval,” Journal of diabetes science and technology, vol. 8, no. 2, pp.
203–208, 2014.
[8] K. Takahashi, K. Doman, Y. Kawanishi, T. Hirayama, I. Ide, D. Deguchi,
and H. Murase, “Estimation of the attractiveness of food photography
focusing on main ingredients,” in Proceedings of the 9th Workshop on
Multimedia for Cooking and Eating Activities in conjunction with The
2017 International Joint Conference on Artificial Intelligence. ACM,
2017, pp. 1–6.
[9] P. Pouladzadeh, S. Shirmohammadi, and R. Al-Maghrabi, “Measuring
calorie and nutrition from food image,” IEEE Transactions on Instrumentation and Measurement, vol. 63, no. 8, pp. 1947–1956, 2014.
[10] K. Aizawa, Y. Maruyama, H. Li, and C. Morikawa, “Food balance
estimation by using personal dietary tendencies in a multimedia food
log,” IEEE Transactions on multimedia, vol. 15, no. 8, pp. 2176–2185,
2013.
[11] S. Mezgec and B. Koroušić Seljak, “Nutrinet: A deep learning food
and drink image recognition system for dietary assessment,” Nutrients,
vol. 9, no. 7, p. 657, 2017.
[12] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,
no. 7553, pp. 436–444, 2015.
[13] J. Schmidhuber, “Deep learning in neural networks: An overview,”
Neural networks, vol. 61, pp. 85–117, 2015.
[14] H. Hassannejad, G. Matrella, P. Ciampolini, I. De Munari, M. Mordonini, and S. Cagnoni, “Food image recognition using very deep convolutional networks,” in Proceedings of the 2nd International Workshop
on Multimedia Assisted Dietary Management. ACM, 2016, pp. 41–49.
[15] Y. Bengio et al., “Learning deep architectures for AI,” Foundations and
trends R in Machine Learning, vol. 2, no. 1, pp. 1–127, 2009.
[16] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet:
A large-scale hierarchical image database,” in Computer Vision and
Pattern Recognition, 2009. CVPR 2009. IEEE Conference on. IEEE,
2009, pp. 248–255.
[17] M.-Y. Chen, Y.-H. Yang, C.-J. Ho, S.-H. Wang, S.-M. Liu, E. Chang, C.H. Yeh, and M. Ouhyoung, “Automatic Chinese food identification and
quantity estimation,” in SIGGRAPH Asia 2012 Technical Briefs. ACM,
2012.
[18] G. M. Farinella, D. Allegra, and F. Stanco, “A benchmark dataset to
study the representation of food images,” in European Conference on
Computer Vision. Springer, 2014, pp. 584–599.
[19] F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf, W. J. Dally,
and K. Keutzer, “Squeezenet: Alexnet-level accuracy with 50x fewer
parameters and < 0.5 mb model size,” arXiv preprint arXiv:1602.07360,
2016.
[20] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
[21] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” in Proceedings of the IEEE conference on computer vision
and pattern recognition, 2016, pp. 770–778.
[22] F. Iandola, M. Moskewicz, S. Karayev, R. Girshick, T. Darrell, and
K. Keutzer, “Densenet: Implementing efficient convnet descriptor pyramids,” arXiv preprint arXiv:1404.1869, 2014.
[23] X. Chen, L. Lin, and Y. Gao, “Parallel nonparametric binarization for
degraded document images,” Neurocomputing, vol. 189, pp. 43–52,
2016.
[24] C. Macdonald and I. Ounis, “Voting for candidates: adapting data fusion
techniques for an expert search task,” in Proceedings of the 15th ACM
international conference on Information and knowledge management.
ACM, 2006, pp. 387–396.
[25] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep
network training by reducing internal covariate shift,” in International
Conference on Machine Learning, 2015, pp. 448–456.
[26] M. Chen, K. Dhingra, W. Wu, L. Yang, R. Sukthankar, and J. Yang,
“PFID: Pittsburgh fast-food image dataset,” in Image Processing (ICIP),
2009 16th IEEE International Conference on. IEEE, 2009, pp. 289–
292.
[27] Y. Matsuda, H. Hoashi, and K. Yanai, “Recognition of multiple-food
images by detecting candidate regions,” in Proc. of IEEE International
Conference on Multimedia and Expo (ICME), 2012.
[28] Y. Kawano and K. Yanai, “Automatic expansion of a food image dataset
leveraging existing categories with domain adaptation,” in Proc. of
ECCV Workshop on Transferring and Adapting Source Knowledge in
Computer Vision (TASK-CV), 2014.
[29] X. Wang, D. Kumar, N. Thome, M. Cord, and F. Precioso, “Recipe
recognition with large multimodal food dataset,” in Multimedia & Expo
Workshops (ICMEW), 2015 IEEE International Conference on. IEEE,
2015, pp. 1–6.
[30] L. Bossard, M. Guillaumin, and L. Van Gool, “Food-101–mining discriminative components with random forests,” in European Conference
on Computer Vision. Springer, 2014, pp. 446–461.
[31] J. Chen and C.-W. Ngo, “Deep-based ingredient recognition for cooking
recipe retrieval,” in Proceedings of the 2016 ACM on Multimedia
Conference. ACM, 2016, pp. 32–41.
[32] S. Yang, M. Chen, D. Pomerleau, and R. Sukthankar, “Food recognition
using statistics of pairwise local features,” in Computer Vision and
Pattern Recognition (CVPR), 2010 IEEE Conference on. IEEE, 2010,
pp. 2249–2256.
[33] F. Zhu, A. Mariappan, C. J. Boushey, D. Kerr, K. D. Lutes, D. S. Ebert,
and E. J. Delp, “Technology-assisted dietary assessment,” in Proceedings
of SPIE, vol. 6814. NIH Public Access, 2008, p. 681411.
[34] F. Kong and J. Tan, “Dietcam: Automatic dietary assessment with mobile
camera phones,” Pervasive and Mobile Computing, vol. 8, no. 1, pp.
147–163, 2012.
[35] M. M. Anthimopoulos, L. Gianola, L. Scarnato, P. Diem, and S. G.
Mougiakakou, “A food recognition system for diabetic patients based
on an optimized bag-of-features model,” IEEE journal of biomedical
and health informatics, vol. 18, no. 4, pp. 1261–1271, 2014.
[36] Y. Matsuda and K. Yanai, “Multiple-food recognition considering
co-occurrence employing manifold ranking,” in Pattern Recognition
(ICPR), 2012 21st International Conference on. IEEE, 2012, pp. 2017–
2020.
[37] G. M. Farinella, M. Moltisanti, and S. Battiato, “Classifying food images
represented as bag of textons,” in Image Processing (ICIP), 2014 IEEE
International Conference on. IEEE, 2014, pp. 5212–5216.
[38] Y. Kawano and K. Yanai, “Foodcam: A real-time mobile food recognition system employing fisher vector,” in International Conference on
Multimedia Modeling. Springer, 2014, pp. 369–373.
[39] S. Christodoulidis, M. Anthimopoulos, and S. Mougiakakou, “Food
recognition for dietary assessment using deep convolutional neural net-
8
[40]
[41]
[42]
[43]
[44]
[45]
[46]
works,” in International Conference on Image Analysis and Processing.
Springer, 2015, pp. 458–465.
N. Martinel, G. L. Foresti, and C. Micheloni, “Wide-slice residual
networks for food recognition,” arXiv preprint arXiv:1612.06543, 2016.
Y. Kawano and K. Yanai, “Food image recognition with deep convolutional features,” in Proceedings of the 2014 ACM International
Joint Conference on Pervasive and Ubiquitous Computing: Adjunct
Publication. ACM, 2014, pp. 589–593.
M. Sonka, V. Hlavac, and R. Boyle, Image processing, analysis, and
machine vision. Cengage Learning, 2014.
A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Advances in neural information processing systems, 2012, pp. 1097–1105.
http://pytorch.org/, Last access on August 16, 2017.
C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
V. Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,”
in Proceedings of the IEEE conference on computer vision and pattern
recognition, 2015, pp. 1–9.
C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking
the inception architecture for computer vision,” in Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, 2016,
pp. 2818–2826.
| 1cs.CV
|
Submitted to the Annals of Statistics
arXiv:1603.00126v2 [math.ST] 10 Sep 2017
MULTICLASS CLASSIFICATION, INFORMATION,
DIVERGENCE, AND SURROGATE RISK
By John Duchi∗ , Khashayar Khosravi, and Feng Ruan∗
Stanford University
We provide a unifying view of statistical information measures,
multi-way Bayesian hypothesis testing, loss functions for multi-class
classification problems, and multi-distribution f -divergences, elaborating equivalence results between all of these objects, and extending
existing results for binary outcome spaces to more general ones. We
consider a generalization of f -divergences to multiple distributions,
and we provide a constructive equivalence between divergences, statistical information (in the sense of DeGroot), and losses for multiclass classification. A major application of our results is in multi-class
classification problems in which we must both infer a discriminant
function γ—for making predictions on a label Y from datum X—
and a data representation (or, in the setting of a hypothesis testing
problem, an experimental design), represented as a quantizer q from
a family of possible quantizers Q. In this setting, we characterize
the equivalence between loss functions, meaning that optimizing either of two losses yields an optimal discriminant and quantizer q,
complementing and extending earlier results of Nguyen et al. [23]
to the multiclass case. Our results provide a more substantial basis
than standard classification calibration results for comparing different losses: we describe the convex losses that are consistent for jointly
choosing a data representation and minimizing the (weighted) probability of error in multiclass classification problems.
1. Introduction. Consider the multiclass classification problem: a decision maker receives a pair of random variables (X, Y ) ∈ X × {1, . . . , k},
where Y is unobserved, and wishes to assign the variable X to one of the
k classes {1, 2, . . . , k} to minimize the probability of a misclassification. We
represent the decision maker via a discriminant function γ : X → Rk , where
each component γy (x), y = 1, . . . , k, represents the margin (or a score or
perceived likelihood) the decision maker assigns to class y for datum x. The
goal is then to minimize the expected loss, or L-risk,
(1)
RL (γ) := E[L(γ(X), Y )],
∗
Partially supported by NSF-CAREER Award 1553086 and the SAIL-Toyota Center
for AI Research. FR additionally supported by the Stanford Graduate Fellowship.
Keywords and phrases: f -divergence, risk, surrogate loss function, hypothesis test
1
2
DUCHI, KHOSRAVI, RUAN
where L(γ(x), y) measures the loss of margins γ(x) ∈ Rk when the true label of x is y and the expectation (1) is taken jointly over (X, Y ). When L
is the 0-1 loss, L(γ(x), y) = 1 {γy (x) ≤ γi (x) for some i 6= y}, the formulation (1) is the misclassification probability P(argmaxy γy (X) 6= Y ). We may
also consider the classical k-category Bayesian experiment: given a random
variable X ∈ X drawn according to one of the k hypotheses
H1 : X ∼ P1 ,
H2 : X ∼ P2 ,
...,
Hk : X ∼ Pk
with prior probabilities π1 , . . . , πk , P
we wish to choose γ minimizing the expected E[L(γ(X), Y )] or posterior y P(Y = y | X = x)L(γ(x), y) loss.
In many applications, making decisions based on the raw X is undesirable—
the vector X may be high-dimensional, carry useless information impinging
on statistical efficiency, or we may need to store or communicate the sample
using limited memory or bandwidth. If all we wish to do is to classify a person
as being taller or shorter than 160 centimeters, it makes little sense to track
his or her blood type and eye color. With the increase in the number and variety of measurements we collect, such careful design choices are important
for maintaining statistical power, interpretability, efficient downstream use,
and mitigating false discovery [3]. This desire to give “better” representations of data X has led to a rich body of work in statistics, machine learning,
and engineering, highlighting the importance of careful measurement, experimental design, and data representation strategies [28, 11, 26, 32].
As Nguyen et al. [23] note in the binary case, one thus frequently extends
the classical formulation (1) to include a stage in which a (data-dependent)
q : X → Z maps the vector X into a vector Z. A number of situations
suggest such an approach. In most practical classification scenarios [29], an
equivalent feature selection reduces the dimension of X or increases its interpretability. As a second motivation, consider the decentralized detection
problem [33, 23] in communication applications in engineering, where remote
sensors communicate data X ∼ Pi through limited bandwidth or memory. In
this case, the central processor can infer the distribution Pi only after communication of the transformed vector Z = q(X), and one chooses a quantizer
q from a family Q of (low complexity) quantizers. In fuller abstraction, we
may treat the problem as a Bayesian experimental design problem, where
the mapping q : X → Z may be stochastic and is chosen from a family Q
of possible experiments (observation channels). In each of the preceding examples, the incorporation of a quantizer q into the classification procedure
poses a more complex problem, as one must simultaneously find a data representation q and discriminant γ. The goal, paralleling that for the risk (1),
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK
3
thus becomes joint minimization of the quantized L-risk
(2)
RL (γ | q) := E [L(γ(q(X)), Y )]
over a prespecified family Q of quantizers q : X → Z, where γ : Z → Rk .
Often—for example, in the zero-one error case—the loss functions L(·, y)
are non-convex (even discontinuous), so population or empirical minimization is intractable. It is thus common to replace the loss with a convex
surrogate and minimize this surrogate instead. A surrogate is Fisher consistent if its minimization yields a Bayes optimal discriminant γ for the original
loss L (for any distribution P on (X, Y )); researchers have characterized the
Fisher consistency of convex surrogates for binary and multiclass classification [36, 21, 2, 31]. A weakness of such results is that they rely strongly on
using the unrestricted class of all measurable discriminants γ : X → Rk ,
and thus most “natural” convex losses are consistent [36, 2]. In this context,
a major difficulty is to understand the consequences of using various surrogate losses, and requiring a restricted quantizer class Q is one approach
to discovering nuanced properties of the relationship between surrogate and
Bayes risk. Nguyen et al. [23] tackle this in the binary case, considering
the problem of joint selection of the discriminant function γ : Z → R and
quantizer q : X → Z. They exhibit a precise correspondence between binary
margin-based loss functions and f -divergences—measures of the similarity
between two probability distributions developed in information theory and
statistics [1, 10, 34]—to give a general characterization of loss equivalence
through classes of divergences. An interesting consequence of their results is
that, in spite of positive results for Fisher consistency in binary classification
problems [36, 21, 2], essentially only hinge-like losses are consistent for the
0-1 loss. We provide the extension of these results to the multiclass case.
Outline and discussion of our contributions. We build on this prior work
to provide a unifying framework that relates statistical information measures, loss functions, and generalized notions of entropy in the context of
multi-class classification. To begin, we recall a generalization of f -divergences
that applies to multiple distributions [16, 13], enumerating analogues of the
positivity properties, data-processing inequalities, and discrete approximation available in the binary case, as multi-way f -divergences may be unfamiliar and they motivate our approach (Section 2). We begin our main contributions in Section 3, where we establish a correspondence between loss
functions L, generalized entropy on discrete distributions [15], and multiway f -divergences. To make this precise, define the probability simplex
∆k := {p ∈ Rk+ | 1T p = 1}. Let π ∈ ∆k be a prior distribution on the
4
DUCHI, KHOSRAVI, RUAN
class label Y and π
e(x) ∈ ∆k be the posterior probabilities for Y conditional
on the observation X = x. For concave H : ∆k → R, DeGroot [11] defines
the information associated with the experiment (X, Y ) as
(3)
IH (X; π) := H(π) − E[H(e
π (X))].
The notion of H as a generalized entropy is clear here, as I is the gap between
prior and posterior entropy and is always non-negative. In this context, the
value H(π) measures the uncertainty of the experimenter (in some appropriate units) about the unknown class label y when his or her prior belief
over y is π, so I is the gap between prior and posterior uncertainty [11].
To relate this type of entropy to loss functions, recall the well-known
result [11, 15] that any loss L : Rk × [k] → R ∪ {+∞} induces an entropy
HL : ∆k → R, also called the pointwise Bayes risk [15, 27, 13, 35], via
(4)
HL (π) = inf
α∈Rk
X
k
i=1
πi L(α, i) − I {π ∈ ∆k }
where I {·} is +∞ if its argument is false (we drop this indicator in the future,
defining HL implicitly on ∆k ). We show an inverse construction, providing
an explicit and constructable mapping from any concave function H to a loss
L inducing H as the pointwise Bayes risk (4), where for each y the loss α 7→
L(α, y) is convex. In Section 3.2, we also develop the natural connections
between these generalized entropies H and classification calibration [36, 21,
2, 31], in that our explicit L is generally calibrated.
In Section 4, we address the comparison of loss functions—building off
of Nguyen et al.’s approach in the binary case [23]—and present our main
results on consistency of joint selection of quantizer (data representation) q
and discriminant γ. Using our correspondence between concave H, losses L,
and f -divergences, we characterize the pairs of losses L(1) and L(2) for which
equivalent quantizers and discriminants minimize the quantized risk (2) in
the sense that there is a continuous concave h with h(0) = 0 such that
RL(2) (γ | q) − inf RL(2) (γ | q) ≤ h RL(1) (γ | q) − inf RL(1) (γ | q)
γ,q∈Q
γ,q∈Q
for any γ and q ∈ Q. Another way to understand our results is as providing
insight into classification calibration when the Bayes’ act (i.e. optimal discriminant γ) does not belong to the class of functions the statistician may
choose in a classification problem. A substantial challenge for and criticism
of the line of work on classification calibration and surrogate risk consistency [36, 21, 2, 31] is that the results say little for restricted families of
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK
5
classifiers. In this context, a corollary of our main contribution is as follows. The loss L(1) is calibrated [36, 2, 31] for L(2) if for any distribution
P on X × Y and sequence γn : X → Rk , RL(1) (γn ) → inf γ RL(2) (γ) implies RL(1) (γn ) → inf γ RL(2) (γ). Now, consider a collection Q of functions
q : X → Z for some set Z, and then define the class of functions
n
o
G(Q) := γ ◦ q | q ∈ Q and γ : Z → Rk is measurable .
Translated to this scenario, our main results—Theorems 1 and 2—imply
Corollary 1. Assume that the loss L(1) is calibrated for L(2) and let
Hi = HL(i) denote the associated pointwise Bayes risk (4). Then
(5)
RL(1) (gn ) → inf RL(1) (g) implies RL(2) (gn ) → inf RL(2) (g)
g∈G(Q)
g∈G(Q)
for any collection Q of mappings X → Z, any set Z, any distribution P on
X × {1, . . . , k}, and sequence gn ∈ G(Q) if and only if there exist a > 0,
b ∈ Rk , and c ∈ R such that H1 (π) = aH2 (π) + bT π + c for all π ∈ ∆k .
This corollary reposes on the connections we develop between losses, uncertainty measures and generalized f -divergences. Such measures of statistical information and divergence have been central to the design of communication and quantization schemes in signal processing [33, 25, 20, 18]; we
characterize the divergence measures that, when optimized, yield optimal
quantizers and detectors. We also provide a result showing when empirical
minimization of a surrogate risk is consistent for the desired (original) risk.
A number of researchers have studied the connections between divergence
measures and risk for binary and multi-category experiments; these point
to the results we present. Indeed, Blackwell [8] shows that if a quantizer
q1 induces class-conditional distributions with larger divergence than those
induced by q2 , then there are prior probabilities such that q1 allows tests
with lower probability of error than q2 . Liese and Vajda [19] give a broad
treatment of f -divergences, using their representation as the difference between prior and posterior risk in a binary experiment [24] to derive a number
of their properties; see also the paper [27]. Garcı́a-Garcı́a and Williamson
[13] show how multi-distribution f -divergences [16] arise naturally in the
context of multi-class classification problems as the gap between prior and
posterior risk in classification, as in the work [19]. In the binary case, these
results elucidate the characterization of Fisher consistency for quantization
and binary classification Nguyen et al. [23] realize. We pursue this line of
research to draw the connections between Fisher consistency, information
measures, multi-class classification, surrogate losses, and divergences.
6
DUCHI, KHOSRAVI, RUAN
Notation. We let 0 and 1 denote the all-zeros and all-ones vectors, respectively. For a vector or collection of objects we define t1:m = {t1 , . . . , tm }. The
indicator function I {·} is +∞ if its argument is false, 0 otherwise, while 1 {·}
is 1 if its argument is true, 0 otherwise. We let ∆k = {v ∈ Rk+ : 1T v = 1}
k
denote the probability
Pm simplex Tin R . For m ∈ N, we set [m] = {1, . . . , m}.
We let aff A = { i=1 λi xi | λ 1 = 1, xi ∈ A, m ∈ N} denote the affine
hull of a set A, and rel int A denotes the interior of A relative to aff A.
We let R = R ∪ {+∞} and R = R ∪ {−∞}. For f : Rk → R, we let
epi f = {(x, t) : f (x) ≤ t} denote the epigraph of f . We say a convex function f is closed if epi f is a closed set, though we abuse notation and say that
a concave f is closed if epi(−f ) is closed. For a convex function f : Rk → R,
we say that f is strictly convex at a point t ∈ Rk if for all λ ∈ (0, 1) and
t1 , t2 6= t such that t = λt1 + (1 − λ)t2 we have f (t) < λf (t1 ) + (1 − λ)f (t2 ).
The (Fenchel) conjugate of a function f : Rk → R is
(6)
f ∗ (s) = sup sT t − f (t) .
t∈Rk
For any f , the conjugate f ∗ is closed convex [17, Chapter X]. For measures ν
and µ, we let dν/dµ denote the Radon-Nikodym derivative of ν with respect
Lp
to µ. For random variables Xn , we say Xn → X∞ if E[|Xn − X∞ |p ] → 0.
2. Multi-distribution f -divergences. Divergence measures for probability distributions have significant statistical, decision-, and informationtheoretic applications, including in optimal testing, minimax rates of convergence, and the design of communication schemes [1, 10, 25, 18]. Central
to this work is the f -divergence, introduced by Ali and Silvey [1] and Csiszár
[10] (see [19] for an overview). Given distributions P, Q defined on a common
set X , a closed convex function f : [0, ∞) → R satisfying f (1) = 0, and any
measure µ dominating P and Q, the f -divergence between P and Q is
Z
Z
p(x)
dP
f
(7)
Df (P ||Q) :=
q(x)dµ(x) = f
dQ.
q(x)
dQ
X
dQ
Here p = dP
dµ and q = dµ denote the densities of P and Q, respectively, and
the value uf (t/u) is defined appropriately for t = 0 and u = 0 (e.g. [19]). A
number of classical divergence
measures arise out of the f -divergence; taking
√
f (t) = t log t, f (t) = 21 ( t − 1)2 , or f (t) = |t − 1| yields (respectively) the
KL-divergence, squared Hellinger distance, or total variation distance.
Central to our study of multi-way hypothesis testing and classification is
an understanding of relationships between multiple distributions. We use the
following generalization [16, 13] of the f -divergence to multiple distributions.
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK
7
Definition 2.1. Let P1 , . . . , Pk be probability distributions on a common
k−1
→ R be a closed convex function
σ-algebra F over a set X . Let f : R+
satisfying f (1) = 0. Let µ be any σ-finite measure such that Pi ≪ µ for all
i, and let pi = dPi /dµ. The f -divergence between P1 , . . . , Pk−1 and Pk is
Z
p1 (x)
pk−1 (x)
(8) Df (P1 , . . . , Pk−1 ||Pk ) := f
,...,
pk (x)dµ(x).
pk (x)
pk (x)
We must specify the value of the integrand (8) when pk (x) = 0. In this case,
the function fe : Rk+ → R defined, for an arbitrary t′ ∈ rel int dom f , by
if u > 0
uf (t/u)
′
(9)
fe(t, u) = lim sf (t − t + t/s) if u = 0
s→0
+∞
otherwise
is a closed convex function with value independent of t′ ; fe is the closure of
the perspective function R+ ×Rk ∋ (u, t) 7→ uf (t/u) of f [17, Prop. IV.2.2.2].
Any time we consider the perspective we implicitly treat it as its closure (9).
We now enumerate a few properties of multi-way f -divergences, showing how they naturally generalize classical binary f -divergences. We focus
on basic properties that are useful for our further results on Bayes risk,
classification, and hypothesis testing and that parallel results in the binary
case (7): they are well-defined, have continuity properties with respect to discrete approximations, and satisfy data-processing inequalities. While Györfi
and Nemetz’s original work [16] essentially contains the results, we carefully
address infinite values (the closure (9)) and strict convexity, and we use
them as definitional building blocks; we defer all proofs to Supplement D.
As our first step, we note that Definition 2.1 is independent of the base
measure µ. (See Supp. D.1 for a proof generalizing [16, Cor. 1].)
Lemma 2.1. In expression (8), the value of the divergence does not depend on the choice of the dominating measure µ. Moreover,
Df (P1 , . . . , Pk−1 ||Pk ) ≥ 0.
The inequality is strict if f is strictly convex at 1 and the Pi are not identical.
Given the importance of quantization to come, we now consider discrete
approximations to the divergence. For an at most countable partition P of
X , we define the partitioned f -divergence
X P1 (A)
Pk−1 (A)
,...,
Pk (A).
Df (P1 , . . . , Pk−1 ||Pk | P) =
f
Pk (A)
Pk (A)
A∈P
8
DUCHI, KHOSRAVI, RUAN
As in the binary case [34, 19], we have the following approximability result generalizing [16, Thm. 6] to possibly infinite integrands: quantizers give
arbitrarily good approximations to f -divergences (see § D.2 for a proof).
Proposition 1.
If f is a closed convex function with f (1) = 0, then
Df (P1 , . . . , Pk−1 ||Pk ) = sup Df (P1 , . . . , Pk−1 ||Pk | P)
P
where the supremum is over finite partitions of X .
In the binary case, f -divergences satisfy data processing inequalities [10,
9, 19] which state that processing or transforming an observation X drawn
from the distributions P1 , P2 , decreases the divergence between them. To
formalize this, recall that Q is a Markov kernel from a set X to Z if Q(· | x)
is a probability distribution on Z for each x ∈ X , and for each measurable
A ⊂ Z, the mapping x 7→ Q(A | x) is measurable. The following general data
processing inequality shows that this holds in the multi-distribution case as
well, generalizing [16, Thm. 4] to possibly infinite f and the closure (9); we
include a proof in Appendix D.3.
Proposition 2. Let f be closed convex with f (1) =R0, Q be a Markov
kernel from X to Z, and define the marginals QP (A) = X Q(A | x)dP (x).
Then
Df QP1 , . . . , QPk−1 ||QPk ≤ Df (P1 , . . . , Pk−1 ||Pk ) .
This proposition is related to the relationships between risk, information,
and quantization we develop in Sections 3 and 4. Defining a quantizer q to
be a measurable mapping q : X → Z between measurable spaces X and Z,
the quantized f divergence is
X
P1 (A)
Pk−1 (A)
f
Df (P1 , . . . , Pk−1 ||Pk | q) := sup
,...,
Pk (A),
Pk (A)
Pk (A)
P
−1
A∈q
(P)
where P ranges over finite partitions of Z and q−1 (P) = {q−1 (B) | B ∈ P}.
Proposition 2 immediately yields that quantization reduces information: the
indicator Q(A | x) = 1 {q(x) ∈ A} defines a Markov kernel, yielding
Corollary 2. Let f be closed convex, satisfy f (1) = 0, and q be a
quantizer of X . Then
Df (P1 , . . . , Pk−1 ||Pk | q) ≤ Df (P1 , . . . , Pk−1 ||Pk ) .
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK
9
We also see that if q1 and q2 are quantizers of X , and q1 induces a finer
partition of X than q2 , meaning that for x, x′ ∈ X the equality q1 (x) = q1 (x′ )
implies q2 (x) = q2 (x′ ), we have
Df (P1 , . . . , Pk−1 ||Pk | q2 ) ≤ Df (P1 , . . . , Pk−1 ||Pk | q1 ) .
This type of ordering is central to this work: any multiclass loss L induces
a unique f -divergence, and consistency of discriminants γ : X → Rk for a
loss L after quantization is intimately tied to the preservation (and relative
ordering) of information as related to the quantized risk (2).
3. Risks, information measures, and f -divergences. Having reviewed the basic properties of f -divergences, we turn to a more detailed
look at their relationships with multi-way hypothesis tests, multi-class classification, generalized entropies, and statistical informations relating multiple distributions. We build a correspondence between these that parallels
that for binary experiments and classification problems [19, 23, 27].
We first recapitulate the probabilistic model for classification and Bayesian
hypothesis testing problems from the introduction. We have a prior π ∈ ∆k
and probability distributions P1 , . . . , Pk defined on a set X . The coordinate
Y ∈ [k] is drawn according to a multinomial with probabilities π, and conditional on Y = y, we draw X ∼ Py . Following DeGroot [11], we refer to this as
an experiment. Associated with any experiment is a family of informations
as follows. Let π
e beP
the posterior distribution on Y given observation X = x,
π
ei (x) = πi dPi (x)/( kj=1 πj dPj (x)). Given any closed concave H : Rk+ → R,
which we refer to as generalized entropy (see [15, § 3.3]; DeGroot [11] calls
H an uncertainty function), the information associated with the experiment
is the reduction of entropy (uncertainty) from prior to posterior (3),
IH (X; π) = H(π) − E[H(e
π (X))].
P
The expectation is taken over X ∼ ki=1 πi Pi . That IH (X; π) ≥ 0 is immediate by concavity; DeGroot [11, Thm. 2.1] shows that IH (X; π) ≥ 0 for all
distributions P1 , . . . , Pk and priors π if and only if H is concave on ∆k .
In this section we develop equivalence results between multiclass classification losses and risk, multi-way f -divergences, and entropy measures.
Concretely, consider L : Rk × [k] → R, and recall the risk (1), defined
as RL (γ) = E[L(γ(X), Y )], where γ ∈ Γ, the set of measurable functions
γ : X → Rk . As in equation (4) in the introduction,
each loss L induces the
P
entropy HL on ∆k via HL (π) = inf α∈Rk ki=1 πi L(α, i), also called the pointwise Bayes risk [15, 13, 27, 35]. In Section 3.1, we give an explicit inverse
mapping showing how each generalized entropy H is induced by (at least
10
DUCHI, KHOSRAVI, RUAN
one) convex loss function L, i.e., L(·, i) is convex for each i. In Section 3.2,
we illustrate consistency properties the entropy H implies about the convex
loss L inducing it. We connect these results in Section 3.3 with multi-way
f -divergences. For any loss L and associated entropy/Bayes risk HL , for all
k−1
→ R with fL,π (1) = 0
π ∈ ∆k there exists a convex function fL,π : R+
such that the gap between the prior Bayes L-risk—the best expected loss
attainable without observing X—and the posterior Bayes risk inf γ RL (γ) is
HL (π) − inf RL (γ) = HL (π) − E[HL (e
π (X))] = DfL,π (P1 , . . . , Pk−1 ||Pk )
γ∈Γ
(see [15, 13]). The inverse direction is new, and given any closed convex
k−1
→ R with f (1) = 0, we construct convex losses L(·, i), an
function f : R+
associated generalized entropy HL , and prior π = 1/k ∈ ∆k satisfying
Df (P1 , . . . , Pk−1 ||Pk ) = inf
α∈Rk
k
X
i=1
πi L(α, i) − inf RL (γ).
γ∈Γ
3.1. Generalized entropies and losses. We construct a natural bidirectional mapping between losses and generalized entropies, giving a few examples to illustrate. For any loss L : Rk × [k] → R, the construction (4) of
HL yields a closed concave function, as HL is the infimum of linear functionals of π. It is thus a generalized entropy [15] (or uncertainty function [11]),
and the gap HL (π) − E[HL (e
π (X))] between prior and posterior entropy is
non-negative. The following two examples with zero-one loss are illustrative.
Example 1 (Zero-one loss).
Consider the zero one loss
Lzo (α, y) = 1 {αy ≤ αj for some j 6= y}
where y ∈ [k]. Then we have
)
( k
X
πi 1 {αi ≤ αj for some j 6= i} = 1 − max πj .
HLzo (π) = inf
α
i=1
j
This generalized entropy is concave, nonnegative, and satisfies HLzo (π) = 0
if and only if π = ei for a standard basis vector ei . ⋄
Example 2 (Cost-weighted classification). In some scenarios, we allow
different costs for classifying certain classes y as others; for example, it may
be less costly to misclassify a benign tumor as cancerous than the opposite.
k×k
, where cyi ≥ 0 is the
In this case, we use a matrix C = [cyi ]ky,i=1 ∈ R+
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 11
cost for classifying an observation of class y as class i (i.e. assigning X ∼ Pi
instead of Py in the experiment). We assume cyy = 0 for each y and define
Lcw (α, y) = max{cyi | αi = max αj }, α ∈ Rk ,
(10)
i
j
the maximal loss for those indices of α attaining maxj αj . Let C = [c1 · · · ck ]
be the column representation of C. If cTy π = minl cTl π, then by choosing any
α such that αy > αj for all j 6= y, we have
HLcw (π) = inf
α
X
k
y=1
πy max{cyi | αi = max αj } = min π T cl .
i
j
l
The entropy HLcw is concave, nonnegative, and satisfies HLcw (ei ) = 0 for
standard basis vectors ei ; Example 1 corresponds to C = 11T − Ik×k . ⋄
The forward mapping (4) from losses L to entropy HL is straightforward,
though it is many-to-one. Using convex duality and conjugacy arguments, we
can show an inverse mapping. This construction is new, though precursors
for proper scoring rules and predictions in the probability simplex exist [15,
14, Thm. 2]; these characterize proper scoring rules, but it is not always clear
how to generate convex losses from these. Before stating the proposition, we
recall the definition (6) of the Fenchel conjugate f ∗ (s) = supt {sT t − f (t)}.
Proposition 3.
For any closed concave H : ∆k → R, the losses
L(·, i) : Rk → R,
(11)
L(α, i) = −αi + (−H)∗ (α),
i ∈ {1, . . . , k}, are closed, convex, and satisfy the equality (4) that H ≡ HL .
Proof. Standard Fenchel conjugacy relationships [17, Chapter X] imply
H(π) = inf −π T α + (−H)∗ (α) where (−H)∗ (α) = sup αT π + H(π) .
α∈Rk
π∈∆k
Defining L(α, i) = −αi + (−H)∗ (α) for i = 1, . . . , k, we can write
H(π) = inf −π T α + (−H)∗ (α) = inf −π T α + π T 1 · (−H)∗ (α)
α∈Rk
α∈Rk
= inf
α∈Rk
X
k
i=1
πi L(α, i) .
Proposition 3 shows that associated with every concave entropy defined on
the simplex, there is at least one set of convex loss functions L(·, i) generating
12
DUCHI, KHOSRAVI, RUAN
the entropy via the infimal representation (4), and there is thus a mapping
from loss functions to entropies and from entropies to convex losses: given
any loss L, we may construct a convex loss Lcvx with HL = HLcvx . The
mapping from entropies H to loss functions generating H as in (4) is one-tomany, as any losses L(1) and L(2) with the same range satisfy HL(1) = HL(2) .
3.2. Surrogate risk consistency and generalized entropies. Our construction (11) of loss functions is a somewhat privileged construction, as it often
yields desirable properties of the convex loss function itself, especially as
related to the non-convex zero-one loss. Indeed, it is often the case that the
convex loss L so generated is Fisher consistent; to make this explicit, we
recall the following definition [36, 31].
Definition 3.1. Let L : Rk × [k] → R. Then L is classification calibrated for the zero-one loss if for any π ∈ ∆k and i∗ such that πi∗ < maxj πj ,
(12)
inf
α∈Rk
( k
X
)
πi L(α, i)
i=1
< inf
α∈Rk
(
k
X
i=1
πi L(α, i) : αi∗ ≥ max αj
j
)
.
k×k
as in Example 2, L is classification calibrated for
Given a matrix C ∈ R+
the cost matrix C if for any π ∈ ∆k and any i∗ with cTi∗ π > minj cTj π,
(13)
inf
α∈Rk
( k
X
)
πi L(α, i)
i=1
< inf
α∈Rk
(
k
X
i=1
πi L(α, i) : αi∗ ≥ max αj
j
)
.
Tewari and Bartlett [31, Thm. 2] and Zhang [36, Thm. 3] show the importance of Definition 3.1: let R(γ) be the zero-one or cost-weighted risk
(Exs. 1–2). If L is lower-bounded, then it is classification calibrated (with
respect to zero-one or the cost-weighted loss) if and only if for any sequence
γn : X → Rk and distribution P on X × Y we have Fisher consistency, i.e.
RL (γn ) → inf RL (γ) implies R(γn ) → inf R(γ).
γ∈Γ
γ∈Γ
That is, classification calibration (with respect to zero-one-risk or the costweighted risk) is equivalent to surrogate risk consistency of the loss L. Because of the predominance of the zero-one loss in the literature, we use
“classification calibration” without any qualification to mean “classification
calibration with respect to zero-one loss”.
We now show how—under minor restrictions on the generalized entropy
function H—the construction (11) yields classification calibrated losses.
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 13
Definition 3.2. A convex function f : Rk → R is (λ, κ, k·k)-uniformly
convex over C ⊂ Rk if it is closed and for all t ∈ [0, 1] and x1 , x2 ∈ C
f (tx1 + (1 − t)x2 )
≤ tf (x1 ) + (1 − t)f (x2 ) −
λ
t(1 − t) kx1 − x2 kκ (1 − t)κ−1 + tκ−1 .
2
We say, without qualification, that f is uniformly convex on C if dom f ⊃ C
and there exist λ > 0, a norm k·k, and constant κ < ∞ such that Definition 3.2 holds; we say f is uniformly concave if −f is uniformly convex.
Definition 3.2 is an extension of the usual notion of strong convexity, which
holds when κ = 2, and is essentially a quantified notion of strict convexity.
With this definition, we have the following two propositions. These two
propositions, whose proofs we provide in Appendix A, show that generalized
entropies naturally give rise to classification calibrated loss functions; we
provide examples of these results in Section 3.4 to come.
Proposition 4. Assume that H is closed concave, symmetric, and has
dom H = ∆k , and let L have definition
(11). Additionally assume that (a)
P
H is strictly concave, and inf α ki=1 πi L(α, i) is attained for all π ∈ ∆k , or
(b) H is uniformly concave. Then L is classification calibrated.
Even when H is not strictly concave, we can give classification calibration
results. Indeed, recall Example 1, which showed that for the zero-one-loss,
we have HL (π) = 1 − maxj πj .
Proposition 5. Let H(π) = 1 − maxj πj . The loss (11) defined by
L(α, i) = −αi + (−H)∗ (α) is classification calibrated. Moreover, we have for
any π ∈ ∆k and α ∈ Rk that
k
X
i=1
πi L(α, i) − inf
α
k
X
i=1
πi L(α, i) ≥
1
k
X
k
i=1
zo
πi L (α, i) − inf
α
k
X
i=1
πi L (α, i) .
zo
3.3. Divergences, risk, and generalized entropies. In this section, we show
that f -divergences as in Definition 2.1 have a precise correspondence with
generalized entropies and losses; Garcı́a-Garcı́a and Williamson [13] establish the correspondence between f -divergences and entropy/pointwise Bayes
risk H; our results show the important link from f directly to the loss L.
We begin as in equation (4) P
with a concave generalized entropy H and loss
L satisfying H(π) = inf α∈Rk ki=1 πi L(α, i); by Proposition 3 it is no loss of
14
DUCHI, KHOSRAVI, RUAN
generality to assume this correspondence. Let Γ be the collection of measurable functions γ : X → Rk . The posterior Bayes risk for L is
(14)
HL (π, P1:k ) := inf
Z X
k
γ∈Γ X
i=1
πi L(γ(x), i)dPi (x) = E[HL (e
π (X))],
where π
e(x) is the posterior distribution on Y conditional on X = x. The
information measure (3) is thus the gap between the prior Bayes L-risk and
posterior Bayes L-risk. We may then write
inf
α∈Rk
=
Z
k
X
i=1
X
πi L(α, i) − inf RL (γ) = HL (π) − HL (π, P1:k ) = IHL (X; π)
γ∈Γ
sup HL (π) −
α
k−1
X
i=1
!
dPi
πi L(α, i)
− πk L(α, k) dPk = DfL,π (P1:k−1 ||Pk ) ,
dPk
k−1
→ R has definition
where the closed convex function fL,π : R+
!
k−1
X
πi L(α, i)ti − πk L(α, k) .
(15)
fL,π (t) := sup HL (π) −
α∈Rk
i=1
As fL,π is the supremum of affine functions of its argument t, it is closed
convex, and fL,π (1) = HL (π) − HL (π) = 0. That is, equation (15) shows
that given any loss L or generalized entropy H, the information measure
IHL (X; π), gap between prior and posterior L-risk, and fL,π -divergence between distributions P1 , . . . , Pk−1 and Pk are identical.
We can also give a converse result that shows that every f -divergence
can be written as the gap between prior and posterior risks for a convex
loss function. We first recall the result that Df (P1:k−1 ||Pk ) is a statistical
information (3) based on an generalized entropy H associated with f . Except
for the closure operation, this result is known [13, Thm. 3].
For closed convex f : Rk−1 → R with f (1) = 0, let
t1
tk−1
H(t1 , . . . , tk ) = −ktk f
,...,
,
tk
tk
Proposition 6.
where we implicitly use the closure of the perspective (Def. (9)). Then
Df (P1 , . . . , Pk−1 ||Pk ) = H(1/k) − E[H(e
π (X))],
where the prior π = 1/k and the expectation is taken according to
P
i
πi Pi .
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 15
By combining Propositions 3 and 6 with the infimal representation (4)
of HL , we immediately obtain the following corollary, which is our explicit
construction of a closed convex loss from an f -divergence.
Corollary 3. Let π 0 = 1/k. For any closed and convex function f :
→ R such that f (1) = 0, the convex losses defined by
π1
πk−1
T
L(α, i) = −αi + sup π α − kπk f
,...,
πk
πk
π∈∆k
Rk−1
satisfy Eq. (15), i.e. f (t) = supα {HL (π 0 ) −
Additionally,
Df (P1:k−1 ||Pk ) = inf
α∈Rk
k
X
i=1
Pk−1
i=1
πi0 L(α, i)ti − πk0 L(α, k)}.
πi0 L(α, i) − inf E[L(γ(X), Y )],
γ
where the expectation is over Y ∼ π 0 and X ∼ Py conditional on Y = y.
For binary classification problems, Nguyen et al. [23, Thm. 1] provide an
explicit construction of a closed convex margin-based loss inducing the f divergence as in Eq. (15); the binary case allows a complete characterization
of all such convex functions, which appears difficult in the multiclass case.
Corollary 3, coupled with the information representation given by the f divergence (15), shows the complete equivalence between f -divergences, loss
functions L, and entropies H. For any f -divergence, there exists a loss function L and prior π = 1/k such that Df (P1:k−1 ||Pk ) = HL (π) − HL (π, P1:k ).
Conversely, for any loss function L and prior π, there exists a multi-way
f -divergence such that the gap HL (π) − HL (π, P1:k ) = Df (P1:k−1 ||Pk ).
3.4. Examples of generalized entropies and loss correspondences. To complement our general results, we illustrate the correspondence between (concave) generalized entropies and the loss construction (11) through several
examples, using Propositions 4 and 5 to guarantee classification calibration.
Example 3 (Zero-one loss, Example 1, continued). We use the generalized entropy H(π) = 1 − maxj πj generated by the zero-one loss to derive a
convex loss function L that gives the same entropy via the representation (4).
The conjugate to −H is
α(1) + α(2) 1
(16) (−H) (α) = 1 + max α(1) − 1,
− ,...,
2
2
∗
Pk
i=1 α(i)
k
1
−
,
k
16
DUCHI, KHOSRAVI, RUAN
where α(1) ≥ α(2) ≥ · · · are the entries of α ∈ Rk in sorted order (see § A.5
for a proof). Then the convex “family-wise” loss, named for its similarity to
family-wise error control in hypothesis tests,
X
l
1
1
Lfw (α, i) = 1 − αi + max
α(j) −
l
l
l∈{1,...,k}
j=1
generates the same entropy HLfw and associated f -divergence as the zeroone loss. Moreover, Proposition 5 guarantees that Lfw is classification calibrated (Def. 3.1). It appears that the loss Lfw is a new convex classificationcalibrated loss function. ⋄
Rather than re-considering Example 2, which we do later in the context of
showing that distinct convex losses can yield the same generalized entropy,
we now consider the multiclass logistic loss. The loss does not correspond
to the zero-one loss, but it generates Shannon entropy and information.
Example
4 (Logistic loss and entropy). For 1 ≤ i ≤ k, define pi (α) =
Pk
α
j
j=1 e . The multi-class logistic loss is then
eαi /
L(α, i) = − log pi (α) = log
X
k
j=1
eαj −αi
for 1 ≤ i ≤ k.
The entropy associated with the loss is the familiar Shannon entropy,
X
k
k
X
πi log πi .
πi log pi (α) = −
(17)
HL (π) = inf
−
α∈Rk
i=1
i=1
The conjugacy calculation (11) (i.e. our inverse construction from H to loss
L) to generate L also yields the multi-class logistic loss. That the multiclass
logistic loss is calibrated for the zero-one loss [36, § 4.4] is now immediate: the negative Shannon entropy is strongly convex over the simplex ∆k
(this is Pinsker’s inequality [9, Ch. 17.3]), so the fact that logistic loss and
Shannon entropy are dual via Eq. (11) and Proposition 4 yield calibration.
The information measure (3) associated with the logistic loss is the mutual
information between the observation X and label Y . Indeed, we have
Z
H(Y | X = x)dP (x)
IH (X; π) = H(π) − E[H(e
π (X))] = H(Y ) −
X
= H(Y ) − H(Y | X) = I(X; Y )
P
where H denotes the Shannon entropy, P = ki=1 πi Pi , and I(X; Y ) is the
usual (Shannon) mutual information between X and Y . ⋄
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 17
We include one final example to show that in some instances, many different convex losses can yield the same generalized entropy H.
Example 5 (Hinge losses). Define the pairwise multiclass hinge loss
X
Lhin (α, i) =
[1 + αj ]+ + I 1T α = 0 .
j6=i
We also consider the slight extension to weighted loss functions to address
asymmetric losses of the form (10) from Example 2. In this case, given the
k×k
, we set
loss matrix C ∈ R+
Lhin (α, i) =
k
X
j=1
P
cij [1 + αj ]+ + I 1T α = 0 .
The loss L(α, i) = j6=i cij [1 + αj − αi ]+ yields a completely identical set
of calculations without the restriction 1T α = 0, as it is invariant to shifts.
A calculation (see § A.6 for completeness) shows the generalized entropy (4)
associated with the hinge loss with loss matrix C = [c1 · · · ck ] is
)
( k
X
πi Lhin (α, i) = k min π T cl .
(18)
HLhin (π) = inf
α∈Rk
l
i=1
Such losses satisfy a number of classification calibration guarantees; we
note one, essentially due to Zhang [36, Thm. 8]. For completeness, we include
a proof in Appendix A.4.
Observation 1. Let φ : R → R be any bounded belowPconvex function,
differentiable on (−∞, 0], with φ′ (0) < 0. Then L(α, y) = ki=1 cyi φ(−αi ) is
classification calibrated for the cost matrix C (Def. 3.1, Eq. (13)).
Taking C = 11T − Ik×k , we see that the hinge loss is calibrated for the
k×k
, the weighted hinge loss is
zero-one loss (Ex. 1); taking arbitrary C ∈ R+
calibrated for the cost matrix C. Even more, we have the following quantitative calibration guarantee in analogy with Proposition 5:
k
X
i=1
πi L
hin
(α, i) − inf′
α
k
X
i=1
πi L
hin
′
(α , i) ≥
k
X
i=1
cw
πi L (α, i) − inf′
α
k
X
πi Lcw (α′ , i)
i=1
for all π ∈ ∆k and α ∈ Rk , strengthening Observation 1. (We prove this as
Lemma A.9 in § A.7.) In the binary case [23, Lem. 2], similar quantitative
guarantees hold for any margin-based classification calibrated loss L for
which HL = HLzo ; we do not know if this extends to the multiclass case. ⋄
18
DUCHI, KHOSRAVI, RUAN
4. Comparison of loss functions. In Section 3, we demonstrated the
correspondence between loss functions, generalized entropies, statistical information, f -divergences, and (in restricted cases) classification calibration.
These correspondences assume that decision makers have access to the entire observation X, which is often not the case; as noted in the introduction,
it is often beneficial to pre-process data to make it lower dimensional, communicate or store it efficiently, or to improve statistical behavior. Thus, we
now explore the impact quantization has on these concepts.
fw
To motivate this further, consider that each of the family-wise
P loss L of
Ex. 3, logistic loss (Ex. 4), and any loss of the form L(α, y) = i6=y φ(−αi )
for φ convex and decreasing with φ′ (0) < 0 (Ex. 5, Observation 1) is classification calibrated. This relates to one of the major criticisms of classification calibration: if the Bayes classifier (minimizer of risk over all functions
X → Rk ) does not belong to the class of functions considered, classification calibration says little. In this context, we shed light on this issue by
identifying losses that are consistent (calibrated) even with the additional
selection of quantizer or data representation—a restriction of the class of
possible functions as in the implication (5) in the introduction.
4.1. A model of quantization and experimental design. Abstractly, we
treat the design of an experiment or choice of data representation as a quantization problem, where a quantizer q maps the space X to a measurable
space Z. Then, for a loss L, prior π ∈ ∆k on the label Y , and discriminant
γ : Z → Rk , we consider the quantized risk (2), which we recall is
RL,π (γ | q) := E [L(γ(q(X)), Y )] .
Given class-conditional distributions P1:k (equivalently, hypotheses Hi : Pi
in the Bayesian testing setting) and collection Q of quantizers, our criterion
is to choose the quantizer q that allows the best attainable risk. That is, we
consider the quantized Bayes L-risk, defined as the infimum of the risk (2)
over discriminants Γ = {γ : Z → Rk },
(19)
inf RL,π (γ | q) =
γ∈Γ
Z
inf
k
X
Z α i=1
πi L(α, i)dPiq (z)
where P q (A) = P (q−1 (A)) denotes the push-forward measure. The risk (19)
measures the best attainable risk for a fixed choice of q ∈ Q; one thus seeks
the design q giving the lowest quantized Bayes L-risk.
Whether for computational or analytic reasons, minimizing the loss (19)
is often intractable; the zero-one loss Lzo (Ex. 1), for example, is non-
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 19
convex and discontinuous. It is thus of interest to understand the asymptotic consequences of using a surrogate loss L in place of the desired loss
(say Lzo ) [36, 21, 2, 31], including the setting in which one incorporates further dimension reduction via the choice q ∈ Q. Nguyen et al. [23] introduce
and study this problem for binary classification, giving a correspondence between f -divergences, loss functions, and surrogate consistency with quantization. The consequences of using a surrogate for consistency of the resulting
quantization and classification procedure in the multiclass case are a-priori
unclear: we do not know when using such a surrogate can be done without penalty. To that end, we now characterize when two loss functions L(1)
and L(2) provide equivalent criteria for choosing quantizers (experimental
designs or data representations) according to the Bayes L-risk (19).
4.2. Universal Equivalence of Loss Functions. Recalling our arguments
in Section 3.3 that statistical information (the gap between prior and posterior risks) is a multi-way f -divergence between distributions P1 , . . . , Pk−1
and Pk , we give a quantized version of this construction. In analogy with
the results of Section 3.3, the quantized statistical information is
IHL (X; π | q) := HL (π) − E[HL (e
π (q(X)))]
(20)
k
X
= inf
α∈Rk
i=1
πi L(α, i) − inf RL,π (γ | q) = DfL,π (P1 , . . . , Pk−1 ||Pk | q) ,
γ
P
where HL (π) = inf α∈Rk ki=1 πi L(α, i) as in (4), the convex function fL,π is
defined as in expression (15) and does not depend on the quantizer q, and
π
e(q(X)) denotes the posterior distribution on Y ∈ [k] conditional on observing q(X). We extend Nguyen et al. [23]’s notion of universal equivalence
from the binary case, defining losses as equivalent if they induce the same
ordering of quantizers q under the information measure (20).
Definition 4.1. Loss functions L(1) and L(2) are universally equivalent
u
for the prior π, denoted L(1) ≡π L(2) , if for any distributions P1 , . . . , Pk on
X and quantizers q1 and q2
I (X, π; HL(1) | q1 ) ≤ I (X, π; HL(1) | q2 ) if and only if
I (X, π; HL(2) | q1 ) ≤ I (X, π; HL(2) | q2 ) .
Definition 4.1 evidently is equivalent to the ordering condition
(21)
inf RL(1) ,π (γ | q1 ) ≤ inf RL(1) ,π (γ | q2 ) if and only if
γ
γ
inf RL(2) ,π (γ | q1 ) ≤ inf RL(2) ,π (γ | q2 ),
γ
γ
20
DUCHI, KHOSRAVI, RUAN
for all distributions P1 , . . . , Pk , on the quantized Bayes L-risk (19). This definition is somewhat stringent: losses are universally equivalent only if they induce the same quantizer ordering for all population distributions. If a quantizer q1 is finer than q2 , all losses yield I (X, π; HL | q2 ) ≤ I (X, π; HL | q1 )
by the data processing inequality (Corollary 2 of Section 2). The stronger
equivalence notion is important for nonparametric classification settings in
which the underlying distribution on (X, Y ) is only weakly constrained and
neither of a pair of quantizers q1 , q2 ∈ Q is finer than the other.
Definition 4.1 and the representation (20) suggest that the entropy function HL associated with the loss L through the infimal representation (4)
and the f -divergence associated with L via the construction (15) are important for the equivalence of two loss functions. This is indeed the case. First,
we have the following result on universal equivalence of loss functions based
on their associated entropies.
Theorem 1. Let L(1) and L(2) be bounded below losses and HL(1) and
HL(2) be the associated generalized entropies as in the construction (4). Then
L(1) and L(2) are universally equivalent with respect to all priors π if and
only if there exist a > 0, b ∈ Rk , and c ∈ R such that for all π ∈ ∆k ,
HL(1) (π) = aHL(2) (π) + bT π + c.
We can also characterize universal equivalence for a prior π.
Theorem 2. Let π ∈ ∆k and as in Theorem 1, let L(1) and L(2) be
(1)
(2)
bounded below loss functions, with fπ and fπ the associated f -divergences
as in the construction (15). Then L(1) and L(2) are universally equivalent
with respect to the prior π if and only if there exist a > 0, b ∈ Rk−1 , and
c ∈ R such that
(22)
fπ(1) (t) = afπ(2) (t) + bT t + c
k−1
.
for all t ∈ R+
Nguyen et al. [23, Thm. 3] prove Theorem 2 for binary classification problems
(k = 2), using convex-conjugacy arguments. We outline our proofs (which
apply for arbitrary k and so require a different set of tools) in Section 5.
4.3. Consistency of empirical risk minimization. A major application of
these theorems is to show that certain non-convex loss functions (such as the
zero-one loss) are universally equivalent to convex loss functions, including
variants of the hinge loss, by showing that their associated entropies are
scalar multiples. As a first application of Theorems 1 and 2, however, we
consider the Bayes consistency of empirical risk minimization for selecting a
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 21
discriminant γ and quantizer q (in analogy with [23, Thm. 2]). In this case,
we receive a sample {(X1 , Y1 ), . . . , (Xn , Yn )} and define the empirical risk
n
X
bL,n (γ | q) := 1
L(γ(q(Xi )), Yi ).
R
n
i=1
Now, let Q1 ⊂ Q2 ⊂ · · · ⊂ Q be a non-decreasing collection of quantizers, indexed by sample size n, and similarly let Γ1 ⊂ Γ2 ⊂ · · · ⊂ Γ be a
non-decreasing collection of discriminant functions, where we assume the
collections satisfy the estimation and approximation error conditions
"
#
b
E
sup
RL,n (γ | q) − RL (γ | q) ≤ ǫest
n
γ∈Γn ,q∈Qn
(23)
RL (γ | q) − inf RL (γ | q) ≤ ǫapp
inf
n ,
γ∈Γn ,q∈Qn
γ∈Γ,q∈Q
app
→ 0 as n → ∞. Additionally, let R be the risk funcwhere ǫest
n → 0 and ǫn
tional for the cost-weighted misclassification loss Lcw (Example 2), where
Lcw (α, y) = maxi {cyi | αi = maxj αj }. Then we have the following result.
Theorem 3. Assume the conditions (23) and that γn and qn are approximate empirical L-risk minimizers satisfying
i
h
bL,n (γ | q) → 0 as n → ∞.
bL,n (γn | qn ) −
R
ǫopt
:=
E
R
inf
n
γ∈Γn ,q∈Qn
Let R⋆ (Q) = inf γ∈Γ,q∈Q R(γ | q). If the loss L is classification calibrated and
universally equivalent to the cost-weighted loss Lcw , then
L
R(γn | qn ) − R⋆ (Q) →1 0.
Theorem 3 guarantees that under the estimation and approximation conditions (23), empirical risk minimization is consistent for minimizing the
quantized Bayes risk whenever the loss L is classification calibrated and
equivalent to the desired loss. The proof of Theorem 3 reposes on the following risk inequality, which may be of independent interest. The lemma
is a consequence of the results on surrogate risk consistency for classification calibration [36, 31, 30] and our universal equivalence guarantees that
exhibits the power of calibration and universal equivalence.
Lemma 4.1. Assume L is classification-calibrated and universally equivk×k
.
alent to the weighted misclassification loss Lcw with cost matrix C ∈ R+
22
DUCHI, KHOSRAVI, RUAN
Then there exists a continuous concave function h with h(0) = 0 such that
⋆
R(γ | q) − inf R(γ | q) ≤ h RL (γ | q) − inf RL (q) .
γ∈Γ,q∈Q
q∈Q
Pk
T
With the choice L(α, y) =
i=1 cyi [1 + αi ]+ + I 1 α = 0 or L(α, y) =
Pk
1
i=1 cyi [1 + αi − αy ]+ , we may take h(ǫ) = (1 + k )ǫ, that is,
1
R(γ | q) − inf R(γ | q) ≤ 1 +
RL (γ | q) − inf RL (γ | q) .
γ∈Γ,q∈Q
γ∈Γ,q∈Q
k
Lemma 4.1 shows that the gap in surrogate risk provides a guaranteed upper bound on the true cost-weighted risk; in the case of the modified hinge
losses of Example 5, this gap is linear. In the binary case, even stronger
results are possible [23, Lemma 2]—one may take h(ǫ) = aǫ (for some
a < ∞) in Lemma 4.1 for any margin-based classification-calibrated loss
universally equivalent to the 0-1 loss—this relies on the specific form any
such binary convex loss must take [23, Eq. (9)]; our Examples 3 (the familywise loss) and 5 show that fairly different-looking losses can be classification
calibrated and universally equivalent to zero-one loss. We provide the proof
of Lemma 4.1 in § C.1. Theorem 3, which we prove in § C.2, is then a
consequence of this lemma and Theorem 1.
4.4. Examples of universal equivalence. In this section, we give several
examples that build off of Theorems 1 and 2, showing that there exist convex losses that allow optimal joint design of quantizers (or measurement
strategies) and discriminant functions, opening the way for potentially efficient convex optimization strategies. To that end, we give two hinge-like loss
functions that are universally equivalent to the zero-one loss for all prior distributions π. We also give examples of classification calibrated loss functions
that are not universally equivalent to the zero-one loss, although minimizing
them without quantization yields Bayes-optimal classifiers.
Example 6 (Cost-weighted
losses). We
Treturn to Example 5, where we
P
hin
have L (α, i) =
j6=i cij [1 + αj ]+ + I 1 α = 0 . In this case, we have
HLhin (π) = k minl π T cl = kHLcw (π), where Lcw denotes the cost-weighted
misclassification error as in Example 2. Theorem 1 immediately guarantees
that the (weighted) hinge loss is universally equivalent to the (weighted) 0-1
loss. The weighted hinge loss Lhin is also, as in Example 5, calibrated for
the cost-weighted misclassification error. ⋄
Example 7 (Max-type losses and zero-one loss). We return to ExamT
α +α
ple 3 and let Lfw (α, i) = 1−αi +max{α(1) −1, (1) 2 (2) − 21 , . . . , 1 k α − k1 }, the
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 23
convex family-wise loss. By Example 3, the associated entropy is HLfw (π) =
1 − maxj πj = HLzo for π ∈ ∆k , and Proposition 5 shows that Lfw is classification calibrated. We thus have that Lfw and the zero-one loss Lzo are
universally equivalent by Theorems 1 and 2. ⋄
For our final example, we consider the logistic loss, which is classification
calibrated but not universally equivalent to the zero-one loss.
P
Example 8 (Logistic loss). The loss Llog (α, i) = log( kj=1 eαj −αi ) has
P
(Shannon) entropy H(π) = − ki=1 πi log πi , as in Ex. 4. There are no a, b, c
such that HLzo (π) = 1 − maxj πj = aHLlog (π) + bT π + c for all π ∈ ∆k .
Theorem 1 shows that the logistic loss is not universally equivalent to the
zero-one loss. That is, in spite of its classification calibration, there are
distributions P1 , . . . , Pk , a collection Q of quantizers X → Z, and a sequence γn : Z → Rk such that RLlog (γn | qn ) → inf γ,q∈Q RLlog (γ | q), but
RLzo (γn | qn ) 6→ inf γ,q∈Q RLzo (γ | q). ⋄
5. Proof of the Theorems 1 and 2. The remainder of the main body
of the paper consists of the major parts of our arguments for Theorems 1
and 2. We divide the proof of the theorems into two parts. The “if” part is
straightforward; the “only if” is substantially more complex.
Proof (if direction). We give the proof for Theorem 2; that for Theorem 1 is
(1)
(2)
identical. Assume that dom fπ = dom fπ and there exist a > 0, b ∈ Rk−1 ,
and c ∈ R such that Eq. (22) holds. By Definition 2.1 of multi-way f divergences, for any quantizer q, we have
Df (1) (P1 , . . . , Pk−1 ||Pk | q) = aDf (2) (P1 , . . . , Pk−1 ||Pk | q) + bT 1 + c,
π
as
R
X
π
dPi = 1. Applying the relationship (20), we obtain
I (X, π; HL(1) | q) = aI (X, π; HL(2) | q) + bT 1 + c.
As a > 0, the universal equivalence of L(1) and L(2) follows immediately.
We turn to the “only if” part of the proofs of Theorems 1 and 2. A
roadmap is as follows: we first define what we call order equivalence of
convex functions, which is related to the equivalence of f -divergences and
generalized entropies (Def. 5.1). Then, for any two loss functions L(1) and
L(2) that are universally equivalent, we show that the associated entropies
HL(1) and HL(2) , as constructed in the infimal representation (4), and the
functions f (1) and f (2) generating the f -divergences via expression (15), are
24
DUCHI, KHOSRAVI, RUAN
order equivalent (Lemmas 5.1 and 5.2). After this, we provide a characterization of order equivalent closed convex functions (Lemma 5.3), which is the
linchpin of our analysis. The lemma shows that for any two order equivalent
closed convex functions f1 and f2 with dom f1 = dom f2 , there are parameters a > 0, b ∈ Rk , and c ∈ R such that f (1) (t) = af (2) (t) + bT t + c for
all t ∈ dom f1 = dom f2 . This proves the “only if” part of the Theorems 1
and 2, yielding the desired result. We present the main parts of the proof in
the body of the paper, deferring technical nuances to the supplement.
5.1. Universal equivalence and order equivalence. By Definition 4.1 (and
its equivalent variant stated (21)), universally equivalent losses L(1) and
L(2) induce the same ordering of quantized information measures and f divergences. The next definition captures this ordering slightly differently.
Definition 5.1. Let f1 : Ω → R and f2 : Ω → R be closed convex
functions, where Ω ⊂ Rk is closed convex. Let m ∈ N be arbitrary and the
matrices A, B ∈ Rk×m satisfy A1 = B1, where A has columns ai ∈ Ω and
B has columns bi ∈ Ω. Then f1 and f2 are order-equivalent if for all m ∈ N
and all such matrices A and B we have
(24)
m
X
j=1
f1 (aj ) ≤
m
X
j=1
f1 (bj ) if and only if
m
X
j=1
f2 (aj ) ≤
m
X
f2 (bj ).
j=1
As the above context suggests, order equivalence has strong connections
with universal equivalence of loss functions L and associated f -divergences
and generalized entropies. The next two lemmas make this explicit.
Lemma 5.1. If losses L(1) and L(2) are lower bounded and universally
equivalent, then the associated entropies of the construction (4) are order
equivalent over ∆k ⊂ Rk+ .
Proof. Let Hi be the entropy (pointwise Bayes risk) associated with
L(i) , noting that dom H1 = dom H2 = ∆k because inf π∈∆k Hi (π) > −∞.
k×m
k×m
satisfy ai , bi ∈ ∆k
and B ∈ R+
Let the matrices A = [a1 · · · am ] ∈ R+
1
1
A1
=
B1
∈
∆
for
each
i
=
1,
.
.
.
,
m,
and
let
v
=
k . We
m
Pm
Pm
Pm
Pmshow that
m
H
(a
)
≤
H
(b
)
if
and
only
if
H
(a
)
≤
j=1 1 j
j=1 1 j
j=1 2 j
j=1 H2 (bj ),
that is, expression (24) holds, by constructing appropriate distributions P1:k
and π, then applying the universal equivalence of L(1) and L(2) .
Let M0 be any integer large enough that v0 = k1 (1 + M10 )1 − M10 v ∈ Rk+ ,
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 25
so that v0 ∈ ∆k . Then define the vectors e
a 1 = v0 , . . . , e
amM0 = v0 , and let
k×M
and
Aext = [a1 · · · am e
a1 · · · e
amM0 ] ∈ R+
k×M
,
B ext = [b1 · · · bm e
a1 · · · e
amM0 ] ∈ R+
ext and
where M = (M0 + 1)m. These satisfy Aext 1 = B ext 1 = M
k 1. We let a
bext denote the columns of these extended matrices.
Now, let the spaces X = [M ] × [M ] and Z = [M ]. Define quantizers
q1 , q2 : X → Z by q1 (i, j) = i and q2 (i, j) = j. For l = 1, . . . , k, define the
distributions Pl on X by
M
M
X
k ext k X ext
k ext
k2
ext ext
Pl (i, j) =
bjl =
ail
a
Pl (i, j) = 2 · ail bjl , so
M
M
M
M il
j=1
j=1
P
k ext
and similarly i Pl (i, j) = M
bjl . Let π = k1 1 be the uniform prior distribution on the label Y ∈ {1, . . . , k}, and note that the posterior probability
#k
"
P
k
πl j Pl (i, j)
aext
−1
il
P
= P ext
= aext
∈ ∆k ,
π
e(q1 ({i})) = P
i
′
π
P
(i,
j)
a
′
′
′
j l
l l
l il
l=1
l=1
P
and similarly π
e(q−1
because
=
2 ({j})) =
j Pl (i, j) =
Pk
ext
bj ∈ ∆k . Taking the expectation over X ∼ l=1 πl Pl , we have
Pl (q−1
1 (i))
E[HL (e
π (q−1
1 (q1 (X))))] =
k ext
M ail ,
M
1 X
1X
−1
Pl (q−1
(i))H
(e
π
(q
(i)))
=
HL (aext
L
i ),
1
1
k
M
i=1
i,l
P
1 PM
ext
π (q−1
because l aext
2 (q2 (X))))] = M
j=1 Hπ (bj ).
il = 1. Similarly, E[HL (e
Recalling the definitions (3) and (20) of the (quantized)
information as1 PM
sociated with H, we have I (X, π; H | q1 ) = H(π) − M i=1 H(aext
i ) and
1 PM
ext
I (X, π; H | q2 ) = H(π) − M i=1 H(bi ). Then the universal equivalence
of losses L(1) and L(2) immediately implies for π = k1 1 that
M
M
1 X
1 X
ext
H1 (ai ) ≤ H1 (π) −
H1 (bext
H1 (π) −
i ) iff
M
M
H2 (π) −
1
M
i=1
M
X
i=1
H2 (aext
i ) ≤ H2 (π) −
1
M
i=1
M
X
H2 (bext
i ).
i=1
Noting that aext
= bext
for each i P
≥ m + 1, we rearrange the preceding
i
i
1
equivalent statements by adding M i≥m+1 H(aext
i ) to each side to obtain
that the Hi satisfy inequality (24).
26
DUCHI, KHOSRAVI, RUAN
For f -divergences, a parallel result is possible; as the techniques are similar
to those we use to prove Lemma 5.1 (by constructing an explicit discrete
space X and quantizers q), we defer the proof to Supplementary § B.1.
Lemma 5.2. If losses L(1) and L(2) are universally equivalent for the
prior π (Def. 4.1) and lower-bounded, the corresponding f -divergences fL(1) ,π
and fL(2) ,π of construction (15) are order equivalent.
5.2. Characterization of the order equivalence of convex functions. Lemmas 5.1 and 5.2 illustrate the intrinsic relationship between the universal
equivalence (Def. 4.1) of losses and the order equivalence (Def. 5.1) of their
associated generalized entropies and f -divergences. Therefore, it is natural
to ask when convex functions are order equivalent. The lemma below characterizes this order equivalence, and coupled with Lemmas 5.1 and 5.2, it
immediately implies Theorems 1 and 2.
Lemma 5.3. Let f1 , f2 : Ω → R be closed convex functions, where Ω ⊂
Rk is a convex set. Then f1 and f2 are order equivalent on Ω if and only if
there exist a > 0, b ∈ Rk , and c ∈ R such that for all t ∈ Ω
f1 (t) = af2 (t) + bT t + c.
(25)
While the proof of Lemma 5.3 is complex, we provide a partial proof
highlighting the most important parts of the argument, deferring technical
details to the supplement. The essential idea is that Lemma 5.3 holds for
simplices (and so it certainly holds for HL ); we can then cover any convex
set Ω with a number of overlapping simplices to extend the result to all of
Ω, which we do fully in Supplement B.2. To demonstrate Lemma 5.3 for
simplices, we require
Definition 5.2.
Vectors u0 , u1 , . . . um are affinely independent if
u1 − u0 , u2 − u0 , . . . , um − u0 ,
are linearly independent. A set E ⊂ Rk is a simplex if E = Conv{u0 , u1 , . . . , uk }
where u0 , . . . , uk are affinely independent.
Then the essential special case of Lemma 5.3 is the following result.
Lemma 5.4. Let E = Conv{u0 , . . . , uk } ⊂ Ω where u0 , . . . , uk are affinely
independent. If f1 and f2 are order equivalent, then there exist a > 0, b ∈ Rk ,
and c ∈ R such that
f1 (t) = af2 (t) + bT t + c for all t ∈ E.
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 27
The proof of Lemma 5.4 proceeds in a series of intermediate results, which
we provide in turn, deferring proofs to Supplement B.3. Our first step is to
argue that we need only prove equivalence results for convex functions on
dense subsets of their domains.
Lemma 5.5 ([17], Prop. IV.1.2.5). Let f1 , f2 : Ω → R be closed convex
and satisfy f1 (t) = f2 (t) for t in a dense subset of Ω. Then f1 = f2 on Ω.
The first technical lemma we prove is essentially a direct consequence of
the definition of order equivalence.
LemmaP5.6. Let u1 , . . . , um ∈ Ω, α ∈ Qm satisfy 1T α = 1, and v ∈ Ω
with v = m
i=1 αi ui . If f1 , f2 : Ω → R are order equivalent, then
m
X
i=1
αi f1 (ui ) ≤ f1 (v) if and only if
m
X
i=1
αi f2 (ui ) ≤ f2 (v).
Thus if α ∈ Qn satisfies 1T α = 1 and u1 , . . . , un ∈ Ω, then
!
!
n
n
n
n
X
X
X
X
αi f2 (ui ).
αi ui =
αi f1 (ui ) iff f2
αi ui =
(26)
f1
i=1
i=1
i=1
i=1
The next lemma shows that we can force equality (25) to hold for the k+1
extreme points and centroid of any simplex in Rk ; it is intuitive because there
are k + 2 free parameters in the choices of a > 0, b ∈ Rk , and c ∈ R.
Lemma 5.7. Let f1 , f2 : Ω → R be closed convex and let u0 , . . . , uk ∈ Ω
be affinely independent. There exist a > 0, b ∈ Rk , and c such P
that f1 (u) =
k
1
af2 (u) + bT u + c for u ∈ {u0 , . . . , uk , ucent }, where ucent = k+1
i=0 ui .
Lastly, we have the following characterization of the linearity of convex functions over convex hulls.
1
m
Lemma 5.8. Let f : Ω →
R be convex with u1 , . . . , um ∈ Ω and ucent =
P
m
1 Pm
u
.
If
f
(u
)
=
i
cent
i=1
i=1 f (ui ), then
m
!
m
m
X
X
T
λi f (ui ) for all λ ∈ Rm
λi ui =
f
+ with 1 λ = 1.
i=1
i=1
With the four lemmas 5.5–5.8, we can now prove Lemma 5.4. By rotating
with ui − u0 and shifting by u0 , it is no loss of generality to assume that the
28
DUCHI, KHOSRAVI, RUAN
functions fi are defined on V = {v ∈ Rk+ | 1T v ≤ 1}, so that f1 and f2 are
continuous, defined, convex, and order equivalent on V . We make one further
reduction. Let ei ∈ Rk for 1 ≤ i ≤ k be the standard basis for Rk P
and e0 = 0
1
be shorthand for the all-zeros vector. Further, let ecenter = k+1 ki=0 ei be
the centroid of V (so V = Conv{e0 , . . . , ek }). Lemma 5.7 guarantees the
existence of a > 0, b ∈ Rk , and c ∈ R such that
f1 (v) = af2 (v) + bT v + c for v ∈ {e0 , e1 , . . . , ek , ecenter }.
Now, let h1 (v) = f1 (v) and h2 (v) = af2 (v)+bT v+c, so h1 and h2 are convex,
order equivalent on V , and satisfy h1 (v) = h2 (v) for v ∈ {e0 , . . . , ek , ecenter }.
Thus, Lemma 5.4 is equivalent to showing that if h1 , h2 are convex, order
equivalent, and equal on the extreme points and centroid of V , then
(27)
h1 (v) = h2 (v) for v ∈ V = {v ∈ Rk+ | 1T v ≤ 1}.
We divide our discussion into two cases.
1 Pk
Linear case. Suppose that h1 (ecenter ) = k+1
i=0 h1 (ei ). Then by order
1 Pk
equivalence of h1 and h2 (Eq. (26)) we have h2 (ecenter ) = k+1
i=0 h2 (ei ).
Lemma 5.8 thus implies that h1 and h2 are linear on V = Conv{e0 , . . . , ek },
equal on the vertices of V , and hence equal on its interior.
1 Pk
Nonlinear case. By convexity we have h1 (ecenter ) < k+1
i=0 h1 (ei ), and
1 Pk
order equivalence (Lemma 5.6) implies h2 (ecenter ) < k+1 i=0 h2 (ei ). For
T
v ∈ V = Conv{e
0 , . . . , ek }, we use v0 = 1 − 1 v for shorthand, so we may
Pk
T
write v = i=0 vi ei and have [v0 v1 · · · vk ] ∈ ∆k+1 . Now, fix an arbitrary
v ∈ V ∩ Qk . We wish to show that h1 (v) = h2 (v). To that end we consider
consider the gaps due convexity of hj (ecenter ) to the values of hj (ei ) relative
to those from hj (v) to hj (ei ), defining the linear functions ϕj : [0, 1] → R by
"
#
#
" k
k
X
1 X
ϕj (r) := (1 − r) hj (ecenter ) −
vi hj (ei ) − hj (v)
hj (ei ) + r
k+1
i=0
i=0
for j = 1, 2. Then
k
1 X
hj (ei ) < 0
ϕj (0) = hj (ecenter ) −
k+1
i=0
by assumption, and by convexity,
ϕj (1) =
k
X
i=0
vi hj (ei ) − hj (v) ≥ 0.
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 29
The key is that the order equivalence of h1 and h2 on V implies that
(28)
sign(ϕ1 (r)) = sign(ϕ2 (r)) for r ∈ [0, 1],
so that ϕ1 and ϕ2 have the same zero crossing r ⋆ > 0, i.e. there exists
0 < r ⋆ ≤ 1 with ϕ1 (r ⋆ ) = ϕ2 (r ⋆ ) = 0. (We prove equality (28) presently.)
At this r ⋆ > 0, we find
0 = ϕ1 (r ⋆ ) − ϕ2 (r ⋆ ) = −r ⋆ h1 (v) + r ⋆ h2 (v),
where we use that h1 (ei ) = h2 (ei ) for i = 0, . . . , k and h1 (ecenter ) = h2 (ecenter ).
That is, h1 (v) = h2 (v), and as v ∈ V ∩ Qk is arbitrary and Qk is dense,
Lemma 5.5 extends the equality h1 = h2 to all of V . Expression (27) holds.
Returning to the sign equivalence (28), for r > 0, we may divide ϕj (r) by
r, and we have ϕj (r) ≤ 0 if and only if
"
#
k
k
X
1 X
1−r
vi hj (ei ) ≤ hj (v).
hj (ecenter ) −
hj (ei ) +
r
k+1
i=0
i=0
1−r
r(k+1)
∈ Q for i = 0, . . . , k and αk+1 = 1−r
r , the inequalPk
ity ϕj (r) ≤ 0 is equivalent to i=0 αi hj (ei ) + αk+1 hj (ecenter ) ≤ hj (v). A
P
calculation yields 1T α = 1 and ki=0 αi ei + αk+1 ecenter = v, and applying
Lemma 5.6 immediately yields that ϕ1 (r) ≤ 0 if and only if ϕ2 (r) ≤ 0 for all
r ∈ (0, 1] ∩ Q. Noting that ϕ1 (0) < 0 and ϕ2 (0) < 0, we obtain equality (28).
Defining αi = vi −
6. Discussion. Rather than recapitulating our contributions, we point
out a few directions we believe will prove interesting for further study. While
Corollary 1 shows that some convex losses are surrogate-risk consistent even
with restricted families of classifiers, it does not apply to the practical case
in which the collection of discriminants γ is a (convex subset of a) finitedimensional vector space. This longstanding problem certainly deserves further work. Another direction, a bit further afield, is to investigate the links
between this work and objective Bayesian approaches and reference priors [5, 4]. In this line of work, one has a family {Pθ }θ∈Θ of probability
models on an observation space X and before performing inference chooses
a prior π on θ to maximize Iπ (X; θ), the (Shannon) information between
X ∼ Pθ and θ ∼ π. For tasks other than minimizing log loss, it may be sensible to use a notion of information and entropy corresponding to the desired
loss. Our notions of loss equivalence, including construction of convex losses
equivalent to non-convex losses, could provide insight in such situations.
30
DUCHI, KHOSRAVI, RUAN
References.
[1] S. M. Ali and S. D. Silvey. A general class of coefficients of divergence of one distribution from another. Journal of the Royal Statistical Society, Series B, 28:131–142,
1966.
[2] P. L. Bartlett, M. I. Jordan, and J. McAuliffe. Convexity, classification, and risk
bounds. Journal of the American Statistical Association, 101:138–156, 2006.
[3] Y. Benjamini and Y. Hochberg. Controlling the false discovery rate: a practical and
powerful approach to multiple testing. Journal of the Royal Statistical Society, Series
B, 57(1):289–300, 1995.
[4] J. O. Berger. The case for objective Bayesian analysis. Bayesian Analysis, 1(3):
385–402, 2006.
[5] J. M. Bernardo. Reference analysis. In D. Day and C. R. Rao, editors, Bayesian
Thinking, Modeling and Computation, volume 25 of Handbook of Statistics, chapter 2,
pages 17–90. Elsevier, 2005.
[6] D. Bertsekas. Nonlinear Programming. Athena Scientific, 1999.
[7] P. Billingsley. Probability and Measure. Wiley, Second edition, 1986.
[8] D. Blackwell. Comparison of experiments. In Proceedings of the 2nd Berkeley Symposium on Probability and Statistics, pages 93–102. University of California Press,
1951.
[9] T. M. Cover and J. A. Thomas. Elements of Information Theory, Second Edition.
Wiley, 2006.
[10] I. Csiszár. Information-type measures of difference of probability distributions and
indirect observation. Studia Scientifica Mathematica Hungary, 2:299–318, 1967.
[11] M. H. DeGroot. Uncertainty, information, and sequential experiments. Annals of
Mathematical Statistics, 33(2):404–419, 1962.
[12] J. C. Duchi, L. Mackey, and M. I. Jordan. The asymptotics of ranking algorithms.
Annals of Statistics, 41(5):2292–2323, 2013.
[13] D. Garcı́a-Garcı́a and R. C. Williamson. Divergences and risks for multiclass experiments. In Proceedings of the Twenty Fifth Annual Conference on Computational
Learning Theory, 2012.
[14] T. Gneiting and A. Raftery. Strictly proper scoring rules, prediction, and estimation.
Journal of the American Statistical Association, 102(477):359–378, 2007.
[15] P. Grünwald and A. P. Dawid. Game theory, maximum entropy, minimum discrepancy, and robust Bayesian decision theory. Annals of Statistics, 32(4):1367–1433,
2004.
[16] L. Györfi and T. Nemetz. f -dissimilarity: A generalization of the affinity of several
distributions. Annals of the Institute of Statistical Mathematics, 30:105–113, 1978.
[17] J. Hiriart-Urruty and C. Lemaréchal. Convex Analysis and Minimization Algorithms
I & II. Springer, New York, 1993.
[18] T. Kailath. The divergence and Bhattacharyya distance measures in signal selection.
IEEE Transactions on Communication Technology, 15(1):52–60, 1967.
[19] F. Liese and I. Vajda. On divergences and informations in statistics and information
theory. IEEE Transactions on Information Theory, 52(10):4394–4412, 2006.
[20] M. Longo, T. D. Lookabaugh, and R. M. Gray. Quantization for decentralized hypothesis testing under communication constraints. IEEE Transactions on Information Theory, 36(2):241–255, 1990.
[21] G. Lugosi and N. Vayatis. On the Bayes-risk consistency of regularized boosting
methods. Annals of Statistics, 32(1):30–55, 2004.
[22] A. W. Marshall, I. Olkin, and B. C. Arnold. Inequalities: Theory of Majorization and
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 31
Its Applications. Springer, second edition, 2011.
[23] X. Nguyen, M. J. Wainwright, and M. I. Jordan. On surrogate loss functions and
f -divergences. Annals of Statistics, 37(2):876–904, 2009.
[24] F. Österreicher and I. Vajda. Statistical information and discrimination. IEEE Transactions on Information Theory, 39(3):1036–1039, 1993.
[25] H. V. Poor and J. B. Thomas. Applications of ali-silvey distance measures in the
design of generalized quantizers for binary decision systems. IEEE Transactions on
Communications, 25(9):893–900, 1977.
[26] F. Pukelsheim. Optimal Design of Experiments. Classics in Applied Mathematics.
SIAM, 1993.
[27] M. Reid and R. Williamson. Information, divergence, and risk for binary experiments.
Journal of Machine Learning Research, 12:731–817, 2011.
[28] H. Robbins. Some aspects of the sequential design of experiments. Bulletin American
Mathematical Society, 55:527–535, 1952.
[29] R. E. Schapire and Y. Freund. Boosting: Foundations and Algorithms. MIT Press,
2012.
[30] I. Steinwart. How to compare different loss functions. Constructive Approximation,
26:225–287, 2007.
[31] A. Tewari and P. L. Bartlett. On the consistency of multiclass classification methods.
Journal of Machine Learning Research, 8:1007–1025, 2007.
[32] N. Tishby, F. Pereira, and W. Bialek. The information bottleneck method. In The
37’th Allerton Conference on Communication, Control, and Computing, 1999.
[33] J. N. Tsitsiklis. Decentralized detection. In Advances in Signal Processing, Vol. 2,
pages 297–344. JAI Press, 1993.
[34] I. Vajda. On the f -divergence and singularity of probability measures. Periodica
Mathematica Hungarica, 2(1–4):223–234, 1972.
[35] R. C. Williamson, E. Vernet, and M. D. Reid. Composite multiclass losses. Journal
of Machine Learning Research, to appear, 2016.
[36] T. Zhang. Statistical analysis of some multi-category large margin classification methods. Journal of Machine Learning Research, 5:1225–1251, 2004.
32
DUCHI, KHOSRAVI, RUAN
APPENDIX A: PROOFS OF CLASSIFICATION CALIBRATION
RESULTS
In this section, we prove Propositions 4 and 5. Before proving the propositions proper, we state several technical lemmas and enumerate continuity
properties of Fenchel conjugates that will prove useful. We also collect a few
important definitions related to convexity and norms here, which we use
without comment in this appendix. For a norm k·k on Rk , we recall the definition of the dual norm k·k∗ as kyk∗ = supkuk≤1 uT y. For a convex function
f : Rk → R, we let
∂f (u) = {g ∈ Rk | f (v) ≥ f (u) + g T (v − u) for all v ∈ Rk }
denote the subgradient set of f at the point u. This set is non-empty if
u ∈ rel int dom f (see [17, Chapter VI]).
A.1. Technical preliminaries. We provide some background on convex functions. We recall Definition 3.2 of uniform convexity, that f is (λ, κ, k·k)uniformly convex over C ⊂ Rk if it is closed and for all t ∈ [0, 1] and
u0 , u1 ∈ C we have
(29)
f (tu0 + (1 − t)u1 )
≤ tf (u0 ) + (1 − t)f (u1 ) −
λ
t(1 − t) ku0 − u1 kκ (1 − t)κ−1 + tκ−1 .
2
We state a related definition of smoothness.
Definition A.1. A function f is (L, β, k·k)-smooth if it has β-Hölder
continuous gradient with respect to the norm k·k, meaning that
k∇f (u0 ) − ∇f (u1 )k∗ ≤ L ku0 − u1 kβ for u0 , u1 ∈ dom f.
Our first technical lemma is an equivalence result for uniform convexity.
Lemma A.1. Let f : Ω → R, where f is closed convex and Ω is a closed
convex set. Then f is (λ, κ, k·k)-uniformly convex over Ω if and only if for
u0 ∈ rel int Ω and all u1 ,
(30)
f (u1 ) ≥ f (u0 ) + sT0 (u1 − u0 ) +
λ
ku0 − u1 kκ for s0 ∈ ∂f (u0 ).
2
If inequality (29) holds, then inequality (30) also holds for any points u0 ∈ Ω
and s0 such that ∂f (u0 ) 6= ∅ and s0 ∈ ∂f (u0 ).
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 33
See Section A.8.1 for a proof of this lemma. There is also a natural duality
between uniform convexity and smoothness of a function’s Fenchel conjugate
f ∗ (v) = supu {v T u − f (u)}; such dualities are common [17, Ch. X.4].
Lemma A.2. Let Ω ⊂ Rk be a closed convex set and f : Ω → R be
1
1
, k·k∗ )-smooth
(λ, κ, k·k)-uniformly convex over Ω. Then f ∗ is (λ− κ−1 , κ−1
∗
k
(Def. A.1) over dom f = R .
See Section A.8.2 for a proof of Lemma A.2. We also have two results on the
properties of smooth functions, whose proofs we provide in Sections A.8.3
and A.8.4, respectively.
Lemma A.3.
Let f be (L, β, k·k) smooth. Then
f (u1 ) ≤ f (u0 ) + ∇f (u0 )T (u1 − u0 ) +
L
ku0 − u1 kβ+1 .
β+1
Lemma A.4. Let f be (L, β, k·k) smooth over Rk and inf x f (x) > −∞.
If the sequence un satisfies limn f (un ) = inf x f (x), then ∇f (un ) → 0.
A.2. Proof of Proposition 4. We state two intermediate lemmas before proving Proposition 4.
Lemma A.5. If H is symmetric, closed and strictly concave, then (−H)∗
is continuously differentiable on Rk . If αi ≥ αj , then p = ∇(−H)∗ (α) satisfies pi ≥ pj .
Proof. As H is strictly concave and (−H)∗ (α) = supπ∈∆k {π T α+H(π)} <
∞ for all α (suprema of closed concave functions over compact sets are attained), so (−H)∗ is continuously differentiable by standard results in convex
analysis [17, Thm X.4.1.1], and ∇(−H)∗ (α) = argmaxp∈∆k {pT α + H(p)}.
Now let α satisfy αi ≥ αj . As ∇(−H)∗ (α) = argmaxp∈∆k {pT α + H(p)},
let us assume for the sake of contradiction that pi < pj . Then letting A be
the permutation matrix swapping entries i and j, the vector p′ = Ap satisfies
H(p′ ) = H(p) but H( 12 (p′ + p)) > 21 H(p′ ) + 21 H(p) = H(p) = H(p′ ), and
αT p − αT p′ = αi pi − αi pj + αj pj − αj pi = (αi − αj )(pi − pj ) ≤ 0.
Thus we have −αT p ≥ −αT p′ , and so
1
1
1 ′
1 ′
1 T
′
T
α (p + p ) + H
p + p ≥ −α p + H
p + p > αT p + H(p),
2
2
2
2
2
a contradiction to the assumed optimality of p. We must have pi ≥ pj
whenever αi ≥ αj .
34
DUCHI, KHOSRAVI, RUAN
P
P
Lemma A.6. If H is symmetric and ki=1 πi L(α⋆ , i) = inf α ki=1 πi L(α, i),
then πi > πj implies that α⋆i ≥ α⋆j . If H is strictly concave, then α⋆i > α⋆j .
Proof. Let π satisfy πi > πj as assumed in the lemma, and suppose
that α⋆i < α⋆j for the sake of contradiction. Let A be the permutation matrix
that swaps α⋆i and α⋆j . Then (−H)∗ (Aα⋆ ) = (−H)∗ (α⋆ ), and (−H)∗ is also
symmetric, and
k
X
i=1
πi L(Aα⋆ , i) −
k
X
i=1
πi L(α⋆ , i) = −π T Aα⋆ + π T α⋆ = (πi − πj )(α⋆i − α⋆j ) < 0,
a contradiction to the optimalityPof α⋆ . If H is strictly concave, Lemma A.5
implies that for α⋆ minimizing ki=1 πi L(α, i), we have π = ∇(−H)∗ (α⋆ ).
Moreover, by Lemma A.5, if πi > πj we must have α⋆i > α⋆j .
P
Proof (of Proposition 4). If the infimum in ki=1 πi L(α, i) is attained,
P then Lemma A.6 gives the result. Otherwise, recall that H(π) =
inf α { ki=1 πi L(α, i)} > −∞, and let πi > πj . Let α(m) be any sequence
P
such that ki=1 πi L(α(m) , i) → H(π).
Using that H is uniformly concave, we have ∇(−H)∗ is Hölder continuous
over Rk (recall Lemma A.2). This implies that that π − ∇(−H)∗ (α(m) ) → 0
as m → ∞ (recall Lemma A.4), or
lim ∇(−H)∗ (α(m) ) = π.
m→∞
(m)
Now, by Lemma A.5, for any p(m) = ∇(−H)∗ (α(m) ), we have that αi ≥
(m)
(m)
(m)
αj implies pi ≥ pj . Thus, if πi > πj , it must be the case that eventually
(m)
we have αi
have
(m)
> αj
(m)
. Moreover, if lim inf |αi
(m)
(m)
lim inf |pi − pj | = 0, which
(m)
(m)
pi − pj → πi − πj . We thus
(m)
− αj
| = 0, then we must
would contradict that πi > πj , as we
(m)
(m)
have
find that lim inf m (αi − αj
the sequence tends to the infimum H(π), which implies that
)
( k
X
πi L(α, i) : αi ≤ αj > H(π).
inf
α
) > 0 if
i=1
The loss (11) is thus classification calibrated (Def. 3.1).
A.3. Proof of Proposition 5. Without loss of generality, we assume
that πk < maxj πj , so that restricting to αk ≥ maxj αj forces α to have larger
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 35
zero-one risk than 1 − maxj πj . We present two lemmas based on convex
duality that imply the result. In each, we let v\i = [v1 · · · vi−1 vi+1 · · · vk ]T ∈
Rk−1 be the vector v ∈ Rk without its ith element.
Let v ∈ Rk . Then
(
0
inf
vT α =
αk ≥maxj αj
−∞
Lemma A.7.
Pk−1
if vk = − i=1
vi , v\k 0
otherwise.
k−1
for the constraints
Proof. By introducing Lagrange multipliers β ∈ R+
αk ≥ αj for j 6= k, we have Lagrangian
(
T
0
if vk = β T 1, v\k = −β
β
L(α, β) = v +
α, so inf L(α, β) =
T
α
−β 1
−∞ otherwise.
Substuting β = −v\k and noting that β 0 gives the result.
Thus, if we define the matrix C = 11T − Ik×k with columns cl , we find
by strong duality that
T
−π α + (−HLzo )∗ (α)
inf
αk ≥maxj αj
(
)
−π T α + sup inf pT α + q T C T p
=
inf
αk ≥maxj αj
= sup
p∈∆k q∈∆k
inf
p∈∆k αk ≥maxj αj
T
T
min cl p + (p − π) α
l
= sup 1 − max pl | π\k p\k , pk = πk + 1T (π\k − p\k ) ,
p∈∆k
l
by Lemma A.7. The next lemma then immediately implies Proposition 5
once we note that HLzo (π) = 1 − maxj πj .
Lemma A.8.
Let Gk (π) = inf αk ≥maxj αj {−π T α + (−HLzo )∗ (α)}. Then
Gk (π) − (1 − max πj ) ≥
j
1
(max πj − πk ).
k j
Proof. We essentially construct the optimal p ∈ ∆k vector for the supremum in the definition of Gk . Without loss of generality, we assume that
π1 = maxj πj > πk , as the result is trivial if πk = maxj πk . For t ∈ R, define
L(t) = πk +
k
X
j=1
[πj − t]+ and R(t) = t.
36
DUCHI, KHOSRAVI, RUAN
By definining tlow = πk and thigh = π1 we have
L(tlow ) = πk +
k
X
j=1
[πj − πk ]+ ≥ πk + [π1 − πk ]+ = π1 > πk = R(tlow ),
L(thigh ) = πk < π1 = R(thigh ),
and the fact that L is strictly decreasing in [tlow , thigh ] and R is strictly
increasing implies that there exists a unique root t⋆ ∈ (tlow , thigh ) such that
L(t⋆ ) = t⋆ = R(t⋆ ). Now, we define the vector p by
pj = min{t⋆ , πj } for j ≤ k − 1, pk = t⋆ .
Pk−1 ⋆
Pk−1
Then we have 1T p = t⋆ + j=1
t ∧ πj = πk + j=1
([πj − t⋆ ]+ + t⋆ ∧ πj ) = 1,
and moreover, we have Gk (π) ≥ 1 − maxl pl = 1 − t⋆ .
It remains to show that 1 − t⋆ − (1 − π1 ) ≥ k1 (π1 − πk ); equivalently, we
must show that t⋆ ≤ (1 − k1 )π1 + k1 πk . Suppose for the sake of contradiction
that this does not hold, that is, that t⋆ > (1 − k1 )π1 + k1 πk . We know that
tlow = πk < t⋆ < π1 = thigh by assumption. With t⋆ satisfying these two
constraints, we have that
L(t⋆ ) = πk +
k
X
j=1
[πj − t⋆ ]+ ≤ πk +
k−1
X
j=1
[π1 − t⋆ ]+ = πk + (k − 1)(π1 − t⋆ )
1
k−1
π1 − πk
< πk + (k − 1) π1 −
k
k
1
k−1
π1 + πk < t⋆ ,
=
k
k
the two strict inequalities by assumption on t⋆ . But then we would have
L(t⋆ ) < R(t⋆ ) = t⋆ , a contradiction to our choice of t⋆ . Because L is decreasing, it is thus the case that t⋆ ≤ (1 − k1 )π1 + k1 πk , giving the result.
A.4. Proof of Observation 1. Our proof is essentially a trivial modification of Zhang [36, Theorem 8]. We assume without loss of generality that
φ(·) ≥ 0. Let π ∈ ∆k , and recalling that the cost matrix C = [c1 · · · ck ], let
L(π, α) =
k
X
y=1
πy
k
X
i=1
cyi φ(−αi ) =
k
X
π T ci φ(−αi ),
i=1
noting that π T ci ≥ 0 for each i. Without loss of generality, we may assume that π T c1 > π T c2 = minl π T cl . If we can show that inf α L(π, α) <
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 37
inf α1 ≥maxj αj L(π, α), then the proof will be complete. Let α(m) ∈ Rk be any
(m)
(m)
sequence satisfying 1T α(m) = 0 and α1 ≥ maxj αj such that L(π, α(m) ) →
inf α1 ≥maxj αj L(π, α).
We first show that it is no loss of generality to assume that α(m) converges.
Suppose for the sake of contradiction that lim supm α(m) = ∞. Then as
(m)
(m)
1T α(m) = 0, we must have lim supm α1 = ∞, so that lim supm φ(−α1 ) =
(m)
∞ because φ′ (0) < 0 and φ is convex. As it must be the case that π T c1 φ(−α1 )
remains bounded (for the convergence of L(π, α(m) )), we would then have
that π T c1 = 0, which is a contradiction because π T c1 > minl π T cl ≥ 0. Thus
we must have lim supm α(m) < ∞, and so there is a subsequence of α(m)
converging; without loss of generality, we assume that α(m) → α⋆ . Then
L(π, α⋆ ) = inf α1 ≥max αj L(π, α) by continuity of φ.
We show that by swapping the value of α⋆1 with the value α⋆2 (or increasing
the latter slightly), we can always improve the value L(π, α⋆ ). We consider
three cases, noting in each that α⋆1 ≥ 0 as 1T α⋆ = 0.
1. Let α⋆1 = α⋆2 ≥ 0. Then φ′ (−α⋆1 ) = φ′ (−α⋆2 ) ≤ φ′ (0) < 0, and additionally
cT1 πφ′ (−α⋆1 ) − cT2 πφ′ (−α⋆2 ) < 0 because cT2 π < cT1 π. For sufficently small
δ > 0, we thus have
cT1 πφ(−α⋆1 + δ) + cT2 πφ(−α⋆2 − δ) < cT1 πφ(−α⋆1 ) + cT2 πφ(−α⋆2 ),
whence L(π, α⋆ ) > L(π, α⋆ − δ(e1 − e2 )).
2. Let α⋆1 > α⋆2 and φ(−α⋆1 ) > φ(−α⋆2 ). Then taking α ∈ Rk such that α1 =
α⋆2 , α2 = α⋆1 , and αi = α⋆i for i ≥ 3, it is clear that L(π, α) < L(π, α⋆ ).
3. Let α⋆1 > α⋆2 and φ(−α⋆1 ) ≤ φ(−α⋆2 ). Using that −α⋆1 < 0, we see that
for sufficiently small δ > 0 we have φ(−α⋆1 + δ) < φ(−α⋆1 ) and φ(−α⋆2 ) ≥
φ(−α⋆2 −δ), because φ must be non-decreasing at the point −α⋆2 as φ′ (0) <
0. Thus we have L(π, α⋆ − δ(e1 − e2 )) < L(π, α⋆ ).
A.5. Biconjugates of the zero-one loss. In this subsection, we calculate the conjugate of the generalized entropy H(π) = 1 − maxj πj (pointwise Bayes risk) associated with the zero-one loss, demonstrating equality (16). Let (−H)∗ (α) = supπ∈∆k {π T α + 1 − kπk∞ }. Formulating the Lagrangian for the supremum with dual variables θ ∈ R, λ ∈ Rk+ , we have
L(π, θ, λ) = π T α − kπk∞ − θ(1T π − 1) + λT π,
which has dual objective
(
θ
sup L(π, θ, λ) =
π
−∞
if kα + λ − θ1k1 ≤ 1
otherwise.
38
DUCHI, KHOSRAVI, RUAN
P
As inf λ≥0 kα + λ − θ1k1 = kj=1 [αj − θ]+ and strong duality obtains (the
problems are linear), we have that the supremum in expression (16) is
X
k
[αj − θ]+ ≤ 1 .
sup π T α − kπk∞ | π T 1 = 1, π ≥ 0 = inf θ |
π
j=1
· · · by symmetry,
Without loss of generality we may assume that α1 ≥ α2 ≥ P
so that over the domain θ ∈ (−∞, α1 ], the function θ 7→ kj=1 [αj − θ]+ is
P
strictly decreasing. Thus, there is a unique smallest θ satisfying kj=1 [αj − θ]+ ≤
1 (attaining the equality), and by inspection, this must be one of
α1 + α2 1 α1 + α2 + α3 1
1T α 1
θ ∈ α1 − 1,
− ,
− ,...,
−
.
2
2
3
3
k
k
(Any θ makes some number of the terms [αj − θ]+ positive; fixing the number of terms and solving for θ gives the preceding equality.) Expression (16)
follows.
A.6. Pointwise infimal risks of hinge losses. We demonstrate equality (18) with C = [c1 · · · ck ]. We note that
)
( k
k
k
X
X
X
π T ci [1 + αi ]+ ,
cyi [1 + αi ]+ = inf
πy
H(π) = inf
αT 1=0
αT 1=0
y=1
i=1
i=1
and formulating the Lagrangian by introducing dual variable θ ∈ R, we have
L(α, θ) =
k
X
i=1
π T ci [1 + αi ]+ − θ1T α.
The generalized KKT conditions [6] for this problem are given by taking
subgradients of the Lagrangian. At optimum, we must have
if αi < −1
0
νi ∈ ∂ [1 + αi ]+ = [0, 1] if αi = −1 and π T ci νi −θ = 0, 1T α = 0, i ∈ [k].
1
if αi > −1,
Without loss of generality, assume that π T c1 = minj π T cj . Set α⋆ ∈ Rk and
θ ⋆ via
α⋆1 = (k − 1), α⋆i = −1 for i ≥ 2,
θ ⋆ = min π T cj = π T c1 .
j
We have 1T α⋆ = (k − 1) − (k − 1) = 0, and setting νi = 1 for i such
min π T c
j
j
∈ [0, 1] otherwise, we see that
that π T ci = minj π T cj and νi =
π T ci
T
⋆
T
T
π ci νi − θ = minj π cj − minj π cj = 0 for all i; the KKT conditions are
satisfied. Thus α⋆ and θ ⋆ are primal-dual optimal, yielding expression (18).
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 39
A.7. Classification calibration of hinge losses. We provide a quantitative guarantee on the classification calibration of hinge-like losses.
k×k
and let Lcw (α, y) = maxi {cyi | αi = maxj αj }
Lemma A.9. Let C ∈ R+
P
be the cost-weighted loss (10). Let L(α, y) = ki=1 cyi [1 + αi ]+ +I 1T α = 0
P
or L(α, y) = ki=1 cyi [1 + αi − αy ]+ . Then for any π ∈ ∆k and α ∈ Rk ,
k
X
i=1
πi L(α, i) − inf′
α
k
X
i=1
πi L(α′ , i) ≥
k
X
i=1
πi Lcw (α, i) − inf′
α
k
X
πi Lcw (α′ , i).
i=1
P
Proof. First, recall from Example 5 that inf α ki=1 πi L(α, i) = k minl π T cl .
Defining y(α) = argmaxj αj (breaking ties in some arbitrary deterministic
P
order), it is sufficient to argue that for L(α, y) = ki=1 cyi [1 + αi ]+ , we have
(31)
X
y
T
πy L(α, y) − k min π cl ≥
l
cTy(α) π
T
− min π cl
l
for any vector π ∈ ∆k and α with 1T α = 0.
To show inequality (31), assume without loss of generality that there
exists an index l⋆ < k such that cT1 π = cT2 π = · · · = cTl⋆ π = minl cTl π, while
cTl π >PcT1 π for l > l⋆ . (If l⋆ = k, then inequality (31) is trivial.) We always
have y πy L(α, y) ≥ k minl cTl π; let us suppose that αl ≥ maxj αj for some
l > l⋆ ; without loss of generality take l = k. Then we have
k
X
y=1
πy L(α, y) ≥
=
(32)
inf
αk ≥maxj αj ,1T α=0
inf
αk ≥maxj αj ,1T α=0
k
X
y=1
k
X
πy
k
X
cyj [1 + αj ]+
j=1
cTl π [1 + αl ]+ .
l=1
Writing the Lagrangian for this problem and introducing variables λ ≥ 0 for
the inequality αk ≥ maxj αj and θ ∈ R for the equality 1T α = 0, we have
L(α, λ, θ) =
k
X
l=1
π T cl [1 + αl ]+ + θ1T α + λ max αj − αk .
j
⋆
⋆
−1
k−l −1
⋆
Set α⋆1 = · · · = α⋆l⋆ = k−l
l⋆ +1 , αk = l⋆ +1 , and αl = −1 for l > l , l 6= k.
⋆
T
We claim that these are optimal. Indeed, set θ = −(1 − ǫ)π c1 − ǫπ T ck
40
DUCHI, KHOSRAVI, RUAN
for some ǫ ∈ (0, 1), whose value we specify later. Taking subgradients of the
Lagrangian with respect to α, we have
T
c1
..
⋆
⋆
∂α L(α , λ, θ ) = diag(ν) . π + θ ⋆ 1 + λ [Conv{e1 , . . . , el⋆ , ek } − ek ] ,
cTk
where νi ∈ [0, 1] are arbitrary scalars satisfying νi ∈ ∂ [1 + α⋆i ]+ , i.e. ν1 =
· · · = νl⋆ = νk = 1, νl ∈ [0, 1] for l ∈ {l⋆ + 1, . . . , k − 1}. Notably, we have
π T cl +θ ⋆ = ǫπ T (c1 −ck ) ≤ 0 for l ∈ {1, . . . , l⋆ } and π T ck +θ ⋆ = (1−ǫ)π T (ck −
c1 ) ≥ 0. Choosing ǫ small enough that ǫπ T ck + (1 − ǫ)π T c1 < π T cl for
T
T
l ∈ {l⋆ +1, . . . , k−1}, it is clear that we may take νl = (1−ǫ)ππTc1c+ǫπ ck ∈ [0, 1]
l
for each l 6∈ {1, . . . , l⋆ , k}, and we show how to choose λ⋆ ≥ 0 so that
0 ∈ ∂α L(α⋆ , λ, θ ⋆ ). Assume without loss of generality (by scaling of λ ≥ 0)
that π T (ck − c1 ) = 1. Eliminating extraneous indices in ∂α L, we see that we
seek a setting of λ and a vector v ∈ ∆l⋆ +1 such that
el⋆ +1 − ǫ1l⋆ +1 + λ(v − el⋆ +1 ) = 0.
1
[ǫ · · · ǫ (1 − ǫ)]T ∈ ∆l⋆ +1 , and set
This is straightforward: take v = l⋆ ǫ+(1−ǫ)
λ⋆ = l⋆ ǫ + (1 − ǫ) ≥ 0.
Summarizing, we find that our preceding choices of α⋆ were optimal for
⋆ −1
the problem (32), that is, α⋆1 = · · · = αl⋆ = k−l
l⋆ +1 = αk , with αl = −1 for
l⋆ < l < k. We thus have
k − l⋆ − 1
k − l⋆ − 1
T
πy L(α, y) ≥
+ π ck 1 +
π cl 1 +
l⋆ + 1
l⋆ + 1
y=1
l=1
k
l⋆
⋆
T
⋆
(k − l − 1) + ⋆
cT π.
= min π cl l + ⋆
l
l +1
l +1 k
⋆
l
X
k
X
T
In particular, we have
k
X
y=1
πy L(α, y) − k min π T cl
l
l⋆
k
⋆
≥ min π cl l + ⋆
(k − l − 1) + ⋆
cT π − k min cTl π
l
l
l +1
l +1 k
k
= ⋆
cT π − min cTl π .
l
l +1 k
T
⋆
Recalling that we must have had l⋆ < k, we obtain inequality (31).
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 41
A.8. Proofs of technical lemmas on smoothness.
A.8.1. Proof of Lemma A.1. Let t ∈ (0, 1) and ut = tu1 + (1 − t)u0 .
Assume that inequality (30) holds and let u0 ∈ rel int Ω. Then
λ t
κ
u − u1
2
λ
t
T
= f (u ) + (1 − t)st (u1 − u0 ) + (1 − t)κ ku0 − u1 kκ ,
2
f (u1 ) ≥ f (ut ) + sTt (u1 − ut ) +
where st ∈ ∂f (ut ), which must exist as ut ∈ rel int Ω (see Lemma B.3 and [17,
Chapter VI]). Similarly, we have
f (u0 ) ≥ f (ut ) + tsTt (u0 − u1 ) +
λ κ
t ku0 − u1 kκ .
2
Multiplying the preceding inequalities by t and (1 − t), respectively, gives
λ
ku0 − u1 kκ [t(1 − t)κ + (1 − t)tκ ]
2
λ
= f (ut ) + t(1 − t) ku0 − u1 kκ (1 − t)κ−1 + tκ−1
2
tf (u1 ) + (1 − t)f (u0 ) ≥ f (ut ) +
for any u0 ∈ rel int Ω.
We now consider the case that u0 ∈ Ω \ rel int Ω, as the preceding display
is equivalent to the uniform convexity condition (29). Let u0 ∈ Ω \ rel int Ω
and u′0 ∈ rel int Ω. For a ∈ [0, 1] define the functions h(a) = tf (u1 ) + (1 −
t)f ((1 − a)u0 + au′0 ) and
ht (a) = f (tu1 + (1 − t)((1 − a)u0 + au′0 ))
λ
+ t(1 − t) (1 − a)u0 + au′0 − u1
2
κ
(1 − t)κ−1 + tκ−1 .
Then h and ht are closed one-dimensional convex functions, which are thus
continuous [17, Chapter I], and we have h(a) ≥ ht (a) for all a ∈ (0, 1) as
(1 − a)u0 + au′0 ∈ rel int Ω. Thus
tf (u0 ) + (1 − t)f (u1 ) = h(0) = lim h(a) ≥ lim ht (a) = ht (0)
a→0
a→0
λ
= f (tu1 + (1 − t)u0 ) + t(1 − t) ku0 − u1 kκ (1 − t)κ−1 + tκ−1 .
2
This is equivalent to the uniform convexity condition (29).
We now prove the converse. Assume the uniform convexity condition (29),
which is equivalent to
f (ut ) − f (u0 ) λ
+ (1 − t) ku0 − u1 kκ [(1 − t)κ−1 + tκ−1 ] ≤ f (u1 ) − f (u0 )
t
2
42
DUCHI, KHOSRAVI, RUAN
(x)
be the
for all u0 , u1 ∈ Ω and t ∈ (0, 1). Let f ′ (x, d) = limt↓0 f (x+td)−f
t
directional derivative of f in direction d, recalling that if ∂f (x) 6= ∅ then
f ′ (x, d) = supg∈∂f (x) hg, di (see [17, Ch. VI.1]). Then taking t ↓ 0, we have
f ′ (u0 , u1 − u0 ) +
λ
ku0 − u1 kκ ≤ f (u1 ) − f (u0 ).
2
Because f ′ (u0 , u1 − u0 ) = supg∈∂f (u0 ) hg, u1 − u0 i, this implies the subgradient condition (30).
A.8.2. Proof of Lemma A.2. First, we note that as dom f = Ω and f is
uniformly convex, it is 1-coercive, meaning that limkuk→∞ f (u)/ kuk = ∞.
Thus dom f ∗ = Rk ; see [17, Proposition X.1.3.8].
As f is strictly convex by assumption, we have that f ∗ is differentiable [17,
Theorem X.4.1.1]. Moreover, as f is closed convex, f = f ∗∗ , and we have
for any s ∈ Rk that u0 = ∇f ∗ (s) if and only if s ∈ ∂f (u0 ), meaning that
f is subdifferentiable on the set Im ∇f ∗ = {∇f ∗ (s) : s ∈ Rk }, whence
Ω ⊃ Im ∇f ∗ . Now, let s0 , s1 ∈ Rk and u0 = ∇f ∗ (s0 ) and u1 = ∇f ∗ (s1 ). We
must then have s0 ∈ ∂f (u0 ) and s1 ∈ ∂f (u1 ) by standard results in convex
analysis [17, Corollary X.1.4.4], so that ∂f (u0 ) 6= ∅ and ∂f (u1 ) 6= ∅.
Now we use the uniform convexity condition (30) of Lemma A.1 to see
λ
ku0 − u1 kκ and
2
λ
f (u0 ) − f (u1 ) ≥ hs1 , u0 − u1 i + ku0 − u1 kκ .
2
f (u1 ) − f (u0 ) ≥ hs0 , u1 − u0 i +
Adding these equations, we find that
hs0 − s1 , u0 − u1 i ≥ λ ku0 − u1 kκ , so ks0 − s1 k∗ ku0 − u1 k ≥ λ ku0 − u1 kκ
by Hölder’s inequality. Dividing each side by ku0 − u1 k, we obtain ku0 − u1 k ≤
1
1
λ− κ−1 ks0 − s1 k∗κ−1 , the desired result once we note that ∇f ∗ (si ) = ui .
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 43
A.8.3. Proof of Lemma A.3. Define ut = u0 + t(u1 − u0 ) for t ∈ [0, 1].
Then using Taylor’s theorem, we have
Z 1
h∇f (tu1 + (1 − t)u0 ), u1 − u0 i dt
f (u1 ) = f (u0 ) +
0
Z 1
h∇f (ut ) − ∇f (u0 ), u1 − u0 i dt
= f (u0 ) + h∇f (u0 ), u1 − u0 i +
0
Z 1
k∇f (ut ) − ∇f (u0 )k∗ ku0 − u1 k dt
≤ f (u0 ) + h∇f (u0 ), u1 − u0 i +
0
Z 1
Ltβ ku0 − u1 kβ+1 dt.
≤ f (u0 ) + h∇f (u0 ), u1 − u0 i +
0
Computing the final integral as
R1
0
tβ dt =
1
β+1
gives the result.
A.8.4. Proof of Lemma A.4. Let gn = ∇f (un ) and suppose for the sake
of contradiction that gn 6→ 0. Then there is a subsequence, which without
loss of generality we take to be the full sequence, such that kgn k∗ ≥ c > 0
for all n. Fix δ > 0, which we will choose later. By Lemma A.3, defining
yn = un − δgn / kgn k∗ , we have
L β
L
β+1
Lδ
≤ f (un ) − δ c −
δ .
f (yn ) ≤ f (un ) − δ kgn k +
β+1
β+1
In particular, we see that if
δ≤
c β+1
·
2
L
1/β
, then f (yn ) ≤ f (un ) −
δc
2
for all n, which contradicts the fact that f (un ) → inf x f (x) > −∞.
APPENDIX B: ORDER EQUIVALENCE OF CONVEX FUNCTIONS
In this appendix, we collect the proofs of our various technical results on
order equivalent functions (Definition 5.1).
B.1. Proof of Lemma 5.2. Before proving the lemma, we give a matrix characterization of non-negative vectors with equal sums similar to the
characterization of majorization via doubly stochastic matrices (cf. [22]; we
temporarily defer the proof of the lemma to Appendix B.1.1).
T
T
Lemma B.1. Vectors a, b ∈ Rm
+ satisfy 1 a = 1 b if and only if there
m×m
T
such that Z1 = a and Z 1 = b.
exists a matrix Z ∈ R+
44
DUCHI, KHOSRAVI, RUAN
Returning to the proof of Lemma 5.2 proper, let fi = fL(i) ,π for shortk−1
by the construction (15), because
hand. Note that dom f1 = dom f2 = R+
(k−1)×m
satisfying
mini inf α∈Rk L(α, i) > −∞. Now, given matrices A, B ∈ R+
A1 = B1 (where A = [a1 · · · am ] and B = [b1 · · · bm ]), we construct
distributions Pi and quantizers q1 and q2 such that for any f we have
Df (P1 , . . . , Pk−1 ||Pk | q1 ) = C
Df (P1 , . . . , Pk−1 ||Pk | q2 ) = C
m
X
j=1
m
X
f (aj ) and
f (bj ),
j=1
where C > 0 is a constant. We then use Definition 4.1 of loss equivalence to
show that f1 and f2 are order equivalent.
With that in mind, take M to be any positive integer such that M >
(k−1)×M
max{kA1k∞ , m}. We enlarge A and B into matrices Aext , B ext ∈ R+
respectively,
To construct these matrices, let aext
i,m+1 =
Pm adding Mext−m columns. P
m
M − j=1 aij and bi,m+1 = M − j=1 bij , for i = 1, . . . , k − 1, and set
ext and
ext
aext
il = bil = 0 for all m + 1 < l ≤ M . The enlarged matrices A
1
1 ext
A 1 = M
B ext 1 = 1, and their columns belong to
B ext thus satisfy M
k−1
dom f1 = dom f2 = R+ .
Let the spaces X = {(i, j) | 1 ≤ i ≤ M, 1 ≤ j ≤ M } and Z =
{1, 2, . . . , M }. Define quantizers q1 , q2 : X → [M ] by q1 (i, j) = i and
q2 (i, j) = j. As Aext 1 = B ext 1 = M 1, Lemma B.1 guarantees the exisl ] ∈ RM ×M such that
tence of matrices Z l = [zij
+
1 ext M
[b ]
∈ RM
+,
M lj j=1
P l
which implies that for all l ∈ [k − 1], the matrix Z l satisfies ij zij
= 1. For
l
each l ∈ [k −1], define the probability distribution Pl on X by Pl ((i, j)) = zij
for i, j ∈ [M ]. Let Pk be the distribution defined by Pk ((i, j)) = M −2 .
Under this quantizer design and choice of distributions Pl , we have for
any prior π ∈ Rk+ (upon which the functions f1 and f2 implicitly depend)
and f = f1 or f = f2 that
Z l1 =
1 ext M
[a ]
∈ RM
+
M lj j=1
and (Z l )T 1 =
Df (P1 , . . . , Pk−1 ||Pk | q1 ) =
Df (P1 , . . . , Pk−1 ||Pk | q2 ) =
M
1 X
f (aext
j ) and
M
j=1
M
1 X
f (bext
j ),
M
j=1
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 45
PM
P
PM l
−1 ,
ext
because M
i=1 Pk ((i, j)) = M
i=1 Pl ((i, j)) =
i=1 zij = alj /M and
and similarly for B ext . By the loss equivalence of L(1) and L(2) (recall
Def. 4.1), we obtain that
M
X
j=1
f1 (aext
j )
≤
M
X
j=1
f1 (bext
j )
if and only if
M
X
j=1
f2 (aext
j )
≤
M
X
f2 (bext
j ).
j=1
Note that aext
= bext
∈ dom fi for each j > m and i = 1, 2. Moreover,
j
j
ext
ext
aj = aj and bj = bj for 1 ≤ j ≤ m, so the preceding display is equivalent
to the order equivalence (24).
B.1.1. Proof of Lemma B.1. One direction of the proof is easy: if Z ∈
and Z1 = a while Z T 1 = b, then 1T Z1 = 1T a = bT 1.
Rm×m
+
We prove the converse using induction. When m = 1, the result is immediate. We claim that it is no loss of generality to assume that a1 ≥ a2 ≥
· · · ≥ am and b1 ≥ · · · ≥ bm ; indeed, let Pa and Pb be permutation matrices
such that Pa a and Pb b are in sorted (decreasing) order. Then if we construct
e = Pa a and ZeT 1 = Pb b, we have Z = P T ZP
e b
such that Z1
Ze ∈ Rm×m
a
+
m×m
T
T
T
T
e = Pa Pa a = a and Z 1 = P Z1
e = b.
and Z1 = Pa Z1
satisfies Z ∈ R+
b
Now, suppose that the statement of the lemma is true for all vectors of
dimension up to m − 1; we argue that the result holds for a, b ∈ Rm
+ . It is
no loss of generality to assume that am ≤ bm . Let zm,m = am , zi,m = 0 for
2 ≤ i ≤ m − 1, and set z1,m = bm − am ≥ 0. Additionally, we set zm,i = 0
m−1×m−1
be the matrix defined by
for 1 ≤ i ≤ m − 1. Now, let Zinner ∈ R+
the upper (m − 1) × (m − 1) sub-matrix of Z, and define the vectors
ainner = [a1 + am − bm a2 a3 · · · am−1 ]T and binner = [b1 b2 · · · bm−1 ]T .
Then as a1 ≥ (1/m)1T a = (1/m)1T b ≥ bm , we have ainner ≥ 0 and binner ≥
0, and moreover, 1T ainner = 1T a − bm = 1T b − bm = 1T binner . In particular,
we have by the inductive hypothesis that we may choose Zinner such that
T
Zinner 1 = ainner and Zinner
1 = binner . By inspection, setting
bm − am
0
Zinner
.
.
Z=
,
.
0
0 ··· 0
am
and Z1 = a and Z T 1 = b.
we have Z ∈ Rm×m
+
46
DUCHI, KHOSRAVI, RUAN
B.2. Proof of Lemma 5.3. As in our discussion in Section 5.2, we
prove the lemma in a sequence of steps and auxiliary lemmas. The roadmap
is as follows: first, we show that we may assume the set Ω in Lemma 5.3 has
non-empty interior (§ B.2.1). With this done, we can consider any affinely
independent subset of k + 1 points from Ω, whence Lemma 5.4 holds. With
this done, we can cover Ω with overlapping simplices (§ B.2.2), and extend
this to all of Ω (§ B.2.3).
B.2.1. It is no loss of generality to assume Ω has non-empty interior in
Lemma 5.3. Let H = aff Ω be the affine hull of Ω, where dim H = l ≥ 1.
(If dim H = 0, then Ω is a single point, and Lemma 5.3 is trivial.) We argue
that if Lemma 5.3 holds for sets Ω such that int Ω 6= ∅ it holds generally;
thus we temporarily assume its truth
for convex Ω with int Ω 6= ∅.
Since dim H = l, we have H = Av + d | v ∈ Rl for some full columnrank matrix A ∈ Rk×l and d ∈ Rk . As Ω has non-empty interior relative
to H (e.g. [17, Theorem 2.1.3]), we have Ω = {Av + d | v ∈ Ω0 } for a
convex set Ω0 ⊂ Rl with int Ω0 6= ∅. Defining fe1 (v) = f1 (Av + d) and
fe2 (v) = f2 (Av + d), where fei (v) = ∞ for v 6∈ Ω0 , we have that fe1 and fe2
are order equivalent on Ω0 ⊂ Rl . By assumption, Lemma 5.3 holds for Ω0 ,
so there exist a > 0, eb ∈ Rl , e
c ∈ R such that
f1 (Av + d) = fe1 (v) = afe2 (v) + heb, vi + c, for v ∈ Ω0 .
As A is full column rank, for all t ∈ Ω there exists a unique vt ∈ Ω0 such
that t = Avt + d, and the mapping t 7→ vt is linear, i.e., vt = Et + g for some
E ∈ Rl×k and g ∈ Rl (we may take E ∈ Rl×k to be any left-inverse of A and
g = −Ed, so v = Et − Eg). We obtain that fi (t) = fei (Et + g) for i = 1, 2,
whence
f1 (t) = fe1 (Et+g) = afe2 (Et+g)+heb, Et+gi+e
c = af2 (t)+hE T eb, ti+heb, gi+e
c
for all t ∈ Ω, which is our desired result. From this point forward, we thus
assume w.l.o.g. that int Ω 6= ∅.
B.2.2. Covering sets with simplices. With Lemma 5.4 in hand, we now
show that the special case for simplices is sufficient to show the general
Lemma 5.3. First, we show that simplices essentially cover convex sets Ω.
Lemma B.2. Let v, w be arbitrary points in int Ω ⊂ Rk . Then, there exist
u0 , u1 , . . . , uk ∈ Ω such that v, w ∈ int E, where E = Conv{u0 , u1 , . . . , uk }.
For any points u2 , . . . , uk that make v, w, u2 , . . . , uk affinely independent
(Def. 5.2), there exist u0 , u1 ∈ Ω such that v, w ∈ int Conv{u0 , u1 , . . . , uk }.
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 47
Before proving Lemma B.2, we state a technical lemma about interior points
of convex sets.
Lemma B.3 (Hiriart-Urruty and Lemaréchal [17], Lemma III.2.1.6). Let
C ⊂ Rk be a convex set, u ∈ rel int C and v ∈ cl C. Then for any λ ∈ [0, 1),
we have λu + (1 − λ)v ∈ rel int C.
Proof (of Lemma B.2). Take u2 , . . . , uk ∈ Ω arbitrarily but in general
Ω
u2
w
u1
PSfrag replacements
v
u0
Fig 1: The construction in Lemma B.2.
position, so that the points v, w, u2 , . . . , uk and w, v, u2 , . .P
. , uk are affinely
1
(v + w + ki=2 ui ), so that
independent (Def. 5.2). Now, define ucent = k+1
ucent ∈ int Ω (Lemma B.3), and choose ǫ > 0 small enough that the points
u0 = v +ǫ(v −ucent ) and u1 = w+ǫ(w−ucent ) are both in Ω. (This is possible
as v, w ∈ int Ω; see Figure 1.) Then we find that v = (u0 + ǫucent )/(1 + ǫ)
and w = (u1 + ǫucent )/(1 + ǫ), so that
!
k
X
u0
u1
2ǫ
1
ui ,
+
+
ucent +
ucent =
k+1 1+ǫ 1+ǫ 1+ǫ
i=2
or by rearranging, ucent = λ0 (u0 + u1 ) + λ1
λ0 =
λ1 =
2ǫ
1−
(k + 1)(1 + ǫ)
1−
2ǫ
(k + 1)(1 + ǫ)
−1
−1
Pk
i=2 ui ,
where
1
and
(k + 1)(1 + ǫ)
1
.
k+1
48
DUCHI, KHOSRAVI, RUAN
Noting that 2λ0 + (k − 1)λ1 = 1 and λi > 0, we obtain that ucent ∈
int Conv{u0 , . . . , uk }. As the points ui are in general position, and as
v=
ǫ
1
ǫ
1
u0 +
ucent and w =
u1 +
ucent ,
1+ǫ
1+ǫ
1+ǫ
1+ǫ
we have v, w ∈ int Conv{u0 , . . . , uk } by Lemma B.3.
B.2.3. Extension from a single simplex to all of Ω. We use Lemma B.2
to show the following lemma, which implies Lemma 5.3.
Lemma B.4. In addition to the conditions of Lemma 5.3, assume that
for all simplices E ⊂ int Ω there exist aE > 0, bE ∈ Rk , and cE ∈ R such
that f1 (t) = aE f2 (t) + bTE t + cE for t ∈ E. Then there exist a > 0, b ∈ Rk ,
and c ∈ R such that
f1 (t) = af2 (t) + bT t + c for t ∈ Ω.
Coupled with Lemma 5.4, Lemma B.4 immediately yields Lemma 5.3; indeed, Lemma 5.4 shows that for any simplex E the conditions of Lemma B.4
holds, so that Lemma 5.3 follows, i.e., f1 (t) = af2 (t) + bT t + c for all t ∈ Ω.
Proof. For i ∈ {1, 2} define the sets
Si = {(u, v) ∈ int Ω × int Ω | ∇fi (u) 6= ∇fi (v)}
∪ {(u, v) ∈ int Ω × int Ω | ∇fi (u) or ∇fi (v) does not exist}
⊂ Ω × Ω.
We divide our discussion into two cases.
Case 1. First we suppose that S1 = S2 = ∅. Then for i = 1, 2, fi are
differentiable in int Ω in this case, we have that ∇fi (u) = ∇fi (v) for all
u, v ∈ int Ω. Then by continuity of the fi on int Ω, we must have fi (t) =
bTi t + ci for i = 1, 2. The result follows by taking a = 1, b = b1 − b2 and
c = c1 − c2 and applying Lemma 5.5.
Case 2. We have that at least one of S1 and S2 is non-empty. Without
loss of generality, say S1 6= ∅. Choose a pair (u⋆ , v ⋆ ) ∈ S1 and consider the
collection of sets
M = {E = Conv{u0 , u1 , . . . , uk } | u⋆ , v ⋆ ∈ int E and u0 , u1 , . . . uk ∈ Ω} .
We show that for any E1 , E2 ∈ M, we have aE1 = aE2 , bE1 = bE2 and
cE1 = cE2 . We know that int E1 ∩ int E2 6= ∅, as u⋆ ∈ int E for all E ∈ M.
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 49
Now, assume for the sake of contradiction that aE1 6= aE2 . For all t ∈ E1 ∩E2
we have
(33)
f1 (t) = aE1 f2 (t) + bTE1 t + cE1 and f1 (t) = aE2 f2 (t) + bTE2 t + cE2 .
By subtracting the preceding equations from one another after multiplying
by aE2 and aE1 , respectively, one obtains that for t ∈ E1 ∩ E2 ,
f1 (t) =
aE 2
1
(aE2 bE1 − aE1 bE2 )T t + (aE2 cE1 − aE1 cE2 ) .
− aE1
This yields a contradiction, since we have assumed that either (i) ∇f1 (u⋆ )
or ∇f2 (v ⋆ ) does not exist or (ii) ∇f1 (u⋆ ) 6= ∇f1 (v ⋆ ), while u⋆ , v ⋆ ∈ int E1 ∩
int E2 = int(E1 ∩ E2 ). Thus aE1 = aE2 . To obtain bE1 = bE2 , note that
int(E1 ∩ E2 ) 6= ∅, as u⋆ , v ⋆ ∈ int(E1 ∩ E2 ). Subtracting the equalities (33),
we find 0 = f1 (t)−f1 (t) = (bE1 −bE2 )T t+cE1 −cE2 for t ∈ E1 ∩E2 , which has
non-empty interior. That bE1 = bE2 immediately implies cE1 = cE2 . Hence,
there exist some a > 0, b ∈ Rk , and c ∈ R such that for all sets E ∈ M, we
have aE = a, bE = b, and cE = c.
S
We complete the proof by showing that {E | E ∈ M} is dense in Ω.
Define Ω◦ ⊂ Ω as
there exist u3 , . . . , uk
◦
Ω = t∈Ω|
.
such that u⋆ , v ⋆ , t, u3 , . . . , uk are affinely independent
The set Ω◦ forms a dense subset of Ω, as u⋆ ∈ int Ω 6= ∅. For any t ∈ Ω◦ ,
Lemma B.2 guarantees the S
existence of E ∈ M such that t ∈ E and u⋆ , v ⋆ ∈
◦
int E. We thus have Ω ⊂ {E | E ∈ M}. As f1 (t) = af2 (t) + bT t + c for
all t ∈ Ω◦ by the previous paragraph, Lemma 5.5 allows us to extend the
equality to f1 (t) = af2 (t) + bT t + c for all t ∈ Ω.
B.3. Proofs of auxiliary lemmas for Lemma 5.4.
B.3.1. Proof of Lemma 5.6. For each i, let λi = [αi ]+ and νi = [αi ]−
be the positive and negative parts of the αi , and let λi = rsi and −νi = qsi
where qi , ri , s ∈ N. Then we have
m
X
i=1
λi ui = v −
m
X
i=1
νi ui or
m
X
ri ui = sv +
i=1
m
X
qi ui ,
i=1
where 1T r = s + 1T q as 1T α = 1. Defining the matrices
· · v} u1 · · · u1 · · · um · · · um ],
A = [u1 · · · u1 · · · um · · · um ] and B = [v| ·{z
| {z } | {z }
| {z } | {z }
r1 times
rm times
s times
q1 times
qm times
50
DUCHI, KHOSRAVI, RUAN
we have A1 = B1 with columns in Ω. Then order equivalence (24) implies
m
X
i=1
ri f1 (ui ) ≤ sf1 (v) +
m
X
qi f1 (ui ) iff
m
X
i=1
i=1
Each of these is equivalent to
Pm
ri f2 (ui ) ≤ sf2 (v) +
i=1 αi fj (ui )
m
X
qi f2 (ui ).
i=1
≤ fj (v) for j = 1, 2.
B.3.2. Proof of Lemma 5.7. We assume that f1 and f2 are non-linear
over Conv{u0 , . . . , uk }, as otherwise the result is trivial. Let the vectors g1
and g2 and matrix H ∈ Rk×k be defined by
T
f2 (u1 ) − f2 (u0 )
f1 (u1 ) − f1 (u0 )
u1 − uT0
..
..
..
H=
.
, g2 =
, g1 =
.
.
.
T
T
f2 (uk ) − f2 (u0 )
f1 (uk ) − f1 (u0 )
uk − u0
For any a > 0, define the vector b(a) ∈ Rk so that
b(a)T (ui − u0 ) = f1 (ui ) − f1 (u0 ) − a(f2 (ui ) − f2 (u0 )) = g1,i − ag2,i ,
that is, as b(a) = H −1 (g1 − ag2 ), which is possible as H is full rank. By
choosing c(a) = f1 (u0 ) − af2 (u0 ) − b(a)T u0 , we then obtain that
f1 (ui ) = af2 (ui ) + b(a)T ui + c(a)
by algebraic manipulations. We now consider ucent . We have
k
T
b(a) ucent
1 X
1
=
1T Hb(a)
b(a)T ui = b(a)T u0 +
k+1
k+1
i=0
= b(a)T u0 +
1
1T (g1 − ag2 ),
k+1
so that
af2 (ucent ) + b(a)T ucent + c(a)
= af2 (ucent ) +
k
X
f1 (ui ) − f1 (u0 ) + af2 (u0 ) − af2 (ui )
k+1
i=1
k
k
i=0
i=0
1 X
1 X
= af2 (ucent ) +
f1 (ui ) − a
f2 (ui ).
k+1
k+1
+ f1 (u0 ) − af2 (u0 )
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 51
Thus we may choose an a > 0 such that our desired equalities hold if and
only if there exists a > 0 such that
k
k
i=0
i=0
1 X
1 X
f1 (ui ) = af2 (ucent ) − a
f2 (ui ).
f1 (ucent ) −
k+1
k+1
By the assumption
that f1 and f2 are non-linear, we
have that (by Lemma 5.8)
1 Pk
1 Pk
f1 (ucent ) < k+1 i=0 f1 (ui ) and f2 (ucent ) < k+1 i=0 f2 (ui ). Thus, setting
f1 (ucent ) −
a=
f2 (ucent ) −
gives the desired result.
1
k+1
1
k+1
Pk
i=0 f1 (ui )
Pk
i=0 f2 (ui )
>0
B.3.3. Proof of Lemma 5.8. Without loss of generality, we assume that
λ1 = kλk∞ and that λ1 > 1/m. Then we have
!
m
m
m
m
X
X
X
X
(λi − λ1 )f (ui )
f (ui ) +
λi f (ui ) = λ1
λi ui ≤
f
i=1
i=1
i=1
(34)
= mλ1 f (ucent ) +
i=1
m
X
i=2
Let Λ =
Pm
i=1 (λ1
(λ1 − λi )f (ui ).
− λi ) = mλ1 − 1 > 0. Then
ucent =
m
m
1 X
Λ X λ1 − λi
λi ui +
ui ,
mλ1
mλ1
Λ
i=1
i=1
and we have
"
!
!#
m
m
X
X
1
Λ
λ1 − λi
mλ1 f (ucent ) ≤ nλ1
λi ui +
f
f
ui
mλ1
mλ1
Λ
i=1
i=2
!
m
m
X
X
λ1 − λi
f (ui ).
λi ui + Λ
≤f
Λ
i=1
i=2
The inequalities (34) must have been equalities, giving the result.
APPENDIX C: PROOFS FOR BAYES CONSISTENCY OF
EMPIRICAL RISK MINIMIZATION
52
DUCHI, KHOSRAVI, RUAN
C.1. Proof of Lemma 4.1. With the definition R⋆ (q) := inf γ R(γ | q),
we have
(35) R(γ | q) −
inf
γ∈Γ,q∈Q
R(γ | q) = R(γ | q) − R⋆ (q) + R⋆ (q) − inf R⋆ (q).
q∈Q
For the second two terms, we note that by Theorem 1, we have
R⋆ (q) = E[HLcw (e
π (q(X)))] = aE[HL (e
π (q(X)))] + bT π + c,
and similarly for inf q R⋆ (q), so that
⋆
⋆
⋆
⋆
R (q) − inf R (q) = a RL (q) − inf RL (q)
q∈Q
q∈Q
≤ a RL (γ | q) − inf RL (γ | q) .
γ∈Γ,q∈Q
Clearly t 7→ at is concave, so that it remains to bound R(γ | q) − R⋆ (q)
in expression (35). To that end, let L(α) ∈ Rk denote the vector of losses
L(α, y) and define the function
T
′
T
cw
cw ′
H(ǫ) = inf
sup π L(α) − L(α ) | sup π L (α) − L (α ) ≥ ǫ
π∈∆k ,α
α′
α′
and let H ∗∗ be its Fenchel biconjugate. Then (see [36, Proposition 25 and
Corollary 26], as well as the papers [31, 30, 12, Proposition 1]) we have that
H ∗∗ (ǫ) > 0 for all ǫ > 0, and defining g(ǫ) = sup{δ : δ ≥ 0, H ∗∗ (δ) ≤ ǫ}
yields the desired concave function by taking h(t) = g(t) + at. In passing,
we note that we may w.l.o.g. replace h(t) with min{h(t), maxij cij }.
Now we give the second result. Without loss of generality, we may assume that the vector γ(z) ∈ Rk has a unique maximal coordinate; we
may otherwise assume a deterministic rule for breaking ties. Let y(γ(z)) =
argmaxj γj (z), assumed w.l.o.g. to be unique. Consider that
⋆
R(γ | q) − R (q) =
Pk
Z
sup
α
k
X
i=1
q
[e
πi (z)Lcw (γ(z), i) − Lcw (α, i)] dP (z)
−1
where P̄ q (A) =
A ⊂ Z) is the pushi=1 πi Pi (q (A)) (for measurable
q
q
k
forward of q and π
e(z) = [πi dPi (z)]i=1 /dP (z) is shorthand for the posterior
of Y conditional on observing z ∈ q(X ). Lemma A.9 immediately implies
⋆
R(γ | q) − R (q) ≤
Z
sup
α
k
X
i=1
q
[e
πi (z)L(γ(z), i) − L(α, i)] dP (z)
⋆
(q)
= RL (γ | q) − RL
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 53
Pk
for either of the hinge-type losses, e.g. L(α, y) =
i=1 cyi [1 + αi − αy ]+ .
In the notation of the previous (general) case, we have a = 1/k because
HL (π) = k1 HLcw (π), whence we may take h(t) = (1 + 1/k)t.
C.2. Proof of Theorem 3. The proof of Theorem 3 is almost immediate from Lemma 4.1. Indeed, by Lemma 4.1, we have
⋆
R(γn | qn ) − R⋆ (Q) ≤ h (RL (γn | qn ) − RL
(Q))
for some h concave, bounded, and satisfying h(0) = 0. It is thus sufficient to
⋆ (Q)] → 0. Now, if we let γ ⋆ and q⋆ minimize
show that E[RL (γn | qn ) − RL
n
n
RL (γ | q) over the set Γn × Qn (or be arbitrarily close to minimizing the
L-risk), we have
RL (γn | qn ) − RL (γn⋆ | q⋆n )
bL,n (γn | qn ) + R
bL,n (γn | qn ) − RL (γ ⋆ | q⋆ )
= RL (γn | qn ) − R
n
n
⋆
⋆
⋆
b
b
≤ RL (γn | qn ) − RL,n (γn | qn ) + RL,n (γn | qn ) − RL (γn | q⋆n )
|
{z
}
bL,n (γ|q)−RL (γ|q)|
≤2 supγ∈Γn ,q∈Qn |R
bL,n (γn | qn ) −
+R
inf
γ∈Γn ,q∈Qn
bL,n (γ | q).
R
Consequently, we have the expectation bound
⋆
E [RL (γn | qn ) − RL
(Q)]
⋆
opt
app
≤ E [RL (γn | qn ) − RL (γn⋆ | q⋆n )] + RL (γn⋆ | q⋆n ) − RL
(Q) ≤ 2ǫest
n + ǫn + ǫn ,
which converges to zero as desired.
APPENDIX D: PROOFS OF BASIC PROPERTIES OF
F -DIVERGENCES
In this section, we collect the proofs of the characterizations of generalized
f -divergences (Def. 2.1).
D.1. Proof of Lemma 2.1. Let µ1 and µ2 be dominating measures;
then µ = µ1 + µ2 also dominates P1 , . . . , Pk as well as µ1 and µ2 . We have
for ν = µ1 or ν = µ2 that
dPi /dν
dPi /dν dν/dµ
dPi /dµ
dPi dν
dPi
=
=
and
=
,
dPk /dν
dPk /dν dν/dµ
dPk /dµ
dν dµ
dµ
54
DUCHI, KHOSRAVI, RUAN
the latter two equalities holding µ-almost surely by definition of the RadonNikodym derivative. Thus we obtain for ν = µ1 or ν = µ2 that
Z
Z
dP1 /dν
dPk−1 /dν dPk dν
dP1:k−1 /dν dPk
dν = f
,...,
dµ
f
dPk /dν
dν
dPk /dν
dPk /dν
dν dµ
Z
dPk−1 /dµ dPk
dP1 /dµ
,...,
dµ,
= f
dPk /dµ
dPk /dµ
dµ
dPi /dµ
again by definition of the Radon-Nikodym derivative. We have ppki = dP
k /dµ
a.s.-µ, which shows that the base measure µPdoes not affect the integral.
To see the positivity, we may take µ = k1 ki=1 Pi , in which case Jensen’s
inequality implies (the perspective function (9) is convex) that
dPk−1
dP1
(X), . . . ,
(X) dPk (X)
Df (P1 , . . . , Pk−1 ||Pk ) = E f
dPk
dPk
E[dP1 ]
E[dPk−1 ]
≥f
,...,
E[dPk ] = f (1) = 0,
E[dPk ]
E[dPk ]
where the expectation is taken under the distribution µ. The inequality is
strict for f strictly convex at 1 as long as dPi /dPk is non-constant for some
i, meaning that there exists an i such that Pi 6= Pk .
D.2. Proof of Proposition 1. Before proving the proposition, we first
establish a more general continuity result for f -divergences. This result is a
direct generalization of results of [34, Thm. 5]. Given a sub-σ-algebra G ⊂ F,
we let P G denote the restriction of the measure P , defined on F, to G.
of F
Lemma D.1. Let F1 ⊂ F2 ⊂ . .. be a sequence of sub-σ-algebras
Fn
||PkFn is non-decreasing
and let F∞ = σ(∪n≥1 Fn ). Then Df P1Fn , . . . , Pk−1
in n and
F∞
Fn
||PkF∞ .
lim Df P1Fn , . . . , Pk−1
||PkFn = Df P1F∞ , . . . , Pk−1
n→∞
Proof. Define the measure ν =
Nikodym derivatives
Vn
1
= ·
k
1
k
Pk
i=1 Pi
and vectors V n via the RadonF
n
dPk−1
dP1Fn dP2Fn
,
,
.
.
.
,
dν Fn dν Fn
dν Fn
!
.
Then (1 − 1T V n )dν Fn = k1 dPkFn , and V n is a martingale adapted to the
filtration Fn by standard properties of conditional expectation (under the
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 55
k−1
| 1T v ≤ 1}, we define g : Ck → R by
measure ν). Letting Ck = {v ∈ R+
v
(1 − 1T v).
g(v) = f
1 − 1T v
We see that g is convex (it is a perspective function), and we have
!
!
Z
k−1
Fn
Fn
X
dP
dP
1 Fn
1
1:k−1
Fn
i
.
dν
=
||P
D
P
Eν [g(V n )] = f
1
−
f
1:k−1 k
k
dν Fn
k
dPkFn
i=1
Because V n ∈ Ck for all n, we see that g(V n ) is a submartingale. This gives
Fn
||PkFn ) is non-decreasing in n.
the first result of the lemma, that Df (P1:k−1
Now, assume that the limit in the second statement is finite, as otherwise
the result is trivial. Using f (1) = 0, we have by convexity that for v ∈ Ck ,
v
g(v) = f
(1 − 1T v) + f (1)1T v ≥ f v + 1(1T v)
T
1−1 v
≥ inf f (v + 1(1T v)) > −∞,
v∈Ck
the final inequality a consequence of the fact that f is closed and hence
attains its infimum. In particular, the sequence g(V n ) − inf v∈Ck g(v) is a
non-negative submartingale, and thus
n
n
sup Eν |g(V ) − inf g(v)| = lim Eν g(V ) − inf g(v) < ∞.
v∈Ck
n
n
v∈Ck
With this integrability guarantee, Doob’s second martingale convergence
theorem [7, Thm. 35.5] yields the existence of a vector V ∞ ∈ F∞ such that
0 ≤ lim Eν g(V n ) − inf g(v) = Eν [g(V ∞ )] − inf g(v) < ∞.
n
v∈Ck
v∈Ck
That inf v∈Ck g(v) > −∞ implies Eν [|g(V ∞ )|] < ∞, giving the lemma.
We P
now give the proof of Proposition 1 proper. Let the the base measure
1
i
µ = k ki=1 Pi and let pi = dP
dµ be the associated densities of the Pi . Define
the increasing sequence of partitions P n of X by sets Aαn for vectors α =
(α1 , . . . , αk−1 ) and B with
αj
pj (x)
αj − 1
< n for j = 1, . . . , k − 1
≤
Aαn = x ∈ X |
2n
pk (x)
2
56
DUCHI, KHOSRAVI, RUAN
where we let each αj range over {−n2n , −n2n + 1, . . . , n2n }, and define and
B = (∪α Aαn )c = X \ ∪α Aαn . Then we have
X
pk−1 (x)
α
p1 (x)
,...,
χ
(x)+(n, . . . , n)χB (x),
= lim
n Aαn
n→∞
pk (x)
pk (x)
2
n
n k−1
α∈{−n2 ,...,n2 }
where χA denotes the {0, 1}-valued characteristic function of the set A.
Each term on the right-hand-side of the previous display is Fn -measurable,
where Fn denotes the sub-σ-field generated by the partition P n . Defining
F∞ = σ(∪n≥1 Fn ), we have
Fn
||PkFn
Df (P1:k−1 ||Pk | P n ) = Df P1Fn , . . . , Pk−1
F∞
||PkF∞ = Df (P1 , . . . , Pk−1 ||Pk ) ,
→ Df P1F∞ , . . . , Pk−1
n
where the limiting operation follows by Lemma D.1 and the final equality
p
because of the measurability containment ( ppk1 , . . . , k−1
pk ) ∈ F∞ .
D.3. Proof of Proposition 2. Before proving the proposition, we
state an inequality that generalizes the classical log-sum inequality (cf. [9,
Theorem 2.7.1]).
Lemma D.2. Let f : Rk+ → R be convex. Let a : X → R+ and bi : X →
R+ for 1 ≤ i ≤ k be nonnegative measurable functions. Then for any finite
measure µ defined on X , we have
R
Z
Z
R
bk dµ
b1 dµ
bk (x)
b1 (x)
,...,
dµ(x).
,..., R
≤ a(x)f
adµ f R
a(x)
a(x)
adµ
adµ
Proof. Recall that the perspective function g(y, t) defined by g(y, t) =
tf (y/t) for t > 0 is jointly convex in y and t. The measure ν = µ/µ(X )
defines a probability measure, so that for X ∼ ν, Jensen’s inequality implies
R
R
Z
b1 dν
bk dν
b1 (X)
bk (X)
adν f R
,..., R
≤ Eν a(X)f
,...,
a(X)
a(X)
adν
adν
Z
1
b1 (x)
bk (x)
=
,...,
a(x)f
dµ(x).
µ(X )
a(x)
a(x)
R
R
R
R
Noting that bi dν/ adν = bi dµ/ adµ gives the result.
We use this lemma and Proposition 1 to give Proposition 2. Proposition 1
implies
Df QP1 , . . . , QPk−1 ||QPk
= sup Df QP1 , . . . , QPk−1 ||QPk | P : P is a finite partition of Z .
P
MULTICLASS CLASSIFICATION, INFORMATION, AND SURROGATE RISK 57
It is consequently no loss of P
generality to assume that Z is finite and
k
Z = {1, 2, . . . , m}. Let µ =
i=1 Pi be a dominating measure and let
pi = dPi /dµ. Letting q(j | x) = Q({j} | x) and qPi (j) = QPi ({j}), we
obtain
m
X
qP (i)
qP1 (i)
, . . . , k−1
qPk (i)f
Df QP1 , . . . , QPk−1 ||QPk =
qPk (i)
qPk (i)
i=1
R
m Z
X
(x), · · · , pk−1 (x)]dµ(x)
X q(i | x)[p
R 1
q(i | x)pk (x)dµ(x) f
=
X
X q(i | x)pk (x)dµ(x)
i=1
m Z
X
p1 (x)
pk−1 (x)
q(i | x)f
≤
,...,
pk (x)dµ(x)
pk (x)
pk (x)
X
i=1
by Lemma D.2. Noting that
Pm
i=1 q(i
| x) = 1, we obtain our desired result.
Stanford University
Stanford, California 94305
E-mail: [email protected]
[email protected]
[email protected]
| 7cs.IT
|
arXiv:1512.07515v6 [math.GN] 16 Aug 2016
COUNTABLE TIGHTNESS AND G-BASES ON FREE TOPOLOGICAL
GROUPS
FUCAI LIN, ALEX RAVSKY, AND JING ZHANG
Abstract. Given a Tychonoff space X, let F (X) and A(X) be respectively the free topological group and the free Abelian topological group over X in the sense of Markov. In this paper,
we consider two topological properties of F (X) or A(X), namely the countable tightness and
G-base. We provide some characterizations of the countable tightness and G-base of F (X)
and A(X) for various special classes of spaces X. Furthermore, we also study the countable
tightness and G-base of some Fn (X) of F (X).
1. Introduction
Let F (X) and A(X) be respectively the free topological group and the free Abelian topological
group over a Tychonoff space X in the sense of Markov [30]. For every n ∈ N, by Fn (X) we
denote the subspace of F (X) that consists of all words of reduced length at most n with respect
to the free basis X. The subspace An (X) is defined similarly. We always use G(X) to denote
F (X) or A(X), and Gn (X) to Fn (X) or An (X) for each n ∈ N. Therefore, any statement about
G(X) applies to F (X) and A(X), and about Gn (X) applies to Fn (X) and An (X).
One of the techniques of studying the topological structure of free topological groups is to
clarify the relations of subspaces X, F (X), A(X), Fn (X) and An (X), where n ∈ N. It is well
known that only when the space X is discrete, F (X) and A(X) can be first-countable. Therefore,
the space F (X) is first-countable if and only if X is discrete [20]. Similarly, the groups F (X)
and A(X) are locally compact if and only if the space X is discrete [10]. More generally, P.
Nickolas and M. Tkachenko proved that if one of the groups F (X) or A(X) is almost metrizable,
then the space X is discrete [31]. Further, K. Yamada gave a characterization for a metrizable
space X such that some the spaces Fn (X) and An (X) are first-countable [40].
Recently, Z. Li et al. in [27] proved that for each stratifiable k-space, the group F (X) is of
countable tightness if and only if the space X is separable or discrete. In Section 3, we refine
this result by giving a characterization of a space X such that the countable tightness of the
space F8 (X) implies the countable tightness of the group F (X). Furthermore, since each space
with the countable fan-tightness or the strong Pytkeev property is of countable tightness, we
also discuss the topological properties of the countable fan-tightness and the strong Pytkeev
property of free topological group F (X) or some Fn (X).
Ferrando et al. in [12] introduced the concept of G-base in the frame of locally convex
spaces. Now the concept of G-base plays an important role in the study of function spaces, see
[7, 9, 15, 16, 17, 18, 19, 26]. From [16], we know that the strong Pytkeev property for general
topological groups is closely related to the notion of a G-base. For instance, each topological
group which is a k-space with a G-base has the strong Pytkeev property. In Section 4, we shall
2000 Mathematics Subject Classification. Primary 54H11, 22A05; Secondary 54E20; 54E35; 54D50; 54D55.
Key words and phrases. Free topological group; free Abelian topological group; countable tightness; countable
fan-tightness; G-base; strong Pytkeev property; universally uniform G-base.
The first author is supported by the NSFC (Nos. 11571158, 11201414, 11471153), the Natural Science
Foundation of Fujian Province (Nos. 2016J05014, 2016J01671, 2016J01672) of China and the project of Abroad
Fund of Minnan Normal University. This paper was partially written when the first author was visiting the
School of Computer and Mathematical Sciences at Auckland University of Technology from March to September
2015, and he wishes to thank the hospitality of his host.
1
2
FUCAI LIN, ALEX RAVSKY, AND JING ZHANG
continuously discuss the properties of free topological groups with a G-base, which are motivated
by the following interesting Questions 1.1 and 1.2.
Question 1.1. [15, Question 4.17] Let X be a submetrizable kω -space. Does the group F (X)
have a G-base?
Question 1.2. [15, Question 4.15] For which X the groups A(X) and F (X) have a G-base?
Recently, S.S. Gabriyelyan and J. Ka̧kol in the paper [17] and A.G. Leiderman, V.G. Pestov
, A.H. Tomita in the paper [26] have given an answer to Questions 1.1 and 1.2 respectively.
2. Notations and Terminology
In this section, we introduce the necessary notations and terminology. Throughout this paper,
all topological spaces are assumed to be Tychonoff, unless otherwise is explicitly stated. For
undefined notations and terminology, refer to [3], [11] and [21]. First of all, let N and Q denote
the sets of positive integers and rational numbers, respectively.
Let X be a topological space and A ⊆ X. The closure of a subspace A of X is denoted by
A. The subspace A is called bounded if every continuous real-valued function f defined on the
subspace A is bounded. If the closure of every bounded set in X is compact, then the space X is
called µ-complete. The space X is called a cf -space if every compact subset of it is finite. The
space X is called a k-space provided that a subset C ⊆ X is closed in X if and only if C ∩ K
is closed in K for each compact subset K of the space X. In particular, the space X is called
a kω -space if there exists a family of countably many compact subsets {Kn : n ∈ N} of X such
that each subset F of the space X is closed in X if and only if F ∩ Kn is closed in Kn for each
n ∈ N. A subset A of the space X is sequentially open if each sequence {xn }n∈N in X converging
to a point of A is eventually in A. The space X is called sequential if every sequentially open
subset of X is open. A space X is of countable tightness if whenever A ⊂ X and x ∈ A, there
exists a countable set B ⊂ A such that x ∈ B. A space X is of countable
fan-tightness [1] if
T
for any countable family {An : n ∈ N} of subsets of X satisfying x ∈ n∈N An , it is possible to
S
select a finite set Kn ⊂ An for each n ∈ N, in such a way that x ∈ n∈N Kn . A sequence {xn },
convergent to a point x of X, is called non-trivial, provided all points xn and x are mutually
distinct.
Let P be a family of subsets of a space X. The family P is called a network at a point
x ∈ X if for each open neighborhood U of x in X there exists an element P ∈ P such that
x ∈ P ⊂ U . The family P is called a cs-network [24] at a point x ∈ X if whenever a sequence
{xn : n ∈ N} converges to the point x and U is an arbitrary open neighborhood of the point x
in X there exist a number m ∈ N and an element P ∈ P such that
{x} ∪ {xn : n > m} ⊆ P ⊆ U.
The space X is called csf-countable if X has a countable cs-network at each point x ∈ X. We
call the family P a cs∗ -network at a point x ∈ X [29] of X if whenever a sequence {xn : n ∈ N}
converges to the point x and U is an arbitrary open neighborhood of the point x in X, there are
an element P ∈ P and a subsequence {xni : i ∈ N} of {xn : n ∈ N} such that {x} ∪ {xni : i ∈
N} ⊆ P ⊆ U . Furthermore, the family P is called a k-network [32] if whenever K is a compact
subset of X and U ⊂ X is
S an arbitrary open set containing K then there is a finite subfamily
P ′ ⊆ P such that K ⊆ P ′ ⊆ U . Recall that a regular space X is ℵ0 if X has a countable
k-network. The family P is called a Pytkeev network [33] at a point x ∈ X if P is a network at
x and for every open set U in X and a set A accumulating at x there exists P ∈ P such that
P ⊂ U and P ∩ A is infinite; the family P is a Pytkeev network in X if P is a Pytkeev network
at each point x ∈ X. The space X is said to have the strong Pytkeev property [37] if at each
point of X there is a countable Pytkeev network. The space X is called P0 if X is regular and
has a countable Pytkeev network.
The following implications follow directly from definitions. However, none of them can be
reversed. By [19, Proposition 4.1], we see that a space is first-countable if and only if it has the
strong Pytkeev property and is of countable fan-tightness.
COUNTABLE TIGHTNESS AND G-BASES ON FREE TOPOLOGICAL GROUPS
✲
csf -countable ✛ countable cs∗ -character ✛
✻
ℵ0 -space
✻
strong Pytkeev property ✛
countable fan-tightness
3
P0 -space
❄
✲ countable tightness
✻
sequential space
✲ k-space
✛
kω -space
Definition 2.1. [21] A topological space X is a stratifiable space if X is T1 and, to each open
U in X, on can assign a sequence {Un }∞
n=1 of open subsets of X such that
⊂
U
;
(a) U
Sn∞
(b) n=1 Un = U ;
(c) Un ⊂ Vn whenever U ⊂ V .
Note: Clearly, each metrizable space is stratifiable [21].
We consider the product NN with the natural partial order, i.e., α ≤ β if αi ≤ βi for each
i ∈ N, where α = (αi )i∈N and β = (βi )i∈N . A topological space (X, τ ) has a small base [18] if
there exist a subset M of NN and a family of open subsets U = {Uα : α ∈ M } in X such that
U is a base for X and Uβ ⊂ Uα for all α, β ∈ M with α ≤ β. In particular, we say that (X, τ )
has a G-base [12] if M = NN . If a space has a G-base, then it has a countable cs∗ -character, see
Proposition 4.7.
Given a group G, the letter eG denotes the neutral element of G. If no confusion occurs, we
simply use e instead of eG to denote the neutral element of G.
Let X be a non-empty Tychonoff space. Throughout this paper, X −1 := {x−1 : x ∈ X}
and −X := {−x : x ∈ X}, which are copies of X. Let e be the neutral element of F (X) (i.e.,
the empty
word)
L −1
L andn 0 be that of A(X). For every n ∈ N and an element (x1 , x2 , · · · , xn )
of (X
X
{e}) we call a word g = x1 x2 · · · xn a form.LIn the
L Abelian case, a word
x1 + x2 · · · + xn is also called a form for (x1 , x2 , · · · , xn ) ∈ (X
−X {0})n . This word g is
called reduced if it does not contains e or any pair of consecutive symbols of the form xx−1 or
x−1 x. It follows that if the word g is reduced and non-empty, then it is different from the neutral
element e of F (X). In particular, each element g ∈ F (X) distinct from the neutral element
can be uniquely written in the form g = xr11 xr22 · · · xrnn , where n ≥ 1, ri ∈ Z \ {0}, xi ∈ X,
r2
rn
and xi 6= xi+1 for each i = 1, · · · , n − 1, and the support of g = xr11 xS
2 · · · xn is defined as
supp(g) := {x1 , · · · , xn }. Given a subset K of F (X), we put supp(K) = g∈K supp(g). Similar
assertions
the obvious changes for commutativity) are valid for A(X). For every n ∈ N, let
L (with
L
in : (X
X −1 {e})n → FL
the natural mapping defined by in (x1 , x2 , ...xn ) = x1 x2 ...xn
n (X) beL
for each (x1 , x2 , ...xn ) ∈ (X
X −1 {e})n . WeL
also useL
the same symbol in the Abelian case,
that is, it means the natural mapping from (X (−X) {0})n onto An (X). Clearly, each in
is a continuous mapping.
3. The characterization of countable tightness in free topological groups
In this section, we mainly discuss the countable tightness and countable fan-tightness of free
topological groups. First, we give a characterization of a stratifiable k-space X such that F8 (X)
has countable tightness (or, equivalently, F (X) has countable tightness). Then we show that a
space X must belong to some special class of spaces if F4 (X) is of countable fan-tightness.
The following theorem generalizes a result in [27].
Theorem 3.1. Let X be a stratifiable k-space. Then the following are equivalent:
(1) F8 (X) is of countable tightness;
(2) F (X) is of countable tightness;
(3) The space X is separable or discrete.
4
FUCAI LIN, ALEX RAVSKY, AND JING ZHANG
Proof. Since the equivalence (2) ⇔ (3) was proved in [27], it suffices to show that (1) ⇒ (3).
Assume that X is neither separable nor discrete. Since each stratifiable space has Gδ diagonal, by Šneı̌der Theorem [35], each compact subspace of X is metrizable. Thus X is
sequential. Since the space X is non-discrete, it contains a non-isolated point x ∈ X. This means
that the set {x} is not sequentially open, that is, there exists a non-trivial convergent sequence.
Hence take an arbitrary a non-trivial convergent sequence C := {xn : n ∈ N} ⊂ X with a limit
point x. Moreover, assume that the space X contains no uncountable closed discrete subset.
This means that the extend of the space X is countable. But, by [25], each stratifiable space
is a σ-space, and a σ-space of countable extent is cosmic (that is, has countable network), and,
therefore, separable. Obtained contradiction shows that there exists an uncountable discrete
closed subset D := {dα : α ∈ ω1 } of X. Without loss of generality, we may assume that
C ∩ D = ∅.
For each α ∈ ω1 let fα : ω1 → ω be a function such that fα ↾α : α → ω is a bijection. For any
distinct α, β ∈ ω1 , put
Eα,β := {dβ xm x−1 dβ−1 dα xfα (β) x−1 d−1
α : m ≤ fα (β)}.
Then put
[
E :=
Eα,β .
α,β∈ω1 ,α6=β
Clearly, e 6∈ E, and e ∈ E by the proof of [41, Proposition 2.2]. In order to obtain a contradiction,
it suffices to show that each countable infinite subset B ⊂ E is closed in F (X). Let Y :=
supp(B). The set Y contains the point x, so Y ⊂ C ∪ D ∪ {x}, which implies that F (Y ) is a
k-space by [4, Theorem 3.7]. Since X is a stratifiable space and Y is closed in X, it follows from
[34] that the subgroup F (Y, X) of F (X) generated L
by Y is naturally topologically isomorphic
to F (Y ). Then F (Y ) is a closed k-subspace in F (C
D). Furthermore, we claim that for each
compact subset K of F8 (Y ), the set K ∩ B is finite. Assume on the contrary that there is a
compact subset K in F8 (Y ) such that K ∩ B is infinite. Clearly, the set K ∩ B is a bounded
subset in F8 (Y ), hence the subspace supp(K ∩B) is bounded in Y by [4, Theorem 1.5]. Since the
space Y is paracompact, supp(K ∩ B) is compact in Y . However, the set supp(K ∩ B) contains
infinite many elements d′α s since B is infinite, which is a contradiction with the compactness
of the subspace supp(K ∩ B). Therefore, the subset B is closed in F8 (Y ), that is, the subset
B is closed in F (X). Hence F8 (X) is not of countable tightness since e ∈ E, which is a
contradiction.
Obviously, we have the following corollary.
Corollary 3.2. Let X be a stratifiable k-space. If F8 (X) has the strong Pytkeev property, then
X is separable or discrete.
L
Remark 3.3. Let X := D K, where D is an uncountable discrete space and K is an infinite
compact metric space. Then F3 (X) is first-countable by Theorem 4.5 in [40], hence it is of
countable tightness. However, the space X is not separable and discrete. We do not know
whether F4 (X) is of countable tightness. Therefore, we have the following Question 3.4.
L
Question 3.4. Let X := C
D, where C is a non-trivial convergent sequence with its limit
point and D is an uncountable discrete space. For each n ∈ {4, 5, 6, 7}, does Fn (X) have the
countable tightness?
Furthermore, we also do not know the answer to the following question.
Question 3.5. Let X be a space. If F2 (X) is of countable tightness, does F3 (X) have the
countable tightness?
It is natural to ask whether Theorem 3.1 holds in the class of free Abelian topological groups.
Next we shall give a partial answer to this question.
Theorem 3.6. Let X be a stratifiable k-space. If A4 (X) is of countable tightness, then the set
of all non-isolated points of X is a separable subspace in X.
COUNTABLE TIGHTNESS AND G-BASES ON FREE TOPOLOGICAL GROUPS
5
Proof. Assume on the contrary that the set of all non-isolated points of X is not separable.
Then the set of all non-isolated points of X is not an ℵ1 -compact space since X is a stratifiable
space [21]. Therefore, there exists an uncountable closed discrete subset D := {dα : α < ω1 } in
X, where each dα is a non-isolated point in X. Moreover, since X is a stratifiable k-space, it is
paracompact (and, hence, collectionwise normal by [11, Th.5.1.17]) and sequential. Hence there
exists a family of mutually disjoint open subsets {Uα : α < ω1 } of X such that dα ∈ Uα for each
α < ω1 . Moreover, for each α < ω1 we L
can take a non-trivial sequence {dα (n) : n ∈ N} ⊂ Uα
convergent to a point dα . Let C(ω1 ) := α<ω1 ({dα (n) : n ∈ N} ∪ {dα }). Since X is stratifiable
and C(ω1 ) is closed in X, it follows from [34] that the subgroup A(C(ω1 ), X) of A(X) generated
by C(ω1 ) is naturally topologically isomorphic to the free Abelian topological group A(C(ω1 )).
Then A4 (C(ω1 )) is topologically isomorphic to A(C(ω1 ), X) ∩ A4 (X), which implies that the
tightness of A4 (C(ω1 )) is countable. However, it follows from [39, Theorem 3.2] and [23] that
the tightness of A4 (C(ω1 )) is uncountable, which is a contradiction.
However, the converse of Theorem 3.6 does not hold, see [39, Theorem 3.2].
Moreover, the proof of the following result is similar to [41, Proposition 2.2], and thus, the
proof is omitted in this paper.
Theorem 3.7. For a stratifiable space X, if F8 (X) is a k-space, then X is separable or discrete.
Next, we shall discuss the countable fan-tightness in free topological groups. In contrast
to Theorem 3.1, we shall find that the situation changes dramatically for the countable fantightness in free topological groups. First, we shall give a characterization of a space X such
that G(X) has the countable fan-tightness.
Theorem 3.8. Let X be a space. Then G(X) has the countable fan-tightness if and only if the
space X is discrete.
Proof. We only consider F (X), as the proof of the case of A(X) is quite similar. Since the
sufficiency is obvious, we shall prove only the necessity. In order to obtain a contradiction,
assume the converse. Suppose that F (X) has a countable fan-tightness but the space X is not
S
T
discrete. Clearly, e ∈ n∈N ( i≥n (Fi (X)) \ Fi−1 (X))). However, if for each natural n we take an
S
S
arbitrary finite subset Dn ⊂ i≥n (Fi (X)) \ Fi−1 (X))), then an intersection ( n∈N Dn ) ∩ Fm (X)
S
will be finite for each natural m. Therefore n∈N Dn is closed and discrete in F (X) by [3,
S
Corollary 7.4.3]. Thus e 6∈ n∈N Dn , a contradiction.
It turns out that the countable fan-tightness of F4 (X) imposes strong restrictions on the space
X. Recall that a subspace Y of a space X is said to be P-embedded in X if each continuous
pseudometric on Y admits a continuous extensions over X.
Theorem 3.9. Let X be a space. If F4 (X) has the countable fan-tightness, then X is either
pseudocompact or a cf -space.
Proof. Suppose X is not a cf -space. Then there exists an infinite compact subset C in X. Next
we shall show that X is pseudocompact.
Assume the converse. Then there exists a discrete family {Un : n ∈ N} of non-empty open
subsetsSof the space X. It can be easily verified that the family {Un : n ∈ N} is also discrete,
hence n∈N Un is closed in X. Since the set C is compact, it can intersect at most finitely many
S
Un ’s. Thus without loss of generality, we may assume that C ∩ n∈N Un = ∅. Then the family
{C} ∪ {Un : n ∈ N} is discrete in X.
Since C is an infinite compact set, it contains a non-isolated point x. For each n ∈ N, pick
yn ∈ Un , and put Cn := yn−1 x−1 Cyn . Let
[
Y := C ∪ {yn : n ∈ N}, and Z :=
Cn .
n∈N
Obviously, the set Y is closed, σ-compact and non-discrete in X. Moreover, Y is P -embedded
in X by [36]. By [34], the subgroup F (Y, X) of F (X) generated by Y is naturally topologically
6
FUCAI LIN, ALEX RAVSKY, AND JING ZHANG
isomorphic to F (Y ). Moreover, since Y is a kω -space, it follows from [3, Theorem 7.4.1] that
F (Y ) is also a kω -space. Hence F4 (Y ) is also a kω -space. Next, we claim that Z is closed in
F4 (Y ).
Indeed, for each n ∈ N, let
Kn := C ∪ C −1 ∪ {yi , yi−1 : i ≤ n} ∪ {e}
and
n
z
}|
{
Xn := Kn · . . . · Kn .
Then it follows from the proof of [3, Theorem 7.4.1] that the topology of F (Y ) is determined by
the family of compact subsets {Xn : n ∈ N}, hence the topology of F4 (Y
S)n is determined by the
family {Xn ∩ F4 (Y ) : n ∈ N}. Therefore, we have Z ∩ Xn ∩ F4 (Y ) = i=1 Ci for each n ∈ N,
which shows that Z is closed in F4 (Y ). Therefore, Z has the countable fan-tightness. Moreover,
T
it is easy to see that e ∈ Cn \ {e} for each n ∈ N, that is, e ∈ n∈N Cn \ {e}. Now, if for each
S
S
natural n, Fn ⊂ Cn \ {e} is a finite subset, then e 6∈ n∈N Fn = n∈N Fn , because the topology
of Z is determined by the family {Cn : n ∈ N}, a contradiction.
Corollary 3.10. Let X be a µ- and k-space. If F4 (X) has the countable fan-tightness, then X
is either compact or discrete.
L
Remark 3.11. By L
Theorems 3.1 and 3.9, it is easy to see that F (C
D) has the countable
tightness and F4 (C
D) does not have the countable fan-tightness, where C is a non-trivial
convergent
L sequence with its limit point and D is a countable infinite discrete space. However,
F3 (C
D) has the countable fan-tightness by Remark 3.3.
L Therefore, we can not replace
L
“F4 (X)” with “F3 (X)” in Theorem 3.9. Furthermore, F (C
D) is a P0 -space since F (C
D)
is an k-space with a countable k-network.
Finally, we shall discuss the strong Pytkeev property in free topological groups. It is well
known that if a space has the strong Pytkeev property then it is of countable tightness and is
csf -countable. In [6], the authors showed that a space is first-countable if and only if it has the
strong Pytkeev property and countable fan-tightness. Therefore, it is interesting to discuss the
strong Pytkeev property on free topological groups. First, we shall give a theorem which has
just been proved in [28].
Recall that a space X is said to be Lašnev if it is the closed image of some metric space.
Theorem 3.12. [28] Let X be a non-discrete Lašnev space. Then F4 (X) is of csf -countable if
and only if F (X) is an ℵ0 -space.
Theorem 3.13. Let X be a non-discrete Lašnev space. Then F (X) has the strong Pytkeev
property if and only if F (X) is a P0 -space.
Proof. Obviously, it suffices to show the necessity. Suppose that F (X) has the strong Pytkeev
property. By Theorem 3.12, the group F (X) is an ℵ0 -space, hence F (X) is separable. It follows
from [17, Theorem 1.7] that F (X) is a P0 -space.
We do not know whether an ℵ0 -space with the the strong Pytkeev property is a P0 -space. If
this answer is positive, then we can replace “F (X) has the strong Pytkeev property” by “F4 (X)
has the strong Pytkeev property” in Theorem 3.13.
In [5], T. Banakh posed the following problem.
Problem 3.14. [5] Let X be a (sequential) P0 -space. Is the free topological group over X a
P0 -space?
But we even do not know an answer to the following question.
Question 3.15. Let X be the rational number subspace Q with the usual topology. Is the free
topological group over X a P0 -space?
COUNTABLE TIGHTNESS AND G-BASES ON FREE TOPOLOGICAL GROUPS
7
However, we have the following result.
Theorem 3.16. Let X be a P0 -space. Then F (X) is the union of countably many P0 -subspaces.
L −1 L
Proof. For each n ∈ N, it follows from [5, Corollary 3.12] that (X
X
{e})n is a P0 -space.
For each n ∈ N, since the mapping
in ↾i−1 (Fn (X)\Fn−1 (X)) : i−1
n (Fn (X) \ Fn−1 (X)) → Fn (X) \ Fn−1 (X)
k
is a homeomorphism,
follows from [5, Corollary 3.12] that i−1
n (Fn (X) \ Fn−1 (X)) is a P0 L −1itL
subspaceSin (X
X
{e})n , hence Fn (X) \ Fn−1 (X) is a P0 -subspace in F (X). Since
F (X) = n∈N (Fn (X) \ Fn−1 (X)), F (X) is the union of disjoint countably many P0 -subspaces.
For closing this section, we discuss the strong Pytkeev property in topological spaces. It
is well known that in the class of regular countably compact spaces the property of countable
tightness is equivalent to the countable fan-tightness [2, Corollary 2]. Moreover, each compact
sequential non-first-countable space is of countable tightness, hence it does not have the strong
Pytkeev property. Hence in the class of regular countably compact spaces the property of
countable tightness is not equivalent to the strong Pytkeev property. Moreover, it is well known
that there exists a countably compact non-metrizable space with a point-countable k-network.
It is natural to ask whether in the class of regular countably compact spaces the existence of
a point-countable k-network implies the existence of a point-countable Pytkeev network. The
answer is also negative, see Example 3.17.
Example 3.17. There exists an infinite countably compact space X which satisfies the following
conditions:
(1) The space X contains no infinite compact subset;
(2) The space X has a point-countable k-network;
(3) The space X does not have the strong Pytkeev property.
Proof. By [14], there exists an infinite, countably compact subspace X of βN, the Stone-Cech
compactification of the of the space of natural numbers endowed with discrete topology, such
that N ⊂ X and every compact subset of the space X is finite. Therefore, the space X has a
point-countable k-network. However, the space X does not have the strong Pytkeev property,
see [5].
From the opposite side, recently, Z. Cai and S. Lin has proved that each sequentially compact space with a point-countable k-network is metrizable [8]. Remark, that the proof of [6,
Proposition 1.4], implies that the space X has a point-countable k-network provided it has a
point-countable Pytkeev network. So the next theorem is a counterpart of this result.
Theorem 3.18. Let X be a Hausdorff countably compact space with a point-countable Pytkeev
network. Then the space X is a metrizable compact space.
Proof. Let N be a point-countable Pytkeev network on the space X. We first show the following
claim.
Claim: For each countably compact subset KSof X and arbitrary open subset U with K ⊂ U ,
there exists a finite subfamily F such that K ⊂ F ⊂ U .
Suppose not. For each x ∈ K, let {N ∈ N : x ∈ N } = {Nn (x) : n ∈ ω}. Inductively choose
xn ∈ K such that xn 6∈ Nj (xi ) for i, j < n. Put A := {xn : n ∈ ω}. Since K is countably
compact and A ⊂ K, the set A has a cluster point x⋆ in K. By the definition of Pytkeev
network, it follows that there exists some N ∈ N such that N contains infinitely many xn ’s.
Therefore, we have N = Nj (xi ) for some i and j, and there exists n > i, j such that xn ∈ Nj (xi ),
contradicting the way the xn ’s were chosen.
By Claim and [22, Theorem 4.1], the space X is metrizable (and thus compact).
8
FUCAI LIN, ALEX RAVSKY, AND JING ZHANG
4. Free topological groups with a G-base
In this section, we shall discuss the properties of free topological groups with a G-base,
which are motivated by Questions 1.1 and 1.2. First, we recall a lemma. Then we shall give
a characterization of free topological groups which are k-spaces having a G-base. Let TGG be
the class of all topological groups having a G-base.
Lemma 4.1. [15]
(1) The group
(2) The group
(3) The group
Let G ∈ TGG . Then the following are equivalent:
G is a k-space;
G is a sequential space;
G is metrizable or contains a submetrizable open kω -subgroup.
By Lemma 4.1, we know that the k-property and sequentiality are equivalent in the class
of all topological groups having a G-base. In [15], the authors also said that “It would be
interesting to know whether the k-property and sequentiality are equivalent for the class of all
topological groups having countable cs∗ -character”. Indeed, the answer is negative, see the
following example.
Example 4.2. There exists a topological group G such that it is a k-space with a countable
cs∗ -character. However, G is not sequential.
Proof. Let X be the Stone-Čech compactification βD of any infinite discrete space D. Let
G := F (X) or G := A(X). Obviously, the group G is a k-space by [3, Theorem 7.4.1]. It
follows from a result of [28] that G is of countable cs∗ -character. However, it is well known
that βD is not a sequential space. Since βD is closed in G, the free topological group G is not
sequential.
However, the topological group G in the proof of Example 4.2 does not have the strong
Pytkeev property by Example 3.17, hence it is natural to pose the following question.
Question 4.3. Let a topological group G be a k-space. If G has the strong Pytkeev property, is
it sequential?
In [17], the authors gave an affirmative answer to Question 1.1. The following theorem
complements it.
Theorem 4.4. Let X be a space. Then F (X) is a k-space with a G-base if and only if either
X is discrete or X is a submetrizable kω -space.
Proof. The sufficiency was proved in [17]. It suffices to show the necessity.
Let F (X) be a k-space with a G-base. Then it follows from Lemma 4.1 that F (X) is
metrizable or contains a submetrizable open kω -subgroup. If F (X) is metrizable, then it is
well known that X is discrete. Hence we may assume that F (X)Lis non-metrizable, and then
F (X) contains a submetrizable open kω -subgroup. Then F (X) = α∈Γ Gα , where each Gα is a
submetrizable open kω -subset in F (X). Since each Gα has a countable k-network, it follows that
F (X) has a σ-compact finite k-network. Moreover, it is obvious that F (X) is locally Lindelöf. It
is well known that a locally Lindelöf topological group is paracompact [3, Problem 3.2.A], then
X is paracompact since X is closed in F (X). Thus X is a paracompact space with a σ-compact
finite k-network. Since F (X) is a non-metrizable k-space, it follows from [27, Theorem 4.14]
that X has a countable k-network, hence X is Lindelöf and submetrizable. Therefore, it is easy
to see that X is a submetrizable kω -space.
Remark 4.5. However, there exists a space L
X such that A(X) is a k-space with a G-base and X
is not a Lindelöf-space. Indeed, let X := C
D, where C is a non-trivial convergent sequence
with its limit point and D is an uncountable discrete space D. Then A(X) ∼
= A(C) × A(D),
thus A(X) is a k-space. Since both A(X) and A(D) have G-bases, it follows from [15] that
A(S) × A(D) has a G-base. However, it is obvious that X is not Lindelöf, hence it is not a
kω -space.
COUNTABLE TIGHTNESS AND G-BASES ON FREE TOPOLOGICAL GROUPS
9
By Remark 4.5, we see that we can not replace “F (X)” by “A(X)” in Theorem 4.4. However,
we have the following theorem when we add some additional assumption on the space X.
Theorem 4.6. Let X be a separable space. Then A(X) is a k-space with a G-base if and only
if X is either countable discrete or a submetrizable kω -space.
Proof. We adapt the proof of Theorem 4.4 for the group A(X) instead of F (X). It suffices to
show that X is a submetrizable kω -space if A(X) is a non-metrizable k-space with a G-base.
Since X is separable, A(X) is separable. Similarly to the proof of Theorem 4.4, we see that the
index set Γ in Theorem 4.4 is countable. Therefore, A(X) has a countable k-network, and then
X is a submetrizable kω -space.
Next we consider the topological properties of X such that the free topological group over
X has a G-base.
By the proof of [15, Theorem 3.12], we can easily obtain the following proposition.
Proposition 4.7. If a space X has a G-base at point x ∈ X, then X is of countable cs∗ character at x.
Therefore, we have the following proposition.
Proposition 4.8. Let X be a space. If each Gn (X) has a G-base at point e, then G(X) is
csf -countable.
Proof. It suffices to note that for each compact subset K in G(X) there exists an n ∈ N such
that K ⊂ Gn (X), see [3, Corollary 7.4.4].
An answer to the following question is still unknown for us.
Question 4.9. Let X be a space. If each Gn (X) has a G-base at e, does G(X) have a G-base?
Theorem 4.10. Let X be a collectionwise normal space containing a non-trivial convergent
sequence. If F (X) has a G-base, then X is ℵ1 -compact.
Proof. Suppose that X is not ℵ1 -compact. Hence there exists an uncountable closed discrete
subset D in X. Moreover, by the assumption, there exists a non-trivial convergent sequence
S with its limit point in X. Without loss of generality, we may assume that S ∩ D = ∅. Let
Y = S ∪ D. Since X is collectionwise normal, the subspace Y is a retract of X, and then Y
is P -embedded in X [3, Exercises 7.7.a]. By [34], the subgroup F (Y, X) of F (X) generated by
Y is naturally topologically isomorphic to F (Y ). However, F (Y ) is not of csf -countable by
a result in [28], thus F (Y ) is not cs∗ -countable. Then F (X) is not cs∗ -countable. However,
since F (X) has a G-base, it follows from Proposition 4.7 that F (X) is cs∗ -countable, which is
a contradiction.
Corollary 4.11. Let X be a stratifiable k-space. If F (X) has a G-base, then X is either discrete
or separable.
Proof. Assume that X is not discrete. Since a stratifiable k-space is paracompact and sequential,
X is ℵ1 -compact by Theorem 4.10. By [25], each stratifiable space is a σ-space, and each ℵ1 compact σ-space is cosmic, and, therefore, separable.
Recently, A.G. Leiderman, V.G. Pestov and A.H. Tomita in [26] showed the following two
results:
Theorem 4.12. [26] The free Abelian topological group A(X) on a uniform space X has a
G-base if and only if X has a G-base.
Corollary 4.13. [26] Let X be a metrizable space and the set of all non-isolated points of X is
a σ-compact subset of X. Then A(X) has a G-base.
10
FUCAI LIN, ALEX RAVSKY, AND JING ZHANG
For a metrizable space X, it follows from [4] that A(X) is a k-space if and only if X is a locally
compact space and the set of all non-isolated points of X is separable. From Corollary 4.13, it
is easy to see that there exists a space Y which is not a k-space such that A(Y ) has a G-base.
However, the situation changes much for (non-Abelian) free topological groups. Let X =
L
C
D, where C is a non-trivial convergent sequence with the limit point and D is a closed
discrete space of cardinality ℵ1 . From [28], F4 (X) is not csf -countable, hence F4 (X) does not
have a G-base. In particular, we see that F (X) does not have a G-base. However, we have the
following Theorem 4.15.
By a similar proof of [15, Proposition 2.7], we can obtain the following proposition.
Proposition 4.14.
Q Suppose that, for each n ∈ N, Xn is a space with a G-base. Then the
countable product n∈N Xn has a G-base.
Given a uniformizable space X there is a finest uniformity on X compatible with the topology
of X called the fine uniformity or universal uniformity. A Tychonoff space X is said to have a
uniform G-base if there exists a uniform structure U on X, which induces the topology of X,
such that U has a G-base. In particular, if U is the universal uniformity on X with a uniform
G-base, then we say that X has an universally uniform G-base.
Theorem 4.15. Let X have an universally uniform G-base. Then F2 (X) has a G-base at each
point.
Proof. Since X has an universally uniform G-base,Lit is easy
L to see X has a local G-base at
each point. By Proposition 4.14, we see that (X
X −1 {e})2 has a local G-base at each
point. Then F2 (X) has a local G-base at each point x ∈ X ∪ X −1 since X ∪ X −1 is open
andL
closed L
in F2 (X). It is well known that F2 (X) \ F1 (X) is homeomorphic to a subspace of
(X
X −1 {e})2 , and then F2 (X) has a local G-base at each point x ∈ F2 (X) \ F1 (X) since
F2 (X) \ F1 (X) is open in F2 (X). It suffices to show that F2 (X) has a G-base at e.
Suppose that U is the universally uniformity on X. Then one can take a basis B = {Uα :
α ∈ NN } for U such that for any α and β in NN with α ≤ β, we have Uβ ⊂ Uα .
For each α ∈ NN , let Wα = {xε y −ε : (x, y) ∈ Uα , ε = ±1}. Then the family {Wα : α ∈ NN }
is a base at e in F2 (X) by [40]. Obviously, {Wα : α ∈ NN } satisfies that for any α and β in NN
with α ≤ β, Wβ ⊂ Wα . Therefore, {Wα : α ∈ NN } is a local G-base at e.
However, the following question is still unknown for us.
Question 4.16. Let X be a space. If F2 (X) has a G-base, does F3 (X) have a G-base?
Acknowledgements. The authors wish to thank professors Salvador Hernández and Boaz
Tsaban for telling us some information of the paper [9]. Moreover, the authors wish to thank
professor Chuan Liu for reading parts of this paper and making comments. Finally, we hope to
thank professor Shou Lin for finding a gap in our proof of Theorem 3.18 in the previous version
and giving some key for us to supplement the proof.
References
[1] A.V. Arhangel’skiı̌, Hurewicz spaces, analytic sets and fan-tightness of spaces of functions, Soviet Math.
Dokl., 33(2)(1986), 396–399.
[2] A.V. Arhangel’skiı̌, A. Bella, Countable fan-tightness versus countable tightness, Comment. Math. Univ.
Carolinae, 37(3)(1996), 567–578.
[3] A.V. Arhangel’skiı̌, M.G. Tkachenko, Topological Groups and Related Structures, Atlantis Press and World
Sci., Paris, 2008.
[4] A.V. Arhangel’skiı̌, O.G. Okunev, V.G. Pestov, Free topological groups over metrizable spaces, Topology
Appl., 33(1989), 63–76.
[5] T. Banakh, P0 -spaces, Topology Appl., 195(2015), 151–173.
[6] T. Banakh, The strong Pytkeev property in topological spaces, http://arXiv:1412.4268v1.
[7] T. Banakh, ω ω -bases in topological and uniform spaces, http://arxiv:1607.07978v1.
[8] Z. Cai, S. Lin, Sequentially compact spaces with a point-countable k-network, Topology Appl., 193(2015),
162–166.
COUNTABLE TIGHTNESS AND G-BASES ON FREE TOPOLOGICAL GROUPS
11
[9] C. Chis, M. Vincenta Ferrer, Salvador Hernández, Boaz Tsaban, The character of topological groups, via
bounded systems, Pontryagin-van Kampen duality and pcf theory, J. Algebra, 420(2014), 86–119.
[10] R.M. Dudley, Continuity of homomorphisms, Duke Math. J., 28(1961), 587–594.
[11] R. Engelking, General Topology (revised and completed edition), Heldermann Verlag, Berlin, 1989.
[12] J.C. Ferrando, J. Ka̧kol, M. López Pellicer, S.A. Saxon, Tightness and distinguished Fréchet spaces, J. Math.
Anal. Appl., 324(2006), 862–881.
[13] P. Fletcher, W.F. Lindgren, Quasi-uniform spaces, Marcel Dekker, New York, 1982.
[14] Z. Frolı́k, Generalizations of the G-property of complete metric spaces, Czech. Math. J., 10(1960), 359–379.
[15] S.S. Gabriyelyan, J. Ka̧kol, A. Leiderman, On topological groups with a small base and metrizability, Fund.
Math., 229(2015), 129–158.
[16] S.S. Gabriyelyan, J. Ka̧kol, A. Leiderman, The strong Pytkeev property for topological groups and topological
vector spaces, Monatsh Math., 175(2014), 519–542.
[17] S.S. Gabriyelyan, J. Ka̧kol, On topological spaces and topological groups with certain local countable networks, Topology Appl., 190(2015), 59–73.
[18] S.S. Gabriyelyan, J. Ka̧kol, A. Kubzdela and M. Lopez Pellicer, On topological properties of Fréchet locally
convex spaces with the weak topology, Topology Appl., 192(2015), 123–137.
[19] S.S. Gabriyelyan, J. Ka̧kol, On B-space and related concepts, Topology Appl., 191(2015), 178–198.
[20] M.I. Graev, Free topological groups, In: Topology and Topological Algebra, Translations Series 1, vol. 8
(1962), pp. 305–364. American Mathematical Society. Russian original in: Izvestiya Akad. Nauk SSSR Ser.
Mat., 12(1948), 279–323.
[21] G. Gruenhage, Generalized metric spaces, In: K. Kunen, J. E. Vaughan(Eds.), Handbook of Set-Theoretic
Topology, Elsevier Science Publishers B.V., Amsterdam, 1984, 423–501.
[22] G. Gruenhage, E.A Michael, Y. Tanaka, Spaces determined by point-countable covvers, Pacific J. Math.,
113(1984), 303–332.
[23] G. Gruenhage, Y. Tanaka, Products of k-spaces and spaces of countable tightness, Trans. Amer. Math. Soc.,
273(1982), 299–308.
[24] J.A. Guthrie, A characterization of ℵ0 -spaces, General Topology Appl., 1(1971), 105–110.
[25] Y. Kanatani, N. Sasaki, J. Nagata, New characterizations of some generalized metric spaces, Math Japonica,
30(1985), 805–820.
[26] A.G. Leiderman, V.G. Pestov , A.H. Tomita, On topological groups admitting a base at indentity indexed
with ω ω , http://arxiv: 1511.07062v1.
[27] Z. Li, F. Lin, C. Liu, Networks on free topological groups, Topology Appl., 180(2015), 186–198.
[28] F. Lin, C. Liu, J. Cao, Weak Countability Axioms in Free Topological Groups, submitted.
[29] S. Lin, Y. Tanaka, Point-countable k-networks, closed maps, and related results, Topology Appl., 59(1994),
79–86.
[30] A.A. Markov, On free topological groups, Izv. Akad. Nauk SSSR Ser. Mat., 9(1945), 3–64 (in Russian);
Amer. Math. Soc. Transl., 8(1962), 195–272.
[31] P. Nickolas, M. Tkachenko, Local compactness in free topological groups, Bull. Austral. Math. Soc.,
68(2)(2003), 243–265.
[32] P. O′ Meara, On paracompactness in function spaces with the compact-open topology, Proc. Amer. Math.
Soc.,29(1971), 183–189.
[33] E.G. Pytkeev, Maximally decomposable spaces, Trudy Mat. Inst. Steklov., 154(1983), 209–213.
[34] O.V. Sipacheva, Free topological groups of spaces and their subspaces, Topology Appl., 101(2000), 181–212.
[35] V. Šneı̌der, Continuous images of Souslin and Borel sets; metrization theorems, Dokl. Acad. Nauk USSR,
50(1945), 77–79.
[36] M. G. Tkachenko, On a spectral decomposition of free topological groups, Usp. Mat. Nauk, 39(2)(1984),
191–192.
[37] Boaz Tsaban, L. Zdomskyy, On the Pytkeev property in spaces of continuous functions (II), Houston J.
Math., 35(2009), 563–571.
[38] K. Yamada, Characterizations of a metrizable space such that every An (X) is a k-space, Topology Appl.,
49(1993), 74–94.
[39] K. Yamada, Tightness of free Abelian topological groups and of finite product of sequntial fans, Topology
Proc., 22(1997), 363–381.
[40] K. Yamada, Metrizable subspaces of free topological groups on metrizable spaces, Topology Proc., 23(1998),
379–409.
[41] K. Yamada, The natural mappings in and k-subspaces of free topological groups on metrizable spaces,
Topology Appl., 146-147(2005), 239–251.
12
FUCAI LIN, ALEX RAVSKY, AND JING ZHANG
(Fucai Lin): School of mathematics and statistics, Minnan Normal University, Zhangzhou 363000,
P. R. China
E-mail address: [email protected]; [email protected]
(Alex Ravsky): Pidstrygach Institute for Applied Problems of Mechanics and Mathematics of
NASU, Naukova 3b, Lviv, 79060, Ukraine
E-mail address: [email protected]
(Jing Zhang): School of mathematics and statistics, Minnan Normal University, Zhangzhou 363000,
P. R. China
E-mail address: [email protected]
| 4math.GR
|
Approximately Optimal Continuous-Time Motion Planning and Control
via Probabilistic Inference
Mustafa Mukadam, Ching-An Cheng, Xinyan Yan, and Byron Boots
arXiv:1702.07335v2 [cs.RO] 27 Feb 2017
2D Environment
Abstract— The problem of optimal motion planing and
control is fundamental in robotics. However, this problem is
intractable for continuous-time stochastic systems in general
and the solution is difficult to approximate if non-instantaneous
nonlinear performance indices are present. In this work, we
provide an efficient algorithm, PIPC (Probabilistic Inference
for Planning and Control), that yields approximately optimal
policies with arbitrary higher-order nonlinear performance
indices. Using probabilistic inference and a Gaussian process
representation of trajectories, PIPC exploits the underlying
sparsity of the problem such that its complexity scales linearly
in the number of nonlinear factors. We demonstrate the
capabilities of our algorithm in a receding horizon setting with
multiple systems in simulation.
Robot View
(a)
I. I NTRODUCTION
A fundamental goal in robotics is to efficiently compute
trajectories of actions that drive a robot to achieve some
desired behavior. We seek a control policy in a multi-stage
decision problem [1] that can maximize performance indices
that describe, for example, the smoothness of motion, energy
consumption, or the likelihood of avoiding an obstacle.
Hierarchical planning and control has been used to solve
this problem in practice [2]. The idea is to first generate a
desired state sequence [3]–[9] without considering full system dynamics, and then design a robust low-level controller
for tracking. Because the dynamic constraints are relaxed, it
becomes possible for an algorithm to plan a trajectory that
satisfies complicated, higher-order performance indices [8]–
[10], offering greater flexibility in system design. Samplingbased planning techniques can even provide formal guarantees such as probabilistically complete solutions [3], [4].
However, recent work has started to challenge this classical
viewpoint by incorporating more dynamic constraints within
trajectory planning in search of solutions with improved
optimality [11], [12].
A theoretically elegant approach would be to address
both the planning and control problems within a stochastic
optimal control framework. Unfortunately, since the states
and actions are coupled through system dynamics, exact
solutions become intractable with the exception of simple
cases known as linearly solvable problems [13].1
These challenges have motivated researchers to find approximate solutions rather than directly approximating the
original problems with hierarchical approaches. One simple
Mustafa Mukadam, Ching-An Cheng, Xinyan Yan, and Byron Boots
are affiliated with the Institute for Robotics and Intelligent Machines,
Georgia Institute of Technology, Atlanta, GA 30332, USA. {mmukadam3,
cacheng, xyan43}@gatech.edu, [email protected].
1 Affine systems with quadratic instantaneous control cost, or fully controllable discrete-time systems.
(b)
(c)
Fig. 1: PIPC used on (a) a 2D holonomic robot (blue) to
reach goal (red) in a 2D environment with dynamic obstacles,
where executed trajectory is in green and current planned
horizon is in black, (b) a 7-DOF WAM arm, and (c) a
PR2’s right arm where the semitransparent arm is the goal
configuration and dotted blue end effector trajectory is the
current planned horizon.
approach is direct policy search [14], [15], which uses
first-order information to find a locally optimal policy. To
improve the convergence rate, differential dynamic programming (DDP) has been widely adopted as the foundation
of locally optimal algorithms [16]–[18], which solve local linear-quadratic Gaussian (LQG) subproblems and iteratively improve these suboptimal solutions. However, for
continuous-time systems, these algorithms would require
inefficient high-frequency sampling to construct the LQG
subproblems, even when the given problem is close to a LQG
(e.g. a performance index with only a small set of nonlinear
factors, or dynamics with a small amount of nonlinearity).
Compared with the hierarchical approach, these algorithms
impose a strict structural assumption: they are only applicable to problems that measure performance as an integral of
instantaneous functions.
In this paper, we propose a novel approximately optimal
approach to continuous-time motion planning and control
that can handle costs expressed as arbitrary higher-order
nonlinear factors and exploit a problem’s underlying sparse
structure. Specifically, we consider problems with a performance index expressed as the product of an exponential-
quadratic factor for instantaneous costs and a finite number
of possibly higher-order nonlinear factors, and provide an algorithm that has linear complexity in the number of nonlinear
factors. Moreover, we show the approximately optimal policy
can be computed by posterior inference on a probabilistic
graphical model, which is a dual to the performance index.
We convert these theoretical results into a practical algorithm called Probabilistic Inference for Planning and Control
(PIPC) that recursively updates the approximately optimal
policy as more information is encountered. To evaluate our
approach, we employ PIPC on both Markov decision processes (MDPs) and partially-observable MDPs (POMDPs)
in dynamic environments with multiple simulated systems
(see Fig. 1).
{z0 , u0 , zδt , · · · , zt } be the history of observations and actions until time t.2 As shorthand, we use boldface to denote
the time trajectory of a variable, and π(u|h) to denote
the collection of time-varying causal (stochastic) policies
πt (ut |ht ) for all t.
We formulate the motion planning and control problem as
a finite-horizon stochastic optimization problem over π. Let
pπ be the distribution of x and u under the stochastic policy
π and system dynamics, and S be a finite set of time indices.
Here the goal is to find an optimal policy π to maximize the
performance index
"
#
Y
max J(x0 ) = max Epπ ψ(x, u)
φS (xS , uS ) . (1)
A. Related Work
The objective function in (1) is defined as the expectation of
the product of two types of factors: a Gaussian process factor
ψ(·) and a higher-order nonlinear factor φS (·). These two
factors, described below, cover many interesting behaviors
that are often desired in planning and control problems.
Our algorithm contributes to a growing set of research that
seeks to reframe planning and control problems as probabilistic inference [19]. Work in this area has formed a new class
of approximately optimal algorithms that leverage tools from
approximate probabilistic inference, including expectation
propagation [20] and expectation maximization [21], [22].
A common framework based on KL-minimization [23], [24]
summarizes the above algorithms as well as approaches like
path-integral control [13].
We contribute to this field in the following ways. First,
we extend the performance index for control algorithms
to incorporate nonlinear factors with arbitrary higher-order
connections in time. In contrast to our approach, the methods
mentioned above generally assume that the performance indices factor into instantaneous terms, and thus require dense
sampling to solve continuous-time problems. Second, we
provide a new approach to derive a Gaussian approximation
based on Laplace approximation and Gaussian processes.
Third, we define a new class of optimal control problems,
called gLEQG (generalized Linear-Exponential-QuadraticGaussian), that are solvable after being transformed into their
dual probabilistic representation. In particular, we show that
gLEQG admits a solution given by posterior inference. This
theoretical result, discussed in Section III-C, closes the gap
in the duality between optimal control and inference.
This rest of the paper is structured as follows. We begin
in Section II by defining the objective function in joint
planning and control problems. Then, in Section III, we
present our main results in approximately optimal motion
planning and control. In Section IV, these theoretical results
are summarized into an online algorithm PIPC that can
perform simultaneous planning and control for partially observable stochastic linear systems in dynamic environments.
To validate our algorithm, we present the implementation
details and experimental results in Section V and Section
VI. Finally, Section VII concludes the paper.
II. T HE P ROBLEM OF M OTION P LANNING AND C ONTROL
We begin by introducing some notation. Let xt , ut , and
zt be the state, action, and observation of a continuoustime partially-observable system at time t, and let ht =
π
π
S∈S
A. Higher-order Nonlinear Factors φS (·)
We define factors of the form
φS (·) = exp(−kfS (·)k2 ),
(2)
to model nonlinear, higher-order couplings frequently used
in planing problems, where fS (·) is a differentiable nonlinear
function defined on a finite number of time indices S ∈ S.
The structure of φS (·) can model many performance indices
in planning: for example, a simple nonlinear cost function
at a single time instance, a penalty based on the difference
between the initial and the terminal states/actions, a penalty
to enforce consistency across landmarks in time, or the cost
of a robot-obstacle collision. As each factor φS (·) depends
only on a finite number of states or actions, we refer to the
corresponding states xS and actions uS as support states or
support actions.
B. Gaussian Process Factors ψ(·)
The Gaussian process factor ψ(·) is a generalization of the
exponential-of-integral cost function in the optimal control
literature [25]. To illustrate, here we consider a special case
ψ(·) = ψ(u). A joint factor between x and u as in (1) can
be defined similarly.
u
Let GPu (ut |mut , Kt,t
0 ) be a Gaussian process [26], where
0
u
u
∀t, t ∈ R, E[ut ] = mut and C[ut , ut0 ] = Kt,t
0 . Let Pt,t0 be
u
the (positive definite) RGreen’s function of Kt,t0 satisfying,
u
u
∀t, t0 ∈ R, δt,t0 = Kt,s
Ps,t
0 ds, where δ is the Dirac
delta distribution and the integral is over the length of the
trajectory. We define the Gaussian process factor ψ(u) as
ZZ
u
u
0
0 (us0 − ms0 )dsds
ψ(u) = exp −
(us − mus )T Ps,s
. (3)
Loosely speaking, we call (3) the probability of a trau
jectory u from GPu (ut |mut , Kt,t
0 ). Note that this notation
2 Here we assume the measurements z are taken in discrete time at time
t
t with sampling interval δt, and ut is a constant continuous-time trajectory
in time [t, t + δt).
does not necessarily imply that u is a sample path of
u
GPu (ut |mut , Kt,t
0 ); rather, we use (3) as a metric between
u
u and m . Intuitively, the maximization in (1) encourages
u to be close to mu in terms of the distance weighted by
u
Pt,t
0.
Solving a stochastic optimization problem with (3) in the
u
objective function is intractable in general, because Pt,t
0 is
only implicitly defined. However, here we show that when
GPu is the sum of a Gaussian white noise process and a linearly transformed Gauss-Markov process, the problem is not
only tractable but can also extend the classical exponentialof-integral cost to model higher-order behaviors.
u
This is realized by defining GPu (ut |mut , Kt,t
0 ) through
a linear stochastic differential equation (SDE). Let yt
be the hidden state of ut (e.g. its higher-order derivatives) and p(y0 ) = N (y0 |my0 , K0y ) be its prior. We set
u
GPu (ut |mut , Kt,t
0 ) as the solution to
dyt = (Dyt + η)dt + Gdω
ut = Hyt + rt + νt
(4)
in which D, η, G, H are (time-varying) system matrices,
rt is control bias, dω is a Wiener process, and νt is
a Gaussian white noise process GPν (0, Qν δt,t0 ). In other
u
words, the Gaussian process GPu (ut |mut , Kt,t
0 ) has mean
and covariance functions:
mut = rt + Hmyt
u
Kt,t
0
= Qν δt,t0 +
y
T
HKt,t
0H
(5)
(6)
y
in which GPy (myt , Kt,t
0 ) is another Gaussian process with
Z t
y
y
mt = Φy (t, t0 )m0 +
Φy (t, s)ηs ds
(7)
t0
y
y
0
T
Kt,t
0 = Φy (t, t0 )K0 Φy (t , t0 ) +
Z min(t,t0 )
Φy (t, s)Gs GTs Φy (t0 , s)T ds
(8)
t0
and Φy (t, s) is the state transition matrix from s to t with
respect to D. For derivations, please refer to [27] and therein.
The definitions (5) and (6) contain the exponential-ofintegral cost [25]
Z
ψ(u) = exp − (us − rs )T Q−1
(u
−
r
)ds
s
s
ν
as a special case, which can be obtained by setting H = 0
u
−1
(i.e. Pt,t
0 = Qν ). In general, it assigns the action ψ(u)
to be close to r, even in terms of higher-order derivatives
(or their hidden states). This leads to a preference toward
smooth control signals. By extension, a joint factor between
x and u would also encourage smooth state trajectories (i.e.
smaller higher-order derivatives of the state).
Constructing the Gaussian process factor by SDE results in
one particularly nice property: If we consider the joint Gaussian process of yt and ut , then its Green’s function is sparse.
To see this, let θt = (ut , yt ) and θ = {θ1 , θ2 , . . . , θN } and
define ψ(θ) as its Gaussian process factor similar to (3).
Then the double integral in ψ(θ) can be broken down into
the sum of smaller double integrals, or factorized as
ψ(θ) = ψ̃(θ0 )
N
−1
Y
ψ̃(θi , θi+1 )
(9)
i=1
where ψ̃(·) has a similar exponential-quadratic form but over
a smaller time interval [ti , ti +δt]. In other words, if we treat
each θi as a coordinate, then the exponent of ψ(θ) can be
written as a quadratic function with a tridiagonal Hessian
matrix (please see [27] for details). This sparse property
will be the foundation of the approximation procedure and
algorithm proposed in Section III and IV.
III. A PPROXIMATE O PTIMIZATION AS I NFERENCE
The mixed features from both planning and control domains in (1) present two major challenges: the optimization over continuous-time trajectories and the higher-order,
nonlinear factors φS (·). The former results in an infinitedimensional problem, which often requires a dense discretization. The latter precludes direct use of algorithms
based on Bellman’s equation, because the factors may not
factorize into instantaneous terms.
In this work, we propose a new approach inspired by
approximate probabilistic inference. The goal here is to
derive an approximation to the problem in (1), in the form
"
#
Y
max Ep̂π ψ(x, u)
φ̂S (xS , uS ) ,
(10)
π
S∈S
where φ̂S (·) is a local exponential-quadratic approximation
of φS (·) and p̂π is a Gaussian process approximation of
pπ . We call the problem in (10) “gLEQG” as it generalizes LEQG (Linear-Exponential-Quadratic-Gaussian) [25] to
incorporate higher-order exponentials in the form of (3).
In the rest of this section, we show how gLEQG can be
derived by using the probabilistic interpretation [20] of the
factors in (1). Further, we show this problem can be solved in
linear time O(|S|) and its solution can be written in closedform as posterior inference.
A. Probabilistic Interpretation of Factors
We begin by representing each factor in (1) with a probability distribution [21]. First, for φS (·), we introduce additional fictitious observations eS such that p(eS |xS , uS ) ∝
φS (xS , uS ). These new variables eS can be interpreted as the
events that we wish the robot to achieve and whose likelihood
of success is reflected proportionally to φS (·). Practically,
they help us keep track of the message propagation over the
support state/action in later derivations. Second, we rewrite
the Gaussian process factor ψ(u) to include the hidden state
yt in (4), as a joint Gaussian process factor q(u, y).3 With
the introduction of yt , the joint Gaussian process q(u, y) has
the sparse property in (9) that we desired.
Now, we rewrite the stochastic optimization (1) in the
new notation. Let eS = {eS }S∈S and ξ = (x, y, u),
3 This
step can be carried similarly as the construction of ψ(u).
and let p(x|u) and p(z|x) be the conditional distributions
defined by the system dynamics and the observation model,
respectively. It can be shown that (1) is equivalent to
Z
(11)
max q(z, ξ|eS )π(u|h)dξdz
π
in which we define a joint distribution
Y
q(z, ξ, eS ) = q(ξ)p(z|x)
p(eS |xS , uS )
(12)
S∈S
with likelihoods p(z|x) and p(eS |xS , uS ), and a prior on the
continuous-time trajectory ξ
q(ξ) = p(x|u)q(u, y).
(13)
Before proceeding, we clarify the notation we use to simplify
writing. We use q to denote the ad hoc constructed Gaussian
process factor (e.g. in (3)) and use p to denote the probability
distribution associated with the real system. As such, q does
not always define an expectation, so the integral notation
(e.g. in (11)) denotes the expectation over p and π that are
well-defined probability distributions. But, with some abuse
of notation, we will call them both Gaussian processes, since
our results depend rather on their algebraic form.
B. Gaussian Approximation
Let ξS = {ξS }S∈S and ξ̄S = ξ\ξS . To derive the gLEQG
approximation to (1), we notice, by (12), q(z, ξ|eS ) in (11)
can be factorized into
q(z, ξ|eS ) = q(z, ξ̄S |ξS )q(ξS |eS )
(14)
where have used the Markovian property in Section IIIA i.e. given ξS , eS is conditionally independent of other
random variables. Therefore, if q(z, ξ̄S |ξS ) and q(ξS |eS )
can be reasonably approximated as Gaussians, then we can
approximate (1) with (10).
However, q(z, ξ̄S |ξS ) and q(ξS |eS ) have notably different
topologies. q(z, ξ̄S |ξS ) is a distribution over continuous-time
trajectories, whereas q(ξS |eS ) is a density function on finite
number of random variables. Therefore, to approximate (1),
we need to find a Gaussian process q̂(z, ξ̄S |ξS ) and a
Gaussian density q̂(ξS |eS ).
1) Gaussian Process Approximation: We derive the Gaussian process approximation q̂(z, ξ) to q(z, ξ). With this
result, the desired conditional Gaussian process q̂(z, ξ̄S |ξS )
is given closed-form.
First we need to define the system dynamics p(x|u) and
the observation model p(z|x). For now, let us assume that
the system is governed by a linear SDE
dx = (Ax + Bu + b)dt + F dw
z = Cx + v
q(u, y) is also Gaussian process, we have a Gaussian process
prior on z and ξ:
q(z, ξ) = p(x, z|u)q(u, y),
(16)
In this case, no approximation is made and therefore
q̂(z, ξ̄S |ξS ) = q(z, ξ̄S |ξS ).
In the case of nonlinear systems, one approach is to
treat (15) as its local linear approximation and derive
q̂(z, ξ) = p̂(x, z|u)q(u, y), where p̂(x, z|u) is the solution
to the linearized system. Alternatively, we can learn the
conditional distribution p̂(x, z|u) from data directly through
Gaussian process regression [26]. However, since our main
purpose here is to show the solution when p̂(x, z|u) is
available, from now on we will assume the system is linear
and given by (15).
2) Gaussian Density Approximation: Unlike q(z, ξ̄S |ξS ),
the approximation to q(ξS |eS ) is more straightforward. First,
because q(ξS |eS ) may not be available in closed form, we
approximate q(ξS |eS ) with q̃(ξS |eS )
Y
q(ξS |eS ) ∝ q(ξS )
p(eS |xS , uS )
S∈S
≈ q̂(ξS )
Y
S∈S
p(eS |xS , uS ) ∝ q̃(ξS |eS )
(17)
where q̂(ξS ) is the marginal distribution of q̂(z, ξ), found
in the previous section. Given (17), we then find a Gaussian
approximation q̂(ξS |eS ) of q̃(ξS |eS ) via a Laplace approximation [28].
For the nonlinear factor from (2), a Laplace approximation
of q̃(ξS |eS ) amounts to solving a nonlinear least-squares
optimization problem. Using the sparsity of the structured
Gaussian processes defined by SDEs, the optimization can
be completed using efficient data structures in O(|S|) [9].
For space constraints, we omit the details here; please see
Appendix A and [9] for details.
3) Summary: The above approximations allow us to approximate (12) with a Gaussian distribution
Y
q̂(z, ξ, eS ) = p̂(z, x|u)q(u, y)
p̂(eS |xS , uS ). (18)
S∈S
In (18), p̂(z, x|u) is the Gaussian process approximation of the system, which is exact when the system is linear, and p̂(eS |xS , uS ) is proportional to the
exponential-quadratic factor φ̂S (xS , uS ) in (10). Moreover,
it can be shown that q̂(z,Qξ, eS ) is a Laplace approximation of p̂(z, x|u)q(u, y) S∈S p(eS |xS , uS ) in terms of
continuous-time trajectory z and ξ.
C. Finding an Approximately Optimal Policy
(15)
in which A, B, b, F , C are (time-varying) system matrices,
dw is a Wiener process, and v is Gaussian noise with
covariance Qv . When a prior is placed on x0 (similar to
Section II-B) it can be shown that the solution to (15)
p(x, z|u) = p(z|x)p(x|u) is a Gaussian process. Since
Substituting the results in Section III-B into (11), we have
the approximated optimization problem
Z
max q̂(z, ξ|eS )π(u|h)dξdz.
(19)
π
By (18), one can show that (19) is equivalent to the problem
in (10), but expressed in probabilistic notation.
However, by writing the problem probabilistically, we can
avoid the algebraic complications arising from attempting
to solve the Bellman’s equation of (10), which, because of
higher-order factors, requires additional state expansion. This
simplicity is reflected in the optimality condition for (19):
πt∗ (ut |ht ) = δ(ut − u∗t (ht ))
Z
∗
ut (ht ) = argmax q̂(z, ξ|eS )π ∗ (ūt |h)dxdydzdūt
ut
= argmax q̂(ut |ht , eS )
t0 ti
t
ti+1
ti + th
tmax
Fig. 2: Time-line with PIPC, where a system that started at
t0 , is currently at time t ∈ [ti , ti+1 ] between support points
ti and ti+1 in δt resolution. In a receding horizon setting, t+
represents the receding horizon window [ti , ti +th ], and tmax
is the (infinite) final time when the algorithm terminates. In
a finite horizon setting, th = tmax − ti .
(20)
ut
in which ūt denotes u\{ut } and δ is Dirac delta distribution.
From the last equality in (20), we see that the solution to
the maximization problem coincides with the mode of the
posterior distribution q̂(ut |ht , eS ). As a result, the optimal
policies for time t can be derived forward in time, by
performing inference without solving for the future policies
first. Please see Appendix B for the proof.
We call this property the duality between gLEQG and
inference. This result seems surprising, but similar ideas can
be traced back to the duality between the optimal control and
estimation [17], [20], in which the optimal value function of
a linear quadratic problem is computed by backward message
propagation without performing maximization.
Compared with previous work, a stronger duality holds
here: gLEQG is dual to the inference problem on the
same probabilistic graphical model defined by the random
variables in Section III-A. This nice property is the result
of the use of an exponential performance index, and enables
us to handle higher-order factors naturally without referring
to ad hoc derivations based on dynamic programming on
extended states.
Our posterior representation of the policy can also be
found in [20], [29], or can be interpreted as one step of
posterior iteration [24]. In [20], this results from the approximation of the optimal value function, but its relationship to
the overall stochastic optimization is unclear. In [29], the
posterior representation is reasoned from the notion of a
predictive policy representation without further justification
of its effects on the whole decision process. Here we derive
the policy based on the assumption that the associated
distribution of (1) can be approximated by a Gaussian
(18). Therefore, the condition on which the approximate
policy remains valid can be more easily understood or even
enforced, as discussed later in Section IV-B.
IV. P ROBABILISTIC M OTION P LANNING AND C ONTROL
In Section III, we show that if q(z, ξ|e) can be approximated well by a Gaussian distribution, the stochastic
optimization in (1) can be approximately solved as posterior
inference (20). This representation suggests that the approximately optimal policy can be updated recursively through
Kalman filtering.
A. Recurrent Policy Inference as Kalman Filtering
The approximately optimal policy in (20) can be viewed
as the belief about the current action ut given the history ht
and the fictitious events eS . Here we exploit the Markovian
structure underlying q̂(z, ξ|eS ) to derive a recursive algorithm for updating the belief q̂(ξt |ht , eS ). Given the belief,
the policy can be derived by marginalization. First, for t = 0,
we write
q̂(ξ0 |h0 , eS ) ∝ p(z0 |ξ0 )q̂(ξ0 |eS )
Z
= q(z0 |ξ0 ) q(ξ0 |ξS )q̂(ξS |eS )dξS
in which q(zt |ξt ) = p(zt |xt ) and q(ξ0 |ξS ) is the conditional
distribution defined by (16). After initialization, the posterior q̂(ξt |ht , eS ) can be propagated through prediction and
correction, summarized together in one step as
q̂(ξt+δt |ht+δt , eS ) ∝ p(zt+δt |ξt+δt )q̂(ξt+δt |eS )
Z
= q(zt+δt |ξt+δt ) q̂(ξt+δt |ξt , eS )q̂(ξt |ht , eS )dξt
(21)
in which the transition is given by
q̂(ξt+δt |ξt , eS ) ∝ q̂(ξk , ξt+δt |eS )
Z
= q(ξk , ξt+δt |ξS )q̂(ξS |eS )dξS
(22)
and q(ξk , ξt+δt |ξS ) is given by (16) [30], [31]. Because
of the Markovian structure in (9), the integral (22) only
depends on two adjacent support states/actions of t and can
be computed in constant time. Note, in q̂(ξt |ht , eS ) in (21),
the action ut is actually conditioned on the action taken
u∗t (ht ). This notation is adopted to simplify the writing.
Thus, we can view (21) as Kalman filtering with transition
dynamics q̂(ξt+δt |ξt , eS ) and observation process q(zt |ξt ).
This formulation gives us the flexibility to switch between
open-loop and closed-loop policies. That is, before a new
observation zt+δt is available, (22) provides a continuoustime open-loop action trajectory during the interval (t, t+δt).
This recurrent policy inference is based on the assumption
that q̂(z, ξ|eS ) is accurate. Although this assumption is not
necessarily true in general, it is a practical approximation if
the belief about current state p(xt |ht ) is concentrated and
the horizon within which (20) applies is short.
B. Online Motion Planning and Control
We now summarize everything into the PIPC algorithm.
Let ti ∈ S. We compensate for the local nature of (20) by recomputing a new Laplace approximation q(zt+ , ξt+ |ht , eS )
whenever t ∈ S, and applying filtering to update the policy
by (21) for t ∈ (ti , ti+1 ), in which the subscript t+ denotes
the future trajectory from t (see Fig. 2). This leads to an
Algorithm 1 Receding Horizon PIPC
Input: horizon th , start time t0 , initial belief q(ξt0 )
Output: success/failure
1: while not STOP CRITERIA do
2:
q̂(ξS |eS , hti −δt , uti −δt ) = getLaplaceApprox(ti , th , q(ξt |hti −δt , uti −δt ), ENVIRONMENT)
3:
for t ∈ [ti , ti+1 ] do
4:
zt = makeObservation()
5:
q̂(ξt |ht , eS ) = filterPolicy(zt , q̂(ξt−δt |ht−δt , eS ), q̂(ξS |eS , hti −δt , uti −δt ))
6:
executePolicy(ut = u∗t (ht ))
7:
q(ξt+δt |ht , ut ) = filterState(zt , ut , q(ξt |ht−δt , ut−δt ))
8:
end for
9: end while
10: return checkSuccess()
iterative framework which solves for the new approximation
with up-to-date knowledge about the system.
We can apply this scheme to MDP/POMDP problems in
both finite and receding horizon cases.4 When facing a dynamic environment, PIPC updates environmental information
in the new Laplace approximation in Section III-B.2.
The details of the receding horizon approach are summarized in Algorithm 1 and can be derived similarly for the
finite horizon case. First, at any time step ti , PIPC computes
the Laplace approximation for the current horizon window
[ti , ti + th ] with the latest information about the system and
the environment, where th ≥ ti+1 − ti is length of the
preview horizon. Second, for t ∈ (ti , ti+1 ), PIPC recursively
updates the policy using the most current observation with a
resolution of δt. These two steps repeat until the set criteria
are met or the execution fails (for example, the robot is in
collision).
V. I MPLEMENTATION D ETAILS
We perform experiments with the receding horizon version of PIPC in four different setups, including both MDP
and POMDP scenarios: MDP-CL and POMDP-CL execute
the receding horizon PIPC in Algorithm 1; MDP-OL and
POMDP-OL ignore the policy filtering step, but instead
recursively apply the open-loop policy given as the mode
found in the Laplace approximation. This open-loop baseline
can be viewed as the direct generalization of [9] to include
action trajectories.
The Laplace approximation is implemented using
GPMP25 and the GTSAM6 C++ library, which solves posterior maximization as a nonlinear least-squared optimization
defined on a factor graph with the Levenberg-Marquardt
algorithm. Note that in implementation we consider yt = ut
(i.e. ξt = (xt , ut )) and a constant time difference ∆t between
any two support states or actions.
We evaluate our algorithms on three different systems:
a 2D holonomic robot, a 7-DOF WAM, and a PR2 arm.
The state dynamics, following (15), is defined as a double
4 The receding-horizon version solves a new finite-horizon problem at
each iteration of the Laplace approximation.
5 Available at https://github.com/gtrll/gpmp2
6 Available at https://bitbucket.org/gtborg/gtsam
integrator with the state consisting of position and velocity
and
0 I
0
0
0
0
A = 0 0 ,B = I ,b = 0 ,FFT = 0 Q I
x
and following (4) we define the Gaussian process factor by
H = 0, D = 0, η = 0, GGT = Qu I
where 0 and I are d × d zero and identity matrices, where
d = 2 for the 2D holonomic robot and d = 7 for the 7DOF WAM arm and PR2 arm, and Qx and Qu are positive
scalars. The observation process in the POMDP is modeled
as a state observation with additive zero-mean Gaussian noise
with covariance Qv = σm I2d×2d . The state dynamics for
both the arms are assumed to be feedback linearized. On a
real system, the control would to be mapped back to real
torques using inverse dynamics.
VI. E VALUATION
We conduct benchmark experiments7 with our receding
horizon algorithm on the 2D holonomic robot in a dynamic
environment, and on the WAM arm and the PR2’s right
arm in a static environment (see Fig. 1). In each case, we
compare the closed-loop and open-loop algorithms for both
MDP and POMDP settings across different Qx (and number
of dynamic obstacles Nobs in the 2D case) with respect to
success rate, time to reach the goal, path length, and path
cost.8 Each setting is run for K times with a unique random
generator seed to account for stochasticity, which is kept
the same across all four algorithms for a fair comparison.
A trial is marked “successful” if the robot reaches the goal
within a Euclidean distance gdist, and is marked “failed” if
at any point the robot runs into collision or runs out of the
maximum allotted time tmax .
A. 2D Robot Benchmark
We simulate a 2D holonomic robot (radius = 0.5m) in a
2D environment (30m × 20m) with moving obstacles (see
Fig. 1 (a)). The robot’s sensor returns a limited view of a
5m × 5m square centered at the robot’s current position.
7A
video of experiments is available at https://youtu.be/8rQcg1O-6aU
cost is calculated as the negative log of the product of the factors.
8 Path
TABLE I: Success rate across increasing Qx and Nobs on the 2D holonomic robot.
10
Qx
CL
0.975
0.95
0.95
0.975
0.95
0.975
0.01
0.04
0.07
0.01
0.04
0.07
MDP
POMDP
20
OL
0.975
0.975
0.85
0.975
0.975
0.875
CL
0.85
0.85
0.875
0.925
0.875
0.825
Q x = 0.01
CL
0.7
0.55
0.725
0.725
0.525
0.7
40
OL
0.675
0.525
0.475
0.7
0.475
0.425
CL
0.4
0.525
0.45
0.375
0.425
0.45
20
10
10
10
10
20
30
40
50
0
10
20
30
40
0
50
60
10
10
10
5
5
5
0
0
40
20
60
0
Path Length (m)
0
60
40
0
×10 6
0
10
10
20
20
30
30
40
40
Number of obstacles
50
50
20
CL
0.25
0.325
0.225
0.15
0.4
0.2
OL
0.325
0.325
0.125
0.225
0.25
0.075
Q x = 0.07
20
0
50
OL
0.375
0.375
0.2
0.4
0.45
0.25
Q x = 0.04
20
Path Cost
Time (s)
30
OL
0.85
0.8
0.575
0.875
0.825
0.55
0
10
20
30
40
50
10
20
30
40
50
10
20
30
40
50
40
0
×10 6
0
10
10
20
20
30
30
40
40
20
50
0
50
0
×10 6
0
Number of obstacles
Number of obstacles
Fig. 3: Results of successful runs with increasing Qx and Nobs on the 2D holonomic robot.
The moving obstacles (squares of 1m × 1m) start at random locations and follow a 2D stochastic jump process,
where a noisy acceleration aobs is uniformly sampled within
[−2.5, 2.5]m/s2 at every time step. Their velocities vobs are
restricted within, [−1.3, 1.3]m/s. All obstacles are confined
inside the boundary during simulation.
Table I summarizes the success rates for this benchmark,9
and Fig. 3 shows the aggregate results of successful runs.
From Table I, we see that, for both MDP and POMDP cases,
the closed-loop algorithms have higher success rates than
the open-loop algorithms, especially in difficult problems
with larger stochasticity in the system (Qx ) or increased
complexity in the environment (Nobs ). Similar increasing
trends can also be observed in the difference of the success
rates between the closed-loop and open-loop algorithms.
The majority of failed open-loop cases arise from collision;
only a few are due to hitting the maximum run time. The
performance in POMDP cases are slightly worse than that
in the MDP cases on average. All three metrics (time, path
length, and path cost) in Fig. 3 increase in general with more
noise and obstacles. It is important that these plots should
be interpreted alongside the success rates, since the sample
size of successful trails is comparatively sparse for the harder
problems.
B. WAM and PR2 Benchmark
We demonstrate the scalability of PIPC to higher dimensional systems by performing a benchmark on the WAM and
9 Parameters
for this benchmark are set as follows: K = 40, gdist = 0.2,
tmax = 20, ∆t = 0.2, th = 2, nip = 20, σm = 0.01, σg = 1,
σf ix = 10−4 , Qu = 10, σobs = 0.02, = 1.
TABLE II: Success rate across increasing Qx on the WAM
and the PR2 robot arms.
Qx
MDP
POMDP
0.01
0.02
0.03
0.01
0.02
0.03
WAM
CL
OL
1
1
1
1
1
0.85
1
1
1
0.9
0.9
0.75
CL
1
1
1
1
1
1
PR2
OL
1
0.95
0.5
1
0.8
0.8
the PR2 robot arms. Here the WAM and the PR2 robot arms
are set up in lab and industrial environments [6], [7], [9]
respectively, in OpenRAVE. Here the task is to drive the
robot arm from a given start to a goal configuration (see
Fig. 1 (b) and (c)). The environments are static and fully
observable at all times. We compare the algorithms with
respect to increasing Qx . Table II summarizes the success
rates for this benchmark,10 and Fig. 4 shows the aggregate
results of successful runs. Similar to the 2D robot benchmark,
the results show that the closed-loop algorithms have higher
success rate than the open-loop ones, and all three metrics
increase with noise. In particular, POMDP-CL performs even
better than MDP-OL.
VII. C ONCLUSION
We consider the problem of motion planning and control
as probabilistic inference, and we propose an algorithm PIPC
10 Parameters for this benchmark are set as follows: K = 20, gdist =
0.065, tmax = 15, ∆t = 0.1, th = 2, nip = 10, σm = 0.005, σg =
0.03, σf ix = 10−4 , Qu = 10, = 0.1, σobs = 0.008 (WAM), σobs =
0.005 (PR2).
3.5
3.5
3.5
10
5
×10
6
3
10
×10
Time (s)
4
3
6
10
5
×10
6
5
(a)
Path Length
4
Q = 0.02
x
Q = 0.03
x
x
15
15
15
10
10
10
5
5
5
8
8
8
7
7
7
6
6
5
×10
Path Cost
4
Path Cost
6
4
2
3
Q = 0.01
Q x = 0.03
Time (s)
Q x = 0.02
6
4
2
Path Length
Q x = 0.01
6
4
2
6
6
5
×10
5
6
15
15
15
10
10
10
5
5
5
×10 6
(b)
Fig. 4: Results of successful runs with increasing Qx on (a) the WAM and (b) the PR2 robot arms.
for solving this problem that can exploit intrinsic sparsity
in continuous-time stochastic systems. In particular, PIPC
can address performance indices given by arbitrary, higherorder nonlinear factors and a general exponential-integralquadratic factor. Despite PIPC solving a continuous-time
problem, its complexity scales only linearly in the number of
nonlinear factors, thus making online simultaneous planning
and control possible in receding/finite horizon MDP/POMDP
problems.
ACKNOWLEDGMENTS
The authors would like to thank Jing Dong for help with
the GTSAM interface. This material is based upon work
supported by NSF CRII Award No. 1464219 and NSF NRI
Award No. 1637758.
R EFERENCES
[1] D. P. Bertsekas, Dynamic programming and optimal control. Athena
Scientific Belmont, MA, 1995, vol. 1, no. 2.
[2] R. C. Arkin, Behavior-based robotics. MIT press, 1998.
[3] L. E. Kavraki, P. Svestka, J.-C. Latombe, and M. H. Overmars, “Probabilistic roadmaps for path planning in high-dimensional configuration
spaces,” Robotics and Automation, IEEE Transactions on, vol. 12,
no. 4, pp. 566–580, 1996.
[4] J. J. Kuffner and S. M. LaValle, “RRT-connect: An efficient approach
to single-query path planning,” in Robotics and Automation, 2000.
Proceedings. ICRA’00. IEEE International Conference on, vol. 2.
IEEE, 2000, pp. 995–1001.
[5] A. Byravan, B. Boots, S. S. Srinivasa, and D. Fox, “Space-time
functional gradient optimization for motion planning,” in Robotics and
Automation (ICRA), 2014 IEEE International Conference on. IEEE,
2014, pp. 6499–6506.
[6] J. Schulman, Y. Duan, J. Ho, A. Lee, I. Awwal, H. Bradlow, J. Pan,
S. Patil, K. Goldberg, and P. Abbeel, “Motion planning with sequential
convex optimization and convex collision checking,” The International
Journal of Robotics Research, vol. 33, no. 9, pp. 1251–1270, 2014.
[7] M. Mukadam, X. Yan, and B. Boots, “Gaussian process motion
planning,” in 2016 IEEE International Conference on Robotics and
Automation (ICRA), May 2016, pp. 9–15.
[8] Z. Marinho, B. Boots, A. Dragan, A. Byravan, G. J. Gordon, and
S. Srinivasa, “Functional gradient motion planning in reproducing kernel hilbert spaces,” in Proceedings of Robotics: Science and Systems
(RSS-2016), 2016.
[9] J. Dong, M. Mukadam, F. Dellaert, and B. Boots, “Motion planning
as probabilistic inference using Gaussian processes and factor graphs,”
in Proceedings of Robotics: Science and Systems (RSS-2016), 2016.
[10] M. Toussaint, “Newton methods for k-order Markov constrained
motion problems,” arXiv preprint arXiv:1407.0414, 2014.
[11] S. M. LaValle and J. J. Kuffner, “Randomized kinodynamic planning,”
The International Journal of Robotics Research, vol. 20, no. 5, pp.
378–400, 2001.
[12] R. Tedrake, I. R. Manchester, M. Tobenkin, and J. W. Roberts, “LQRtrees: Feedback motion planning via sums-of-squares verification,” The
International Journal of Robotics Research, 2010.
[13] H. J. Kappen, “Linear theory for control of nonlinear stochastic
systems,” Physical review letters, vol. 95, no. 20, p. 200201, 2005.
[14] S. Levine and V. Koltun, “Guided policy search.” in ICML (3), 2013,
pp. 1–9.
[15] M. Deisenroth and C. E. Rasmussen, “PILCO: A model-based and
data-efficient approach to policy search,” in Proceedings of the 28th
International Conference on machine learning (ICML-11), 2011, pp.
465–472.
[16] D. Mayne, “A second-order gradient method for determining optimal
trajectories of non-linear discrete-time systems,” International Journal
of Control, vol. 3, no. 1, pp. 85–95, 1966.
[17] E. Todorov and W. Li, “A generalized iterative LQG method for
locally-optimal feedback control of constrained nonlinear stochastic
systems,” in American Control Conference, 2005. Proceedings of the
2005. IEEE, 2005, pp. 300–306.
[18] E. Todorov and Y. Tassa, “Iterative local dynamic programming,”
in 2009 IEEE Symposium on Adaptive Dynamic Programming and
Reinforcement Learning. IEEE, 2009, pp. 90–95.
[19] H. Attias, “Planning by probabilistic inference,” in AISTATS, 2003.
[20] M. Toussaint, “Robot trajectory optimization using approximate inference,” in Proceedings of the 26th annual international conference on
machine learning. ACM, 2009, pp. 1049–1056.
[21] M. Toussaint and A. Storkey, “Probabilistic inference for solving discrete and continuous state Markov decision processes,” in Proceedings
of the 23rd international conference on Machine learning. ACM,
2006, pp. 945–952.
[22] S. Levine and V. Koltun, “Variational policy search via trajectory
optimization,” in Advances in Neural Information Processing Systems,
2013, pp. 207–215.
[23] H. J. Kappen, V. Gómez, and M. Opper, “Optimal control
as a graphical model inference problem,” Machine Learning,
vol. 87, no. 2, pp. 159–182, 2012. [Online]. Available:
http://dx.doi.org/10.1007/s10994-012-5278-7
[24] K. Rawlik, M. Toussaint, and S. Vijayakumar, “On stochastic optimal
control and reinforcement learning by approximate inference,” Proceedings of Robotics: Science and Systems VIII, 2012.
[25] P. Kumar and J. Van Schuppen, “On the optimal control of stochastic
systems with an exponential-of-integral performance index,” Journal
of mathematical analysis and applications, vol. 80, no. 2, pp. 312–332,
1981.
[26] C. E. Rasmussen, Gaussian processes for machine learning, 2006.
[27] S. Sarkka, A. Solin, and J. Hartikainen, “Spatiotemporal learning
via infinite-dimensional Bayesian filtering and smoothing: A look at
Gaussian process regression through Kalman filtering,” IEEE Signal
Processing Magazine, vol. 30, no. 4, pp. 51–61, 2013.
[28] C. M. Bishop, “Pattern recognition,” Machine Learning, vol. 128, pp.
1–58, 2006.
[29] A. Boularias, “A predictive model for imitation learning in partially
observable environments,” in Machine Learning and Applications,
2008. ICMLA’08. Seventh International Conference on. IEEE, 2008,
pp. 83–90.
[30] T. Barfoot, C. H. Tong, and S. Sarkka, “Batch continuous-time
trajectory estimation as exactly sparse Gaussian process regression,”
Proceedings of Robotics: Science and Systems, Berkeley, USA, 2014.
[31] X. Yan, V. Indelman, and B. Boots, “Incremental sparse GP regression
for continuous-time trajectory estimation & mapping,” in Proceedings
of the International Symposium on Robotics Research (ISRR-2015),
2015.
A PPENDIX
A. Laplace Approximation with Factor Graphs
PIPC updates the Laplace approximation whenever t =
ti by efficiently solving a nonlinear least-squares problem
defined on a bipartite factor graph G = {ξS , fS , E},
Y
q(ξS |eS ) ∝
fS (ξS ).
(23)
Obstacle Factor:
fiobs = exp{− 12 eTi Q−1
obs ei },
ei = h(ξi )
Interpoaltion Factor:
fiintp = exp{− 12 eTi Q−1
obs ei },
ei = h(GPinterpolate(ξi , ξi+1 ))
GP Prior Factor:
figp = exp{− 12 eTi Q−1
gp,i ei },
ei = Φξ (ti+1 , ti )ξi − ξi+1
S∈S
where recall that ξS is the set of support augmented states,
and fS = {fS }S denotes the set of factors, and E are edges
connected to between ξS and fS .
An example factor graph is shown in Fig. 5 for a trajectory
starting from ti with a length equal to th . The sparse set
of support augmented states ξS are uniformly ∆t apart and
are connected to their neighbours via the Gaussian process
factors, forming a Gauss-Markov chain. Note that in our
implementation ξt = (xt , ut ).
1) Details of Factor Implementation:
a) Prior Factor: For each Laplace approximation, a
prior factor is placed on the first hidden state ξt , reflecting its
current belief given past history ht . In the MDP setting, the
covariance for state xt is set as a diagonal matrix σf2 ix I2d×2d ,
in which σf ix is a small number to indicate high confidence;
for control ut , we use the original Gaussian process factor
given by (3). Together they define Qprior . In the POMDP
setting, the belief of the hidden augmented state is obtained
via Kalman filtering, and we heuristically set the covariance
for the state, x to σf2 ix I2d×2d as mentioned previously.
b) Gaussian Process Factors: Analogous to defining
GPu for (4), we can define GPξ , which in turns define.
q(ξS ) in (17). In Fig. 5, this corresponds to Gaussian process
factors with
T
Z ti+1
F F
Qgp,i =
Φξ (ti+1 , s)
Φξ (ti+1 , s)T ds,
G
G
ti
B
[A
0 H
where Φξ is the state transition matrix associated with
]
that takes the system from ti to ti+1 .
c) Obstacle and Interpolation Factors: For obstacle
avoidance, we use a hinge loss function h with safety distance to compute a signed distance field as in GPMP2 [9].
In effect, it defines the obstacle factors and interpolation
2
factors in Fig. 5, which both use Qobs = σobs
I. Though
abstracted as a single factor in Fig. 5, between any two
support points ti and ti+1 , multiple (nip ) interpolated factors
can be constructed with indexes evenly spaced in time (δt
apart) to ensure path safety. See [9] for details.
d) Goal Factor: To drive the system to a desired goal
configuration ξgoal (for example, a particular position in
configuration space with zero velocity and action), we add
a goal factor to every support point except the current state.
This encourages the optimizer to drive all states in the current
horizon window closer to the goal. Because the covariance
of this factor acts as a weighting term, we define it as
||ξti −ξgoal ||2
Qgoal,i = σg2 ||ξstart
−ξgoal ||2 I such that it monotonically
decreases with the Euclidean distance to the goal.
Prior Factor:
f m = exp{− 12 ξi Q−1
prior ξi },
Goal Factor:
figoal = exp{− 12 eTi Q−1
goal,i ei },
ei = ξi − ξgoal
Fig. 5: A factor graph of an example Laplace approximation problem showing states (white circle) (ξi is used
as a shorthand for ξti ) and different kinds of factors: GP
Prior (black circle), obstacle and interpolation (white square),
measurement (gray square) and goal (black square).
2) Update of Laplace Approximation: The same Laplace
approximation is used to recursively update the policy for
t ∈ [ti , ti+1 ) with a resolution of δt, and, when t = ti+1 ,
the graph is updated to construct a new nonlinear leastsquare optimization problem. This is done by shifting the
horizon window ahead by ∆t and update the factors to
include any environmental changes during [ti , ti+1 ). In the
updated graph, the prior factor on the first state is given by
an additional Kalman filter based on (15) and (4) with (x, y)
as hidden states and (z, u) as observations. For POMDP
problems, we treat the estimation of current state as perfect
knowledge without uncertainty. This extra heuristic step is
a compromise which makes the assumption accurate at the
mean of the current belief.
B. Proof of (20)
Here we prove that the solution (20) to the approximate
optimization problem (19)
Z
max q̂(z, ξ|eS )π(u|h)dξdz
π
can be written as posterior inference: ∀t,
πt∗ (ut |ht ) = δ(ut − u∗t (ht ))
Z
∗
ut (ht ) = argmax q̂(z, ξ|eS )π ∗ (ūt |h)dxdydzdūt
ut
= argmax q̂(ut |ht , eS )
ut
where q̂(z, ξ|eS ) Q ∝
q(ξ)p(z|x)p̂(eS |xS , uS ) and
p̂(eS |xS , uS ) =
p̂(e
S |xS , uS ) is found by the
S∈S
exponential-quadratic approximate factor given by the
Laplace approximation.
Proof: We assume the length of the trajectory is T .
In the following, we first show that the optimal policy is
deterministic and then show that it corresponds to the mode
of the posterior distribution q̂(ut |ht , eS ).
a) The Optimal Policy is Deterministic: For any t, we
can write the objective function (19) as
Z
q̂(z, ξ|eS )π(u|h)dξdz
Z
∝ q(ξ)p(z|x)p̂(eS |xS , uS )π(u|h)dξdz
Z
= πt (ut |ht )fht (ut )dut dht
in which
Z
fht (ut ) =
q(ξ)p(z|x)p̂(eS |xS , uS )π(ūt |h̄t )dθdzt+ dut+
and
π(ūt |h̄t ) :=
π(u|h)
πt (ut |ht )
Therefore, equivalently, (19) can be formulated explicitly as
a variational problem of density function πt :
Z
(24)
max πt (ut |ht )fht (ut )dut
πt
s.t.
Z
π(ut |ht )dut = 1
π(ut |ht ) ≥ 0,
∀ut
To deal with the inequality, let gt2 (ut ) = πt (ut |ht ), and we
can further write (24) as
Z
max gt2 (ut )fht (ut )dut
(25)
gt
s.t.
Z
gt2 (ut )dut = 1
Let λ ∈ R be a Lagrangian multiplier. We can turn the (25)
into an unconstrained optimization and use calculus of variations to derive the solution:
min max L(gt , λ)
gt
λ
Z
Z
2
= min max gt (ut )fht (ut )dut + λ( gt2 (ut )dut − 1)
λ
gt
Suppose gt∗ (λ) is the optimum. Let gt = gt∗ + η, where
η is an arbitrary continuous function and → 0. Then the
optimality condition is given by
Z
∂L(gt , λ)
= 2gt (ut )η(ut )(λ + fht (ut ))dut = 0.
∂
Since η is arbitrary, it implies ∀ut ,
gt (ut )(λ + fht (ut )) = 0
Given that λ is a scalar and g(un ) is non-zero, we can
conclude that πt∗ (ut |ht ) = δ(ut − u∗t (ht )) satisfying
u∗t (ht ) = arg max fht (ut )
ut (ht )
b) The Optimal Policy is the Mode of Posterior: From
the previous proof, we know that the policy corresponds to
the mode of fht (ut ) for any t. Therefore, to complete the
proof, we only need to show that fh∗t (ut ) ∝ q̂(ut |eS , ht ),
where fh∗t (ut ) is fht (ut ) when the policies are optimal.
First, let fˆht (ut ) denote fht (ut ) when all policies are
deterministic, and define, for all t,
fˆht (ut ) := πt− (ut− |ht− )q̂(zt , zt− , eS , ut− , ut )
(26)
Next, we introduce a lemma:
Lemma 1: Let z = (x, y) ∈ Rn . If f (x, y) ∝
N (z|m, S), then, for all y,
Z
max f (x, y) = C f (x, y)dx
(27)
x(y)
for some constant C independent of y, in which m and S
are the mean and covariance of a Gaussian.
Now we can show fh∗t (ut ) ∝ q̂(ut |eS , ht ) by induction
in backward order. To start with, for the last policy at τ =
T − δt, we can write (26) as
fhτ (uτ ) = fˆhτ (uτ ) = πτ− (uτ− |hτ− )q̂(zτ , zτ− , eS , uτ− , uτ )
= πτ− (uτ− |hτ− )q̂(hτ , eS , uτ )
∝ q̂(uτ |hτ , eS )
in which we purposefully omit the dependency of uτ− on
hτ− , because the exact value of uτ− is observed when
performing the optimization. Then we have u∗τ (hτ ) =
argmaxuτ fˆhτ (uτ ) = argmaxuτ q̂(uτ |hτ , eS ).
Now we propagate the objective function one step backward from τ to τ − δt. Given hτ −δt , the maximization at
τ − δt is given as
Z
max max πτ (uτ |hτ )fˆhτ (uτ )duτ dzτ duτ −δt
πτ −δt πτ
Z
= max
max fˆhτ (uτ )dzτ duτ −δt
πτ −δt
uτ (hτ )
Z
= max πτ− (uτ− |hτ− ) max q̂(hτ , eS , uτ )dzτ duτ −δt
πτ −δt
uτ (hτ )
Z
∝ max πτ− (uτ− |hτ− )q̂(hτ , eS , uτ )dzτ duτ duτ −δt
πτ −δt
Z
= max πτ− (uτ− |hτ− )q̂(hτ −δt , eS , uτ −δt )duτ −δt
πτ −δt
Z
= max πτ −δt (uτ −δt |hτ −δt )fˆhτ −δt (uτ −δt )duτ −δt
πτ −δt
The second equality is due to the policy is deterministic;
the third proportionality is given by (27); the last equality is given by the definition (26). Therefore, the backward iteration
maintains the policy optimization problem
R
maxπt πt (ut |ht )fˆht (ut )dut in the same algebraic form as
the last step for all t. Since fˆht (ut ) ∝ q̂(ut |ht , eS ), this
completes the proof.
| 3cs.SY
|
Scoring and Classifying with Gated
Auto-encoders
arXiv:1412.6610v5 [cs.LG] 15 Jun 2015
Daniel Jiwoong Im, and Graham W. Taylor
School of Engineering
University of Guelph
Guelph, On, Canada
{imj,gwtaylor}@uoguelph.ca
Abstract. Auto-encoders are perhaps the best-known non-probabilistic
methods for representation learning. They are conceptually simple and
easy to train. Recent theoretical work has shed light on their ability to
capture manifold structure, and drawn connections to density modeling. This has motivated researchers to seek ways of auto-encoder scoring, which has furthered their use in classification. Gated auto-encoders
(GAEs) are an interesting and flexible extension of auto-encoders which
can learn transformations among different images or pixel covariances
within images. However, they have been much less studied, theoretically
or empirically. In this work, we apply a dynamical systems view to GAEs,
deriving a scoring function, and drawing connections to Restricted Boltzmann Machines. On a set of deep learning benchmarks, we also demonstrate their effectiveness for single and multi-label classification.
1
Introduction
Representation learning algorithms are machine learning algorithms which involve the learning of features or explanatory factors. Deep learning techniques,
which employ several layers of representation learning, have achieved much recent success in machine learning benchmarks and competitions, however, most
of these successes have been achieved with purely supervised learning methods
and have relied on large amounts of labeled data [10,22]. Though progress has
been slower, it is likely that unsupervised learning will be important to future
advances in deep learning [1].
The most successful and well-known example of non-probabilistic unsupervised learning is the auto-encoder. Conceptually simple and easy to train via
backpropagation, various regularized variants of the model have recently been
proposed [20,25,21] as well as theoretical insights into their operation [6,24].
In practice, the latent representation learned by auto-encoders has typically
been used to solve a secondary problem, often classification. The most common
setup is to train a single auto-encoder on data from all classes and then a classifier
is tasked to discriminate among classes. However, this contrasts with the way
probabilistic models have typically been used in the past: in that literature, it is
more common to train one model per class and use Bayes’ rule for classification.
2
There are two challenges to classifying using per-class auto-encoders. First, up
until very recently, it was not known how to obtain the score of data under
an auto-encoder, meaning how much the model “likes” an input. Second, autoencoders are non-probabilistic, so even if they can be scored, the scores do not
integrate to 1 and therefore the per-class models need to be calibrated.
Kamyshanska and Memisevic have recently shown how scores can be computed from an auto-encoder by interpreting it as a dynamical system [7]. Although the scores do not integrate to 1, they show how one can combine the
unnormalized scores into a generative classifier by learning class-specific normalizing constants from labeled data.
In this paper we turn our interest towards a variant of auto-encoders which
are capable of learning higher-order features from data [15]. The main idea is to
learn relations between pixel intensities rather than the pixel intensities themselves by structuring the model as a tri-partite graph which connects hidden
units to pairs of images. If the images are different, the hidden units learn how
the images transform. If the images are the same, the hidden units encode withinimage pixel covariances. Learning such higher-order features can yield improved
results on recognition and generative tasks.
We adopt a dynamical systems view of gated auto-encoders, demonstrating
that they can be scored similarly to the classical auto-encoder. We adopt the
framework of [7] both conceptually and formally in developing a theory which
yields insights into the operation of gated auto-encoders. In addition to the
theory, we show in our experiments that a classification model based on gated
auto-encoder scoring can outperform a number of other representation learning
architectures, including classical auto-encoder scoring. We also demonstrate that
scoring can be useful for the structured output task of multi-label classification.
2
Gated Auto-encoders
In this section, we review the gated auto-encoder (GAE). Due to space constraints, we will not review the classical auto-encoder. Instead, we direct the
reader to the reviews in [15,8] with which we share notation. Similar to the classical auto-encoder, the GAE consists of an encoder h(·) and decoder r(·). While
the standard auto-encoder processes a datapoint x, the GAE processes inputoutput pairs (x, y). The GAE is usually trained to reconstruct y given x, though
it can also be trained symmetrically, that is, to reconstruct both y from x and
x from y. Intuitively, the GAE learns relations between the inputs, rather than
representations of the inputs themselves1 . If x 6= y, for example, they represent
sequential frames of a video, intuitively, the mapping units h learn transformations. In the case that x = y (i.e. the input is copied), the mapping units learn
pixel covariances.
In the simplest form of the GAE, the M hidden (mapping) units are given
by a basis expansion of x and y. However, this leads to a parameterization
1
Relational features can be mixed with standard features by simply adding connections that are not gated.
3
that it is at least quadratic in the number of inputs and thus, prohibitively
large. Therefore, in practice, x, y, and h are projected onto matrices or (“latent
factors”), W X , W Y , and W H , respectively. The number of factors, F , must be
the same for X, Y , and H. Thus, the model is completely parameterized by
θ = {W X , W Y , W H } such that W X and W Y are F × D matrices (assuming
both x and y are D-dimensional) and W H is an M × F matrix. The encoder
function is defined by
h(x, y) = σ(W H ((W X x)
(W Y y)))
(1)
where
is element-wise multiplication and σ(·) is an activation function. The
decoder function is defined by
r(y|x, h) = (W Y )T ((W X x)
X T
Y
r(x|y, h) = (W ) ((W y)
(W H )T h(x, y)).
H T
(W ) h(x, y)),
(2)
(3)
Note that the parameters are usually shared between the encoder and decoder.
The choice of whether to apply a nonlinearity to the output, and the specific
form of objective function will depend on the nature of the inputs, for example,
binary, categorical, or real-valued. Here, we have assumed real-valued inputs for
simplicity of presentation, therefore, Eqs. 2 and 3 are bi-linear functions of h
and we use a squared-error objective:
J=
1
kr(y|x) − yk2 .
2
(4)
We can also constrain the GAE to be a symmetric model by training it to
reconstruct both x given y and y given x [15]:
J=
1
1
kr(y|x) − yk2 + kr(x|y) − xk2 .
2
2
(5)
The symmetric objective can be thought of as the non-probabilistic analogue
of modeling a joint distribution over x and y as opposed to a conditional [15].
3
Gated Auto-Encoder Scoring
In [7], the authors showed that data could be scored under an auto-encoder
by interpreting the model as a dynamical system. In contrast to the probabilistic views based on score matching [21,24,6] and regularization, the dynamical
systems approach permits scoring under models with either linear (real-valued
data) or sigmoid (binary data) outputs, as well as arbitrary hidden unit activation functions. The method is also agnostic to the learning procedure used to
train the model, meaning that it is suitable for the various types of regularized
auto-encoders which have been proposed recently. In this section, we demonstrate how the dynamical systems view can be extended to the GAE.
4
3.1
Vector field representation
Similar to [7], we will view the GAE as a dynamical system with the vector field
defined by
F (y|x) = r(y|x) − y.
The vector field represents the local transformation that y|x undergoes as a
result of applying the reconstruction function r(y|x). Repeatedly applying the
reconstruction function to an input y|x → r(y|x) → r(r(y|x)|x) → · · · →
r(r · · · r(y|x)|x) yields a trajectory whose dynamics, from a physics perspective,
can be viewed as a force field. At any point, the potential force acting on a point
is the gradient of some potential energy (negative goodness) at that point. In
this light, the GAE reconstruction may be viewed as pushing pairs of inputs x, y
in the direction of lower energy.
Our goal is to derive the energy function, which we call a scoring function,
and which measures how much a GAE “likes” a given pair of inputs (x, y) up to
normalizing constant. In order to find an expression for the potential energy, the
vector field must be able to be written as the derivative of a scalar field [7]. To
check this, we can submit to Poincaré’s integrability criterion: For some open,
simple connected set U, a continuously differentiable function F : U → <m
defines a gradient field if and only if
∂Fj (y)
∂Fi (y)
=
, ∀i, j = 1 · · · n.
∂yj
∂yi
The vector field defined by the GAE indeed satisfies Poincaré’s integrability criterion; therefore it can be written as the derivative of a scalar field. A derivation
is given in the Appendix A.2 This also applies to the GAE with a symmetric
objective function (Eq. 5) by setting the input as ξ|γ such that ξ = [y; x] and
γ = [x; y] and following the exact same procedure.
3.2
Scoring the GAE
As mentioned in Section 3.1, our goal is to find an energy surface, so that we
can express the energy for a specific pair (x, y). From the previous section, we
showed that Poincaré’s criterion is satisfied and this implies that we can write
the vector field as the derivative of a scalar field. Moreover, it illustrates that
this vector field is a conservative field and this means that the vector field is a
gradient of some scalar function, which in this case is the energy function of a
GAE:
r(y|x) − y = ∇E.
Hence, by integrating out the trajectory of the GAE (x, y), we can measure the
energy along a path. Moreover, the line integral of a conservative vector field
5
is path independent, which allows us to take the anti-derivative of the scalar
function:
Z
Z
Z
E(y|x) = (r(y|x) − y)dy = W Y W X x) W H h(u dy − ydy
Z
Z
(6)
=W Y
W Xx
W H h (u) dy − ydy,
where u is an auxiliary variable such that u = W H ((W Y y) (W X x)) and
du
H
Y
(W X x ⊗ 1D )), and ⊗ is the Kronecker product. Moreover, the
dy = W (W
decoder can be re-formulated as
r(y|x) = (W Y )T (W X x
= (W Y )T
(W H )T h(y, x))
(W X x ⊗ 1D ) (W H )T h(y, x).
Re-writing Eq. 6 in terms of the auxiliary variable u, we get
E(y|x) = (W Y )T (W Y x ⊗ 1D ) (W H )T
Z
Z
−1
H
Y
X
h(u) W
W
(W x ⊗ 1D )
du − ydy
Z
1
= h(u)du − y2 + const.
2
(7)
(8)
A more detailed derivation from Eq. 6 to Eq. 8 is provided in the Appendix A.2.
Identical to [7], if h(u) is an element-wise activation function and we know its
anti-derivative, then it is very simple to compute E(x, y).
4
Relationship to Restricted Boltzmann Machines
In this section, we relate GAEs through the scoring function to other types of
Restricted Boltzmann Machines, such as the Factored Gated Conditional RBM
[23] and the Mean-covariance RBM [19].
4.1
Gated Auto-encoder and Factored Gated Conditional
Restricted Boltzmann Machines
Kamyshanska and Memisevic showed that several hidden activation functions
defined gradient fields, including sigmoid, softmax, tanh, linear, rectified linear
function (ReLU), modulus, and squaring. These activation functions are applicable to GAEs as well.
In the case of the sigmoid activation function, σ = h(u) = 1+exp1 (−u) , our
energy function becomes
Z
1
Eσ =2 (1 + exp −(u))−1 du − (x2 + y2 ) + const,
2
X
1
H
=2
log (1 + exp (Wk·
(W X x W X y))) − (x2 + y2 ) + const.
2
k
6
Note that if we consider the conditional GAE we reconstruct x given y only, this
yields
X
y2
Y
X
Eσ (y|x) =
log (1 + exp (W H (Wk·
y Wk·
x))) −
+ const.
(9)
2
k
This expression is identical, up to a constant, to the free energy in a Factored
Gated Conditional Restricted Boltzmann Machine (FCRBM) with Gaussian visible units and Bernoulli hidden units. We have ignored biases for simplicity. A
derivation including biases is provided in the Appendix B.1.
4.2
Mean-Covariance Auto-encoder and Mean-covariance
Restricted Boltzmann Machines
The Covariance auto-encoder (cAE) was introduced in [15]. It is a specific form of
symmetrically trained auto-encoder with identical inputs: x = y, and tied input
weights: W X = W Y . It maintains a set of relational mapping units to model
covariance between pixels. One can introduce a separate set of mapping units
connected pairwise to only one of the inputs which model the mean intensity. In
this case, the model becomes a Mean-covariance auto-encoder (mcAE).
Theorem 1. Consider a cAE with encoder and decoder:
h(x) = h(W H ((W X x)2 ) + b)
r(x|h) = (W X )T (W X x
(W H )T h(x)) + a,
where θ = {W X , W H , a, b} are the parameters of the model, and h(z) = 1+exp1 (−z)
is a sigmoid. Moreover, consider a Covariance RBM [19] with Gaussian-distributed
visibles and Bernoulli-distributed hiddens, with an energy function defined by
(a − x)2 X
E c (x, h) =
−
P h(Cx)2 − bh.
σ2
f
Then the energy function of the cAE with dynamics r(x|y) − x is equivalent to
the free energy of Covariance RBM up to a constant:
X
x2
E(x, x) =
log 1 + exp W H (W X x)2 + b −
+ const.
(10)
2
k
The proof is given in the Appendix B.2. We can extend this analysis to the
mcAE by using the above theorem and the results from [7].
Corollary 1. The energy function of a mcAE and the free energy of a Meancovariance RBM (mcRBM) with Gaussian-distributed visibles and Bernoullidistributed hiddens are equivalent up to a constant. The energy of the mcAE
is:
X
X
E=
log 1 + exp −W H (W X x)2 − b +
log (1 + exp(W x + c))−x2 +const
k
k
(11)
7
where θm = {W, c} parameterizes the mean mapping units and θc = {W X , W H ,
a, b} parameterizes the covariance mapping units.
Proof. The proof is very simple. Let Emc = Em + Ec , where Em is the energy
of the mean auto-encoder, Ec is the energy of the covariance auto-encoder, and
Emc is the energy of the mcAE. We know from Theorem 1 that Ec is equivalent
to the free energy of a covariance RBM, and the results from [7] show that that
Em is equivalent to the free energy of mean (classical) RBM. As shown in [19],
the free energy of a mcRBM is equal to summing the free energies of a mean
RBM and a covariance RBM.
5
Classification with Gated Auto-encoders
Kamyshanska and Memisevic demonstrated that one application of the ability
to assign energy or scores to auto-encoders was in constructing a classifier from
class-specific auto-encoders. In this section, we explore two different paradigms
for classification. Similar to that work, we consider the usual multi-class problem
by first training class-specific auto-encoders, and using their energy functions
as confidence scores. We also consider the more challenging structured output
problem, specifically, the case of multi-label prediction where a data point may
have more than one associated label, and there may be correlations among the
labels.
5.1
Classification using class-specific gated auto-encoders
One approach to classification is to take several class-specific models and assemble them into a classifier. The best-known example of this approach is to
fit several directed graphical models and use Bayes’ rule to combine them. The
process is simple because the models are normalized, or calibrated. While it is
possible to apply a similar technique to undirected or non-normalized models
such as auto-encoders, one must take care to calibrate them.
The approach proposed in [7] is to train K class-specific auto-encoders, each
of which assigns a non-normalized energy to the data Ei (x) , i = 1 . . . , K, and
then define the conditional distribution over classes zi as
exp (Ei (x) + Bi )
P (zi |x) = P
,
j exp (Ej (x) + Bj )
(12)
where Bi is a learned bias for class i. The bias terms take the role of calibrating
the unnormalized energies. Note that we can similarly combine the energies from
a symmetric gated auto-encoder where x = y (i.e. a covariance auto-encoder)
and apply Eq. 12. If, for each class, we train both a covariance auto-encoder and
a classical auto-encoder (i.e. a “mean” auto-encoder) then we can combine both
sets of unnormalized energies as follows
exp(EiM (x) + EiC (x) + Bi )
PmcAE (zi |x) = P
,
M
C
j exp(Ej (x) + Ej (x) + Bj )
(13)
8
where EiM (x) is the energy which comes from the “mean” (standard) autoencoder trained on class i and EiC (x) the energy which comes from the “covariance” (gated) auto-encoder trained on class i. We call the classifiers in Eq. 12
and Eq. 13 “Covariance Auto-encoder Scoring” (cAES) and “Mean-Covariance
Auto-encoder Scoring” (mcAES), respectively.
The training procedure is summarized as follows:
1. Train a (mean)-covariance auto-encoder individually for each class. Both the
mean and covariance auto-encoder have tied weights in the encoder and decoder. The covariance auto-encoder is a gated auto-encoder with tied inputs.
2. Learn the Bi calibration terms using maximum likelihood, and backpropagate to the GAE parameters.
Experimental results We followed the same experimental setup as [16] where
we used a standard set of “Deep Learning Benchmarks” [11]. We used mini-batch
stochastic gradient descent to optimize parameters during training. The hyperparameters: number of hiddens, number of factors, corruption level, learning
rate, weight-decay, momentum rate, and batch sizes were chosen based on a
held-out validation set. Corruption levels and weight-decay were selected from
{0, 0.1, 0.2, 0.3, 0.4, 0.5}, and number of hidden and factors were selected from
100,300,500. We selected the learning rate and weight-decay from the range
(0.001, 0.0001).
Classification error results are shown in Table 1. First, the error rates of
auto-encoder scoring variant methods illustrate that across all datasets AES
outperforms cAES and mcAES outperforms both AES and cAES. AE models
pixel means and cAE models pixel covariance, while mcAE models both mean
and covariance, making it naturally more expressive. We observe that cAES and
mcAES achieve lower error rates by a large margin on rotated MNIST with
backgrounds (final row). On the other hand, both cAES and mcAES perform
poorly on MNIST with random white noise background (second row from bottom). We believe this phenomenon is due to the inability to model covariance in
this dataset. In MNIST with random white noise the pixels are typically uncorrelated, where in rotated MNIST with backgrounds the correlations are present
and consistent.
5.2
Multi-label classification via optimization in label space
The dominant application of deep learning approaches to vision has been the
assignment of images to discrete classes (e.g. object recognition). Many applications, however, involve “structured outputs” where the output variable is highdimensional and has a complex, multi-modal joint distribution. Structured output prediction may include tasks such as multi-label classification where there
are regularities to be learned in the output, and segmentation, where the output
is as high-dimensional as the input. A key challenge to such approaches lies in
developing models that are able to capture complex, high level structure like
shape, while still remaining tractable.
9
DATA
SVM RBM DEEP GSM AES cAES mcAES
RECT
RECTIMG
CONVEX
MNISTSMALL
MNISTROT
MNISTRAND
MNISTROTIM
2.15
24.04
19.13
3.03
11.11
14.58
55.18
RBF
SAA3
4.71
23.69
19.92
3.94
14.69
9.80
52.21
2.14
24.05
18.41
3.46
10.30
11.28
51.93
0.56
22.51
17.08
3.70
11.75
10.48
55.16
0.84
21.45
21.52
2.61
11.25
9.70
47.14
0.61
22.85
21.6
3.65
16.5
18.65
39.98
0.54
21.41
20.63
3.65
13.42
16.73
35.52
Table 1: Classification error rates on the Deep Learning Benchmark dataset.
SAA3 stands for three-layer Stacked Auto-encoder. SVM and RBM results are
from [24], DEEP and GSM are results from [15], and AES is from [7].
Though our proposed work is based on a deterministic model, we have shown
that the energy, or scoring function of the GAE is equivalent, up to a constant, to
that of a conditional RBM, a model that has already seen some use in structured
prediction problems [18,12].
GAE scoring can be applied to structured output problems as a type of
“post-classification” [17]. The idea is to let a naiv̈e, non-structured classifier
make an initial prediction of the outputs in a fast, feed-forward manner, and
then allow a second model (in our case, a GAE) clean up the outputs of the first
model. Since GAEs can model the relationship between input x and structured
output y, we can initialize the output with the output of the naiv̈e model, and
then optimize its energy function with respect to the outputs. Input x is held
constant throughout the optimization.
Li et al recently proposed Compositional High Order Pattern Potentials, a hybrid of Conditional Random Fields (CRF) and Restricted Boltzmann Machines.
The RBM provides a global shape information prior to the locally-connected
CRF. Adopting the idea of learning structured relationships between outputs,
we propose an alternate approach which the inputs of the GAE are not (x, y)
but (y, y). In other words, the post-classification model is a covariance autoencoder. The intuition behind the first approach is to use a GAE to learn the
relationship between the input x and the output y, whereas the second method
aims to learn the correlations between the outputs y.
We denote our two proposed methods GAEXY and GAEY 2 . GAEXY corresponds to a GAE, trained conditionally, whose mapping units directly model the
relationship between input and output and GAEY 2 corresponds to a GAE which
models correlations between output dimensions. GAEXY defines E (y|x), while
GAEY 2 defines E (y|y) = E(y). They differ only in terms of the data vectors
that they consume. The training and test procedures are detailed in Algorithm
1.
Experimental results We consider multi-label classification, where the problem is to classify instances which can take on more than one label at a time. We
10
Algorithm 1 Structured Output Prediction with GAE scoring
1: procedure Multi-label Classification(D = {(xi , yi ) ∈ Xtrain × Ytrain } )
2:
Train a Multi-layer Perceptron (MLP) to learn an input-output mapping f (·):
X
argmin l(x, y; θ1 ) =
loss1 ((f (xi ; θ1 ) − yi )
(14)
θ1
3:
i
where loss1 is an appropriate loss function for the MLP.2
Train a Gated Auto-encoder with inputs (xi , yi ); For the case of GAEY 2 , set
xi = yi .
X
argmin l(x, y; θ2 ) =
loss2 (r(yi |xi , θ2 ) − yi )
(15)
θ2
4:
5:
i
where loss2 is an appropriate reconstructive loss for the auto-encoder.
for each test data point xi ∈ Xtest do
Initialize the output using the MLP.
y0 = f (xtest )
6:
7:
8:
9:
(16)
while kE(yt+1 |x) − E(yt |x)k > or ≤ max. iter. do
Compute Oyt E
Update yt+1 = yt − λOyt E
where is the tolerance rate with respect to the convergence of the
optimization.
followed the same experimental set up as [18]. Four multi-labeled datasets were
considered: Yeast [5] consists of biological attributes, Scene [2] is image-based,
and MTurk [13] and MajMin [14] are targeted towards tagging music. Yeast consists of 103 biological attributes and has 14 possible labels, Scene consists of 294
image pixels with 6 possible labels, and MTurk and MajMin each consist of 389
audio features extracted from music and have 92 and 96 possible tags, respectively. Figure 1 visualizes the covariance matrix for the label dimensions in each
dataset. We can see from this that there are correlations present in the labels
which suggests that a structured approach may improve on a non-structured
predictor.
Yeast
Scene
0.21
0.18
0.15
0.12
0.09
0.06
0.03
0.00
0.03
Mturk
0.21
0.18
0.15
0.12
0.09
0.06
0.03
0.00
0.03
Majmin
0.21
0.18
0.15
0.12
0.09
0.06
0.03
0.00
0.03
0.21
0.18
0.15
0.12
0.09
0.06
0.03
0.00
0.03
Fig. 1: Covariance matrices for the multi-label datasets: Yeast, Scene, MTurk,
and MajMin.
2
In our experiments, we used the cross-entropy loss function for loss1 and loss2 .
11
We compared our proposed approaches to logistic regression, a standard
MLP, and the two structured CRBM training algorithms presented in [18]. To
permit a fair comparison, we followed the same procedure for training and reporting errors as in that paper, where we cross validated over 10 folds and training,
validation, test examples are randomly separated into 80%, 10%, and 10% in
each fold. The error rate was measured by averaging the errors on each label
dimension.
Method
LogReg
HashCRBM∗
MLP
GAESXY
GAESY 2
Yeast
20.16
20.02
19.79
19.27
19.58
Scene
10.11
8.80
8.99
6.83
6.81
MTurk MajMin
8.10
4.34
7.24
4.24
7.13
4.23
6.59
3.96
6.59
4.29
Table 2: Error rate on multi-label datasets. As in previous work, we report the
mean across 10 repeated runs with different random weight initializations.
The performance on four multi-label datasets is shown in Table 2. We observed that adding a small amount of Gaussian noise to the input y improved
the performance for GAEXY . However, adding noise to the input x did not have
as much of an effect. We suspect that adding noise makes the GAE more robust
to the input provided by the MLP. Interestingly, we found that the performance
of GAEY 2 was negatively affected by adding noise. Both of our proposed methods, GAESXY and GAESY 2 generally outperformed the other methods except
for GAESY 2 on the MajMin dataset. At least for these datasets, there is no
clear winner between the two. GAESXY achieved lower error than GAESY 2 for
Yeast and MajMin, and the same error rate on the MTurk dataset. However,
GAESY 2 outperforms GAESXY on the Scene dataset. Overall, the results show
that GAE scoring may be a promising means of post-classification in structured
output prediction.
6
Conclusion
There have been many theoretical and empirical studies on auto-encoders [25,20,21,24,6,7],
however, the theoretical study of gated auto-encoders is limited apart from [15,4].
The GAE has several intriguing properties that a classical auto-encoder does not,
based on its ability to model relations among pixel intensities rather than just
the intensities themselves. This opens up a broader set of applications. In this
paper, we derive some theoretical results for the GAE that enable us to gain
more insight and understanding of its operation.
We cast the GAE as a dynamical system driven by a vector field in order to
analyze the model. In the first part of the paper, by following the same procedure
as [7], we showed that the GAE could be scored according to an energy function.
12
From this perspective, we demonstrated the equivalency of the GAE energy to
the free energy of a FCRBM with Gaussian visible units, Bernoulli hidden units,
and sigmoid hidden activations. In the same manner, we also showed that the
covariance auto-encoder can be formulated in a way such that its energy function
is the same as the free energy of a covariance RBM, and this naturally led to
a connection between the mean-covariance auto-encoder and mean-covariance
RBM. One interesting observation is that Gaussian-Bernoulli RBMs have been
reported to be difficult to train [9,3], and the success of training RBMs is highly
dependent on the training setup [26]. Auto-encoders are an attractive alternative,
even when an energy function is required.
Structured output prediction is a natural next step for representation learning. The main advantage of our approach compared to other popular approaches
such as Markov Random Fields, is that inference is extremely fast, using a
gradient-based optimization of the auto-encoder scoring function. In the future,
we plan on tackling more challenging structured output prediction problems.
References
1. Bengio, Y., Thibodeau-Laufer, É.: Deep generative stochastic networks trainable
by backprop. arXiv preprint arXiv:1306.1091 (2013)
2. Boutell, M.R., Luob, J., Shen, X., Brown, C.M.: Learning multi-label scene classification. Pattern Recognition 37, 1757–1771 (2004)
3. Cho, K., Ilin, A., Raiko, T.: Improved learning of Gaussian-Bernoulli restricted
Boltzmann machines. In: ICANN. pp. 10–17 (2011)
4. Droniou, A., Sigaud, O.: Gated autoencoders with tied input weights. In: ICML
(2013)
5. Elisseeff, A., Weston, J.: A kernel method for multi-labelled classification. In: NIPS
(2002)
6. Guillaume, A., Bengio, Y.: What regularized auto-encoders learn from the data
generating distribution. In: ICLR (2013)
7. Kamyshanska, H., Memisevic, R.: On autoencoder scoring. In: ICML. pp. 720–728
(2013)
8. Kamyshanska, H., Memisevic, R.: The potential energy of an auto-encoder. IEEE
Transactions on Pattern Analysis and Machine Intelligence 37(6), 1261–1273 (2014)
9. Krizhevsky, A.: Learning multiple layers of features from tiny images. Tech. rep.,
Department of Computer Science, University of Toronto (2009)
10. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: NIPS (2012)
11. Larochelle, H., Erhan, D., Courville, A., Bergstra, J., Bengio, Y.: An empirical
evaluation of deep architectures on problems with many factors of variation. In:
ICML (2007)
12. Li, Y., Tarlow, D., Zemel, R.: Exploring compositional high order pattern potentials
for structured output learning. In: CVPR (2013)
13. Mandel, M.I., Eck, D., Bengio, Y.: Learning tags that vary within a song. In:
ISMIR (2010)
14. Mandel, M.I., Ellis, D.P.W.: A web-based game for collecting music metadata.
Journal New of Music Research 37, 151–165 (2008)
13
15. Memisevic, R.: Gradient-based learning of higher-order image features. In: ICCV
(2011)
16. Memisevic, R., Zach, C., Hinton, G., Pollefeys, M.: Gated softmax classification.
In: NIPS (2010)
17. Mnih, V., Hinton, G.: Learning to detect roads in high-resolution aerial images.
In: Proceedings of the 11th European Conference on Computer Vision (ECCV)
(2010)
18. Mnih, V., Larochelle, H., Hinton, G.E.: Conditional restricted Boltzmann machines
for structured output prediction. In: UAI (2011)
19. Ranzato, M., Hinton, G.E.: Modeling pixel means and covariances using factorized
third-order Boltzmann machines. In: CVPR (2010)
20. Rifai, S.: Contractive auto-encoders: Explicit invariance during feature extraction.
In: ICML (2011)
21. Swersky, K., Ranzato, M., Buchman, D., Freitas, N.D., Marlin, B.M.: On autoencoders and score matching for energy based models. In: ICML. pp. 1201–1208
(2011)
22. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D.,
Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. arXiv preprint
arXiv:1409.4842 (2014)
23. Taylor, G.W., Hinton, G.E.: Factored conditional restricted Boltzmann machines
for modeling motion style. In: ICML. pp. 1025–1032 (2009)
24. Vincent, P.: A connection between score matching and denoising auto-encoders.
Neural Computation 23(7), 1661–1674 (2010)
25. Vincent, P., Larochelle, H., Bengio, Y., Manzagol, P.: Extracting and composing
robust features with denoising autoencoders. In: ICML (2008)
26. Wang, N., Melchior, J., Wiskott, L.: Gaussian-binary restricted Boltzmann machines on modeling natural image statistics. Tech. rep., Institut fur Neuroinformatik Ruhr-Universitat Bochum, Bochum, 44780, Germany (2014)
A
A.1
Gated Auto-encoder Scoring
Vector field representation
To check that the vector field can be written as the derivative of a scalar field, we
can submit to Poincaré’s integrability criterion: For some open, simple connected
set U, a continuously differentiable function F : U → <m defines a gradient field
if and only if
∂Fi (y)
∂Fj (y)
=
, ∀i, j = 1 · · · n.
∂yj
∂yi
Considering the GAE, note that ith component of the decoder ri (y|x) can be
rewritten as
ri (y|x) = (W·iY )T (W X x
(W H )T h(y, x)) = (W·iY
W X x)T (W H )T h(y, x).
14
The derivatives of ri (y|x) − yi with respect to yj are
∂ri (y|x)
∂h(x, y)
∂rj (y|x)
=(W·iY W X x)T (W H )T
=
∂yj
∂yj
∂yi
∂h(y, x) ∂h(u) H
=
W (W·jY W X x)
∂yj
∂u
where u = W H ((W Y y)
we have
(W X x)). By substituting Equation 17 into
(17)
∂Fi ∂Fj
∂yj , ∂yi ,
∂rj (y|x)
∂Fj
∂Fi ∂ri (y|x)
=
−δij =
−δij =
∂yj
∂yj
∂yi
∂yi
where δij = 1 for i = j and 0 for i 6= j. Similarly, the derivatives of ri (y|x) − yi
with respect to xj are
∂ri (y|x)
=(W·iY W·jX )T (W H )T h(x, y) + (W·iY
∂xj
∂h(y, x) ∂h(u) H
=
W (W·jY W X x).
∂xj
∂u
By substituting Equation 18 into
∂Fi ∂Fj
∂xj , ∂xi ,
W X x)(W H )T
∂h
∂rj (y|x)
=
,
∂xj
∂xi
(18)
this yields
∂Fi ∂ri (x|y) ∂rj (x|y) ∂Fj
=
=
=
.
∂xj
∂xj
∂xi
∂xi
A.2
Deriving an Energy Function
Integrating out the GAE’s trajectory, we have
Z
E(y|x) = (r(y|x) − y)dy
ZC
Z
= W Y W X x) W H h(u dy − ydy
Z
Z
Y
X
H
=W
W x
W
h (u) du − ydy,
(19)
where u is an auxiliary variable such that u = W H ((W Y y) (W X x)) and
du
H
Y
(W X x ⊗ 1D )), where ⊗ is the Kronecker product. Consider
dy = W (W
the symmetric objective function, which is defined in Equation 5. Then we have
to also consider the vector field system where both symmetric cases x|y and y|x
are valid. As mentioned in Section 3.1, let ξ = [x; y] and γ = [y; x]. As well, let
W ξ = diag(W X , W Y ) and W γ = diag(W Y , W X ) where they are block diagonal
matrices. Consequently, the vector field becomes
F (ξ|γ) = r(ξ|γ) − ξ,
(20)
15
and the energy function becomes
Z
E(ξ|γ) = (r(ξ|γ) − ξ)dξ
Z
= (W ξ )T ((W γ γ)
=(W ξ )T ((W γ γ)
Z
(W H )T h(u))dξ − ξdξ
Z
Z
H T
h(u)du) − ξdξ
(W )
where u is an auxiliary variable such that u = W H (W ξ ξ)
du
= WH Wξ
dξ
(W γ γ) . Then
(W γ γ ⊗ 1D ) .
Moreover, note that the decoder can be re-formulated as
r(ξ|γ) = (W ξ )T (W γ γ
= (W ξ )T
(W H )T h(ξ, γ))
(W γ γ ⊗ 1D ) (W H )T h(ξ, γ).
Re-writing the first term of Equation 19 in terms of the auxiliary variable u, the
energy reduces to
Z
Z
−1
E(ξ|γ) = (W ξ )T (W γ γ ⊗ 1D ) (W H )T h(u) W H (W ξ (W γ γ ⊗ 1D ))
du − ξdξ
Z
Z
−T
= (W ξ )T (W γ γ ⊗ 1D ) (W H )T (W ξ (W γ γ ⊗ 1D ))W H
h(u)du − ξdξ
Z
Z
= h(u)du − ξdξ
Z
1
= h(u)du − ξ 2 + const.
2
B
B.1
Relation to other types of Restricted Boltzmann
Machines
Gated Auto-encoder and Factored Gated Conditional
Restricted Boltzmann Machines
Suppose that the hidden activation function is a sigmoid. Moreover, we define
our Gated Auto-encoder to consists of an encoder h(·) and decoder r(·) such
that
h(x, y) = h(W H ((W X x)
r(x|y, h) = (W X )T ((W Y y)
(W Y y)) + b)
(W H )T h(x, y)) + a,
where θ = {W H , W X , W Y , b} is the parameters of the model. Note that the
weights are not tied in this case. The energy function for the Gated Auto-encoder
16
will be:
Z
Eσ (x|y) =
=
(1 + exp (−W H (W X x)
X
x2
+ ax + const
2
x2
(W Y y) − bk )) −
+ ax + const.
2
(W Y y) − b))−1 du −
H
log (1 + exp (−Wk·
(W X x)
k
Now consider the free energy of a Factored Gated Conditional Restricted
Boltzmann Machine (FCRBM).
The energy function of a FCRBM with Gaussian visible units and Bernoulli
hidden units is defined by
E(x, h|y) =
X
(a − x)2
− bh −
WfX· x
2
2σ
WfY· y
WfH· h.
f
Given y, the conditional probability density assigned by the FCRBM to data
point x is
P
exp −(E(x, h|y))
exp (−F (x|y))
=
Z(y)
Z(y)
!
X
−F (x|y) = log
exp (−E(x, h|y))
h
p(x|y) =
h
P
where Z(y) = x,h exp (E(x, h|y)) is the partition function and F (x|y) is the
free energy function. Expanding the free energy function, we get
−F (x|y) = log
X
exp (−E(x, h|y))
h
2
−(a − x)
+ bh +
WfX· x WfY· y WfH· h
2σ 2
f
h
2
X
X
(a − x)
=−
+ log
exp bh +
WfX· x WfY· y WfH· h
2σ 2
h
f
2
X
Y
X
(a − x)
=−
+ log
exp bk hk +
(WfX· x WfY· y) WfHk hk
2σ 2
h
k
f
2
X
X
(a − x)
=−
+
log 1 + exp bk +
(WfHk )T (W X x W Y y) .
2σ 2
= log
X
exp
k
X
f
17
Note that we can center the data by subtracting mean of x and dividing by its
standard deviation, and therefore assume that σ 2 = 1. Substituting, we have
X
(a − x)2 X
−F (x|y) = −
+
log 1 + exp −bk −
(WfHk )T (W X x W Y y)
2
k
f
X
X
x2
=
log 1 + exp bk +
(WfHk )T (W X x W Y y) − a2 + ax −
2
k
f
X
X
x2
=
log 1 + exp bk +
(WfHk )T (W X x W Y y) + ax −
+ const
2
k
f
Letting W H = (W H )T , we get
=
X
log 1 + exp bk +
k
X
H
Wkf
(W X x
W Y y) + ax −
f
x2
+ const
2
Hence, the Conditional Gated Auto-encoder and the FCRBM are equal up to a
constant.
B.2
Gated Auto-encoder and mean-covariance Restricted
Boltzmann Machines
Theorem 2. Consider a covariance auto-encoder with an encoder and decoder,
h(x, x) = h(W H ((W F x)2 ) + b)
r(x|y = x, h) = (W F )T (W F y
(W H )T h(x, y)) + a,
where θ = {W F , W H , a, b} are the parameters of the model. Moreover, consider
a covariance Restricted Boltzmann Machine with Gaussian distribution over the
visibles and Bernoulli distribution over the hiddens, such that its energy function
is defined by
(a − x)2 X
E c (x, h) =
−
P h(Cx)2 − bh,
σ2
f
where θ = {P, C, a, b} are its parameters. Then the energy function for a covariance Auto-encoder with dynamics r(x|y) − x is equivalent to the free energy of a
covariance Restricted Boltzmann Machine. The energy function of the covariance
Auto-encoder is
X
E(x, x) =
log(1 + exp(W H (W F x)2 + b)) − x2 + const
(21)
k
18
Proof. Note that the covariance auto-encoder is the same as a regular Gated
Auto-encoder, but setting y = x and making the factor loading matrices the
same, i.e. W F = W Y = W X . Then applying the general energy equation for
GAE, Equation 8, to the covariance auto-encoder, we get
Z
1
E(x, x) = h(u)du − x2 + const
2
X
=
log(1 + exp(W H (W F x)2 + b)) − x2 + ax + const,
(22)
k
where u = W H (W F x)2 + b.
Now consider the free energy of the mean-covariance Restricted Boltzmann
Machine (mcRBM) with Gaussian distribution over the visible units and Bernoulli
distribution over the hidden units:
X
−F (x|y) = log
exp (−E(x, h|y))
h
(a − x)2
2
+ (P h)(Cx) + bh
= log
exp −
σ2
h
2
XY
X
(a
−
x)
= log
exp −
+
(Pf k hk )(Cx)2 + bk hk
σ2
h
k
f
X
X
(a − x)2
=
log 1 + exp (Pf k hk )(Cx)2 −
.
σ2
X
k
f
As before, we can center the data by subtracting mean of x and dividing by its
standard deviation, and therefore assume that σ 2 = 1. Substituting, we have
X
X
=
log 1 + exp (Pf k hk )(Cx)2 − (a − x)2 .
(23)
k
f
Letting W H = P T and W F = C, we get
X
X
=
log 1 + exp (Pf k hk )(Cx)2 − x2 + ax + const.
k
f
Therefore, the two equations are equivalent.
(24)
| 9cs.NE
|
A new class of tests for multinormality
with i.i.d. and garch data based on the
arXiv:1711.07199v1 [math.ST] 20 Nov 2017
empirical moment generating function
Norbert Henze1 , Marı́a Dolores Jiménez–Gamero2 ,
1
2
Institute of Stochastics, Karlsruhe Institute of Technology, Karlsruhe, Germany
Department of Statistics and Operations Research, University of Seville, Seville,
Spain
Abstract. We generalize a recent class of tests for univariate normality that are based
on the empirical moment generating function to the multivariate setting, thus obtaining a
class of affine invariant, consistent and easy-to-use goodness-of-fit tests for multinormality.
The test statistics are suitably weighted L2 -statistics, and we provide their asymptotic behavior both for i.i.d. observations as well as in the context of testing that the innovation
distribution of a multivariate GARCH model is Gaussian. We study the finite-sample behavior of the new tests, compare the criteria with alternative existing procedures, and apply the
new procedure to a data set of monthly log returns.
Keywords. Moment generating function; Goodness-of-fit test; multivariate normality; Gaussian GARCH model
AMS 2000 classification numbers: 62H15, 62G20.
1. Introduction
As evidenced by the papers Arcones (2007), Batsidis et al. (2013), Cardoso de
Oliveira and Ferreira (2010), Ebner (2012), Enomoto et al. (2012), Farrel et al. (2007),
Hanusz and Tarasińska (2008, 2012), Henze et al. (2017), Joenssen and Vogel (2014),
Jönsson (2011), Kim (2016), Koizumi et al. (2014), Mecklin and Mundfrom (2005),
Pudelko (2005), Székely and Rizzo (2005), Tenreiro (2011, 2017), Thulin (2014), Villa1
señor-Alva and Estrada (2009), Voinov et al. (2016), Yanada et al. (2015), and Zhou
and Shao (2014), there is an ongoing interest in the problem of testing for multivariate
normality. Without claiming to be exhaustive, the above list probably covers most of
the publications in this field since the review paper Henze (2002).
Recently, Henze and Koch (2017) provided the lacking theory for a test for univariate normality suggested by Zghoul (2010). The purpose of this paper is twofold.
First, we generalize the results of Henze and Koch (2017) to the multivariate case, thus
obtaining a class of affine invariant and consistent tests for multivariate normality. Second, in contrast to that paper (and most of the other publications), which considered
only independent and identically distributed (i.i.d.) observations, we also provide the
asymptotics of our test statistics in the context of GARCH-type dependence.
To be more specific, let (for the time being) X, X1 , X2 , . . . be a sequence of i.i.d.
d-variate random column vectors that are defined on a common probability space
(Ω, A, P). We assume that the distribution PX is absolutely continuous with respect to
Lebesgue measure. Let Nd (µ, Σ) denote the d-variate normal distribution with mean
vector µ and non-degenerate covariance matrix Σ, and write Nd for the class of all
non-degenerate d-dimensional normal distributions. A test for multivariate normality
is a test of the null hypothesis
H0 : PX ∈ Nd ,
and usually such a test should be consistent against any fixed non-normal alternative
distribution. Since the class Nd is closed with respect to full rank affine transformations,
any genuine test statistic Tn = Tn (X1 , . . . , Xn ) based on X1 , . . . , Xn should also be
affine invariant, i.e., we should have Tn (AX1 + b, . . . , AXn + b) = Tn (X1 , . . . , Xn ) for
each nonsingular d×d-matrix A and each b ∈ Rd , see Henze (2002) for a critical account
on affine invariant tests for multivariate normality.
P
P
In what follows, let X n = n−1 nj=1 Xj , Sn = n−1 nj=1 (Xj − X n )(Xj − X n )>
denote the sample mean and the sample covariance matrix of X1 , . . . , Xn , respectively,
where > means transposition of vectors and matrices. Furthermore, let
Yn,j = Sn−1/2 (Xj − X n ),
j = 1, . . . , n,
be the so-called scaled residuals of X1 , . . . , Xn , which provide an empirical standard2
−1/2
ization of X1 , . . . , Xn . Here, Sn
denotes the unique symmetric square root of Sn .
Notice that Sn is invertible with probability one provided that n ≥ d + 1, see Eaton
and Perlman (1973). The latter condition is tacitly assumed to hold in what follows.
Letting
n
1X
exp t> Yn,j ,
Mn (t) =
n j=1
t ∈ Rd ,
(1.1)
denote the empirical moment generating function of Yn,1 , . . . , Yn,n , Mn (t) should be
close to
m(t) = exp(ktk2 /2),
which is the moment generating function of the standard normal distribution Nd (0, Id ).
Here and in the sequel, k · k stands for the Euclidean norm on Rd , and Id is the unit
matrix of order d.
The statistic proposed in this paper is the weighted L2 -statistic
Z
(Mn (t) − m(t))2 wβ (t) dt,
Tn,β = n
(1.2)
Rd
where
wβ (t) = exp −βktk2 ,
(1.3)
and β > 1 is some fixed parameter, the role of which will be discussed later. Notice
that Tn,β is the ’moment generating function analogue’ to the BHEP-statistics for
testing for multivariate normality (see, e.g., Baringhaus and Henze (1988), Henze and
Zirkler (1990), Henze and Wagner (1997)). The latter statistics originate if one replaces
Mn (t) with the empirical characteristic function of the scaled residuals and m(t) with
the characteristic function exp(−ktk2 /2) of the standard normal distribution Nd (0, Id ).
For a general account on weighted L2 -statistics see, e.g., Baringhaus et al. (2017).
In principle, one could replace wβ in (1.3) with a more general weight function
satisfying some general conditions. The above special choice, however, leads to a test
criterion with certain extremely appealing features, since straightforward calculations
yield the representation
Tn,β = π
d/2
n
1 X 1
kYn,i + Yn,j k2
n
exp
+
d/2
n i,j=1 β
4β
(β − 1)d/2
!
n
X
1
kYn,j k2
−2
exp
,
d/2
(β
−
1/2)
4β
−
2
j=1
3
(1.4)
which is amenable to computational purposes. Notice that the condition β > 1 is
necessary for the integral in (1.2) to be finite. Later, we have to impose the further
restriction β > 2 to prove that Tnβ has a non-degenerate limit null distribution as n →
∞. We remark that Tn,β is affine invariant since it only depends on the Mahanalobis
>
Yn,j , 1 ≤ i, j ≤ n. Rejection of H0 is for large values of Tn,β .
angles and distances Yn,i
The rest of the paper unfolds as follows. The next section shows that letting β
tend to infinity in (1.2) yields a linear combination of two well-known measures of
multivariate skewness. In Section 3 we derive the limit null distribution of Tn,β in
the i.i.d. setting. Section 4 addresses the question of consistency of the new tests
against general alternatives, while Section 5 considers the new criterion in the context
of multivariate GARCH models in order to test for normality of innovations, and it
provides the pertaining large-sample theory. Section 6 presents a Monte Carlo study
that compares the new tests with competing ones, and it considers a real data set from
the financial market. The article concludes with discussions in Section 7.
2. The case β → ∞
In this section, we show that the statistic Tn,β , after a suitable scaling, approaches
a linear combination of two well-known measures of multivariate skewness as β → ∞.
Theorem 2.1 We have
lim β 3+d/2
β→∞
96Tn,β
= 2b1,d + 3eb1,d ,
nπ d/2
where
b1,d
n
3
1 X
>
= 2
Yn,j
Yn,k ,
n j,k=1
n
X
>
eb1,d = 1
Yn,j
Yn,k kYn,j k2 kYn,k k2
2
n j,k=1
are multivariate sample skewness in the sense of Mardia (1970) and Móri, Rohatgi and
Székely (1993), respectively.
Proof. Let b2,d = n−1
Pn
j=1
kYn,j k4 denote multivariate sample kurtosis in the sense
of Mardia (1970). From (1.4) and
exp(y) = 1 + y +
4
y2 y3
+
+ O(y 4 )
2
6
as y → 0, the result follows by very tedious but straightforward calculations, using the
P
P
P
relations nj=1 Yn,j = 0, nj=1 kYn,j k2 = nd, nj,k=1 kYn,j + Yn,k k2 = 2n2 d,
n
X
kYn,j + Yn,k k4 = 2n2 b2,d + d2 + 2d ,
j,k=1
n
X
>
kYn,j + Yn,k k4 Yn,j
Yn,k = 8n2 b2,d + 4n2 b1,d + 2n2eb1,d ,
j,k=1
n
X
kYn,j + Yn,k k6 = 2n
n
X
kYn,j k6 + 6(d+4)n2 b2,d + 8n2 b1,d + 12n2eb1,d .
j=1
j,k=1
For the derivation of the second but last expression, see the proof of Theorem 4.1 of
P
Henze et al. (2017). We stress that although b2,d and nj=1 kYn,j k6 show up in some of
the equations above, these terms cancel out in the derivation of the final result.
Remark 2.2 Interestingly, Tn,β exhibits the same limit behavior as β → ∞ as both
the statistic studied by Henze et al. (2017), which is based on a weighted L2 -distance
involving both the empirical characteristic function and the empirical moment generating function, and the BHEP-statistic for testing for multivariate normality, which is
based on the empirical characteristic function, see Theorem 2.1 of Henze (1997). At
first sight, Theorem 2.1 seems to differ from Theorem 4 of Henze and Koch (2017)
which covers the special case d = 1, but a careful analysis shows that – with the
notation τ (β) in that paper – we have limβ→∞ β 7/2 τ (β) = 0.
3. Asymptotic null distribution in the i.i.d. case
In this section we consider the case that X1 , X2 , . . . are i.i.d. d-dimensional random vectors with some non-degenerate normal distribution. The key observation for
deriving the limit distribution of Tn,β is the fact that
Z
Tn,β =
Wn (t)2 wβ (t) dt,
Rd
where
Wn (t) =
√
n (Mn (t) − m(t)) ,
t ∈ Rd ,
(3.1)
with Mn (t) given in (1.1). Notice that Wn is a random element of the Hilbert space
L2β := L2 (Rd , B d , wβ (t)dt)
5
(3.2)
of (equivalence classes of) measurable functions f : Rd → R that are square integrable
with respect to the finite measure on the σ-field B d of Borel sets of Rd given by the
weight function wβ defined in (1.3). The resulting norm in L2β will be denoted by
p
kf kL2β = hf, f i. With this notation, Tn,β takes the form
Tn,β = kWn k2L2 .
(3.3)
β
D
Writing ”−→” for convergence in distribution of random vectors and stochastic processes, the main result of this section is as follows.
Theorem 3.1 (Convergence of Wn under H0 )
Suppose that X has some non-degenerate d-variate normal distribution, and that β > 2
in (1.3). Then there is a centred Gaussian random element W of L2 having covariance
kernel
C(s, t) = exp
ksk2 + ktk2
2
s> t
e
s> t
−1−s t−
2
2 !
>
,
s, t ∈ Rd ,
D
so that Wn −→ W as n → ∞.
In view of (3.3), the Continuous Mapping Theorem yields the following result.
Corollary 3.2 If β > 2, then, under the null hypothesis H0 ,
D
Tn,β −→ kW k2L2 as n → ∞.
β
Remark 3.3 The distribution of T∞,β := kW k2L2 (say) is that of
β
P∞
j=1
λj Nj2 , where
λ1 , λ2 , . . . are the positive eigenvalues of the integral operator f 7→ Af on L2β associated
R
with the kernel C given in Theorem 3.1, i.e., (Af )(t) = C(s, t)f (s) exp(−βksk2 )ds,
and N1 , N2 , . . . are i.i.d. standard normal random variables. We did not succeed in
obtaining explicit solutions of this equation. However, since
Z
E(T∞,β ) =
C(t, t) wβ (t) dt,
d
R
Z Z
C 2 (s, t)wβ (s)wβ (t) dsdt
V(T∞,β ) = 2
Rd
Rd
(see Shorack and Wellner, 1986, p. 213), tedious but straighforward manipulations of
integrals yield the following result, which generalizes Theorem 2 of Henze and Koch
(2017).
6
Theorem 3.4 If β > 2 we have
a)
E(T∞,β ) = π
d/2
1
1
d
d(d + 2)
−
−
−
d/2
d/2
d/2+1
(β − 2)
(β − 1)
2(β − 1)
8(β − 1)d/2+2
,
b)
V(T∞,β ) = 2π
d
1
2d+1 (1 + 2d)2d d(d + 2)2d
−
−
−
(β(β − 2))d/2
η d/2
η d/2+1
η d/2+2
d
3d(d + 2)
1
,
+
+
+
(β − 1)d 2(β − 1)d+2 64(β − 1)d+4
where η = 4(β − 1)2 − 1.
Proof of Theorem 3.1. In view of affine invariance, we assume w.l.o.g. that the
distribution of X is Nd (0, Id ). In Henze et al. (2017), the authors considered the
“exponentially down-weighted empirical moment generating function process”
ktk2
An (t) = exp −
Mn (t), t ∈ Rd .
2
(3.4)
Notice that, with the notation given in (3.2), we have
kAn k2L2 = kMn k2L2γ ,
β
where γ = β − 1 From display (10.5) and Propositions 10.3 and 10.4 of Henze et al.
(2017) we have
ktk2
An (t) = exp −
2
where
R
Rd
√
n
!
n
1 X t> Xj
e
− m(t) + Vn (t) + Rn (t),
n j=1
Rn2 (t)wγ (t)dt = oP (1), and
n
n
1 X >
1 X >
Vn (t) = − √
(t Xj )2 − ktk2 − √
t Xj .
2 n j=1
n j=1
Display (3.4) and the representation of An as a sum yield
n
1 X
Wn (t) = √
Zj (t) + m(t)Rn (t),
n j=1
7
where
>X
j
Zj (t) = et
− m(t) −
m(t) >
(t Xj )2 − ktk2 − m(t)t> Xj .
2
Notice that Z1 , Z2 , . . . are i.i.d. centred random elements of L2β . Since
Z
Z
2
(m(t)Rn (t)) wβ (t) dt =
Rn2 (t)wγ (t) dt = oP (1),
Rd
Rd
a Central Limit Theorem in Hilbert spaces (see e.g., Bosq, 2000) shows that there is a
D
centered Gaussian random element W of L2β , so that Wn −→ W. Using the fact that
t> X has the normal distribution N(0, ktk2 )
h >
i
E es X (t> X)2 =
h >
i
s X >
E e
t X =
> 2 > 2
E (s X) (t X)
=
and the relations
m(s) (s> t)2 + ktk2 ,
m(s)s> t,
2(s> t)2 + ksk2 ktk2 ,
some straightforward algebra shows that the covariance kernel C(s, t) figuring in the
statement of Theorem 3.1 equals EZ1 (s)Z1 (t).
4. Consistency
The next result shows that the test for multivariate normality based on Tn,β is
consistent against general alternatives.
Theorem 4.1 Suppose X has some absolutely continuous distribution, and that MX (t) =
e = Σ−1/2 (X − µ), where µ = E(X) and
E[exp(t> X)] < ∞, t ∈ Rd . Furthermore, let X
Σ−1/2 is the symmetric square root of the inverse of the covariance matrix Σ of X.
e we have
Letting MXe (t) = E[exp(t> X)],
Z
2
Tn,β
lim inf
≥
MXe (t) − m(t) wβ (t) dt
n→∞
n
Rd
almost surely.
Proof. Because of affine invariance we may w.l.o.g. assume EX = 0 and Σ = Id . Fix
P
K > 0 and put Mn◦ (t) = n−1 nj=1 exp(t> Xj ). From the proof of Theorem 6.1 of Henze
et al. (2017) we have
lim max Mn (t) − Mn◦ (t) = 0
n→∞ ktk≤K
8
P-almost surely. Now, the strong law of large numbers in the Banach space of continuous functions on B(K) := {t ∈ Rd : ktk ≤ K} and Fatou’s lemma yield
Z
Tn,β
lim inf
≥ lim inf
(Mn (t) − m(t))2 wβ (t) dt
n→∞
n→∞
n
B(K)
Z
>
2
≥
Eet X − m(t) wβ (t) dt
B(K)
P-almost surely. Since K is arbitrary, the assertion follows.
Now, suppose that X has an alternative distribution (which is assumed to be
standardized) satisfying the conditions of Theorem 4.1. Since E exp(t> X) − m(t) 6= 0
for at least one t, Theorem 4.1 shows that limn→∞ Tn,β = ∞ P-almost surely. Since, for
any given nominal level α ∈ (0, 1), the sequence of critical values of a level-α-test based
on Tnβ that rejects H0 for large values of Tn,β converges according to Theorem 3.1, this
test is consistent against such an alternative. It should be ’all the more consistent’
against any distribution not satisfying the conditions of Theorem 4.1 but, in view of
the reasoning given in Csörgő (1986), the behavior of Tn,β against such alternatives is
a difficult problem.
5. Testing for normality in GARCH models
In this section we consider the multivariate GARCH (MGARCH) model
1/2
Xj = Σj (θ)εj ,
j ∈ Z,
(5.1)
where θ ∈ Θ ⊆ Rv is a v-dimensional vector of unknown parameters. The unobservable
random errors or innovations {εj , j ∈ Z} are i.i.d. copies of a d-dimensional random
vector ε, which is assumed to have mean zero and unit covariance matrix. Hence
Σj (θ) = Σ(θ; Xj−1 , Xj−2 , . . .)
is the conditional variance of Xj , given Xj−1 , Xj−2 , . . .. The explicit expression of Σj (θ)
depends on the assumed MGARCH model (see, e.g., Francq and Zakoı̈an, 2010, for a
detailed description of several relevant models). The interest in testing for normality
of the innovations stems from the fact that this distributional assumption is made in
9
some applications, and that, if erroneously accepted, some inferential procedures can
lead to wrong conclusions (see, e.g., Spierdijk, 2016, for the effect on the assessment of
standard risk measures such as the value at risk).
Therefore, an important step in the analysis of GARCH models is to check whether
the data support the distributional hypotheses made on the innovations. Because of
this reason, a number of goodness-of-fit tests have been proposed for the innovation
distribution. The papers by Klar et al. (2012) and Ghoudi and Rémillard (2014)
contain an extensive review of such tests as well as some numerical comparisons between
them for the special case of testing for univariate normality. The proposals for testing
goodness-of-fit in the multivariate case are rather scarce.
The class of GARCH models has been proved to be particularly valuable in modeling financial data. As discussed, among others in Rydberg (2000), one of the stylized
features of financial data is that they are heavy-tailed. From an extensive simulation
study (a summary is reported in Section 6), we learnt that, for i.i.d. data, the test
of normality based on Tn,β exhibits a high power against heavy-tailed distributions.
Because of these reasons, this section is devoted to adapt that procedure to testing
for normality of the innovations based on data X1 , . . . , Xn that are driven by equation
(5.1). Therefore, on the basis of the observations, we wish to test the null hypothesis
H0,G : The law of ε is Nd (0, Id ).
against general alternatives. Notice that H0,G is equivalent to the hypothesis that,
conditionally on {Xj−1 , Xj−2 , . . .}, the law of Xj is Nd (0, Σj (θ)), for some θ ∈ Θ. Two
main differences with respect to the i.i.d. case are: (a) the innovations in (5.1) are
assumed to be centered at zero with unit covariance matrix; and (b) the conditional
covariance matrix Σj (θ) of Xj is time-varying in a way that depends on the unknown
parameter θ and on past observations.
Notice that although H0,G is about the distribution of ε, the innovations are unobservable in the context of model (5.1). Hence any inference on the distribution of
the innovations should be based on residuals
e −1/2 (θbn )Xj ,
εej (θbn ) = Σ
j
1 ≤ j ≤ n.
(5.2)
Recall that Σj (θ) = Σ(θ; Xj−1 , Xj−2 , . . . ), but we only observe X1 , . . . , Xn . Therefore,
10
to estimate Σj (θ), apart from a suitable estimator θbn of θ, we also need to specify values
ej , j ≤ 0}. So we write Σ
e j (θ) for Σ(θ; Xj−1 , . . . , X1 , X
e0 , X
e−1 . . .).
for {Xj , j ≤ 0}, say {X
Under certain conditions, these arbitrarily fixed initial values are asymptotically irrelevant.
Taking into account that the innovations have mean zero and unit covvariance
matrix, we will work directly with the residuals, without standardizing them. Let MnG
be defined as Mn in (1.1) by replacing Yn,j with εej (θbn ), 1 ≤ j ≤ n, and define T G as
n,β
Tn,β in (3.3) with Wn changed for
WnG ,
where
WnG
is defined as Wn in (3.1) with Mn
replaced by MnG . In order to derive the asymptotic null distribution of WnG we will
make the assumptions (A.1)–(A.6) below. In the sequel, C > 0 and %, 0 < % < 1,
denote generic constants, the values of which may vary across the text, θ0 stands for
P
the true value of θ, and for any matrix A = (akj ), kAk = k,j |akj | denotes the l1 -norm
(we use the same notation as for the Euclidean norm of vectors).
P
√
(A.1) The estimator θbn satisfies n(θbn − θ0 ) = n−1/2 nj=1 Lj + oP (1), where Lj = hj gj ,
gj = g(θ0 ; εj ) is a vector of d2 measurable functions such that E(gj ) = 0 and
E(gj> gj ) < ∞, and hj = h(θ0 ; εj−1 , εj−2 . . .) is a v × d2 -matrix of measurable
2
functions satisfying E(khj h>
j k ) < ∞,
e −1/2 (θ) ≤ C,
(A.2) supθ∈Θ Σ
j
−1/2
supθ∈Θ Σj
(θ) ≤ C
P-a.s.,
1/2
e 1/2 (θ)k ≤ C%j ,
(A.3) supθ∈Θ kΣj (θ) − Σ
j
1/2
(A.4) E kXj kς < ∞ and E Σj (θ0 )
ς
< ∞ for some ς > 0,
(A.5) for any sequence x1 , x2 , . . . of vectors of Rd , the function θ 7→ Σ1/2 (θ; x1 , x2 , . . . )
admits continuous second-order derivatives,
(A.6) for some neighborhood V (θ0 ) of θ0 , there exist p > 1, q > 2 and r > 1 so that
11
2p−1 + 2r−1 = 1 and 4q −1 + 2r−1 = 1, and
v
X
E sup
θ∈V (Θ)
k,`=1
v
X
E sup
θ∈V (Θ)
∂
−1/2
Σj (θ)
k=1
p
1/2
Σj (θ)
∂θk ∂θ`
< ∞,
q
1/2
∂Σj (θ)
−1/2
Σj (θ)
∂θk
−1/2
1/2
E sup
2
Σj (θ0 )Σj
< ∞,
r
(θ)
< ∞.
θ∈V (Θ)
The next result gives the asymptotic null distribution of WnG .
Theorem 5.1 (Convergence of WnG under H0,G )
Let {Xj } be a strictly stationary process satisfying (5.1), with Xj being measurable
with respect to the sigma-field generated by {εu , u ≤ j}. Assume that (A.1)–(A.6) hold
and that β > 2. Then under the null hypothesis H0,G , there is a centered Gaussian
random element WG of L2β , having covariance kernel CG (s, t) = cov(U (t), U (s)), so
D
that WnG −→ WG as n → ∞, where
U (t) = exp(t> ε1 ) − m(t) − m(t)a(t)> L1 ,
−1/2
a(t)> = (t> µ1 t, . . . , t> µv t), µk = E[A1k (θ0 )], A1k (θ) = Σ1
1/2
(θ) ∂θ∂k Σ1 (θ), 1 ≤ k ≤ v.
From Theorem 5.1 and the Continuous Mapping Theorem we have the following
corollary.
Corollary 5.2 Under the assumptions of Theorem 5.1, we have
D
G
Tn,β
−→ kWG k2L2 as n → ∞.
β
The standard estimation method for the parameter θ in GARCH models is the
quasi maximum likelihood estimator (QMLE), defined as
θbn = arg max Ln (θ),
θ∈Θ
where
n
1 Xe
Ln (θ) = −
`j (θ),
2 j=1
e j (θ)−1 Xj + log Σ
e j (θ) .
`ej (θ) = Xj> Σ
12
Comte and Leiberman (2003) and Bardet and Wintenberger (2009), among others,
have shown that under certain mild regularity conditions the QMLE satisfies (A.1) for
general MGARCH models.
As observed before, there are many MGARCH parametrizations for the matrix
Σj (θ). Nevertheless, there exist only partial theoretical results for such models. The
Constant Conditional Correlation model, proposed by Bollerslev (1990) and extended
by Jeantheau (1998), is an exception, since its properties have been thoroughly studied. This model decomposes the conditional covariance matrix figuring in (5.1) into
conditional standard deviations and a conditional correlation matrix, according to
Σj (θ0 ) = Dj (θ0 )R0 Dj (θ0 ), where Dj (θ0 ) and R0 are d × d-matrices, R0 is a correla
tion matrix, and Dj (θ0 ) is a diagonal matrix so that σj2 (θ) = diag Dj2 (θ) with
σj2 (θ)
=b+
p
X
(2)
Bk Xj−k
k=1
(2)
Here, Xj
= Xj
Xj , where
+
q
X
2
Γk σj−k
(θ).
(5.3)
k=1
denotes the Hadamard product, that is, the element by
element product, b is a vector of dimension d with positive elements, and {Bk }pk=1 and
{Γk }qk=1 are d × d matrices with non-negative elements. This model will be referred to
as CCC-GARCH(p,q). Under certain weak assumptions, the QMLE for the parameters
in this model satisfies (A.1), and (A.2)–(A.6) also hold, see Francq and Zakoı̈an (2010)
and Francq et al. (2017).
G
depends on the equation defining the
The asymptotic null distribution of Tn,β
GARCH model and on θ0 through the quantities µ1 , . . . , µv , as well as on which estimator of θ has been employed. Therefore, the asymptotic null distribution cannot
G
be used to approximate the null distribution of Tn,β
. Following Klar et al. (2012), we
G
will estimate the null distribution of Tn,β
by using the following parametric bootstrap
algorithm:
(i) Calculate θbn = θbn (X1 , . . . , Xn ), the residuals εe1 , . . . , εen and the test statistic
G
G
Tn,β
= Tn,β
(e
ε1 , . . . , εen ).
b ∗,
e 1/2 (θ)ε
(ii) Generate vectors ε∗1 , . . . , ε∗n i.i.d. from a Nd (0, Id ) distribution. Let Xj∗ = Σ
j
j
j = 1, . . . , n.
(iii) Calculate θbn∗ = θbn (X1∗ , . . . , Xn∗ ), the residuals εe∗1 , . . . , εe∗n , and approximate the null
13
G
distribution of Tn,β
by means of the conditional distribution, given the data, of
G
G∗
(e
ε∗1 , . . . , εe∗n ).
= Tn,β
Tn,β
In practice, the approximation in step (iii) is carried out by generating a large
G
number of bootstrap replications of the test statistic Tn,β
, whose empirical distribution
G
. Similar steps to those given
function is used to estimate the null distribution of Tn,β
in the proof of Theorem 5.1 show that if one assumes that (A.1)–(A.6) continue to
hold when θ0 is replaced by θn , with θn → θ0 as n → ∞, and ε ∼ Nd (0, Id ), then
G∗
, given the data, converges in law to kWG k2L2 , with
the conditional distribution of Tn,β
β
WG as defined in Theorem 5.1. Therefore, the above bootstrap procedure provides a
consistent null distribution estimator.
Remark 5.3 The practical application of the above bootstrap null distribution estimator entails that the parameter estimator of θ and the residuals must be calculated for
each bootstrap resample, which results in a time-consuming procedure. Following the
approaches in Ghoudi and Rémillard (2014) and Jiménez-Gamero and Pardo-Fernández
(2017) for other goodness-of-fit tests for univariate GARCH models, we could use a
weighted bootstrap null distribution estimator in the sense of Burke (2000). From a
computational point of view, it provides a more efficient estimator. Nevertheless, it can
be verified that the consistency of the weighted bootstrap null distribution estimator of
G
Tn,β
requires the existence of the moment generating function of the true distribution
generating the innovations, which is a rather strong condition, specially taking into
account that the alternatives of interest are heavy-tailed.
As in the i.i.d. case, the next result shows that the test for multivariate normality
G
based on Tn,β
is consistent against general alternatives.
Theorem 5.4 Let {Xj } be a strictly stationary process satisfying (5.1), with Xj being measurable with respect to the sigma-field generated by {εu , u ≤ j}. Assume
that (A.1)–(A.6) hold, that ε has some absolutely continuous distribution, and that
Mε (t) = E[exp(t> ε)] < ∞, t ∈ Rd . We then have
Z
G
Tn,β
lim inf
≥
(Mε (t) − m(t))2 wβ (t) dt
n→∞
n
d
R
in probability.
14
Similar comments to those made after Theorem 4.1 for the i.i.d. case can be done
in this setting.
Proof of Theorem 5.1. From the proof of Theorem 7.1 in Henze et al. (2017), it
P
G
G
(t) = n−1/2 nj=1 Vj (t),
(t) + rn,1 (t), with W1,n
follows that WnG (t) = W1,n
√
Vj (t) = exp(t> εj ) − m(t)a(t)> n(θbn − θ0 ) − m(t),
G
G
G
(t) =
(t) + rn,2 (t), with W2,n
(t) = W2,n
and krn,1 kL2β = oP (1). By Assumption A.1, W1,n
Pn
−1/2
n
j=1 Uj (t),
Uj (t) = exp(t> εj ) − exp(−ktk2 /2)a(t)> Lj − exp(−ktk2 /2),
and krn,2 kL2β = oP (1).
G
(t), t ∈
To prove the result we will apply Theorem 4.2 in Billingsley (1968) to {W2,n
G
(t), t ∈ B(K)} converges in law to
Rd } by showing that (a) for each positive M , {W2,n
{W G (t), t ∈ B(K)} in C(B(K)), the Banach space of real-valued continuous functions
on B(K) := {t ∈ Rd : ktk ≤ K}, endowed with the supremum norm; (b) for each
ε > 0, there is a positive K so that
Z
G 2
E W2,n
(t) wβ (t) dt < ε,
(5.4)
Rd \B(K)
Z
E W G (t)2 wβ (t) dt < ε.
(5.5)
Rd \B(K)
Proof of (a): By applying the central limit theorem for martingale differences, the
G
(t), t ∈ Rd } converge to those of {WG (t), t ∈
finite-dimensional distributions of {W2,n
G
Rd }. Hence, to prove (a) we must show that {W2,n
(t), t ∈ B(K)} is tight. With this
P
G
G
G
G
aim we write W2,n
(t) = W3,n
(t)−W4,n
(t), with W3,n
(t) = n−1/2 nj=1 {exp(t> εj )−m(t)}
P
G
and W4,n
(t) = m(t)a(t)> n−1/2 nj=1 Lj . The mean value theorem gives
E {exp(t> ε) − exp(s> ε)}2 ≤ κkt − sk2 ,
s, t ∈ B(K),
G
for some positive κ. From Theorem 12.3 in Billingsley (1968), the process {W3,n
(t), t ∈
P
B(K)} is tight. By the central limit theorem for martingale differences, n−1/2 nj=1 Lj
G
converges in law to a v-variate zero mean normal random vector. Hence {W4,n
(t), t ∈
B(K)}, being a product of a continuous function and a term which is OP (1), is tight,
G
and the same property holds for {W2,n
(t), t ∈ B(K)}.
15
G 2
Proof of (b): In view of E W2,n
(t) = E [U1 (t)2 ] < ∞, for each ε > 0 there is some
positive constant K so that (5.4) holds. Likewise, (5.5) holds, which completes the
proof.
−1/2
fG (t) =
Proof of Theorem 5.4. Let εj (θ) = Σj (θ)Xj . Notice that εj (θ0 ) = εj . Let M
n
Pn
Pn
Pn
−1
>
G
−1
>
◦
−1
>
c
n
ej (θ̂n )}, Mn (t) = n
j=1 exp{t ε
j=1 exp{t εj (θ̂n )}, Mn (t) = n
j=1 exp{t εj }
and B(K) := {t ∈ Rd : ktk ≤ K}. To show the result we will prove
cG (t) − M ◦ (t)| = oP (1),
(a) supt∈B(K) |M
n
n
fG (t) − M
cG (t)| = oP (1),
(b) supt∈B(K) |M
n
n
and the result will follow by using the same proof as in the i.i.d. case.
−1/2
1/2
Proof of (a): Let θbn = (θbn1 , . . . , θbnv )> , θ0 = (θ01 , . . . , θ0v )> and Ajk (θ) = Σj (θ) ∂θ∂k Σj (θ).
P
We have εj (θbn ) = εj + ∆n,j , with ∆n,j = − vk=1 Ajk (θen,j )εj (θbnk − θ0k ), for some θen,j
between θbn and θ0 . Observe that exp(t> ∆n,j ) − 1 = t> ∆n,j exp(αn,j t> ∆n,j ) for some
αn,j ∈ (0, 1). Now (A.1) and (A.6) yield k∆n,j k ≤ Dj kεj kkθbn − θ0 k for large enough n,
where E(Dj2 ) < ∞. The Cauchy–Schwarz inequality gives
n
X
cnG (t) − Mn◦ (t)| = 1
|M
exp(t> εj ) exp(t> ∆n, j) − 1
n j=1
≤ r1,n (t)1/2 r2,n (t)1/2 ,
where r1,n (t) = Mn (2t), and
X
n
1
b
b
r2,n (t) = ktk kθn − θ0 k exp 2ktkkθn − θ0 k max Dj kεj k
Dj2 kεj k2 .
1≤j≤n
n j=1
2
2
From the strong law of large numbers in the Banach space of continuous functions on
B(K), we have
sup r1,n (t) ≤ sup Mε (2t) + sup |MnG (2t) + Mε (2t)| < K1
t∈B(K)
t∈B(K)
P-a.s.
t∈B(K)
for some positive constant K1 . From the ergodic theorem, n−1
Pn
j=1
Dj2 kεj k2 < K2
P-almost surely for some positive constant K2 . Using stationarity and finite second
√
moments, if follows that max1≤j≤n Dj kεj k/ n → 0, P-almost surely. Hence (A.1)
yields supt∈B(K) r2,n (t) → 0, in probability. This concludes the proof of (a).
Proof of (b): The reasoning follows similar steps as the proof of fact (c.1) in the proof
of Theorem 7.1 in Henze et al. (2017) and is thus omitted.
16
6. Monte Carlo results
This section describes and summarizes the results of an extensive simulation experiment carried out to study the finite-sample performance of the proposed tests.
Moreover, we consider a real data set of monthly log returns. All computations have
been performed using programs written in the R language.
6.1. Numerical experiments for i.i.d. data
Upper quantiles of the null distribution of Tn,β have been approximated by generating 100,000 samples from a law Nd (0, Id ). Table 1 displays some critical values with
the convention that an entry like
−4
1.17 stands for 1.17 × 10−4 . The results show that
large sample sizes are required to approximate the critical values by their corresponding
asymptotic values.
A natural competitor of the test based on Tn,β is the CF-based test studied in
Henze and Wagner (1997) (HW-test). The latter procedure is simple to compute as
well as affine invariant, and it has revealed good power performance with regard to
competitors. The behaviour of the test based on Tn,β in relation to the HW-test depends
on whether the distribution is heavy-tailed or not. We tried a number of non-heavytailed distributions (specifically, the multivariate Laplace distribution, finite mixtures
of normal distributions, the skew-normal distribution, the multivariate χ2 -distribution,
the Khintchine distribution, the uniform distribution on [0, 1]d and the Pearson type
II family). For these distributions we observed that the power of the proposed test is
either similar or smaller than that of the HW-test; for very heavy-tailed distributions,
the new test outperforms the HW-test. This observation can be appreciated by looking
at Table 2, which displays the empirical power calculated by generating 10,000 samples
(in each case), for the significance level α = 0.05, from the following heavy-tailed
alternatives: (ASEθ ) the θ-stable and elliptically-contoured distribution and the (Tθ )
multivariate Student’s t with θ degrees of freedom. The same fact was also observed
in Zghoul (2010), who numerically studied the test based on Tn,β for univariate data.
In our simulations we tried a large number of values for β for the proposed test as
well as for the HW-test. The tables display the results for those values of β giving the
highest power in most of the cases considered. The same can be said for the simulations
17
Table 1: Critical points for π −d/2 Tn,β .
β
d
n
α
2.5
3.0
2
20
0.05
0.10
0.05
0.10
0.05
0.10
0.05
0.10
0.05
0.10
0.05
0.10
0.213
0.339
0.391
0.661
0.511
0.868
0.612
1.028
0.679
1.132
0.701
1.148
−1 0.751
−2 3.269
−2 1.639
−1 1.147
−2 4.857
−1 1.246
0.05
0.10
0.05
0.10
0.05
0.10
0.05
0.10
0.05
0.10
0.05
0.10
0.05
0.10
0.05
0.10
0.05
0.10
0.05
0.10
0.05
0.10
0.05
0.10
50
100
200
300
400
3
20
50
100
200
300
400
5
20
50
100
200
300
400
3.5
4.0
6.0
10.0
−3 5.408
−3 2.266
−4 2.241
−2 2.380
−3 7.638
−3 3.150
−4 3.025
−2 5.098
−2 2.436
−3 7.594
−3 3.078
−4 2.875
−1 1.997
−2 7.802
−2 3.624
−3 10.917
−3 4.330
−4 3.897
−1 1.539
−2 6.073
−2 2.838
−3 8.620
−3 3.429
−4 3.111
−1 2.432
−2 9.168
−2 4.153
−3 12.094
−3 4.724
−4 4.143
−1 1.757
−2 6.719
−2 3.085
−3 9.181
−3 3.616
−4 3.232
−1 2.726
−2 9.908
−2 4.382
−3 12.528
−3 4.845
−4 4.221
−1 1.894
−2 7.114
−2 3.223
−3 9.466
−3 3.719
−4 3.296
−1 2.878
−2 10.259
−2 4.518
−3 12.748
−3 4.905
−4 4.232
−1 1.925
−2 7.165
−2 3.248
−3 9.502
−3 3.721
−4 3.283
−1 2.868
−2 10.084
−2 4.417
−3 12.521
−3 4.843
−4 4.187
0.356
0.520
0.719
1.153
0.988
1.646
1.232
2.046
1.332
2.187
1.397
2.245
−1 1.066
−2 4.095
−2 1.851
−3 5.218
−3 1.942
−4 1.413
−1 1.504
−2 5.629
−2 2.503
−3 6.886
−3 2.518
−4 1.773
−1 1.906
−2 6.760
−2 2.894
−3 7.598
−3 2.709
−4 1.828
−1 2.879
−2 9.789
−2 4.073
−3 10.317
−3 3.593
−4 2.334
−1 2.433
−2 8.258
−2 3.426
−3 8.696
−3 3.043
−4 1.992
−1 3.732
−2 11.943
−2 4.788
−3 11.572
−3 3.945
−4 2.489
−1 2.851
−2 9.322
−2 3.781
−3 9.365
−3 3.231
−4 2.078
−1 4.319
−2 13.243
−2 5.167
−3 12.210
−3 4.123
−4 2.567
−1 2.979
−2 9.555
−2 3.849
−3 9.431
−3 3.242
−4 2.072
−1 4.441
−2 13.364
−2 5.156
−3 12.105
−3 4.073
−4 2.527
−1 3.061
−2 9.725
−2 3.893
−3 9.509
−3 3.260
−4 2.084
−1 4.481
−2 13.341
−2 5.122
−3 12.010
−3 4.046
−4 2.519
0.597
0.774
1.519
2.332
2.315
3.782
3.047
4.969
3.346
5.445
3.608
5.838
−1 1.347
−2 4.130
−2 1.554
−3 3.275
−3 0.971
−5 3.884
−1 1.691
−2 5.089
−2 1.886
−3 3.900
−3 1.142
−5 4.474
−1 2.868
−2 7.862
−2 2.731
−3 5.215
−3 1.460
−5 5.283
−1 4.130
−2 10.801
−2 3.633
−3 6.633
−3 1.809
−5 6.260
−1 3.947
−2 10.132
−2 3.381
−3 6.134
−3 1.667
−5 5.768
−1 5.884
−2 14.199
−2 4.530
−3 7.779
−3 2.051
−5 6.782
−1 4.744
−2 11.541
−2 3.736
−3 6.565
−3 1.755
−5 5.962
−1 6.964
−2 15.880
−2 4.896
−3 8.131
−3 2.112
−5 6.875
−1 5.016
−2 11.974
−2 3.829
−3 6.636
−3 1.769
−5 5.985
−1 7.343
−2 16.307
−2 4.960
−3 8.119
−3 2.100
−5 6.832
−1 5.234
−2 12.292
−2 3.889
−3 6.679
−3 1.776
−5 5.997
−1 7.586
−2 16.477
−2 4.958
−3 8.085
−3 2.085
−5 6.821
18
5.0
Table 2: Percentage of rejection for nominal level α = 0.05 and n = 50.
ASE1.75
ASE1.85
ASE1.95
T5
T7
T10
Test based on Tn,β
HW-test
β
β
d
3
3.5
4.0
5.0
6.0
10.0
0.1
0.5
1.0
2
3
5
2
3
5
2
3
5
2
3
5
2
3
5
2
3
5
72.47
82.70
90.51
54.43
62.72
75.31
24.67
29.31
38.28
58.77
40.59
87.14
42.33
55.23
71.51
28.80
38.34
51.64
72.62
82.78
90.86
54.39
62.67
75.52
24.62
29.37
38.39
58.82
40.76
87.80
42.21
55.49
72.50
28.82
38.56
52.38
72.43
82.76
90.95
54.35
62.61
75.65
24.52
29.47
38.37
58.74
40.98
88.43
42.16
55.51
73.17
28.73
38.58
53.01
72.08
82.69
91.25
53.91
62.46
75.96
24.22
29.12
38.27
58.26
41.11
89.17
41.86
55.62
74.07
28.29
38.51
53.91
71.59
82.52
91.32
53.55
62.39
76.03
24.11
28.91
38.00
57.82
41.34
89.59
41.43
55.30
74.47
27.99
38.33
54.25
70.34
81.92
91.05
52.59
61.58
75.63
23.56
28.45
37.55
56.21
41.31
89.76
39.68
54.38
74.30
27.12
37.00
54.33
67.29
79.07
88.89
50.00
57.95
71.81
22.44
26.37
33.99
51.44
39.69
86.36
36.19
49.20
69.35
24.48
32.97
48.85
67.75
78.16
87.46
48.17
54.67
66.66
20.78
24.04
29.35
54.20
37.79
87.21
36.30
48.82
67.75
22.94
30.55
45.36
59.91
68.60
75.71
39.35
42.44
47.82
15.38
16.79
17.39
47.58
28.99
77.92
28.97
37.89
51.33
16.23
20.84
28.41
in the next subsection.
6.2. Numerical experiments for GARCH data
In our simulations we considered a bivariate CCC–GARCH(1,1) model with
0.1
0.1 0.1
γ 0.01
1 r
, B1 =
, Γ1 =
, R =
,
b=
0.1
0.1 0.1
0.01 γ
r 1
for γ = 0.3, 0.4, 0.5 and r = 0, 0.3, and a trivariate CCC–GARCH(1,1) model with
b = (0.1, 0.1, 0.1)0 ,
γ 0.01 0.01
1 r r
0.1 0.1 0.1
B1 =
0.1 0.1 0.1 , Γ1 = 0.01 γ 0.01 , R = r 1 r
0.1 0.1 0.1
0.01 0.01 γ
r r 1
and γ and r as before. The parameters in the CCC-GARCH models were estimated
by their QMLE using the package ccgarch of the language R. For the distribution of
the innovations, we took ε1 , . . . , εn i.i.d. from the distribution of ε with ε having a (N )
multivariate normal distribution, in order to study the level of the resulting bootstrap
19
test. To assess the power we considered the following heavy-tailed distributions: Tθ ,
the multivariate β-generalized distribution (GNθ ), that coincides with the normal distribution for θ = 2 and has heavy tails for 0 < θ < 2 (Goodman and Kotz, 1973), and
the asymmetric exponential power distribution (AEP ), whereby (X1 , . . . , Xd )> , with
X1 , . . . , Xd i.i.d. from a univariate AEP distribution (Zhu and Zinde-Walsh, 2009)
with parameters α = 0.4, p1 = 1.182 and p2 = 1.820 (these settings gave useful results
in practical applications for the errors in GARCH type models). As in the previous
subsection, we also calculated the HW-test.
Table 3 reports the percentages of rejections for nominal significance level α = 0.05
and sample size n = 300, for r = 0, 0.3 and γ = 0.4. The resulting pictures for
γ = 0.3, 0.5 are quite similar so, to save space, we omit the results for these values of
γ. In order to reduce the computational burden we adopted the warp-speed method of
Giacomini et al. (2013), which works as follows: rather than computing critical points
for each Monte Carlo sample, one resample is generated for each Monte Carlo sample,
and the resampling test statistic is computed for that sample; then the resampling
G
critical values for Tn,β
are computed from the empirical distribution determined by the
G∗
resampling replications of Tn,β
. In our simulations we generated 10, 000 Monte Carlo
samples for the level and 2, 000 for the power. Looking at Table 3, we conclude that:
the actual level of the proposed bootstrap test is very close to the nominal level, and
this is also true for the HW-test (although to the best of our knowledge, the consistency
of the bootstrap null distribution estimator of the HW-test statistic has been proved
only for the univariate case in Jiménez-Gamero, 2014); and with respect to the power,
the proposed test in most cases outperforms the HW-test.
6.3. A real data set application
As an illustration, we consider the monthly log returns of IBM stock and the
S&P 500 index from January 1926 to December 2008 with 888 observations. This
data set was analyzed in Example 10.5 of Tsay(2010), where it is showed that a CCCGARCH(1,1) model provides a adequate description of the data, which is available from
the website http://faculty.chicagobooth.edu/ruey.tsay/teaching/fts/) of the
author. We applied the proposed test and the HW test for testing H0,G . The pvalues were obtained by generating 1000 bootstrap samples. For all values of β in
20
Table 3: Percentage of rejections for nominal level α = 0.05, γ = 0.4 and n = 300.
N
d
r
2
0.0
0.3
0.0
0.3
0.0
0.3
0.0
0.3
0.0
0.3
0.0
0.3
0.0
0.3
0.0
0.3
3
T10
2
3
GN1.65
2
3
AEP
2
3
G
Test based on Tn,β
HW-test
β
β
2.1
2.2
2.3
2.4
2.5
1.0
1.5
2.0
2.5
4.96
4.14
4.54
4.96
61.85
66.95
81.45
78.30
22.40
18.30
17.55
20.00
56.75
52.70
55.40
59.55
4.85
4.33
4.71
4.85
61.20
66.80
80.95
78.05
21.05
17.80
18.40
19.65
55.50
51.20
55.85
59.30
4.81
4.38
4.73
4.81
59.25
65.85
80.15
78.20
20.10
17.70
18.10
19.85
53.35
49.65
54.85
58.75
4.79
4.40
4.74
4.79
57.55
64.15
79.65
77.20
18.95
16.80
17.80
19.80
51.10
47.90
53.75
57.15
4.73
4.27
4.73
4.73
55.50
61.35
78.20
77.15
17.85
16.10
16.90
18.90
49.00
45.75
51.65
57.00
5.06
4.95
4.64
5.06
26.70
20.50
45.75
42.40
8.65
8.00
9.10
9.70
29.55
26.35
38.25
33.15
4.80
5.45
4.64
4.80
36.70
31.70
55.40
55.70
15.20
14.00
14.85
13.95
49.85
45.20
54.25
53.65
4.97
5.36
4.88
4.97
37.20
32.10
50.95
52.85
16.45
16.00
15.35
15.55
52.85
50.00
55.45
53.90
4.82
5.29
4.51
4.82
34.85
30.60
43.80
44.00
16.75
14.30
15.60
15.15
51.45
49.20
49.25
49.70
Table 3 we get the same p-value, 0.000, which leads us to reject H0,G , as expected by
looking at Figure 1, which displays the scatter plot of the residuals after fitting a CCCGARCH(1,1) model to the log returns, and Figure 2, that represents the histograms of
the marginal residuals with the probability density function of a standard normal law
superimposed.
7. Conclusions
We have studied a class of affine invariant tests for multivariate normality both
in an i.i.d. setting and in the context of testing that the innovation distribution of a
multivariate GARCH model is Gaussian, thus generalizing results of Henze and Koch
(2017) in two ways. The test statistics are suitably weighted L2 -statistics based on
the difference between the empirical moment generating function of scaled residuals
of the data and the moment generating function of the standard normal distribution
in Rd . As such, they can be considered as ’moment generating function analogues’ to
21
4
●
●
−4
−2
0
S&P500 returns
2
●
●
● ● ● ● ● ●●
●● ● ● ●
●●
●●
●
●
● ●●● ●●● ●●●
● ●
●
●
●●
●●●
●●
●
●● ●
●
●
●
●
●
●
● ●● ●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●
●●
●●
●
● ●●●
●
●●
●●●●
●
●
●●
●
●
● ●
●
●
●●
●●
●
●
●
●
●
●
● ●●●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●● ●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●● ● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ● ●
●●
●
●●
●
●●●
●●
●
●●
●●
● ●●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●●
●
● ● ●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●
●
●
●
●●
●
●●●
●●
●●●
●● ●
●●●
●
●
●● ● ●
●
●●
●
●●● ●
●
●●
●●
●● ● ●
●●
●
● ●
●
●●
● ●●● ●
●
●● ●
●
●●
●
● ●●
● ●
●
●
●
●
●
●●
●●●● ●
●
● ●●●●●● ● ●●
●●
●
● ● ● ● ●
●
●
● ●
●
●
●
●
●
●
−4
−2
0
2
4
IBM returns
0.0
0.0
0.1
0.1
0.2
0.2
0.3
0.3
0.4
0.4
Figure 1: Scatter plot of the residuals.
−4
−2
0
2
4
−6
IBM returns
−4
−2
0
S&P500 returns
Figure 2: Histograms of the residuals.
22
2
4
the time-honored class of BHEP tests that use the empirical characteristic function.
As the decay of a weight function figuring in the test statistic tends to infinity, the
test statistic approaches a certain linear combination of two well-known measures of
multivariate skewness. The tests are easy to implement, and they turn out to be
consistent against a wide range of alternatives. In contrast to a recently studied L2 statistic of Henze et al. (2017) that uses both the empirical moment generating and
the empirical characteristic function, our test is also feasible for larger sample sizes
since the computational complexity is of order O(n2 ). Regarding power, the new tests
outperform the BHEP-tests against heavy-tailed distributions.
Acknowledgements
M.D. Jiménez-Gamero was partially supported by grant MTM2014-55966-P of the
Spanish Ministry of Economy and Competitiveness.
References
Arcones, M. (2007), Two tests for multivariate normality based on the characteristic function,
Math. Methods Statist., 16, 177–201.
Bardet, J. M. and Wintenberger, O. (2009), Asymptotic normality of the quasi-maximum
likelihood estimator for multidimensional causal processes, Ann. Statist., 37, 2730–
2759.
Baringhaus, L., and Henze, N. (1988), A consistent test for multivariate normality based on
the empirical characteristic function, Metrika, 35, 339–348.
Baringhaus, L., Ebner, B., and Henze, N. (2017), The limit distribution of weighted L2 goodness-of fit statistics under fixed alternatives, with applications, Ann. Inst. Stat.
Math., 69, 969–995.
Batsidis, A., Martin, N., Pardo, L., and Zografos, K. (2013), A necessary power divergence
type family of tests for multivariate normality, Comm. Statist. – Simul. Comput., 42,
2253–2271.
Billingsley, P. (1968), Convergence of Probability Measures. John Wiley & Sons, New YorkLondon-Sidney.
Bollerslev, T. (1990), Modelling the coherence in short-run nominal exchange rates: a multivariate generalized ARCH model, Rev. Econ. Stat., 72, 498–505.
23
Bosq, D. (2000), Linear Processes in Function Spaces. Springer, New York.
Burke M.D. (2000), Multivariate tests-of-fit and uniform confidence bands using a weighted
bootstrap, Statist. Probab. Lett., 46, 13–20.
Cardoso de Oliveira, I.R., and Ferreira, D.F. (2010), Multivariate extension of chi-squared
univariate normality test, J. Statist. Comput. Simul., 80, 513–526.
Comte, F. and Lieberman, O. (2003), Asymptotic theory for multivariate GARCH processes,
J. Multiv. Anal., 84, 61–84.
Csörgő, S. (1989), Consistency of some tests for multivariate normality, Metrika, 36, 107–116.
Eaton, M.L., and Perlman, M.D. (1973), The non-singularity of generalized sample covariance
matrices, Ann. Statist., 1, 710–717.
Ebner, B. (2012), Asymptotic theory for the tests of multivariate normality by Cox and Small,
J. Multiv. Anal., 111, 368–379.
Enomoto, R., Okamoto, N., and Seo, T. (2012) Multivariate normality test using Srivastava’s
skewness and kurtosis, SUT J. Math., 48, 103–115.
Farrel, P.J., Salibian–Barrera, M., and Naczk, K. (2007) On tests for multivariate normality
and associated simulation studies, J. Statist. Comput. Simul., 77, 1053–1068.
Francq, C., Jiménez-Gamero, M.D., and Meintanis, S.G. (2017), Tests for sphericity in multivariate GARCH models, J. Econometrics, 196, 305–319.
Francq, C. and Zakoı̈an, J.M. (2010), GARCH Models: Structure, Statistical Inference and
Applications. Wiley, London.
Francq, C. and Zakoı̈an, J.M. (2012), QML estimation of a class of multivariate asymmetric
GARCH models, Econometric Theory, 28, 179–206.
Ghoudi K., and Rémillard, B. (2014), Comparison of specification tests for GARCH models,
Computat. Statist. Data Anal., 76, 291–300.
Giacomini, R., Politis, D.N., and White, H. (2013), A warp-speed method for conducting
Monte Carlo experiments involving bootstrap estimators, Econometric Theory, 29,
567–589.
Goodman, I.R., Kotz, S. (1973), Multivariate θ-generalized normal distributions, J. Multiv.
Anal., 3, 204–219.
Hanusz, Z., and Tarasińska, J. (2008), A note on Srivastava and Hui’s test of multivariate
normality, J. Multiv. Anal., 99, 2364–2367.
Hanusz, Z., and Tarasińska, J. (2012). New test for multivariate normality based on Small’s
and Srivastava’s graphical methods, J. Statist. Comput. Simul., 82, 1743–1752.
24
Henze, N. (1997), Extreme smoothing and testing for multivariate normality, Statist. Probab.
Lett., 35, 203–213.
Henze, N. (2002), Invariant tests for multivariate normality: a critical review, Statist. Papers,
43, 467–506.
Henze, N., Jiménez-Gamero, M.D., and Meintanis, S.G. (2017), Characterizations of
multinormality and and corresponding tests of fit, including for Garch models,
arXiv:1706.03029.
Henze, N., and Koch, S. (2017), On a test of normality based on the empirical moment
generating function, Statist. Papers, doi:10.1007/s00362-017-0923-7.
Henze, N., and Wagner, T. (1997), A new approach to the BHEP tests for multivariate
normality, J. Multiv. Anal., 62, 1–23.
Henze, N., and Zirkler, B. (1990), A class of invariant consistent tests for multivariate normality, Comm. Statist. Theory Methods, 19, 3595–3617.
Jeantheau, T. (1998), Strong consistency of estimators for multivariate ARCH models, Econometric Theory, 14, 70–86.
Jiménez-Gamero, M.D. (2014), On the empirical characteristic function process of the residuals in GARCH models and applications, TEST, 23, 409–432.
Jiménez-Gamero, M.D., and Pardo-Fernández, J.C. (2017), Empirical characteristic function
tests for GARCH innovation distribution using multipliers, J. Stat. Comput. Simul.,
87, 2069–2093.
Joenssen, D.W., and Vogel, J. (2014), A power study of goodness-of-fit tests for multivariate
normality implemented in R, J. Stat. Comput. Simul., 84, 1055–1078.
Jönsson, K., (2011), A robust test for multivariate normality, Econom. Lett., 113, 199–201.
Kim, N. (2016), A robustified Jarque–Bera test for multivariate normality, Econom. Lett.,
140, 48–52,
Klar, B., Lindner, F. and Meintanis, S.G. (2012), Specification tests for the error distribution
in GARCH models, Comput. Statist. Data Anal., 56, 3587–3598.
Koizumi, K., Hyodo, M., and Pavlenko, T. (2014), Modified Jarque–Bera tests for multivariate normality in a high-dimensional framework, J. Statist. Th. Pract., 8, 382–399.
Mardia, K.V. (1970), Measures of multivariate skewness and kurtosis with applications,
Biometrika, 57, 519–530.
25
Mecklin, Ch. J., and Mundfrom, D.J. (2005), A Monte Carlo comparison of the type I and
type II error rates of tests of multivariate normality, J. Statist. Comput. Simul., 75,
93–107.
Móri, T.F., Rohatgi, V.K., and Székely, G.J. (1993), On multivariate skewness and kurtosis,
Theory Probab. Appl. 38, 547–551.
Pudelko, J. (2005), On a new affine invariant and consistent test for multivariate normality,
Probab. Math. Statist., 25, 43–54.
Rydberg, T.H. (2000), Realistic statistical modelling of financial data, Int. Stat. Rev., 68,
233–258.
Shorack, G. R., and Wellner, J. A. (1986), Empirical processes with applications to statistics,
Wiley Series in Probability and Mathematical Statistics. John Wiley & Sons, Inc., New
York, 1986.
Spierdijk, L. (2016), Confidence intervals for ARMA-GARCH Value-at-Risk: The case of
heavy tails and skewness, Computat. Statist. Data Anal., 100, 545–559.
Székeley, G.J., and Rizzo, M.L. (2005), A new test for multivariate normality, J. Multiv.
Anal., 93, 58–80.
Tenreiro, C. (2011), An affine invariant multiple test procedure for assessing multivariate
normality, Comput. Statist. Data Anal., 55, 1980–1992.
Tenreiro, C. (2017), A new test for multivariate normality by combining extreme and nonextreme BHEP tests, Commun. Statist. –Th. Meth., 46, 1746–1759.
Thulin, M. (2014), Tests for multivariate normality based on canonical correlations, Stat.
Meth. Appl., 23, 189–208.
Tsay, R.S. (2010). Analysis of Financial Time Series. Wiley, Hoboken, New Jersey.
Villaseñor-Alva, J.A., and Estrada, E.G. (2009), A generalization of Shapiro–Wilk’s test for
multivariate normality, Commun. Statist. – Th. Meth., 38, 1870–1883.
Voinov, V., Pya, N., Makarov, R. and Voinov, Y. (2016), New invariant and consistent chisquared type goodness-of-fit tests for multivariate normality and a related comparative
simulation study, Commun. Statist. – Th. Meth. 45, 3249–3263.
Yanada, T., Romer, M.M., and Richards, D. St. P. (2015), Kurtosis tests for multivariate
normality with monotone incomplete data, TEST 24, 532–557.
Zghoul, A.A. (2010), A goodness-of-fit test for normality based on the empirical moment
generating function, Comm. Statist. – Simul. Comput., 39, 1292–1304.
26
Zhou, S., and Shao, Y. (2014), A powerful test for multivariate normality, J. Appl. Statist.,
41, 351–363.
Zhu, D., and Zinde-Walsh, V. (2009), Properties and estimation of asymmetric exponential
power distribution. J. Econometrics, 148, 86–99.
27
| 10math.ST
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.